diff --git "a/falcon_edge_3b_lm_head.mlmodelc/model.mil" "b/falcon_edge_3b_lm_head.mlmodelc/model.mil"
new file mode 100644--- /dev/null
+++ "b/falcon_edge_3b_lm_head.mlmodelc/model.mil"
@@ -0,0 +1,6084 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3405.2.1"}, {"coremlc-version", "3405.2.1"}})]
+{
+    func lm_head_length_1<ios18>(tensor<fp16, [1, 2048, 1, 1]> hidden_states) {
+            tensor<fp16, [1, 2048, 1, 1]> final_norm_rmsnorm_abs = abs(x = hidden_states)[name = string("final_norm_rmsnorm_abs")];
+            tensor<int32, [1]> final_norm_rmsnorm_maxval_axes_1 = const()[name = string("final_norm_rmsnorm_maxval_axes_1"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_maxval_keep_dims_1 = const()[name = string("final_norm_rmsnorm_maxval_keep_dims_1"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> final_norm_rmsnorm_maxval = reduce_max(axes = final_norm_rmsnorm_maxval_axes_1, keep_dims = final_norm_rmsnorm_maxval_keep_dims_1, x = final_norm_rmsnorm_abs)[name = string("final_norm_rmsnorm_maxval")];
+            fp16 final_norm_rmsnorm_maxval_clipped_alpha_1 = const()[name = string("final_norm_rmsnorm_maxval_clipped_alpha_1"), val = fp16(0x1p-14)];
+            fp16 final_norm_rmsnorm_maxval_clipped_beta_1 = const()[name = string("final_norm_rmsnorm_maxval_clipped_beta_1"), val = fp16(inf)];
+            tensor<fp16, [1, 1, 1, 1]> final_norm_rmsnorm_maxval_clipped = clip(alpha = final_norm_rmsnorm_maxval_clipped_alpha_1, beta = final_norm_rmsnorm_maxval_clipped_beta_1, x = final_norm_rmsnorm_maxval)[name = string("final_norm_rmsnorm_maxval_clipped")];
+            tensor<fp16, [1, 2048, 1, 1]> final_norm_rmsnorm_scaled = real_div(x = hidden_states, y = final_norm_rmsnorm_maxval_clipped)[name = string("final_norm_rmsnorm_scaled")];
+            tensor<int32, [1]> final_norm_rmsnorm_squared_sum_axes_1 = const()[name = string("final_norm_rmsnorm_squared_sum_axes_1"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_squared_sum_keep_dims_1 = const()[name = string("final_norm_rmsnorm_squared_sum_keep_dims_1"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> final_norm_rmsnorm_squared_sum = reduce_sum_square(axes = final_norm_rmsnorm_squared_sum_axes_1, keep_dims = final_norm_rmsnorm_squared_sum_keep_dims_1, x = final_norm_rmsnorm_scaled)[name = string("final_norm_rmsnorm_squared_sum")];
+            fp16 final_norm_rmsnorm_rsqrt_epsilon_1 = const()[name = string("final_norm_rmsnorm_rsqrt_epsilon_1"), val = fp16(0x1p-14)];
+            tensor<fp16, [1, 1, 1, 1]> final_norm_rmsnorm_rsqrt = rsqrt(epsilon = final_norm_rmsnorm_rsqrt_epsilon_1, x = final_norm_rmsnorm_squared_sum)[name = string("final_norm_rmsnorm_rsqrt")];
+            fp16 final_norm_rmsnorm_dim_scaled_y_1 = const()[name = string("final_norm_rmsnorm_dim_scaled_y_1"), val = fp16(0x1.6ap+5)];
+            tensor<fp16, [1, 2048, 1, 1]> final_norm_rmsnorm_dim_scaled = mul(x = final_norm_rmsnorm_scaled, y = final_norm_rmsnorm_dim_scaled_y_1)[name = string("final_norm_rmsnorm_dim_scaled")];
+            tensor<fp16, [1, 2048, 1, 1]> final_norm_rmsnorm_normalized = mul(x = final_norm_rmsnorm_dim_scaled, y = final_norm_rmsnorm_rsqrt)[name = string("final_norm_rmsnorm_normalized")];
+            tensor<fp16, [2048, 1, 1]> final_norm_rmsnorm_y_1 = const()[name = string("final_norm_rmsnorm_y_1"), val = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 2048, 1, 1]> final_norm_rmsnorm = mul(x = final_norm_rmsnorm_normalized, y = final_norm_rmsnorm_y_1)[name = string("final_norm_rmsnorm")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_0_weight_1 = const()[name = string("logits_chunk_0_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4224)))];
+            tensor<int32, [2]> logits_chunk_0_strides_1 = const()[name = string("logits_chunk_0_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_0_pad_type_1 = const()[name = string("logits_chunk_0_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_0_pad_1 = const()[name = string("logits_chunk_0_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_0_dilations_1 = const()[name = string("logits_chunk_0_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_0_groups_1 = const()[name = string("logits_chunk_0_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_0 = conv(dilations = logits_chunk_0_dilations_1, groups = logits_chunk_0_groups_1, pad = logits_chunk_0_pad_1, pad_type = logits_chunk_0_pad_type_1, strides = logits_chunk_0_strides_1, weight = logits_chunk_0_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_0")];
+            int32 argmax_chunk_0_axis_0 = const()[name = string("argmax_chunk_0_axis_0"), val = int32(1)];
+            bool argmax_chunk_0_keep_dims_0 = const()[name = string("argmax_chunk_0_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_0_output_dtype_0 = const()[name = string("argmax_chunk_0_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> argmax_chunk_0 = reduce_argmax(axis = argmax_chunk_0_axis_0, keep_dims = argmax_chunk_0_keep_dims_0, output_dtype = argmax_chunk_0_output_dtype_0, x = logits_chunk_0)[name = string("argmax_chunk_0")];
+            tensor<int32, [1]> max_chunk_0_axes_0 = const()[name = string("max_chunk_0_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_0_keep_dims_0 = const()[name = string("max_chunk_0_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> max_chunk_0 = reduce_max(axes = max_chunk_0_axes_0, keep_dims = max_chunk_0_keep_dims_0, x = logits_chunk_0)[name = string("max_chunk_0")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_1_weight_1 = const()[name = string("logits_chunk_1_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8392896)))];
+            tensor<int32, [2]> logits_chunk_1_strides_1 = const()[name = string("logits_chunk_1_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_1_pad_type_1 = const()[name = string("logits_chunk_1_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_1_pad_1 = const()[name = string("logits_chunk_1_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_1_dilations_1 = const()[name = string("logits_chunk_1_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_1_groups_1 = const()[name = string("logits_chunk_1_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_1 = conv(dilations = logits_chunk_1_dilations_1, groups = logits_chunk_1_groups_1, pad = logits_chunk_1_pad_1, pad_type = logits_chunk_1_pad_type_1, strides = logits_chunk_1_strides_1, weight = logits_chunk_1_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_1")];
+            int32 argmax_chunk_1_axis_0 = const()[name = string("argmax_chunk_1_axis_0"), val = int32(1)];
+            bool argmax_chunk_1_keep_dims_0 = const()[name = string("argmax_chunk_1_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_1_output_dtype_0 = const()[name = string("argmax_chunk_1_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> argmax_chunk_1 = reduce_argmax(axis = argmax_chunk_1_axis_0, keep_dims = argmax_chunk_1_keep_dims_0, output_dtype = argmax_chunk_1_output_dtype_0, x = logits_chunk_1)[name = string("argmax_chunk_1")];
+            tensor<int32, [1]> max_chunk_1_axes_0 = const()[name = string("max_chunk_1_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_1_keep_dims_0 = const()[name = string("max_chunk_1_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> max_chunk_1 = reduce_max(axes = max_chunk_1_axes_0, keep_dims = max_chunk_1_keep_dims_0, x = logits_chunk_1)[name = string("max_chunk_1")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_2_weight_1 = const()[name = string("logits_chunk_2_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16781568)))];
+            tensor<int32, [2]> logits_chunk_2_strides_1 = const()[name = string("logits_chunk_2_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_2_pad_type_1 = const()[name = string("logits_chunk_2_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_2_pad_1 = const()[name = string("logits_chunk_2_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_2_dilations_1 = const()[name = string("logits_chunk_2_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_2_groups_1 = const()[name = string("logits_chunk_2_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_2 = conv(dilations = logits_chunk_2_dilations_1, groups = logits_chunk_2_groups_1, pad = logits_chunk_2_pad_1, pad_type = logits_chunk_2_pad_type_1, strides = logits_chunk_2_strides_1, weight = logits_chunk_2_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_2")];
+            int32 argmax_chunk_2_axis_0 = const()[name = string("argmax_chunk_2_axis_0"), val = int32(1)];
+            bool argmax_chunk_2_keep_dims_0 = const()[name = string("argmax_chunk_2_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_2_output_dtype_0 = const()[name = string("argmax_chunk_2_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> argmax_chunk_2 = reduce_argmax(axis = argmax_chunk_2_axis_0, keep_dims = argmax_chunk_2_keep_dims_0, output_dtype = argmax_chunk_2_output_dtype_0, x = logits_chunk_2)[name = string("argmax_chunk_2")];
+            tensor<int32, [1]> max_chunk_2_axes_0 = const()[name = string("max_chunk_2_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_2_keep_dims_0 = const()[name = string("max_chunk_2_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> max_chunk_2 = reduce_max(axes = max_chunk_2_axes_0, keep_dims = max_chunk_2_keep_dims_0, x = logits_chunk_2)[name = string("max_chunk_2")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_3_weight_1 = const()[name = string("logits_chunk_3_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25170240)))];
+            tensor<int32, [2]> logits_chunk_3_strides_1 = const()[name = string("logits_chunk_3_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_3_pad_type_1 = const()[name = string("logits_chunk_3_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_3_pad_1 = const()[name = string("logits_chunk_3_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_3_dilations_1 = const()[name = string("logits_chunk_3_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_3_groups_1 = const()[name = string("logits_chunk_3_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_3 = conv(dilations = logits_chunk_3_dilations_1, groups = logits_chunk_3_groups_1, pad = logits_chunk_3_pad_1, pad_type = logits_chunk_3_pad_type_1, strides = logits_chunk_3_strides_1, weight = logits_chunk_3_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_3")];
+            int32 argmax_chunk_3_axis_0 = const()[name = string("argmax_chunk_3_axis_0"), val = int32(1)];
+            bool argmax_chunk_3_keep_dims_0 = const()[name = string("argmax_chunk_3_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_3_output_dtype_0 = const()[name = string("argmax_chunk_3_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> argmax_chunk_3 = reduce_argmax(axis = argmax_chunk_3_axis_0, keep_dims = argmax_chunk_3_keep_dims_0, output_dtype = argmax_chunk_3_output_dtype_0, x = logits_chunk_3)[name = string("argmax_chunk_3")];
+            tensor<int32, [1]> max_chunk_3_axes_0 = const()[name = string("max_chunk_3_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_3_keep_dims_0 = const()[name = string("max_chunk_3_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> max_chunk_3 = reduce_max(axes = max_chunk_3_axes_0, keep_dims = max_chunk_3_keep_dims_0, x = logits_chunk_3)[name = string("max_chunk_3")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_4_weight_1 = const()[name = string("logits_chunk_4_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33558912)))];
+            tensor<int32, [2]> logits_chunk_4_strides_1 = const()[name = string("logits_chunk_4_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_4_pad_type_1 = const()[name = string("logits_chunk_4_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_4_pad_1 = const()[name = string("logits_chunk_4_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_4_dilations_1 = const()[name = string("logits_chunk_4_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_4_groups_1 = const()[name = string("logits_chunk_4_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_4 = conv(dilations = logits_chunk_4_dilations_1, groups = logits_chunk_4_groups_1, pad = logits_chunk_4_pad_1, pad_type = logits_chunk_4_pad_type_1, strides = logits_chunk_4_strides_1, weight = logits_chunk_4_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_4")];
+            int32 argmax_chunk_4_axis_0 = const()[name = string("argmax_chunk_4_axis_0"), val = int32(1)];
+            bool argmax_chunk_4_keep_dims_0 = const()[name = string("argmax_chunk_4_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_4_output_dtype_0 = const()[name = string("argmax_chunk_4_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> argmax_chunk_4 = reduce_argmax(axis = argmax_chunk_4_axis_0, keep_dims = argmax_chunk_4_keep_dims_0, output_dtype = argmax_chunk_4_output_dtype_0, x = logits_chunk_4)[name = string("argmax_chunk_4")];
+            tensor<int32, [1]> max_chunk_4_axes_0 = const()[name = string("max_chunk_4_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_4_keep_dims_0 = const()[name = string("max_chunk_4_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> max_chunk_4 = reduce_max(axes = max_chunk_4_axes_0, keep_dims = max_chunk_4_keep_dims_0, x = logits_chunk_4)[name = string("max_chunk_4")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_5_weight_1 = const()[name = string("logits_chunk_5_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41947584)))];
+            tensor<int32, [2]> logits_chunk_5_strides_1 = const()[name = string("logits_chunk_5_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_5_pad_type_1 = const()[name = string("logits_chunk_5_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_5_pad_1 = const()[name = string("logits_chunk_5_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_5_dilations_1 = const()[name = string("logits_chunk_5_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_5_groups_1 = const()[name = string("logits_chunk_5_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_5 = conv(dilations = logits_chunk_5_dilations_1, groups = logits_chunk_5_groups_1, pad = logits_chunk_5_pad_1, pad_type = logits_chunk_5_pad_type_1, strides = logits_chunk_5_strides_1, weight = logits_chunk_5_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_5")];
+            int32 argmax_chunk_5_axis_0 = const()[name = string("argmax_chunk_5_axis_0"), val = int32(1)];
+            bool argmax_chunk_5_keep_dims_0 = const()[name = string("argmax_chunk_5_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_5_output_dtype_0 = const()[name = string("argmax_chunk_5_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> argmax_chunk_5 = reduce_argmax(axis = argmax_chunk_5_axis_0, keep_dims = argmax_chunk_5_keep_dims_0, output_dtype = argmax_chunk_5_output_dtype_0, x = logits_chunk_5)[name = string("argmax_chunk_5")];
+            tensor<int32, [1]> max_chunk_5_axes_0 = const()[name = string("max_chunk_5_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_5_keep_dims_0 = const()[name = string("max_chunk_5_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> max_chunk_5 = reduce_max(axes = max_chunk_5_axes_0, keep_dims = max_chunk_5_keep_dims_0, x = logits_chunk_5)[name = string("max_chunk_5")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_6_weight_1 = const()[name = string("logits_chunk_6_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50336256)))];
+            tensor<int32, [2]> logits_chunk_6_strides_1 = const()[name = string("logits_chunk_6_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_6_pad_type_1 = const()[name = string("logits_chunk_6_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_6_pad_1 = const()[name = string("logits_chunk_6_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_6_dilations_1 = const()[name = string("logits_chunk_6_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_6_groups_1 = const()[name = string("logits_chunk_6_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_6 = conv(dilations = logits_chunk_6_dilations_1, groups = logits_chunk_6_groups_1, pad = logits_chunk_6_pad_1, pad_type = logits_chunk_6_pad_type_1, strides = logits_chunk_6_strides_1, weight = logits_chunk_6_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_6")];
+            int32 argmax_chunk_6_axis_0 = const()[name = string("argmax_chunk_6_axis_0"), val = int32(1)];
+            bool argmax_chunk_6_keep_dims_0 = const()[name = string("argmax_chunk_6_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_6_output_dtype_0 = const()[name = string("argmax_chunk_6_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> argmax_chunk_6 = reduce_argmax(axis = argmax_chunk_6_axis_0, keep_dims = argmax_chunk_6_keep_dims_0, output_dtype = argmax_chunk_6_output_dtype_0, x = logits_chunk_6)[name = string("argmax_chunk_6")];
+            tensor<int32, [1]> max_chunk_6_axes_0 = const()[name = string("max_chunk_6_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_6_keep_dims_0 = const()[name = string("max_chunk_6_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> max_chunk_6 = reduce_max(axes = max_chunk_6_axes_0, keep_dims = max_chunk_6_keep_dims_0, x = logits_chunk_6)[name = string("max_chunk_6")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_7_weight_1 = const()[name = string("logits_chunk_7_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58724928)))];
+            tensor<int32, [2]> logits_chunk_7_strides_1 = const()[name = string("logits_chunk_7_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_7_pad_type_1 = const()[name = string("logits_chunk_7_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_7_pad_1 = const()[name = string("logits_chunk_7_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_7_dilations_1 = const()[name = string("logits_chunk_7_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_7_groups_1 = const()[name = string("logits_chunk_7_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_7 = conv(dilations = logits_chunk_7_dilations_1, groups = logits_chunk_7_groups_1, pad = logits_chunk_7_pad_1, pad_type = logits_chunk_7_pad_type_1, strides = logits_chunk_7_strides_1, weight = logits_chunk_7_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_7")];
+            int32 argmax_chunk_7_axis_0 = const()[name = string("argmax_chunk_7_axis_0"), val = int32(1)];
+            bool argmax_chunk_7_keep_dims_0 = const()[name = string("argmax_chunk_7_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_7_output_dtype_0 = const()[name = string("argmax_chunk_7_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> argmax_chunk_7 = reduce_argmax(axis = argmax_chunk_7_axis_0, keep_dims = argmax_chunk_7_keep_dims_0, output_dtype = argmax_chunk_7_output_dtype_0, x = logits_chunk_7)[name = string("argmax_chunk_7")];
+            tensor<int32, [1]> max_chunk_7_axes_0 = const()[name = string("max_chunk_7_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_7_keep_dims_0 = const()[name = string("max_chunk_7_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> max_chunk_7 = reduce_max(axes = max_chunk_7_axes_0, keep_dims = max_chunk_7_keep_dims_0, x = logits_chunk_7)[name = string("max_chunk_7")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_8_weight_1 = const()[name = string("logits_chunk_8_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67113600)))];
+            tensor<int32, [2]> logits_chunk_8_strides_1 = const()[name = string("logits_chunk_8_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_8_pad_type_1 = const()[name = string("logits_chunk_8_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_8_pad_1 = const()[name = string("logits_chunk_8_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_8_dilations_1 = const()[name = string("logits_chunk_8_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_8_groups_1 = const()[name = string("logits_chunk_8_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_8 = conv(dilations = logits_chunk_8_dilations_1, groups = logits_chunk_8_groups_1, pad = logits_chunk_8_pad_1, pad_type = logits_chunk_8_pad_type_1, strides = logits_chunk_8_strides_1, weight = logits_chunk_8_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_8")];
+            int32 argmax_chunk_8_axis_0 = const()[name = string("argmax_chunk_8_axis_0"), val = int32(1)];
+            bool argmax_chunk_8_keep_dims_0 = const()[name = string("argmax_chunk_8_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_8_output_dtype_0 = const()[name = string("argmax_chunk_8_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> argmax_chunk_8 = reduce_argmax(axis = argmax_chunk_8_axis_0, keep_dims = argmax_chunk_8_keep_dims_0, output_dtype = argmax_chunk_8_output_dtype_0, x = logits_chunk_8)[name = string("argmax_chunk_8")];
+            tensor<int32, [1]> max_chunk_8_axes_0 = const()[name = string("max_chunk_8_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_8_keep_dims_0 = const()[name = string("max_chunk_8_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> max_chunk_8 = reduce_max(axes = max_chunk_8_axes_0, keep_dims = max_chunk_8_keep_dims_0, x = logits_chunk_8)[name = string("max_chunk_8")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_9_weight_1 = const()[name = string("logits_chunk_9_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75502272)))];
+            tensor<int32, [2]> logits_chunk_9_strides_1 = const()[name = string("logits_chunk_9_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_9_pad_type_1 = const()[name = string("logits_chunk_9_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_9_pad_1 = const()[name = string("logits_chunk_9_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_9_dilations_1 = const()[name = string("logits_chunk_9_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_9_groups_1 = const()[name = string("logits_chunk_9_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_9 = conv(dilations = logits_chunk_9_dilations_1, groups = logits_chunk_9_groups_1, pad = logits_chunk_9_pad_1, pad_type = logits_chunk_9_pad_type_1, strides = logits_chunk_9_strides_1, weight = logits_chunk_9_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_9")];
+            int32 argmax_chunk_9_axis_0 = const()[name = string("argmax_chunk_9_axis_0"), val = int32(1)];
+            bool argmax_chunk_9_keep_dims_0 = const()[name = string("argmax_chunk_9_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_9_output_dtype_0 = const()[name = string("argmax_chunk_9_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> argmax_chunk_9 = reduce_argmax(axis = argmax_chunk_9_axis_0, keep_dims = argmax_chunk_9_keep_dims_0, output_dtype = argmax_chunk_9_output_dtype_0, x = logits_chunk_9)[name = string("argmax_chunk_9")];
+            tensor<int32, [1]> max_chunk_9_axes_0 = const()[name = string("max_chunk_9_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_9_keep_dims_0 = const()[name = string("max_chunk_9_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> max_chunk_9 = reduce_max(axes = max_chunk_9_axes_0, keep_dims = max_chunk_9_keep_dims_0, x = logits_chunk_9)[name = string("max_chunk_9")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_10_weight_1 = const()[name = string("logits_chunk_10_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83890944)))];
+            tensor<int32, [2]> logits_chunk_10_strides_1 = const()[name = string("logits_chunk_10_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_10_pad_type_1 = const()[name = string("logits_chunk_10_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_10_pad_1 = const()[name = string("logits_chunk_10_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_10_dilations_1 = const()[name = string("logits_chunk_10_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_10_groups_1 = const()[name = string("logits_chunk_10_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_10 = conv(dilations = logits_chunk_10_dilations_1, groups = logits_chunk_10_groups_1, pad = logits_chunk_10_pad_1, pad_type = logits_chunk_10_pad_type_1, strides = logits_chunk_10_strides_1, weight = logits_chunk_10_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_10")];
+            int32 argmax_chunk_10_axis_0 = const()[name = string("argmax_chunk_10_axis_0"), val = int32(1)];
+            bool argmax_chunk_10_keep_dims_0 = const()[name = string("argmax_chunk_10_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_10_output_dtype_0 = const()[name = string("argmax_chunk_10_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> argmax_chunk_10 = reduce_argmax(axis = argmax_chunk_10_axis_0, keep_dims = argmax_chunk_10_keep_dims_0, output_dtype = argmax_chunk_10_output_dtype_0, x = logits_chunk_10)[name = string("argmax_chunk_10")];
+            tensor<int32, [1]> max_chunk_10_axes_0 = const()[name = string("max_chunk_10_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_10_keep_dims_0 = const()[name = string("max_chunk_10_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> max_chunk_10 = reduce_max(axes = max_chunk_10_axes_0, keep_dims = max_chunk_10_keep_dims_0, x = logits_chunk_10)[name = string("max_chunk_10")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_11_weight_1 = const()[name = string("logits_chunk_11_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92279616)))];
+            tensor<int32, [2]> logits_chunk_11_strides_1 = const()[name = string("logits_chunk_11_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_11_pad_type_1 = const()[name = string("logits_chunk_11_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_11_pad_1 = const()[name = string("logits_chunk_11_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_11_dilations_1 = const()[name = string("logits_chunk_11_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_11_groups_1 = const()[name = string("logits_chunk_11_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_11 = conv(dilations = logits_chunk_11_dilations_1, groups = logits_chunk_11_groups_1, pad = logits_chunk_11_pad_1, pad_type = logits_chunk_11_pad_type_1, strides = logits_chunk_11_strides_1, weight = logits_chunk_11_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_11")];
+            int32 argmax_chunk_11_axis_0 = const()[name = string("argmax_chunk_11_axis_0"), val = int32(1)];
+            bool argmax_chunk_11_keep_dims_0 = const()[name = string("argmax_chunk_11_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_11_output_dtype_0 = const()[name = string("argmax_chunk_11_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> argmax_chunk_11 = reduce_argmax(axis = argmax_chunk_11_axis_0, keep_dims = argmax_chunk_11_keep_dims_0, output_dtype = argmax_chunk_11_output_dtype_0, x = logits_chunk_11)[name = string("argmax_chunk_11")];
+            tensor<int32, [1]> max_chunk_11_axes_0 = const()[name = string("max_chunk_11_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_11_keep_dims_0 = const()[name = string("max_chunk_11_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> max_chunk_11 = reduce_max(axes = max_chunk_11_axes_0, keep_dims = max_chunk_11_keep_dims_0, x = logits_chunk_11)[name = string("max_chunk_11")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_12_weight_1 = const()[name = string("logits_chunk_12_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100668288)))];
+            tensor<int32, [2]> logits_chunk_12_strides_1 = const()[name = string("logits_chunk_12_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_12_pad_type_1 = const()[name = string("logits_chunk_12_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_12_pad_1 = const()[name = string("logits_chunk_12_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_12_dilations_1 = const()[name = string("logits_chunk_12_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_12_groups_1 = const()[name = string("logits_chunk_12_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_12 = conv(dilations = logits_chunk_12_dilations_1, groups = logits_chunk_12_groups_1, pad = logits_chunk_12_pad_1, pad_type = logits_chunk_12_pad_type_1, strides = logits_chunk_12_strides_1, weight = logits_chunk_12_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_12")];
+            int32 argmax_chunk_12_axis_0 = const()[name = string("argmax_chunk_12_axis_0"), val = int32(1)];
+            bool argmax_chunk_12_keep_dims_0 = const()[name = string("argmax_chunk_12_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_12_output_dtype_0 = const()[name = string("argmax_chunk_12_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> argmax_chunk_12 = reduce_argmax(axis = argmax_chunk_12_axis_0, keep_dims = argmax_chunk_12_keep_dims_0, output_dtype = argmax_chunk_12_output_dtype_0, x = logits_chunk_12)[name = string("argmax_chunk_12")];
+            tensor<int32, [1]> max_chunk_12_axes_0 = const()[name = string("max_chunk_12_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_12_keep_dims_0 = const()[name = string("max_chunk_12_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> max_chunk_12 = reduce_max(axes = max_chunk_12_axes_0, keep_dims = max_chunk_12_keep_dims_0, x = logits_chunk_12)[name = string("max_chunk_12")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_13_weight_1 = const()[name = string("logits_chunk_13_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109056960)))];
+            tensor<int32, [2]> logits_chunk_13_strides_1 = const()[name = string("logits_chunk_13_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_13_pad_type_1 = const()[name = string("logits_chunk_13_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_13_pad_1 = const()[name = string("logits_chunk_13_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_13_dilations_1 = const()[name = string("logits_chunk_13_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_13_groups_1 = const()[name = string("logits_chunk_13_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_13 = conv(dilations = logits_chunk_13_dilations_1, groups = logits_chunk_13_groups_1, pad = logits_chunk_13_pad_1, pad_type = logits_chunk_13_pad_type_1, strides = logits_chunk_13_strides_1, weight = logits_chunk_13_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_13")];
+            int32 argmax_chunk_13_axis_0 = const()[name = string("argmax_chunk_13_axis_0"), val = int32(1)];
+            bool argmax_chunk_13_keep_dims_0 = const()[name = string("argmax_chunk_13_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_13_output_dtype_0 = const()[name = string("argmax_chunk_13_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> argmax_chunk_13 = reduce_argmax(axis = argmax_chunk_13_axis_0, keep_dims = argmax_chunk_13_keep_dims_0, output_dtype = argmax_chunk_13_output_dtype_0, x = logits_chunk_13)[name = string("argmax_chunk_13")];
+            tensor<int32, [1]> max_chunk_13_axes_0 = const()[name = string("max_chunk_13_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_13_keep_dims_0 = const()[name = string("max_chunk_13_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> max_chunk_13 = reduce_max(axes = max_chunk_13_axes_0, keep_dims = max_chunk_13_keep_dims_0, x = logits_chunk_13)[name = string("max_chunk_13")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_14_weight_1 = const()[name = string("logits_chunk_14_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117445632)))];
+            tensor<int32, [2]> logits_chunk_14_strides_1 = const()[name = string("logits_chunk_14_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_14_pad_type_1 = const()[name = string("logits_chunk_14_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_14_pad_1 = const()[name = string("logits_chunk_14_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_14_dilations_1 = const()[name = string("logits_chunk_14_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_14_groups_1 = const()[name = string("logits_chunk_14_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_14 = conv(dilations = logits_chunk_14_dilations_1, groups = logits_chunk_14_groups_1, pad = logits_chunk_14_pad_1, pad_type = logits_chunk_14_pad_type_1, strides = logits_chunk_14_strides_1, weight = logits_chunk_14_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_14")];
+            int32 argmax_chunk_14_axis_0 = const()[name = string("argmax_chunk_14_axis_0"), val = int32(1)];
+            bool argmax_chunk_14_keep_dims_0 = const()[name = string("argmax_chunk_14_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_14_output_dtype_0 = const()[name = string("argmax_chunk_14_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> argmax_chunk_14 = reduce_argmax(axis = argmax_chunk_14_axis_0, keep_dims = argmax_chunk_14_keep_dims_0, output_dtype = argmax_chunk_14_output_dtype_0, x = logits_chunk_14)[name = string("argmax_chunk_14")];
+            tensor<int32, [1]> max_chunk_14_axes_0 = const()[name = string("max_chunk_14_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_14_keep_dims_0 = const()[name = string("max_chunk_14_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> max_chunk_14 = reduce_max(axes = max_chunk_14_axes_0, keep_dims = max_chunk_14_keep_dims_0, x = logits_chunk_14)[name = string("max_chunk_14")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_15_weight_1 = const()[name = string("logits_chunk_15_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125834304)))];
+            tensor<int32, [2]> logits_chunk_15_strides_1 = const()[name = string("logits_chunk_15_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_15_pad_type_1 = const()[name = string("logits_chunk_15_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_15_pad_1 = const()[name = string("logits_chunk_15_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_15_dilations_1 = const()[name = string("logits_chunk_15_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_15_groups_1 = const()[name = string("logits_chunk_15_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_15 = conv(dilations = logits_chunk_15_dilations_1, groups = logits_chunk_15_groups_1, pad = logits_chunk_15_pad_1, pad_type = logits_chunk_15_pad_type_1, strides = logits_chunk_15_strides_1, weight = logits_chunk_15_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_15")];
+            int32 argmax_chunk_15_axis_0 = const()[name = string("argmax_chunk_15_axis_0"), val = int32(1)];
+            bool argmax_chunk_15_keep_dims_0 = const()[name = string("argmax_chunk_15_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_15_output_dtype_0 = const()[name = string("argmax_chunk_15_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> argmax_chunk_15 = reduce_argmax(axis = argmax_chunk_15_axis_0, keep_dims = argmax_chunk_15_keep_dims_0, output_dtype = argmax_chunk_15_output_dtype_0, x = logits_chunk_15)[name = string("argmax_chunk_15")];
+            tensor<int32, [1]> max_chunk_15_axes_0 = const()[name = string("max_chunk_15_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_15_keep_dims_0 = const()[name = string("max_chunk_15_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> max_chunk_15 = reduce_max(axes = max_chunk_15_axes_0, keep_dims = max_chunk_15_keep_dims_0, x = logits_chunk_15)[name = string("max_chunk_15")];
+            int32 logits_axis_0 = const()[name = string("logits_axis_0"), val = int32(1)];
+            bool logits_interleave_0 = const()[name = string("logits_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 32768, 1, 1]> logits = concat(axis = logits_axis_0, interleave = logits_interleave_0, values = (logits_chunk_0, logits_chunk_1, logits_chunk_2, logits_chunk_3, logits_chunk_4, logits_chunk_5, logits_chunk_6, logits_chunk_7, logits_chunk_8, logits_chunk_9, logits_chunk_10, logits_chunk_11, logits_chunk_12, logits_chunk_13, logits_chunk_14, logits_chunk_15))[name = string("logits")];
+            int32 values_axis_0 = const()[name = string("values_axis_0"), val = int32(1)];
+            bool values_interleave_0 = const()[name = string("values_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 1]> values = concat(axis = values_axis_0, interleave = values_interleave_0, values = (max_chunk_0, max_chunk_1, max_chunk_2, max_chunk_3, max_chunk_4, max_chunk_5, max_chunk_6, max_chunk_7, max_chunk_8, max_chunk_9, max_chunk_10, max_chunk_11, max_chunk_12, max_chunk_13, max_chunk_14, max_chunk_15))[name = string("values")];
+            tensor<int32, [1]> max_value_axes_0 = const()[name = string("max_value_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_value_keep_dims_0 = const()[name = string("max_value_keep_dims_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 1]> max_value = reduce_max(axes = max_value_axes_0, keep_dims = max_value_keep_dims_0, x = values)[name = string("max_value")];
+            int32 max_value_index_axis_0 = const()[name = string("max_value_index_axis_0"), val = int32(1)];
+            bool max_value_index_keep_dims_0 = const()[name = string("max_value_index_keep_dims_0"), val = bool(true)];
+            string max_value_index_output_dtype_0 = const()[name = string("max_value_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 1, 1, 1]> max_value_index = reduce_argmax(axis = max_value_index_axis_0, keep_dims = max_value_index_keep_dims_0, output_dtype = max_value_index_output_dtype_0, x = values)[name = string("max_value_index")];
+            string indices_chunk_0_int32_dtype_1 = const()[name = string("indices_chunk_0_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_1_int32_dtype_1 = const()[name = string("indices_chunk_1_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_2_int32_dtype_1 = const()[name = string("indices_chunk_2_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_3_int32_dtype_1 = const()[name = string("indices_chunk_3_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_4_int32_dtype_1 = const()[name = string("indices_chunk_4_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_5_int32_dtype_1 = const()[name = string("indices_chunk_5_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_6_int32_dtype_1 = const()[name = string("indices_chunk_6_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_7_int32_dtype_1 = const()[name = string("indices_chunk_7_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_8_int32_dtype_1 = const()[name = string("indices_chunk_8_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_9_int32_dtype_1 = const()[name = string("indices_chunk_9_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_10_int32_dtype_1 = const()[name = string("indices_chunk_10_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_11_int32_dtype_1 = const()[name = string("indices_chunk_11_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_12_int32_dtype_1 = const()[name = string("indices_chunk_12_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_13_int32_dtype_1 = const()[name = string("indices_chunk_13_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_14_int32_dtype_1 = const()[name = string("indices_chunk_14_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_15_int32_dtype_1 = const()[name = string("indices_chunk_15_int32_dtype_1"), val = string("int32")];
+            int32 indices_axis_1 = const()[name = string("indices_axis_1"), val = int32(1)];
+            bool indices_interleave_1 = const()[name = string("indices_interleave_1"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_15_int32 = cast(dtype = indices_chunk_15_int32_dtype_1, x = argmax_chunk_15)[name = string("cast_16")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_14_int32 = cast(dtype = indices_chunk_14_int32_dtype_1, x = argmax_chunk_14)[name = string("cast_17")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_13_int32 = cast(dtype = indices_chunk_13_int32_dtype_1, x = argmax_chunk_13)[name = string("cast_18")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_12_int32 = cast(dtype = indices_chunk_12_int32_dtype_1, x = argmax_chunk_12)[name = string("cast_19")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_11_int32 = cast(dtype = indices_chunk_11_int32_dtype_1, x = argmax_chunk_11)[name = string("cast_20")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_10_int32 = cast(dtype = indices_chunk_10_int32_dtype_1, x = argmax_chunk_10)[name = string("cast_21")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_9_int32 = cast(dtype = indices_chunk_9_int32_dtype_1, x = argmax_chunk_9)[name = string("cast_22")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_8_int32 = cast(dtype = indices_chunk_8_int32_dtype_1, x = argmax_chunk_8)[name = string("cast_23")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_7_int32 = cast(dtype = indices_chunk_7_int32_dtype_1, x = argmax_chunk_7)[name = string("cast_24")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_6_int32 = cast(dtype = indices_chunk_6_int32_dtype_1, x = argmax_chunk_6)[name = string("cast_25")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_5_int32 = cast(dtype = indices_chunk_5_int32_dtype_1, x = argmax_chunk_5)[name = string("cast_26")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_4_int32 = cast(dtype = indices_chunk_4_int32_dtype_1, x = argmax_chunk_4)[name = string("cast_27")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_3_int32 = cast(dtype = indices_chunk_3_int32_dtype_1, x = argmax_chunk_3)[name = string("cast_28")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_2_int32 = cast(dtype = indices_chunk_2_int32_dtype_1, x = argmax_chunk_2)[name = string("cast_29")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_1_int32 = cast(dtype = indices_chunk_1_int32_dtype_1, x = argmax_chunk_1)[name = string("cast_30")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_0_int32 = cast(dtype = indices_chunk_0_int32_dtype_1, x = argmax_chunk_0)[name = string("cast_31")];
+            tensor<int32, [1, 16, 1, 1]> indices = concat(axis = indices_axis_1, interleave = indices_interleave_1, values = (indices_chunk_0_int32, indices_chunk_1_int32, indices_chunk_2_int32, indices_chunk_3_int32, indices_chunk_4_int32, indices_chunk_5_int32, indices_chunk_6_int32, indices_chunk_7_int32, indices_chunk_8_int32, indices_chunk_9_int32, indices_chunk_10_int32, indices_chunk_11_int32, indices_chunk_12_int32, indices_chunk_13_int32, indices_chunk_14_int32, indices_chunk_15_int32))[name = string("indices")];
+            int32 argmax_chunks_axis_1 = const()[name = string("argmax_chunks_axis_1"), val = int32(1)];
+            bool argmax_chunks_validate_indices_1 = const()[name = string("argmax_chunks_validate_indices_1"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 1]> argmax_chunks = gather_along_axis(axis = argmax_chunks_axis_1, indices = max_value_index, validate_indices = argmax_chunks_validate_indices_1, x = indices)[name = string("argmax_chunks")];
+            int32 mul_1_x_0 = const()[name = string("mul_1_x_0"), val = int32(2048)];
+            tensor<int32, [1, 1, 1, 1]> mul_1 = mul(x = mul_1_x_0, y = max_value_index)[name = string("mul_1")];
+            tensor<int32, [1, 1, 1, 1]> argmax = add(x = argmax_chunks, y = mul_1)[name = string("argmax")];
+        } -> (logits, argmax, max_value);
+    func lm_head_length_128<ios18>(tensor<fp16, [1, 2048, 1, 128]> hidden_states) {
+            tensor<fp16, [1, 2048, 1, 128]> final_norm_rmsnorm_abs = abs(x = hidden_states)[name = string("final_norm_rmsnorm_abs")];
+            tensor<int32, [1]> final_norm_rmsnorm_maxval_axes_1 = const()[name = string("final_norm_rmsnorm_maxval_axes_1"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_maxval_keep_dims_1 = const()[name = string("final_norm_rmsnorm_maxval_keep_dims_1"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> final_norm_rmsnorm_maxval = reduce_max(axes = final_norm_rmsnorm_maxval_axes_1, keep_dims = final_norm_rmsnorm_maxval_keep_dims_1, x = final_norm_rmsnorm_abs)[name = string("final_norm_rmsnorm_maxval")];
+            fp16 final_norm_rmsnorm_maxval_clipped_alpha_1 = const()[name = string("final_norm_rmsnorm_maxval_clipped_alpha_1"), val = fp16(0x1p-14)];
+            fp16 final_norm_rmsnorm_maxval_clipped_beta_1 = const()[name = string("final_norm_rmsnorm_maxval_clipped_beta_1"), val = fp16(inf)];
+            tensor<fp16, [1, 1, 1, 128]> final_norm_rmsnorm_maxval_clipped = clip(alpha = final_norm_rmsnorm_maxval_clipped_alpha_1, beta = final_norm_rmsnorm_maxval_clipped_beta_1, x = final_norm_rmsnorm_maxval)[name = string("final_norm_rmsnorm_maxval_clipped")];
+            tensor<fp16, [1, 2048, 1, 128]> final_norm_rmsnorm_scaled = real_div(x = hidden_states, y = final_norm_rmsnorm_maxval_clipped)[name = string("final_norm_rmsnorm_scaled")];
+            tensor<int32, [1]> final_norm_rmsnorm_squared_sum_axes_1 = const()[name = string("final_norm_rmsnorm_squared_sum_axes_1"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_squared_sum_keep_dims_1 = const()[name = string("final_norm_rmsnorm_squared_sum_keep_dims_1"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> final_norm_rmsnorm_squared_sum = reduce_sum_square(axes = final_norm_rmsnorm_squared_sum_axes_1, keep_dims = final_norm_rmsnorm_squared_sum_keep_dims_1, x = final_norm_rmsnorm_scaled)[name = string("final_norm_rmsnorm_squared_sum")];
+            fp16 final_norm_rmsnorm_rsqrt_epsilon_1 = const()[name = string("final_norm_rmsnorm_rsqrt_epsilon_1"), val = fp16(0x1p-14)];
+            tensor<fp16, [1, 1, 1, 128]> final_norm_rmsnorm_rsqrt = rsqrt(epsilon = final_norm_rmsnorm_rsqrt_epsilon_1, x = final_norm_rmsnorm_squared_sum)[name = string("final_norm_rmsnorm_rsqrt")];
+            fp16 final_norm_rmsnorm_dim_scaled_y_1 = const()[name = string("final_norm_rmsnorm_dim_scaled_y_1"), val = fp16(0x1.6ap+5)];
+            tensor<fp16, [1, 2048, 1, 128]> final_norm_rmsnorm_dim_scaled = mul(x = final_norm_rmsnorm_scaled, y = final_norm_rmsnorm_dim_scaled_y_1)[name = string("final_norm_rmsnorm_dim_scaled")];
+            tensor<fp16, [1, 2048, 1, 128]> final_norm_rmsnorm_normalized = mul(x = final_norm_rmsnorm_dim_scaled, y = final_norm_rmsnorm_rsqrt)[name = string("final_norm_rmsnorm_normalized")];
+            tensor<fp16, [2048, 1, 1]> final_norm_rmsnorm_y_1 = const()[name = string("final_norm_rmsnorm_y_1"), val = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 2048, 1, 128]> final_norm_rmsnorm = mul(x = final_norm_rmsnorm_normalized, y = final_norm_rmsnorm_y_1)[name = string("final_norm_rmsnorm")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_0_weight_1 = const()[name = string("logits_chunk_0_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4224)))];
+            tensor<int32, [2]> logits_chunk_0_strides_1 = const()[name = string("logits_chunk_0_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_0_pad_type_1 = const()[name = string("logits_chunk_0_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_0_pad_1 = const()[name = string("logits_chunk_0_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_0_dilations_1 = const()[name = string("logits_chunk_0_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_0_groups_1 = const()[name = string("logits_chunk_0_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_0 = conv(dilations = logits_chunk_0_dilations_1, groups = logits_chunk_0_groups_1, pad = logits_chunk_0_pad_1, pad_type = logits_chunk_0_pad_type_1, strides = logits_chunk_0_strides_1, weight = logits_chunk_0_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_0")];
+            int32 argmax_chunk_0_axis_0 = const()[name = string("argmax_chunk_0_axis_0"), val = int32(1)];
+            bool argmax_chunk_0_keep_dims_0 = const()[name = string("argmax_chunk_0_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_0_output_dtype_0 = const()[name = string("argmax_chunk_0_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> argmax_chunk_0 = reduce_argmax(axis = argmax_chunk_0_axis_0, keep_dims = argmax_chunk_0_keep_dims_0, output_dtype = argmax_chunk_0_output_dtype_0, x = logits_chunk_0)[name = string("argmax_chunk_0")];
+            tensor<int32, [1]> max_chunk_0_axes_0 = const()[name = string("max_chunk_0_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_0_keep_dims_0 = const()[name = string("max_chunk_0_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> max_chunk_0 = reduce_max(axes = max_chunk_0_axes_0, keep_dims = max_chunk_0_keep_dims_0, x = logits_chunk_0)[name = string("max_chunk_0")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_1_weight_1 = const()[name = string("logits_chunk_1_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8392896)))];
+            tensor<int32, [2]> logits_chunk_1_strides_1 = const()[name = string("logits_chunk_1_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_1_pad_type_1 = const()[name = string("logits_chunk_1_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_1_pad_1 = const()[name = string("logits_chunk_1_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_1_dilations_1 = const()[name = string("logits_chunk_1_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_1_groups_1 = const()[name = string("logits_chunk_1_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_1 = conv(dilations = logits_chunk_1_dilations_1, groups = logits_chunk_1_groups_1, pad = logits_chunk_1_pad_1, pad_type = logits_chunk_1_pad_type_1, strides = logits_chunk_1_strides_1, weight = logits_chunk_1_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_1")];
+            int32 argmax_chunk_1_axis_0 = const()[name = string("argmax_chunk_1_axis_0"), val = int32(1)];
+            bool argmax_chunk_1_keep_dims_0 = const()[name = string("argmax_chunk_1_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_1_output_dtype_0 = const()[name = string("argmax_chunk_1_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> argmax_chunk_1 = reduce_argmax(axis = argmax_chunk_1_axis_0, keep_dims = argmax_chunk_1_keep_dims_0, output_dtype = argmax_chunk_1_output_dtype_0, x = logits_chunk_1)[name = string("argmax_chunk_1")];
+            tensor<int32, [1]> max_chunk_1_axes_0 = const()[name = string("max_chunk_1_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_1_keep_dims_0 = const()[name = string("max_chunk_1_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> max_chunk_1 = reduce_max(axes = max_chunk_1_axes_0, keep_dims = max_chunk_1_keep_dims_0, x = logits_chunk_1)[name = string("max_chunk_1")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_2_weight_1 = const()[name = string("logits_chunk_2_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16781568)))];
+            tensor<int32, [2]> logits_chunk_2_strides_1 = const()[name = string("logits_chunk_2_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_2_pad_type_1 = const()[name = string("logits_chunk_2_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_2_pad_1 = const()[name = string("logits_chunk_2_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_2_dilations_1 = const()[name = string("logits_chunk_2_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_2_groups_1 = const()[name = string("logits_chunk_2_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_2 = conv(dilations = logits_chunk_2_dilations_1, groups = logits_chunk_2_groups_1, pad = logits_chunk_2_pad_1, pad_type = logits_chunk_2_pad_type_1, strides = logits_chunk_2_strides_1, weight = logits_chunk_2_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_2")];
+            int32 argmax_chunk_2_axis_0 = const()[name = string("argmax_chunk_2_axis_0"), val = int32(1)];
+            bool argmax_chunk_2_keep_dims_0 = const()[name = string("argmax_chunk_2_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_2_output_dtype_0 = const()[name = string("argmax_chunk_2_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> argmax_chunk_2 = reduce_argmax(axis = argmax_chunk_2_axis_0, keep_dims = argmax_chunk_2_keep_dims_0, output_dtype = argmax_chunk_2_output_dtype_0, x = logits_chunk_2)[name = string("argmax_chunk_2")];
+            tensor<int32, [1]> max_chunk_2_axes_0 = const()[name = string("max_chunk_2_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_2_keep_dims_0 = const()[name = string("max_chunk_2_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> max_chunk_2 = reduce_max(axes = max_chunk_2_axes_0, keep_dims = max_chunk_2_keep_dims_0, x = logits_chunk_2)[name = string("max_chunk_2")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_3_weight_1 = const()[name = string("logits_chunk_3_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25170240)))];
+            tensor<int32, [2]> logits_chunk_3_strides_1 = const()[name = string("logits_chunk_3_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_3_pad_type_1 = const()[name = string("logits_chunk_3_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_3_pad_1 = const()[name = string("logits_chunk_3_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_3_dilations_1 = const()[name = string("logits_chunk_3_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_3_groups_1 = const()[name = string("logits_chunk_3_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_3 = conv(dilations = logits_chunk_3_dilations_1, groups = logits_chunk_3_groups_1, pad = logits_chunk_3_pad_1, pad_type = logits_chunk_3_pad_type_1, strides = logits_chunk_3_strides_1, weight = logits_chunk_3_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_3")];
+            int32 argmax_chunk_3_axis_0 = const()[name = string("argmax_chunk_3_axis_0"), val = int32(1)];
+            bool argmax_chunk_3_keep_dims_0 = const()[name = string("argmax_chunk_3_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_3_output_dtype_0 = const()[name = string("argmax_chunk_3_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> argmax_chunk_3 = reduce_argmax(axis = argmax_chunk_3_axis_0, keep_dims = argmax_chunk_3_keep_dims_0, output_dtype = argmax_chunk_3_output_dtype_0, x = logits_chunk_3)[name = string("argmax_chunk_3")];
+            tensor<int32, [1]> max_chunk_3_axes_0 = const()[name = string("max_chunk_3_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_3_keep_dims_0 = const()[name = string("max_chunk_3_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> max_chunk_3 = reduce_max(axes = max_chunk_3_axes_0, keep_dims = max_chunk_3_keep_dims_0, x = logits_chunk_3)[name = string("max_chunk_3")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_4_weight_1 = const()[name = string("logits_chunk_4_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33558912)))];
+            tensor<int32, [2]> logits_chunk_4_strides_1 = const()[name = string("logits_chunk_4_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_4_pad_type_1 = const()[name = string("logits_chunk_4_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_4_pad_1 = const()[name = string("logits_chunk_4_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_4_dilations_1 = const()[name = string("logits_chunk_4_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_4_groups_1 = const()[name = string("logits_chunk_4_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_4 = conv(dilations = logits_chunk_4_dilations_1, groups = logits_chunk_4_groups_1, pad = logits_chunk_4_pad_1, pad_type = logits_chunk_4_pad_type_1, strides = logits_chunk_4_strides_1, weight = logits_chunk_4_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_4")];
+            int32 argmax_chunk_4_axis_0 = const()[name = string("argmax_chunk_4_axis_0"), val = int32(1)];
+            bool argmax_chunk_4_keep_dims_0 = const()[name = string("argmax_chunk_4_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_4_output_dtype_0 = const()[name = string("argmax_chunk_4_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> argmax_chunk_4 = reduce_argmax(axis = argmax_chunk_4_axis_0, keep_dims = argmax_chunk_4_keep_dims_0, output_dtype = argmax_chunk_4_output_dtype_0, x = logits_chunk_4)[name = string("argmax_chunk_4")];
+            tensor<int32, [1]> max_chunk_4_axes_0 = const()[name = string("max_chunk_4_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_4_keep_dims_0 = const()[name = string("max_chunk_4_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> max_chunk_4 = reduce_max(axes = max_chunk_4_axes_0, keep_dims = max_chunk_4_keep_dims_0, x = logits_chunk_4)[name = string("max_chunk_4")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_5_weight_1 = const()[name = string("logits_chunk_5_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41947584)))];
+            tensor<int32, [2]> logits_chunk_5_strides_1 = const()[name = string("logits_chunk_5_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_5_pad_type_1 = const()[name = string("logits_chunk_5_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_5_pad_1 = const()[name = string("logits_chunk_5_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_5_dilations_1 = const()[name = string("logits_chunk_5_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_5_groups_1 = const()[name = string("logits_chunk_5_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_5 = conv(dilations = logits_chunk_5_dilations_1, groups = logits_chunk_5_groups_1, pad = logits_chunk_5_pad_1, pad_type = logits_chunk_5_pad_type_1, strides = logits_chunk_5_strides_1, weight = logits_chunk_5_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_5")];
+            int32 argmax_chunk_5_axis_0 = const()[name = string("argmax_chunk_5_axis_0"), val = int32(1)];
+            bool argmax_chunk_5_keep_dims_0 = const()[name = string("argmax_chunk_5_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_5_output_dtype_0 = const()[name = string("argmax_chunk_5_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> argmax_chunk_5 = reduce_argmax(axis = argmax_chunk_5_axis_0, keep_dims = argmax_chunk_5_keep_dims_0, output_dtype = argmax_chunk_5_output_dtype_0, x = logits_chunk_5)[name = string("argmax_chunk_5")];
+            tensor<int32, [1]> max_chunk_5_axes_0 = const()[name = string("max_chunk_5_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_5_keep_dims_0 = const()[name = string("max_chunk_5_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> max_chunk_5 = reduce_max(axes = max_chunk_5_axes_0, keep_dims = max_chunk_5_keep_dims_0, x = logits_chunk_5)[name = string("max_chunk_5")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_6_weight_1 = const()[name = string("logits_chunk_6_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50336256)))];
+            tensor<int32, [2]> logits_chunk_6_strides_1 = const()[name = string("logits_chunk_6_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_6_pad_type_1 = const()[name = string("logits_chunk_6_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_6_pad_1 = const()[name = string("logits_chunk_6_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_6_dilations_1 = const()[name = string("logits_chunk_6_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_6_groups_1 = const()[name = string("logits_chunk_6_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_6 = conv(dilations = logits_chunk_6_dilations_1, groups = logits_chunk_6_groups_1, pad = logits_chunk_6_pad_1, pad_type = logits_chunk_6_pad_type_1, strides = logits_chunk_6_strides_1, weight = logits_chunk_6_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_6")];
+            int32 argmax_chunk_6_axis_0 = const()[name = string("argmax_chunk_6_axis_0"), val = int32(1)];
+            bool argmax_chunk_6_keep_dims_0 = const()[name = string("argmax_chunk_6_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_6_output_dtype_0 = const()[name = string("argmax_chunk_6_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> argmax_chunk_6 = reduce_argmax(axis = argmax_chunk_6_axis_0, keep_dims = argmax_chunk_6_keep_dims_0, output_dtype = argmax_chunk_6_output_dtype_0, x = logits_chunk_6)[name = string("argmax_chunk_6")];
+            tensor<int32, [1]> max_chunk_6_axes_0 = const()[name = string("max_chunk_6_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_6_keep_dims_0 = const()[name = string("max_chunk_6_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> max_chunk_6 = reduce_max(axes = max_chunk_6_axes_0, keep_dims = max_chunk_6_keep_dims_0, x = logits_chunk_6)[name = string("max_chunk_6")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_7_weight_1 = const()[name = string("logits_chunk_7_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58724928)))];
+            tensor<int32, [2]> logits_chunk_7_strides_1 = const()[name = string("logits_chunk_7_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_7_pad_type_1 = const()[name = string("logits_chunk_7_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_7_pad_1 = const()[name = string("logits_chunk_7_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_7_dilations_1 = const()[name = string("logits_chunk_7_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_7_groups_1 = const()[name = string("logits_chunk_7_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_7 = conv(dilations = logits_chunk_7_dilations_1, groups = logits_chunk_7_groups_1, pad = logits_chunk_7_pad_1, pad_type = logits_chunk_7_pad_type_1, strides = logits_chunk_7_strides_1, weight = logits_chunk_7_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_7")];
+            int32 argmax_chunk_7_axis_0 = const()[name = string("argmax_chunk_7_axis_0"), val = int32(1)];
+            bool argmax_chunk_7_keep_dims_0 = const()[name = string("argmax_chunk_7_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_7_output_dtype_0 = const()[name = string("argmax_chunk_7_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> argmax_chunk_7 = reduce_argmax(axis = argmax_chunk_7_axis_0, keep_dims = argmax_chunk_7_keep_dims_0, output_dtype = argmax_chunk_7_output_dtype_0, x = logits_chunk_7)[name = string("argmax_chunk_7")];
+            tensor<int32, [1]> max_chunk_7_axes_0 = const()[name = string("max_chunk_7_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_7_keep_dims_0 = const()[name = string("max_chunk_7_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> max_chunk_7 = reduce_max(axes = max_chunk_7_axes_0, keep_dims = max_chunk_7_keep_dims_0, x = logits_chunk_7)[name = string("max_chunk_7")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_8_weight_1 = const()[name = string("logits_chunk_8_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67113600)))];
+            tensor<int32, [2]> logits_chunk_8_strides_1 = const()[name = string("logits_chunk_8_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_8_pad_type_1 = const()[name = string("logits_chunk_8_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_8_pad_1 = const()[name = string("logits_chunk_8_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_8_dilations_1 = const()[name = string("logits_chunk_8_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_8_groups_1 = const()[name = string("logits_chunk_8_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_8 = conv(dilations = logits_chunk_8_dilations_1, groups = logits_chunk_8_groups_1, pad = logits_chunk_8_pad_1, pad_type = logits_chunk_8_pad_type_1, strides = logits_chunk_8_strides_1, weight = logits_chunk_8_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_8")];
+            int32 argmax_chunk_8_axis_0 = const()[name = string("argmax_chunk_8_axis_0"), val = int32(1)];
+            bool argmax_chunk_8_keep_dims_0 = const()[name = string("argmax_chunk_8_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_8_output_dtype_0 = const()[name = string("argmax_chunk_8_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> argmax_chunk_8 = reduce_argmax(axis = argmax_chunk_8_axis_0, keep_dims = argmax_chunk_8_keep_dims_0, output_dtype = argmax_chunk_8_output_dtype_0, x = logits_chunk_8)[name = string("argmax_chunk_8")];
+            tensor<int32, [1]> max_chunk_8_axes_0 = const()[name = string("max_chunk_8_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_8_keep_dims_0 = const()[name = string("max_chunk_8_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> max_chunk_8 = reduce_max(axes = max_chunk_8_axes_0, keep_dims = max_chunk_8_keep_dims_0, x = logits_chunk_8)[name = string("max_chunk_8")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_9_weight_1 = const()[name = string("logits_chunk_9_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75502272)))];
+            tensor<int32, [2]> logits_chunk_9_strides_1 = const()[name = string("logits_chunk_9_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_9_pad_type_1 = const()[name = string("logits_chunk_9_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_9_pad_1 = const()[name = string("logits_chunk_9_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_9_dilations_1 = const()[name = string("logits_chunk_9_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_9_groups_1 = const()[name = string("logits_chunk_9_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_9 = conv(dilations = logits_chunk_9_dilations_1, groups = logits_chunk_9_groups_1, pad = logits_chunk_9_pad_1, pad_type = logits_chunk_9_pad_type_1, strides = logits_chunk_9_strides_1, weight = logits_chunk_9_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_9")];
+            int32 argmax_chunk_9_axis_0 = const()[name = string("argmax_chunk_9_axis_0"), val = int32(1)];
+            bool argmax_chunk_9_keep_dims_0 = const()[name = string("argmax_chunk_9_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_9_output_dtype_0 = const()[name = string("argmax_chunk_9_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> argmax_chunk_9 = reduce_argmax(axis = argmax_chunk_9_axis_0, keep_dims = argmax_chunk_9_keep_dims_0, output_dtype = argmax_chunk_9_output_dtype_0, x = logits_chunk_9)[name = string("argmax_chunk_9")];
+            tensor<int32, [1]> max_chunk_9_axes_0 = const()[name = string("max_chunk_9_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_9_keep_dims_0 = const()[name = string("max_chunk_9_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> max_chunk_9 = reduce_max(axes = max_chunk_9_axes_0, keep_dims = max_chunk_9_keep_dims_0, x = logits_chunk_9)[name = string("max_chunk_9")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_10_weight_1 = const()[name = string("logits_chunk_10_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83890944)))];
+            tensor<int32, [2]> logits_chunk_10_strides_1 = const()[name = string("logits_chunk_10_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_10_pad_type_1 = const()[name = string("logits_chunk_10_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_10_pad_1 = const()[name = string("logits_chunk_10_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_10_dilations_1 = const()[name = string("logits_chunk_10_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_10_groups_1 = const()[name = string("logits_chunk_10_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_10 = conv(dilations = logits_chunk_10_dilations_1, groups = logits_chunk_10_groups_1, pad = logits_chunk_10_pad_1, pad_type = logits_chunk_10_pad_type_1, strides = logits_chunk_10_strides_1, weight = logits_chunk_10_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_10")];
+            int32 argmax_chunk_10_axis_0 = const()[name = string("argmax_chunk_10_axis_0"), val = int32(1)];
+            bool argmax_chunk_10_keep_dims_0 = const()[name = string("argmax_chunk_10_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_10_output_dtype_0 = const()[name = string("argmax_chunk_10_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> argmax_chunk_10 = reduce_argmax(axis = argmax_chunk_10_axis_0, keep_dims = argmax_chunk_10_keep_dims_0, output_dtype = argmax_chunk_10_output_dtype_0, x = logits_chunk_10)[name = string("argmax_chunk_10")];
+            tensor<int32, [1]> max_chunk_10_axes_0 = const()[name = string("max_chunk_10_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_10_keep_dims_0 = const()[name = string("max_chunk_10_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> max_chunk_10 = reduce_max(axes = max_chunk_10_axes_0, keep_dims = max_chunk_10_keep_dims_0, x = logits_chunk_10)[name = string("max_chunk_10")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_11_weight_1 = const()[name = string("logits_chunk_11_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92279616)))];
+            tensor<int32, [2]> logits_chunk_11_strides_1 = const()[name = string("logits_chunk_11_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_11_pad_type_1 = const()[name = string("logits_chunk_11_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_11_pad_1 = const()[name = string("logits_chunk_11_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_11_dilations_1 = const()[name = string("logits_chunk_11_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_11_groups_1 = const()[name = string("logits_chunk_11_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_11 = conv(dilations = logits_chunk_11_dilations_1, groups = logits_chunk_11_groups_1, pad = logits_chunk_11_pad_1, pad_type = logits_chunk_11_pad_type_1, strides = logits_chunk_11_strides_1, weight = logits_chunk_11_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_11")];
+            int32 argmax_chunk_11_axis_0 = const()[name = string("argmax_chunk_11_axis_0"), val = int32(1)];
+            bool argmax_chunk_11_keep_dims_0 = const()[name = string("argmax_chunk_11_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_11_output_dtype_0 = const()[name = string("argmax_chunk_11_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> argmax_chunk_11 = reduce_argmax(axis = argmax_chunk_11_axis_0, keep_dims = argmax_chunk_11_keep_dims_0, output_dtype = argmax_chunk_11_output_dtype_0, x = logits_chunk_11)[name = string("argmax_chunk_11")];
+            tensor<int32, [1]> max_chunk_11_axes_0 = const()[name = string("max_chunk_11_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_11_keep_dims_0 = const()[name = string("max_chunk_11_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> max_chunk_11 = reduce_max(axes = max_chunk_11_axes_0, keep_dims = max_chunk_11_keep_dims_0, x = logits_chunk_11)[name = string("max_chunk_11")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_12_weight_1 = const()[name = string("logits_chunk_12_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100668288)))];
+            tensor<int32, [2]> logits_chunk_12_strides_1 = const()[name = string("logits_chunk_12_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_12_pad_type_1 = const()[name = string("logits_chunk_12_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_12_pad_1 = const()[name = string("logits_chunk_12_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_12_dilations_1 = const()[name = string("logits_chunk_12_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_12_groups_1 = const()[name = string("logits_chunk_12_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_12 = conv(dilations = logits_chunk_12_dilations_1, groups = logits_chunk_12_groups_1, pad = logits_chunk_12_pad_1, pad_type = logits_chunk_12_pad_type_1, strides = logits_chunk_12_strides_1, weight = logits_chunk_12_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_12")];
+            int32 argmax_chunk_12_axis_0 = const()[name = string("argmax_chunk_12_axis_0"), val = int32(1)];
+            bool argmax_chunk_12_keep_dims_0 = const()[name = string("argmax_chunk_12_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_12_output_dtype_0 = const()[name = string("argmax_chunk_12_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> argmax_chunk_12 = reduce_argmax(axis = argmax_chunk_12_axis_0, keep_dims = argmax_chunk_12_keep_dims_0, output_dtype = argmax_chunk_12_output_dtype_0, x = logits_chunk_12)[name = string("argmax_chunk_12")];
+            tensor<int32, [1]> max_chunk_12_axes_0 = const()[name = string("max_chunk_12_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_12_keep_dims_0 = const()[name = string("max_chunk_12_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> max_chunk_12 = reduce_max(axes = max_chunk_12_axes_0, keep_dims = max_chunk_12_keep_dims_0, x = logits_chunk_12)[name = string("max_chunk_12")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_13_weight_1 = const()[name = string("logits_chunk_13_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109056960)))];
+            tensor<int32, [2]> logits_chunk_13_strides_1 = const()[name = string("logits_chunk_13_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_13_pad_type_1 = const()[name = string("logits_chunk_13_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_13_pad_1 = const()[name = string("logits_chunk_13_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_13_dilations_1 = const()[name = string("logits_chunk_13_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_13_groups_1 = const()[name = string("logits_chunk_13_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_13 = conv(dilations = logits_chunk_13_dilations_1, groups = logits_chunk_13_groups_1, pad = logits_chunk_13_pad_1, pad_type = logits_chunk_13_pad_type_1, strides = logits_chunk_13_strides_1, weight = logits_chunk_13_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_13")];
+            int32 argmax_chunk_13_axis_0 = const()[name = string("argmax_chunk_13_axis_0"), val = int32(1)];
+            bool argmax_chunk_13_keep_dims_0 = const()[name = string("argmax_chunk_13_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_13_output_dtype_0 = const()[name = string("argmax_chunk_13_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> argmax_chunk_13 = reduce_argmax(axis = argmax_chunk_13_axis_0, keep_dims = argmax_chunk_13_keep_dims_0, output_dtype = argmax_chunk_13_output_dtype_0, x = logits_chunk_13)[name = string("argmax_chunk_13")];
+            tensor<int32, [1]> max_chunk_13_axes_0 = const()[name = string("max_chunk_13_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_13_keep_dims_0 = const()[name = string("max_chunk_13_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> max_chunk_13 = reduce_max(axes = max_chunk_13_axes_0, keep_dims = max_chunk_13_keep_dims_0, x = logits_chunk_13)[name = string("max_chunk_13")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_14_weight_1 = const()[name = string("logits_chunk_14_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117445632)))];
+            tensor<int32, [2]> logits_chunk_14_strides_1 = const()[name = string("logits_chunk_14_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_14_pad_type_1 = const()[name = string("logits_chunk_14_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_14_pad_1 = const()[name = string("logits_chunk_14_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_14_dilations_1 = const()[name = string("logits_chunk_14_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_14_groups_1 = const()[name = string("logits_chunk_14_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_14 = conv(dilations = logits_chunk_14_dilations_1, groups = logits_chunk_14_groups_1, pad = logits_chunk_14_pad_1, pad_type = logits_chunk_14_pad_type_1, strides = logits_chunk_14_strides_1, weight = logits_chunk_14_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_14")];
+            int32 argmax_chunk_14_axis_0 = const()[name = string("argmax_chunk_14_axis_0"), val = int32(1)];
+            bool argmax_chunk_14_keep_dims_0 = const()[name = string("argmax_chunk_14_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_14_output_dtype_0 = const()[name = string("argmax_chunk_14_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> argmax_chunk_14 = reduce_argmax(axis = argmax_chunk_14_axis_0, keep_dims = argmax_chunk_14_keep_dims_0, output_dtype = argmax_chunk_14_output_dtype_0, x = logits_chunk_14)[name = string("argmax_chunk_14")];
+            tensor<int32, [1]> max_chunk_14_axes_0 = const()[name = string("max_chunk_14_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_14_keep_dims_0 = const()[name = string("max_chunk_14_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> max_chunk_14 = reduce_max(axes = max_chunk_14_axes_0, keep_dims = max_chunk_14_keep_dims_0, x = logits_chunk_14)[name = string("max_chunk_14")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_15_weight_1 = const()[name = string("logits_chunk_15_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125834304)))];
+            tensor<int32, [2]> logits_chunk_15_strides_1 = const()[name = string("logits_chunk_15_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_15_pad_type_1 = const()[name = string("logits_chunk_15_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_15_pad_1 = const()[name = string("logits_chunk_15_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_15_dilations_1 = const()[name = string("logits_chunk_15_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_15_groups_1 = const()[name = string("logits_chunk_15_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_15 = conv(dilations = logits_chunk_15_dilations_1, groups = logits_chunk_15_groups_1, pad = logits_chunk_15_pad_1, pad_type = logits_chunk_15_pad_type_1, strides = logits_chunk_15_strides_1, weight = logits_chunk_15_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_15")];
+            int32 argmax_chunk_15_axis_0 = const()[name = string("argmax_chunk_15_axis_0"), val = int32(1)];
+            bool argmax_chunk_15_keep_dims_0 = const()[name = string("argmax_chunk_15_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_15_output_dtype_0 = const()[name = string("argmax_chunk_15_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> argmax_chunk_15 = reduce_argmax(axis = argmax_chunk_15_axis_0, keep_dims = argmax_chunk_15_keep_dims_0, output_dtype = argmax_chunk_15_output_dtype_0, x = logits_chunk_15)[name = string("argmax_chunk_15")];
+            tensor<int32, [1]> max_chunk_15_axes_0 = const()[name = string("max_chunk_15_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_15_keep_dims_0 = const()[name = string("max_chunk_15_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> max_chunk_15 = reduce_max(axes = max_chunk_15_axes_0, keep_dims = max_chunk_15_keep_dims_0, x = logits_chunk_15)[name = string("max_chunk_15")];
+            int32 logits_axis_0 = const()[name = string("logits_axis_0"), val = int32(1)];
+            bool logits_interleave_0 = const()[name = string("logits_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 32768, 1, 128]> logits = concat(axis = logits_axis_0, interleave = logits_interleave_0, values = (logits_chunk_0, logits_chunk_1, logits_chunk_2, logits_chunk_3, logits_chunk_4, logits_chunk_5, logits_chunk_6, logits_chunk_7, logits_chunk_8, logits_chunk_9, logits_chunk_10, logits_chunk_11, logits_chunk_12, logits_chunk_13, logits_chunk_14, logits_chunk_15))[name = string("logits")];
+            int32 values_axis_0 = const()[name = string("values_axis_0"), val = int32(1)];
+            bool values_interleave_0 = const()[name = string("values_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 128]> values = concat(axis = values_axis_0, interleave = values_interleave_0, values = (max_chunk_0, max_chunk_1, max_chunk_2, max_chunk_3, max_chunk_4, max_chunk_5, max_chunk_6, max_chunk_7, max_chunk_8, max_chunk_9, max_chunk_10, max_chunk_11, max_chunk_12, max_chunk_13, max_chunk_14, max_chunk_15))[name = string("values")];
+            tensor<int32, [1]> max_value_axes_0 = const()[name = string("max_value_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_value_keep_dims_0 = const()[name = string("max_value_keep_dims_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 128]> max_value = reduce_max(axes = max_value_axes_0, keep_dims = max_value_keep_dims_0, x = values)[name = string("max_value")];
+            int32 max_value_index_axis_0 = const()[name = string("max_value_index_axis_0"), val = int32(1)];
+            bool max_value_index_keep_dims_0 = const()[name = string("max_value_index_keep_dims_0"), val = bool(true)];
+            string max_value_index_output_dtype_0 = const()[name = string("max_value_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 1, 1, 128]> max_value_index = reduce_argmax(axis = max_value_index_axis_0, keep_dims = max_value_index_keep_dims_0, output_dtype = max_value_index_output_dtype_0, x = values)[name = string("max_value_index")];
+            string indices_chunk_0_int32_dtype_1 = const()[name = string("indices_chunk_0_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_1_int32_dtype_1 = const()[name = string("indices_chunk_1_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_2_int32_dtype_1 = const()[name = string("indices_chunk_2_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_3_int32_dtype_1 = const()[name = string("indices_chunk_3_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_4_int32_dtype_1 = const()[name = string("indices_chunk_4_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_5_int32_dtype_1 = const()[name = string("indices_chunk_5_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_6_int32_dtype_1 = const()[name = string("indices_chunk_6_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_7_int32_dtype_1 = const()[name = string("indices_chunk_7_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_8_int32_dtype_1 = const()[name = string("indices_chunk_8_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_9_int32_dtype_1 = const()[name = string("indices_chunk_9_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_10_int32_dtype_1 = const()[name = string("indices_chunk_10_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_11_int32_dtype_1 = const()[name = string("indices_chunk_11_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_12_int32_dtype_1 = const()[name = string("indices_chunk_12_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_13_int32_dtype_1 = const()[name = string("indices_chunk_13_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_14_int32_dtype_1 = const()[name = string("indices_chunk_14_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_15_int32_dtype_1 = const()[name = string("indices_chunk_15_int32_dtype_1"), val = string("int32")];
+            int32 indices_axis_1 = const()[name = string("indices_axis_1"), val = int32(1)];
+            bool indices_interleave_1 = const()[name = string("indices_interleave_1"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_15_int32 = cast(dtype = indices_chunk_15_int32_dtype_1, x = argmax_chunk_15)[name = string("cast_112")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_14_int32 = cast(dtype = indices_chunk_14_int32_dtype_1, x = argmax_chunk_14)[name = string("cast_113")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_13_int32 = cast(dtype = indices_chunk_13_int32_dtype_1, x = argmax_chunk_13)[name = string("cast_114")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_12_int32 = cast(dtype = indices_chunk_12_int32_dtype_1, x = argmax_chunk_12)[name = string("cast_115")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_11_int32 = cast(dtype = indices_chunk_11_int32_dtype_1, x = argmax_chunk_11)[name = string("cast_116")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_10_int32 = cast(dtype = indices_chunk_10_int32_dtype_1, x = argmax_chunk_10)[name = string("cast_117")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_9_int32 = cast(dtype = indices_chunk_9_int32_dtype_1, x = argmax_chunk_9)[name = string("cast_118")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_8_int32 = cast(dtype = indices_chunk_8_int32_dtype_1, x = argmax_chunk_8)[name = string("cast_119")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_7_int32 = cast(dtype = indices_chunk_7_int32_dtype_1, x = argmax_chunk_7)[name = string("cast_120")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_6_int32 = cast(dtype = indices_chunk_6_int32_dtype_1, x = argmax_chunk_6)[name = string("cast_121")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_5_int32 = cast(dtype = indices_chunk_5_int32_dtype_1, x = argmax_chunk_5)[name = string("cast_122")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_4_int32 = cast(dtype = indices_chunk_4_int32_dtype_1, x = argmax_chunk_4)[name = string("cast_123")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_3_int32 = cast(dtype = indices_chunk_3_int32_dtype_1, x = argmax_chunk_3)[name = string("cast_124")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_2_int32 = cast(dtype = indices_chunk_2_int32_dtype_1, x = argmax_chunk_2)[name = string("cast_125")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_1_int32 = cast(dtype = indices_chunk_1_int32_dtype_1, x = argmax_chunk_1)[name = string("cast_126")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_0_int32 = cast(dtype = indices_chunk_0_int32_dtype_1, x = argmax_chunk_0)[name = string("cast_127")];
+            tensor<int32, [1, 16, 1, 128]> indices = concat(axis = indices_axis_1, interleave = indices_interleave_1, values = (indices_chunk_0_int32, indices_chunk_1_int32, indices_chunk_2_int32, indices_chunk_3_int32, indices_chunk_4_int32, indices_chunk_5_int32, indices_chunk_6_int32, indices_chunk_7_int32, indices_chunk_8_int32, indices_chunk_9_int32, indices_chunk_10_int32, indices_chunk_11_int32, indices_chunk_12_int32, indices_chunk_13_int32, indices_chunk_14_int32, indices_chunk_15_int32))[name = string("indices")];
+            int32 argmax_chunks_axis_1 = const()[name = string("argmax_chunks_axis_1"), val = int32(1)];
+            bool argmax_chunks_validate_indices_1 = const()[name = string("argmax_chunks_validate_indices_1"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 128]> argmax_chunks = gather_along_axis(axis = argmax_chunks_axis_1, indices = max_value_index, validate_indices = argmax_chunks_validate_indices_1, x = indices)[name = string("argmax_chunks")];
+            int32 mul_1_x_0 = const()[name = string("mul_1_x_0"), val = int32(2048)];
+            tensor<int32, [1, 1, 1, 128]> mul_1 = mul(x = mul_1_x_0, y = max_value_index)[name = string("mul_1")];
+            tensor<int32, [1, 1, 1, 128]> argmax = add(x = argmax_chunks, y = mul_1)[name = string("argmax")];
+        } -> (logits, argmax, max_value);
+    func lm_head_length_16<ios18>(tensor<fp16, [1, 2048, 1, 16]> hidden_states) {
+            tensor<fp16, [1, 2048, 1, 16]> final_norm_rmsnorm_abs = abs(x = hidden_states)[name = string("final_norm_rmsnorm_abs")];
+            tensor<int32, [1]> final_norm_rmsnorm_maxval_axes_1 = const()[name = string("final_norm_rmsnorm_maxval_axes_1"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_maxval_keep_dims_1 = const()[name = string("final_norm_rmsnorm_maxval_keep_dims_1"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> final_norm_rmsnorm_maxval = reduce_max(axes = final_norm_rmsnorm_maxval_axes_1, keep_dims = final_norm_rmsnorm_maxval_keep_dims_1, x = final_norm_rmsnorm_abs)[name = string("final_norm_rmsnorm_maxval")];
+            fp16 final_norm_rmsnorm_maxval_clipped_alpha_1 = const()[name = string("final_norm_rmsnorm_maxval_clipped_alpha_1"), val = fp16(0x1p-14)];
+            fp16 final_norm_rmsnorm_maxval_clipped_beta_1 = const()[name = string("final_norm_rmsnorm_maxval_clipped_beta_1"), val = fp16(inf)];
+            tensor<fp16, [1, 1, 1, 16]> final_norm_rmsnorm_maxval_clipped = clip(alpha = final_norm_rmsnorm_maxval_clipped_alpha_1, beta = final_norm_rmsnorm_maxval_clipped_beta_1, x = final_norm_rmsnorm_maxval)[name = string("final_norm_rmsnorm_maxval_clipped")];
+            tensor<fp16, [1, 2048, 1, 16]> final_norm_rmsnorm_scaled = real_div(x = hidden_states, y = final_norm_rmsnorm_maxval_clipped)[name = string("final_norm_rmsnorm_scaled")];
+            tensor<int32, [1]> final_norm_rmsnorm_squared_sum_axes_1 = const()[name = string("final_norm_rmsnorm_squared_sum_axes_1"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_squared_sum_keep_dims_1 = const()[name = string("final_norm_rmsnorm_squared_sum_keep_dims_1"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> final_norm_rmsnorm_squared_sum = reduce_sum_square(axes = final_norm_rmsnorm_squared_sum_axes_1, keep_dims = final_norm_rmsnorm_squared_sum_keep_dims_1, x = final_norm_rmsnorm_scaled)[name = string("final_norm_rmsnorm_squared_sum")];
+            fp16 final_norm_rmsnorm_rsqrt_epsilon_1 = const()[name = string("final_norm_rmsnorm_rsqrt_epsilon_1"), val = fp16(0x1p-14)];
+            tensor<fp16, [1, 1, 1, 16]> final_norm_rmsnorm_rsqrt = rsqrt(epsilon = final_norm_rmsnorm_rsqrt_epsilon_1, x = final_norm_rmsnorm_squared_sum)[name = string("final_norm_rmsnorm_rsqrt")];
+            fp16 final_norm_rmsnorm_dim_scaled_y_1 = const()[name = string("final_norm_rmsnorm_dim_scaled_y_1"), val = fp16(0x1.6ap+5)];
+            tensor<fp16, [1, 2048, 1, 16]> final_norm_rmsnorm_dim_scaled = mul(x = final_norm_rmsnorm_scaled, y = final_norm_rmsnorm_dim_scaled_y_1)[name = string("final_norm_rmsnorm_dim_scaled")];
+            tensor<fp16, [1, 2048, 1, 16]> final_norm_rmsnorm_normalized = mul(x = final_norm_rmsnorm_dim_scaled, y = final_norm_rmsnorm_rsqrt)[name = string("final_norm_rmsnorm_normalized")];
+            tensor<fp16, [2048, 1, 1]> final_norm_rmsnorm_y_1 = const()[name = string("final_norm_rmsnorm_y_1"), val = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 2048, 1, 16]> final_norm_rmsnorm = mul(x = final_norm_rmsnorm_normalized, y = final_norm_rmsnorm_y_1)[name = string("final_norm_rmsnorm")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_0_weight_1 = const()[name = string("logits_chunk_0_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4224)))];
+            tensor<int32, [2]> logits_chunk_0_strides_1 = const()[name = string("logits_chunk_0_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_0_pad_type_1 = const()[name = string("logits_chunk_0_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_0_pad_1 = const()[name = string("logits_chunk_0_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_0_dilations_1 = const()[name = string("logits_chunk_0_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_0_groups_1 = const()[name = string("logits_chunk_0_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_0 = conv(dilations = logits_chunk_0_dilations_1, groups = logits_chunk_0_groups_1, pad = logits_chunk_0_pad_1, pad_type = logits_chunk_0_pad_type_1, strides = logits_chunk_0_strides_1, weight = logits_chunk_0_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_0")];
+            int32 argmax_chunk_0_axis_0 = const()[name = string("argmax_chunk_0_axis_0"), val = int32(1)];
+            bool argmax_chunk_0_keep_dims_0 = const()[name = string("argmax_chunk_0_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_0_output_dtype_0 = const()[name = string("argmax_chunk_0_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> argmax_chunk_0 = reduce_argmax(axis = argmax_chunk_0_axis_0, keep_dims = argmax_chunk_0_keep_dims_0, output_dtype = argmax_chunk_0_output_dtype_0, x = logits_chunk_0)[name = string("argmax_chunk_0")];
+            tensor<int32, [1]> max_chunk_0_axes_0 = const()[name = string("max_chunk_0_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_0_keep_dims_0 = const()[name = string("max_chunk_0_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> max_chunk_0 = reduce_max(axes = max_chunk_0_axes_0, keep_dims = max_chunk_0_keep_dims_0, x = logits_chunk_0)[name = string("max_chunk_0")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_1_weight_1 = const()[name = string("logits_chunk_1_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8392896)))];
+            tensor<int32, [2]> logits_chunk_1_strides_1 = const()[name = string("logits_chunk_1_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_1_pad_type_1 = const()[name = string("logits_chunk_1_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_1_pad_1 = const()[name = string("logits_chunk_1_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_1_dilations_1 = const()[name = string("logits_chunk_1_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_1_groups_1 = const()[name = string("logits_chunk_1_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_1 = conv(dilations = logits_chunk_1_dilations_1, groups = logits_chunk_1_groups_1, pad = logits_chunk_1_pad_1, pad_type = logits_chunk_1_pad_type_1, strides = logits_chunk_1_strides_1, weight = logits_chunk_1_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_1")];
+            int32 argmax_chunk_1_axis_0 = const()[name = string("argmax_chunk_1_axis_0"), val = int32(1)];
+            bool argmax_chunk_1_keep_dims_0 = const()[name = string("argmax_chunk_1_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_1_output_dtype_0 = const()[name = string("argmax_chunk_1_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> argmax_chunk_1 = reduce_argmax(axis = argmax_chunk_1_axis_0, keep_dims = argmax_chunk_1_keep_dims_0, output_dtype = argmax_chunk_1_output_dtype_0, x = logits_chunk_1)[name = string("argmax_chunk_1")];
+            tensor<int32, [1]> max_chunk_1_axes_0 = const()[name = string("max_chunk_1_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_1_keep_dims_0 = const()[name = string("max_chunk_1_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> max_chunk_1 = reduce_max(axes = max_chunk_1_axes_0, keep_dims = max_chunk_1_keep_dims_0, x = logits_chunk_1)[name = string("max_chunk_1")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_2_weight_1 = const()[name = string("logits_chunk_2_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16781568)))];
+            tensor<int32, [2]> logits_chunk_2_strides_1 = const()[name = string("logits_chunk_2_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_2_pad_type_1 = const()[name = string("logits_chunk_2_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_2_pad_1 = const()[name = string("logits_chunk_2_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_2_dilations_1 = const()[name = string("logits_chunk_2_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_2_groups_1 = const()[name = string("logits_chunk_2_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_2 = conv(dilations = logits_chunk_2_dilations_1, groups = logits_chunk_2_groups_1, pad = logits_chunk_2_pad_1, pad_type = logits_chunk_2_pad_type_1, strides = logits_chunk_2_strides_1, weight = logits_chunk_2_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_2")];
+            int32 argmax_chunk_2_axis_0 = const()[name = string("argmax_chunk_2_axis_0"), val = int32(1)];
+            bool argmax_chunk_2_keep_dims_0 = const()[name = string("argmax_chunk_2_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_2_output_dtype_0 = const()[name = string("argmax_chunk_2_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> argmax_chunk_2 = reduce_argmax(axis = argmax_chunk_2_axis_0, keep_dims = argmax_chunk_2_keep_dims_0, output_dtype = argmax_chunk_2_output_dtype_0, x = logits_chunk_2)[name = string("argmax_chunk_2")];
+            tensor<int32, [1]> max_chunk_2_axes_0 = const()[name = string("max_chunk_2_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_2_keep_dims_0 = const()[name = string("max_chunk_2_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> max_chunk_2 = reduce_max(axes = max_chunk_2_axes_0, keep_dims = max_chunk_2_keep_dims_0, x = logits_chunk_2)[name = string("max_chunk_2")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_3_weight_1 = const()[name = string("logits_chunk_3_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25170240)))];
+            tensor<int32, [2]> logits_chunk_3_strides_1 = const()[name = string("logits_chunk_3_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_3_pad_type_1 = const()[name = string("logits_chunk_3_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_3_pad_1 = const()[name = string("logits_chunk_3_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_3_dilations_1 = const()[name = string("logits_chunk_3_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_3_groups_1 = const()[name = string("logits_chunk_3_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_3 = conv(dilations = logits_chunk_3_dilations_1, groups = logits_chunk_3_groups_1, pad = logits_chunk_3_pad_1, pad_type = logits_chunk_3_pad_type_1, strides = logits_chunk_3_strides_1, weight = logits_chunk_3_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_3")];
+            int32 argmax_chunk_3_axis_0 = const()[name = string("argmax_chunk_3_axis_0"), val = int32(1)];
+            bool argmax_chunk_3_keep_dims_0 = const()[name = string("argmax_chunk_3_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_3_output_dtype_0 = const()[name = string("argmax_chunk_3_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> argmax_chunk_3 = reduce_argmax(axis = argmax_chunk_3_axis_0, keep_dims = argmax_chunk_3_keep_dims_0, output_dtype = argmax_chunk_3_output_dtype_0, x = logits_chunk_3)[name = string("argmax_chunk_3")];
+            tensor<int32, [1]> max_chunk_3_axes_0 = const()[name = string("max_chunk_3_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_3_keep_dims_0 = const()[name = string("max_chunk_3_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> max_chunk_3 = reduce_max(axes = max_chunk_3_axes_0, keep_dims = max_chunk_3_keep_dims_0, x = logits_chunk_3)[name = string("max_chunk_3")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_4_weight_1 = const()[name = string("logits_chunk_4_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33558912)))];
+            tensor<int32, [2]> logits_chunk_4_strides_1 = const()[name = string("logits_chunk_4_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_4_pad_type_1 = const()[name = string("logits_chunk_4_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_4_pad_1 = const()[name = string("logits_chunk_4_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_4_dilations_1 = const()[name = string("logits_chunk_4_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_4_groups_1 = const()[name = string("logits_chunk_4_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_4 = conv(dilations = logits_chunk_4_dilations_1, groups = logits_chunk_4_groups_1, pad = logits_chunk_4_pad_1, pad_type = logits_chunk_4_pad_type_1, strides = logits_chunk_4_strides_1, weight = logits_chunk_4_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_4")];
+            int32 argmax_chunk_4_axis_0 = const()[name = string("argmax_chunk_4_axis_0"), val = int32(1)];
+            bool argmax_chunk_4_keep_dims_0 = const()[name = string("argmax_chunk_4_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_4_output_dtype_0 = const()[name = string("argmax_chunk_4_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> argmax_chunk_4 = reduce_argmax(axis = argmax_chunk_4_axis_0, keep_dims = argmax_chunk_4_keep_dims_0, output_dtype = argmax_chunk_4_output_dtype_0, x = logits_chunk_4)[name = string("argmax_chunk_4")];
+            tensor<int32, [1]> max_chunk_4_axes_0 = const()[name = string("max_chunk_4_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_4_keep_dims_0 = const()[name = string("max_chunk_4_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> max_chunk_4 = reduce_max(axes = max_chunk_4_axes_0, keep_dims = max_chunk_4_keep_dims_0, x = logits_chunk_4)[name = string("max_chunk_4")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_5_weight_1 = const()[name = string("logits_chunk_5_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41947584)))];
+            tensor<int32, [2]> logits_chunk_5_strides_1 = const()[name = string("logits_chunk_5_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_5_pad_type_1 = const()[name = string("logits_chunk_5_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_5_pad_1 = const()[name = string("logits_chunk_5_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_5_dilations_1 = const()[name = string("logits_chunk_5_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_5_groups_1 = const()[name = string("logits_chunk_5_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_5 = conv(dilations = logits_chunk_5_dilations_1, groups = logits_chunk_5_groups_1, pad = logits_chunk_5_pad_1, pad_type = logits_chunk_5_pad_type_1, strides = logits_chunk_5_strides_1, weight = logits_chunk_5_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_5")];
+            int32 argmax_chunk_5_axis_0 = const()[name = string("argmax_chunk_5_axis_0"), val = int32(1)];
+            bool argmax_chunk_5_keep_dims_0 = const()[name = string("argmax_chunk_5_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_5_output_dtype_0 = const()[name = string("argmax_chunk_5_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> argmax_chunk_5 = reduce_argmax(axis = argmax_chunk_5_axis_0, keep_dims = argmax_chunk_5_keep_dims_0, output_dtype = argmax_chunk_5_output_dtype_0, x = logits_chunk_5)[name = string("argmax_chunk_5")];
+            tensor<int32, [1]> max_chunk_5_axes_0 = const()[name = string("max_chunk_5_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_5_keep_dims_0 = const()[name = string("max_chunk_5_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> max_chunk_5 = reduce_max(axes = max_chunk_5_axes_0, keep_dims = max_chunk_5_keep_dims_0, x = logits_chunk_5)[name = string("max_chunk_5")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_6_weight_1 = const()[name = string("logits_chunk_6_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50336256)))];
+            tensor<int32, [2]> logits_chunk_6_strides_1 = const()[name = string("logits_chunk_6_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_6_pad_type_1 = const()[name = string("logits_chunk_6_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_6_pad_1 = const()[name = string("logits_chunk_6_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_6_dilations_1 = const()[name = string("logits_chunk_6_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_6_groups_1 = const()[name = string("logits_chunk_6_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_6 = conv(dilations = logits_chunk_6_dilations_1, groups = logits_chunk_6_groups_1, pad = logits_chunk_6_pad_1, pad_type = logits_chunk_6_pad_type_1, strides = logits_chunk_6_strides_1, weight = logits_chunk_6_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_6")];
+            int32 argmax_chunk_6_axis_0 = const()[name = string("argmax_chunk_6_axis_0"), val = int32(1)];
+            bool argmax_chunk_6_keep_dims_0 = const()[name = string("argmax_chunk_6_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_6_output_dtype_0 = const()[name = string("argmax_chunk_6_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> argmax_chunk_6 = reduce_argmax(axis = argmax_chunk_6_axis_0, keep_dims = argmax_chunk_6_keep_dims_0, output_dtype = argmax_chunk_6_output_dtype_0, x = logits_chunk_6)[name = string("argmax_chunk_6")];
+            tensor<int32, [1]> max_chunk_6_axes_0 = const()[name = string("max_chunk_6_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_6_keep_dims_0 = const()[name = string("max_chunk_6_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> max_chunk_6 = reduce_max(axes = max_chunk_6_axes_0, keep_dims = max_chunk_6_keep_dims_0, x = logits_chunk_6)[name = string("max_chunk_6")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_7_weight_1 = const()[name = string("logits_chunk_7_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58724928)))];
+            tensor<int32, [2]> logits_chunk_7_strides_1 = const()[name = string("logits_chunk_7_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_7_pad_type_1 = const()[name = string("logits_chunk_7_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_7_pad_1 = const()[name = string("logits_chunk_7_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_7_dilations_1 = const()[name = string("logits_chunk_7_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_7_groups_1 = const()[name = string("logits_chunk_7_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_7 = conv(dilations = logits_chunk_7_dilations_1, groups = logits_chunk_7_groups_1, pad = logits_chunk_7_pad_1, pad_type = logits_chunk_7_pad_type_1, strides = logits_chunk_7_strides_1, weight = logits_chunk_7_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_7")];
+            int32 argmax_chunk_7_axis_0 = const()[name = string("argmax_chunk_7_axis_0"), val = int32(1)];
+            bool argmax_chunk_7_keep_dims_0 = const()[name = string("argmax_chunk_7_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_7_output_dtype_0 = const()[name = string("argmax_chunk_7_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> argmax_chunk_7 = reduce_argmax(axis = argmax_chunk_7_axis_0, keep_dims = argmax_chunk_7_keep_dims_0, output_dtype = argmax_chunk_7_output_dtype_0, x = logits_chunk_7)[name = string("argmax_chunk_7")];
+            tensor<int32, [1]> max_chunk_7_axes_0 = const()[name = string("max_chunk_7_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_7_keep_dims_0 = const()[name = string("max_chunk_7_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> max_chunk_7 = reduce_max(axes = max_chunk_7_axes_0, keep_dims = max_chunk_7_keep_dims_0, x = logits_chunk_7)[name = string("max_chunk_7")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_8_weight_1 = const()[name = string("logits_chunk_8_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67113600)))];
+            tensor<int32, [2]> logits_chunk_8_strides_1 = const()[name = string("logits_chunk_8_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_8_pad_type_1 = const()[name = string("logits_chunk_8_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_8_pad_1 = const()[name = string("logits_chunk_8_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_8_dilations_1 = const()[name = string("logits_chunk_8_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_8_groups_1 = const()[name = string("logits_chunk_8_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_8 = conv(dilations = logits_chunk_8_dilations_1, groups = logits_chunk_8_groups_1, pad = logits_chunk_8_pad_1, pad_type = logits_chunk_8_pad_type_1, strides = logits_chunk_8_strides_1, weight = logits_chunk_8_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_8")];
+            int32 argmax_chunk_8_axis_0 = const()[name = string("argmax_chunk_8_axis_0"), val = int32(1)];
+            bool argmax_chunk_8_keep_dims_0 = const()[name = string("argmax_chunk_8_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_8_output_dtype_0 = const()[name = string("argmax_chunk_8_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> argmax_chunk_8 = reduce_argmax(axis = argmax_chunk_8_axis_0, keep_dims = argmax_chunk_8_keep_dims_0, output_dtype = argmax_chunk_8_output_dtype_0, x = logits_chunk_8)[name = string("argmax_chunk_8")];
+            tensor<int32, [1]> max_chunk_8_axes_0 = const()[name = string("max_chunk_8_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_8_keep_dims_0 = const()[name = string("max_chunk_8_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> max_chunk_8 = reduce_max(axes = max_chunk_8_axes_0, keep_dims = max_chunk_8_keep_dims_0, x = logits_chunk_8)[name = string("max_chunk_8")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_9_weight_1 = const()[name = string("logits_chunk_9_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75502272)))];
+            tensor<int32, [2]> logits_chunk_9_strides_1 = const()[name = string("logits_chunk_9_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_9_pad_type_1 = const()[name = string("logits_chunk_9_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_9_pad_1 = const()[name = string("logits_chunk_9_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_9_dilations_1 = const()[name = string("logits_chunk_9_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_9_groups_1 = const()[name = string("logits_chunk_9_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_9 = conv(dilations = logits_chunk_9_dilations_1, groups = logits_chunk_9_groups_1, pad = logits_chunk_9_pad_1, pad_type = logits_chunk_9_pad_type_1, strides = logits_chunk_9_strides_1, weight = logits_chunk_9_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_9")];
+            int32 argmax_chunk_9_axis_0 = const()[name = string("argmax_chunk_9_axis_0"), val = int32(1)];
+            bool argmax_chunk_9_keep_dims_0 = const()[name = string("argmax_chunk_9_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_9_output_dtype_0 = const()[name = string("argmax_chunk_9_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> argmax_chunk_9 = reduce_argmax(axis = argmax_chunk_9_axis_0, keep_dims = argmax_chunk_9_keep_dims_0, output_dtype = argmax_chunk_9_output_dtype_0, x = logits_chunk_9)[name = string("argmax_chunk_9")];
+            tensor<int32, [1]> max_chunk_9_axes_0 = const()[name = string("max_chunk_9_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_9_keep_dims_0 = const()[name = string("max_chunk_9_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> max_chunk_9 = reduce_max(axes = max_chunk_9_axes_0, keep_dims = max_chunk_9_keep_dims_0, x = logits_chunk_9)[name = string("max_chunk_9")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_10_weight_1 = const()[name = string("logits_chunk_10_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83890944)))];
+            tensor<int32, [2]> logits_chunk_10_strides_1 = const()[name = string("logits_chunk_10_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_10_pad_type_1 = const()[name = string("logits_chunk_10_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_10_pad_1 = const()[name = string("logits_chunk_10_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_10_dilations_1 = const()[name = string("logits_chunk_10_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_10_groups_1 = const()[name = string("logits_chunk_10_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_10 = conv(dilations = logits_chunk_10_dilations_1, groups = logits_chunk_10_groups_1, pad = logits_chunk_10_pad_1, pad_type = logits_chunk_10_pad_type_1, strides = logits_chunk_10_strides_1, weight = logits_chunk_10_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_10")];
+            int32 argmax_chunk_10_axis_0 = const()[name = string("argmax_chunk_10_axis_0"), val = int32(1)];
+            bool argmax_chunk_10_keep_dims_0 = const()[name = string("argmax_chunk_10_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_10_output_dtype_0 = const()[name = string("argmax_chunk_10_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> argmax_chunk_10 = reduce_argmax(axis = argmax_chunk_10_axis_0, keep_dims = argmax_chunk_10_keep_dims_0, output_dtype = argmax_chunk_10_output_dtype_0, x = logits_chunk_10)[name = string("argmax_chunk_10")];
+            tensor<int32, [1]> max_chunk_10_axes_0 = const()[name = string("max_chunk_10_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_10_keep_dims_0 = const()[name = string("max_chunk_10_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> max_chunk_10 = reduce_max(axes = max_chunk_10_axes_0, keep_dims = max_chunk_10_keep_dims_0, x = logits_chunk_10)[name = string("max_chunk_10")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_11_weight_1 = const()[name = string("logits_chunk_11_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92279616)))];
+            tensor<int32, [2]> logits_chunk_11_strides_1 = const()[name = string("logits_chunk_11_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_11_pad_type_1 = const()[name = string("logits_chunk_11_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_11_pad_1 = const()[name = string("logits_chunk_11_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_11_dilations_1 = const()[name = string("logits_chunk_11_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_11_groups_1 = const()[name = string("logits_chunk_11_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_11 = conv(dilations = logits_chunk_11_dilations_1, groups = logits_chunk_11_groups_1, pad = logits_chunk_11_pad_1, pad_type = logits_chunk_11_pad_type_1, strides = logits_chunk_11_strides_1, weight = logits_chunk_11_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_11")];
+            int32 argmax_chunk_11_axis_0 = const()[name = string("argmax_chunk_11_axis_0"), val = int32(1)];
+            bool argmax_chunk_11_keep_dims_0 = const()[name = string("argmax_chunk_11_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_11_output_dtype_0 = const()[name = string("argmax_chunk_11_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> argmax_chunk_11 = reduce_argmax(axis = argmax_chunk_11_axis_0, keep_dims = argmax_chunk_11_keep_dims_0, output_dtype = argmax_chunk_11_output_dtype_0, x = logits_chunk_11)[name = string("argmax_chunk_11")];
+            tensor<int32, [1]> max_chunk_11_axes_0 = const()[name = string("max_chunk_11_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_11_keep_dims_0 = const()[name = string("max_chunk_11_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> max_chunk_11 = reduce_max(axes = max_chunk_11_axes_0, keep_dims = max_chunk_11_keep_dims_0, x = logits_chunk_11)[name = string("max_chunk_11")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_12_weight_1 = const()[name = string("logits_chunk_12_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100668288)))];
+            tensor<int32, [2]> logits_chunk_12_strides_1 = const()[name = string("logits_chunk_12_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_12_pad_type_1 = const()[name = string("logits_chunk_12_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_12_pad_1 = const()[name = string("logits_chunk_12_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_12_dilations_1 = const()[name = string("logits_chunk_12_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_12_groups_1 = const()[name = string("logits_chunk_12_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_12 = conv(dilations = logits_chunk_12_dilations_1, groups = logits_chunk_12_groups_1, pad = logits_chunk_12_pad_1, pad_type = logits_chunk_12_pad_type_1, strides = logits_chunk_12_strides_1, weight = logits_chunk_12_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_12")];
+            int32 argmax_chunk_12_axis_0 = const()[name = string("argmax_chunk_12_axis_0"), val = int32(1)];
+            bool argmax_chunk_12_keep_dims_0 = const()[name = string("argmax_chunk_12_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_12_output_dtype_0 = const()[name = string("argmax_chunk_12_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> argmax_chunk_12 = reduce_argmax(axis = argmax_chunk_12_axis_0, keep_dims = argmax_chunk_12_keep_dims_0, output_dtype = argmax_chunk_12_output_dtype_0, x = logits_chunk_12)[name = string("argmax_chunk_12")];
+            tensor<int32, [1]> max_chunk_12_axes_0 = const()[name = string("max_chunk_12_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_12_keep_dims_0 = const()[name = string("max_chunk_12_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> max_chunk_12 = reduce_max(axes = max_chunk_12_axes_0, keep_dims = max_chunk_12_keep_dims_0, x = logits_chunk_12)[name = string("max_chunk_12")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_13_weight_1 = const()[name = string("logits_chunk_13_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109056960)))];
+            tensor<int32, [2]> logits_chunk_13_strides_1 = const()[name = string("logits_chunk_13_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_13_pad_type_1 = const()[name = string("logits_chunk_13_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_13_pad_1 = const()[name = string("logits_chunk_13_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_13_dilations_1 = const()[name = string("logits_chunk_13_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_13_groups_1 = const()[name = string("logits_chunk_13_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_13 = conv(dilations = logits_chunk_13_dilations_1, groups = logits_chunk_13_groups_1, pad = logits_chunk_13_pad_1, pad_type = logits_chunk_13_pad_type_1, strides = logits_chunk_13_strides_1, weight = logits_chunk_13_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_13")];
+            int32 argmax_chunk_13_axis_0 = const()[name = string("argmax_chunk_13_axis_0"), val = int32(1)];
+            bool argmax_chunk_13_keep_dims_0 = const()[name = string("argmax_chunk_13_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_13_output_dtype_0 = const()[name = string("argmax_chunk_13_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> argmax_chunk_13 = reduce_argmax(axis = argmax_chunk_13_axis_0, keep_dims = argmax_chunk_13_keep_dims_0, output_dtype = argmax_chunk_13_output_dtype_0, x = logits_chunk_13)[name = string("argmax_chunk_13")];
+            tensor<int32, [1]> max_chunk_13_axes_0 = const()[name = string("max_chunk_13_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_13_keep_dims_0 = const()[name = string("max_chunk_13_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> max_chunk_13 = reduce_max(axes = max_chunk_13_axes_0, keep_dims = max_chunk_13_keep_dims_0, x = logits_chunk_13)[name = string("max_chunk_13")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_14_weight_1 = const()[name = string("logits_chunk_14_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117445632)))];
+            tensor<int32, [2]> logits_chunk_14_strides_1 = const()[name = string("logits_chunk_14_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_14_pad_type_1 = const()[name = string("logits_chunk_14_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_14_pad_1 = const()[name = string("logits_chunk_14_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_14_dilations_1 = const()[name = string("logits_chunk_14_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_14_groups_1 = const()[name = string("logits_chunk_14_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_14 = conv(dilations = logits_chunk_14_dilations_1, groups = logits_chunk_14_groups_1, pad = logits_chunk_14_pad_1, pad_type = logits_chunk_14_pad_type_1, strides = logits_chunk_14_strides_1, weight = logits_chunk_14_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_14")];
+            int32 argmax_chunk_14_axis_0 = const()[name = string("argmax_chunk_14_axis_0"), val = int32(1)];
+            bool argmax_chunk_14_keep_dims_0 = const()[name = string("argmax_chunk_14_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_14_output_dtype_0 = const()[name = string("argmax_chunk_14_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> argmax_chunk_14 = reduce_argmax(axis = argmax_chunk_14_axis_0, keep_dims = argmax_chunk_14_keep_dims_0, output_dtype = argmax_chunk_14_output_dtype_0, x = logits_chunk_14)[name = string("argmax_chunk_14")];
+            tensor<int32, [1]> max_chunk_14_axes_0 = const()[name = string("max_chunk_14_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_14_keep_dims_0 = const()[name = string("max_chunk_14_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> max_chunk_14 = reduce_max(axes = max_chunk_14_axes_0, keep_dims = max_chunk_14_keep_dims_0, x = logits_chunk_14)[name = string("max_chunk_14")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_15_weight_1 = const()[name = string("logits_chunk_15_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125834304)))];
+            tensor<int32, [2]> logits_chunk_15_strides_1 = const()[name = string("logits_chunk_15_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_15_pad_type_1 = const()[name = string("logits_chunk_15_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_15_pad_1 = const()[name = string("logits_chunk_15_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_15_dilations_1 = const()[name = string("logits_chunk_15_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_15_groups_1 = const()[name = string("logits_chunk_15_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_15 = conv(dilations = logits_chunk_15_dilations_1, groups = logits_chunk_15_groups_1, pad = logits_chunk_15_pad_1, pad_type = logits_chunk_15_pad_type_1, strides = logits_chunk_15_strides_1, weight = logits_chunk_15_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_15")];
+            int32 argmax_chunk_15_axis_0 = const()[name = string("argmax_chunk_15_axis_0"), val = int32(1)];
+            bool argmax_chunk_15_keep_dims_0 = const()[name = string("argmax_chunk_15_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_15_output_dtype_0 = const()[name = string("argmax_chunk_15_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> argmax_chunk_15 = reduce_argmax(axis = argmax_chunk_15_axis_0, keep_dims = argmax_chunk_15_keep_dims_0, output_dtype = argmax_chunk_15_output_dtype_0, x = logits_chunk_15)[name = string("argmax_chunk_15")];
+            tensor<int32, [1]> max_chunk_15_axes_0 = const()[name = string("max_chunk_15_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_15_keep_dims_0 = const()[name = string("max_chunk_15_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> max_chunk_15 = reduce_max(axes = max_chunk_15_axes_0, keep_dims = max_chunk_15_keep_dims_0, x = logits_chunk_15)[name = string("max_chunk_15")];
+            int32 logits_axis_0 = const()[name = string("logits_axis_0"), val = int32(1)];
+            bool logits_interleave_0 = const()[name = string("logits_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 32768, 1, 16]> logits = concat(axis = logits_axis_0, interleave = logits_interleave_0, values = (logits_chunk_0, logits_chunk_1, logits_chunk_2, logits_chunk_3, logits_chunk_4, logits_chunk_5, logits_chunk_6, logits_chunk_7, logits_chunk_8, logits_chunk_9, logits_chunk_10, logits_chunk_11, logits_chunk_12, logits_chunk_13, logits_chunk_14, logits_chunk_15))[name = string("logits")];
+            int32 values_axis_0 = const()[name = string("values_axis_0"), val = int32(1)];
+            bool values_interleave_0 = const()[name = string("values_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 16]> values = concat(axis = values_axis_0, interleave = values_interleave_0, values = (max_chunk_0, max_chunk_1, max_chunk_2, max_chunk_3, max_chunk_4, max_chunk_5, max_chunk_6, max_chunk_7, max_chunk_8, max_chunk_9, max_chunk_10, max_chunk_11, max_chunk_12, max_chunk_13, max_chunk_14, max_chunk_15))[name = string("values")];
+            tensor<int32, [1]> max_value_axes_0 = const()[name = string("max_value_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_value_keep_dims_0 = const()[name = string("max_value_keep_dims_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 16]> max_value = reduce_max(axes = max_value_axes_0, keep_dims = max_value_keep_dims_0, x = values)[name = string("max_value")];
+            int32 max_value_index_axis_0 = const()[name = string("max_value_index_axis_0"), val = int32(1)];
+            bool max_value_index_keep_dims_0 = const()[name = string("max_value_index_keep_dims_0"), val = bool(true)];
+            string max_value_index_output_dtype_0 = const()[name = string("max_value_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 1, 1, 16]> max_value_index = reduce_argmax(axis = max_value_index_axis_0, keep_dims = max_value_index_keep_dims_0, output_dtype = max_value_index_output_dtype_0, x = values)[name = string("max_value_index")];
+            string indices_chunk_0_int32_dtype_1 = const()[name = string("indices_chunk_0_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_1_int32_dtype_1 = const()[name = string("indices_chunk_1_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_2_int32_dtype_1 = const()[name = string("indices_chunk_2_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_3_int32_dtype_1 = const()[name = string("indices_chunk_3_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_4_int32_dtype_1 = const()[name = string("indices_chunk_4_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_5_int32_dtype_1 = const()[name = string("indices_chunk_5_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_6_int32_dtype_1 = const()[name = string("indices_chunk_6_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_7_int32_dtype_1 = const()[name = string("indices_chunk_7_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_8_int32_dtype_1 = const()[name = string("indices_chunk_8_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_9_int32_dtype_1 = const()[name = string("indices_chunk_9_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_10_int32_dtype_1 = const()[name = string("indices_chunk_10_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_11_int32_dtype_1 = const()[name = string("indices_chunk_11_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_12_int32_dtype_1 = const()[name = string("indices_chunk_12_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_13_int32_dtype_1 = const()[name = string("indices_chunk_13_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_14_int32_dtype_1 = const()[name = string("indices_chunk_14_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_15_int32_dtype_1 = const()[name = string("indices_chunk_15_int32_dtype_1"), val = string("int32")];
+            int32 indices_axis_1 = const()[name = string("indices_axis_1"), val = int32(1)];
+            bool indices_interleave_1 = const()[name = string("indices_interleave_1"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_15_int32 = cast(dtype = indices_chunk_15_int32_dtype_1, x = argmax_chunk_15)[name = string("cast_48")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_14_int32 = cast(dtype = indices_chunk_14_int32_dtype_1, x = argmax_chunk_14)[name = string("cast_49")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_13_int32 = cast(dtype = indices_chunk_13_int32_dtype_1, x = argmax_chunk_13)[name = string("cast_50")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_12_int32 = cast(dtype = indices_chunk_12_int32_dtype_1, x = argmax_chunk_12)[name = string("cast_51")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_11_int32 = cast(dtype = indices_chunk_11_int32_dtype_1, x = argmax_chunk_11)[name = string("cast_52")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_10_int32 = cast(dtype = indices_chunk_10_int32_dtype_1, x = argmax_chunk_10)[name = string("cast_53")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_9_int32 = cast(dtype = indices_chunk_9_int32_dtype_1, x = argmax_chunk_9)[name = string("cast_54")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_8_int32 = cast(dtype = indices_chunk_8_int32_dtype_1, x = argmax_chunk_8)[name = string("cast_55")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_7_int32 = cast(dtype = indices_chunk_7_int32_dtype_1, x = argmax_chunk_7)[name = string("cast_56")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_6_int32 = cast(dtype = indices_chunk_6_int32_dtype_1, x = argmax_chunk_6)[name = string("cast_57")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_5_int32 = cast(dtype = indices_chunk_5_int32_dtype_1, x = argmax_chunk_5)[name = string("cast_58")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_4_int32 = cast(dtype = indices_chunk_4_int32_dtype_1, x = argmax_chunk_4)[name = string("cast_59")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_3_int32 = cast(dtype = indices_chunk_3_int32_dtype_1, x = argmax_chunk_3)[name = string("cast_60")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_2_int32 = cast(dtype = indices_chunk_2_int32_dtype_1, x = argmax_chunk_2)[name = string("cast_61")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_1_int32 = cast(dtype = indices_chunk_1_int32_dtype_1, x = argmax_chunk_1)[name = string("cast_62")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_0_int32 = cast(dtype = indices_chunk_0_int32_dtype_1, x = argmax_chunk_0)[name = string("cast_63")];
+            tensor<int32, [1, 16, 1, 16]> indices = concat(axis = indices_axis_1, interleave = indices_interleave_1, values = (indices_chunk_0_int32, indices_chunk_1_int32, indices_chunk_2_int32, indices_chunk_3_int32, indices_chunk_4_int32, indices_chunk_5_int32, indices_chunk_6_int32, indices_chunk_7_int32, indices_chunk_8_int32, indices_chunk_9_int32, indices_chunk_10_int32, indices_chunk_11_int32, indices_chunk_12_int32, indices_chunk_13_int32, indices_chunk_14_int32, indices_chunk_15_int32))[name = string("indices")];
+            int32 argmax_chunks_axis_1 = const()[name = string("argmax_chunks_axis_1"), val = int32(1)];
+            bool argmax_chunks_validate_indices_1 = const()[name = string("argmax_chunks_validate_indices_1"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 16]> argmax_chunks = gather_along_axis(axis = argmax_chunks_axis_1, indices = max_value_index, validate_indices = argmax_chunks_validate_indices_1, x = indices)[name = string("argmax_chunks")];
+            int32 mul_1_x_0 = const()[name = string("mul_1_x_0"), val = int32(2048)];
+            tensor<int32, [1, 1, 1, 16]> mul_1 = mul(x = mul_1_x_0, y = max_value_index)[name = string("mul_1")];
+            tensor<int32, [1, 1, 1, 16]> argmax = add(x = argmax_chunks, y = mul_1)[name = string("argmax")];
+        } -> (logits, argmax, max_value);
+    func lm_head_length_32<ios18>(tensor<fp16, [1, 2048, 1, 32]> hidden_states) {
+            tensor<fp16, [1, 2048, 1, 32]> final_norm_rmsnorm_abs = abs(x = hidden_states)[name = string("final_norm_rmsnorm_abs")];
+            tensor<int32, [1]> final_norm_rmsnorm_maxval_axes_1 = const()[name = string("final_norm_rmsnorm_maxval_axes_1"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_maxval_keep_dims_1 = const()[name = string("final_norm_rmsnorm_maxval_keep_dims_1"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> final_norm_rmsnorm_maxval = reduce_max(axes = final_norm_rmsnorm_maxval_axes_1, keep_dims = final_norm_rmsnorm_maxval_keep_dims_1, x = final_norm_rmsnorm_abs)[name = string("final_norm_rmsnorm_maxval")];
+            fp16 final_norm_rmsnorm_maxval_clipped_alpha_1 = const()[name = string("final_norm_rmsnorm_maxval_clipped_alpha_1"), val = fp16(0x1p-14)];
+            fp16 final_norm_rmsnorm_maxval_clipped_beta_1 = const()[name = string("final_norm_rmsnorm_maxval_clipped_beta_1"), val = fp16(inf)];
+            tensor<fp16, [1, 1, 1, 32]> final_norm_rmsnorm_maxval_clipped = clip(alpha = final_norm_rmsnorm_maxval_clipped_alpha_1, beta = final_norm_rmsnorm_maxval_clipped_beta_1, x = final_norm_rmsnorm_maxval)[name = string("final_norm_rmsnorm_maxval_clipped")];
+            tensor<fp16, [1, 2048, 1, 32]> final_norm_rmsnorm_scaled = real_div(x = hidden_states, y = final_norm_rmsnorm_maxval_clipped)[name = string("final_norm_rmsnorm_scaled")];
+            tensor<int32, [1]> final_norm_rmsnorm_squared_sum_axes_1 = const()[name = string("final_norm_rmsnorm_squared_sum_axes_1"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_squared_sum_keep_dims_1 = const()[name = string("final_norm_rmsnorm_squared_sum_keep_dims_1"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> final_norm_rmsnorm_squared_sum = reduce_sum_square(axes = final_norm_rmsnorm_squared_sum_axes_1, keep_dims = final_norm_rmsnorm_squared_sum_keep_dims_1, x = final_norm_rmsnorm_scaled)[name = string("final_norm_rmsnorm_squared_sum")];
+            fp16 final_norm_rmsnorm_rsqrt_epsilon_1 = const()[name = string("final_norm_rmsnorm_rsqrt_epsilon_1"), val = fp16(0x1p-14)];
+            tensor<fp16, [1, 1, 1, 32]> final_norm_rmsnorm_rsqrt = rsqrt(epsilon = final_norm_rmsnorm_rsqrt_epsilon_1, x = final_norm_rmsnorm_squared_sum)[name = string("final_norm_rmsnorm_rsqrt")];
+            fp16 final_norm_rmsnorm_dim_scaled_y_1 = const()[name = string("final_norm_rmsnorm_dim_scaled_y_1"), val = fp16(0x1.6ap+5)];
+            tensor<fp16, [1, 2048, 1, 32]> final_norm_rmsnorm_dim_scaled = mul(x = final_norm_rmsnorm_scaled, y = final_norm_rmsnorm_dim_scaled_y_1)[name = string("final_norm_rmsnorm_dim_scaled")];
+            tensor<fp16, [1, 2048, 1, 32]> final_norm_rmsnorm_normalized = mul(x = final_norm_rmsnorm_dim_scaled, y = final_norm_rmsnorm_rsqrt)[name = string("final_norm_rmsnorm_normalized")];
+            tensor<fp16, [2048, 1, 1]> final_norm_rmsnorm_y_1 = const()[name = string("final_norm_rmsnorm_y_1"), val = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 2048, 1, 32]> final_norm_rmsnorm = mul(x = final_norm_rmsnorm_normalized, y = final_norm_rmsnorm_y_1)[name = string("final_norm_rmsnorm")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_0_weight_1 = const()[name = string("logits_chunk_0_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4224)))];
+            tensor<int32, [2]> logits_chunk_0_strides_1 = const()[name = string("logits_chunk_0_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_0_pad_type_1 = const()[name = string("logits_chunk_0_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_0_pad_1 = const()[name = string("logits_chunk_0_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_0_dilations_1 = const()[name = string("logits_chunk_0_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_0_groups_1 = const()[name = string("logits_chunk_0_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_0 = conv(dilations = logits_chunk_0_dilations_1, groups = logits_chunk_0_groups_1, pad = logits_chunk_0_pad_1, pad_type = logits_chunk_0_pad_type_1, strides = logits_chunk_0_strides_1, weight = logits_chunk_0_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_0")];
+            int32 argmax_chunk_0_axis_0 = const()[name = string("argmax_chunk_0_axis_0"), val = int32(1)];
+            bool argmax_chunk_0_keep_dims_0 = const()[name = string("argmax_chunk_0_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_0_output_dtype_0 = const()[name = string("argmax_chunk_0_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> argmax_chunk_0 = reduce_argmax(axis = argmax_chunk_0_axis_0, keep_dims = argmax_chunk_0_keep_dims_0, output_dtype = argmax_chunk_0_output_dtype_0, x = logits_chunk_0)[name = string("argmax_chunk_0")];
+            tensor<int32, [1]> max_chunk_0_axes_0 = const()[name = string("max_chunk_0_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_0_keep_dims_0 = const()[name = string("max_chunk_0_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> max_chunk_0 = reduce_max(axes = max_chunk_0_axes_0, keep_dims = max_chunk_0_keep_dims_0, x = logits_chunk_0)[name = string("max_chunk_0")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_1_weight_1 = const()[name = string("logits_chunk_1_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8392896)))];
+            tensor<int32, [2]> logits_chunk_1_strides_1 = const()[name = string("logits_chunk_1_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_1_pad_type_1 = const()[name = string("logits_chunk_1_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_1_pad_1 = const()[name = string("logits_chunk_1_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_1_dilations_1 = const()[name = string("logits_chunk_1_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_1_groups_1 = const()[name = string("logits_chunk_1_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_1 = conv(dilations = logits_chunk_1_dilations_1, groups = logits_chunk_1_groups_1, pad = logits_chunk_1_pad_1, pad_type = logits_chunk_1_pad_type_1, strides = logits_chunk_1_strides_1, weight = logits_chunk_1_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_1")];
+            int32 argmax_chunk_1_axis_0 = const()[name = string("argmax_chunk_1_axis_0"), val = int32(1)];
+            bool argmax_chunk_1_keep_dims_0 = const()[name = string("argmax_chunk_1_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_1_output_dtype_0 = const()[name = string("argmax_chunk_1_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> argmax_chunk_1 = reduce_argmax(axis = argmax_chunk_1_axis_0, keep_dims = argmax_chunk_1_keep_dims_0, output_dtype = argmax_chunk_1_output_dtype_0, x = logits_chunk_1)[name = string("argmax_chunk_1")];
+            tensor<int32, [1]> max_chunk_1_axes_0 = const()[name = string("max_chunk_1_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_1_keep_dims_0 = const()[name = string("max_chunk_1_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> max_chunk_1 = reduce_max(axes = max_chunk_1_axes_0, keep_dims = max_chunk_1_keep_dims_0, x = logits_chunk_1)[name = string("max_chunk_1")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_2_weight_1 = const()[name = string("logits_chunk_2_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16781568)))];
+            tensor<int32, [2]> logits_chunk_2_strides_1 = const()[name = string("logits_chunk_2_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_2_pad_type_1 = const()[name = string("logits_chunk_2_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_2_pad_1 = const()[name = string("logits_chunk_2_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_2_dilations_1 = const()[name = string("logits_chunk_2_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_2_groups_1 = const()[name = string("logits_chunk_2_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_2 = conv(dilations = logits_chunk_2_dilations_1, groups = logits_chunk_2_groups_1, pad = logits_chunk_2_pad_1, pad_type = logits_chunk_2_pad_type_1, strides = logits_chunk_2_strides_1, weight = logits_chunk_2_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_2")];
+            int32 argmax_chunk_2_axis_0 = const()[name = string("argmax_chunk_2_axis_0"), val = int32(1)];
+            bool argmax_chunk_2_keep_dims_0 = const()[name = string("argmax_chunk_2_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_2_output_dtype_0 = const()[name = string("argmax_chunk_2_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> argmax_chunk_2 = reduce_argmax(axis = argmax_chunk_2_axis_0, keep_dims = argmax_chunk_2_keep_dims_0, output_dtype = argmax_chunk_2_output_dtype_0, x = logits_chunk_2)[name = string("argmax_chunk_2")];
+            tensor<int32, [1]> max_chunk_2_axes_0 = const()[name = string("max_chunk_2_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_2_keep_dims_0 = const()[name = string("max_chunk_2_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> max_chunk_2 = reduce_max(axes = max_chunk_2_axes_0, keep_dims = max_chunk_2_keep_dims_0, x = logits_chunk_2)[name = string("max_chunk_2")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_3_weight_1 = const()[name = string("logits_chunk_3_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25170240)))];
+            tensor<int32, [2]> logits_chunk_3_strides_1 = const()[name = string("logits_chunk_3_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_3_pad_type_1 = const()[name = string("logits_chunk_3_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_3_pad_1 = const()[name = string("logits_chunk_3_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_3_dilations_1 = const()[name = string("logits_chunk_3_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_3_groups_1 = const()[name = string("logits_chunk_3_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_3 = conv(dilations = logits_chunk_3_dilations_1, groups = logits_chunk_3_groups_1, pad = logits_chunk_3_pad_1, pad_type = logits_chunk_3_pad_type_1, strides = logits_chunk_3_strides_1, weight = logits_chunk_3_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_3")];
+            int32 argmax_chunk_3_axis_0 = const()[name = string("argmax_chunk_3_axis_0"), val = int32(1)];
+            bool argmax_chunk_3_keep_dims_0 = const()[name = string("argmax_chunk_3_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_3_output_dtype_0 = const()[name = string("argmax_chunk_3_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> argmax_chunk_3 = reduce_argmax(axis = argmax_chunk_3_axis_0, keep_dims = argmax_chunk_3_keep_dims_0, output_dtype = argmax_chunk_3_output_dtype_0, x = logits_chunk_3)[name = string("argmax_chunk_3")];
+            tensor<int32, [1]> max_chunk_3_axes_0 = const()[name = string("max_chunk_3_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_3_keep_dims_0 = const()[name = string("max_chunk_3_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> max_chunk_3 = reduce_max(axes = max_chunk_3_axes_0, keep_dims = max_chunk_3_keep_dims_0, x = logits_chunk_3)[name = string("max_chunk_3")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_4_weight_1 = const()[name = string("logits_chunk_4_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33558912)))];
+            tensor<int32, [2]> logits_chunk_4_strides_1 = const()[name = string("logits_chunk_4_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_4_pad_type_1 = const()[name = string("logits_chunk_4_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_4_pad_1 = const()[name = string("logits_chunk_4_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_4_dilations_1 = const()[name = string("logits_chunk_4_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_4_groups_1 = const()[name = string("logits_chunk_4_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_4 = conv(dilations = logits_chunk_4_dilations_1, groups = logits_chunk_4_groups_1, pad = logits_chunk_4_pad_1, pad_type = logits_chunk_4_pad_type_1, strides = logits_chunk_4_strides_1, weight = logits_chunk_4_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_4")];
+            int32 argmax_chunk_4_axis_0 = const()[name = string("argmax_chunk_4_axis_0"), val = int32(1)];
+            bool argmax_chunk_4_keep_dims_0 = const()[name = string("argmax_chunk_4_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_4_output_dtype_0 = const()[name = string("argmax_chunk_4_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> argmax_chunk_4 = reduce_argmax(axis = argmax_chunk_4_axis_0, keep_dims = argmax_chunk_4_keep_dims_0, output_dtype = argmax_chunk_4_output_dtype_0, x = logits_chunk_4)[name = string("argmax_chunk_4")];
+            tensor<int32, [1]> max_chunk_4_axes_0 = const()[name = string("max_chunk_4_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_4_keep_dims_0 = const()[name = string("max_chunk_4_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> max_chunk_4 = reduce_max(axes = max_chunk_4_axes_0, keep_dims = max_chunk_4_keep_dims_0, x = logits_chunk_4)[name = string("max_chunk_4")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_5_weight_1 = const()[name = string("logits_chunk_5_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41947584)))];
+            tensor<int32, [2]> logits_chunk_5_strides_1 = const()[name = string("logits_chunk_5_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_5_pad_type_1 = const()[name = string("logits_chunk_5_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_5_pad_1 = const()[name = string("logits_chunk_5_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_5_dilations_1 = const()[name = string("logits_chunk_5_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_5_groups_1 = const()[name = string("logits_chunk_5_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_5 = conv(dilations = logits_chunk_5_dilations_1, groups = logits_chunk_5_groups_1, pad = logits_chunk_5_pad_1, pad_type = logits_chunk_5_pad_type_1, strides = logits_chunk_5_strides_1, weight = logits_chunk_5_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_5")];
+            int32 argmax_chunk_5_axis_0 = const()[name = string("argmax_chunk_5_axis_0"), val = int32(1)];
+            bool argmax_chunk_5_keep_dims_0 = const()[name = string("argmax_chunk_5_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_5_output_dtype_0 = const()[name = string("argmax_chunk_5_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> argmax_chunk_5 = reduce_argmax(axis = argmax_chunk_5_axis_0, keep_dims = argmax_chunk_5_keep_dims_0, output_dtype = argmax_chunk_5_output_dtype_0, x = logits_chunk_5)[name = string("argmax_chunk_5")];
+            tensor<int32, [1]> max_chunk_5_axes_0 = const()[name = string("max_chunk_5_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_5_keep_dims_0 = const()[name = string("max_chunk_5_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> max_chunk_5 = reduce_max(axes = max_chunk_5_axes_0, keep_dims = max_chunk_5_keep_dims_0, x = logits_chunk_5)[name = string("max_chunk_5")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_6_weight_1 = const()[name = string("logits_chunk_6_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50336256)))];
+            tensor<int32, [2]> logits_chunk_6_strides_1 = const()[name = string("logits_chunk_6_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_6_pad_type_1 = const()[name = string("logits_chunk_6_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_6_pad_1 = const()[name = string("logits_chunk_6_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_6_dilations_1 = const()[name = string("logits_chunk_6_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_6_groups_1 = const()[name = string("logits_chunk_6_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_6 = conv(dilations = logits_chunk_6_dilations_1, groups = logits_chunk_6_groups_1, pad = logits_chunk_6_pad_1, pad_type = logits_chunk_6_pad_type_1, strides = logits_chunk_6_strides_1, weight = logits_chunk_6_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_6")];
+            int32 argmax_chunk_6_axis_0 = const()[name = string("argmax_chunk_6_axis_0"), val = int32(1)];
+            bool argmax_chunk_6_keep_dims_0 = const()[name = string("argmax_chunk_6_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_6_output_dtype_0 = const()[name = string("argmax_chunk_6_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> argmax_chunk_6 = reduce_argmax(axis = argmax_chunk_6_axis_0, keep_dims = argmax_chunk_6_keep_dims_0, output_dtype = argmax_chunk_6_output_dtype_0, x = logits_chunk_6)[name = string("argmax_chunk_6")];
+            tensor<int32, [1]> max_chunk_6_axes_0 = const()[name = string("max_chunk_6_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_6_keep_dims_0 = const()[name = string("max_chunk_6_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> max_chunk_6 = reduce_max(axes = max_chunk_6_axes_0, keep_dims = max_chunk_6_keep_dims_0, x = logits_chunk_6)[name = string("max_chunk_6")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_7_weight_1 = const()[name = string("logits_chunk_7_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58724928)))];
+            tensor<int32, [2]> logits_chunk_7_strides_1 = const()[name = string("logits_chunk_7_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_7_pad_type_1 = const()[name = string("logits_chunk_7_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_7_pad_1 = const()[name = string("logits_chunk_7_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_7_dilations_1 = const()[name = string("logits_chunk_7_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_7_groups_1 = const()[name = string("logits_chunk_7_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_7 = conv(dilations = logits_chunk_7_dilations_1, groups = logits_chunk_7_groups_1, pad = logits_chunk_7_pad_1, pad_type = logits_chunk_7_pad_type_1, strides = logits_chunk_7_strides_1, weight = logits_chunk_7_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_7")];
+            int32 argmax_chunk_7_axis_0 = const()[name = string("argmax_chunk_7_axis_0"), val = int32(1)];
+            bool argmax_chunk_7_keep_dims_0 = const()[name = string("argmax_chunk_7_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_7_output_dtype_0 = const()[name = string("argmax_chunk_7_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> argmax_chunk_7 = reduce_argmax(axis = argmax_chunk_7_axis_0, keep_dims = argmax_chunk_7_keep_dims_0, output_dtype = argmax_chunk_7_output_dtype_0, x = logits_chunk_7)[name = string("argmax_chunk_7")];
+            tensor<int32, [1]> max_chunk_7_axes_0 = const()[name = string("max_chunk_7_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_7_keep_dims_0 = const()[name = string("max_chunk_7_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> max_chunk_7 = reduce_max(axes = max_chunk_7_axes_0, keep_dims = max_chunk_7_keep_dims_0, x = logits_chunk_7)[name = string("max_chunk_7")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_8_weight_1 = const()[name = string("logits_chunk_8_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67113600)))];
+            tensor<int32, [2]> logits_chunk_8_strides_1 = const()[name = string("logits_chunk_8_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_8_pad_type_1 = const()[name = string("logits_chunk_8_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_8_pad_1 = const()[name = string("logits_chunk_8_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_8_dilations_1 = const()[name = string("logits_chunk_8_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_8_groups_1 = const()[name = string("logits_chunk_8_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_8 = conv(dilations = logits_chunk_8_dilations_1, groups = logits_chunk_8_groups_1, pad = logits_chunk_8_pad_1, pad_type = logits_chunk_8_pad_type_1, strides = logits_chunk_8_strides_1, weight = logits_chunk_8_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_8")];
+            int32 argmax_chunk_8_axis_0 = const()[name = string("argmax_chunk_8_axis_0"), val = int32(1)];
+            bool argmax_chunk_8_keep_dims_0 = const()[name = string("argmax_chunk_8_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_8_output_dtype_0 = const()[name = string("argmax_chunk_8_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> argmax_chunk_8 = reduce_argmax(axis = argmax_chunk_8_axis_0, keep_dims = argmax_chunk_8_keep_dims_0, output_dtype = argmax_chunk_8_output_dtype_0, x = logits_chunk_8)[name = string("argmax_chunk_8")];
+            tensor<int32, [1]> max_chunk_8_axes_0 = const()[name = string("max_chunk_8_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_8_keep_dims_0 = const()[name = string("max_chunk_8_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> max_chunk_8 = reduce_max(axes = max_chunk_8_axes_0, keep_dims = max_chunk_8_keep_dims_0, x = logits_chunk_8)[name = string("max_chunk_8")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_9_weight_1 = const()[name = string("logits_chunk_9_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75502272)))];
+            tensor<int32, [2]> logits_chunk_9_strides_1 = const()[name = string("logits_chunk_9_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_9_pad_type_1 = const()[name = string("logits_chunk_9_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_9_pad_1 = const()[name = string("logits_chunk_9_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_9_dilations_1 = const()[name = string("logits_chunk_9_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_9_groups_1 = const()[name = string("logits_chunk_9_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_9 = conv(dilations = logits_chunk_9_dilations_1, groups = logits_chunk_9_groups_1, pad = logits_chunk_9_pad_1, pad_type = logits_chunk_9_pad_type_1, strides = logits_chunk_9_strides_1, weight = logits_chunk_9_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_9")];
+            int32 argmax_chunk_9_axis_0 = const()[name = string("argmax_chunk_9_axis_0"), val = int32(1)];
+            bool argmax_chunk_9_keep_dims_0 = const()[name = string("argmax_chunk_9_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_9_output_dtype_0 = const()[name = string("argmax_chunk_9_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> argmax_chunk_9 = reduce_argmax(axis = argmax_chunk_9_axis_0, keep_dims = argmax_chunk_9_keep_dims_0, output_dtype = argmax_chunk_9_output_dtype_0, x = logits_chunk_9)[name = string("argmax_chunk_9")];
+            tensor<int32, [1]> max_chunk_9_axes_0 = const()[name = string("max_chunk_9_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_9_keep_dims_0 = const()[name = string("max_chunk_9_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> max_chunk_9 = reduce_max(axes = max_chunk_9_axes_0, keep_dims = max_chunk_9_keep_dims_0, x = logits_chunk_9)[name = string("max_chunk_9")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_10_weight_1 = const()[name = string("logits_chunk_10_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83890944)))];
+            tensor<int32, [2]> logits_chunk_10_strides_1 = const()[name = string("logits_chunk_10_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_10_pad_type_1 = const()[name = string("logits_chunk_10_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_10_pad_1 = const()[name = string("logits_chunk_10_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_10_dilations_1 = const()[name = string("logits_chunk_10_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_10_groups_1 = const()[name = string("logits_chunk_10_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_10 = conv(dilations = logits_chunk_10_dilations_1, groups = logits_chunk_10_groups_1, pad = logits_chunk_10_pad_1, pad_type = logits_chunk_10_pad_type_1, strides = logits_chunk_10_strides_1, weight = logits_chunk_10_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_10")];
+            int32 argmax_chunk_10_axis_0 = const()[name = string("argmax_chunk_10_axis_0"), val = int32(1)];
+            bool argmax_chunk_10_keep_dims_0 = const()[name = string("argmax_chunk_10_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_10_output_dtype_0 = const()[name = string("argmax_chunk_10_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> argmax_chunk_10 = reduce_argmax(axis = argmax_chunk_10_axis_0, keep_dims = argmax_chunk_10_keep_dims_0, output_dtype = argmax_chunk_10_output_dtype_0, x = logits_chunk_10)[name = string("argmax_chunk_10")];
+            tensor<int32, [1]> max_chunk_10_axes_0 = const()[name = string("max_chunk_10_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_10_keep_dims_0 = const()[name = string("max_chunk_10_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> max_chunk_10 = reduce_max(axes = max_chunk_10_axes_0, keep_dims = max_chunk_10_keep_dims_0, x = logits_chunk_10)[name = string("max_chunk_10")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_11_weight_1 = const()[name = string("logits_chunk_11_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92279616)))];
+            tensor<int32, [2]> logits_chunk_11_strides_1 = const()[name = string("logits_chunk_11_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_11_pad_type_1 = const()[name = string("logits_chunk_11_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_11_pad_1 = const()[name = string("logits_chunk_11_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_11_dilations_1 = const()[name = string("logits_chunk_11_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_11_groups_1 = const()[name = string("logits_chunk_11_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_11 = conv(dilations = logits_chunk_11_dilations_1, groups = logits_chunk_11_groups_1, pad = logits_chunk_11_pad_1, pad_type = logits_chunk_11_pad_type_1, strides = logits_chunk_11_strides_1, weight = logits_chunk_11_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_11")];
+            int32 argmax_chunk_11_axis_0 = const()[name = string("argmax_chunk_11_axis_0"), val = int32(1)];
+            bool argmax_chunk_11_keep_dims_0 = const()[name = string("argmax_chunk_11_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_11_output_dtype_0 = const()[name = string("argmax_chunk_11_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> argmax_chunk_11 = reduce_argmax(axis = argmax_chunk_11_axis_0, keep_dims = argmax_chunk_11_keep_dims_0, output_dtype = argmax_chunk_11_output_dtype_0, x = logits_chunk_11)[name = string("argmax_chunk_11")];
+            tensor<int32, [1]> max_chunk_11_axes_0 = const()[name = string("max_chunk_11_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_11_keep_dims_0 = const()[name = string("max_chunk_11_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> max_chunk_11 = reduce_max(axes = max_chunk_11_axes_0, keep_dims = max_chunk_11_keep_dims_0, x = logits_chunk_11)[name = string("max_chunk_11")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_12_weight_1 = const()[name = string("logits_chunk_12_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100668288)))];
+            tensor<int32, [2]> logits_chunk_12_strides_1 = const()[name = string("logits_chunk_12_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_12_pad_type_1 = const()[name = string("logits_chunk_12_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_12_pad_1 = const()[name = string("logits_chunk_12_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_12_dilations_1 = const()[name = string("logits_chunk_12_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_12_groups_1 = const()[name = string("logits_chunk_12_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_12 = conv(dilations = logits_chunk_12_dilations_1, groups = logits_chunk_12_groups_1, pad = logits_chunk_12_pad_1, pad_type = logits_chunk_12_pad_type_1, strides = logits_chunk_12_strides_1, weight = logits_chunk_12_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_12")];
+            int32 argmax_chunk_12_axis_0 = const()[name = string("argmax_chunk_12_axis_0"), val = int32(1)];
+            bool argmax_chunk_12_keep_dims_0 = const()[name = string("argmax_chunk_12_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_12_output_dtype_0 = const()[name = string("argmax_chunk_12_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> argmax_chunk_12 = reduce_argmax(axis = argmax_chunk_12_axis_0, keep_dims = argmax_chunk_12_keep_dims_0, output_dtype = argmax_chunk_12_output_dtype_0, x = logits_chunk_12)[name = string("argmax_chunk_12")];
+            tensor<int32, [1]> max_chunk_12_axes_0 = const()[name = string("max_chunk_12_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_12_keep_dims_0 = const()[name = string("max_chunk_12_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> max_chunk_12 = reduce_max(axes = max_chunk_12_axes_0, keep_dims = max_chunk_12_keep_dims_0, x = logits_chunk_12)[name = string("max_chunk_12")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_13_weight_1 = const()[name = string("logits_chunk_13_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109056960)))];
+            tensor<int32, [2]> logits_chunk_13_strides_1 = const()[name = string("logits_chunk_13_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_13_pad_type_1 = const()[name = string("logits_chunk_13_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_13_pad_1 = const()[name = string("logits_chunk_13_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_13_dilations_1 = const()[name = string("logits_chunk_13_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_13_groups_1 = const()[name = string("logits_chunk_13_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_13 = conv(dilations = logits_chunk_13_dilations_1, groups = logits_chunk_13_groups_1, pad = logits_chunk_13_pad_1, pad_type = logits_chunk_13_pad_type_1, strides = logits_chunk_13_strides_1, weight = logits_chunk_13_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_13")];
+            int32 argmax_chunk_13_axis_0 = const()[name = string("argmax_chunk_13_axis_0"), val = int32(1)];
+            bool argmax_chunk_13_keep_dims_0 = const()[name = string("argmax_chunk_13_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_13_output_dtype_0 = const()[name = string("argmax_chunk_13_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> argmax_chunk_13 = reduce_argmax(axis = argmax_chunk_13_axis_0, keep_dims = argmax_chunk_13_keep_dims_0, output_dtype = argmax_chunk_13_output_dtype_0, x = logits_chunk_13)[name = string("argmax_chunk_13")];
+            tensor<int32, [1]> max_chunk_13_axes_0 = const()[name = string("max_chunk_13_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_13_keep_dims_0 = const()[name = string("max_chunk_13_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> max_chunk_13 = reduce_max(axes = max_chunk_13_axes_0, keep_dims = max_chunk_13_keep_dims_0, x = logits_chunk_13)[name = string("max_chunk_13")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_14_weight_1 = const()[name = string("logits_chunk_14_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117445632)))];
+            tensor<int32, [2]> logits_chunk_14_strides_1 = const()[name = string("logits_chunk_14_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_14_pad_type_1 = const()[name = string("logits_chunk_14_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_14_pad_1 = const()[name = string("logits_chunk_14_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_14_dilations_1 = const()[name = string("logits_chunk_14_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_14_groups_1 = const()[name = string("logits_chunk_14_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_14 = conv(dilations = logits_chunk_14_dilations_1, groups = logits_chunk_14_groups_1, pad = logits_chunk_14_pad_1, pad_type = logits_chunk_14_pad_type_1, strides = logits_chunk_14_strides_1, weight = logits_chunk_14_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_14")];
+            int32 argmax_chunk_14_axis_0 = const()[name = string("argmax_chunk_14_axis_0"), val = int32(1)];
+            bool argmax_chunk_14_keep_dims_0 = const()[name = string("argmax_chunk_14_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_14_output_dtype_0 = const()[name = string("argmax_chunk_14_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> argmax_chunk_14 = reduce_argmax(axis = argmax_chunk_14_axis_0, keep_dims = argmax_chunk_14_keep_dims_0, output_dtype = argmax_chunk_14_output_dtype_0, x = logits_chunk_14)[name = string("argmax_chunk_14")];
+            tensor<int32, [1]> max_chunk_14_axes_0 = const()[name = string("max_chunk_14_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_14_keep_dims_0 = const()[name = string("max_chunk_14_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> max_chunk_14 = reduce_max(axes = max_chunk_14_axes_0, keep_dims = max_chunk_14_keep_dims_0, x = logits_chunk_14)[name = string("max_chunk_14")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_15_weight_1 = const()[name = string("logits_chunk_15_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125834304)))];
+            tensor<int32, [2]> logits_chunk_15_strides_1 = const()[name = string("logits_chunk_15_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_15_pad_type_1 = const()[name = string("logits_chunk_15_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_15_pad_1 = const()[name = string("logits_chunk_15_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_15_dilations_1 = const()[name = string("logits_chunk_15_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_15_groups_1 = const()[name = string("logits_chunk_15_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_15 = conv(dilations = logits_chunk_15_dilations_1, groups = logits_chunk_15_groups_1, pad = logits_chunk_15_pad_1, pad_type = logits_chunk_15_pad_type_1, strides = logits_chunk_15_strides_1, weight = logits_chunk_15_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_15")];
+            int32 argmax_chunk_15_axis_0 = const()[name = string("argmax_chunk_15_axis_0"), val = int32(1)];
+            bool argmax_chunk_15_keep_dims_0 = const()[name = string("argmax_chunk_15_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_15_output_dtype_0 = const()[name = string("argmax_chunk_15_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> argmax_chunk_15 = reduce_argmax(axis = argmax_chunk_15_axis_0, keep_dims = argmax_chunk_15_keep_dims_0, output_dtype = argmax_chunk_15_output_dtype_0, x = logits_chunk_15)[name = string("argmax_chunk_15")];
+            tensor<int32, [1]> max_chunk_15_axes_0 = const()[name = string("max_chunk_15_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_15_keep_dims_0 = const()[name = string("max_chunk_15_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> max_chunk_15 = reduce_max(axes = max_chunk_15_axes_0, keep_dims = max_chunk_15_keep_dims_0, x = logits_chunk_15)[name = string("max_chunk_15")];
+            int32 logits_axis_0 = const()[name = string("logits_axis_0"), val = int32(1)];
+            bool logits_interleave_0 = const()[name = string("logits_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 32768, 1, 32]> logits = concat(axis = logits_axis_0, interleave = logits_interleave_0, values = (logits_chunk_0, logits_chunk_1, logits_chunk_2, logits_chunk_3, logits_chunk_4, logits_chunk_5, logits_chunk_6, logits_chunk_7, logits_chunk_8, logits_chunk_9, logits_chunk_10, logits_chunk_11, logits_chunk_12, logits_chunk_13, logits_chunk_14, logits_chunk_15))[name = string("logits")];
+            int32 values_axis_0 = const()[name = string("values_axis_0"), val = int32(1)];
+            bool values_interleave_0 = const()[name = string("values_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 32]> values = concat(axis = values_axis_0, interleave = values_interleave_0, values = (max_chunk_0, max_chunk_1, max_chunk_2, max_chunk_3, max_chunk_4, max_chunk_5, max_chunk_6, max_chunk_7, max_chunk_8, max_chunk_9, max_chunk_10, max_chunk_11, max_chunk_12, max_chunk_13, max_chunk_14, max_chunk_15))[name = string("values")];
+            tensor<int32, [1]> max_value_axes_0 = const()[name = string("max_value_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_value_keep_dims_0 = const()[name = string("max_value_keep_dims_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 32]> max_value = reduce_max(axes = max_value_axes_0, keep_dims = max_value_keep_dims_0, x = values)[name = string("max_value")];
+            int32 max_value_index_axis_0 = const()[name = string("max_value_index_axis_0"), val = int32(1)];
+            bool max_value_index_keep_dims_0 = const()[name = string("max_value_index_keep_dims_0"), val = bool(true)];
+            string max_value_index_output_dtype_0 = const()[name = string("max_value_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 1, 1, 32]> max_value_index = reduce_argmax(axis = max_value_index_axis_0, keep_dims = max_value_index_keep_dims_0, output_dtype = max_value_index_output_dtype_0, x = values)[name = string("max_value_index")];
+            string indices_chunk_0_int32_dtype_1 = const()[name = string("indices_chunk_0_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_1_int32_dtype_1 = const()[name = string("indices_chunk_1_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_2_int32_dtype_1 = const()[name = string("indices_chunk_2_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_3_int32_dtype_1 = const()[name = string("indices_chunk_3_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_4_int32_dtype_1 = const()[name = string("indices_chunk_4_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_5_int32_dtype_1 = const()[name = string("indices_chunk_5_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_6_int32_dtype_1 = const()[name = string("indices_chunk_6_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_7_int32_dtype_1 = const()[name = string("indices_chunk_7_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_8_int32_dtype_1 = const()[name = string("indices_chunk_8_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_9_int32_dtype_1 = const()[name = string("indices_chunk_9_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_10_int32_dtype_1 = const()[name = string("indices_chunk_10_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_11_int32_dtype_1 = const()[name = string("indices_chunk_11_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_12_int32_dtype_1 = const()[name = string("indices_chunk_12_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_13_int32_dtype_1 = const()[name = string("indices_chunk_13_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_14_int32_dtype_1 = const()[name = string("indices_chunk_14_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_15_int32_dtype_1 = const()[name = string("indices_chunk_15_int32_dtype_1"), val = string("int32")];
+            int32 indices_axis_1 = const()[name = string("indices_axis_1"), val = int32(1)];
+            bool indices_interleave_1 = const()[name = string("indices_interleave_1"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_15_int32 = cast(dtype = indices_chunk_15_int32_dtype_1, x = argmax_chunk_15)[name = string("cast_64")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_14_int32 = cast(dtype = indices_chunk_14_int32_dtype_1, x = argmax_chunk_14)[name = string("cast_65")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_13_int32 = cast(dtype = indices_chunk_13_int32_dtype_1, x = argmax_chunk_13)[name = string("cast_66")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_12_int32 = cast(dtype = indices_chunk_12_int32_dtype_1, x = argmax_chunk_12)[name = string("cast_67")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_11_int32 = cast(dtype = indices_chunk_11_int32_dtype_1, x = argmax_chunk_11)[name = string("cast_68")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_10_int32 = cast(dtype = indices_chunk_10_int32_dtype_1, x = argmax_chunk_10)[name = string("cast_69")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_9_int32 = cast(dtype = indices_chunk_9_int32_dtype_1, x = argmax_chunk_9)[name = string("cast_70")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_8_int32 = cast(dtype = indices_chunk_8_int32_dtype_1, x = argmax_chunk_8)[name = string("cast_71")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_7_int32 = cast(dtype = indices_chunk_7_int32_dtype_1, x = argmax_chunk_7)[name = string("cast_72")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_6_int32 = cast(dtype = indices_chunk_6_int32_dtype_1, x = argmax_chunk_6)[name = string("cast_73")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_5_int32 = cast(dtype = indices_chunk_5_int32_dtype_1, x = argmax_chunk_5)[name = string("cast_74")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_4_int32 = cast(dtype = indices_chunk_4_int32_dtype_1, x = argmax_chunk_4)[name = string("cast_75")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_3_int32 = cast(dtype = indices_chunk_3_int32_dtype_1, x = argmax_chunk_3)[name = string("cast_76")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_2_int32 = cast(dtype = indices_chunk_2_int32_dtype_1, x = argmax_chunk_2)[name = string("cast_77")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_1_int32 = cast(dtype = indices_chunk_1_int32_dtype_1, x = argmax_chunk_1)[name = string("cast_78")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_0_int32 = cast(dtype = indices_chunk_0_int32_dtype_1, x = argmax_chunk_0)[name = string("cast_79")];
+            tensor<int32, [1, 16, 1, 32]> indices = concat(axis = indices_axis_1, interleave = indices_interleave_1, values = (indices_chunk_0_int32, indices_chunk_1_int32, indices_chunk_2_int32, indices_chunk_3_int32, indices_chunk_4_int32, indices_chunk_5_int32, indices_chunk_6_int32, indices_chunk_7_int32, indices_chunk_8_int32, indices_chunk_9_int32, indices_chunk_10_int32, indices_chunk_11_int32, indices_chunk_12_int32, indices_chunk_13_int32, indices_chunk_14_int32, indices_chunk_15_int32))[name = string("indices")];
+            int32 argmax_chunks_axis_1 = const()[name = string("argmax_chunks_axis_1"), val = int32(1)];
+            bool argmax_chunks_validate_indices_1 = const()[name = string("argmax_chunks_validate_indices_1"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 32]> argmax_chunks = gather_along_axis(axis = argmax_chunks_axis_1, indices = max_value_index, validate_indices = argmax_chunks_validate_indices_1, x = indices)[name = string("argmax_chunks")];
+            int32 mul_1_x_0 = const()[name = string("mul_1_x_0"), val = int32(2048)];
+            tensor<int32, [1, 1, 1, 32]> mul_1 = mul(x = mul_1_x_0, y = max_value_index)[name = string("mul_1")];
+            tensor<int32, [1, 1, 1, 32]> argmax = add(x = argmax_chunks, y = mul_1)[name = string("argmax")];
+        } -> (logits, argmax, max_value);
+    func lm_head_length_48<ios18>(tensor<fp16, [1, 2048, 1, 48]> hidden_states) {
+            tensor<fp16, [1, 2048, 1, 48]> final_norm_rmsnorm_abs = abs(x = hidden_states)[name = string("final_norm_rmsnorm_abs")];
+            tensor<int32, [1]> final_norm_rmsnorm_maxval_axes_1 = const()[name = string("final_norm_rmsnorm_maxval_axes_1"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_maxval_keep_dims_1 = const()[name = string("final_norm_rmsnorm_maxval_keep_dims_1"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> final_norm_rmsnorm_maxval = reduce_max(axes = final_norm_rmsnorm_maxval_axes_1, keep_dims = final_norm_rmsnorm_maxval_keep_dims_1, x = final_norm_rmsnorm_abs)[name = string("final_norm_rmsnorm_maxval")];
+            fp16 final_norm_rmsnorm_maxval_clipped_alpha_1 = const()[name = string("final_norm_rmsnorm_maxval_clipped_alpha_1"), val = fp16(0x1p-14)];
+            fp16 final_norm_rmsnorm_maxval_clipped_beta_1 = const()[name = string("final_norm_rmsnorm_maxval_clipped_beta_1"), val = fp16(inf)];
+            tensor<fp16, [1, 1, 1, 48]> final_norm_rmsnorm_maxval_clipped = clip(alpha = final_norm_rmsnorm_maxval_clipped_alpha_1, beta = final_norm_rmsnorm_maxval_clipped_beta_1, x = final_norm_rmsnorm_maxval)[name = string("final_norm_rmsnorm_maxval_clipped")];
+            tensor<fp16, [1, 2048, 1, 48]> final_norm_rmsnorm_scaled = real_div(x = hidden_states, y = final_norm_rmsnorm_maxval_clipped)[name = string("final_norm_rmsnorm_scaled")];
+            tensor<int32, [1]> final_norm_rmsnorm_squared_sum_axes_1 = const()[name = string("final_norm_rmsnorm_squared_sum_axes_1"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_squared_sum_keep_dims_1 = const()[name = string("final_norm_rmsnorm_squared_sum_keep_dims_1"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> final_norm_rmsnorm_squared_sum = reduce_sum_square(axes = final_norm_rmsnorm_squared_sum_axes_1, keep_dims = final_norm_rmsnorm_squared_sum_keep_dims_1, x = final_norm_rmsnorm_scaled)[name = string("final_norm_rmsnorm_squared_sum")];
+            fp16 final_norm_rmsnorm_rsqrt_epsilon_1 = const()[name = string("final_norm_rmsnorm_rsqrt_epsilon_1"), val = fp16(0x1p-14)];
+            tensor<fp16, [1, 1, 1, 48]> final_norm_rmsnorm_rsqrt = rsqrt(epsilon = final_norm_rmsnorm_rsqrt_epsilon_1, x = final_norm_rmsnorm_squared_sum)[name = string("final_norm_rmsnorm_rsqrt")];
+            fp16 final_norm_rmsnorm_dim_scaled_y_1 = const()[name = string("final_norm_rmsnorm_dim_scaled_y_1"), val = fp16(0x1.6ap+5)];
+            tensor<fp16, [1, 2048, 1, 48]> final_norm_rmsnorm_dim_scaled = mul(x = final_norm_rmsnorm_scaled, y = final_norm_rmsnorm_dim_scaled_y_1)[name = string("final_norm_rmsnorm_dim_scaled")];
+            tensor<fp16, [1, 2048, 1, 48]> final_norm_rmsnorm_normalized = mul(x = final_norm_rmsnorm_dim_scaled, y = final_norm_rmsnorm_rsqrt)[name = string("final_norm_rmsnorm_normalized")];
+            tensor<fp16, [2048, 1, 1]> final_norm_rmsnorm_y_1 = const()[name = string("final_norm_rmsnorm_y_1"), val = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 2048, 1, 48]> final_norm_rmsnorm = mul(x = final_norm_rmsnorm_normalized, y = final_norm_rmsnorm_y_1)[name = string("final_norm_rmsnorm")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_0_weight_1 = const()[name = string("logits_chunk_0_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4224)))];
+            tensor<int32, [2]> logits_chunk_0_strides_1 = const()[name = string("logits_chunk_0_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_0_pad_type_1 = const()[name = string("logits_chunk_0_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_0_pad_1 = const()[name = string("logits_chunk_0_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_0_dilations_1 = const()[name = string("logits_chunk_0_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_0_groups_1 = const()[name = string("logits_chunk_0_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_0 = conv(dilations = logits_chunk_0_dilations_1, groups = logits_chunk_0_groups_1, pad = logits_chunk_0_pad_1, pad_type = logits_chunk_0_pad_type_1, strides = logits_chunk_0_strides_1, weight = logits_chunk_0_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_0")];
+            int32 argmax_chunk_0_axis_0 = const()[name = string("argmax_chunk_0_axis_0"), val = int32(1)];
+            bool argmax_chunk_0_keep_dims_0 = const()[name = string("argmax_chunk_0_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_0_output_dtype_0 = const()[name = string("argmax_chunk_0_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> argmax_chunk_0 = reduce_argmax(axis = argmax_chunk_0_axis_0, keep_dims = argmax_chunk_0_keep_dims_0, output_dtype = argmax_chunk_0_output_dtype_0, x = logits_chunk_0)[name = string("argmax_chunk_0")];
+            tensor<int32, [1]> max_chunk_0_axes_0 = const()[name = string("max_chunk_0_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_0_keep_dims_0 = const()[name = string("max_chunk_0_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> max_chunk_0 = reduce_max(axes = max_chunk_0_axes_0, keep_dims = max_chunk_0_keep_dims_0, x = logits_chunk_0)[name = string("max_chunk_0")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_1_weight_1 = const()[name = string("logits_chunk_1_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8392896)))];
+            tensor<int32, [2]> logits_chunk_1_strides_1 = const()[name = string("logits_chunk_1_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_1_pad_type_1 = const()[name = string("logits_chunk_1_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_1_pad_1 = const()[name = string("logits_chunk_1_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_1_dilations_1 = const()[name = string("logits_chunk_1_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_1_groups_1 = const()[name = string("logits_chunk_1_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_1 = conv(dilations = logits_chunk_1_dilations_1, groups = logits_chunk_1_groups_1, pad = logits_chunk_1_pad_1, pad_type = logits_chunk_1_pad_type_1, strides = logits_chunk_1_strides_1, weight = logits_chunk_1_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_1")];
+            int32 argmax_chunk_1_axis_0 = const()[name = string("argmax_chunk_1_axis_0"), val = int32(1)];
+            bool argmax_chunk_1_keep_dims_0 = const()[name = string("argmax_chunk_1_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_1_output_dtype_0 = const()[name = string("argmax_chunk_1_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> argmax_chunk_1 = reduce_argmax(axis = argmax_chunk_1_axis_0, keep_dims = argmax_chunk_1_keep_dims_0, output_dtype = argmax_chunk_1_output_dtype_0, x = logits_chunk_1)[name = string("argmax_chunk_1")];
+            tensor<int32, [1]> max_chunk_1_axes_0 = const()[name = string("max_chunk_1_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_1_keep_dims_0 = const()[name = string("max_chunk_1_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> max_chunk_1 = reduce_max(axes = max_chunk_1_axes_0, keep_dims = max_chunk_1_keep_dims_0, x = logits_chunk_1)[name = string("max_chunk_1")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_2_weight_1 = const()[name = string("logits_chunk_2_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16781568)))];
+            tensor<int32, [2]> logits_chunk_2_strides_1 = const()[name = string("logits_chunk_2_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_2_pad_type_1 = const()[name = string("logits_chunk_2_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_2_pad_1 = const()[name = string("logits_chunk_2_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_2_dilations_1 = const()[name = string("logits_chunk_2_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_2_groups_1 = const()[name = string("logits_chunk_2_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_2 = conv(dilations = logits_chunk_2_dilations_1, groups = logits_chunk_2_groups_1, pad = logits_chunk_2_pad_1, pad_type = logits_chunk_2_pad_type_1, strides = logits_chunk_2_strides_1, weight = logits_chunk_2_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_2")];
+            int32 argmax_chunk_2_axis_0 = const()[name = string("argmax_chunk_2_axis_0"), val = int32(1)];
+            bool argmax_chunk_2_keep_dims_0 = const()[name = string("argmax_chunk_2_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_2_output_dtype_0 = const()[name = string("argmax_chunk_2_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> argmax_chunk_2 = reduce_argmax(axis = argmax_chunk_2_axis_0, keep_dims = argmax_chunk_2_keep_dims_0, output_dtype = argmax_chunk_2_output_dtype_0, x = logits_chunk_2)[name = string("argmax_chunk_2")];
+            tensor<int32, [1]> max_chunk_2_axes_0 = const()[name = string("max_chunk_2_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_2_keep_dims_0 = const()[name = string("max_chunk_2_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> max_chunk_2 = reduce_max(axes = max_chunk_2_axes_0, keep_dims = max_chunk_2_keep_dims_0, x = logits_chunk_2)[name = string("max_chunk_2")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_3_weight_1 = const()[name = string("logits_chunk_3_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25170240)))];
+            tensor<int32, [2]> logits_chunk_3_strides_1 = const()[name = string("logits_chunk_3_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_3_pad_type_1 = const()[name = string("logits_chunk_3_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_3_pad_1 = const()[name = string("logits_chunk_3_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_3_dilations_1 = const()[name = string("logits_chunk_3_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_3_groups_1 = const()[name = string("logits_chunk_3_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_3 = conv(dilations = logits_chunk_3_dilations_1, groups = logits_chunk_3_groups_1, pad = logits_chunk_3_pad_1, pad_type = logits_chunk_3_pad_type_1, strides = logits_chunk_3_strides_1, weight = logits_chunk_3_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_3")];
+            int32 argmax_chunk_3_axis_0 = const()[name = string("argmax_chunk_3_axis_0"), val = int32(1)];
+            bool argmax_chunk_3_keep_dims_0 = const()[name = string("argmax_chunk_3_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_3_output_dtype_0 = const()[name = string("argmax_chunk_3_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> argmax_chunk_3 = reduce_argmax(axis = argmax_chunk_3_axis_0, keep_dims = argmax_chunk_3_keep_dims_0, output_dtype = argmax_chunk_3_output_dtype_0, x = logits_chunk_3)[name = string("argmax_chunk_3")];
+            tensor<int32, [1]> max_chunk_3_axes_0 = const()[name = string("max_chunk_3_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_3_keep_dims_0 = const()[name = string("max_chunk_3_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> max_chunk_3 = reduce_max(axes = max_chunk_3_axes_0, keep_dims = max_chunk_3_keep_dims_0, x = logits_chunk_3)[name = string("max_chunk_3")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_4_weight_1 = const()[name = string("logits_chunk_4_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33558912)))];
+            tensor<int32, [2]> logits_chunk_4_strides_1 = const()[name = string("logits_chunk_4_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_4_pad_type_1 = const()[name = string("logits_chunk_4_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_4_pad_1 = const()[name = string("logits_chunk_4_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_4_dilations_1 = const()[name = string("logits_chunk_4_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_4_groups_1 = const()[name = string("logits_chunk_4_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_4 = conv(dilations = logits_chunk_4_dilations_1, groups = logits_chunk_4_groups_1, pad = logits_chunk_4_pad_1, pad_type = logits_chunk_4_pad_type_1, strides = logits_chunk_4_strides_1, weight = logits_chunk_4_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_4")];
+            int32 argmax_chunk_4_axis_0 = const()[name = string("argmax_chunk_4_axis_0"), val = int32(1)];
+            bool argmax_chunk_4_keep_dims_0 = const()[name = string("argmax_chunk_4_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_4_output_dtype_0 = const()[name = string("argmax_chunk_4_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> argmax_chunk_4 = reduce_argmax(axis = argmax_chunk_4_axis_0, keep_dims = argmax_chunk_4_keep_dims_0, output_dtype = argmax_chunk_4_output_dtype_0, x = logits_chunk_4)[name = string("argmax_chunk_4")];
+            tensor<int32, [1]> max_chunk_4_axes_0 = const()[name = string("max_chunk_4_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_4_keep_dims_0 = const()[name = string("max_chunk_4_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> max_chunk_4 = reduce_max(axes = max_chunk_4_axes_0, keep_dims = max_chunk_4_keep_dims_0, x = logits_chunk_4)[name = string("max_chunk_4")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_5_weight_1 = const()[name = string("logits_chunk_5_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41947584)))];
+            tensor<int32, [2]> logits_chunk_5_strides_1 = const()[name = string("logits_chunk_5_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_5_pad_type_1 = const()[name = string("logits_chunk_5_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_5_pad_1 = const()[name = string("logits_chunk_5_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_5_dilations_1 = const()[name = string("logits_chunk_5_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_5_groups_1 = const()[name = string("logits_chunk_5_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_5 = conv(dilations = logits_chunk_5_dilations_1, groups = logits_chunk_5_groups_1, pad = logits_chunk_5_pad_1, pad_type = logits_chunk_5_pad_type_1, strides = logits_chunk_5_strides_1, weight = logits_chunk_5_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_5")];
+            int32 argmax_chunk_5_axis_0 = const()[name = string("argmax_chunk_5_axis_0"), val = int32(1)];
+            bool argmax_chunk_5_keep_dims_0 = const()[name = string("argmax_chunk_5_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_5_output_dtype_0 = const()[name = string("argmax_chunk_5_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> argmax_chunk_5 = reduce_argmax(axis = argmax_chunk_5_axis_0, keep_dims = argmax_chunk_5_keep_dims_0, output_dtype = argmax_chunk_5_output_dtype_0, x = logits_chunk_5)[name = string("argmax_chunk_5")];
+            tensor<int32, [1]> max_chunk_5_axes_0 = const()[name = string("max_chunk_5_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_5_keep_dims_0 = const()[name = string("max_chunk_5_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> max_chunk_5 = reduce_max(axes = max_chunk_5_axes_0, keep_dims = max_chunk_5_keep_dims_0, x = logits_chunk_5)[name = string("max_chunk_5")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_6_weight_1 = const()[name = string("logits_chunk_6_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50336256)))];
+            tensor<int32, [2]> logits_chunk_6_strides_1 = const()[name = string("logits_chunk_6_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_6_pad_type_1 = const()[name = string("logits_chunk_6_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_6_pad_1 = const()[name = string("logits_chunk_6_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_6_dilations_1 = const()[name = string("logits_chunk_6_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_6_groups_1 = const()[name = string("logits_chunk_6_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_6 = conv(dilations = logits_chunk_6_dilations_1, groups = logits_chunk_6_groups_1, pad = logits_chunk_6_pad_1, pad_type = logits_chunk_6_pad_type_1, strides = logits_chunk_6_strides_1, weight = logits_chunk_6_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_6")];
+            int32 argmax_chunk_6_axis_0 = const()[name = string("argmax_chunk_6_axis_0"), val = int32(1)];
+            bool argmax_chunk_6_keep_dims_0 = const()[name = string("argmax_chunk_6_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_6_output_dtype_0 = const()[name = string("argmax_chunk_6_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> argmax_chunk_6 = reduce_argmax(axis = argmax_chunk_6_axis_0, keep_dims = argmax_chunk_6_keep_dims_0, output_dtype = argmax_chunk_6_output_dtype_0, x = logits_chunk_6)[name = string("argmax_chunk_6")];
+            tensor<int32, [1]> max_chunk_6_axes_0 = const()[name = string("max_chunk_6_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_6_keep_dims_0 = const()[name = string("max_chunk_6_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> max_chunk_6 = reduce_max(axes = max_chunk_6_axes_0, keep_dims = max_chunk_6_keep_dims_0, x = logits_chunk_6)[name = string("max_chunk_6")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_7_weight_1 = const()[name = string("logits_chunk_7_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58724928)))];
+            tensor<int32, [2]> logits_chunk_7_strides_1 = const()[name = string("logits_chunk_7_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_7_pad_type_1 = const()[name = string("logits_chunk_7_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_7_pad_1 = const()[name = string("logits_chunk_7_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_7_dilations_1 = const()[name = string("logits_chunk_7_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_7_groups_1 = const()[name = string("logits_chunk_7_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_7 = conv(dilations = logits_chunk_7_dilations_1, groups = logits_chunk_7_groups_1, pad = logits_chunk_7_pad_1, pad_type = logits_chunk_7_pad_type_1, strides = logits_chunk_7_strides_1, weight = logits_chunk_7_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_7")];
+            int32 argmax_chunk_7_axis_0 = const()[name = string("argmax_chunk_7_axis_0"), val = int32(1)];
+            bool argmax_chunk_7_keep_dims_0 = const()[name = string("argmax_chunk_7_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_7_output_dtype_0 = const()[name = string("argmax_chunk_7_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> argmax_chunk_7 = reduce_argmax(axis = argmax_chunk_7_axis_0, keep_dims = argmax_chunk_7_keep_dims_0, output_dtype = argmax_chunk_7_output_dtype_0, x = logits_chunk_7)[name = string("argmax_chunk_7")];
+            tensor<int32, [1]> max_chunk_7_axes_0 = const()[name = string("max_chunk_7_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_7_keep_dims_0 = const()[name = string("max_chunk_7_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> max_chunk_7 = reduce_max(axes = max_chunk_7_axes_0, keep_dims = max_chunk_7_keep_dims_0, x = logits_chunk_7)[name = string("max_chunk_7")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_8_weight_1 = const()[name = string("logits_chunk_8_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67113600)))];
+            tensor<int32, [2]> logits_chunk_8_strides_1 = const()[name = string("logits_chunk_8_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_8_pad_type_1 = const()[name = string("logits_chunk_8_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_8_pad_1 = const()[name = string("logits_chunk_8_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_8_dilations_1 = const()[name = string("logits_chunk_8_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_8_groups_1 = const()[name = string("logits_chunk_8_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_8 = conv(dilations = logits_chunk_8_dilations_1, groups = logits_chunk_8_groups_1, pad = logits_chunk_8_pad_1, pad_type = logits_chunk_8_pad_type_1, strides = logits_chunk_8_strides_1, weight = logits_chunk_8_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_8")];
+            int32 argmax_chunk_8_axis_0 = const()[name = string("argmax_chunk_8_axis_0"), val = int32(1)];
+            bool argmax_chunk_8_keep_dims_0 = const()[name = string("argmax_chunk_8_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_8_output_dtype_0 = const()[name = string("argmax_chunk_8_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> argmax_chunk_8 = reduce_argmax(axis = argmax_chunk_8_axis_0, keep_dims = argmax_chunk_8_keep_dims_0, output_dtype = argmax_chunk_8_output_dtype_0, x = logits_chunk_8)[name = string("argmax_chunk_8")];
+            tensor<int32, [1]> max_chunk_8_axes_0 = const()[name = string("max_chunk_8_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_8_keep_dims_0 = const()[name = string("max_chunk_8_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> max_chunk_8 = reduce_max(axes = max_chunk_8_axes_0, keep_dims = max_chunk_8_keep_dims_0, x = logits_chunk_8)[name = string("max_chunk_8")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_9_weight_1 = const()[name = string("logits_chunk_9_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75502272)))];
+            tensor<int32, [2]> logits_chunk_9_strides_1 = const()[name = string("logits_chunk_9_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_9_pad_type_1 = const()[name = string("logits_chunk_9_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_9_pad_1 = const()[name = string("logits_chunk_9_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_9_dilations_1 = const()[name = string("logits_chunk_9_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_9_groups_1 = const()[name = string("logits_chunk_9_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_9 = conv(dilations = logits_chunk_9_dilations_1, groups = logits_chunk_9_groups_1, pad = logits_chunk_9_pad_1, pad_type = logits_chunk_9_pad_type_1, strides = logits_chunk_9_strides_1, weight = logits_chunk_9_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_9")];
+            int32 argmax_chunk_9_axis_0 = const()[name = string("argmax_chunk_9_axis_0"), val = int32(1)];
+            bool argmax_chunk_9_keep_dims_0 = const()[name = string("argmax_chunk_9_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_9_output_dtype_0 = const()[name = string("argmax_chunk_9_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> argmax_chunk_9 = reduce_argmax(axis = argmax_chunk_9_axis_0, keep_dims = argmax_chunk_9_keep_dims_0, output_dtype = argmax_chunk_9_output_dtype_0, x = logits_chunk_9)[name = string("argmax_chunk_9")];
+            tensor<int32, [1]> max_chunk_9_axes_0 = const()[name = string("max_chunk_9_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_9_keep_dims_0 = const()[name = string("max_chunk_9_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> max_chunk_9 = reduce_max(axes = max_chunk_9_axes_0, keep_dims = max_chunk_9_keep_dims_0, x = logits_chunk_9)[name = string("max_chunk_9")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_10_weight_1 = const()[name = string("logits_chunk_10_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83890944)))];
+            tensor<int32, [2]> logits_chunk_10_strides_1 = const()[name = string("logits_chunk_10_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_10_pad_type_1 = const()[name = string("logits_chunk_10_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_10_pad_1 = const()[name = string("logits_chunk_10_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_10_dilations_1 = const()[name = string("logits_chunk_10_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_10_groups_1 = const()[name = string("logits_chunk_10_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_10 = conv(dilations = logits_chunk_10_dilations_1, groups = logits_chunk_10_groups_1, pad = logits_chunk_10_pad_1, pad_type = logits_chunk_10_pad_type_1, strides = logits_chunk_10_strides_1, weight = logits_chunk_10_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_10")];
+            int32 argmax_chunk_10_axis_0 = const()[name = string("argmax_chunk_10_axis_0"), val = int32(1)];
+            bool argmax_chunk_10_keep_dims_0 = const()[name = string("argmax_chunk_10_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_10_output_dtype_0 = const()[name = string("argmax_chunk_10_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> argmax_chunk_10 = reduce_argmax(axis = argmax_chunk_10_axis_0, keep_dims = argmax_chunk_10_keep_dims_0, output_dtype = argmax_chunk_10_output_dtype_0, x = logits_chunk_10)[name = string("argmax_chunk_10")];
+            tensor<int32, [1]> max_chunk_10_axes_0 = const()[name = string("max_chunk_10_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_10_keep_dims_0 = const()[name = string("max_chunk_10_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> max_chunk_10 = reduce_max(axes = max_chunk_10_axes_0, keep_dims = max_chunk_10_keep_dims_0, x = logits_chunk_10)[name = string("max_chunk_10")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_11_weight_1 = const()[name = string("logits_chunk_11_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92279616)))];
+            tensor<int32, [2]> logits_chunk_11_strides_1 = const()[name = string("logits_chunk_11_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_11_pad_type_1 = const()[name = string("logits_chunk_11_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_11_pad_1 = const()[name = string("logits_chunk_11_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_11_dilations_1 = const()[name = string("logits_chunk_11_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_11_groups_1 = const()[name = string("logits_chunk_11_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_11 = conv(dilations = logits_chunk_11_dilations_1, groups = logits_chunk_11_groups_1, pad = logits_chunk_11_pad_1, pad_type = logits_chunk_11_pad_type_1, strides = logits_chunk_11_strides_1, weight = logits_chunk_11_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_11")];
+            int32 argmax_chunk_11_axis_0 = const()[name = string("argmax_chunk_11_axis_0"), val = int32(1)];
+            bool argmax_chunk_11_keep_dims_0 = const()[name = string("argmax_chunk_11_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_11_output_dtype_0 = const()[name = string("argmax_chunk_11_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> argmax_chunk_11 = reduce_argmax(axis = argmax_chunk_11_axis_0, keep_dims = argmax_chunk_11_keep_dims_0, output_dtype = argmax_chunk_11_output_dtype_0, x = logits_chunk_11)[name = string("argmax_chunk_11")];
+            tensor<int32, [1]> max_chunk_11_axes_0 = const()[name = string("max_chunk_11_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_11_keep_dims_0 = const()[name = string("max_chunk_11_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> max_chunk_11 = reduce_max(axes = max_chunk_11_axes_0, keep_dims = max_chunk_11_keep_dims_0, x = logits_chunk_11)[name = string("max_chunk_11")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_12_weight_1 = const()[name = string("logits_chunk_12_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100668288)))];
+            tensor<int32, [2]> logits_chunk_12_strides_1 = const()[name = string("logits_chunk_12_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_12_pad_type_1 = const()[name = string("logits_chunk_12_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_12_pad_1 = const()[name = string("logits_chunk_12_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_12_dilations_1 = const()[name = string("logits_chunk_12_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_12_groups_1 = const()[name = string("logits_chunk_12_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_12 = conv(dilations = logits_chunk_12_dilations_1, groups = logits_chunk_12_groups_1, pad = logits_chunk_12_pad_1, pad_type = logits_chunk_12_pad_type_1, strides = logits_chunk_12_strides_1, weight = logits_chunk_12_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_12")];
+            int32 argmax_chunk_12_axis_0 = const()[name = string("argmax_chunk_12_axis_0"), val = int32(1)];
+            bool argmax_chunk_12_keep_dims_0 = const()[name = string("argmax_chunk_12_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_12_output_dtype_0 = const()[name = string("argmax_chunk_12_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> argmax_chunk_12 = reduce_argmax(axis = argmax_chunk_12_axis_0, keep_dims = argmax_chunk_12_keep_dims_0, output_dtype = argmax_chunk_12_output_dtype_0, x = logits_chunk_12)[name = string("argmax_chunk_12")];
+            tensor<int32, [1]> max_chunk_12_axes_0 = const()[name = string("max_chunk_12_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_12_keep_dims_0 = const()[name = string("max_chunk_12_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> max_chunk_12 = reduce_max(axes = max_chunk_12_axes_0, keep_dims = max_chunk_12_keep_dims_0, x = logits_chunk_12)[name = string("max_chunk_12")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_13_weight_1 = const()[name = string("logits_chunk_13_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109056960)))];
+            tensor<int32, [2]> logits_chunk_13_strides_1 = const()[name = string("logits_chunk_13_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_13_pad_type_1 = const()[name = string("logits_chunk_13_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_13_pad_1 = const()[name = string("logits_chunk_13_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_13_dilations_1 = const()[name = string("logits_chunk_13_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_13_groups_1 = const()[name = string("logits_chunk_13_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_13 = conv(dilations = logits_chunk_13_dilations_1, groups = logits_chunk_13_groups_1, pad = logits_chunk_13_pad_1, pad_type = logits_chunk_13_pad_type_1, strides = logits_chunk_13_strides_1, weight = logits_chunk_13_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_13")];
+            int32 argmax_chunk_13_axis_0 = const()[name = string("argmax_chunk_13_axis_0"), val = int32(1)];
+            bool argmax_chunk_13_keep_dims_0 = const()[name = string("argmax_chunk_13_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_13_output_dtype_0 = const()[name = string("argmax_chunk_13_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> argmax_chunk_13 = reduce_argmax(axis = argmax_chunk_13_axis_0, keep_dims = argmax_chunk_13_keep_dims_0, output_dtype = argmax_chunk_13_output_dtype_0, x = logits_chunk_13)[name = string("argmax_chunk_13")];
+            tensor<int32, [1]> max_chunk_13_axes_0 = const()[name = string("max_chunk_13_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_13_keep_dims_0 = const()[name = string("max_chunk_13_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> max_chunk_13 = reduce_max(axes = max_chunk_13_axes_0, keep_dims = max_chunk_13_keep_dims_0, x = logits_chunk_13)[name = string("max_chunk_13")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_14_weight_1 = const()[name = string("logits_chunk_14_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117445632)))];
+            tensor<int32, [2]> logits_chunk_14_strides_1 = const()[name = string("logits_chunk_14_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_14_pad_type_1 = const()[name = string("logits_chunk_14_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_14_pad_1 = const()[name = string("logits_chunk_14_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_14_dilations_1 = const()[name = string("logits_chunk_14_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_14_groups_1 = const()[name = string("logits_chunk_14_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_14 = conv(dilations = logits_chunk_14_dilations_1, groups = logits_chunk_14_groups_1, pad = logits_chunk_14_pad_1, pad_type = logits_chunk_14_pad_type_1, strides = logits_chunk_14_strides_1, weight = logits_chunk_14_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_14")];
+            int32 argmax_chunk_14_axis_0 = const()[name = string("argmax_chunk_14_axis_0"), val = int32(1)];
+            bool argmax_chunk_14_keep_dims_0 = const()[name = string("argmax_chunk_14_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_14_output_dtype_0 = const()[name = string("argmax_chunk_14_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> argmax_chunk_14 = reduce_argmax(axis = argmax_chunk_14_axis_0, keep_dims = argmax_chunk_14_keep_dims_0, output_dtype = argmax_chunk_14_output_dtype_0, x = logits_chunk_14)[name = string("argmax_chunk_14")];
+            tensor<int32, [1]> max_chunk_14_axes_0 = const()[name = string("max_chunk_14_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_14_keep_dims_0 = const()[name = string("max_chunk_14_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> max_chunk_14 = reduce_max(axes = max_chunk_14_axes_0, keep_dims = max_chunk_14_keep_dims_0, x = logits_chunk_14)[name = string("max_chunk_14")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_15_weight_1 = const()[name = string("logits_chunk_15_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125834304)))];
+            tensor<int32, [2]> logits_chunk_15_strides_1 = const()[name = string("logits_chunk_15_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_15_pad_type_1 = const()[name = string("logits_chunk_15_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_15_pad_1 = const()[name = string("logits_chunk_15_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_15_dilations_1 = const()[name = string("logits_chunk_15_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_15_groups_1 = const()[name = string("logits_chunk_15_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_15 = conv(dilations = logits_chunk_15_dilations_1, groups = logits_chunk_15_groups_1, pad = logits_chunk_15_pad_1, pad_type = logits_chunk_15_pad_type_1, strides = logits_chunk_15_strides_1, weight = logits_chunk_15_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_15")];
+            int32 argmax_chunk_15_axis_0 = const()[name = string("argmax_chunk_15_axis_0"), val = int32(1)];
+            bool argmax_chunk_15_keep_dims_0 = const()[name = string("argmax_chunk_15_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_15_output_dtype_0 = const()[name = string("argmax_chunk_15_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> argmax_chunk_15 = reduce_argmax(axis = argmax_chunk_15_axis_0, keep_dims = argmax_chunk_15_keep_dims_0, output_dtype = argmax_chunk_15_output_dtype_0, x = logits_chunk_15)[name = string("argmax_chunk_15")];
+            tensor<int32, [1]> max_chunk_15_axes_0 = const()[name = string("max_chunk_15_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_15_keep_dims_0 = const()[name = string("max_chunk_15_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> max_chunk_15 = reduce_max(axes = max_chunk_15_axes_0, keep_dims = max_chunk_15_keep_dims_0, x = logits_chunk_15)[name = string("max_chunk_15")];
+            int32 logits_axis_0 = const()[name = string("logits_axis_0"), val = int32(1)];
+            bool logits_interleave_0 = const()[name = string("logits_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 32768, 1, 48]> logits = concat(axis = logits_axis_0, interleave = logits_interleave_0, values = (logits_chunk_0, logits_chunk_1, logits_chunk_2, logits_chunk_3, logits_chunk_4, logits_chunk_5, logits_chunk_6, logits_chunk_7, logits_chunk_8, logits_chunk_9, logits_chunk_10, logits_chunk_11, logits_chunk_12, logits_chunk_13, logits_chunk_14, logits_chunk_15))[name = string("logits")];
+            int32 values_axis_0 = const()[name = string("values_axis_0"), val = int32(1)];
+            bool values_interleave_0 = const()[name = string("values_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 48]> values = concat(axis = values_axis_0, interleave = values_interleave_0, values = (max_chunk_0, max_chunk_1, max_chunk_2, max_chunk_3, max_chunk_4, max_chunk_5, max_chunk_6, max_chunk_7, max_chunk_8, max_chunk_9, max_chunk_10, max_chunk_11, max_chunk_12, max_chunk_13, max_chunk_14, max_chunk_15))[name = string("values")];
+            tensor<int32, [1]> max_value_axes_0 = const()[name = string("max_value_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_value_keep_dims_0 = const()[name = string("max_value_keep_dims_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 48]> max_value = reduce_max(axes = max_value_axes_0, keep_dims = max_value_keep_dims_0, x = values)[name = string("max_value")];
+            int32 max_value_index_axis_0 = const()[name = string("max_value_index_axis_0"), val = int32(1)];
+            bool max_value_index_keep_dims_0 = const()[name = string("max_value_index_keep_dims_0"), val = bool(true)];
+            string max_value_index_output_dtype_0 = const()[name = string("max_value_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 1, 1, 48]> max_value_index = reduce_argmax(axis = max_value_index_axis_0, keep_dims = max_value_index_keep_dims_0, output_dtype = max_value_index_output_dtype_0, x = values)[name = string("max_value_index")];
+            string indices_chunk_0_int32_dtype_1 = const()[name = string("indices_chunk_0_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_1_int32_dtype_1 = const()[name = string("indices_chunk_1_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_2_int32_dtype_1 = const()[name = string("indices_chunk_2_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_3_int32_dtype_1 = const()[name = string("indices_chunk_3_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_4_int32_dtype_1 = const()[name = string("indices_chunk_4_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_5_int32_dtype_1 = const()[name = string("indices_chunk_5_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_6_int32_dtype_1 = const()[name = string("indices_chunk_6_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_7_int32_dtype_1 = const()[name = string("indices_chunk_7_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_8_int32_dtype_1 = const()[name = string("indices_chunk_8_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_9_int32_dtype_1 = const()[name = string("indices_chunk_9_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_10_int32_dtype_1 = const()[name = string("indices_chunk_10_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_11_int32_dtype_1 = const()[name = string("indices_chunk_11_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_12_int32_dtype_1 = const()[name = string("indices_chunk_12_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_13_int32_dtype_1 = const()[name = string("indices_chunk_13_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_14_int32_dtype_1 = const()[name = string("indices_chunk_14_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_15_int32_dtype_1 = const()[name = string("indices_chunk_15_int32_dtype_1"), val = string("int32")];
+            int32 indices_axis_1 = const()[name = string("indices_axis_1"), val = int32(1)];
+            bool indices_interleave_1 = const()[name = string("indices_interleave_1"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_15_int32 = cast(dtype = indices_chunk_15_int32_dtype_1, x = argmax_chunk_15)[name = string("cast_80")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_14_int32 = cast(dtype = indices_chunk_14_int32_dtype_1, x = argmax_chunk_14)[name = string("cast_81")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_13_int32 = cast(dtype = indices_chunk_13_int32_dtype_1, x = argmax_chunk_13)[name = string("cast_82")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_12_int32 = cast(dtype = indices_chunk_12_int32_dtype_1, x = argmax_chunk_12)[name = string("cast_83")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_11_int32 = cast(dtype = indices_chunk_11_int32_dtype_1, x = argmax_chunk_11)[name = string("cast_84")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_10_int32 = cast(dtype = indices_chunk_10_int32_dtype_1, x = argmax_chunk_10)[name = string("cast_85")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_9_int32 = cast(dtype = indices_chunk_9_int32_dtype_1, x = argmax_chunk_9)[name = string("cast_86")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_8_int32 = cast(dtype = indices_chunk_8_int32_dtype_1, x = argmax_chunk_8)[name = string("cast_87")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_7_int32 = cast(dtype = indices_chunk_7_int32_dtype_1, x = argmax_chunk_7)[name = string("cast_88")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_6_int32 = cast(dtype = indices_chunk_6_int32_dtype_1, x = argmax_chunk_6)[name = string("cast_89")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_5_int32 = cast(dtype = indices_chunk_5_int32_dtype_1, x = argmax_chunk_5)[name = string("cast_90")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_4_int32 = cast(dtype = indices_chunk_4_int32_dtype_1, x = argmax_chunk_4)[name = string("cast_91")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_3_int32 = cast(dtype = indices_chunk_3_int32_dtype_1, x = argmax_chunk_3)[name = string("cast_92")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_2_int32 = cast(dtype = indices_chunk_2_int32_dtype_1, x = argmax_chunk_2)[name = string("cast_93")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_1_int32 = cast(dtype = indices_chunk_1_int32_dtype_1, x = argmax_chunk_1)[name = string("cast_94")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_0_int32 = cast(dtype = indices_chunk_0_int32_dtype_1, x = argmax_chunk_0)[name = string("cast_95")];
+            tensor<int32, [1, 16, 1, 48]> indices = concat(axis = indices_axis_1, interleave = indices_interleave_1, values = (indices_chunk_0_int32, indices_chunk_1_int32, indices_chunk_2_int32, indices_chunk_3_int32, indices_chunk_4_int32, indices_chunk_5_int32, indices_chunk_6_int32, indices_chunk_7_int32, indices_chunk_8_int32, indices_chunk_9_int32, indices_chunk_10_int32, indices_chunk_11_int32, indices_chunk_12_int32, indices_chunk_13_int32, indices_chunk_14_int32, indices_chunk_15_int32))[name = string("indices")];
+            int32 argmax_chunks_axis_1 = const()[name = string("argmax_chunks_axis_1"), val = int32(1)];
+            bool argmax_chunks_validate_indices_1 = const()[name = string("argmax_chunks_validate_indices_1"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 48]> argmax_chunks = gather_along_axis(axis = argmax_chunks_axis_1, indices = max_value_index, validate_indices = argmax_chunks_validate_indices_1, x = indices)[name = string("argmax_chunks")];
+            int32 mul_1_x_0 = const()[name = string("mul_1_x_0"), val = int32(2048)];
+            tensor<int32, [1, 1, 1, 48]> mul_1 = mul(x = mul_1_x_0, y = max_value_index)[name = string("mul_1")];
+            tensor<int32, [1, 1, 1, 48]> argmax = add(x = argmax_chunks, y = mul_1)[name = string("argmax")];
+        } -> (logits, argmax, max_value);
+    func lm_head_length_64<ios18>(tensor<fp16, [1, 2048, 1, 64]> hidden_states) {
+            tensor<fp16, [1, 2048, 1, 64]> final_norm_rmsnorm_abs = abs(x = hidden_states)[name = string("final_norm_rmsnorm_abs")];
+            tensor<int32, [1]> final_norm_rmsnorm_maxval_axes_1 = const()[name = string("final_norm_rmsnorm_maxval_axes_1"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_maxval_keep_dims_1 = const()[name = string("final_norm_rmsnorm_maxval_keep_dims_1"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> final_norm_rmsnorm_maxval = reduce_max(axes = final_norm_rmsnorm_maxval_axes_1, keep_dims = final_norm_rmsnorm_maxval_keep_dims_1, x = final_norm_rmsnorm_abs)[name = string("final_norm_rmsnorm_maxval")];
+            fp16 final_norm_rmsnorm_maxval_clipped_alpha_1 = const()[name = string("final_norm_rmsnorm_maxval_clipped_alpha_1"), val = fp16(0x1p-14)];
+            fp16 final_norm_rmsnorm_maxval_clipped_beta_1 = const()[name = string("final_norm_rmsnorm_maxval_clipped_beta_1"), val = fp16(inf)];
+            tensor<fp16, [1, 1, 1, 64]> final_norm_rmsnorm_maxval_clipped = clip(alpha = final_norm_rmsnorm_maxval_clipped_alpha_1, beta = final_norm_rmsnorm_maxval_clipped_beta_1, x = final_norm_rmsnorm_maxval)[name = string("final_norm_rmsnorm_maxval_clipped")];
+            tensor<fp16, [1, 2048, 1, 64]> final_norm_rmsnorm_scaled = real_div(x = hidden_states, y = final_norm_rmsnorm_maxval_clipped)[name = string("final_norm_rmsnorm_scaled")];
+            tensor<int32, [1]> final_norm_rmsnorm_squared_sum_axes_1 = const()[name = string("final_norm_rmsnorm_squared_sum_axes_1"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_squared_sum_keep_dims_1 = const()[name = string("final_norm_rmsnorm_squared_sum_keep_dims_1"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> final_norm_rmsnorm_squared_sum = reduce_sum_square(axes = final_norm_rmsnorm_squared_sum_axes_1, keep_dims = final_norm_rmsnorm_squared_sum_keep_dims_1, x = final_norm_rmsnorm_scaled)[name = string("final_norm_rmsnorm_squared_sum")];
+            fp16 final_norm_rmsnorm_rsqrt_epsilon_1 = const()[name = string("final_norm_rmsnorm_rsqrt_epsilon_1"), val = fp16(0x1p-14)];
+            tensor<fp16, [1, 1, 1, 64]> final_norm_rmsnorm_rsqrt = rsqrt(epsilon = final_norm_rmsnorm_rsqrt_epsilon_1, x = final_norm_rmsnorm_squared_sum)[name = string("final_norm_rmsnorm_rsqrt")];
+            fp16 final_norm_rmsnorm_dim_scaled_y_1 = const()[name = string("final_norm_rmsnorm_dim_scaled_y_1"), val = fp16(0x1.6ap+5)];
+            tensor<fp16, [1, 2048, 1, 64]> final_norm_rmsnorm_dim_scaled = mul(x = final_norm_rmsnorm_scaled, y = final_norm_rmsnorm_dim_scaled_y_1)[name = string("final_norm_rmsnorm_dim_scaled")];
+            tensor<fp16, [1, 2048, 1, 64]> final_norm_rmsnorm_normalized = mul(x = final_norm_rmsnorm_dim_scaled, y = final_norm_rmsnorm_rsqrt)[name = string("final_norm_rmsnorm_normalized")];
+            tensor<fp16, [2048, 1, 1]> final_norm_rmsnorm_y_1 = const()[name = string("final_norm_rmsnorm_y_1"), val = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 2048, 1, 64]> final_norm_rmsnorm = mul(x = final_norm_rmsnorm_normalized, y = final_norm_rmsnorm_y_1)[name = string("final_norm_rmsnorm")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_0_weight_1 = const()[name = string("logits_chunk_0_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4224)))];
+            tensor<int32, [2]> logits_chunk_0_strides_1 = const()[name = string("logits_chunk_0_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_0_pad_type_1 = const()[name = string("logits_chunk_0_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_0_pad_1 = const()[name = string("logits_chunk_0_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_0_dilations_1 = const()[name = string("logits_chunk_0_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_0_groups_1 = const()[name = string("logits_chunk_0_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_0 = conv(dilations = logits_chunk_0_dilations_1, groups = logits_chunk_0_groups_1, pad = logits_chunk_0_pad_1, pad_type = logits_chunk_0_pad_type_1, strides = logits_chunk_0_strides_1, weight = logits_chunk_0_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_0")];
+            int32 argmax_chunk_0_axis_0 = const()[name = string("argmax_chunk_0_axis_0"), val = int32(1)];
+            bool argmax_chunk_0_keep_dims_0 = const()[name = string("argmax_chunk_0_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_0_output_dtype_0 = const()[name = string("argmax_chunk_0_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> argmax_chunk_0 = reduce_argmax(axis = argmax_chunk_0_axis_0, keep_dims = argmax_chunk_0_keep_dims_0, output_dtype = argmax_chunk_0_output_dtype_0, x = logits_chunk_0)[name = string("argmax_chunk_0")];
+            tensor<int32, [1]> max_chunk_0_axes_0 = const()[name = string("max_chunk_0_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_0_keep_dims_0 = const()[name = string("max_chunk_0_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> max_chunk_0 = reduce_max(axes = max_chunk_0_axes_0, keep_dims = max_chunk_0_keep_dims_0, x = logits_chunk_0)[name = string("max_chunk_0")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_1_weight_1 = const()[name = string("logits_chunk_1_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8392896)))];
+            tensor<int32, [2]> logits_chunk_1_strides_1 = const()[name = string("logits_chunk_1_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_1_pad_type_1 = const()[name = string("logits_chunk_1_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_1_pad_1 = const()[name = string("logits_chunk_1_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_1_dilations_1 = const()[name = string("logits_chunk_1_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_1_groups_1 = const()[name = string("logits_chunk_1_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_1 = conv(dilations = logits_chunk_1_dilations_1, groups = logits_chunk_1_groups_1, pad = logits_chunk_1_pad_1, pad_type = logits_chunk_1_pad_type_1, strides = logits_chunk_1_strides_1, weight = logits_chunk_1_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_1")];
+            int32 argmax_chunk_1_axis_0 = const()[name = string("argmax_chunk_1_axis_0"), val = int32(1)];
+            bool argmax_chunk_1_keep_dims_0 = const()[name = string("argmax_chunk_1_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_1_output_dtype_0 = const()[name = string("argmax_chunk_1_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> argmax_chunk_1 = reduce_argmax(axis = argmax_chunk_1_axis_0, keep_dims = argmax_chunk_1_keep_dims_0, output_dtype = argmax_chunk_1_output_dtype_0, x = logits_chunk_1)[name = string("argmax_chunk_1")];
+            tensor<int32, [1]> max_chunk_1_axes_0 = const()[name = string("max_chunk_1_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_1_keep_dims_0 = const()[name = string("max_chunk_1_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> max_chunk_1 = reduce_max(axes = max_chunk_1_axes_0, keep_dims = max_chunk_1_keep_dims_0, x = logits_chunk_1)[name = string("max_chunk_1")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_2_weight_1 = const()[name = string("logits_chunk_2_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16781568)))];
+            tensor<int32, [2]> logits_chunk_2_strides_1 = const()[name = string("logits_chunk_2_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_2_pad_type_1 = const()[name = string("logits_chunk_2_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_2_pad_1 = const()[name = string("logits_chunk_2_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_2_dilations_1 = const()[name = string("logits_chunk_2_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_2_groups_1 = const()[name = string("logits_chunk_2_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_2 = conv(dilations = logits_chunk_2_dilations_1, groups = logits_chunk_2_groups_1, pad = logits_chunk_2_pad_1, pad_type = logits_chunk_2_pad_type_1, strides = logits_chunk_2_strides_1, weight = logits_chunk_2_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_2")];
+            int32 argmax_chunk_2_axis_0 = const()[name = string("argmax_chunk_2_axis_0"), val = int32(1)];
+            bool argmax_chunk_2_keep_dims_0 = const()[name = string("argmax_chunk_2_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_2_output_dtype_0 = const()[name = string("argmax_chunk_2_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> argmax_chunk_2 = reduce_argmax(axis = argmax_chunk_2_axis_0, keep_dims = argmax_chunk_2_keep_dims_0, output_dtype = argmax_chunk_2_output_dtype_0, x = logits_chunk_2)[name = string("argmax_chunk_2")];
+            tensor<int32, [1]> max_chunk_2_axes_0 = const()[name = string("max_chunk_2_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_2_keep_dims_0 = const()[name = string("max_chunk_2_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> max_chunk_2 = reduce_max(axes = max_chunk_2_axes_0, keep_dims = max_chunk_2_keep_dims_0, x = logits_chunk_2)[name = string("max_chunk_2")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_3_weight_1 = const()[name = string("logits_chunk_3_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25170240)))];
+            tensor<int32, [2]> logits_chunk_3_strides_1 = const()[name = string("logits_chunk_3_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_3_pad_type_1 = const()[name = string("logits_chunk_3_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_3_pad_1 = const()[name = string("logits_chunk_3_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_3_dilations_1 = const()[name = string("logits_chunk_3_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_3_groups_1 = const()[name = string("logits_chunk_3_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_3 = conv(dilations = logits_chunk_3_dilations_1, groups = logits_chunk_3_groups_1, pad = logits_chunk_3_pad_1, pad_type = logits_chunk_3_pad_type_1, strides = logits_chunk_3_strides_1, weight = logits_chunk_3_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_3")];
+            int32 argmax_chunk_3_axis_0 = const()[name = string("argmax_chunk_3_axis_0"), val = int32(1)];
+            bool argmax_chunk_3_keep_dims_0 = const()[name = string("argmax_chunk_3_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_3_output_dtype_0 = const()[name = string("argmax_chunk_3_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> argmax_chunk_3 = reduce_argmax(axis = argmax_chunk_3_axis_0, keep_dims = argmax_chunk_3_keep_dims_0, output_dtype = argmax_chunk_3_output_dtype_0, x = logits_chunk_3)[name = string("argmax_chunk_3")];
+            tensor<int32, [1]> max_chunk_3_axes_0 = const()[name = string("max_chunk_3_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_3_keep_dims_0 = const()[name = string("max_chunk_3_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> max_chunk_3 = reduce_max(axes = max_chunk_3_axes_0, keep_dims = max_chunk_3_keep_dims_0, x = logits_chunk_3)[name = string("max_chunk_3")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_4_weight_1 = const()[name = string("logits_chunk_4_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33558912)))];
+            tensor<int32, [2]> logits_chunk_4_strides_1 = const()[name = string("logits_chunk_4_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_4_pad_type_1 = const()[name = string("logits_chunk_4_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_4_pad_1 = const()[name = string("logits_chunk_4_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_4_dilations_1 = const()[name = string("logits_chunk_4_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_4_groups_1 = const()[name = string("logits_chunk_4_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_4 = conv(dilations = logits_chunk_4_dilations_1, groups = logits_chunk_4_groups_1, pad = logits_chunk_4_pad_1, pad_type = logits_chunk_4_pad_type_1, strides = logits_chunk_4_strides_1, weight = logits_chunk_4_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_4")];
+            int32 argmax_chunk_4_axis_0 = const()[name = string("argmax_chunk_4_axis_0"), val = int32(1)];
+            bool argmax_chunk_4_keep_dims_0 = const()[name = string("argmax_chunk_4_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_4_output_dtype_0 = const()[name = string("argmax_chunk_4_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> argmax_chunk_4 = reduce_argmax(axis = argmax_chunk_4_axis_0, keep_dims = argmax_chunk_4_keep_dims_0, output_dtype = argmax_chunk_4_output_dtype_0, x = logits_chunk_4)[name = string("argmax_chunk_4")];
+            tensor<int32, [1]> max_chunk_4_axes_0 = const()[name = string("max_chunk_4_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_4_keep_dims_0 = const()[name = string("max_chunk_4_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> max_chunk_4 = reduce_max(axes = max_chunk_4_axes_0, keep_dims = max_chunk_4_keep_dims_0, x = logits_chunk_4)[name = string("max_chunk_4")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_5_weight_1 = const()[name = string("logits_chunk_5_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41947584)))];
+            tensor<int32, [2]> logits_chunk_5_strides_1 = const()[name = string("logits_chunk_5_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_5_pad_type_1 = const()[name = string("logits_chunk_5_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_5_pad_1 = const()[name = string("logits_chunk_5_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_5_dilations_1 = const()[name = string("logits_chunk_5_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_5_groups_1 = const()[name = string("logits_chunk_5_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_5 = conv(dilations = logits_chunk_5_dilations_1, groups = logits_chunk_5_groups_1, pad = logits_chunk_5_pad_1, pad_type = logits_chunk_5_pad_type_1, strides = logits_chunk_5_strides_1, weight = logits_chunk_5_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_5")];
+            int32 argmax_chunk_5_axis_0 = const()[name = string("argmax_chunk_5_axis_0"), val = int32(1)];
+            bool argmax_chunk_5_keep_dims_0 = const()[name = string("argmax_chunk_5_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_5_output_dtype_0 = const()[name = string("argmax_chunk_5_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> argmax_chunk_5 = reduce_argmax(axis = argmax_chunk_5_axis_0, keep_dims = argmax_chunk_5_keep_dims_0, output_dtype = argmax_chunk_5_output_dtype_0, x = logits_chunk_5)[name = string("argmax_chunk_5")];
+            tensor<int32, [1]> max_chunk_5_axes_0 = const()[name = string("max_chunk_5_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_5_keep_dims_0 = const()[name = string("max_chunk_5_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> max_chunk_5 = reduce_max(axes = max_chunk_5_axes_0, keep_dims = max_chunk_5_keep_dims_0, x = logits_chunk_5)[name = string("max_chunk_5")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_6_weight_1 = const()[name = string("logits_chunk_6_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50336256)))];
+            tensor<int32, [2]> logits_chunk_6_strides_1 = const()[name = string("logits_chunk_6_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_6_pad_type_1 = const()[name = string("logits_chunk_6_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_6_pad_1 = const()[name = string("logits_chunk_6_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_6_dilations_1 = const()[name = string("logits_chunk_6_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_6_groups_1 = const()[name = string("logits_chunk_6_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_6 = conv(dilations = logits_chunk_6_dilations_1, groups = logits_chunk_6_groups_1, pad = logits_chunk_6_pad_1, pad_type = logits_chunk_6_pad_type_1, strides = logits_chunk_6_strides_1, weight = logits_chunk_6_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_6")];
+            int32 argmax_chunk_6_axis_0 = const()[name = string("argmax_chunk_6_axis_0"), val = int32(1)];
+            bool argmax_chunk_6_keep_dims_0 = const()[name = string("argmax_chunk_6_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_6_output_dtype_0 = const()[name = string("argmax_chunk_6_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> argmax_chunk_6 = reduce_argmax(axis = argmax_chunk_6_axis_0, keep_dims = argmax_chunk_6_keep_dims_0, output_dtype = argmax_chunk_6_output_dtype_0, x = logits_chunk_6)[name = string("argmax_chunk_6")];
+            tensor<int32, [1]> max_chunk_6_axes_0 = const()[name = string("max_chunk_6_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_6_keep_dims_0 = const()[name = string("max_chunk_6_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> max_chunk_6 = reduce_max(axes = max_chunk_6_axes_0, keep_dims = max_chunk_6_keep_dims_0, x = logits_chunk_6)[name = string("max_chunk_6")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_7_weight_1 = const()[name = string("logits_chunk_7_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58724928)))];
+            tensor<int32, [2]> logits_chunk_7_strides_1 = const()[name = string("logits_chunk_7_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_7_pad_type_1 = const()[name = string("logits_chunk_7_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_7_pad_1 = const()[name = string("logits_chunk_7_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_7_dilations_1 = const()[name = string("logits_chunk_7_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_7_groups_1 = const()[name = string("logits_chunk_7_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_7 = conv(dilations = logits_chunk_7_dilations_1, groups = logits_chunk_7_groups_1, pad = logits_chunk_7_pad_1, pad_type = logits_chunk_7_pad_type_1, strides = logits_chunk_7_strides_1, weight = logits_chunk_7_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_7")];
+            int32 argmax_chunk_7_axis_0 = const()[name = string("argmax_chunk_7_axis_0"), val = int32(1)];
+            bool argmax_chunk_7_keep_dims_0 = const()[name = string("argmax_chunk_7_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_7_output_dtype_0 = const()[name = string("argmax_chunk_7_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> argmax_chunk_7 = reduce_argmax(axis = argmax_chunk_7_axis_0, keep_dims = argmax_chunk_7_keep_dims_0, output_dtype = argmax_chunk_7_output_dtype_0, x = logits_chunk_7)[name = string("argmax_chunk_7")];
+            tensor<int32, [1]> max_chunk_7_axes_0 = const()[name = string("max_chunk_7_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_7_keep_dims_0 = const()[name = string("max_chunk_7_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> max_chunk_7 = reduce_max(axes = max_chunk_7_axes_0, keep_dims = max_chunk_7_keep_dims_0, x = logits_chunk_7)[name = string("max_chunk_7")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_8_weight_1 = const()[name = string("logits_chunk_8_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67113600)))];
+            tensor<int32, [2]> logits_chunk_8_strides_1 = const()[name = string("logits_chunk_8_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_8_pad_type_1 = const()[name = string("logits_chunk_8_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_8_pad_1 = const()[name = string("logits_chunk_8_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_8_dilations_1 = const()[name = string("logits_chunk_8_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_8_groups_1 = const()[name = string("logits_chunk_8_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_8 = conv(dilations = logits_chunk_8_dilations_1, groups = logits_chunk_8_groups_1, pad = logits_chunk_8_pad_1, pad_type = logits_chunk_8_pad_type_1, strides = logits_chunk_8_strides_1, weight = logits_chunk_8_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_8")];
+            int32 argmax_chunk_8_axis_0 = const()[name = string("argmax_chunk_8_axis_0"), val = int32(1)];
+            bool argmax_chunk_8_keep_dims_0 = const()[name = string("argmax_chunk_8_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_8_output_dtype_0 = const()[name = string("argmax_chunk_8_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> argmax_chunk_8 = reduce_argmax(axis = argmax_chunk_8_axis_0, keep_dims = argmax_chunk_8_keep_dims_0, output_dtype = argmax_chunk_8_output_dtype_0, x = logits_chunk_8)[name = string("argmax_chunk_8")];
+            tensor<int32, [1]> max_chunk_8_axes_0 = const()[name = string("max_chunk_8_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_8_keep_dims_0 = const()[name = string("max_chunk_8_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> max_chunk_8 = reduce_max(axes = max_chunk_8_axes_0, keep_dims = max_chunk_8_keep_dims_0, x = logits_chunk_8)[name = string("max_chunk_8")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_9_weight_1 = const()[name = string("logits_chunk_9_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75502272)))];
+            tensor<int32, [2]> logits_chunk_9_strides_1 = const()[name = string("logits_chunk_9_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_9_pad_type_1 = const()[name = string("logits_chunk_9_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_9_pad_1 = const()[name = string("logits_chunk_9_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_9_dilations_1 = const()[name = string("logits_chunk_9_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_9_groups_1 = const()[name = string("logits_chunk_9_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_9 = conv(dilations = logits_chunk_9_dilations_1, groups = logits_chunk_9_groups_1, pad = logits_chunk_9_pad_1, pad_type = logits_chunk_9_pad_type_1, strides = logits_chunk_9_strides_1, weight = logits_chunk_9_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_9")];
+            int32 argmax_chunk_9_axis_0 = const()[name = string("argmax_chunk_9_axis_0"), val = int32(1)];
+            bool argmax_chunk_9_keep_dims_0 = const()[name = string("argmax_chunk_9_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_9_output_dtype_0 = const()[name = string("argmax_chunk_9_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> argmax_chunk_9 = reduce_argmax(axis = argmax_chunk_9_axis_0, keep_dims = argmax_chunk_9_keep_dims_0, output_dtype = argmax_chunk_9_output_dtype_0, x = logits_chunk_9)[name = string("argmax_chunk_9")];
+            tensor<int32, [1]> max_chunk_9_axes_0 = const()[name = string("max_chunk_9_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_9_keep_dims_0 = const()[name = string("max_chunk_9_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> max_chunk_9 = reduce_max(axes = max_chunk_9_axes_0, keep_dims = max_chunk_9_keep_dims_0, x = logits_chunk_9)[name = string("max_chunk_9")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_10_weight_1 = const()[name = string("logits_chunk_10_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83890944)))];
+            tensor<int32, [2]> logits_chunk_10_strides_1 = const()[name = string("logits_chunk_10_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_10_pad_type_1 = const()[name = string("logits_chunk_10_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_10_pad_1 = const()[name = string("logits_chunk_10_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_10_dilations_1 = const()[name = string("logits_chunk_10_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_10_groups_1 = const()[name = string("logits_chunk_10_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_10 = conv(dilations = logits_chunk_10_dilations_1, groups = logits_chunk_10_groups_1, pad = logits_chunk_10_pad_1, pad_type = logits_chunk_10_pad_type_1, strides = logits_chunk_10_strides_1, weight = logits_chunk_10_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_10")];
+            int32 argmax_chunk_10_axis_0 = const()[name = string("argmax_chunk_10_axis_0"), val = int32(1)];
+            bool argmax_chunk_10_keep_dims_0 = const()[name = string("argmax_chunk_10_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_10_output_dtype_0 = const()[name = string("argmax_chunk_10_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> argmax_chunk_10 = reduce_argmax(axis = argmax_chunk_10_axis_0, keep_dims = argmax_chunk_10_keep_dims_0, output_dtype = argmax_chunk_10_output_dtype_0, x = logits_chunk_10)[name = string("argmax_chunk_10")];
+            tensor<int32, [1]> max_chunk_10_axes_0 = const()[name = string("max_chunk_10_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_10_keep_dims_0 = const()[name = string("max_chunk_10_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> max_chunk_10 = reduce_max(axes = max_chunk_10_axes_0, keep_dims = max_chunk_10_keep_dims_0, x = logits_chunk_10)[name = string("max_chunk_10")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_11_weight_1 = const()[name = string("logits_chunk_11_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92279616)))];
+            tensor<int32, [2]> logits_chunk_11_strides_1 = const()[name = string("logits_chunk_11_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_11_pad_type_1 = const()[name = string("logits_chunk_11_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_11_pad_1 = const()[name = string("logits_chunk_11_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_11_dilations_1 = const()[name = string("logits_chunk_11_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_11_groups_1 = const()[name = string("logits_chunk_11_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_11 = conv(dilations = logits_chunk_11_dilations_1, groups = logits_chunk_11_groups_1, pad = logits_chunk_11_pad_1, pad_type = logits_chunk_11_pad_type_1, strides = logits_chunk_11_strides_1, weight = logits_chunk_11_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_11")];
+            int32 argmax_chunk_11_axis_0 = const()[name = string("argmax_chunk_11_axis_0"), val = int32(1)];
+            bool argmax_chunk_11_keep_dims_0 = const()[name = string("argmax_chunk_11_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_11_output_dtype_0 = const()[name = string("argmax_chunk_11_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> argmax_chunk_11 = reduce_argmax(axis = argmax_chunk_11_axis_0, keep_dims = argmax_chunk_11_keep_dims_0, output_dtype = argmax_chunk_11_output_dtype_0, x = logits_chunk_11)[name = string("argmax_chunk_11")];
+            tensor<int32, [1]> max_chunk_11_axes_0 = const()[name = string("max_chunk_11_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_11_keep_dims_0 = const()[name = string("max_chunk_11_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> max_chunk_11 = reduce_max(axes = max_chunk_11_axes_0, keep_dims = max_chunk_11_keep_dims_0, x = logits_chunk_11)[name = string("max_chunk_11")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_12_weight_1 = const()[name = string("logits_chunk_12_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100668288)))];
+            tensor<int32, [2]> logits_chunk_12_strides_1 = const()[name = string("logits_chunk_12_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_12_pad_type_1 = const()[name = string("logits_chunk_12_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_12_pad_1 = const()[name = string("logits_chunk_12_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_12_dilations_1 = const()[name = string("logits_chunk_12_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_12_groups_1 = const()[name = string("logits_chunk_12_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_12 = conv(dilations = logits_chunk_12_dilations_1, groups = logits_chunk_12_groups_1, pad = logits_chunk_12_pad_1, pad_type = logits_chunk_12_pad_type_1, strides = logits_chunk_12_strides_1, weight = logits_chunk_12_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_12")];
+            int32 argmax_chunk_12_axis_0 = const()[name = string("argmax_chunk_12_axis_0"), val = int32(1)];
+            bool argmax_chunk_12_keep_dims_0 = const()[name = string("argmax_chunk_12_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_12_output_dtype_0 = const()[name = string("argmax_chunk_12_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> argmax_chunk_12 = reduce_argmax(axis = argmax_chunk_12_axis_0, keep_dims = argmax_chunk_12_keep_dims_0, output_dtype = argmax_chunk_12_output_dtype_0, x = logits_chunk_12)[name = string("argmax_chunk_12")];
+            tensor<int32, [1]> max_chunk_12_axes_0 = const()[name = string("max_chunk_12_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_12_keep_dims_0 = const()[name = string("max_chunk_12_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> max_chunk_12 = reduce_max(axes = max_chunk_12_axes_0, keep_dims = max_chunk_12_keep_dims_0, x = logits_chunk_12)[name = string("max_chunk_12")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_13_weight_1 = const()[name = string("logits_chunk_13_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109056960)))];
+            tensor<int32, [2]> logits_chunk_13_strides_1 = const()[name = string("logits_chunk_13_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_13_pad_type_1 = const()[name = string("logits_chunk_13_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_13_pad_1 = const()[name = string("logits_chunk_13_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_13_dilations_1 = const()[name = string("logits_chunk_13_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_13_groups_1 = const()[name = string("logits_chunk_13_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_13 = conv(dilations = logits_chunk_13_dilations_1, groups = logits_chunk_13_groups_1, pad = logits_chunk_13_pad_1, pad_type = logits_chunk_13_pad_type_1, strides = logits_chunk_13_strides_1, weight = logits_chunk_13_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_13")];
+            int32 argmax_chunk_13_axis_0 = const()[name = string("argmax_chunk_13_axis_0"), val = int32(1)];
+            bool argmax_chunk_13_keep_dims_0 = const()[name = string("argmax_chunk_13_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_13_output_dtype_0 = const()[name = string("argmax_chunk_13_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> argmax_chunk_13 = reduce_argmax(axis = argmax_chunk_13_axis_0, keep_dims = argmax_chunk_13_keep_dims_0, output_dtype = argmax_chunk_13_output_dtype_0, x = logits_chunk_13)[name = string("argmax_chunk_13")];
+            tensor<int32, [1]> max_chunk_13_axes_0 = const()[name = string("max_chunk_13_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_13_keep_dims_0 = const()[name = string("max_chunk_13_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> max_chunk_13 = reduce_max(axes = max_chunk_13_axes_0, keep_dims = max_chunk_13_keep_dims_0, x = logits_chunk_13)[name = string("max_chunk_13")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_14_weight_1 = const()[name = string("logits_chunk_14_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117445632)))];
+            tensor<int32, [2]> logits_chunk_14_strides_1 = const()[name = string("logits_chunk_14_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_14_pad_type_1 = const()[name = string("logits_chunk_14_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_14_pad_1 = const()[name = string("logits_chunk_14_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_14_dilations_1 = const()[name = string("logits_chunk_14_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_14_groups_1 = const()[name = string("logits_chunk_14_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_14 = conv(dilations = logits_chunk_14_dilations_1, groups = logits_chunk_14_groups_1, pad = logits_chunk_14_pad_1, pad_type = logits_chunk_14_pad_type_1, strides = logits_chunk_14_strides_1, weight = logits_chunk_14_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_14")];
+            int32 argmax_chunk_14_axis_0 = const()[name = string("argmax_chunk_14_axis_0"), val = int32(1)];
+            bool argmax_chunk_14_keep_dims_0 = const()[name = string("argmax_chunk_14_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_14_output_dtype_0 = const()[name = string("argmax_chunk_14_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> argmax_chunk_14 = reduce_argmax(axis = argmax_chunk_14_axis_0, keep_dims = argmax_chunk_14_keep_dims_0, output_dtype = argmax_chunk_14_output_dtype_0, x = logits_chunk_14)[name = string("argmax_chunk_14")];
+            tensor<int32, [1]> max_chunk_14_axes_0 = const()[name = string("max_chunk_14_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_14_keep_dims_0 = const()[name = string("max_chunk_14_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> max_chunk_14 = reduce_max(axes = max_chunk_14_axes_0, keep_dims = max_chunk_14_keep_dims_0, x = logits_chunk_14)[name = string("max_chunk_14")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_15_weight_1 = const()[name = string("logits_chunk_15_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125834304)))];
+            tensor<int32, [2]> logits_chunk_15_strides_1 = const()[name = string("logits_chunk_15_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_15_pad_type_1 = const()[name = string("logits_chunk_15_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_15_pad_1 = const()[name = string("logits_chunk_15_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_15_dilations_1 = const()[name = string("logits_chunk_15_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_15_groups_1 = const()[name = string("logits_chunk_15_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_15 = conv(dilations = logits_chunk_15_dilations_1, groups = logits_chunk_15_groups_1, pad = logits_chunk_15_pad_1, pad_type = logits_chunk_15_pad_type_1, strides = logits_chunk_15_strides_1, weight = logits_chunk_15_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_15")];
+            int32 argmax_chunk_15_axis_0 = const()[name = string("argmax_chunk_15_axis_0"), val = int32(1)];
+            bool argmax_chunk_15_keep_dims_0 = const()[name = string("argmax_chunk_15_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_15_output_dtype_0 = const()[name = string("argmax_chunk_15_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> argmax_chunk_15 = reduce_argmax(axis = argmax_chunk_15_axis_0, keep_dims = argmax_chunk_15_keep_dims_0, output_dtype = argmax_chunk_15_output_dtype_0, x = logits_chunk_15)[name = string("argmax_chunk_15")];
+            tensor<int32, [1]> max_chunk_15_axes_0 = const()[name = string("max_chunk_15_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_15_keep_dims_0 = const()[name = string("max_chunk_15_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> max_chunk_15 = reduce_max(axes = max_chunk_15_axes_0, keep_dims = max_chunk_15_keep_dims_0, x = logits_chunk_15)[name = string("max_chunk_15")];
+            int32 logits_axis_0 = const()[name = string("logits_axis_0"), val = int32(1)];
+            bool logits_interleave_0 = const()[name = string("logits_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 32768, 1, 64]> logits = concat(axis = logits_axis_0, interleave = logits_interleave_0, values = (logits_chunk_0, logits_chunk_1, logits_chunk_2, logits_chunk_3, logits_chunk_4, logits_chunk_5, logits_chunk_6, logits_chunk_7, logits_chunk_8, logits_chunk_9, logits_chunk_10, logits_chunk_11, logits_chunk_12, logits_chunk_13, logits_chunk_14, logits_chunk_15))[name = string("logits")];
+            int32 values_axis_0 = const()[name = string("values_axis_0"), val = int32(1)];
+            bool values_interleave_0 = const()[name = string("values_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 64]> values = concat(axis = values_axis_0, interleave = values_interleave_0, values = (max_chunk_0, max_chunk_1, max_chunk_2, max_chunk_3, max_chunk_4, max_chunk_5, max_chunk_6, max_chunk_7, max_chunk_8, max_chunk_9, max_chunk_10, max_chunk_11, max_chunk_12, max_chunk_13, max_chunk_14, max_chunk_15))[name = string("values")];
+            tensor<int32, [1]> max_value_axes_0 = const()[name = string("max_value_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_value_keep_dims_0 = const()[name = string("max_value_keep_dims_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 64]> max_value = reduce_max(axes = max_value_axes_0, keep_dims = max_value_keep_dims_0, x = values)[name = string("max_value")];
+            int32 max_value_index_axis_0 = const()[name = string("max_value_index_axis_0"), val = int32(1)];
+            bool max_value_index_keep_dims_0 = const()[name = string("max_value_index_keep_dims_0"), val = bool(true)];
+            string max_value_index_output_dtype_0 = const()[name = string("max_value_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 1, 1, 64]> max_value_index = reduce_argmax(axis = max_value_index_axis_0, keep_dims = max_value_index_keep_dims_0, output_dtype = max_value_index_output_dtype_0, x = values)[name = string("max_value_index")];
+            string indices_chunk_0_int32_dtype_1 = const()[name = string("indices_chunk_0_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_1_int32_dtype_1 = const()[name = string("indices_chunk_1_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_2_int32_dtype_1 = const()[name = string("indices_chunk_2_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_3_int32_dtype_1 = const()[name = string("indices_chunk_3_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_4_int32_dtype_1 = const()[name = string("indices_chunk_4_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_5_int32_dtype_1 = const()[name = string("indices_chunk_5_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_6_int32_dtype_1 = const()[name = string("indices_chunk_6_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_7_int32_dtype_1 = const()[name = string("indices_chunk_7_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_8_int32_dtype_1 = const()[name = string("indices_chunk_8_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_9_int32_dtype_1 = const()[name = string("indices_chunk_9_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_10_int32_dtype_1 = const()[name = string("indices_chunk_10_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_11_int32_dtype_1 = const()[name = string("indices_chunk_11_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_12_int32_dtype_1 = const()[name = string("indices_chunk_12_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_13_int32_dtype_1 = const()[name = string("indices_chunk_13_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_14_int32_dtype_1 = const()[name = string("indices_chunk_14_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_15_int32_dtype_1 = const()[name = string("indices_chunk_15_int32_dtype_1"), val = string("int32")];
+            int32 indices_axis_1 = const()[name = string("indices_axis_1"), val = int32(1)];
+            bool indices_interleave_1 = const()[name = string("indices_interleave_1"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_15_int32 = cast(dtype = indices_chunk_15_int32_dtype_1, x = argmax_chunk_15)[name = string("cast_96")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_14_int32 = cast(dtype = indices_chunk_14_int32_dtype_1, x = argmax_chunk_14)[name = string("cast_97")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_13_int32 = cast(dtype = indices_chunk_13_int32_dtype_1, x = argmax_chunk_13)[name = string("cast_98")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_12_int32 = cast(dtype = indices_chunk_12_int32_dtype_1, x = argmax_chunk_12)[name = string("cast_99")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_11_int32 = cast(dtype = indices_chunk_11_int32_dtype_1, x = argmax_chunk_11)[name = string("cast_100")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_10_int32 = cast(dtype = indices_chunk_10_int32_dtype_1, x = argmax_chunk_10)[name = string("cast_101")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_9_int32 = cast(dtype = indices_chunk_9_int32_dtype_1, x = argmax_chunk_9)[name = string("cast_102")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_8_int32 = cast(dtype = indices_chunk_8_int32_dtype_1, x = argmax_chunk_8)[name = string("cast_103")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_7_int32 = cast(dtype = indices_chunk_7_int32_dtype_1, x = argmax_chunk_7)[name = string("cast_104")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_6_int32 = cast(dtype = indices_chunk_6_int32_dtype_1, x = argmax_chunk_6)[name = string("cast_105")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_5_int32 = cast(dtype = indices_chunk_5_int32_dtype_1, x = argmax_chunk_5)[name = string("cast_106")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_4_int32 = cast(dtype = indices_chunk_4_int32_dtype_1, x = argmax_chunk_4)[name = string("cast_107")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_3_int32 = cast(dtype = indices_chunk_3_int32_dtype_1, x = argmax_chunk_3)[name = string("cast_108")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_2_int32 = cast(dtype = indices_chunk_2_int32_dtype_1, x = argmax_chunk_2)[name = string("cast_109")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_1_int32 = cast(dtype = indices_chunk_1_int32_dtype_1, x = argmax_chunk_1)[name = string("cast_110")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_0_int32 = cast(dtype = indices_chunk_0_int32_dtype_1, x = argmax_chunk_0)[name = string("cast_111")];
+            tensor<int32, [1, 16, 1, 64]> indices = concat(axis = indices_axis_1, interleave = indices_interleave_1, values = (indices_chunk_0_int32, indices_chunk_1_int32, indices_chunk_2_int32, indices_chunk_3_int32, indices_chunk_4_int32, indices_chunk_5_int32, indices_chunk_6_int32, indices_chunk_7_int32, indices_chunk_8_int32, indices_chunk_9_int32, indices_chunk_10_int32, indices_chunk_11_int32, indices_chunk_12_int32, indices_chunk_13_int32, indices_chunk_14_int32, indices_chunk_15_int32))[name = string("indices")];
+            int32 argmax_chunks_axis_1 = const()[name = string("argmax_chunks_axis_1"), val = int32(1)];
+            bool argmax_chunks_validate_indices_1 = const()[name = string("argmax_chunks_validate_indices_1"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 64]> argmax_chunks = gather_along_axis(axis = argmax_chunks_axis_1, indices = max_value_index, validate_indices = argmax_chunks_validate_indices_1, x = indices)[name = string("argmax_chunks")];
+            int32 mul_1_x_0 = const()[name = string("mul_1_x_0"), val = int32(2048)];
+            tensor<int32, [1, 1, 1, 64]> mul_1 = mul(x = mul_1_x_0, y = max_value_index)[name = string("mul_1")];
+            tensor<int32, [1, 1, 1, 64]> argmax = add(x = argmax_chunks, y = mul_1)[name = string("argmax")];
+        } -> (logits, argmax, max_value);
+    func lm_head_length_8<ios18>(tensor<fp16, [1, 2048, 1, 8]> hidden_states) {
+            tensor<fp16, [1, 2048, 1, 8]> final_norm_rmsnorm_abs = abs(x = hidden_states)[name = string("final_norm_rmsnorm_abs")];
+            tensor<int32, [1]> final_norm_rmsnorm_maxval_axes_1 = const()[name = string("final_norm_rmsnorm_maxval_axes_1"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_maxval_keep_dims_1 = const()[name = string("final_norm_rmsnorm_maxval_keep_dims_1"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> final_norm_rmsnorm_maxval = reduce_max(axes = final_norm_rmsnorm_maxval_axes_1, keep_dims = final_norm_rmsnorm_maxval_keep_dims_1, x = final_norm_rmsnorm_abs)[name = string("final_norm_rmsnorm_maxval")];
+            fp16 final_norm_rmsnorm_maxval_clipped_alpha_1 = const()[name = string("final_norm_rmsnorm_maxval_clipped_alpha_1"), val = fp16(0x1p-14)];
+            fp16 final_norm_rmsnorm_maxval_clipped_beta_1 = const()[name = string("final_norm_rmsnorm_maxval_clipped_beta_1"), val = fp16(inf)];
+            tensor<fp16, [1, 1, 1, 8]> final_norm_rmsnorm_maxval_clipped = clip(alpha = final_norm_rmsnorm_maxval_clipped_alpha_1, beta = final_norm_rmsnorm_maxval_clipped_beta_1, x = final_norm_rmsnorm_maxval)[name = string("final_norm_rmsnorm_maxval_clipped")];
+            tensor<fp16, [1, 2048, 1, 8]> final_norm_rmsnorm_scaled = real_div(x = hidden_states, y = final_norm_rmsnorm_maxval_clipped)[name = string("final_norm_rmsnorm_scaled")];
+            tensor<int32, [1]> final_norm_rmsnorm_squared_sum_axes_1 = const()[name = string("final_norm_rmsnorm_squared_sum_axes_1"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_squared_sum_keep_dims_1 = const()[name = string("final_norm_rmsnorm_squared_sum_keep_dims_1"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> final_norm_rmsnorm_squared_sum = reduce_sum_square(axes = final_norm_rmsnorm_squared_sum_axes_1, keep_dims = final_norm_rmsnorm_squared_sum_keep_dims_1, x = final_norm_rmsnorm_scaled)[name = string("final_norm_rmsnorm_squared_sum")];
+            fp16 final_norm_rmsnorm_rsqrt_epsilon_1 = const()[name = string("final_norm_rmsnorm_rsqrt_epsilon_1"), val = fp16(0x1p-14)];
+            tensor<fp16, [1, 1, 1, 8]> final_norm_rmsnorm_rsqrt = rsqrt(epsilon = final_norm_rmsnorm_rsqrt_epsilon_1, x = final_norm_rmsnorm_squared_sum)[name = string("final_norm_rmsnorm_rsqrt")];
+            fp16 final_norm_rmsnorm_dim_scaled_y_1 = const()[name = string("final_norm_rmsnorm_dim_scaled_y_1"), val = fp16(0x1.6ap+5)];
+            tensor<fp16, [1, 2048, 1, 8]> final_norm_rmsnorm_dim_scaled = mul(x = final_norm_rmsnorm_scaled, y = final_norm_rmsnorm_dim_scaled_y_1)[name = string("final_norm_rmsnorm_dim_scaled")];
+            tensor<fp16, [1, 2048, 1, 8]> final_norm_rmsnorm_normalized = mul(x = final_norm_rmsnorm_dim_scaled, y = final_norm_rmsnorm_rsqrt)[name = string("final_norm_rmsnorm_normalized")];
+            tensor<fp16, [2048, 1, 1]> final_norm_rmsnorm_y_1 = const()[name = string("final_norm_rmsnorm_y_1"), val = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 2048, 1, 8]> final_norm_rmsnorm = mul(x = final_norm_rmsnorm_normalized, y = final_norm_rmsnorm_y_1)[name = string("final_norm_rmsnorm")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_0_weight_1 = const()[name = string("logits_chunk_0_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4224)))];
+            tensor<int32, [2]> logits_chunk_0_strides_1 = const()[name = string("logits_chunk_0_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_0_pad_type_1 = const()[name = string("logits_chunk_0_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_0_pad_1 = const()[name = string("logits_chunk_0_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_0_dilations_1 = const()[name = string("logits_chunk_0_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_0_groups_1 = const()[name = string("logits_chunk_0_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_0 = conv(dilations = logits_chunk_0_dilations_1, groups = logits_chunk_0_groups_1, pad = logits_chunk_0_pad_1, pad_type = logits_chunk_0_pad_type_1, strides = logits_chunk_0_strides_1, weight = logits_chunk_0_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_0")];
+            int32 argmax_chunk_0_axis_0 = const()[name = string("argmax_chunk_0_axis_0"), val = int32(1)];
+            bool argmax_chunk_0_keep_dims_0 = const()[name = string("argmax_chunk_0_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_0_output_dtype_0 = const()[name = string("argmax_chunk_0_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> argmax_chunk_0 = reduce_argmax(axis = argmax_chunk_0_axis_0, keep_dims = argmax_chunk_0_keep_dims_0, output_dtype = argmax_chunk_0_output_dtype_0, x = logits_chunk_0)[name = string("argmax_chunk_0")];
+            tensor<int32, [1]> max_chunk_0_axes_0 = const()[name = string("max_chunk_0_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_0_keep_dims_0 = const()[name = string("max_chunk_0_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> max_chunk_0 = reduce_max(axes = max_chunk_0_axes_0, keep_dims = max_chunk_0_keep_dims_0, x = logits_chunk_0)[name = string("max_chunk_0")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_1_weight_1 = const()[name = string("logits_chunk_1_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8392896)))];
+            tensor<int32, [2]> logits_chunk_1_strides_1 = const()[name = string("logits_chunk_1_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_1_pad_type_1 = const()[name = string("logits_chunk_1_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_1_pad_1 = const()[name = string("logits_chunk_1_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_1_dilations_1 = const()[name = string("logits_chunk_1_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_1_groups_1 = const()[name = string("logits_chunk_1_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_1 = conv(dilations = logits_chunk_1_dilations_1, groups = logits_chunk_1_groups_1, pad = logits_chunk_1_pad_1, pad_type = logits_chunk_1_pad_type_1, strides = logits_chunk_1_strides_1, weight = logits_chunk_1_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_1")];
+            int32 argmax_chunk_1_axis_0 = const()[name = string("argmax_chunk_1_axis_0"), val = int32(1)];
+            bool argmax_chunk_1_keep_dims_0 = const()[name = string("argmax_chunk_1_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_1_output_dtype_0 = const()[name = string("argmax_chunk_1_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> argmax_chunk_1 = reduce_argmax(axis = argmax_chunk_1_axis_0, keep_dims = argmax_chunk_1_keep_dims_0, output_dtype = argmax_chunk_1_output_dtype_0, x = logits_chunk_1)[name = string("argmax_chunk_1")];
+            tensor<int32, [1]> max_chunk_1_axes_0 = const()[name = string("max_chunk_1_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_1_keep_dims_0 = const()[name = string("max_chunk_1_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> max_chunk_1 = reduce_max(axes = max_chunk_1_axes_0, keep_dims = max_chunk_1_keep_dims_0, x = logits_chunk_1)[name = string("max_chunk_1")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_2_weight_1 = const()[name = string("logits_chunk_2_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16781568)))];
+            tensor<int32, [2]> logits_chunk_2_strides_1 = const()[name = string("logits_chunk_2_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_2_pad_type_1 = const()[name = string("logits_chunk_2_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_2_pad_1 = const()[name = string("logits_chunk_2_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_2_dilations_1 = const()[name = string("logits_chunk_2_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_2_groups_1 = const()[name = string("logits_chunk_2_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_2 = conv(dilations = logits_chunk_2_dilations_1, groups = logits_chunk_2_groups_1, pad = logits_chunk_2_pad_1, pad_type = logits_chunk_2_pad_type_1, strides = logits_chunk_2_strides_1, weight = logits_chunk_2_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_2")];
+            int32 argmax_chunk_2_axis_0 = const()[name = string("argmax_chunk_2_axis_0"), val = int32(1)];
+            bool argmax_chunk_2_keep_dims_0 = const()[name = string("argmax_chunk_2_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_2_output_dtype_0 = const()[name = string("argmax_chunk_2_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> argmax_chunk_2 = reduce_argmax(axis = argmax_chunk_2_axis_0, keep_dims = argmax_chunk_2_keep_dims_0, output_dtype = argmax_chunk_2_output_dtype_0, x = logits_chunk_2)[name = string("argmax_chunk_2")];
+            tensor<int32, [1]> max_chunk_2_axes_0 = const()[name = string("max_chunk_2_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_2_keep_dims_0 = const()[name = string("max_chunk_2_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> max_chunk_2 = reduce_max(axes = max_chunk_2_axes_0, keep_dims = max_chunk_2_keep_dims_0, x = logits_chunk_2)[name = string("max_chunk_2")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_3_weight_1 = const()[name = string("logits_chunk_3_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25170240)))];
+            tensor<int32, [2]> logits_chunk_3_strides_1 = const()[name = string("logits_chunk_3_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_3_pad_type_1 = const()[name = string("logits_chunk_3_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_3_pad_1 = const()[name = string("logits_chunk_3_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_3_dilations_1 = const()[name = string("logits_chunk_3_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_3_groups_1 = const()[name = string("logits_chunk_3_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_3 = conv(dilations = logits_chunk_3_dilations_1, groups = logits_chunk_3_groups_1, pad = logits_chunk_3_pad_1, pad_type = logits_chunk_3_pad_type_1, strides = logits_chunk_3_strides_1, weight = logits_chunk_3_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_3")];
+            int32 argmax_chunk_3_axis_0 = const()[name = string("argmax_chunk_3_axis_0"), val = int32(1)];
+            bool argmax_chunk_3_keep_dims_0 = const()[name = string("argmax_chunk_3_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_3_output_dtype_0 = const()[name = string("argmax_chunk_3_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> argmax_chunk_3 = reduce_argmax(axis = argmax_chunk_3_axis_0, keep_dims = argmax_chunk_3_keep_dims_0, output_dtype = argmax_chunk_3_output_dtype_0, x = logits_chunk_3)[name = string("argmax_chunk_3")];
+            tensor<int32, [1]> max_chunk_3_axes_0 = const()[name = string("max_chunk_3_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_3_keep_dims_0 = const()[name = string("max_chunk_3_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> max_chunk_3 = reduce_max(axes = max_chunk_3_axes_0, keep_dims = max_chunk_3_keep_dims_0, x = logits_chunk_3)[name = string("max_chunk_3")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_4_weight_1 = const()[name = string("logits_chunk_4_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33558912)))];
+            tensor<int32, [2]> logits_chunk_4_strides_1 = const()[name = string("logits_chunk_4_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_4_pad_type_1 = const()[name = string("logits_chunk_4_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_4_pad_1 = const()[name = string("logits_chunk_4_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_4_dilations_1 = const()[name = string("logits_chunk_4_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_4_groups_1 = const()[name = string("logits_chunk_4_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_4 = conv(dilations = logits_chunk_4_dilations_1, groups = logits_chunk_4_groups_1, pad = logits_chunk_4_pad_1, pad_type = logits_chunk_4_pad_type_1, strides = logits_chunk_4_strides_1, weight = logits_chunk_4_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_4")];
+            int32 argmax_chunk_4_axis_0 = const()[name = string("argmax_chunk_4_axis_0"), val = int32(1)];
+            bool argmax_chunk_4_keep_dims_0 = const()[name = string("argmax_chunk_4_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_4_output_dtype_0 = const()[name = string("argmax_chunk_4_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> argmax_chunk_4 = reduce_argmax(axis = argmax_chunk_4_axis_0, keep_dims = argmax_chunk_4_keep_dims_0, output_dtype = argmax_chunk_4_output_dtype_0, x = logits_chunk_4)[name = string("argmax_chunk_4")];
+            tensor<int32, [1]> max_chunk_4_axes_0 = const()[name = string("max_chunk_4_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_4_keep_dims_0 = const()[name = string("max_chunk_4_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> max_chunk_4 = reduce_max(axes = max_chunk_4_axes_0, keep_dims = max_chunk_4_keep_dims_0, x = logits_chunk_4)[name = string("max_chunk_4")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_5_weight_1 = const()[name = string("logits_chunk_5_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41947584)))];
+            tensor<int32, [2]> logits_chunk_5_strides_1 = const()[name = string("logits_chunk_5_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_5_pad_type_1 = const()[name = string("logits_chunk_5_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_5_pad_1 = const()[name = string("logits_chunk_5_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_5_dilations_1 = const()[name = string("logits_chunk_5_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_5_groups_1 = const()[name = string("logits_chunk_5_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_5 = conv(dilations = logits_chunk_5_dilations_1, groups = logits_chunk_5_groups_1, pad = logits_chunk_5_pad_1, pad_type = logits_chunk_5_pad_type_1, strides = logits_chunk_5_strides_1, weight = logits_chunk_5_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_5")];
+            int32 argmax_chunk_5_axis_0 = const()[name = string("argmax_chunk_5_axis_0"), val = int32(1)];
+            bool argmax_chunk_5_keep_dims_0 = const()[name = string("argmax_chunk_5_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_5_output_dtype_0 = const()[name = string("argmax_chunk_5_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> argmax_chunk_5 = reduce_argmax(axis = argmax_chunk_5_axis_0, keep_dims = argmax_chunk_5_keep_dims_0, output_dtype = argmax_chunk_5_output_dtype_0, x = logits_chunk_5)[name = string("argmax_chunk_5")];
+            tensor<int32, [1]> max_chunk_5_axes_0 = const()[name = string("max_chunk_5_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_5_keep_dims_0 = const()[name = string("max_chunk_5_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> max_chunk_5 = reduce_max(axes = max_chunk_5_axes_0, keep_dims = max_chunk_5_keep_dims_0, x = logits_chunk_5)[name = string("max_chunk_5")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_6_weight_1 = const()[name = string("logits_chunk_6_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50336256)))];
+            tensor<int32, [2]> logits_chunk_6_strides_1 = const()[name = string("logits_chunk_6_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_6_pad_type_1 = const()[name = string("logits_chunk_6_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_6_pad_1 = const()[name = string("logits_chunk_6_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_6_dilations_1 = const()[name = string("logits_chunk_6_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_6_groups_1 = const()[name = string("logits_chunk_6_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_6 = conv(dilations = logits_chunk_6_dilations_1, groups = logits_chunk_6_groups_1, pad = logits_chunk_6_pad_1, pad_type = logits_chunk_6_pad_type_1, strides = logits_chunk_6_strides_1, weight = logits_chunk_6_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_6")];
+            int32 argmax_chunk_6_axis_0 = const()[name = string("argmax_chunk_6_axis_0"), val = int32(1)];
+            bool argmax_chunk_6_keep_dims_0 = const()[name = string("argmax_chunk_6_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_6_output_dtype_0 = const()[name = string("argmax_chunk_6_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> argmax_chunk_6 = reduce_argmax(axis = argmax_chunk_6_axis_0, keep_dims = argmax_chunk_6_keep_dims_0, output_dtype = argmax_chunk_6_output_dtype_0, x = logits_chunk_6)[name = string("argmax_chunk_6")];
+            tensor<int32, [1]> max_chunk_6_axes_0 = const()[name = string("max_chunk_6_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_6_keep_dims_0 = const()[name = string("max_chunk_6_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> max_chunk_6 = reduce_max(axes = max_chunk_6_axes_0, keep_dims = max_chunk_6_keep_dims_0, x = logits_chunk_6)[name = string("max_chunk_6")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_7_weight_1 = const()[name = string("logits_chunk_7_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58724928)))];
+            tensor<int32, [2]> logits_chunk_7_strides_1 = const()[name = string("logits_chunk_7_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_7_pad_type_1 = const()[name = string("logits_chunk_7_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_7_pad_1 = const()[name = string("logits_chunk_7_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_7_dilations_1 = const()[name = string("logits_chunk_7_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_7_groups_1 = const()[name = string("logits_chunk_7_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_7 = conv(dilations = logits_chunk_7_dilations_1, groups = logits_chunk_7_groups_1, pad = logits_chunk_7_pad_1, pad_type = logits_chunk_7_pad_type_1, strides = logits_chunk_7_strides_1, weight = logits_chunk_7_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_7")];
+            int32 argmax_chunk_7_axis_0 = const()[name = string("argmax_chunk_7_axis_0"), val = int32(1)];
+            bool argmax_chunk_7_keep_dims_0 = const()[name = string("argmax_chunk_7_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_7_output_dtype_0 = const()[name = string("argmax_chunk_7_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> argmax_chunk_7 = reduce_argmax(axis = argmax_chunk_7_axis_0, keep_dims = argmax_chunk_7_keep_dims_0, output_dtype = argmax_chunk_7_output_dtype_0, x = logits_chunk_7)[name = string("argmax_chunk_7")];
+            tensor<int32, [1]> max_chunk_7_axes_0 = const()[name = string("max_chunk_7_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_7_keep_dims_0 = const()[name = string("max_chunk_7_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> max_chunk_7 = reduce_max(axes = max_chunk_7_axes_0, keep_dims = max_chunk_7_keep_dims_0, x = logits_chunk_7)[name = string("max_chunk_7")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_8_weight_1 = const()[name = string("logits_chunk_8_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67113600)))];
+            tensor<int32, [2]> logits_chunk_8_strides_1 = const()[name = string("logits_chunk_8_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_8_pad_type_1 = const()[name = string("logits_chunk_8_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_8_pad_1 = const()[name = string("logits_chunk_8_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_8_dilations_1 = const()[name = string("logits_chunk_8_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_8_groups_1 = const()[name = string("logits_chunk_8_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_8 = conv(dilations = logits_chunk_8_dilations_1, groups = logits_chunk_8_groups_1, pad = logits_chunk_8_pad_1, pad_type = logits_chunk_8_pad_type_1, strides = logits_chunk_8_strides_1, weight = logits_chunk_8_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_8")];
+            int32 argmax_chunk_8_axis_0 = const()[name = string("argmax_chunk_8_axis_0"), val = int32(1)];
+            bool argmax_chunk_8_keep_dims_0 = const()[name = string("argmax_chunk_8_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_8_output_dtype_0 = const()[name = string("argmax_chunk_8_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> argmax_chunk_8 = reduce_argmax(axis = argmax_chunk_8_axis_0, keep_dims = argmax_chunk_8_keep_dims_0, output_dtype = argmax_chunk_8_output_dtype_0, x = logits_chunk_8)[name = string("argmax_chunk_8")];
+            tensor<int32, [1]> max_chunk_8_axes_0 = const()[name = string("max_chunk_8_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_8_keep_dims_0 = const()[name = string("max_chunk_8_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> max_chunk_8 = reduce_max(axes = max_chunk_8_axes_0, keep_dims = max_chunk_8_keep_dims_0, x = logits_chunk_8)[name = string("max_chunk_8")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_9_weight_1 = const()[name = string("logits_chunk_9_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75502272)))];
+            tensor<int32, [2]> logits_chunk_9_strides_1 = const()[name = string("logits_chunk_9_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_9_pad_type_1 = const()[name = string("logits_chunk_9_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_9_pad_1 = const()[name = string("logits_chunk_9_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_9_dilations_1 = const()[name = string("logits_chunk_9_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_9_groups_1 = const()[name = string("logits_chunk_9_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_9 = conv(dilations = logits_chunk_9_dilations_1, groups = logits_chunk_9_groups_1, pad = logits_chunk_9_pad_1, pad_type = logits_chunk_9_pad_type_1, strides = logits_chunk_9_strides_1, weight = logits_chunk_9_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_9")];
+            int32 argmax_chunk_9_axis_0 = const()[name = string("argmax_chunk_9_axis_0"), val = int32(1)];
+            bool argmax_chunk_9_keep_dims_0 = const()[name = string("argmax_chunk_9_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_9_output_dtype_0 = const()[name = string("argmax_chunk_9_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> argmax_chunk_9 = reduce_argmax(axis = argmax_chunk_9_axis_0, keep_dims = argmax_chunk_9_keep_dims_0, output_dtype = argmax_chunk_9_output_dtype_0, x = logits_chunk_9)[name = string("argmax_chunk_9")];
+            tensor<int32, [1]> max_chunk_9_axes_0 = const()[name = string("max_chunk_9_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_9_keep_dims_0 = const()[name = string("max_chunk_9_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> max_chunk_9 = reduce_max(axes = max_chunk_9_axes_0, keep_dims = max_chunk_9_keep_dims_0, x = logits_chunk_9)[name = string("max_chunk_9")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_10_weight_1 = const()[name = string("logits_chunk_10_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83890944)))];
+            tensor<int32, [2]> logits_chunk_10_strides_1 = const()[name = string("logits_chunk_10_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_10_pad_type_1 = const()[name = string("logits_chunk_10_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_10_pad_1 = const()[name = string("logits_chunk_10_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_10_dilations_1 = const()[name = string("logits_chunk_10_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_10_groups_1 = const()[name = string("logits_chunk_10_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_10 = conv(dilations = logits_chunk_10_dilations_1, groups = logits_chunk_10_groups_1, pad = logits_chunk_10_pad_1, pad_type = logits_chunk_10_pad_type_1, strides = logits_chunk_10_strides_1, weight = logits_chunk_10_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_10")];
+            int32 argmax_chunk_10_axis_0 = const()[name = string("argmax_chunk_10_axis_0"), val = int32(1)];
+            bool argmax_chunk_10_keep_dims_0 = const()[name = string("argmax_chunk_10_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_10_output_dtype_0 = const()[name = string("argmax_chunk_10_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> argmax_chunk_10 = reduce_argmax(axis = argmax_chunk_10_axis_0, keep_dims = argmax_chunk_10_keep_dims_0, output_dtype = argmax_chunk_10_output_dtype_0, x = logits_chunk_10)[name = string("argmax_chunk_10")];
+            tensor<int32, [1]> max_chunk_10_axes_0 = const()[name = string("max_chunk_10_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_10_keep_dims_0 = const()[name = string("max_chunk_10_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> max_chunk_10 = reduce_max(axes = max_chunk_10_axes_0, keep_dims = max_chunk_10_keep_dims_0, x = logits_chunk_10)[name = string("max_chunk_10")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_11_weight_1 = const()[name = string("logits_chunk_11_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92279616)))];
+            tensor<int32, [2]> logits_chunk_11_strides_1 = const()[name = string("logits_chunk_11_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_11_pad_type_1 = const()[name = string("logits_chunk_11_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_11_pad_1 = const()[name = string("logits_chunk_11_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_11_dilations_1 = const()[name = string("logits_chunk_11_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_11_groups_1 = const()[name = string("logits_chunk_11_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_11 = conv(dilations = logits_chunk_11_dilations_1, groups = logits_chunk_11_groups_1, pad = logits_chunk_11_pad_1, pad_type = logits_chunk_11_pad_type_1, strides = logits_chunk_11_strides_1, weight = logits_chunk_11_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_11")];
+            int32 argmax_chunk_11_axis_0 = const()[name = string("argmax_chunk_11_axis_0"), val = int32(1)];
+            bool argmax_chunk_11_keep_dims_0 = const()[name = string("argmax_chunk_11_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_11_output_dtype_0 = const()[name = string("argmax_chunk_11_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> argmax_chunk_11 = reduce_argmax(axis = argmax_chunk_11_axis_0, keep_dims = argmax_chunk_11_keep_dims_0, output_dtype = argmax_chunk_11_output_dtype_0, x = logits_chunk_11)[name = string("argmax_chunk_11")];
+            tensor<int32, [1]> max_chunk_11_axes_0 = const()[name = string("max_chunk_11_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_11_keep_dims_0 = const()[name = string("max_chunk_11_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> max_chunk_11 = reduce_max(axes = max_chunk_11_axes_0, keep_dims = max_chunk_11_keep_dims_0, x = logits_chunk_11)[name = string("max_chunk_11")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_12_weight_1 = const()[name = string("logits_chunk_12_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100668288)))];
+            tensor<int32, [2]> logits_chunk_12_strides_1 = const()[name = string("logits_chunk_12_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_12_pad_type_1 = const()[name = string("logits_chunk_12_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_12_pad_1 = const()[name = string("logits_chunk_12_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_12_dilations_1 = const()[name = string("logits_chunk_12_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_12_groups_1 = const()[name = string("logits_chunk_12_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_12 = conv(dilations = logits_chunk_12_dilations_1, groups = logits_chunk_12_groups_1, pad = logits_chunk_12_pad_1, pad_type = logits_chunk_12_pad_type_1, strides = logits_chunk_12_strides_1, weight = logits_chunk_12_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_12")];
+            int32 argmax_chunk_12_axis_0 = const()[name = string("argmax_chunk_12_axis_0"), val = int32(1)];
+            bool argmax_chunk_12_keep_dims_0 = const()[name = string("argmax_chunk_12_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_12_output_dtype_0 = const()[name = string("argmax_chunk_12_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> argmax_chunk_12 = reduce_argmax(axis = argmax_chunk_12_axis_0, keep_dims = argmax_chunk_12_keep_dims_0, output_dtype = argmax_chunk_12_output_dtype_0, x = logits_chunk_12)[name = string("argmax_chunk_12")];
+            tensor<int32, [1]> max_chunk_12_axes_0 = const()[name = string("max_chunk_12_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_12_keep_dims_0 = const()[name = string("max_chunk_12_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> max_chunk_12 = reduce_max(axes = max_chunk_12_axes_0, keep_dims = max_chunk_12_keep_dims_0, x = logits_chunk_12)[name = string("max_chunk_12")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_13_weight_1 = const()[name = string("logits_chunk_13_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109056960)))];
+            tensor<int32, [2]> logits_chunk_13_strides_1 = const()[name = string("logits_chunk_13_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_13_pad_type_1 = const()[name = string("logits_chunk_13_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_13_pad_1 = const()[name = string("logits_chunk_13_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_13_dilations_1 = const()[name = string("logits_chunk_13_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_13_groups_1 = const()[name = string("logits_chunk_13_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_13 = conv(dilations = logits_chunk_13_dilations_1, groups = logits_chunk_13_groups_1, pad = logits_chunk_13_pad_1, pad_type = logits_chunk_13_pad_type_1, strides = logits_chunk_13_strides_1, weight = logits_chunk_13_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_13")];
+            int32 argmax_chunk_13_axis_0 = const()[name = string("argmax_chunk_13_axis_0"), val = int32(1)];
+            bool argmax_chunk_13_keep_dims_0 = const()[name = string("argmax_chunk_13_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_13_output_dtype_0 = const()[name = string("argmax_chunk_13_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> argmax_chunk_13 = reduce_argmax(axis = argmax_chunk_13_axis_0, keep_dims = argmax_chunk_13_keep_dims_0, output_dtype = argmax_chunk_13_output_dtype_0, x = logits_chunk_13)[name = string("argmax_chunk_13")];
+            tensor<int32, [1]> max_chunk_13_axes_0 = const()[name = string("max_chunk_13_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_13_keep_dims_0 = const()[name = string("max_chunk_13_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> max_chunk_13 = reduce_max(axes = max_chunk_13_axes_0, keep_dims = max_chunk_13_keep_dims_0, x = logits_chunk_13)[name = string("max_chunk_13")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_14_weight_1 = const()[name = string("logits_chunk_14_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117445632)))];
+            tensor<int32, [2]> logits_chunk_14_strides_1 = const()[name = string("logits_chunk_14_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_14_pad_type_1 = const()[name = string("logits_chunk_14_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_14_pad_1 = const()[name = string("logits_chunk_14_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_14_dilations_1 = const()[name = string("logits_chunk_14_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_14_groups_1 = const()[name = string("logits_chunk_14_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_14 = conv(dilations = logits_chunk_14_dilations_1, groups = logits_chunk_14_groups_1, pad = logits_chunk_14_pad_1, pad_type = logits_chunk_14_pad_type_1, strides = logits_chunk_14_strides_1, weight = logits_chunk_14_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_14")];
+            int32 argmax_chunk_14_axis_0 = const()[name = string("argmax_chunk_14_axis_0"), val = int32(1)];
+            bool argmax_chunk_14_keep_dims_0 = const()[name = string("argmax_chunk_14_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_14_output_dtype_0 = const()[name = string("argmax_chunk_14_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> argmax_chunk_14 = reduce_argmax(axis = argmax_chunk_14_axis_0, keep_dims = argmax_chunk_14_keep_dims_0, output_dtype = argmax_chunk_14_output_dtype_0, x = logits_chunk_14)[name = string("argmax_chunk_14")];
+            tensor<int32, [1]> max_chunk_14_axes_0 = const()[name = string("max_chunk_14_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_14_keep_dims_0 = const()[name = string("max_chunk_14_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> max_chunk_14 = reduce_max(axes = max_chunk_14_axes_0, keep_dims = max_chunk_14_keep_dims_0, x = logits_chunk_14)[name = string("max_chunk_14")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_15_weight_1 = const()[name = string("logits_chunk_15_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125834304)))];
+            tensor<int32, [2]> logits_chunk_15_strides_1 = const()[name = string("logits_chunk_15_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_15_pad_type_1 = const()[name = string("logits_chunk_15_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_15_pad_1 = const()[name = string("logits_chunk_15_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_15_dilations_1 = const()[name = string("logits_chunk_15_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_15_groups_1 = const()[name = string("logits_chunk_15_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_15 = conv(dilations = logits_chunk_15_dilations_1, groups = logits_chunk_15_groups_1, pad = logits_chunk_15_pad_1, pad_type = logits_chunk_15_pad_type_1, strides = logits_chunk_15_strides_1, weight = logits_chunk_15_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_15")];
+            int32 argmax_chunk_15_axis_0 = const()[name = string("argmax_chunk_15_axis_0"), val = int32(1)];
+            bool argmax_chunk_15_keep_dims_0 = const()[name = string("argmax_chunk_15_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_15_output_dtype_0 = const()[name = string("argmax_chunk_15_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> argmax_chunk_15 = reduce_argmax(axis = argmax_chunk_15_axis_0, keep_dims = argmax_chunk_15_keep_dims_0, output_dtype = argmax_chunk_15_output_dtype_0, x = logits_chunk_15)[name = string("argmax_chunk_15")];
+            tensor<int32, [1]> max_chunk_15_axes_0 = const()[name = string("max_chunk_15_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_15_keep_dims_0 = const()[name = string("max_chunk_15_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> max_chunk_15 = reduce_max(axes = max_chunk_15_axes_0, keep_dims = max_chunk_15_keep_dims_0, x = logits_chunk_15)[name = string("max_chunk_15")];
+            int32 logits_axis_0 = const()[name = string("logits_axis_0"), val = int32(1)];
+            bool logits_interleave_0 = const()[name = string("logits_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 32768, 1, 8]> logits = concat(axis = logits_axis_0, interleave = logits_interleave_0, values = (logits_chunk_0, logits_chunk_1, logits_chunk_2, logits_chunk_3, logits_chunk_4, logits_chunk_5, logits_chunk_6, logits_chunk_7, logits_chunk_8, logits_chunk_9, logits_chunk_10, logits_chunk_11, logits_chunk_12, logits_chunk_13, logits_chunk_14, logits_chunk_15))[name = string("logits")];
+            int32 values_axis_0 = const()[name = string("values_axis_0"), val = int32(1)];
+            bool values_interleave_0 = const()[name = string("values_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 8]> values = concat(axis = values_axis_0, interleave = values_interleave_0, values = (max_chunk_0, max_chunk_1, max_chunk_2, max_chunk_3, max_chunk_4, max_chunk_5, max_chunk_6, max_chunk_7, max_chunk_8, max_chunk_9, max_chunk_10, max_chunk_11, max_chunk_12, max_chunk_13, max_chunk_14, max_chunk_15))[name = string("values")];
+            tensor<int32, [1]> max_value_axes_0 = const()[name = string("max_value_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_value_keep_dims_0 = const()[name = string("max_value_keep_dims_0"), val = bool(false)];
+            tensor<fp16, [1, 1, 8]> max_value = reduce_max(axes = max_value_axes_0, keep_dims = max_value_keep_dims_0, x = values)[name = string("max_value")];
+            int32 max_value_index_axis_0 = const()[name = string("max_value_index_axis_0"), val = int32(1)];
+            bool max_value_index_keep_dims_0 = const()[name = string("max_value_index_keep_dims_0"), val = bool(true)];
+            string max_value_index_output_dtype_0 = const()[name = string("max_value_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 1, 1, 8]> max_value_index = reduce_argmax(axis = max_value_index_axis_0, keep_dims = max_value_index_keep_dims_0, output_dtype = max_value_index_output_dtype_0, x = values)[name = string("max_value_index")];
+            string indices_chunk_0_int32_dtype_1 = const()[name = string("indices_chunk_0_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_1_int32_dtype_1 = const()[name = string("indices_chunk_1_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_2_int32_dtype_1 = const()[name = string("indices_chunk_2_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_3_int32_dtype_1 = const()[name = string("indices_chunk_3_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_4_int32_dtype_1 = const()[name = string("indices_chunk_4_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_5_int32_dtype_1 = const()[name = string("indices_chunk_5_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_6_int32_dtype_1 = const()[name = string("indices_chunk_6_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_7_int32_dtype_1 = const()[name = string("indices_chunk_7_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_8_int32_dtype_1 = const()[name = string("indices_chunk_8_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_9_int32_dtype_1 = const()[name = string("indices_chunk_9_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_10_int32_dtype_1 = const()[name = string("indices_chunk_10_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_11_int32_dtype_1 = const()[name = string("indices_chunk_11_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_12_int32_dtype_1 = const()[name = string("indices_chunk_12_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_13_int32_dtype_1 = const()[name = string("indices_chunk_13_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_14_int32_dtype_1 = const()[name = string("indices_chunk_14_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_15_int32_dtype_1 = const()[name = string("indices_chunk_15_int32_dtype_1"), val = string("int32")];
+            int32 indices_axis_1 = const()[name = string("indices_axis_1"), val = int32(1)];
+            bool indices_interleave_1 = const()[name = string("indices_interleave_1"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_15_int32 = cast(dtype = indices_chunk_15_int32_dtype_1, x = argmax_chunk_15)[name = string("cast_32")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_14_int32 = cast(dtype = indices_chunk_14_int32_dtype_1, x = argmax_chunk_14)[name = string("cast_33")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_13_int32 = cast(dtype = indices_chunk_13_int32_dtype_1, x = argmax_chunk_13)[name = string("cast_34")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_12_int32 = cast(dtype = indices_chunk_12_int32_dtype_1, x = argmax_chunk_12)[name = string("cast_35")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_11_int32 = cast(dtype = indices_chunk_11_int32_dtype_1, x = argmax_chunk_11)[name = string("cast_36")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_10_int32 = cast(dtype = indices_chunk_10_int32_dtype_1, x = argmax_chunk_10)[name = string("cast_37")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_9_int32 = cast(dtype = indices_chunk_9_int32_dtype_1, x = argmax_chunk_9)[name = string("cast_38")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_8_int32 = cast(dtype = indices_chunk_8_int32_dtype_1, x = argmax_chunk_8)[name = string("cast_39")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_7_int32 = cast(dtype = indices_chunk_7_int32_dtype_1, x = argmax_chunk_7)[name = string("cast_40")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_6_int32 = cast(dtype = indices_chunk_6_int32_dtype_1, x = argmax_chunk_6)[name = string("cast_41")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_5_int32 = cast(dtype = indices_chunk_5_int32_dtype_1, x = argmax_chunk_5)[name = string("cast_42")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_4_int32 = cast(dtype = indices_chunk_4_int32_dtype_1, x = argmax_chunk_4)[name = string("cast_43")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_3_int32 = cast(dtype = indices_chunk_3_int32_dtype_1, x = argmax_chunk_3)[name = string("cast_44")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_2_int32 = cast(dtype = indices_chunk_2_int32_dtype_1, x = argmax_chunk_2)[name = string("cast_45")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_1_int32 = cast(dtype = indices_chunk_1_int32_dtype_1, x = argmax_chunk_1)[name = string("cast_46")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_0_int32 = cast(dtype = indices_chunk_0_int32_dtype_1, x = argmax_chunk_0)[name = string("cast_47")];
+            tensor<int32, [1, 16, 1, 8]> indices = concat(axis = indices_axis_1, interleave = indices_interleave_1, values = (indices_chunk_0_int32, indices_chunk_1_int32, indices_chunk_2_int32, indices_chunk_3_int32, indices_chunk_4_int32, indices_chunk_5_int32, indices_chunk_6_int32, indices_chunk_7_int32, indices_chunk_8_int32, indices_chunk_9_int32, indices_chunk_10_int32, indices_chunk_11_int32, indices_chunk_12_int32, indices_chunk_13_int32, indices_chunk_14_int32, indices_chunk_15_int32))[name = string("indices")];
+            int32 argmax_chunks_axis_1 = const()[name = string("argmax_chunks_axis_1"), val = int32(1)];
+            bool argmax_chunks_validate_indices_1 = const()[name = string("argmax_chunks_validate_indices_1"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 8]> argmax_chunks = gather_along_axis(axis = argmax_chunks_axis_1, indices = max_value_index, validate_indices = argmax_chunks_validate_indices_1, x = indices)[name = string("argmax_chunks")];
+            int32 mul_1_x_0 = const()[name = string("mul_1_x_0"), val = int32(2048)];
+            tensor<int32, [1, 1, 1, 8]> mul_1 = mul(x = mul_1_x_0, y = max_value_index)[name = string("mul_1")];
+            tensor<int32, [1, 1, 1, 8]> argmax = add(x = argmax_chunks, y = mul_1)[name = string("argmax")];
+        } -> (logits, argmax, max_value);
+    func main<ios18>(tensor<fp16, [1, 2048, 1, ?]> hidden_states) [FlexibleShapeInformation = tuple<tuple<string, dict<string, tensor<int32, [?]>>>, tuple<string, dict<string, dict<string, tensor<int32, [?]>>>>>((("DefaultShapes", {{"hidden_states", [1, 2048, 1, 1]}}), ("EnumeratedShapes", {{"00cbfba1", {{"hidden_states", [1, 2048, 1, 128]}}}, {"67738071", {{"hidden_states", [1, 2048, 1, 48]}}}, {"6f8a4ce5", {{"hidden_states", [1, 2048, 1, 8]}}}, {"744d043e", {{"hidden_states", [1, 2048, 1, 64]}}}, {"956807f6", {{"hidden_states", [1, 2048, 1, 16]}}}, {"ae87ff54", {{"hidden_states", [1, 2048, 1, 32]}}}, {"c695af0b", {{"hidden_states", [1, 2048, 1, 1]}}}})))] {
+            tensor<fp16, [1, 2048, 1, ?]> final_norm_rmsnorm_abs = abs(x = hidden_states)[name = string("final_norm_rmsnorm_abs")];
+            tensor<int32, [1]> final_norm_rmsnorm_maxval_axes_1 = const()[name = string("final_norm_rmsnorm_maxval_axes_1"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_maxval_keep_dims_1 = const()[name = string("final_norm_rmsnorm_maxval_keep_dims_1"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, ?]> final_norm_rmsnorm_maxval = reduce_max(axes = final_norm_rmsnorm_maxval_axes_1, keep_dims = final_norm_rmsnorm_maxval_keep_dims_1, x = final_norm_rmsnorm_abs)[name = string("final_norm_rmsnorm_maxval")];
+            fp16 final_norm_rmsnorm_maxval_clipped_alpha_1 = const()[name = string("final_norm_rmsnorm_maxval_clipped_alpha_1"), val = fp16(0x1p-14)];
+            fp16 final_norm_rmsnorm_maxval_clipped_beta_1 = const()[name = string("final_norm_rmsnorm_maxval_clipped_beta_1"), val = fp16(inf)];
+            tensor<fp16, [1, 1, 1, ?]> final_norm_rmsnorm_maxval_clipped = clip(alpha = final_norm_rmsnorm_maxval_clipped_alpha_1, beta = final_norm_rmsnorm_maxval_clipped_beta_1, x = final_norm_rmsnorm_maxval)[name = string("final_norm_rmsnorm_maxval_clipped")];
+            tensor<fp16, [1, 2048, 1, ?]> final_norm_rmsnorm_scaled = real_div(x = hidden_states, y = final_norm_rmsnorm_maxval_clipped)[name = string("final_norm_rmsnorm_scaled")];
+            tensor<int32, [1]> final_norm_rmsnorm_squared_sum_axes_1 = const()[name = string("final_norm_rmsnorm_squared_sum_axes_1"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_squared_sum_keep_dims_1 = const()[name = string("final_norm_rmsnorm_squared_sum_keep_dims_1"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, ?]> final_norm_rmsnorm_squared_sum = reduce_sum_square(axes = final_norm_rmsnorm_squared_sum_axes_1, keep_dims = final_norm_rmsnorm_squared_sum_keep_dims_1, x = final_norm_rmsnorm_scaled)[name = string("final_norm_rmsnorm_squared_sum")];
+            fp16 final_norm_rmsnorm_rsqrt_epsilon_1 = const()[name = string("final_norm_rmsnorm_rsqrt_epsilon_1"), val = fp16(0x1p-14)];
+            tensor<fp16, [1, 1, 1, ?]> final_norm_rmsnorm_rsqrt = rsqrt(epsilon = final_norm_rmsnorm_rsqrt_epsilon_1, x = final_norm_rmsnorm_squared_sum)[name = string("final_norm_rmsnorm_rsqrt")];
+            fp16 final_norm_rmsnorm_dim_scaled_y_1 = const()[name = string("final_norm_rmsnorm_dim_scaled_y_1"), val = fp16(0x1.6ap+5)];
+            tensor<fp16, [1, 2048, 1, ?]> final_norm_rmsnorm_dim_scaled = mul(x = final_norm_rmsnorm_scaled, y = final_norm_rmsnorm_dim_scaled_y_1)[name = string("final_norm_rmsnorm_dim_scaled")];
+            tensor<fp16, [1, 2048, 1, ?]> final_norm_rmsnorm_normalized = mul(x = final_norm_rmsnorm_dim_scaled, y = final_norm_rmsnorm_rsqrt)[name = string("final_norm_rmsnorm_normalized")];
+            tensor<fp16, [2048, 1, 1]> final_norm_rmsnorm_y_1 = const()[name = string("final_norm_rmsnorm_y_1"), val = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 2048, 1, ?]> final_norm_rmsnorm = mul(x = final_norm_rmsnorm_normalized, y = final_norm_rmsnorm_y_1)[name = string("final_norm_rmsnorm")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_0_weight_1 = const()[name = string("logits_chunk_0_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4224)))];
+            tensor<int32, [2]> logits_chunk_0_strides_1 = const()[name = string("logits_chunk_0_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_0_pad_type_1 = const()[name = string("logits_chunk_0_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_0_pad_1 = const()[name = string("logits_chunk_0_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_0_dilations_1 = const()[name = string("logits_chunk_0_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_0_groups_1 = const()[name = string("logits_chunk_0_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, ?]> logits_chunk_0 = conv(dilations = logits_chunk_0_dilations_1, groups = logits_chunk_0_groups_1, pad = logits_chunk_0_pad_1, pad_type = logits_chunk_0_pad_type_1, strides = logits_chunk_0_strides_1, weight = logits_chunk_0_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_0")];
+            int32 argmax_chunk_0_axis_0 = const()[name = string("argmax_chunk_0_axis_0"), val = int32(1)];
+            bool argmax_chunk_0_keep_dims_0 = const()[name = string("argmax_chunk_0_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_0_output_dtype_0 = const()[name = string("argmax_chunk_0_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, ?]> argmax_chunk_0 = reduce_argmax(axis = argmax_chunk_0_axis_0, keep_dims = argmax_chunk_0_keep_dims_0, output_dtype = argmax_chunk_0_output_dtype_0, x = logits_chunk_0)[name = string("argmax_chunk_0")];
+            tensor<int32, [1]> max_chunk_0_axes_0 = const()[name = string("max_chunk_0_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_0_keep_dims_0 = const()[name = string("max_chunk_0_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, ?]> max_chunk_0 = reduce_max(axes = max_chunk_0_axes_0, keep_dims = max_chunk_0_keep_dims_0, x = logits_chunk_0)[name = string("max_chunk_0")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_1_weight_1 = const()[name = string("logits_chunk_1_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8392896)))];
+            tensor<int32, [2]> logits_chunk_1_strides_1 = const()[name = string("logits_chunk_1_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_1_pad_type_1 = const()[name = string("logits_chunk_1_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_1_pad_1 = const()[name = string("logits_chunk_1_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_1_dilations_1 = const()[name = string("logits_chunk_1_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_1_groups_1 = const()[name = string("logits_chunk_1_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, ?]> logits_chunk_1 = conv(dilations = logits_chunk_1_dilations_1, groups = logits_chunk_1_groups_1, pad = logits_chunk_1_pad_1, pad_type = logits_chunk_1_pad_type_1, strides = logits_chunk_1_strides_1, weight = logits_chunk_1_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_1")];
+            int32 argmax_chunk_1_axis_0 = const()[name = string("argmax_chunk_1_axis_0"), val = int32(1)];
+            bool argmax_chunk_1_keep_dims_0 = const()[name = string("argmax_chunk_1_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_1_output_dtype_0 = const()[name = string("argmax_chunk_1_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, ?]> argmax_chunk_1 = reduce_argmax(axis = argmax_chunk_1_axis_0, keep_dims = argmax_chunk_1_keep_dims_0, output_dtype = argmax_chunk_1_output_dtype_0, x = logits_chunk_1)[name = string("argmax_chunk_1")];
+            tensor<int32, [1]> max_chunk_1_axes_0 = const()[name = string("max_chunk_1_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_1_keep_dims_0 = const()[name = string("max_chunk_1_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, ?]> max_chunk_1 = reduce_max(axes = max_chunk_1_axes_0, keep_dims = max_chunk_1_keep_dims_0, x = logits_chunk_1)[name = string("max_chunk_1")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_2_weight_1 = const()[name = string("logits_chunk_2_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16781568)))];
+            tensor<int32, [2]> logits_chunk_2_strides_1 = const()[name = string("logits_chunk_2_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_2_pad_type_1 = const()[name = string("logits_chunk_2_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_2_pad_1 = const()[name = string("logits_chunk_2_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_2_dilations_1 = const()[name = string("logits_chunk_2_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_2_groups_1 = const()[name = string("logits_chunk_2_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, ?]> logits_chunk_2 = conv(dilations = logits_chunk_2_dilations_1, groups = logits_chunk_2_groups_1, pad = logits_chunk_2_pad_1, pad_type = logits_chunk_2_pad_type_1, strides = logits_chunk_2_strides_1, weight = logits_chunk_2_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_2")];
+            int32 argmax_chunk_2_axis_0 = const()[name = string("argmax_chunk_2_axis_0"), val = int32(1)];
+            bool argmax_chunk_2_keep_dims_0 = const()[name = string("argmax_chunk_2_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_2_output_dtype_0 = const()[name = string("argmax_chunk_2_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, ?]> argmax_chunk_2 = reduce_argmax(axis = argmax_chunk_2_axis_0, keep_dims = argmax_chunk_2_keep_dims_0, output_dtype = argmax_chunk_2_output_dtype_0, x = logits_chunk_2)[name = string("argmax_chunk_2")];
+            tensor<int32, [1]> max_chunk_2_axes_0 = const()[name = string("max_chunk_2_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_2_keep_dims_0 = const()[name = string("max_chunk_2_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, ?]> max_chunk_2 = reduce_max(axes = max_chunk_2_axes_0, keep_dims = max_chunk_2_keep_dims_0, x = logits_chunk_2)[name = string("max_chunk_2")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_3_weight_1 = const()[name = string("logits_chunk_3_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25170240)))];
+            tensor<int32, [2]> logits_chunk_3_strides_1 = const()[name = string("logits_chunk_3_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_3_pad_type_1 = const()[name = string("logits_chunk_3_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_3_pad_1 = const()[name = string("logits_chunk_3_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_3_dilations_1 = const()[name = string("logits_chunk_3_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_3_groups_1 = const()[name = string("logits_chunk_3_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, ?]> logits_chunk_3 = conv(dilations = logits_chunk_3_dilations_1, groups = logits_chunk_3_groups_1, pad = logits_chunk_3_pad_1, pad_type = logits_chunk_3_pad_type_1, strides = logits_chunk_3_strides_1, weight = logits_chunk_3_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_3")];
+            int32 argmax_chunk_3_axis_0 = const()[name = string("argmax_chunk_3_axis_0"), val = int32(1)];
+            bool argmax_chunk_3_keep_dims_0 = const()[name = string("argmax_chunk_3_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_3_output_dtype_0 = const()[name = string("argmax_chunk_3_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, ?]> argmax_chunk_3 = reduce_argmax(axis = argmax_chunk_3_axis_0, keep_dims = argmax_chunk_3_keep_dims_0, output_dtype = argmax_chunk_3_output_dtype_0, x = logits_chunk_3)[name = string("argmax_chunk_3")];
+            tensor<int32, [1]> max_chunk_3_axes_0 = const()[name = string("max_chunk_3_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_3_keep_dims_0 = const()[name = string("max_chunk_3_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, ?]> max_chunk_3 = reduce_max(axes = max_chunk_3_axes_0, keep_dims = max_chunk_3_keep_dims_0, x = logits_chunk_3)[name = string("max_chunk_3")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_4_weight_1 = const()[name = string("logits_chunk_4_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33558912)))];
+            tensor<int32, [2]> logits_chunk_4_strides_1 = const()[name = string("logits_chunk_4_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_4_pad_type_1 = const()[name = string("logits_chunk_4_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_4_pad_1 = const()[name = string("logits_chunk_4_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_4_dilations_1 = const()[name = string("logits_chunk_4_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_4_groups_1 = const()[name = string("logits_chunk_4_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, ?]> logits_chunk_4 = conv(dilations = logits_chunk_4_dilations_1, groups = logits_chunk_4_groups_1, pad = logits_chunk_4_pad_1, pad_type = logits_chunk_4_pad_type_1, strides = logits_chunk_4_strides_1, weight = logits_chunk_4_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_4")];
+            int32 argmax_chunk_4_axis_0 = const()[name = string("argmax_chunk_4_axis_0"), val = int32(1)];
+            bool argmax_chunk_4_keep_dims_0 = const()[name = string("argmax_chunk_4_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_4_output_dtype_0 = const()[name = string("argmax_chunk_4_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, ?]> argmax_chunk_4 = reduce_argmax(axis = argmax_chunk_4_axis_0, keep_dims = argmax_chunk_4_keep_dims_0, output_dtype = argmax_chunk_4_output_dtype_0, x = logits_chunk_4)[name = string("argmax_chunk_4")];
+            tensor<int32, [1]> max_chunk_4_axes_0 = const()[name = string("max_chunk_4_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_4_keep_dims_0 = const()[name = string("max_chunk_4_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, ?]> max_chunk_4 = reduce_max(axes = max_chunk_4_axes_0, keep_dims = max_chunk_4_keep_dims_0, x = logits_chunk_4)[name = string("max_chunk_4")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_5_weight_1 = const()[name = string("logits_chunk_5_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41947584)))];
+            tensor<int32, [2]> logits_chunk_5_strides_1 = const()[name = string("logits_chunk_5_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_5_pad_type_1 = const()[name = string("logits_chunk_5_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_5_pad_1 = const()[name = string("logits_chunk_5_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_5_dilations_1 = const()[name = string("logits_chunk_5_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_5_groups_1 = const()[name = string("logits_chunk_5_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, ?]> logits_chunk_5 = conv(dilations = logits_chunk_5_dilations_1, groups = logits_chunk_5_groups_1, pad = logits_chunk_5_pad_1, pad_type = logits_chunk_5_pad_type_1, strides = logits_chunk_5_strides_1, weight = logits_chunk_5_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_5")];
+            int32 argmax_chunk_5_axis_0 = const()[name = string("argmax_chunk_5_axis_0"), val = int32(1)];
+            bool argmax_chunk_5_keep_dims_0 = const()[name = string("argmax_chunk_5_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_5_output_dtype_0 = const()[name = string("argmax_chunk_5_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, ?]> argmax_chunk_5 = reduce_argmax(axis = argmax_chunk_5_axis_0, keep_dims = argmax_chunk_5_keep_dims_0, output_dtype = argmax_chunk_5_output_dtype_0, x = logits_chunk_5)[name = string("argmax_chunk_5")];
+            tensor<int32, [1]> max_chunk_5_axes_0 = const()[name = string("max_chunk_5_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_5_keep_dims_0 = const()[name = string("max_chunk_5_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, ?]> max_chunk_5 = reduce_max(axes = max_chunk_5_axes_0, keep_dims = max_chunk_5_keep_dims_0, x = logits_chunk_5)[name = string("max_chunk_5")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_6_weight_1 = const()[name = string("logits_chunk_6_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50336256)))];
+            tensor<int32, [2]> logits_chunk_6_strides_1 = const()[name = string("logits_chunk_6_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_6_pad_type_1 = const()[name = string("logits_chunk_6_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_6_pad_1 = const()[name = string("logits_chunk_6_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_6_dilations_1 = const()[name = string("logits_chunk_6_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_6_groups_1 = const()[name = string("logits_chunk_6_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, ?]> logits_chunk_6 = conv(dilations = logits_chunk_6_dilations_1, groups = logits_chunk_6_groups_1, pad = logits_chunk_6_pad_1, pad_type = logits_chunk_6_pad_type_1, strides = logits_chunk_6_strides_1, weight = logits_chunk_6_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_6")];
+            int32 argmax_chunk_6_axis_0 = const()[name = string("argmax_chunk_6_axis_0"), val = int32(1)];
+            bool argmax_chunk_6_keep_dims_0 = const()[name = string("argmax_chunk_6_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_6_output_dtype_0 = const()[name = string("argmax_chunk_6_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, ?]> argmax_chunk_6 = reduce_argmax(axis = argmax_chunk_6_axis_0, keep_dims = argmax_chunk_6_keep_dims_0, output_dtype = argmax_chunk_6_output_dtype_0, x = logits_chunk_6)[name = string("argmax_chunk_6")];
+            tensor<int32, [1]> max_chunk_6_axes_0 = const()[name = string("max_chunk_6_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_6_keep_dims_0 = const()[name = string("max_chunk_6_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, ?]> max_chunk_6 = reduce_max(axes = max_chunk_6_axes_0, keep_dims = max_chunk_6_keep_dims_0, x = logits_chunk_6)[name = string("max_chunk_6")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_7_weight_1 = const()[name = string("logits_chunk_7_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58724928)))];
+            tensor<int32, [2]> logits_chunk_7_strides_1 = const()[name = string("logits_chunk_7_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_7_pad_type_1 = const()[name = string("logits_chunk_7_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_7_pad_1 = const()[name = string("logits_chunk_7_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_7_dilations_1 = const()[name = string("logits_chunk_7_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_7_groups_1 = const()[name = string("logits_chunk_7_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, ?]> logits_chunk_7 = conv(dilations = logits_chunk_7_dilations_1, groups = logits_chunk_7_groups_1, pad = logits_chunk_7_pad_1, pad_type = logits_chunk_7_pad_type_1, strides = logits_chunk_7_strides_1, weight = logits_chunk_7_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_7")];
+            int32 argmax_chunk_7_axis_0 = const()[name = string("argmax_chunk_7_axis_0"), val = int32(1)];
+            bool argmax_chunk_7_keep_dims_0 = const()[name = string("argmax_chunk_7_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_7_output_dtype_0 = const()[name = string("argmax_chunk_7_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, ?]> argmax_chunk_7 = reduce_argmax(axis = argmax_chunk_7_axis_0, keep_dims = argmax_chunk_7_keep_dims_0, output_dtype = argmax_chunk_7_output_dtype_0, x = logits_chunk_7)[name = string("argmax_chunk_7")];
+            tensor<int32, [1]> max_chunk_7_axes_0 = const()[name = string("max_chunk_7_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_7_keep_dims_0 = const()[name = string("max_chunk_7_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, ?]> max_chunk_7 = reduce_max(axes = max_chunk_7_axes_0, keep_dims = max_chunk_7_keep_dims_0, x = logits_chunk_7)[name = string("max_chunk_7")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_8_weight_1 = const()[name = string("logits_chunk_8_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67113600)))];
+            tensor<int32, [2]> logits_chunk_8_strides_1 = const()[name = string("logits_chunk_8_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_8_pad_type_1 = const()[name = string("logits_chunk_8_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_8_pad_1 = const()[name = string("logits_chunk_8_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_8_dilations_1 = const()[name = string("logits_chunk_8_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_8_groups_1 = const()[name = string("logits_chunk_8_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, ?]> logits_chunk_8 = conv(dilations = logits_chunk_8_dilations_1, groups = logits_chunk_8_groups_1, pad = logits_chunk_8_pad_1, pad_type = logits_chunk_8_pad_type_1, strides = logits_chunk_8_strides_1, weight = logits_chunk_8_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_8")];
+            int32 argmax_chunk_8_axis_0 = const()[name = string("argmax_chunk_8_axis_0"), val = int32(1)];
+            bool argmax_chunk_8_keep_dims_0 = const()[name = string("argmax_chunk_8_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_8_output_dtype_0 = const()[name = string("argmax_chunk_8_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, ?]> argmax_chunk_8 = reduce_argmax(axis = argmax_chunk_8_axis_0, keep_dims = argmax_chunk_8_keep_dims_0, output_dtype = argmax_chunk_8_output_dtype_0, x = logits_chunk_8)[name = string("argmax_chunk_8")];
+            tensor<int32, [1]> max_chunk_8_axes_0 = const()[name = string("max_chunk_8_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_8_keep_dims_0 = const()[name = string("max_chunk_8_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, ?]> max_chunk_8 = reduce_max(axes = max_chunk_8_axes_0, keep_dims = max_chunk_8_keep_dims_0, x = logits_chunk_8)[name = string("max_chunk_8")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_9_weight_1 = const()[name = string("logits_chunk_9_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75502272)))];
+            tensor<int32, [2]> logits_chunk_9_strides_1 = const()[name = string("logits_chunk_9_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_9_pad_type_1 = const()[name = string("logits_chunk_9_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_9_pad_1 = const()[name = string("logits_chunk_9_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_9_dilations_1 = const()[name = string("logits_chunk_9_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_9_groups_1 = const()[name = string("logits_chunk_9_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, ?]> logits_chunk_9 = conv(dilations = logits_chunk_9_dilations_1, groups = logits_chunk_9_groups_1, pad = logits_chunk_9_pad_1, pad_type = logits_chunk_9_pad_type_1, strides = logits_chunk_9_strides_1, weight = logits_chunk_9_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_9")];
+            int32 argmax_chunk_9_axis_0 = const()[name = string("argmax_chunk_9_axis_0"), val = int32(1)];
+            bool argmax_chunk_9_keep_dims_0 = const()[name = string("argmax_chunk_9_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_9_output_dtype_0 = const()[name = string("argmax_chunk_9_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, ?]> argmax_chunk_9 = reduce_argmax(axis = argmax_chunk_9_axis_0, keep_dims = argmax_chunk_9_keep_dims_0, output_dtype = argmax_chunk_9_output_dtype_0, x = logits_chunk_9)[name = string("argmax_chunk_9")];
+            tensor<int32, [1]> max_chunk_9_axes_0 = const()[name = string("max_chunk_9_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_9_keep_dims_0 = const()[name = string("max_chunk_9_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, ?]> max_chunk_9 = reduce_max(axes = max_chunk_9_axes_0, keep_dims = max_chunk_9_keep_dims_0, x = logits_chunk_9)[name = string("max_chunk_9")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_10_weight_1 = const()[name = string("logits_chunk_10_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83890944)))];
+            tensor<int32, [2]> logits_chunk_10_strides_1 = const()[name = string("logits_chunk_10_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_10_pad_type_1 = const()[name = string("logits_chunk_10_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_10_pad_1 = const()[name = string("logits_chunk_10_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_10_dilations_1 = const()[name = string("logits_chunk_10_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_10_groups_1 = const()[name = string("logits_chunk_10_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, ?]> logits_chunk_10 = conv(dilations = logits_chunk_10_dilations_1, groups = logits_chunk_10_groups_1, pad = logits_chunk_10_pad_1, pad_type = logits_chunk_10_pad_type_1, strides = logits_chunk_10_strides_1, weight = logits_chunk_10_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_10")];
+            int32 argmax_chunk_10_axis_0 = const()[name = string("argmax_chunk_10_axis_0"), val = int32(1)];
+            bool argmax_chunk_10_keep_dims_0 = const()[name = string("argmax_chunk_10_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_10_output_dtype_0 = const()[name = string("argmax_chunk_10_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, ?]> argmax_chunk_10 = reduce_argmax(axis = argmax_chunk_10_axis_0, keep_dims = argmax_chunk_10_keep_dims_0, output_dtype = argmax_chunk_10_output_dtype_0, x = logits_chunk_10)[name = string("argmax_chunk_10")];
+            tensor<int32, [1]> max_chunk_10_axes_0 = const()[name = string("max_chunk_10_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_10_keep_dims_0 = const()[name = string("max_chunk_10_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, ?]> max_chunk_10 = reduce_max(axes = max_chunk_10_axes_0, keep_dims = max_chunk_10_keep_dims_0, x = logits_chunk_10)[name = string("max_chunk_10")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_11_weight_1 = const()[name = string("logits_chunk_11_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92279616)))];
+            tensor<int32, [2]> logits_chunk_11_strides_1 = const()[name = string("logits_chunk_11_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_11_pad_type_1 = const()[name = string("logits_chunk_11_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_11_pad_1 = const()[name = string("logits_chunk_11_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_11_dilations_1 = const()[name = string("logits_chunk_11_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_11_groups_1 = const()[name = string("logits_chunk_11_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, ?]> logits_chunk_11 = conv(dilations = logits_chunk_11_dilations_1, groups = logits_chunk_11_groups_1, pad = logits_chunk_11_pad_1, pad_type = logits_chunk_11_pad_type_1, strides = logits_chunk_11_strides_1, weight = logits_chunk_11_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_11")];
+            int32 argmax_chunk_11_axis_0 = const()[name = string("argmax_chunk_11_axis_0"), val = int32(1)];
+            bool argmax_chunk_11_keep_dims_0 = const()[name = string("argmax_chunk_11_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_11_output_dtype_0 = const()[name = string("argmax_chunk_11_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, ?]> argmax_chunk_11 = reduce_argmax(axis = argmax_chunk_11_axis_0, keep_dims = argmax_chunk_11_keep_dims_0, output_dtype = argmax_chunk_11_output_dtype_0, x = logits_chunk_11)[name = string("argmax_chunk_11")];
+            tensor<int32, [1]> max_chunk_11_axes_0 = const()[name = string("max_chunk_11_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_11_keep_dims_0 = const()[name = string("max_chunk_11_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, ?]> max_chunk_11 = reduce_max(axes = max_chunk_11_axes_0, keep_dims = max_chunk_11_keep_dims_0, x = logits_chunk_11)[name = string("max_chunk_11")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_12_weight_1 = const()[name = string("logits_chunk_12_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100668288)))];
+            tensor<int32, [2]> logits_chunk_12_strides_1 = const()[name = string("logits_chunk_12_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_12_pad_type_1 = const()[name = string("logits_chunk_12_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_12_pad_1 = const()[name = string("logits_chunk_12_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_12_dilations_1 = const()[name = string("logits_chunk_12_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_12_groups_1 = const()[name = string("logits_chunk_12_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, ?]> logits_chunk_12 = conv(dilations = logits_chunk_12_dilations_1, groups = logits_chunk_12_groups_1, pad = logits_chunk_12_pad_1, pad_type = logits_chunk_12_pad_type_1, strides = logits_chunk_12_strides_1, weight = logits_chunk_12_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_12")];
+            int32 argmax_chunk_12_axis_0 = const()[name = string("argmax_chunk_12_axis_0"), val = int32(1)];
+            bool argmax_chunk_12_keep_dims_0 = const()[name = string("argmax_chunk_12_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_12_output_dtype_0 = const()[name = string("argmax_chunk_12_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, ?]> argmax_chunk_12 = reduce_argmax(axis = argmax_chunk_12_axis_0, keep_dims = argmax_chunk_12_keep_dims_0, output_dtype = argmax_chunk_12_output_dtype_0, x = logits_chunk_12)[name = string("argmax_chunk_12")];
+            tensor<int32, [1]> max_chunk_12_axes_0 = const()[name = string("max_chunk_12_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_12_keep_dims_0 = const()[name = string("max_chunk_12_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, ?]> max_chunk_12 = reduce_max(axes = max_chunk_12_axes_0, keep_dims = max_chunk_12_keep_dims_0, x = logits_chunk_12)[name = string("max_chunk_12")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_13_weight_1 = const()[name = string("logits_chunk_13_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109056960)))];
+            tensor<int32, [2]> logits_chunk_13_strides_1 = const()[name = string("logits_chunk_13_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_13_pad_type_1 = const()[name = string("logits_chunk_13_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_13_pad_1 = const()[name = string("logits_chunk_13_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_13_dilations_1 = const()[name = string("logits_chunk_13_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_13_groups_1 = const()[name = string("logits_chunk_13_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, ?]> logits_chunk_13 = conv(dilations = logits_chunk_13_dilations_1, groups = logits_chunk_13_groups_1, pad = logits_chunk_13_pad_1, pad_type = logits_chunk_13_pad_type_1, strides = logits_chunk_13_strides_1, weight = logits_chunk_13_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_13")];
+            int32 argmax_chunk_13_axis_0 = const()[name = string("argmax_chunk_13_axis_0"), val = int32(1)];
+            bool argmax_chunk_13_keep_dims_0 = const()[name = string("argmax_chunk_13_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_13_output_dtype_0 = const()[name = string("argmax_chunk_13_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, ?]> argmax_chunk_13 = reduce_argmax(axis = argmax_chunk_13_axis_0, keep_dims = argmax_chunk_13_keep_dims_0, output_dtype = argmax_chunk_13_output_dtype_0, x = logits_chunk_13)[name = string("argmax_chunk_13")];
+            tensor<int32, [1]> max_chunk_13_axes_0 = const()[name = string("max_chunk_13_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_13_keep_dims_0 = const()[name = string("max_chunk_13_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, ?]> max_chunk_13 = reduce_max(axes = max_chunk_13_axes_0, keep_dims = max_chunk_13_keep_dims_0, x = logits_chunk_13)[name = string("max_chunk_13")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_14_weight_1 = const()[name = string("logits_chunk_14_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117445632)))];
+            tensor<int32, [2]> logits_chunk_14_strides_1 = const()[name = string("logits_chunk_14_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_14_pad_type_1 = const()[name = string("logits_chunk_14_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_14_pad_1 = const()[name = string("logits_chunk_14_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_14_dilations_1 = const()[name = string("logits_chunk_14_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_14_groups_1 = const()[name = string("logits_chunk_14_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, ?]> logits_chunk_14 = conv(dilations = logits_chunk_14_dilations_1, groups = logits_chunk_14_groups_1, pad = logits_chunk_14_pad_1, pad_type = logits_chunk_14_pad_type_1, strides = logits_chunk_14_strides_1, weight = logits_chunk_14_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_14")];
+            int32 argmax_chunk_14_axis_0 = const()[name = string("argmax_chunk_14_axis_0"), val = int32(1)];
+            bool argmax_chunk_14_keep_dims_0 = const()[name = string("argmax_chunk_14_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_14_output_dtype_0 = const()[name = string("argmax_chunk_14_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, ?]> argmax_chunk_14 = reduce_argmax(axis = argmax_chunk_14_axis_0, keep_dims = argmax_chunk_14_keep_dims_0, output_dtype = argmax_chunk_14_output_dtype_0, x = logits_chunk_14)[name = string("argmax_chunk_14")];
+            tensor<int32, [1]> max_chunk_14_axes_0 = const()[name = string("max_chunk_14_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_14_keep_dims_0 = const()[name = string("max_chunk_14_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, ?]> max_chunk_14 = reduce_max(axes = max_chunk_14_axes_0, keep_dims = max_chunk_14_keep_dims_0, x = logits_chunk_14)[name = string("max_chunk_14")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_15_weight_1 = const()[name = string("logits_chunk_15_weight_1"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125834304)))];
+            tensor<int32, [2]> logits_chunk_15_strides_1 = const()[name = string("logits_chunk_15_strides_1"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_15_pad_type_1 = const()[name = string("logits_chunk_15_pad_type_1"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_15_pad_1 = const()[name = string("logits_chunk_15_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_15_dilations_1 = const()[name = string("logits_chunk_15_dilations_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_15_groups_1 = const()[name = string("logits_chunk_15_groups_1"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, ?]> logits_chunk_15 = conv(dilations = logits_chunk_15_dilations_1, groups = logits_chunk_15_groups_1, pad = logits_chunk_15_pad_1, pad_type = logits_chunk_15_pad_type_1, strides = logits_chunk_15_strides_1, weight = logits_chunk_15_weight_1, x = final_norm_rmsnorm)[name = string("logits_chunk_15")];
+            int32 argmax_chunk_15_axis_0 = const()[name = string("argmax_chunk_15_axis_0"), val = int32(1)];
+            bool argmax_chunk_15_keep_dims_0 = const()[name = string("argmax_chunk_15_keep_dims_0"), val = bool(true)];
+            string argmax_chunk_15_output_dtype_0 = const()[name = string("argmax_chunk_15_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, ?]> argmax_chunk_15 = reduce_argmax(axis = argmax_chunk_15_axis_0, keep_dims = argmax_chunk_15_keep_dims_0, output_dtype = argmax_chunk_15_output_dtype_0, x = logits_chunk_15)[name = string("argmax_chunk_15")];
+            tensor<int32, [1]> max_chunk_15_axes_0 = const()[name = string("max_chunk_15_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_chunk_15_keep_dims_0 = const()[name = string("max_chunk_15_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, ?]> max_chunk_15 = reduce_max(axes = max_chunk_15_axes_0, keep_dims = max_chunk_15_keep_dims_0, x = logits_chunk_15)[name = string("max_chunk_15")];
+            int32 logits_axis_0 = const()[name = string("logits_axis_0"), val = int32(1)];
+            bool logits_interleave_0 = const()[name = string("logits_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 32768, 1, ?]> logits = concat(axis = logits_axis_0, interleave = logits_interleave_0, values = (logits_chunk_0, logits_chunk_1, logits_chunk_2, logits_chunk_3, logits_chunk_4, logits_chunk_5, logits_chunk_6, logits_chunk_7, logits_chunk_8, logits_chunk_9, logits_chunk_10, logits_chunk_11, logits_chunk_12, logits_chunk_13, logits_chunk_14, logits_chunk_15))[name = string("logits")];
+            int32 values_axis_0 = const()[name = string("values_axis_0"), val = int32(1)];
+            bool values_interleave_0 = const()[name = string("values_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, ?]> values = concat(axis = values_axis_0, interleave = values_interleave_0, values = (max_chunk_0, max_chunk_1, max_chunk_2, max_chunk_3, max_chunk_4, max_chunk_5, max_chunk_6, max_chunk_7, max_chunk_8, max_chunk_9, max_chunk_10, max_chunk_11, max_chunk_12, max_chunk_13, max_chunk_14, max_chunk_15))[name = string("values")];
+            tensor<int32, [1]> max_value_axes_0 = const()[name = string("max_value_axes_0"), val = tensor<int32, [1]>([1])];
+            bool max_value_keep_dims_0 = const()[name = string("max_value_keep_dims_0"), val = bool(false)];
+            tensor<fp16, [1, 1, ?]> max_value = reduce_max(axes = max_value_axes_0, keep_dims = max_value_keep_dims_0, x = values)[name = string("max_value")];
+            int32 max_value_index_axis_0 = const()[name = string("max_value_index_axis_0"), val = int32(1)];
+            bool max_value_index_keep_dims_0 = const()[name = string("max_value_index_keep_dims_0"), val = bool(true)];
+            string max_value_index_output_dtype_0 = const()[name = string("max_value_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 1, 1, ?]> max_value_index = reduce_argmax(axis = max_value_index_axis_0, keep_dims = max_value_index_keep_dims_0, output_dtype = max_value_index_output_dtype_0, x = values)[name = string("max_value_index")];
+            string indices_chunk_0_int32_dtype_1 = const()[name = string("indices_chunk_0_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_1_int32_dtype_1 = const()[name = string("indices_chunk_1_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_2_int32_dtype_1 = const()[name = string("indices_chunk_2_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_3_int32_dtype_1 = const()[name = string("indices_chunk_3_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_4_int32_dtype_1 = const()[name = string("indices_chunk_4_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_5_int32_dtype_1 = const()[name = string("indices_chunk_5_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_6_int32_dtype_1 = const()[name = string("indices_chunk_6_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_7_int32_dtype_1 = const()[name = string("indices_chunk_7_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_8_int32_dtype_1 = const()[name = string("indices_chunk_8_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_9_int32_dtype_1 = const()[name = string("indices_chunk_9_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_10_int32_dtype_1 = const()[name = string("indices_chunk_10_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_11_int32_dtype_1 = const()[name = string("indices_chunk_11_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_12_int32_dtype_1 = const()[name = string("indices_chunk_12_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_13_int32_dtype_1 = const()[name = string("indices_chunk_13_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_14_int32_dtype_1 = const()[name = string("indices_chunk_14_int32_dtype_1"), val = string("int32")];
+            string indices_chunk_15_int32_dtype_1 = const()[name = string("indices_chunk_15_int32_dtype_1"), val = string("int32")];
+            int32 indices_axis_1 = const()[name = string("indices_axis_1"), val = int32(1)];
+            bool indices_interleave_1 = const()[name = string("indices_interleave_1"), val = bool(false)];
+            tensor<int32, [1, 1, 1, ?]> indices_chunk_15_int32 = cast(dtype = indices_chunk_15_int32_dtype_1, x = argmax_chunk_15)[name = string("cast_0")];
+            tensor<int32, [1, 1, 1, ?]> indices_chunk_14_int32 = cast(dtype = indices_chunk_14_int32_dtype_1, x = argmax_chunk_14)[name = string("cast_1")];
+            tensor<int32, [1, 1, 1, ?]> indices_chunk_13_int32 = cast(dtype = indices_chunk_13_int32_dtype_1, x = argmax_chunk_13)[name = string("cast_2")];
+            tensor<int32, [1, 1, 1, ?]> indices_chunk_12_int32 = cast(dtype = indices_chunk_12_int32_dtype_1, x = argmax_chunk_12)[name = string("cast_3")];
+            tensor<int32, [1, 1, 1, ?]> indices_chunk_11_int32 = cast(dtype = indices_chunk_11_int32_dtype_1, x = argmax_chunk_11)[name = string("cast_4")];
+            tensor<int32, [1, 1, 1, ?]> indices_chunk_10_int32 = cast(dtype = indices_chunk_10_int32_dtype_1, x = argmax_chunk_10)[name = string("cast_5")];
+            tensor<int32, [1, 1, 1, ?]> indices_chunk_9_int32 = cast(dtype = indices_chunk_9_int32_dtype_1, x = argmax_chunk_9)[name = string("cast_6")];
+            tensor<int32, [1, 1, 1, ?]> indices_chunk_8_int32 = cast(dtype = indices_chunk_8_int32_dtype_1, x = argmax_chunk_8)[name = string("cast_7")];
+            tensor<int32, [1, 1, 1, ?]> indices_chunk_7_int32 = cast(dtype = indices_chunk_7_int32_dtype_1, x = argmax_chunk_7)[name = string("cast_8")];
+            tensor<int32, [1, 1, 1, ?]> indices_chunk_6_int32 = cast(dtype = indices_chunk_6_int32_dtype_1, x = argmax_chunk_6)[name = string("cast_9")];
+            tensor<int32, [1, 1, 1, ?]> indices_chunk_5_int32 = cast(dtype = indices_chunk_5_int32_dtype_1, x = argmax_chunk_5)[name = string("cast_10")];
+            tensor<int32, [1, 1, 1, ?]> indices_chunk_4_int32 = cast(dtype = indices_chunk_4_int32_dtype_1, x = argmax_chunk_4)[name = string("cast_11")];
+            tensor<int32, [1, 1, 1, ?]> indices_chunk_3_int32 = cast(dtype = indices_chunk_3_int32_dtype_1, x = argmax_chunk_3)[name = string("cast_12")];
+            tensor<int32, [1, 1, 1, ?]> indices_chunk_2_int32 = cast(dtype = indices_chunk_2_int32_dtype_1, x = argmax_chunk_2)[name = string("cast_13")];
+            tensor<int32, [1, 1, 1, ?]> indices_chunk_1_int32 = cast(dtype = indices_chunk_1_int32_dtype_1, x = argmax_chunk_1)[name = string("cast_14")];
+            tensor<int32, [1, 1, 1, ?]> indices_chunk_0_int32 = cast(dtype = indices_chunk_0_int32_dtype_1, x = argmax_chunk_0)[name = string("cast_15")];
+            tensor<int32, [1, 16, 1, ?]> indices = concat(axis = indices_axis_1, interleave = indices_interleave_1, values = (indices_chunk_0_int32, indices_chunk_1_int32, indices_chunk_2_int32, indices_chunk_3_int32, indices_chunk_4_int32, indices_chunk_5_int32, indices_chunk_6_int32, indices_chunk_7_int32, indices_chunk_8_int32, indices_chunk_9_int32, indices_chunk_10_int32, indices_chunk_11_int32, indices_chunk_12_int32, indices_chunk_13_int32, indices_chunk_14_int32, indices_chunk_15_int32))[name = string("indices")];
+            int32 argmax_chunks_axis_1 = const()[name = string("argmax_chunks_axis_1"), val = int32(1)];
+            bool argmax_chunks_validate_indices_1 = const()[name = string("argmax_chunks_validate_indices_1"), val = bool(false)];
+            tensor<int32, [1, 1, 1, ?]> argmax_chunks = gather_along_axis(axis = argmax_chunks_axis_1, indices = max_value_index, validate_indices = argmax_chunks_validate_indices_1, x = indices)[name = string("argmax_chunks")];
+            int32 mul_1_x_0 = const()[name = string("mul_1_x_0"), val = int32(2048)];
+            tensor<int32, [1, 1, 1, ?]> mul_1 = mul(x = mul_1_x_0, y = max_value_index)[name = string("mul_1")];
+            tensor<int32, [1, 1, 1, ?]> argmax = add(x = argmax_chunks, y = mul_1)[name = string("argmax")];
+        } -> (logits, argmax, max_value);
+    func min_p_length_1<ios18>(tensor<fp16, [1, 2048, 1, 1]> hidden_states, tensor<fp16, [1]> p, tensor<fp32, [1]> random_number, tensor<fp16, [1]> temp) {
+            tensor<fp16, [1, 2048, 1, 1]> final_norm_rmsnorm_abs = abs(x = hidden_states)[name = string("final_norm_rmsnorm_abs")];
+            tensor<int32, [1]> final_norm_rmsnorm_maxval_axes_0 = const()[name = string("final_norm_rmsnorm_maxval_axes_0"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_maxval_keep_dims_0 = const()[name = string("final_norm_rmsnorm_maxval_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> final_norm_rmsnorm_maxval = reduce_max(axes = final_norm_rmsnorm_maxval_axes_0, keep_dims = final_norm_rmsnorm_maxval_keep_dims_0, x = final_norm_rmsnorm_abs)[name = string("final_norm_rmsnorm_maxval")];
+            fp16 final_norm_rmsnorm_maxval_clipped_alpha_0 = const()[name = string("final_norm_rmsnorm_maxval_clipped_alpha_0"), val = fp16(0x1p-14)];
+            fp16 final_norm_rmsnorm_maxval_clipped_beta_0 = const()[name = string("final_norm_rmsnorm_maxval_clipped_beta_0"), val = fp16(inf)];
+            tensor<fp16, [1, 1, 1, 1]> final_norm_rmsnorm_maxval_clipped = clip(alpha = final_norm_rmsnorm_maxval_clipped_alpha_0, beta = final_norm_rmsnorm_maxval_clipped_beta_0, x = final_norm_rmsnorm_maxval)[name = string("final_norm_rmsnorm_maxval_clipped")];
+            tensor<fp16, [1, 2048, 1, 1]> final_norm_rmsnorm_scaled = real_div(x = hidden_states, y = final_norm_rmsnorm_maxval_clipped)[name = string("final_norm_rmsnorm_scaled")];
+            tensor<int32, [1]> final_norm_rmsnorm_squared_sum_axes_0 = const()[name = string("final_norm_rmsnorm_squared_sum_axes_0"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_squared_sum_keep_dims_0 = const()[name = string("final_norm_rmsnorm_squared_sum_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> final_norm_rmsnorm_squared_sum = reduce_sum_square(axes = final_norm_rmsnorm_squared_sum_axes_0, keep_dims = final_norm_rmsnorm_squared_sum_keep_dims_0, x = final_norm_rmsnorm_scaled)[name = string("final_norm_rmsnorm_squared_sum")];
+            fp16 final_norm_rmsnorm_rsqrt_epsilon_0 = const()[name = string("final_norm_rmsnorm_rsqrt_epsilon_0"), val = fp16(0x1p-14)];
+            tensor<fp16, [1, 1, 1, 1]> final_norm_rmsnorm_rsqrt = rsqrt(epsilon = final_norm_rmsnorm_rsqrt_epsilon_0, x = final_norm_rmsnorm_squared_sum)[name = string("final_norm_rmsnorm_rsqrt")];
+            fp16 final_norm_rmsnorm_dim_scaled_y_0 = const()[name = string("final_norm_rmsnorm_dim_scaled_y_0"), val = fp16(0x1.6ap+5)];
+            tensor<fp16, [1, 2048, 1, 1]> final_norm_rmsnorm_dim_scaled = mul(x = final_norm_rmsnorm_scaled, y = final_norm_rmsnorm_dim_scaled_y_0)[name = string("final_norm_rmsnorm_dim_scaled")];
+            tensor<fp16, [1, 2048, 1, 1]> final_norm_rmsnorm_normalized = mul(x = final_norm_rmsnorm_dim_scaled, y = final_norm_rmsnorm_rsqrt)[name = string("final_norm_rmsnorm_normalized")];
+            tensor<fp16, [2048, 1, 1]> final_norm_rmsnorm_y_0 = const()[name = string("final_norm_rmsnorm_y_0"), val = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 2048, 1, 1]> final_norm_rmsnorm = mul(x = final_norm_rmsnorm_normalized, y = final_norm_rmsnorm_y_0)[name = string("final_norm_rmsnorm")];
+            fp16 temp_inverse_epsilon_0 = const()[name = string("temp_inverse_epsilon_0"), val = fp16(0x0p+0)];
+            tensor<fp16, [1]> temp_inverse = inverse(epsilon = temp_inverse_epsilon_0, x = temp)[name = string("temp_inverse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_0_weight_0 = const()[name = string("logits_chunk_0_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4224)))];
+            tensor<int32, [2]> logits_chunk_0_strides_0 = const()[name = string("logits_chunk_0_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_0_pad_type_0 = const()[name = string("logits_chunk_0_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_0_pad_0 = const()[name = string("logits_chunk_0_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_0_dilations_0 = const()[name = string("logits_chunk_0_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_0_groups_0 = const()[name = string("logits_chunk_0_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_0 = conv(dilations = logits_chunk_0_dilations_0, groups = logits_chunk_0_groups_0, pad = logits_chunk_0_pad_0, pad_type = logits_chunk_0_pad_type_0, strides = logits_chunk_0_strides_0, weight = logits_chunk_0_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_0")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_0_mul = mul(x = logits_chunk_0, y = temp_inverse)[name = string("logits_chunk_0_mul")];
+            tensor<int32, [1]> logits_chunk_0_max_axes_0 = const()[name = string("logits_chunk_0_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_0_max_keep_dims_0 = const()[name = string("logits_chunk_0_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_0_max = reduce_max(axes = logits_chunk_0_max_axes_0, keep_dims = logits_chunk_0_max_keep_dims_0, x = logits_chunk_0_mul)[name = string("logits_chunk_0_max")];
+            int32 logits_chunk_0_argmax_axis_0 = const()[name = string("logits_chunk_0_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_0_argmax_keep_dims_0 = const()[name = string("logits_chunk_0_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_0_argmax_output_dtype_0 = const()[name = string("logits_chunk_0_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> logits_chunk_0_argmax = reduce_argmax(axis = logits_chunk_0_argmax_axis_0, keep_dims = logits_chunk_0_argmax_keep_dims_0, output_dtype = logits_chunk_0_argmax_output_dtype_0, x = logits_chunk_0_mul)[name = string("logits_chunk_0_argmax")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_0_sub = sub(x = logits_chunk_0_mul, y = logits_chunk_0_max)[name = string("logits_chunk_0_sub")];
+            tensor<int32, [1]> logits_chunk_0_lse_sub_axes_0 = const()[name = string("logits_chunk_0_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_0_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_0_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_0_lse_sub = reduce_log_sum_exp(axes = logits_chunk_0_lse_sub_axes_0, keep_dims = logits_chunk_0_lse_sub_keep_dims_0, x = logits_chunk_0_sub)[name = string("logits_chunk_0_lse_sub")];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_0_lse = add(x = logits_chunk_0_lse_sub, y = logits_chunk_0_max)[name = string("logits_chunk_0_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_1_weight_0 = const()[name = string("logits_chunk_1_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8392896)))];
+            tensor<int32, [2]> logits_chunk_1_strides_0 = const()[name = string("logits_chunk_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_1_pad_type_0 = const()[name = string("logits_chunk_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_1_pad_0 = const()[name = string("logits_chunk_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_1_dilations_0 = const()[name = string("logits_chunk_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_1_groups_0 = const()[name = string("logits_chunk_1_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_1 = conv(dilations = logits_chunk_1_dilations_0, groups = logits_chunk_1_groups_0, pad = logits_chunk_1_pad_0, pad_type = logits_chunk_1_pad_type_0, strides = logits_chunk_1_strides_0, weight = logits_chunk_1_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_1")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_1_mul = mul(x = logits_chunk_1, y = temp_inverse)[name = string("logits_chunk_1_mul")];
+            tensor<int32, [1]> logits_chunk_1_max_axes_0 = const()[name = string("logits_chunk_1_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_1_max_keep_dims_0 = const()[name = string("logits_chunk_1_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_1_max = reduce_max(axes = logits_chunk_1_max_axes_0, keep_dims = logits_chunk_1_max_keep_dims_0, x = logits_chunk_1_mul)[name = string("logits_chunk_1_max")];
+            int32 logits_chunk_1_argmax_axis_0 = const()[name = string("logits_chunk_1_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_1_argmax_keep_dims_0 = const()[name = string("logits_chunk_1_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_1_argmax_output_dtype_0 = const()[name = string("logits_chunk_1_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> logits_chunk_1_argmax = reduce_argmax(axis = logits_chunk_1_argmax_axis_0, keep_dims = logits_chunk_1_argmax_keep_dims_0, output_dtype = logits_chunk_1_argmax_output_dtype_0, x = logits_chunk_1_mul)[name = string("logits_chunk_1_argmax")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_1_sub = sub(x = logits_chunk_1_mul, y = logits_chunk_1_max)[name = string("logits_chunk_1_sub")];
+            tensor<int32, [1]> logits_chunk_1_lse_sub_axes_0 = const()[name = string("logits_chunk_1_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_1_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_1_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_1_lse_sub = reduce_log_sum_exp(axes = logits_chunk_1_lse_sub_axes_0, keep_dims = logits_chunk_1_lse_sub_keep_dims_0, x = logits_chunk_1_sub)[name = string("logits_chunk_1_lse_sub")];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_1_lse = add(x = logits_chunk_1_lse_sub, y = logits_chunk_1_max)[name = string("logits_chunk_1_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_2_weight_0 = const()[name = string("logits_chunk_2_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16781568)))];
+            tensor<int32, [2]> logits_chunk_2_strides_0 = const()[name = string("logits_chunk_2_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_2_pad_type_0 = const()[name = string("logits_chunk_2_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_2_pad_0 = const()[name = string("logits_chunk_2_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_2_dilations_0 = const()[name = string("logits_chunk_2_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_2_groups_0 = const()[name = string("logits_chunk_2_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_2 = conv(dilations = logits_chunk_2_dilations_0, groups = logits_chunk_2_groups_0, pad = logits_chunk_2_pad_0, pad_type = logits_chunk_2_pad_type_0, strides = logits_chunk_2_strides_0, weight = logits_chunk_2_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_2")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_2_mul = mul(x = logits_chunk_2, y = temp_inverse)[name = string("logits_chunk_2_mul")];
+            tensor<int32, [1]> logits_chunk_2_max_axes_0 = const()[name = string("logits_chunk_2_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_2_max_keep_dims_0 = const()[name = string("logits_chunk_2_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_2_max = reduce_max(axes = logits_chunk_2_max_axes_0, keep_dims = logits_chunk_2_max_keep_dims_0, x = logits_chunk_2_mul)[name = string("logits_chunk_2_max")];
+            int32 logits_chunk_2_argmax_axis_0 = const()[name = string("logits_chunk_2_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_2_argmax_keep_dims_0 = const()[name = string("logits_chunk_2_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_2_argmax_output_dtype_0 = const()[name = string("logits_chunk_2_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> logits_chunk_2_argmax = reduce_argmax(axis = logits_chunk_2_argmax_axis_0, keep_dims = logits_chunk_2_argmax_keep_dims_0, output_dtype = logits_chunk_2_argmax_output_dtype_0, x = logits_chunk_2_mul)[name = string("logits_chunk_2_argmax")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_2_sub = sub(x = logits_chunk_2_mul, y = logits_chunk_2_max)[name = string("logits_chunk_2_sub")];
+            tensor<int32, [1]> logits_chunk_2_lse_sub_axes_0 = const()[name = string("logits_chunk_2_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_2_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_2_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_2_lse_sub = reduce_log_sum_exp(axes = logits_chunk_2_lse_sub_axes_0, keep_dims = logits_chunk_2_lse_sub_keep_dims_0, x = logits_chunk_2_sub)[name = string("logits_chunk_2_lse_sub")];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_2_lse = add(x = logits_chunk_2_lse_sub, y = logits_chunk_2_max)[name = string("logits_chunk_2_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_3_weight_0 = const()[name = string("logits_chunk_3_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25170240)))];
+            tensor<int32, [2]> logits_chunk_3_strides_0 = const()[name = string("logits_chunk_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_3_pad_type_0 = const()[name = string("logits_chunk_3_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_3_pad_0 = const()[name = string("logits_chunk_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_3_dilations_0 = const()[name = string("logits_chunk_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_3_groups_0 = const()[name = string("logits_chunk_3_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_3 = conv(dilations = logits_chunk_3_dilations_0, groups = logits_chunk_3_groups_0, pad = logits_chunk_3_pad_0, pad_type = logits_chunk_3_pad_type_0, strides = logits_chunk_3_strides_0, weight = logits_chunk_3_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_3")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_3_mul = mul(x = logits_chunk_3, y = temp_inverse)[name = string("logits_chunk_3_mul")];
+            tensor<int32, [1]> logits_chunk_3_max_axes_0 = const()[name = string("logits_chunk_3_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_3_max_keep_dims_0 = const()[name = string("logits_chunk_3_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_3_max = reduce_max(axes = logits_chunk_3_max_axes_0, keep_dims = logits_chunk_3_max_keep_dims_0, x = logits_chunk_3_mul)[name = string("logits_chunk_3_max")];
+            int32 logits_chunk_3_argmax_axis_0 = const()[name = string("logits_chunk_3_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_3_argmax_keep_dims_0 = const()[name = string("logits_chunk_3_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_3_argmax_output_dtype_0 = const()[name = string("logits_chunk_3_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> logits_chunk_3_argmax = reduce_argmax(axis = logits_chunk_3_argmax_axis_0, keep_dims = logits_chunk_3_argmax_keep_dims_0, output_dtype = logits_chunk_3_argmax_output_dtype_0, x = logits_chunk_3_mul)[name = string("logits_chunk_3_argmax")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_3_sub = sub(x = logits_chunk_3_mul, y = logits_chunk_3_max)[name = string("logits_chunk_3_sub")];
+            tensor<int32, [1]> logits_chunk_3_lse_sub_axes_0 = const()[name = string("logits_chunk_3_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_3_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_3_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_3_lse_sub = reduce_log_sum_exp(axes = logits_chunk_3_lse_sub_axes_0, keep_dims = logits_chunk_3_lse_sub_keep_dims_0, x = logits_chunk_3_sub)[name = string("logits_chunk_3_lse_sub")];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_3_lse = add(x = logits_chunk_3_lse_sub, y = logits_chunk_3_max)[name = string("logits_chunk_3_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_4_weight_0 = const()[name = string("logits_chunk_4_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33558912)))];
+            tensor<int32, [2]> logits_chunk_4_strides_0 = const()[name = string("logits_chunk_4_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_4_pad_type_0 = const()[name = string("logits_chunk_4_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_4_pad_0 = const()[name = string("logits_chunk_4_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_4_dilations_0 = const()[name = string("logits_chunk_4_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_4_groups_0 = const()[name = string("logits_chunk_4_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_4 = conv(dilations = logits_chunk_4_dilations_0, groups = logits_chunk_4_groups_0, pad = logits_chunk_4_pad_0, pad_type = logits_chunk_4_pad_type_0, strides = logits_chunk_4_strides_0, weight = logits_chunk_4_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_4")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_4_mul = mul(x = logits_chunk_4, y = temp_inverse)[name = string("logits_chunk_4_mul")];
+            tensor<int32, [1]> logits_chunk_4_max_axes_0 = const()[name = string("logits_chunk_4_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_4_max_keep_dims_0 = const()[name = string("logits_chunk_4_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_4_max = reduce_max(axes = logits_chunk_4_max_axes_0, keep_dims = logits_chunk_4_max_keep_dims_0, x = logits_chunk_4_mul)[name = string("logits_chunk_4_max")];
+            int32 logits_chunk_4_argmax_axis_0 = const()[name = string("logits_chunk_4_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_4_argmax_keep_dims_0 = const()[name = string("logits_chunk_4_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_4_argmax_output_dtype_0 = const()[name = string("logits_chunk_4_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> logits_chunk_4_argmax = reduce_argmax(axis = logits_chunk_4_argmax_axis_0, keep_dims = logits_chunk_4_argmax_keep_dims_0, output_dtype = logits_chunk_4_argmax_output_dtype_0, x = logits_chunk_4_mul)[name = string("logits_chunk_4_argmax")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_4_sub = sub(x = logits_chunk_4_mul, y = logits_chunk_4_max)[name = string("logits_chunk_4_sub")];
+            tensor<int32, [1]> logits_chunk_4_lse_sub_axes_0 = const()[name = string("logits_chunk_4_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_4_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_4_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_4_lse_sub = reduce_log_sum_exp(axes = logits_chunk_4_lse_sub_axes_0, keep_dims = logits_chunk_4_lse_sub_keep_dims_0, x = logits_chunk_4_sub)[name = string("logits_chunk_4_lse_sub")];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_4_lse = add(x = logits_chunk_4_lse_sub, y = logits_chunk_4_max)[name = string("logits_chunk_4_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_5_weight_0 = const()[name = string("logits_chunk_5_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41947584)))];
+            tensor<int32, [2]> logits_chunk_5_strides_0 = const()[name = string("logits_chunk_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_5_pad_type_0 = const()[name = string("logits_chunk_5_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_5_pad_0 = const()[name = string("logits_chunk_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_5_dilations_0 = const()[name = string("logits_chunk_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_5_groups_0 = const()[name = string("logits_chunk_5_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_5 = conv(dilations = logits_chunk_5_dilations_0, groups = logits_chunk_5_groups_0, pad = logits_chunk_5_pad_0, pad_type = logits_chunk_5_pad_type_0, strides = logits_chunk_5_strides_0, weight = logits_chunk_5_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_5")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_5_mul = mul(x = logits_chunk_5, y = temp_inverse)[name = string("logits_chunk_5_mul")];
+            tensor<int32, [1]> logits_chunk_5_max_axes_0 = const()[name = string("logits_chunk_5_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_5_max_keep_dims_0 = const()[name = string("logits_chunk_5_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_5_max = reduce_max(axes = logits_chunk_5_max_axes_0, keep_dims = logits_chunk_5_max_keep_dims_0, x = logits_chunk_5_mul)[name = string("logits_chunk_5_max")];
+            int32 logits_chunk_5_argmax_axis_0 = const()[name = string("logits_chunk_5_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_5_argmax_keep_dims_0 = const()[name = string("logits_chunk_5_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_5_argmax_output_dtype_0 = const()[name = string("logits_chunk_5_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> logits_chunk_5_argmax = reduce_argmax(axis = logits_chunk_5_argmax_axis_0, keep_dims = logits_chunk_5_argmax_keep_dims_0, output_dtype = logits_chunk_5_argmax_output_dtype_0, x = logits_chunk_5_mul)[name = string("logits_chunk_5_argmax")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_5_sub = sub(x = logits_chunk_5_mul, y = logits_chunk_5_max)[name = string("logits_chunk_5_sub")];
+            tensor<int32, [1]> logits_chunk_5_lse_sub_axes_0 = const()[name = string("logits_chunk_5_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_5_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_5_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_5_lse_sub = reduce_log_sum_exp(axes = logits_chunk_5_lse_sub_axes_0, keep_dims = logits_chunk_5_lse_sub_keep_dims_0, x = logits_chunk_5_sub)[name = string("logits_chunk_5_lse_sub")];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_5_lse = add(x = logits_chunk_5_lse_sub, y = logits_chunk_5_max)[name = string("logits_chunk_5_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_6_weight_0 = const()[name = string("logits_chunk_6_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50336256)))];
+            tensor<int32, [2]> logits_chunk_6_strides_0 = const()[name = string("logits_chunk_6_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_6_pad_type_0 = const()[name = string("logits_chunk_6_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_6_pad_0 = const()[name = string("logits_chunk_6_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_6_dilations_0 = const()[name = string("logits_chunk_6_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_6_groups_0 = const()[name = string("logits_chunk_6_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_6 = conv(dilations = logits_chunk_6_dilations_0, groups = logits_chunk_6_groups_0, pad = logits_chunk_6_pad_0, pad_type = logits_chunk_6_pad_type_0, strides = logits_chunk_6_strides_0, weight = logits_chunk_6_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_6")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_6_mul = mul(x = logits_chunk_6, y = temp_inverse)[name = string("logits_chunk_6_mul")];
+            tensor<int32, [1]> logits_chunk_6_max_axes_0 = const()[name = string("logits_chunk_6_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_6_max_keep_dims_0 = const()[name = string("logits_chunk_6_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_6_max = reduce_max(axes = logits_chunk_6_max_axes_0, keep_dims = logits_chunk_6_max_keep_dims_0, x = logits_chunk_6_mul)[name = string("logits_chunk_6_max")];
+            int32 logits_chunk_6_argmax_axis_0 = const()[name = string("logits_chunk_6_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_6_argmax_keep_dims_0 = const()[name = string("logits_chunk_6_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_6_argmax_output_dtype_0 = const()[name = string("logits_chunk_6_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> logits_chunk_6_argmax = reduce_argmax(axis = logits_chunk_6_argmax_axis_0, keep_dims = logits_chunk_6_argmax_keep_dims_0, output_dtype = logits_chunk_6_argmax_output_dtype_0, x = logits_chunk_6_mul)[name = string("logits_chunk_6_argmax")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_6_sub = sub(x = logits_chunk_6_mul, y = logits_chunk_6_max)[name = string("logits_chunk_6_sub")];
+            tensor<int32, [1]> logits_chunk_6_lse_sub_axes_0 = const()[name = string("logits_chunk_6_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_6_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_6_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_6_lse_sub = reduce_log_sum_exp(axes = logits_chunk_6_lse_sub_axes_0, keep_dims = logits_chunk_6_lse_sub_keep_dims_0, x = logits_chunk_6_sub)[name = string("logits_chunk_6_lse_sub")];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_6_lse = add(x = logits_chunk_6_lse_sub, y = logits_chunk_6_max)[name = string("logits_chunk_6_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_7_weight_0 = const()[name = string("logits_chunk_7_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58724928)))];
+            tensor<int32, [2]> logits_chunk_7_strides_0 = const()[name = string("logits_chunk_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_7_pad_type_0 = const()[name = string("logits_chunk_7_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_7_pad_0 = const()[name = string("logits_chunk_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_7_dilations_0 = const()[name = string("logits_chunk_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_7_groups_0 = const()[name = string("logits_chunk_7_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_7 = conv(dilations = logits_chunk_7_dilations_0, groups = logits_chunk_7_groups_0, pad = logits_chunk_7_pad_0, pad_type = logits_chunk_7_pad_type_0, strides = logits_chunk_7_strides_0, weight = logits_chunk_7_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_7")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_7_mul = mul(x = logits_chunk_7, y = temp_inverse)[name = string("logits_chunk_7_mul")];
+            tensor<int32, [1]> logits_chunk_7_max_axes_0 = const()[name = string("logits_chunk_7_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_7_max_keep_dims_0 = const()[name = string("logits_chunk_7_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_7_max = reduce_max(axes = logits_chunk_7_max_axes_0, keep_dims = logits_chunk_7_max_keep_dims_0, x = logits_chunk_7_mul)[name = string("logits_chunk_7_max")];
+            int32 logits_chunk_7_argmax_axis_0 = const()[name = string("logits_chunk_7_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_7_argmax_keep_dims_0 = const()[name = string("logits_chunk_7_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_7_argmax_output_dtype_0 = const()[name = string("logits_chunk_7_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> logits_chunk_7_argmax = reduce_argmax(axis = logits_chunk_7_argmax_axis_0, keep_dims = logits_chunk_7_argmax_keep_dims_0, output_dtype = logits_chunk_7_argmax_output_dtype_0, x = logits_chunk_7_mul)[name = string("logits_chunk_7_argmax")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_7_sub = sub(x = logits_chunk_7_mul, y = logits_chunk_7_max)[name = string("logits_chunk_7_sub")];
+            tensor<int32, [1]> logits_chunk_7_lse_sub_axes_0 = const()[name = string("logits_chunk_7_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_7_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_7_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_7_lse_sub = reduce_log_sum_exp(axes = logits_chunk_7_lse_sub_axes_0, keep_dims = logits_chunk_7_lse_sub_keep_dims_0, x = logits_chunk_7_sub)[name = string("logits_chunk_7_lse_sub")];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_7_lse = add(x = logits_chunk_7_lse_sub, y = logits_chunk_7_max)[name = string("logits_chunk_7_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_8_weight_0 = const()[name = string("logits_chunk_8_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67113600)))];
+            tensor<int32, [2]> logits_chunk_8_strides_0 = const()[name = string("logits_chunk_8_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_8_pad_type_0 = const()[name = string("logits_chunk_8_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_8_pad_0 = const()[name = string("logits_chunk_8_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_8_dilations_0 = const()[name = string("logits_chunk_8_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_8_groups_0 = const()[name = string("logits_chunk_8_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_8 = conv(dilations = logits_chunk_8_dilations_0, groups = logits_chunk_8_groups_0, pad = logits_chunk_8_pad_0, pad_type = logits_chunk_8_pad_type_0, strides = logits_chunk_8_strides_0, weight = logits_chunk_8_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_8")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_8_mul = mul(x = logits_chunk_8, y = temp_inverse)[name = string("logits_chunk_8_mul")];
+            tensor<int32, [1]> logits_chunk_8_max_axes_0 = const()[name = string("logits_chunk_8_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_8_max_keep_dims_0 = const()[name = string("logits_chunk_8_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_8_max = reduce_max(axes = logits_chunk_8_max_axes_0, keep_dims = logits_chunk_8_max_keep_dims_0, x = logits_chunk_8_mul)[name = string("logits_chunk_8_max")];
+            int32 logits_chunk_8_argmax_axis_0 = const()[name = string("logits_chunk_8_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_8_argmax_keep_dims_0 = const()[name = string("logits_chunk_8_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_8_argmax_output_dtype_0 = const()[name = string("logits_chunk_8_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> logits_chunk_8_argmax = reduce_argmax(axis = logits_chunk_8_argmax_axis_0, keep_dims = logits_chunk_8_argmax_keep_dims_0, output_dtype = logits_chunk_8_argmax_output_dtype_0, x = logits_chunk_8_mul)[name = string("logits_chunk_8_argmax")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_8_sub = sub(x = logits_chunk_8_mul, y = logits_chunk_8_max)[name = string("logits_chunk_8_sub")];
+            tensor<int32, [1]> logits_chunk_8_lse_sub_axes_0 = const()[name = string("logits_chunk_8_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_8_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_8_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_8_lse_sub = reduce_log_sum_exp(axes = logits_chunk_8_lse_sub_axes_0, keep_dims = logits_chunk_8_lse_sub_keep_dims_0, x = logits_chunk_8_sub)[name = string("logits_chunk_8_lse_sub")];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_8_lse = add(x = logits_chunk_8_lse_sub, y = logits_chunk_8_max)[name = string("logits_chunk_8_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_9_weight_0 = const()[name = string("logits_chunk_9_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75502272)))];
+            tensor<int32, [2]> logits_chunk_9_strides_0 = const()[name = string("logits_chunk_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_9_pad_type_0 = const()[name = string("logits_chunk_9_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_9_pad_0 = const()[name = string("logits_chunk_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_9_dilations_0 = const()[name = string("logits_chunk_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_9_groups_0 = const()[name = string("logits_chunk_9_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_9 = conv(dilations = logits_chunk_9_dilations_0, groups = logits_chunk_9_groups_0, pad = logits_chunk_9_pad_0, pad_type = logits_chunk_9_pad_type_0, strides = logits_chunk_9_strides_0, weight = logits_chunk_9_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_9")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_9_mul = mul(x = logits_chunk_9, y = temp_inverse)[name = string("logits_chunk_9_mul")];
+            tensor<int32, [1]> logits_chunk_9_max_axes_0 = const()[name = string("logits_chunk_9_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_9_max_keep_dims_0 = const()[name = string("logits_chunk_9_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_9_max = reduce_max(axes = logits_chunk_9_max_axes_0, keep_dims = logits_chunk_9_max_keep_dims_0, x = logits_chunk_9_mul)[name = string("logits_chunk_9_max")];
+            int32 logits_chunk_9_argmax_axis_0 = const()[name = string("logits_chunk_9_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_9_argmax_keep_dims_0 = const()[name = string("logits_chunk_9_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_9_argmax_output_dtype_0 = const()[name = string("logits_chunk_9_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> logits_chunk_9_argmax = reduce_argmax(axis = logits_chunk_9_argmax_axis_0, keep_dims = logits_chunk_9_argmax_keep_dims_0, output_dtype = logits_chunk_9_argmax_output_dtype_0, x = logits_chunk_9_mul)[name = string("logits_chunk_9_argmax")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_9_sub = sub(x = logits_chunk_9_mul, y = logits_chunk_9_max)[name = string("logits_chunk_9_sub")];
+            tensor<int32, [1]> logits_chunk_9_lse_sub_axes_0 = const()[name = string("logits_chunk_9_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_9_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_9_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_9_lse_sub = reduce_log_sum_exp(axes = logits_chunk_9_lse_sub_axes_0, keep_dims = logits_chunk_9_lse_sub_keep_dims_0, x = logits_chunk_9_sub)[name = string("logits_chunk_9_lse_sub")];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_9_lse = add(x = logits_chunk_9_lse_sub, y = logits_chunk_9_max)[name = string("logits_chunk_9_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_10_weight_0 = const()[name = string("logits_chunk_10_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83890944)))];
+            tensor<int32, [2]> logits_chunk_10_strides_0 = const()[name = string("logits_chunk_10_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_10_pad_type_0 = const()[name = string("logits_chunk_10_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_10_pad_0 = const()[name = string("logits_chunk_10_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_10_dilations_0 = const()[name = string("logits_chunk_10_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_10_groups_0 = const()[name = string("logits_chunk_10_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_10 = conv(dilations = logits_chunk_10_dilations_0, groups = logits_chunk_10_groups_0, pad = logits_chunk_10_pad_0, pad_type = logits_chunk_10_pad_type_0, strides = logits_chunk_10_strides_0, weight = logits_chunk_10_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_10")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_10_mul = mul(x = logits_chunk_10, y = temp_inverse)[name = string("logits_chunk_10_mul")];
+            tensor<int32, [1]> logits_chunk_10_max_axes_0 = const()[name = string("logits_chunk_10_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_10_max_keep_dims_0 = const()[name = string("logits_chunk_10_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_10_max = reduce_max(axes = logits_chunk_10_max_axes_0, keep_dims = logits_chunk_10_max_keep_dims_0, x = logits_chunk_10_mul)[name = string("logits_chunk_10_max")];
+            int32 logits_chunk_10_argmax_axis_0 = const()[name = string("logits_chunk_10_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_10_argmax_keep_dims_0 = const()[name = string("logits_chunk_10_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_10_argmax_output_dtype_0 = const()[name = string("logits_chunk_10_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> logits_chunk_10_argmax = reduce_argmax(axis = logits_chunk_10_argmax_axis_0, keep_dims = logits_chunk_10_argmax_keep_dims_0, output_dtype = logits_chunk_10_argmax_output_dtype_0, x = logits_chunk_10_mul)[name = string("logits_chunk_10_argmax")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_10_sub = sub(x = logits_chunk_10_mul, y = logits_chunk_10_max)[name = string("logits_chunk_10_sub")];
+            tensor<int32, [1]> logits_chunk_10_lse_sub_axes_0 = const()[name = string("logits_chunk_10_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_10_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_10_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_10_lse_sub = reduce_log_sum_exp(axes = logits_chunk_10_lse_sub_axes_0, keep_dims = logits_chunk_10_lse_sub_keep_dims_0, x = logits_chunk_10_sub)[name = string("logits_chunk_10_lse_sub")];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_10_lse = add(x = logits_chunk_10_lse_sub, y = logits_chunk_10_max)[name = string("logits_chunk_10_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_11_weight_0 = const()[name = string("logits_chunk_11_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92279616)))];
+            tensor<int32, [2]> logits_chunk_11_strides_0 = const()[name = string("logits_chunk_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_11_pad_type_0 = const()[name = string("logits_chunk_11_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_11_pad_0 = const()[name = string("logits_chunk_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_11_dilations_0 = const()[name = string("logits_chunk_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_11_groups_0 = const()[name = string("logits_chunk_11_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_11 = conv(dilations = logits_chunk_11_dilations_0, groups = logits_chunk_11_groups_0, pad = logits_chunk_11_pad_0, pad_type = logits_chunk_11_pad_type_0, strides = logits_chunk_11_strides_0, weight = logits_chunk_11_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_11")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_11_mul = mul(x = logits_chunk_11, y = temp_inverse)[name = string("logits_chunk_11_mul")];
+            tensor<int32, [1]> logits_chunk_11_max_axes_0 = const()[name = string("logits_chunk_11_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_11_max_keep_dims_0 = const()[name = string("logits_chunk_11_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_11_max = reduce_max(axes = logits_chunk_11_max_axes_0, keep_dims = logits_chunk_11_max_keep_dims_0, x = logits_chunk_11_mul)[name = string("logits_chunk_11_max")];
+            int32 logits_chunk_11_argmax_axis_0 = const()[name = string("logits_chunk_11_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_11_argmax_keep_dims_0 = const()[name = string("logits_chunk_11_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_11_argmax_output_dtype_0 = const()[name = string("logits_chunk_11_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> logits_chunk_11_argmax = reduce_argmax(axis = logits_chunk_11_argmax_axis_0, keep_dims = logits_chunk_11_argmax_keep_dims_0, output_dtype = logits_chunk_11_argmax_output_dtype_0, x = logits_chunk_11_mul)[name = string("logits_chunk_11_argmax")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_11_sub = sub(x = logits_chunk_11_mul, y = logits_chunk_11_max)[name = string("logits_chunk_11_sub")];
+            tensor<int32, [1]> logits_chunk_11_lse_sub_axes_0 = const()[name = string("logits_chunk_11_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_11_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_11_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_11_lse_sub = reduce_log_sum_exp(axes = logits_chunk_11_lse_sub_axes_0, keep_dims = logits_chunk_11_lse_sub_keep_dims_0, x = logits_chunk_11_sub)[name = string("logits_chunk_11_lse_sub")];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_11_lse = add(x = logits_chunk_11_lse_sub, y = logits_chunk_11_max)[name = string("logits_chunk_11_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_12_weight_0 = const()[name = string("logits_chunk_12_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100668288)))];
+            tensor<int32, [2]> logits_chunk_12_strides_0 = const()[name = string("logits_chunk_12_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_12_pad_type_0 = const()[name = string("logits_chunk_12_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_12_pad_0 = const()[name = string("logits_chunk_12_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_12_dilations_0 = const()[name = string("logits_chunk_12_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_12_groups_0 = const()[name = string("logits_chunk_12_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_12 = conv(dilations = logits_chunk_12_dilations_0, groups = logits_chunk_12_groups_0, pad = logits_chunk_12_pad_0, pad_type = logits_chunk_12_pad_type_0, strides = logits_chunk_12_strides_0, weight = logits_chunk_12_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_12")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_12_mul = mul(x = logits_chunk_12, y = temp_inverse)[name = string("logits_chunk_12_mul")];
+            tensor<int32, [1]> logits_chunk_12_max_axes_0 = const()[name = string("logits_chunk_12_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_12_max_keep_dims_0 = const()[name = string("logits_chunk_12_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_12_max = reduce_max(axes = logits_chunk_12_max_axes_0, keep_dims = logits_chunk_12_max_keep_dims_0, x = logits_chunk_12_mul)[name = string("logits_chunk_12_max")];
+            int32 logits_chunk_12_argmax_axis_0 = const()[name = string("logits_chunk_12_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_12_argmax_keep_dims_0 = const()[name = string("logits_chunk_12_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_12_argmax_output_dtype_0 = const()[name = string("logits_chunk_12_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> logits_chunk_12_argmax = reduce_argmax(axis = logits_chunk_12_argmax_axis_0, keep_dims = logits_chunk_12_argmax_keep_dims_0, output_dtype = logits_chunk_12_argmax_output_dtype_0, x = logits_chunk_12_mul)[name = string("logits_chunk_12_argmax")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_12_sub = sub(x = logits_chunk_12_mul, y = logits_chunk_12_max)[name = string("logits_chunk_12_sub")];
+            tensor<int32, [1]> logits_chunk_12_lse_sub_axes_0 = const()[name = string("logits_chunk_12_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_12_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_12_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_12_lse_sub = reduce_log_sum_exp(axes = logits_chunk_12_lse_sub_axes_0, keep_dims = logits_chunk_12_lse_sub_keep_dims_0, x = logits_chunk_12_sub)[name = string("logits_chunk_12_lse_sub")];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_12_lse = add(x = logits_chunk_12_lse_sub, y = logits_chunk_12_max)[name = string("logits_chunk_12_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_13_weight_0 = const()[name = string("logits_chunk_13_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109056960)))];
+            tensor<int32, [2]> logits_chunk_13_strides_0 = const()[name = string("logits_chunk_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_13_pad_type_0 = const()[name = string("logits_chunk_13_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_13_pad_0 = const()[name = string("logits_chunk_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_13_dilations_0 = const()[name = string("logits_chunk_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_13_groups_0 = const()[name = string("logits_chunk_13_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_13 = conv(dilations = logits_chunk_13_dilations_0, groups = logits_chunk_13_groups_0, pad = logits_chunk_13_pad_0, pad_type = logits_chunk_13_pad_type_0, strides = logits_chunk_13_strides_0, weight = logits_chunk_13_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_13")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_13_mul = mul(x = logits_chunk_13, y = temp_inverse)[name = string("logits_chunk_13_mul")];
+            tensor<int32, [1]> logits_chunk_13_max_axes_0 = const()[name = string("logits_chunk_13_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_13_max_keep_dims_0 = const()[name = string("logits_chunk_13_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_13_max = reduce_max(axes = logits_chunk_13_max_axes_0, keep_dims = logits_chunk_13_max_keep_dims_0, x = logits_chunk_13_mul)[name = string("logits_chunk_13_max")];
+            int32 logits_chunk_13_argmax_axis_0 = const()[name = string("logits_chunk_13_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_13_argmax_keep_dims_0 = const()[name = string("logits_chunk_13_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_13_argmax_output_dtype_0 = const()[name = string("logits_chunk_13_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> logits_chunk_13_argmax = reduce_argmax(axis = logits_chunk_13_argmax_axis_0, keep_dims = logits_chunk_13_argmax_keep_dims_0, output_dtype = logits_chunk_13_argmax_output_dtype_0, x = logits_chunk_13_mul)[name = string("logits_chunk_13_argmax")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_13_sub = sub(x = logits_chunk_13_mul, y = logits_chunk_13_max)[name = string("logits_chunk_13_sub")];
+            tensor<int32, [1]> logits_chunk_13_lse_sub_axes_0 = const()[name = string("logits_chunk_13_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_13_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_13_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_13_lse_sub = reduce_log_sum_exp(axes = logits_chunk_13_lse_sub_axes_0, keep_dims = logits_chunk_13_lse_sub_keep_dims_0, x = logits_chunk_13_sub)[name = string("logits_chunk_13_lse_sub")];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_13_lse = add(x = logits_chunk_13_lse_sub, y = logits_chunk_13_max)[name = string("logits_chunk_13_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_14_weight_0 = const()[name = string("logits_chunk_14_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117445632)))];
+            tensor<int32, [2]> logits_chunk_14_strides_0 = const()[name = string("logits_chunk_14_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_14_pad_type_0 = const()[name = string("logits_chunk_14_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_14_pad_0 = const()[name = string("logits_chunk_14_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_14_dilations_0 = const()[name = string("logits_chunk_14_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_14_groups_0 = const()[name = string("logits_chunk_14_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_14 = conv(dilations = logits_chunk_14_dilations_0, groups = logits_chunk_14_groups_0, pad = logits_chunk_14_pad_0, pad_type = logits_chunk_14_pad_type_0, strides = logits_chunk_14_strides_0, weight = logits_chunk_14_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_14")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_14_mul = mul(x = logits_chunk_14, y = temp_inverse)[name = string("logits_chunk_14_mul")];
+            tensor<int32, [1]> logits_chunk_14_max_axes_0 = const()[name = string("logits_chunk_14_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_14_max_keep_dims_0 = const()[name = string("logits_chunk_14_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_14_max = reduce_max(axes = logits_chunk_14_max_axes_0, keep_dims = logits_chunk_14_max_keep_dims_0, x = logits_chunk_14_mul)[name = string("logits_chunk_14_max")];
+            int32 logits_chunk_14_argmax_axis_0 = const()[name = string("logits_chunk_14_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_14_argmax_keep_dims_0 = const()[name = string("logits_chunk_14_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_14_argmax_output_dtype_0 = const()[name = string("logits_chunk_14_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> logits_chunk_14_argmax = reduce_argmax(axis = logits_chunk_14_argmax_axis_0, keep_dims = logits_chunk_14_argmax_keep_dims_0, output_dtype = logits_chunk_14_argmax_output_dtype_0, x = logits_chunk_14_mul)[name = string("logits_chunk_14_argmax")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_14_sub = sub(x = logits_chunk_14_mul, y = logits_chunk_14_max)[name = string("logits_chunk_14_sub")];
+            tensor<int32, [1]> logits_chunk_14_lse_sub_axes_0 = const()[name = string("logits_chunk_14_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_14_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_14_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_14_lse_sub = reduce_log_sum_exp(axes = logits_chunk_14_lse_sub_axes_0, keep_dims = logits_chunk_14_lse_sub_keep_dims_0, x = logits_chunk_14_sub)[name = string("logits_chunk_14_lse_sub")];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_14_lse = add(x = logits_chunk_14_lse_sub, y = logits_chunk_14_max)[name = string("logits_chunk_14_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_15_weight_0 = const()[name = string("logits_chunk_15_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125834304)))];
+            tensor<int32, [2]> logits_chunk_15_strides_0 = const()[name = string("logits_chunk_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_15_pad_type_0 = const()[name = string("logits_chunk_15_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_15_pad_0 = const()[name = string("logits_chunk_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_15_dilations_0 = const()[name = string("logits_chunk_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_15_groups_0 = const()[name = string("logits_chunk_15_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_15 = conv(dilations = logits_chunk_15_dilations_0, groups = logits_chunk_15_groups_0, pad = logits_chunk_15_pad_0, pad_type = logits_chunk_15_pad_type_0, strides = logits_chunk_15_strides_0, weight = logits_chunk_15_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_15")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_15_mul = mul(x = logits_chunk_15, y = temp_inverse)[name = string("logits_chunk_15_mul")];
+            tensor<int32, [1]> logits_chunk_15_max_axes_0 = const()[name = string("logits_chunk_15_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_15_max_keep_dims_0 = const()[name = string("logits_chunk_15_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_15_max = reduce_max(axes = logits_chunk_15_max_axes_0, keep_dims = logits_chunk_15_max_keep_dims_0, x = logits_chunk_15_mul)[name = string("logits_chunk_15_max")];
+            int32 logits_chunk_15_argmax_axis_0 = const()[name = string("logits_chunk_15_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_15_argmax_keep_dims_0 = const()[name = string("logits_chunk_15_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_15_argmax_output_dtype_0 = const()[name = string("logits_chunk_15_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 1]> logits_chunk_15_argmax = reduce_argmax(axis = logits_chunk_15_argmax_axis_0, keep_dims = logits_chunk_15_argmax_keep_dims_0, output_dtype = logits_chunk_15_argmax_output_dtype_0, x = logits_chunk_15_mul)[name = string("logits_chunk_15_argmax")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_15_sub = sub(x = logits_chunk_15_mul, y = logits_chunk_15_max)[name = string("logits_chunk_15_sub")];
+            tensor<int32, [1]> logits_chunk_15_lse_sub_axes_0 = const()[name = string("logits_chunk_15_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_15_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_15_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_15_lse_sub = reduce_log_sum_exp(axes = logits_chunk_15_lse_sub_axes_0, keep_dims = logits_chunk_15_lse_sub_keep_dims_0, x = logits_chunk_15_sub)[name = string("logits_chunk_15_lse_sub")];
+            tensor<fp16, [1, 1, 1, 1]> logits_chunk_15_lse = add(x = logits_chunk_15_lse_sub, y = logits_chunk_15_max)[name = string("logits_chunk_15_lse")];
+            int32 logits_lses_axis_0 = const()[name = string("logits_lses_axis_0"), val = int32(1)];
+            bool logits_lses_interleave_0 = const()[name = string("logits_lses_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 1]> logits_lses = concat(axis = logits_lses_axis_0, interleave = logits_lses_interleave_0, values = (logits_chunk_0_lse, logits_chunk_1_lse, logits_chunk_2_lse, logits_chunk_3_lse, logits_chunk_4_lse, logits_chunk_5_lse, logits_chunk_6_lse, logits_chunk_7_lse, logits_chunk_8_lse, logits_chunk_9_lse, logits_chunk_10_lse, logits_chunk_11_lse, logits_chunk_12_lse, logits_chunk_13_lse, logits_chunk_14_lse, logits_chunk_15_lse))[name = string("logits_lses")];
+            tensor<int32, [1]> logits_lses_max_axes_0 = const()[name = string("logits_lses_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_lses_max_keep_dims_0 = const()[name = string("logits_lses_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_lses_max = reduce_max(axes = logits_lses_max_axes_0, keep_dims = logits_lses_max_keep_dims_0, x = logits_lses)[name = string("logits_lses_max")];
+            tensor<fp16, [1, 16, 1, 1]> logits_lses_sub = sub(x = logits_lses, y = logits_lses_max)[name = string("logits_lses_sub")];
+            tensor<int32, [1]> logits_lses_logsumexp_axes_0 = const()[name = string("logits_lses_logsumexp_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_lses_logsumexp_keep_dims_0 = const()[name = string("logits_lses_logsumexp_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_lses_logsumexp = reduce_log_sum_exp(axes = logits_lses_logsumexp_axes_0, keep_dims = logits_lses_logsumexp_keep_dims_0, x = logits_lses_sub)[name = string("logits_lses_logsumexp")];
+            tensor<fp16, [1, 1, 1, 1]> logits_lse = add(x = logits_lses_logsumexp, y = logits_lses_max)[name = string("logits_lse")];
+            int32 logits_max_logits_chunks_axis_0 = const()[name = string("logits_max_logits_chunks_axis_0"), val = int32(1)];
+            bool logits_max_logits_chunks_interleave_0 = const()[name = string("logits_max_logits_chunks_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 1]> logits_max_logits_chunks = concat(axis = logits_max_logits_chunks_axis_0, interleave = logits_max_logits_chunks_interleave_0, values = (logits_chunk_0_max, logits_chunk_1_max, logits_chunk_2_max, logits_chunk_3_max, logits_chunk_4_max, logits_chunk_5_max, logits_chunk_6_max, logits_chunk_7_max, logits_chunk_8_max, logits_chunk_9_max, logits_chunk_10_max, logits_chunk_11_max, logits_chunk_12_max, logits_chunk_13_max, logits_chunk_14_max, logits_chunk_15_max))[name = string("logits_max_logits_chunks")];
+            tensor<int32, [1]> logits_max_logit_axes_0 = const()[name = string("logits_max_logit_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_max_logit_keep_dims_0 = const()[name = string("logits_max_logit_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 1]> logits_max_logit = reduce_max(axes = logits_max_logit_axes_0, keep_dims = logits_max_logit_keep_dims_0, x = logits_max_logits_chunks)[name = string("logits_max_logit")];
+            tensor<fp16, [1, 1, 1, 1]> logits_max_logit_sub = sub(x = logits_max_logit, y = logits_lse)[name = string("logits_max_logit_sub")];
+            tensor<fp16, [1, 1, 1, 1]> max_prob = exp(x = logits_max_logit_sub)[name = string("max_prob")];
+            tensor<fp16, [1, 1, 1, 1]> min_p_thresh = mul(x = max_prob, y = p)[name = string("min_p_thresh")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_0_sub_1 = sub(x = logits_chunk_0_mul, y = logits_lse)[name = string("logits_chunk_0_sub_1")];
+            tensor<fp16, [1, 2048, 1, 1]> probs_chunk_0 = exp(x = logits_chunk_0_sub_1)[name = string("probs_chunk_0")];
+            tensor<bool, [1, 2048, 1, 1]> mask_probs_chunk_0 = greater_equal(x = probs_chunk_0, y = min_p_thresh)[name = string("mask_probs_chunk_0")];
+            string mask_chunk_0_fp16_dtype_0 = const()[name = string("mask_chunk_0_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 1]> mask_chunk_0_fp16 = cast(dtype = mask_chunk_0_fp16_dtype_0, x = mask_probs_chunk_0)[name = string("cast_67")];
+            tensor<fp16, [1, 2048, 1, 1]> masked_probs_chunk_0 = select(a = probs_chunk_0, b = mask_chunk_0_fp16, cond = mask_probs_chunk_0)[name = string("masked_probs_chunk_0")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_1_sub_1 = sub(x = logits_chunk_1_mul, y = logits_lse)[name = string("logits_chunk_1_sub_1")];
+            tensor<fp16, [1, 2048, 1, 1]> probs_chunk_1 = exp(x = logits_chunk_1_sub_1)[name = string("probs_chunk_1")];
+            tensor<bool, [1, 2048, 1, 1]> mask_probs_chunk_1 = greater_equal(x = probs_chunk_1, y = min_p_thresh)[name = string("mask_probs_chunk_1")];
+            string mask_chunk_1_fp16_dtype_0 = const()[name = string("mask_chunk_1_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 1]> mask_chunk_1_fp16 = cast(dtype = mask_chunk_1_fp16_dtype_0, x = mask_probs_chunk_1)[name = string("cast_66")];
+            tensor<fp16, [1, 2048, 1, 1]> masked_probs_chunk_1 = select(a = probs_chunk_1, b = mask_chunk_1_fp16, cond = mask_probs_chunk_1)[name = string("masked_probs_chunk_1")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_2_sub_1 = sub(x = logits_chunk_2_mul, y = logits_lse)[name = string("logits_chunk_2_sub_1")];
+            tensor<fp16, [1, 2048, 1, 1]> probs_chunk_2 = exp(x = logits_chunk_2_sub_1)[name = string("probs_chunk_2")];
+            tensor<bool, [1, 2048, 1, 1]> mask_probs_chunk_2 = greater_equal(x = probs_chunk_2, y = min_p_thresh)[name = string("mask_probs_chunk_2")];
+            string mask_chunk_2_fp16_dtype_0 = const()[name = string("mask_chunk_2_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 1]> mask_chunk_2_fp16 = cast(dtype = mask_chunk_2_fp16_dtype_0, x = mask_probs_chunk_2)[name = string("cast_65")];
+            tensor<fp16, [1, 2048, 1, 1]> masked_probs_chunk_2 = select(a = probs_chunk_2, b = mask_chunk_2_fp16, cond = mask_probs_chunk_2)[name = string("masked_probs_chunk_2")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_3_sub_1 = sub(x = logits_chunk_3_mul, y = logits_lse)[name = string("logits_chunk_3_sub_1")];
+            tensor<fp16, [1, 2048, 1, 1]> probs_chunk_3 = exp(x = logits_chunk_3_sub_1)[name = string("probs_chunk_3")];
+            tensor<bool, [1, 2048, 1, 1]> mask_probs_chunk_3 = greater_equal(x = probs_chunk_3, y = min_p_thresh)[name = string("mask_probs_chunk_3")];
+            string mask_chunk_3_fp16_dtype_0 = const()[name = string("mask_chunk_3_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 1]> mask_chunk_3_fp16 = cast(dtype = mask_chunk_3_fp16_dtype_0, x = mask_probs_chunk_3)[name = string("cast_64")];
+            tensor<fp16, [1, 2048, 1, 1]> masked_probs_chunk_3 = select(a = probs_chunk_3, b = mask_chunk_3_fp16, cond = mask_probs_chunk_3)[name = string("masked_probs_chunk_3")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_4_sub_1 = sub(x = logits_chunk_4_mul, y = logits_lse)[name = string("logits_chunk_4_sub_1")];
+            tensor<fp16, [1, 2048, 1, 1]> probs_chunk_4 = exp(x = logits_chunk_4_sub_1)[name = string("probs_chunk_4")];
+            tensor<bool, [1, 2048, 1, 1]> mask_probs_chunk_4 = greater_equal(x = probs_chunk_4, y = min_p_thresh)[name = string("mask_probs_chunk_4")];
+            string mask_chunk_4_fp16_dtype_0 = const()[name = string("mask_chunk_4_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 1]> mask_chunk_4_fp16 = cast(dtype = mask_chunk_4_fp16_dtype_0, x = mask_probs_chunk_4)[name = string("cast_63")];
+            tensor<fp16, [1, 2048, 1, 1]> masked_probs_chunk_4 = select(a = probs_chunk_4, b = mask_chunk_4_fp16, cond = mask_probs_chunk_4)[name = string("masked_probs_chunk_4")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_5_sub_1 = sub(x = logits_chunk_5_mul, y = logits_lse)[name = string("logits_chunk_5_sub_1")];
+            tensor<fp16, [1, 2048, 1, 1]> probs_chunk_5 = exp(x = logits_chunk_5_sub_1)[name = string("probs_chunk_5")];
+            tensor<bool, [1, 2048, 1, 1]> mask_probs_chunk_5 = greater_equal(x = probs_chunk_5, y = min_p_thresh)[name = string("mask_probs_chunk_5")];
+            string mask_chunk_5_fp16_dtype_0 = const()[name = string("mask_chunk_5_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 1]> mask_chunk_5_fp16 = cast(dtype = mask_chunk_5_fp16_dtype_0, x = mask_probs_chunk_5)[name = string("cast_62")];
+            tensor<fp16, [1, 2048, 1, 1]> masked_probs_chunk_5 = select(a = probs_chunk_5, b = mask_chunk_5_fp16, cond = mask_probs_chunk_5)[name = string("masked_probs_chunk_5")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_6_sub_1 = sub(x = logits_chunk_6_mul, y = logits_lse)[name = string("logits_chunk_6_sub_1")];
+            tensor<fp16, [1, 2048, 1, 1]> probs_chunk_6 = exp(x = logits_chunk_6_sub_1)[name = string("probs_chunk_6")];
+            tensor<bool, [1, 2048, 1, 1]> mask_probs_chunk_6 = greater_equal(x = probs_chunk_6, y = min_p_thresh)[name = string("mask_probs_chunk_6")];
+            string mask_chunk_6_fp16_dtype_0 = const()[name = string("mask_chunk_6_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 1]> mask_chunk_6_fp16 = cast(dtype = mask_chunk_6_fp16_dtype_0, x = mask_probs_chunk_6)[name = string("cast_61")];
+            tensor<fp16, [1, 2048, 1, 1]> masked_probs_chunk_6 = select(a = probs_chunk_6, b = mask_chunk_6_fp16, cond = mask_probs_chunk_6)[name = string("masked_probs_chunk_6")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_7_sub_1 = sub(x = logits_chunk_7_mul, y = logits_lse)[name = string("logits_chunk_7_sub_1")];
+            tensor<fp16, [1, 2048, 1, 1]> probs_chunk_7 = exp(x = logits_chunk_7_sub_1)[name = string("probs_chunk_7")];
+            tensor<bool, [1, 2048, 1, 1]> mask_probs_chunk_7 = greater_equal(x = probs_chunk_7, y = min_p_thresh)[name = string("mask_probs_chunk_7")];
+            string mask_chunk_7_fp16_dtype_0 = const()[name = string("mask_chunk_7_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 1]> mask_chunk_7_fp16 = cast(dtype = mask_chunk_7_fp16_dtype_0, x = mask_probs_chunk_7)[name = string("cast_60")];
+            tensor<fp16, [1, 2048, 1, 1]> masked_probs_chunk_7 = select(a = probs_chunk_7, b = mask_chunk_7_fp16, cond = mask_probs_chunk_7)[name = string("masked_probs_chunk_7")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_8_sub_1 = sub(x = logits_chunk_8_mul, y = logits_lse)[name = string("logits_chunk_8_sub_1")];
+            tensor<fp16, [1, 2048, 1, 1]> probs_chunk_8 = exp(x = logits_chunk_8_sub_1)[name = string("probs_chunk_8")];
+            tensor<bool, [1, 2048, 1, 1]> mask_probs_chunk_8 = greater_equal(x = probs_chunk_8, y = min_p_thresh)[name = string("mask_probs_chunk_8")];
+            string mask_chunk_8_fp16_dtype_0 = const()[name = string("mask_chunk_8_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 1]> mask_chunk_8_fp16 = cast(dtype = mask_chunk_8_fp16_dtype_0, x = mask_probs_chunk_8)[name = string("cast_59")];
+            tensor<fp16, [1, 2048, 1, 1]> masked_probs_chunk_8 = select(a = probs_chunk_8, b = mask_chunk_8_fp16, cond = mask_probs_chunk_8)[name = string("masked_probs_chunk_8")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_9_sub_1 = sub(x = logits_chunk_9_mul, y = logits_lse)[name = string("logits_chunk_9_sub_1")];
+            tensor<fp16, [1, 2048, 1, 1]> probs_chunk_9 = exp(x = logits_chunk_9_sub_1)[name = string("probs_chunk_9")];
+            tensor<bool, [1, 2048, 1, 1]> mask_probs_chunk_9 = greater_equal(x = probs_chunk_9, y = min_p_thresh)[name = string("mask_probs_chunk_9")];
+            string mask_chunk_9_fp16_dtype_0 = const()[name = string("mask_chunk_9_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 1]> mask_chunk_9_fp16 = cast(dtype = mask_chunk_9_fp16_dtype_0, x = mask_probs_chunk_9)[name = string("cast_58")];
+            tensor<fp16, [1, 2048, 1, 1]> masked_probs_chunk_9 = select(a = probs_chunk_9, b = mask_chunk_9_fp16, cond = mask_probs_chunk_9)[name = string("masked_probs_chunk_9")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_10_sub_1 = sub(x = logits_chunk_10_mul, y = logits_lse)[name = string("logits_chunk_10_sub_1")];
+            tensor<fp16, [1, 2048, 1, 1]> probs_chunk_10 = exp(x = logits_chunk_10_sub_1)[name = string("probs_chunk_10")];
+            tensor<bool, [1, 2048, 1, 1]> mask_probs_chunk_10 = greater_equal(x = probs_chunk_10, y = min_p_thresh)[name = string("mask_probs_chunk_10")];
+            string mask_chunk_10_fp16_dtype_0 = const()[name = string("mask_chunk_10_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 1]> mask_chunk_10_fp16 = cast(dtype = mask_chunk_10_fp16_dtype_0, x = mask_probs_chunk_10)[name = string("cast_57")];
+            tensor<fp16, [1, 2048, 1, 1]> masked_probs_chunk_10 = select(a = probs_chunk_10, b = mask_chunk_10_fp16, cond = mask_probs_chunk_10)[name = string("masked_probs_chunk_10")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_11_sub_1 = sub(x = logits_chunk_11_mul, y = logits_lse)[name = string("logits_chunk_11_sub_1")];
+            tensor<fp16, [1, 2048, 1, 1]> probs_chunk_11 = exp(x = logits_chunk_11_sub_1)[name = string("probs_chunk_11")];
+            tensor<bool, [1, 2048, 1, 1]> mask_probs_chunk_11 = greater_equal(x = probs_chunk_11, y = min_p_thresh)[name = string("mask_probs_chunk_11")];
+            string mask_chunk_11_fp16_dtype_0 = const()[name = string("mask_chunk_11_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 1]> mask_chunk_11_fp16 = cast(dtype = mask_chunk_11_fp16_dtype_0, x = mask_probs_chunk_11)[name = string("cast_56")];
+            tensor<fp16, [1, 2048, 1, 1]> masked_probs_chunk_11 = select(a = probs_chunk_11, b = mask_chunk_11_fp16, cond = mask_probs_chunk_11)[name = string("masked_probs_chunk_11")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_12_sub_1 = sub(x = logits_chunk_12_mul, y = logits_lse)[name = string("logits_chunk_12_sub_1")];
+            tensor<fp16, [1, 2048, 1, 1]> probs_chunk_12 = exp(x = logits_chunk_12_sub_1)[name = string("probs_chunk_12")];
+            tensor<bool, [1, 2048, 1, 1]> mask_probs_chunk_12 = greater_equal(x = probs_chunk_12, y = min_p_thresh)[name = string("mask_probs_chunk_12")];
+            string mask_chunk_12_fp16_dtype_0 = const()[name = string("mask_chunk_12_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 1]> mask_chunk_12_fp16 = cast(dtype = mask_chunk_12_fp16_dtype_0, x = mask_probs_chunk_12)[name = string("cast_55")];
+            tensor<fp16, [1, 2048, 1, 1]> masked_probs_chunk_12 = select(a = probs_chunk_12, b = mask_chunk_12_fp16, cond = mask_probs_chunk_12)[name = string("masked_probs_chunk_12")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_13_sub_1 = sub(x = logits_chunk_13_mul, y = logits_lse)[name = string("logits_chunk_13_sub_1")];
+            tensor<fp16, [1, 2048, 1, 1]> probs_chunk_13 = exp(x = logits_chunk_13_sub_1)[name = string("probs_chunk_13")];
+            tensor<bool, [1, 2048, 1, 1]> mask_probs_chunk_13 = greater_equal(x = probs_chunk_13, y = min_p_thresh)[name = string("mask_probs_chunk_13")];
+            string mask_chunk_13_fp16_dtype_0 = const()[name = string("mask_chunk_13_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 1]> mask_chunk_13_fp16 = cast(dtype = mask_chunk_13_fp16_dtype_0, x = mask_probs_chunk_13)[name = string("cast_54")];
+            tensor<fp16, [1, 2048, 1, 1]> masked_probs_chunk_13 = select(a = probs_chunk_13, b = mask_chunk_13_fp16, cond = mask_probs_chunk_13)[name = string("masked_probs_chunk_13")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_14_sub_1 = sub(x = logits_chunk_14_mul, y = logits_lse)[name = string("logits_chunk_14_sub_1")];
+            tensor<fp16, [1, 2048, 1, 1]> probs_chunk_14 = exp(x = logits_chunk_14_sub_1)[name = string("probs_chunk_14")];
+            tensor<bool, [1, 2048, 1, 1]> mask_probs_chunk_14 = greater_equal(x = probs_chunk_14, y = min_p_thresh)[name = string("mask_probs_chunk_14")];
+            string mask_chunk_14_fp16_dtype_0 = const()[name = string("mask_chunk_14_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 1]> mask_chunk_14_fp16 = cast(dtype = mask_chunk_14_fp16_dtype_0, x = mask_probs_chunk_14)[name = string("cast_53")];
+            tensor<fp16, [1, 2048, 1, 1]> masked_probs_chunk_14 = select(a = probs_chunk_14, b = mask_chunk_14_fp16, cond = mask_probs_chunk_14)[name = string("masked_probs_chunk_14")];
+            tensor<fp16, [1, 2048, 1, 1]> logits_chunk_15_sub_1 = sub(x = logits_chunk_15_mul, y = logits_lse)[name = string("logits_chunk_15_sub_1")];
+            tensor<fp16, [1, 2048, 1, 1]> probs_chunk_15 = exp(x = logits_chunk_15_sub_1)[name = string("probs_chunk_15")];
+            tensor<bool, [1, 2048, 1, 1]> mask_probs_chunk_15 = greater_equal(x = probs_chunk_15, y = min_p_thresh)[name = string("mask_probs_chunk_15")];
+            string mask_chunk_15_fp16_dtype_0 = const()[name = string("mask_chunk_15_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 1]> mask_chunk_15_fp16 = cast(dtype = mask_chunk_15_fp16_dtype_0, x = mask_probs_chunk_15)[name = string("cast_52")];
+            tensor<fp16, [1, 2048, 1, 1]> masked_probs_chunk_15 = select(a = probs_chunk_15, b = mask_chunk_15_fp16, cond = mask_probs_chunk_15)[name = string("masked_probs_chunk_15")];
+            int32 probs_axis_0 = const()[name = string("probs_axis_0"), val = int32(1)];
+            bool probs_interleave_0 = const()[name = string("probs_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 32768, 1, 1]> probs = concat(axis = probs_axis_0, interleave = probs_interleave_0, values = (masked_probs_chunk_0, masked_probs_chunk_1, masked_probs_chunk_2, masked_probs_chunk_3, masked_probs_chunk_4, masked_probs_chunk_5, masked_probs_chunk_6, masked_probs_chunk_7, masked_probs_chunk_8, masked_probs_chunk_9, masked_probs_chunk_10, masked_probs_chunk_11, masked_probs_chunk_12, masked_probs_chunk_13, masked_probs_chunk_14, masked_probs_chunk_15))[name = string("probs")];
+            string probs_fp32_dtype_0 = const()[name = string("probs_fp32_dtype_0"), val = string("fp32")];
+            int32 probs_cumsum_axis_0 = const()[name = string("probs_cumsum_axis_0"), val = int32(1)];
+            bool probs_cumsum_exclusive_0 = const()[name = string("probs_cumsum_exclusive_0"), val = bool(false)];
+            bool probs_cumsum_reverse_0 = const()[name = string("probs_cumsum_reverse_0"), val = bool(false)];
+            tensor<fp32, [1, 32768, 1, 1]> probs_fp32 = cast(dtype = probs_fp32_dtype_0, x = probs)[name = string("cast_51")];
+            tensor<fp32, [1, 32768, 1, 1]> probs_cumsum = cumsum(axis = probs_cumsum_axis_0, exclusive = probs_cumsum_exclusive_0, reverse = probs_cumsum_reverse_0, x = probs_fp32)[name = string("probs_cumsum")];
+            tensor<int32, [1]> probs_sum_indices_0 = const()[name = string("probs_sum_indices_0"), val = tensor<int32, [1]>([32767])];
+            int32 probs_sum_axis_0 = const()[name = string("probs_sum_axis_0"), val = int32(1)];
+            int32 probs_sum_batch_dims_0 = const()[name = string("probs_sum_batch_dims_0"), val = int32(0)];
+            bool probs_sum_validate_indices_0 = const()[name = string("probs_sum_validate_indices_0"), val = bool(false)];
+            tensor<fp32, [1, 1, 1, 1]> probs_sum = gather(axis = probs_sum_axis_0, batch_dims = probs_sum_batch_dims_0, indices = probs_sum_indices_0, validate_indices = probs_sum_validate_indices_0, x = probs_cumsum)[name = string("probs_sum")];
+            tensor<fp32, [1, 1, 1, 1]> random_number_scaled = mul(x = random_number, y = probs_sum)[name = string("random_number_scaled")];
+            tensor<bool, [1, 32768, 1, 1]> probs_greater = greater(x = probs_cumsum, y = random_number_scaled)[name = string("probs_greater")];
+            string probs_greater_int32_dtype_0 = const()[name = string("probs_greater_int32_dtype_0"), val = string("int32")];
+            int32 sampled_index_axis_0 = const()[name = string("sampled_index_axis_0"), val = int32(1)];
+            bool sampled_index_keep_dims_0 = const()[name = string("sampled_index_keep_dims_0"), val = bool(true)];
+            string sampled_index_output_dtype_0 = const()[name = string("sampled_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 32768, 1, 1]> probs_greater_int32 = cast(dtype = probs_greater_int32_dtype_0, x = probs_greater)[name = string("cast_50")];
+            tensor<int32, [1, 1, 1, 1]> sampled_index = reduce_argmax(axis = sampled_index_axis_0, keep_dims = sampled_index_keep_dims_0, output_dtype = sampled_index_output_dtype_0, x = probs_greater_int32)[name = string("sampled_index")];
+            int32 sampled_index_probability_axis_0 = const()[name = string("sampled_index_probability_axis_0"), val = int32(1)];
+            bool sampled_index_probability_validate_indices_0 = const()[name = string("sampled_index_probability_validate_indices_0"), val = bool(false)];
+            tensor<fp32, [1, 1, 1, 1]> sampled_index_probability = gather_along_axis(axis = sampled_index_probability_axis_0, indices = sampled_index, validate_indices = sampled_index_probability_validate_indices_0, x = probs_fp32)[name = string("sampled_index_probability")];
+            int32 max_logit_index_axis_0 = const()[name = string("max_logit_index_axis_0"), val = int32(1)];
+            bool max_logit_index_keep_dims_0 = const()[name = string("max_logit_index_keep_dims_0"), val = bool(true)];
+            string max_logit_index_output_dtype_0 = const()[name = string("max_logit_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 1, 1, 1]> max_logit_index = reduce_argmax(axis = max_logit_index_axis_0, keep_dims = max_logit_index_keep_dims_0, output_dtype = max_logit_index_output_dtype_0, x = logits_max_logits_chunks)[name = string("max_logit_index")];
+            string indices_chunk_0_int32_dtype_0 = const()[name = string("indices_chunk_0_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_1_int32_dtype_0 = const()[name = string("indices_chunk_1_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_2_int32_dtype_0 = const()[name = string("indices_chunk_2_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_3_int32_dtype_0 = const()[name = string("indices_chunk_3_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_4_int32_dtype_0 = const()[name = string("indices_chunk_4_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_5_int32_dtype_0 = const()[name = string("indices_chunk_5_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_6_int32_dtype_0 = const()[name = string("indices_chunk_6_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_7_int32_dtype_0 = const()[name = string("indices_chunk_7_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_8_int32_dtype_0 = const()[name = string("indices_chunk_8_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_9_int32_dtype_0 = const()[name = string("indices_chunk_9_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_10_int32_dtype_0 = const()[name = string("indices_chunk_10_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_11_int32_dtype_0 = const()[name = string("indices_chunk_11_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_12_int32_dtype_0 = const()[name = string("indices_chunk_12_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_13_int32_dtype_0 = const()[name = string("indices_chunk_13_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_14_int32_dtype_0 = const()[name = string("indices_chunk_14_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_15_int32_dtype_0 = const()[name = string("indices_chunk_15_int32_dtype_0"), val = string("int32")];
+            int32 indices_axis_0 = const()[name = string("indices_axis_0"), val = int32(1)];
+            bool indices_interleave_0 = const()[name = string("indices_interleave_0"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_15_int32 = cast(dtype = indices_chunk_15_int32_dtype_0, x = logits_chunk_15_argmax)[name = string("cast_34")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_14_int32 = cast(dtype = indices_chunk_14_int32_dtype_0, x = logits_chunk_14_argmax)[name = string("cast_35")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_13_int32 = cast(dtype = indices_chunk_13_int32_dtype_0, x = logits_chunk_13_argmax)[name = string("cast_36")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_12_int32 = cast(dtype = indices_chunk_12_int32_dtype_0, x = logits_chunk_12_argmax)[name = string("cast_37")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_11_int32 = cast(dtype = indices_chunk_11_int32_dtype_0, x = logits_chunk_11_argmax)[name = string("cast_38")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_10_int32 = cast(dtype = indices_chunk_10_int32_dtype_0, x = logits_chunk_10_argmax)[name = string("cast_39")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_9_int32 = cast(dtype = indices_chunk_9_int32_dtype_0, x = logits_chunk_9_argmax)[name = string("cast_40")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_8_int32 = cast(dtype = indices_chunk_8_int32_dtype_0, x = logits_chunk_8_argmax)[name = string("cast_41")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_7_int32 = cast(dtype = indices_chunk_7_int32_dtype_0, x = logits_chunk_7_argmax)[name = string("cast_42")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_6_int32 = cast(dtype = indices_chunk_6_int32_dtype_0, x = logits_chunk_6_argmax)[name = string("cast_43")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_5_int32 = cast(dtype = indices_chunk_5_int32_dtype_0, x = logits_chunk_5_argmax)[name = string("cast_44")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_4_int32 = cast(dtype = indices_chunk_4_int32_dtype_0, x = logits_chunk_4_argmax)[name = string("cast_45")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_3_int32 = cast(dtype = indices_chunk_3_int32_dtype_0, x = logits_chunk_3_argmax)[name = string("cast_46")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_2_int32 = cast(dtype = indices_chunk_2_int32_dtype_0, x = logits_chunk_2_argmax)[name = string("cast_47")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_1_int32 = cast(dtype = indices_chunk_1_int32_dtype_0, x = logits_chunk_1_argmax)[name = string("cast_48")];
+            tensor<int32, [1, 1, 1, 1]> indices_chunk_0_int32 = cast(dtype = indices_chunk_0_int32_dtype_0, x = logits_chunk_0_argmax)[name = string("cast_49")];
+            tensor<int32, [1, 16, 1, 1]> indices = concat(axis = indices_axis_0, interleave = indices_interleave_0, values = (indices_chunk_0_int32, indices_chunk_1_int32, indices_chunk_2_int32, indices_chunk_3_int32, indices_chunk_4_int32, indices_chunk_5_int32, indices_chunk_6_int32, indices_chunk_7_int32, indices_chunk_8_int32, indices_chunk_9_int32, indices_chunk_10_int32, indices_chunk_11_int32, indices_chunk_12_int32, indices_chunk_13_int32, indices_chunk_14_int32, indices_chunk_15_int32))[name = string("indices")];
+            int32 argmax_chunks_axis_0 = const()[name = string("argmax_chunks_axis_0"), val = int32(1)];
+            bool argmax_chunks_validate_indices_0 = const()[name = string("argmax_chunks_validate_indices_0"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 1]> argmax_chunks = gather_along_axis(axis = argmax_chunks_axis_0, indices = max_logit_index, validate_indices = argmax_chunks_validate_indices_0, x = indices)[name = string("argmax_chunks")];
+            int32 mul_0_x_0 = const()[name = string("mul_0_x_0"), val = int32(2048)];
+            tensor<int32, [1, 1, 1, 1]> mul_0 = mul(x = mul_0_x_0, y = max_logit_index)[name = string("mul_0")];
+            tensor<int32, [1, 1, 1, 1]> argmax = add(x = argmax_chunks, y = mul_0)[name = string("argmax")];
+        } -> (sampled_index, sampled_index_probability, argmax, max_prob);
+    func min_p_length_128<ios18>(tensor<fp16, [1, 2048, 1, 128]> hidden_states, tensor<fp16, [1]> p, tensor<fp32, [128]> random_number, tensor<fp16, [1]> temp) {
+            tensor<fp16, [1, 2048, 1, 128]> final_norm_rmsnorm_abs = abs(x = hidden_states)[name = string("final_norm_rmsnorm_abs")];
+            tensor<int32, [1]> final_norm_rmsnorm_maxval_axes_0 = const()[name = string("final_norm_rmsnorm_maxval_axes_0"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_maxval_keep_dims_0 = const()[name = string("final_norm_rmsnorm_maxval_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> final_norm_rmsnorm_maxval = reduce_max(axes = final_norm_rmsnorm_maxval_axes_0, keep_dims = final_norm_rmsnorm_maxval_keep_dims_0, x = final_norm_rmsnorm_abs)[name = string("final_norm_rmsnorm_maxval")];
+            fp16 final_norm_rmsnorm_maxval_clipped_alpha_0 = const()[name = string("final_norm_rmsnorm_maxval_clipped_alpha_0"), val = fp16(0x1p-14)];
+            fp16 final_norm_rmsnorm_maxval_clipped_beta_0 = const()[name = string("final_norm_rmsnorm_maxval_clipped_beta_0"), val = fp16(inf)];
+            tensor<fp16, [1, 1, 1, 128]> final_norm_rmsnorm_maxval_clipped = clip(alpha = final_norm_rmsnorm_maxval_clipped_alpha_0, beta = final_norm_rmsnorm_maxval_clipped_beta_0, x = final_norm_rmsnorm_maxval)[name = string("final_norm_rmsnorm_maxval_clipped")];
+            tensor<fp16, [1, 2048, 1, 128]> final_norm_rmsnorm_scaled = real_div(x = hidden_states, y = final_norm_rmsnorm_maxval_clipped)[name = string("final_norm_rmsnorm_scaled")];
+            tensor<int32, [1]> final_norm_rmsnorm_squared_sum_axes_0 = const()[name = string("final_norm_rmsnorm_squared_sum_axes_0"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_squared_sum_keep_dims_0 = const()[name = string("final_norm_rmsnorm_squared_sum_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> final_norm_rmsnorm_squared_sum = reduce_sum_square(axes = final_norm_rmsnorm_squared_sum_axes_0, keep_dims = final_norm_rmsnorm_squared_sum_keep_dims_0, x = final_norm_rmsnorm_scaled)[name = string("final_norm_rmsnorm_squared_sum")];
+            fp16 final_norm_rmsnorm_rsqrt_epsilon_0 = const()[name = string("final_norm_rmsnorm_rsqrt_epsilon_0"), val = fp16(0x1p-14)];
+            tensor<fp16, [1, 1, 1, 128]> final_norm_rmsnorm_rsqrt = rsqrt(epsilon = final_norm_rmsnorm_rsqrt_epsilon_0, x = final_norm_rmsnorm_squared_sum)[name = string("final_norm_rmsnorm_rsqrt")];
+            fp16 final_norm_rmsnorm_dim_scaled_y_0 = const()[name = string("final_norm_rmsnorm_dim_scaled_y_0"), val = fp16(0x1.6ap+5)];
+            tensor<fp16, [1, 2048, 1, 128]> final_norm_rmsnorm_dim_scaled = mul(x = final_norm_rmsnorm_scaled, y = final_norm_rmsnorm_dim_scaled_y_0)[name = string("final_norm_rmsnorm_dim_scaled")];
+            tensor<fp16, [1, 2048, 1, 128]> final_norm_rmsnorm_normalized = mul(x = final_norm_rmsnorm_dim_scaled, y = final_norm_rmsnorm_rsqrt)[name = string("final_norm_rmsnorm_normalized")];
+            tensor<fp16, [2048, 1, 1]> final_norm_rmsnorm_y_0 = const()[name = string("final_norm_rmsnorm_y_0"), val = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 2048, 1, 128]> final_norm_rmsnorm = mul(x = final_norm_rmsnorm_normalized, y = final_norm_rmsnorm_y_0)[name = string("final_norm_rmsnorm")];
+            fp16 temp_inverse_epsilon_0 = const()[name = string("temp_inverse_epsilon_0"), val = fp16(0x0p+0)];
+            tensor<fp16, [1]> temp_inverse = inverse(epsilon = temp_inverse_epsilon_0, x = temp)[name = string("temp_inverse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_0_weight_0 = const()[name = string("logits_chunk_0_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4224)))];
+            tensor<int32, [2]> logits_chunk_0_strides_0 = const()[name = string("logits_chunk_0_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_0_pad_type_0 = const()[name = string("logits_chunk_0_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_0_pad_0 = const()[name = string("logits_chunk_0_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_0_dilations_0 = const()[name = string("logits_chunk_0_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_0_groups_0 = const()[name = string("logits_chunk_0_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_0 = conv(dilations = logits_chunk_0_dilations_0, groups = logits_chunk_0_groups_0, pad = logits_chunk_0_pad_0, pad_type = logits_chunk_0_pad_type_0, strides = logits_chunk_0_strides_0, weight = logits_chunk_0_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_0")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_0_mul = mul(x = logits_chunk_0, y = temp_inverse)[name = string("logits_chunk_0_mul")];
+            tensor<int32, [1]> logits_chunk_0_max_axes_0 = const()[name = string("logits_chunk_0_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_0_max_keep_dims_0 = const()[name = string("logits_chunk_0_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_0_max = reduce_max(axes = logits_chunk_0_max_axes_0, keep_dims = logits_chunk_0_max_keep_dims_0, x = logits_chunk_0_mul)[name = string("logits_chunk_0_max")];
+            int32 logits_chunk_0_argmax_axis_0 = const()[name = string("logits_chunk_0_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_0_argmax_keep_dims_0 = const()[name = string("logits_chunk_0_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_0_argmax_output_dtype_0 = const()[name = string("logits_chunk_0_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> logits_chunk_0_argmax = reduce_argmax(axis = logits_chunk_0_argmax_axis_0, keep_dims = logits_chunk_0_argmax_keep_dims_0, output_dtype = logits_chunk_0_argmax_output_dtype_0, x = logits_chunk_0_mul)[name = string("logits_chunk_0_argmax")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_0_sub = sub(x = logits_chunk_0_mul, y = logits_chunk_0_max)[name = string("logits_chunk_0_sub")];
+            tensor<int32, [1]> logits_chunk_0_lse_sub_axes_0 = const()[name = string("logits_chunk_0_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_0_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_0_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_0_lse_sub = reduce_log_sum_exp(axes = logits_chunk_0_lse_sub_axes_0, keep_dims = logits_chunk_0_lse_sub_keep_dims_0, x = logits_chunk_0_sub)[name = string("logits_chunk_0_lse_sub")];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_0_lse = add(x = logits_chunk_0_lse_sub, y = logits_chunk_0_max)[name = string("logits_chunk_0_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_1_weight_0 = const()[name = string("logits_chunk_1_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8392896)))];
+            tensor<int32, [2]> logits_chunk_1_strides_0 = const()[name = string("logits_chunk_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_1_pad_type_0 = const()[name = string("logits_chunk_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_1_pad_0 = const()[name = string("logits_chunk_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_1_dilations_0 = const()[name = string("logits_chunk_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_1_groups_0 = const()[name = string("logits_chunk_1_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_1 = conv(dilations = logits_chunk_1_dilations_0, groups = logits_chunk_1_groups_0, pad = logits_chunk_1_pad_0, pad_type = logits_chunk_1_pad_type_0, strides = logits_chunk_1_strides_0, weight = logits_chunk_1_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_1")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_1_mul = mul(x = logits_chunk_1, y = temp_inverse)[name = string("logits_chunk_1_mul")];
+            tensor<int32, [1]> logits_chunk_1_max_axes_0 = const()[name = string("logits_chunk_1_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_1_max_keep_dims_0 = const()[name = string("logits_chunk_1_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_1_max = reduce_max(axes = logits_chunk_1_max_axes_0, keep_dims = logits_chunk_1_max_keep_dims_0, x = logits_chunk_1_mul)[name = string("logits_chunk_1_max")];
+            int32 logits_chunk_1_argmax_axis_0 = const()[name = string("logits_chunk_1_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_1_argmax_keep_dims_0 = const()[name = string("logits_chunk_1_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_1_argmax_output_dtype_0 = const()[name = string("logits_chunk_1_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> logits_chunk_1_argmax = reduce_argmax(axis = logits_chunk_1_argmax_axis_0, keep_dims = logits_chunk_1_argmax_keep_dims_0, output_dtype = logits_chunk_1_argmax_output_dtype_0, x = logits_chunk_1_mul)[name = string("logits_chunk_1_argmax")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_1_sub = sub(x = logits_chunk_1_mul, y = logits_chunk_1_max)[name = string("logits_chunk_1_sub")];
+            tensor<int32, [1]> logits_chunk_1_lse_sub_axes_0 = const()[name = string("logits_chunk_1_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_1_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_1_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_1_lse_sub = reduce_log_sum_exp(axes = logits_chunk_1_lse_sub_axes_0, keep_dims = logits_chunk_1_lse_sub_keep_dims_0, x = logits_chunk_1_sub)[name = string("logits_chunk_1_lse_sub")];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_1_lse = add(x = logits_chunk_1_lse_sub, y = logits_chunk_1_max)[name = string("logits_chunk_1_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_2_weight_0 = const()[name = string("logits_chunk_2_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16781568)))];
+            tensor<int32, [2]> logits_chunk_2_strides_0 = const()[name = string("logits_chunk_2_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_2_pad_type_0 = const()[name = string("logits_chunk_2_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_2_pad_0 = const()[name = string("logits_chunk_2_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_2_dilations_0 = const()[name = string("logits_chunk_2_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_2_groups_0 = const()[name = string("logits_chunk_2_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_2 = conv(dilations = logits_chunk_2_dilations_0, groups = logits_chunk_2_groups_0, pad = logits_chunk_2_pad_0, pad_type = logits_chunk_2_pad_type_0, strides = logits_chunk_2_strides_0, weight = logits_chunk_2_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_2")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_2_mul = mul(x = logits_chunk_2, y = temp_inverse)[name = string("logits_chunk_2_mul")];
+            tensor<int32, [1]> logits_chunk_2_max_axes_0 = const()[name = string("logits_chunk_2_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_2_max_keep_dims_0 = const()[name = string("logits_chunk_2_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_2_max = reduce_max(axes = logits_chunk_2_max_axes_0, keep_dims = logits_chunk_2_max_keep_dims_0, x = logits_chunk_2_mul)[name = string("logits_chunk_2_max")];
+            int32 logits_chunk_2_argmax_axis_0 = const()[name = string("logits_chunk_2_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_2_argmax_keep_dims_0 = const()[name = string("logits_chunk_2_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_2_argmax_output_dtype_0 = const()[name = string("logits_chunk_2_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> logits_chunk_2_argmax = reduce_argmax(axis = logits_chunk_2_argmax_axis_0, keep_dims = logits_chunk_2_argmax_keep_dims_0, output_dtype = logits_chunk_2_argmax_output_dtype_0, x = logits_chunk_2_mul)[name = string("logits_chunk_2_argmax")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_2_sub = sub(x = logits_chunk_2_mul, y = logits_chunk_2_max)[name = string("logits_chunk_2_sub")];
+            tensor<int32, [1]> logits_chunk_2_lse_sub_axes_0 = const()[name = string("logits_chunk_2_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_2_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_2_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_2_lse_sub = reduce_log_sum_exp(axes = logits_chunk_2_lse_sub_axes_0, keep_dims = logits_chunk_2_lse_sub_keep_dims_0, x = logits_chunk_2_sub)[name = string("logits_chunk_2_lse_sub")];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_2_lse = add(x = logits_chunk_2_lse_sub, y = logits_chunk_2_max)[name = string("logits_chunk_2_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_3_weight_0 = const()[name = string("logits_chunk_3_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25170240)))];
+            tensor<int32, [2]> logits_chunk_3_strides_0 = const()[name = string("logits_chunk_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_3_pad_type_0 = const()[name = string("logits_chunk_3_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_3_pad_0 = const()[name = string("logits_chunk_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_3_dilations_0 = const()[name = string("logits_chunk_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_3_groups_0 = const()[name = string("logits_chunk_3_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_3 = conv(dilations = logits_chunk_3_dilations_0, groups = logits_chunk_3_groups_0, pad = logits_chunk_3_pad_0, pad_type = logits_chunk_3_pad_type_0, strides = logits_chunk_3_strides_0, weight = logits_chunk_3_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_3")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_3_mul = mul(x = logits_chunk_3, y = temp_inverse)[name = string("logits_chunk_3_mul")];
+            tensor<int32, [1]> logits_chunk_3_max_axes_0 = const()[name = string("logits_chunk_3_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_3_max_keep_dims_0 = const()[name = string("logits_chunk_3_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_3_max = reduce_max(axes = logits_chunk_3_max_axes_0, keep_dims = logits_chunk_3_max_keep_dims_0, x = logits_chunk_3_mul)[name = string("logits_chunk_3_max")];
+            int32 logits_chunk_3_argmax_axis_0 = const()[name = string("logits_chunk_3_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_3_argmax_keep_dims_0 = const()[name = string("logits_chunk_3_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_3_argmax_output_dtype_0 = const()[name = string("logits_chunk_3_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> logits_chunk_3_argmax = reduce_argmax(axis = logits_chunk_3_argmax_axis_0, keep_dims = logits_chunk_3_argmax_keep_dims_0, output_dtype = logits_chunk_3_argmax_output_dtype_0, x = logits_chunk_3_mul)[name = string("logits_chunk_3_argmax")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_3_sub = sub(x = logits_chunk_3_mul, y = logits_chunk_3_max)[name = string("logits_chunk_3_sub")];
+            tensor<int32, [1]> logits_chunk_3_lse_sub_axes_0 = const()[name = string("logits_chunk_3_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_3_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_3_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_3_lse_sub = reduce_log_sum_exp(axes = logits_chunk_3_lse_sub_axes_0, keep_dims = logits_chunk_3_lse_sub_keep_dims_0, x = logits_chunk_3_sub)[name = string("logits_chunk_3_lse_sub")];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_3_lse = add(x = logits_chunk_3_lse_sub, y = logits_chunk_3_max)[name = string("logits_chunk_3_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_4_weight_0 = const()[name = string("logits_chunk_4_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33558912)))];
+            tensor<int32, [2]> logits_chunk_4_strides_0 = const()[name = string("logits_chunk_4_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_4_pad_type_0 = const()[name = string("logits_chunk_4_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_4_pad_0 = const()[name = string("logits_chunk_4_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_4_dilations_0 = const()[name = string("logits_chunk_4_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_4_groups_0 = const()[name = string("logits_chunk_4_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_4 = conv(dilations = logits_chunk_4_dilations_0, groups = logits_chunk_4_groups_0, pad = logits_chunk_4_pad_0, pad_type = logits_chunk_4_pad_type_0, strides = logits_chunk_4_strides_0, weight = logits_chunk_4_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_4")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_4_mul = mul(x = logits_chunk_4, y = temp_inverse)[name = string("logits_chunk_4_mul")];
+            tensor<int32, [1]> logits_chunk_4_max_axes_0 = const()[name = string("logits_chunk_4_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_4_max_keep_dims_0 = const()[name = string("logits_chunk_4_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_4_max = reduce_max(axes = logits_chunk_4_max_axes_0, keep_dims = logits_chunk_4_max_keep_dims_0, x = logits_chunk_4_mul)[name = string("logits_chunk_4_max")];
+            int32 logits_chunk_4_argmax_axis_0 = const()[name = string("logits_chunk_4_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_4_argmax_keep_dims_0 = const()[name = string("logits_chunk_4_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_4_argmax_output_dtype_0 = const()[name = string("logits_chunk_4_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> logits_chunk_4_argmax = reduce_argmax(axis = logits_chunk_4_argmax_axis_0, keep_dims = logits_chunk_4_argmax_keep_dims_0, output_dtype = logits_chunk_4_argmax_output_dtype_0, x = logits_chunk_4_mul)[name = string("logits_chunk_4_argmax")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_4_sub = sub(x = logits_chunk_4_mul, y = logits_chunk_4_max)[name = string("logits_chunk_4_sub")];
+            tensor<int32, [1]> logits_chunk_4_lse_sub_axes_0 = const()[name = string("logits_chunk_4_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_4_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_4_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_4_lse_sub = reduce_log_sum_exp(axes = logits_chunk_4_lse_sub_axes_0, keep_dims = logits_chunk_4_lse_sub_keep_dims_0, x = logits_chunk_4_sub)[name = string("logits_chunk_4_lse_sub")];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_4_lse = add(x = logits_chunk_4_lse_sub, y = logits_chunk_4_max)[name = string("logits_chunk_4_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_5_weight_0 = const()[name = string("logits_chunk_5_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41947584)))];
+            tensor<int32, [2]> logits_chunk_5_strides_0 = const()[name = string("logits_chunk_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_5_pad_type_0 = const()[name = string("logits_chunk_5_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_5_pad_0 = const()[name = string("logits_chunk_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_5_dilations_0 = const()[name = string("logits_chunk_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_5_groups_0 = const()[name = string("logits_chunk_5_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_5 = conv(dilations = logits_chunk_5_dilations_0, groups = logits_chunk_5_groups_0, pad = logits_chunk_5_pad_0, pad_type = logits_chunk_5_pad_type_0, strides = logits_chunk_5_strides_0, weight = logits_chunk_5_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_5")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_5_mul = mul(x = logits_chunk_5, y = temp_inverse)[name = string("logits_chunk_5_mul")];
+            tensor<int32, [1]> logits_chunk_5_max_axes_0 = const()[name = string("logits_chunk_5_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_5_max_keep_dims_0 = const()[name = string("logits_chunk_5_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_5_max = reduce_max(axes = logits_chunk_5_max_axes_0, keep_dims = logits_chunk_5_max_keep_dims_0, x = logits_chunk_5_mul)[name = string("logits_chunk_5_max")];
+            int32 logits_chunk_5_argmax_axis_0 = const()[name = string("logits_chunk_5_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_5_argmax_keep_dims_0 = const()[name = string("logits_chunk_5_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_5_argmax_output_dtype_0 = const()[name = string("logits_chunk_5_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> logits_chunk_5_argmax = reduce_argmax(axis = logits_chunk_5_argmax_axis_0, keep_dims = logits_chunk_5_argmax_keep_dims_0, output_dtype = logits_chunk_5_argmax_output_dtype_0, x = logits_chunk_5_mul)[name = string("logits_chunk_5_argmax")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_5_sub = sub(x = logits_chunk_5_mul, y = logits_chunk_5_max)[name = string("logits_chunk_5_sub")];
+            tensor<int32, [1]> logits_chunk_5_lse_sub_axes_0 = const()[name = string("logits_chunk_5_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_5_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_5_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_5_lse_sub = reduce_log_sum_exp(axes = logits_chunk_5_lse_sub_axes_0, keep_dims = logits_chunk_5_lse_sub_keep_dims_0, x = logits_chunk_5_sub)[name = string("logits_chunk_5_lse_sub")];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_5_lse = add(x = logits_chunk_5_lse_sub, y = logits_chunk_5_max)[name = string("logits_chunk_5_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_6_weight_0 = const()[name = string("logits_chunk_6_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50336256)))];
+            tensor<int32, [2]> logits_chunk_6_strides_0 = const()[name = string("logits_chunk_6_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_6_pad_type_0 = const()[name = string("logits_chunk_6_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_6_pad_0 = const()[name = string("logits_chunk_6_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_6_dilations_0 = const()[name = string("logits_chunk_6_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_6_groups_0 = const()[name = string("logits_chunk_6_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_6 = conv(dilations = logits_chunk_6_dilations_0, groups = logits_chunk_6_groups_0, pad = logits_chunk_6_pad_0, pad_type = logits_chunk_6_pad_type_0, strides = logits_chunk_6_strides_0, weight = logits_chunk_6_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_6")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_6_mul = mul(x = logits_chunk_6, y = temp_inverse)[name = string("logits_chunk_6_mul")];
+            tensor<int32, [1]> logits_chunk_6_max_axes_0 = const()[name = string("logits_chunk_6_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_6_max_keep_dims_0 = const()[name = string("logits_chunk_6_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_6_max = reduce_max(axes = logits_chunk_6_max_axes_0, keep_dims = logits_chunk_6_max_keep_dims_0, x = logits_chunk_6_mul)[name = string("logits_chunk_6_max")];
+            int32 logits_chunk_6_argmax_axis_0 = const()[name = string("logits_chunk_6_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_6_argmax_keep_dims_0 = const()[name = string("logits_chunk_6_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_6_argmax_output_dtype_0 = const()[name = string("logits_chunk_6_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> logits_chunk_6_argmax = reduce_argmax(axis = logits_chunk_6_argmax_axis_0, keep_dims = logits_chunk_6_argmax_keep_dims_0, output_dtype = logits_chunk_6_argmax_output_dtype_0, x = logits_chunk_6_mul)[name = string("logits_chunk_6_argmax")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_6_sub = sub(x = logits_chunk_6_mul, y = logits_chunk_6_max)[name = string("logits_chunk_6_sub")];
+            tensor<int32, [1]> logits_chunk_6_lse_sub_axes_0 = const()[name = string("logits_chunk_6_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_6_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_6_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_6_lse_sub = reduce_log_sum_exp(axes = logits_chunk_6_lse_sub_axes_0, keep_dims = logits_chunk_6_lse_sub_keep_dims_0, x = logits_chunk_6_sub)[name = string("logits_chunk_6_lse_sub")];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_6_lse = add(x = logits_chunk_6_lse_sub, y = logits_chunk_6_max)[name = string("logits_chunk_6_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_7_weight_0 = const()[name = string("logits_chunk_7_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58724928)))];
+            tensor<int32, [2]> logits_chunk_7_strides_0 = const()[name = string("logits_chunk_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_7_pad_type_0 = const()[name = string("logits_chunk_7_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_7_pad_0 = const()[name = string("logits_chunk_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_7_dilations_0 = const()[name = string("logits_chunk_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_7_groups_0 = const()[name = string("logits_chunk_7_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_7 = conv(dilations = logits_chunk_7_dilations_0, groups = logits_chunk_7_groups_0, pad = logits_chunk_7_pad_0, pad_type = logits_chunk_7_pad_type_0, strides = logits_chunk_7_strides_0, weight = logits_chunk_7_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_7")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_7_mul = mul(x = logits_chunk_7, y = temp_inverse)[name = string("logits_chunk_7_mul")];
+            tensor<int32, [1]> logits_chunk_7_max_axes_0 = const()[name = string("logits_chunk_7_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_7_max_keep_dims_0 = const()[name = string("logits_chunk_7_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_7_max = reduce_max(axes = logits_chunk_7_max_axes_0, keep_dims = logits_chunk_7_max_keep_dims_0, x = logits_chunk_7_mul)[name = string("logits_chunk_7_max")];
+            int32 logits_chunk_7_argmax_axis_0 = const()[name = string("logits_chunk_7_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_7_argmax_keep_dims_0 = const()[name = string("logits_chunk_7_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_7_argmax_output_dtype_0 = const()[name = string("logits_chunk_7_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> logits_chunk_7_argmax = reduce_argmax(axis = logits_chunk_7_argmax_axis_0, keep_dims = logits_chunk_7_argmax_keep_dims_0, output_dtype = logits_chunk_7_argmax_output_dtype_0, x = logits_chunk_7_mul)[name = string("logits_chunk_7_argmax")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_7_sub = sub(x = logits_chunk_7_mul, y = logits_chunk_7_max)[name = string("logits_chunk_7_sub")];
+            tensor<int32, [1]> logits_chunk_7_lse_sub_axes_0 = const()[name = string("logits_chunk_7_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_7_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_7_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_7_lse_sub = reduce_log_sum_exp(axes = logits_chunk_7_lse_sub_axes_0, keep_dims = logits_chunk_7_lse_sub_keep_dims_0, x = logits_chunk_7_sub)[name = string("logits_chunk_7_lse_sub")];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_7_lse = add(x = logits_chunk_7_lse_sub, y = logits_chunk_7_max)[name = string("logits_chunk_7_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_8_weight_0 = const()[name = string("logits_chunk_8_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67113600)))];
+            tensor<int32, [2]> logits_chunk_8_strides_0 = const()[name = string("logits_chunk_8_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_8_pad_type_0 = const()[name = string("logits_chunk_8_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_8_pad_0 = const()[name = string("logits_chunk_8_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_8_dilations_0 = const()[name = string("logits_chunk_8_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_8_groups_0 = const()[name = string("logits_chunk_8_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_8 = conv(dilations = logits_chunk_8_dilations_0, groups = logits_chunk_8_groups_0, pad = logits_chunk_8_pad_0, pad_type = logits_chunk_8_pad_type_0, strides = logits_chunk_8_strides_0, weight = logits_chunk_8_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_8")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_8_mul = mul(x = logits_chunk_8, y = temp_inverse)[name = string("logits_chunk_8_mul")];
+            tensor<int32, [1]> logits_chunk_8_max_axes_0 = const()[name = string("logits_chunk_8_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_8_max_keep_dims_0 = const()[name = string("logits_chunk_8_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_8_max = reduce_max(axes = logits_chunk_8_max_axes_0, keep_dims = logits_chunk_8_max_keep_dims_0, x = logits_chunk_8_mul)[name = string("logits_chunk_8_max")];
+            int32 logits_chunk_8_argmax_axis_0 = const()[name = string("logits_chunk_8_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_8_argmax_keep_dims_0 = const()[name = string("logits_chunk_8_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_8_argmax_output_dtype_0 = const()[name = string("logits_chunk_8_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> logits_chunk_8_argmax = reduce_argmax(axis = logits_chunk_8_argmax_axis_0, keep_dims = logits_chunk_8_argmax_keep_dims_0, output_dtype = logits_chunk_8_argmax_output_dtype_0, x = logits_chunk_8_mul)[name = string("logits_chunk_8_argmax")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_8_sub = sub(x = logits_chunk_8_mul, y = logits_chunk_8_max)[name = string("logits_chunk_8_sub")];
+            tensor<int32, [1]> logits_chunk_8_lse_sub_axes_0 = const()[name = string("logits_chunk_8_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_8_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_8_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_8_lse_sub = reduce_log_sum_exp(axes = logits_chunk_8_lse_sub_axes_0, keep_dims = logits_chunk_8_lse_sub_keep_dims_0, x = logits_chunk_8_sub)[name = string("logits_chunk_8_lse_sub")];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_8_lse = add(x = logits_chunk_8_lse_sub, y = logits_chunk_8_max)[name = string("logits_chunk_8_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_9_weight_0 = const()[name = string("logits_chunk_9_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75502272)))];
+            tensor<int32, [2]> logits_chunk_9_strides_0 = const()[name = string("logits_chunk_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_9_pad_type_0 = const()[name = string("logits_chunk_9_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_9_pad_0 = const()[name = string("logits_chunk_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_9_dilations_0 = const()[name = string("logits_chunk_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_9_groups_0 = const()[name = string("logits_chunk_9_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_9 = conv(dilations = logits_chunk_9_dilations_0, groups = logits_chunk_9_groups_0, pad = logits_chunk_9_pad_0, pad_type = logits_chunk_9_pad_type_0, strides = logits_chunk_9_strides_0, weight = logits_chunk_9_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_9")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_9_mul = mul(x = logits_chunk_9, y = temp_inverse)[name = string("logits_chunk_9_mul")];
+            tensor<int32, [1]> logits_chunk_9_max_axes_0 = const()[name = string("logits_chunk_9_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_9_max_keep_dims_0 = const()[name = string("logits_chunk_9_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_9_max = reduce_max(axes = logits_chunk_9_max_axes_0, keep_dims = logits_chunk_9_max_keep_dims_0, x = logits_chunk_9_mul)[name = string("logits_chunk_9_max")];
+            int32 logits_chunk_9_argmax_axis_0 = const()[name = string("logits_chunk_9_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_9_argmax_keep_dims_0 = const()[name = string("logits_chunk_9_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_9_argmax_output_dtype_0 = const()[name = string("logits_chunk_9_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> logits_chunk_9_argmax = reduce_argmax(axis = logits_chunk_9_argmax_axis_0, keep_dims = logits_chunk_9_argmax_keep_dims_0, output_dtype = logits_chunk_9_argmax_output_dtype_0, x = logits_chunk_9_mul)[name = string("logits_chunk_9_argmax")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_9_sub = sub(x = logits_chunk_9_mul, y = logits_chunk_9_max)[name = string("logits_chunk_9_sub")];
+            tensor<int32, [1]> logits_chunk_9_lse_sub_axes_0 = const()[name = string("logits_chunk_9_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_9_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_9_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_9_lse_sub = reduce_log_sum_exp(axes = logits_chunk_9_lse_sub_axes_0, keep_dims = logits_chunk_9_lse_sub_keep_dims_0, x = logits_chunk_9_sub)[name = string("logits_chunk_9_lse_sub")];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_9_lse = add(x = logits_chunk_9_lse_sub, y = logits_chunk_9_max)[name = string("logits_chunk_9_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_10_weight_0 = const()[name = string("logits_chunk_10_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83890944)))];
+            tensor<int32, [2]> logits_chunk_10_strides_0 = const()[name = string("logits_chunk_10_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_10_pad_type_0 = const()[name = string("logits_chunk_10_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_10_pad_0 = const()[name = string("logits_chunk_10_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_10_dilations_0 = const()[name = string("logits_chunk_10_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_10_groups_0 = const()[name = string("logits_chunk_10_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_10 = conv(dilations = logits_chunk_10_dilations_0, groups = logits_chunk_10_groups_0, pad = logits_chunk_10_pad_0, pad_type = logits_chunk_10_pad_type_0, strides = logits_chunk_10_strides_0, weight = logits_chunk_10_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_10")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_10_mul = mul(x = logits_chunk_10, y = temp_inverse)[name = string("logits_chunk_10_mul")];
+            tensor<int32, [1]> logits_chunk_10_max_axes_0 = const()[name = string("logits_chunk_10_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_10_max_keep_dims_0 = const()[name = string("logits_chunk_10_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_10_max = reduce_max(axes = logits_chunk_10_max_axes_0, keep_dims = logits_chunk_10_max_keep_dims_0, x = logits_chunk_10_mul)[name = string("logits_chunk_10_max")];
+            int32 logits_chunk_10_argmax_axis_0 = const()[name = string("logits_chunk_10_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_10_argmax_keep_dims_0 = const()[name = string("logits_chunk_10_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_10_argmax_output_dtype_0 = const()[name = string("logits_chunk_10_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> logits_chunk_10_argmax = reduce_argmax(axis = logits_chunk_10_argmax_axis_0, keep_dims = logits_chunk_10_argmax_keep_dims_0, output_dtype = logits_chunk_10_argmax_output_dtype_0, x = logits_chunk_10_mul)[name = string("logits_chunk_10_argmax")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_10_sub = sub(x = logits_chunk_10_mul, y = logits_chunk_10_max)[name = string("logits_chunk_10_sub")];
+            tensor<int32, [1]> logits_chunk_10_lse_sub_axes_0 = const()[name = string("logits_chunk_10_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_10_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_10_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_10_lse_sub = reduce_log_sum_exp(axes = logits_chunk_10_lse_sub_axes_0, keep_dims = logits_chunk_10_lse_sub_keep_dims_0, x = logits_chunk_10_sub)[name = string("logits_chunk_10_lse_sub")];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_10_lse = add(x = logits_chunk_10_lse_sub, y = logits_chunk_10_max)[name = string("logits_chunk_10_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_11_weight_0 = const()[name = string("logits_chunk_11_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92279616)))];
+            tensor<int32, [2]> logits_chunk_11_strides_0 = const()[name = string("logits_chunk_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_11_pad_type_0 = const()[name = string("logits_chunk_11_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_11_pad_0 = const()[name = string("logits_chunk_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_11_dilations_0 = const()[name = string("logits_chunk_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_11_groups_0 = const()[name = string("logits_chunk_11_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_11 = conv(dilations = logits_chunk_11_dilations_0, groups = logits_chunk_11_groups_0, pad = logits_chunk_11_pad_0, pad_type = logits_chunk_11_pad_type_0, strides = logits_chunk_11_strides_0, weight = logits_chunk_11_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_11")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_11_mul = mul(x = logits_chunk_11, y = temp_inverse)[name = string("logits_chunk_11_mul")];
+            tensor<int32, [1]> logits_chunk_11_max_axes_0 = const()[name = string("logits_chunk_11_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_11_max_keep_dims_0 = const()[name = string("logits_chunk_11_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_11_max = reduce_max(axes = logits_chunk_11_max_axes_0, keep_dims = logits_chunk_11_max_keep_dims_0, x = logits_chunk_11_mul)[name = string("logits_chunk_11_max")];
+            int32 logits_chunk_11_argmax_axis_0 = const()[name = string("logits_chunk_11_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_11_argmax_keep_dims_0 = const()[name = string("logits_chunk_11_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_11_argmax_output_dtype_0 = const()[name = string("logits_chunk_11_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> logits_chunk_11_argmax = reduce_argmax(axis = logits_chunk_11_argmax_axis_0, keep_dims = logits_chunk_11_argmax_keep_dims_0, output_dtype = logits_chunk_11_argmax_output_dtype_0, x = logits_chunk_11_mul)[name = string("logits_chunk_11_argmax")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_11_sub = sub(x = logits_chunk_11_mul, y = logits_chunk_11_max)[name = string("logits_chunk_11_sub")];
+            tensor<int32, [1]> logits_chunk_11_lse_sub_axes_0 = const()[name = string("logits_chunk_11_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_11_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_11_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_11_lse_sub = reduce_log_sum_exp(axes = logits_chunk_11_lse_sub_axes_0, keep_dims = logits_chunk_11_lse_sub_keep_dims_0, x = logits_chunk_11_sub)[name = string("logits_chunk_11_lse_sub")];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_11_lse = add(x = logits_chunk_11_lse_sub, y = logits_chunk_11_max)[name = string("logits_chunk_11_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_12_weight_0 = const()[name = string("logits_chunk_12_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100668288)))];
+            tensor<int32, [2]> logits_chunk_12_strides_0 = const()[name = string("logits_chunk_12_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_12_pad_type_0 = const()[name = string("logits_chunk_12_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_12_pad_0 = const()[name = string("logits_chunk_12_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_12_dilations_0 = const()[name = string("logits_chunk_12_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_12_groups_0 = const()[name = string("logits_chunk_12_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_12 = conv(dilations = logits_chunk_12_dilations_0, groups = logits_chunk_12_groups_0, pad = logits_chunk_12_pad_0, pad_type = logits_chunk_12_pad_type_0, strides = logits_chunk_12_strides_0, weight = logits_chunk_12_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_12")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_12_mul = mul(x = logits_chunk_12, y = temp_inverse)[name = string("logits_chunk_12_mul")];
+            tensor<int32, [1]> logits_chunk_12_max_axes_0 = const()[name = string("logits_chunk_12_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_12_max_keep_dims_0 = const()[name = string("logits_chunk_12_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_12_max = reduce_max(axes = logits_chunk_12_max_axes_0, keep_dims = logits_chunk_12_max_keep_dims_0, x = logits_chunk_12_mul)[name = string("logits_chunk_12_max")];
+            int32 logits_chunk_12_argmax_axis_0 = const()[name = string("logits_chunk_12_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_12_argmax_keep_dims_0 = const()[name = string("logits_chunk_12_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_12_argmax_output_dtype_0 = const()[name = string("logits_chunk_12_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> logits_chunk_12_argmax = reduce_argmax(axis = logits_chunk_12_argmax_axis_0, keep_dims = logits_chunk_12_argmax_keep_dims_0, output_dtype = logits_chunk_12_argmax_output_dtype_0, x = logits_chunk_12_mul)[name = string("logits_chunk_12_argmax")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_12_sub = sub(x = logits_chunk_12_mul, y = logits_chunk_12_max)[name = string("logits_chunk_12_sub")];
+            tensor<int32, [1]> logits_chunk_12_lse_sub_axes_0 = const()[name = string("logits_chunk_12_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_12_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_12_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_12_lse_sub = reduce_log_sum_exp(axes = logits_chunk_12_lse_sub_axes_0, keep_dims = logits_chunk_12_lse_sub_keep_dims_0, x = logits_chunk_12_sub)[name = string("logits_chunk_12_lse_sub")];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_12_lse = add(x = logits_chunk_12_lse_sub, y = logits_chunk_12_max)[name = string("logits_chunk_12_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_13_weight_0 = const()[name = string("logits_chunk_13_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109056960)))];
+            tensor<int32, [2]> logits_chunk_13_strides_0 = const()[name = string("logits_chunk_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_13_pad_type_0 = const()[name = string("logits_chunk_13_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_13_pad_0 = const()[name = string("logits_chunk_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_13_dilations_0 = const()[name = string("logits_chunk_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_13_groups_0 = const()[name = string("logits_chunk_13_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_13 = conv(dilations = logits_chunk_13_dilations_0, groups = logits_chunk_13_groups_0, pad = logits_chunk_13_pad_0, pad_type = logits_chunk_13_pad_type_0, strides = logits_chunk_13_strides_0, weight = logits_chunk_13_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_13")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_13_mul = mul(x = logits_chunk_13, y = temp_inverse)[name = string("logits_chunk_13_mul")];
+            tensor<int32, [1]> logits_chunk_13_max_axes_0 = const()[name = string("logits_chunk_13_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_13_max_keep_dims_0 = const()[name = string("logits_chunk_13_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_13_max = reduce_max(axes = logits_chunk_13_max_axes_0, keep_dims = logits_chunk_13_max_keep_dims_0, x = logits_chunk_13_mul)[name = string("logits_chunk_13_max")];
+            int32 logits_chunk_13_argmax_axis_0 = const()[name = string("logits_chunk_13_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_13_argmax_keep_dims_0 = const()[name = string("logits_chunk_13_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_13_argmax_output_dtype_0 = const()[name = string("logits_chunk_13_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> logits_chunk_13_argmax = reduce_argmax(axis = logits_chunk_13_argmax_axis_0, keep_dims = logits_chunk_13_argmax_keep_dims_0, output_dtype = logits_chunk_13_argmax_output_dtype_0, x = logits_chunk_13_mul)[name = string("logits_chunk_13_argmax")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_13_sub = sub(x = logits_chunk_13_mul, y = logits_chunk_13_max)[name = string("logits_chunk_13_sub")];
+            tensor<int32, [1]> logits_chunk_13_lse_sub_axes_0 = const()[name = string("logits_chunk_13_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_13_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_13_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_13_lse_sub = reduce_log_sum_exp(axes = logits_chunk_13_lse_sub_axes_0, keep_dims = logits_chunk_13_lse_sub_keep_dims_0, x = logits_chunk_13_sub)[name = string("logits_chunk_13_lse_sub")];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_13_lse = add(x = logits_chunk_13_lse_sub, y = logits_chunk_13_max)[name = string("logits_chunk_13_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_14_weight_0 = const()[name = string("logits_chunk_14_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117445632)))];
+            tensor<int32, [2]> logits_chunk_14_strides_0 = const()[name = string("logits_chunk_14_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_14_pad_type_0 = const()[name = string("logits_chunk_14_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_14_pad_0 = const()[name = string("logits_chunk_14_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_14_dilations_0 = const()[name = string("logits_chunk_14_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_14_groups_0 = const()[name = string("logits_chunk_14_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_14 = conv(dilations = logits_chunk_14_dilations_0, groups = logits_chunk_14_groups_0, pad = logits_chunk_14_pad_0, pad_type = logits_chunk_14_pad_type_0, strides = logits_chunk_14_strides_0, weight = logits_chunk_14_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_14")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_14_mul = mul(x = logits_chunk_14, y = temp_inverse)[name = string("logits_chunk_14_mul")];
+            tensor<int32, [1]> logits_chunk_14_max_axes_0 = const()[name = string("logits_chunk_14_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_14_max_keep_dims_0 = const()[name = string("logits_chunk_14_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_14_max = reduce_max(axes = logits_chunk_14_max_axes_0, keep_dims = logits_chunk_14_max_keep_dims_0, x = logits_chunk_14_mul)[name = string("logits_chunk_14_max")];
+            int32 logits_chunk_14_argmax_axis_0 = const()[name = string("logits_chunk_14_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_14_argmax_keep_dims_0 = const()[name = string("logits_chunk_14_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_14_argmax_output_dtype_0 = const()[name = string("logits_chunk_14_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> logits_chunk_14_argmax = reduce_argmax(axis = logits_chunk_14_argmax_axis_0, keep_dims = logits_chunk_14_argmax_keep_dims_0, output_dtype = logits_chunk_14_argmax_output_dtype_0, x = logits_chunk_14_mul)[name = string("logits_chunk_14_argmax")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_14_sub = sub(x = logits_chunk_14_mul, y = logits_chunk_14_max)[name = string("logits_chunk_14_sub")];
+            tensor<int32, [1]> logits_chunk_14_lse_sub_axes_0 = const()[name = string("logits_chunk_14_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_14_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_14_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_14_lse_sub = reduce_log_sum_exp(axes = logits_chunk_14_lse_sub_axes_0, keep_dims = logits_chunk_14_lse_sub_keep_dims_0, x = logits_chunk_14_sub)[name = string("logits_chunk_14_lse_sub")];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_14_lse = add(x = logits_chunk_14_lse_sub, y = logits_chunk_14_max)[name = string("logits_chunk_14_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_15_weight_0 = const()[name = string("logits_chunk_15_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125834304)))];
+            tensor<int32, [2]> logits_chunk_15_strides_0 = const()[name = string("logits_chunk_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_15_pad_type_0 = const()[name = string("logits_chunk_15_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_15_pad_0 = const()[name = string("logits_chunk_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_15_dilations_0 = const()[name = string("logits_chunk_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_15_groups_0 = const()[name = string("logits_chunk_15_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_15 = conv(dilations = logits_chunk_15_dilations_0, groups = logits_chunk_15_groups_0, pad = logits_chunk_15_pad_0, pad_type = logits_chunk_15_pad_type_0, strides = logits_chunk_15_strides_0, weight = logits_chunk_15_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_15")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_15_mul = mul(x = logits_chunk_15, y = temp_inverse)[name = string("logits_chunk_15_mul")];
+            tensor<int32, [1]> logits_chunk_15_max_axes_0 = const()[name = string("logits_chunk_15_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_15_max_keep_dims_0 = const()[name = string("logits_chunk_15_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_15_max = reduce_max(axes = logits_chunk_15_max_axes_0, keep_dims = logits_chunk_15_max_keep_dims_0, x = logits_chunk_15_mul)[name = string("logits_chunk_15_max")];
+            int32 logits_chunk_15_argmax_axis_0 = const()[name = string("logits_chunk_15_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_15_argmax_keep_dims_0 = const()[name = string("logits_chunk_15_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_15_argmax_output_dtype_0 = const()[name = string("logits_chunk_15_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 128]> logits_chunk_15_argmax = reduce_argmax(axis = logits_chunk_15_argmax_axis_0, keep_dims = logits_chunk_15_argmax_keep_dims_0, output_dtype = logits_chunk_15_argmax_output_dtype_0, x = logits_chunk_15_mul)[name = string("logits_chunk_15_argmax")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_15_sub = sub(x = logits_chunk_15_mul, y = logits_chunk_15_max)[name = string("logits_chunk_15_sub")];
+            tensor<int32, [1]> logits_chunk_15_lse_sub_axes_0 = const()[name = string("logits_chunk_15_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_15_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_15_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_15_lse_sub = reduce_log_sum_exp(axes = logits_chunk_15_lse_sub_axes_0, keep_dims = logits_chunk_15_lse_sub_keep_dims_0, x = logits_chunk_15_sub)[name = string("logits_chunk_15_lse_sub")];
+            tensor<fp16, [1, 1, 1, 128]> logits_chunk_15_lse = add(x = logits_chunk_15_lse_sub, y = logits_chunk_15_max)[name = string("logits_chunk_15_lse")];
+            int32 logits_lses_axis_0 = const()[name = string("logits_lses_axis_0"), val = int32(1)];
+            bool logits_lses_interleave_0 = const()[name = string("logits_lses_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 128]> logits_lses = concat(axis = logits_lses_axis_0, interleave = logits_lses_interleave_0, values = (logits_chunk_0_lse, logits_chunk_1_lse, logits_chunk_2_lse, logits_chunk_3_lse, logits_chunk_4_lse, logits_chunk_5_lse, logits_chunk_6_lse, logits_chunk_7_lse, logits_chunk_8_lse, logits_chunk_9_lse, logits_chunk_10_lse, logits_chunk_11_lse, logits_chunk_12_lse, logits_chunk_13_lse, logits_chunk_14_lse, logits_chunk_15_lse))[name = string("logits_lses")];
+            tensor<int32, [1]> logits_lses_max_axes_0 = const()[name = string("logits_lses_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_lses_max_keep_dims_0 = const()[name = string("logits_lses_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_lses_max = reduce_max(axes = logits_lses_max_axes_0, keep_dims = logits_lses_max_keep_dims_0, x = logits_lses)[name = string("logits_lses_max")];
+            tensor<fp16, [1, 16, 1, 128]> logits_lses_sub = sub(x = logits_lses, y = logits_lses_max)[name = string("logits_lses_sub")];
+            tensor<int32, [1]> logits_lses_logsumexp_axes_0 = const()[name = string("logits_lses_logsumexp_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_lses_logsumexp_keep_dims_0 = const()[name = string("logits_lses_logsumexp_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_lses_logsumexp = reduce_log_sum_exp(axes = logits_lses_logsumexp_axes_0, keep_dims = logits_lses_logsumexp_keep_dims_0, x = logits_lses_sub)[name = string("logits_lses_logsumexp")];
+            tensor<fp16, [1, 1, 1, 128]> logits_lse = add(x = logits_lses_logsumexp, y = logits_lses_max)[name = string("logits_lse")];
+            int32 logits_max_logits_chunks_axis_0 = const()[name = string("logits_max_logits_chunks_axis_0"), val = int32(1)];
+            bool logits_max_logits_chunks_interleave_0 = const()[name = string("logits_max_logits_chunks_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 128]> logits_max_logits_chunks = concat(axis = logits_max_logits_chunks_axis_0, interleave = logits_max_logits_chunks_interleave_0, values = (logits_chunk_0_max, logits_chunk_1_max, logits_chunk_2_max, logits_chunk_3_max, logits_chunk_4_max, logits_chunk_5_max, logits_chunk_6_max, logits_chunk_7_max, logits_chunk_8_max, logits_chunk_9_max, logits_chunk_10_max, logits_chunk_11_max, logits_chunk_12_max, logits_chunk_13_max, logits_chunk_14_max, logits_chunk_15_max))[name = string("logits_max_logits_chunks")];
+            tensor<int32, [1]> logits_max_logit_axes_0 = const()[name = string("logits_max_logit_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_max_logit_keep_dims_0 = const()[name = string("logits_max_logit_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 128]> logits_max_logit = reduce_max(axes = logits_max_logit_axes_0, keep_dims = logits_max_logit_keep_dims_0, x = logits_max_logits_chunks)[name = string("logits_max_logit")];
+            tensor<fp16, [1, 1, 1, 128]> logits_max_logit_sub = sub(x = logits_max_logit, y = logits_lse)[name = string("logits_max_logit_sub")];
+            tensor<fp16, [1, 1, 1, 128]> max_prob = exp(x = logits_max_logit_sub)[name = string("max_prob")];
+            tensor<fp16, [1, 1, 1, 128]> min_p_thresh = mul(x = max_prob, y = p)[name = string("min_p_thresh")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_0_sub_1 = sub(x = logits_chunk_0_mul, y = logits_lse)[name = string("logits_chunk_0_sub_1")];
+            tensor<fp16, [1, 2048, 1, 128]> probs_chunk_0 = exp(x = logits_chunk_0_sub_1)[name = string("probs_chunk_0")];
+            tensor<bool, [1, 2048, 1, 128]> mask_probs_chunk_0 = greater_equal(x = probs_chunk_0, y = min_p_thresh)[name = string("mask_probs_chunk_0")];
+            string mask_chunk_0_fp16_dtype_0 = const()[name = string("mask_chunk_0_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 128]> mask_chunk_0_fp16 = cast(dtype = mask_chunk_0_fp16_dtype_0, x = mask_probs_chunk_0)[name = string("cast_271")];
+            tensor<fp16, [1, 2048, 1, 128]> masked_probs_chunk_0 = select(a = probs_chunk_0, b = mask_chunk_0_fp16, cond = mask_probs_chunk_0)[name = string("masked_probs_chunk_0")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_1_sub_1 = sub(x = logits_chunk_1_mul, y = logits_lse)[name = string("logits_chunk_1_sub_1")];
+            tensor<fp16, [1, 2048, 1, 128]> probs_chunk_1 = exp(x = logits_chunk_1_sub_1)[name = string("probs_chunk_1")];
+            tensor<bool, [1, 2048, 1, 128]> mask_probs_chunk_1 = greater_equal(x = probs_chunk_1, y = min_p_thresh)[name = string("mask_probs_chunk_1")];
+            string mask_chunk_1_fp16_dtype_0 = const()[name = string("mask_chunk_1_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 128]> mask_chunk_1_fp16 = cast(dtype = mask_chunk_1_fp16_dtype_0, x = mask_probs_chunk_1)[name = string("cast_270")];
+            tensor<fp16, [1, 2048, 1, 128]> masked_probs_chunk_1 = select(a = probs_chunk_1, b = mask_chunk_1_fp16, cond = mask_probs_chunk_1)[name = string("masked_probs_chunk_1")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_2_sub_1 = sub(x = logits_chunk_2_mul, y = logits_lse)[name = string("logits_chunk_2_sub_1")];
+            tensor<fp16, [1, 2048, 1, 128]> probs_chunk_2 = exp(x = logits_chunk_2_sub_1)[name = string("probs_chunk_2")];
+            tensor<bool, [1, 2048, 1, 128]> mask_probs_chunk_2 = greater_equal(x = probs_chunk_2, y = min_p_thresh)[name = string("mask_probs_chunk_2")];
+            string mask_chunk_2_fp16_dtype_0 = const()[name = string("mask_chunk_2_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 128]> mask_chunk_2_fp16 = cast(dtype = mask_chunk_2_fp16_dtype_0, x = mask_probs_chunk_2)[name = string("cast_269")];
+            tensor<fp16, [1, 2048, 1, 128]> masked_probs_chunk_2 = select(a = probs_chunk_2, b = mask_chunk_2_fp16, cond = mask_probs_chunk_2)[name = string("masked_probs_chunk_2")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_3_sub_1 = sub(x = logits_chunk_3_mul, y = logits_lse)[name = string("logits_chunk_3_sub_1")];
+            tensor<fp16, [1, 2048, 1, 128]> probs_chunk_3 = exp(x = logits_chunk_3_sub_1)[name = string("probs_chunk_3")];
+            tensor<bool, [1, 2048, 1, 128]> mask_probs_chunk_3 = greater_equal(x = probs_chunk_3, y = min_p_thresh)[name = string("mask_probs_chunk_3")];
+            string mask_chunk_3_fp16_dtype_0 = const()[name = string("mask_chunk_3_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 128]> mask_chunk_3_fp16 = cast(dtype = mask_chunk_3_fp16_dtype_0, x = mask_probs_chunk_3)[name = string("cast_268")];
+            tensor<fp16, [1, 2048, 1, 128]> masked_probs_chunk_3 = select(a = probs_chunk_3, b = mask_chunk_3_fp16, cond = mask_probs_chunk_3)[name = string("masked_probs_chunk_3")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_4_sub_1 = sub(x = logits_chunk_4_mul, y = logits_lse)[name = string("logits_chunk_4_sub_1")];
+            tensor<fp16, [1, 2048, 1, 128]> probs_chunk_4 = exp(x = logits_chunk_4_sub_1)[name = string("probs_chunk_4")];
+            tensor<bool, [1, 2048, 1, 128]> mask_probs_chunk_4 = greater_equal(x = probs_chunk_4, y = min_p_thresh)[name = string("mask_probs_chunk_4")];
+            string mask_chunk_4_fp16_dtype_0 = const()[name = string("mask_chunk_4_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 128]> mask_chunk_4_fp16 = cast(dtype = mask_chunk_4_fp16_dtype_0, x = mask_probs_chunk_4)[name = string("cast_267")];
+            tensor<fp16, [1, 2048, 1, 128]> masked_probs_chunk_4 = select(a = probs_chunk_4, b = mask_chunk_4_fp16, cond = mask_probs_chunk_4)[name = string("masked_probs_chunk_4")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_5_sub_1 = sub(x = logits_chunk_5_mul, y = logits_lse)[name = string("logits_chunk_5_sub_1")];
+            tensor<fp16, [1, 2048, 1, 128]> probs_chunk_5 = exp(x = logits_chunk_5_sub_1)[name = string("probs_chunk_5")];
+            tensor<bool, [1, 2048, 1, 128]> mask_probs_chunk_5 = greater_equal(x = probs_chunk_5, y = min_p_thresh)[name = string("mask_probs_chunk_5")];
+            string mask_chunk_5_fp16_dtype_0 = const()[name = string("mask_chunk_5_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 128]> mask_chunk_5_fp16 = cast(dtype = mask_chunk_5_fp16_dtype_0, x = mask_probs_chunk_5)[name = string("cast_266")];
+            tensor<fp16, [1, 2048, 1, 128]> masked_probs_chunk_5 = select(a = probs_chunk_5, b = mask_chunk_5_fp16, cond = mask_probs_chunk_5)[name = string("masked_probs_chunk_5")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_6_sub_1 = sub(x = logits_chunk_6_mul, y = logits_lse)[name = string("logits_chunk_6_sub_1")];
+            tensor<fp16, [1, 2048, 1, 128]> probs_chunk_6 = exp(x = logits_chunk_6_sub_1)[name = string("probs_chunk_6")];
+            tensor<bool, [1, 2048, 1, 128]> mask_probs_chunk_6 = greater_equal(x = probs_chunk_6, y = min_p_thresh)[name = string("mask_probs_chunk_6")];
+            string mask_chunk_6_fp16_dtype_0 = const()[name = string("mask_chunk_6_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 128]> mask_chunk_6_fp16 = cast(dtype = mask_chunk_6_fp16_dtype_0, x = mask_probs_chunk_6)[name = string("cast_265")];
+            tensor<fp16, [1, 2048, 1, 128]> masked_probs_chunk_6 = select(a = probs_chunk_6, b = mask_chunk_6_fp16, cond = mask_probs_chunk_6)[name = string("masked_probs_chunk_6")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_7_sub_1 = sub(x = logits_chunk_7_mul, y = logits_lse)[name = string("logits_chunk_7_sub_1")];
+            tensor<fp16, [1, 2048, 1, 128]> probs_chunk_7 = exp(x = logits_chunk_7_sub_1)[name = string("probs_chunk_7")];
+            tensor<bool, [1, 2048, 1, 128]> mask_probs_chunk_7 = greater_equal(x = probs_chunk_7, y = min_p_thresh)[name = string("mask_probs_chunk_7")];
+            string mask_chunk_7_fp16_dtype_0 = const()[name = string("mask_chunk_7_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 128]> mask_chunk_7_fp16 = cast(dtype = mask_chunk_7_fp16_dtype_0, x = mask_probs_chunk_7)[name = string("cast_264")];
+            tensor<fp16, [1, 2048, 1, 128]> masked_probs_chunk_7 = select(a = probs_chunk_7, b = mask_chunk_7_fp16, cond = mask_probs_chunk_7)[name = string("masked_probs_chunk_7")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_8_sub_1 = sub(x = logits_chunk_8_mul, y = logits_lse)[name = string("logits_chunk_8_sub_1")];
+            tensor<fp16, [1, 2048, 1, 128]> probs_chunk_8 = exp(x = logits_chunk_8_sub_1)[name = string("probs_chunk_8")];
+            tensor<bool, [1, 2048, 1, 128]> mask_probs_chunk_8 = greater_equal(x = probs_chunk_8, y = min_p_thresh)[name = string("mask_probs_chunk_8")];
+            string mask_chunk_8_fp16_dtype_0 = const()[name = string("mask_chunk_8_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 128]> mask_chunk_8_fp16 = cast(dtype = mask_chunk_8_fp16_dtype_0, x = mask_probs_chunk_8)[name = string("cast_263")];
+            tensor<fp16, [1, 2048, 1, 128]> masked_probs_chunk_8 = select(a = probs_chunk_8, b = mask_chunk_8_fp16, cond = mask_probs_chunk_8)[name = string("masked_probs_chunk_8")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_9_sub_1 = sub(x = logits_chunk_9_mul, y = logits_lse)[name = string("logits_chunk_9_sub_1")];
+            tensor<fp16, [1, 2048, 1, 128]> probs_chunk_9 = exp(x = logits_chunk_9_sub_1)[name = string("probs_chunk_9")];
+            tensor<bool, [1, 2048, 1, 128]> mask_probs_chunk_9 = greater_equal(x = probs_chunk_9, y = min_p_thresh)[name = string("mask_probs_chunk_9")];
+            string mask_chunk_9_fp16_dtype_0 = const()[name = string("mask_chunk_9_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 128]> mask_chunk_9_fp16 = cast(dtype = mask_chunk_9_fp16_dtype_0, x = mask_probs_chunk_9)[name = string("cast_262")];
+            tensor<fp16, [1, 2048, 1, 128]> masked_probs_chunk_9 = select(a = probs_chunk_9, b = mask_chunk_9_fp16, cond = mask_probs_chunk_9)[name = string("masked_probs_chunk_9")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_10_sub_1 = sub(x = logits_chunk_10_mul, y = logits_lse)[name = string("logits_chunk_10_sub_1")];
+            tensor<fp16, [1, 2048, 1, 128]> probs_chunk_10 = exp(x = logits_chunk_10_sub_1)[name = string("probs_chunk_10")];
+            tensor<bool, [1, 2048, 1, 128]> mask_probs_chunk_10 = greater_equal(x = probs_chunk_10, y = min_p_thresh)[name = string("mask_probs_chunk_10")];
+            string mask_chunk_10_fp16_dtype_0 = const()[name = string("mask_chunk_10_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 128]> mask_chunk_10_fp16 = cast(dtype = mask_chunk_10_fp16_dtype_0, x = mask_probs_chunk_10)[name = string("cast_261")];
+            tensor<fp16, [1, 2048, 1, 128]> masked_probs_chunk_10 = select(a = probs_chunk_10, b = mask_chunk_10_fp16, cond = mask_probs_chunk_10)[name = string("masked_probs_chunk_10")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_11_sub_1 = sub(x = logits_chunk_11_mul, y = logits_lse)[name = string("logits_chunk_11_sub_1")];
+            tensor<fp16, [1, 2048, 1, 128]> probs_chunk_11 = exp(x = logits_chunk_11_sub_1)[name = string("probs_chunk_11")];
+            tensor<bool, [1, 2048, 1, 128]> mask_probs_chunk_11 = greater_equal(x = probs_chunk_11, y = min_p_thresh)[name = string("mask_probs_chunk_11")];
+            string mask_chunk_11_fp16_dtype_0 = const()[name = string("mask_chunk_11_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 128]> mask_chunk_11_fp16 = cast(dtype = mask_chunk_11_fp16_dtype_0, x = mask_probs_chunk_11)[name = string("cast_260")];
+            tensor<fp16, [1, 2048, 1, 128]> masked_probs_chunk_11 = select(a = probs_chunk_11, b = mask_chunk_11_fp16, cond = mask_probs_chunk_11)[name = string("masked_probs_chunk_11")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_12_sub_1 = sub(x = logits_chunk_12_mul, y = logits_lse)[name = string("logits_chunk_12_sub_1")];
+            tensor<fp16, [1, 2048, 1, 128]> probs_chunk_12 = exp(x = logits_chunk_12_sub_1)[name = string("probs_chunk_12")];
+            tensor<bool, [1, 2048, 1, 128]> mask_probs_chunk_12 = greater_equal(x = probs_chunk_12, y = min_p_thresh)[name = string("mask_probs_chunk_12")];
+            string mask_chunk_12_fp16_dtype_0 = const()[name = string("mask_chunk_12_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 128]> mask_chunk_12_fp16 = cast(dtype = mask_chunk_12_fp16_dtype_0, x = mask_probs_chunk_12)[name = string("cast_259")];
+            tensor<fp16, [1, 2048, 1, 128]> masked_probs_chunk_12 = select(a = probs_chunk_12, b = mask_chunk_12_fp16, cond = mask_probs_chunk_12)[name = string("masked_probs_chunk_12")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_13_sub_1 = sub(x = logits_chunk_13_mul, y = logits_lse)[name = string("logits_chunk_13_sub_1")];
+            tensor<fp16, [1, 2048, 1, 128]> probs_chunk_13 = exp(x = logits_chunk_13_sub_1)[name = string("probs_chunk_13")];
+            tensor<bool, [1, 2048, 1, 128]> mask_probs_chunk_13 = greater_equal(x = probs_chunk_13, y = min_p_thresh)[name = string("mask_probs_chunk_13")];
+            string mask_chunk_13_fp16_dtype_0 = const()[name = string("mask_chunk_13_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 128]> mask_chunk_13_fp16 = cast(dtype = mask_chunk_13_fp16_dtype_0, x = mask_probs_chunk_13)[name = string("cast_258")];
+            tensor<fp16, [1, 2048, 1, 128]> masked_probs_chunk_13 = select(a = probs_chunk_13, b = mask_chunk_13_fp16, cond = mask_probs_chunk_13)[name = string("masked_probs_chunk_13")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_14_sub_1 = sub(x = logits_chunk_14_mul, y = logits_lse)[name = string("logits_chunk_14_sub_1")];
+            tensor<fp16, [1, 2048, 1, 128]> probs_chunk_14 = exp(x = logits_chunk_14_sub_1)[name = string("probs_chunk_14")];
+            tensor<bool, [1, 2048, 1, 128]> mask_probs_chunk_14 = greater_equal(x = probs_chunk_14, y = min_p_thresh)[name = string("mask_probs_chunk_14")];
+            string mask_chunk_14_fp16_dtype_0 = const()[name = string("mask_chunk_14_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 128]> mask_chunk_14_fp16 = cast(dtype = mask_chunk_14_fp16_dtype_0, x = mask_probs_chunk_14)[name = string("cast_257")];
+            tensor<fp16, [1, 2048, 1, 128]> masked_probs_chunk_14 = select(a = probs_chunk_14, b = mask_chunk_14_fp16, cond = mask_probs_chunk_14)[name = string("masked_probs_chunk_14")];
+            tensor<fp16, [1, 2048, 1, 128]> logits_chunk_15_sub_1 = sub(x = logits_chunk_15_mul, y = logits_lse)[name = string("logits_chunk_15_sub_1")];
+            tensor<fp16, [1, 2048, 1, 128]> probs_chunk_15 = exp(x = logits_chunk_15_sub_1)[name = string("probs_chunk_15")];
+            tensor<bool, [1, 2048, 1, 128]> mask_probs_chunk_15 = greater_equal(x = probs_chunk_15, y = min_p_thresh)[name = string("mask_probs_chunk_15")];
+            string mask_chunk_15_fp16_dtype_0 = const()[name = string("mask_chunk_15_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 128]> mask_chunk_15_fp16 = cast(dtype = mask_chunk_15_fp16_dtype_0, x = mask_probs_chunk_15)[name = string("cast_256")];
+            tensor<fp16, [1, 2048, 1, 128]> masked_probs_chunk_15 = select(a = probs_chunk_15, b = mask_chunk_15_fp16, cond = mask_probs_chunk_15)[name = string("masked_probs_chunk_15")];
+            int32 probs_axis_0 = const()[name = string("probs_axis_0"), val = int32(1)];
+            bool probs_interleave_0 = const()[name = string("probs_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 32768, 1, 128]> probs = concat(axis = probs_axis_0, interleave = probs_interleave_0, values = (masked_probs_chunk_0, masked_probs_chunk_1, masked_probs_chunk_2, masked_probs_chunk_3, masked_probs_chunk_4, masked_probs_chunk_5, masked_probs_chunk_6, masked_probs_chunk_7, masked_probs_chunk_8, masked_probs_chunk_9, masked_probs_chunk_10, masked_probs_chunk_11, masked_probs_chunk_12, masked_probs_chunk_13, masked_probs_chunk_14, masked_probs_chunk_15))[name = string("probs")];
+            string probs_fp32_dtype_0 = const()[name = string("probs_fp32_dtype_0"), val = string("fp32")];
+            int32 probs_cumsum_axis_0 = const()[name = string("probs_cumsum_axis_0"), val = int32(1)];
+            bool probs_cumsum_exclusive_0 = const()[name = string("probs_cumsum_exclusive_0"), val = bool(false)];
+            bool probs_cumsum_reverse_0 = const()[name = string("probs_cumsum_reverse_0"), val = bool(false)];
+            tensor<fp32, [1, 32768, 1, 128]> probs_fp32 = cast(dtype = probs_fp32_dtype_0, x = probs)[name = string("cast_255")];
+            tensor<fp32, [1, 32768, 1, 128]> probs_cumsum = cumsum(axis = probs_cumsum_axis_0, exclusive = probs_cumsum_exclusive_0, reverse = probs_cumsum_reverse_0, x = probs_fp32)[name = string("probs_cumsum")];
+            tensor<int32, [1]> probs_sum_indices_0 = const()[name = string("probs_sum_indices_0"), val = tensor<int32, [1]>([32767])];
+            int32 probs_sum_axis_0 = const()[name = string("probs_sum_axis_0"), val = int32(1)];
+            int32 probs_sum_batch_dims_0 = const()[name = string("probs_sum_batch_dims_0"), val = int32(0)];
+            bool probs_sum_validate_indices_0 = const()[name = string("probs_sum_validate_indices_0"), val = bool(false)];
+            tensor<fp32, [1, 1, 1, 128]> probs_sum = gather(axis = probs_sum_axis_0, batch_dims = probs_sum_batch_dims_0, indices = probs_sum_indices_0, validate_indices = probs_sum_validate_indices_0, x = probs_cumsum)[name = string("probs_sum")];
+            tensor<fp32, [1, 1, 1, 128]> random_number_scaled = mul(x = random_number, y = probs_sum)[name = string("random_number_scaled")];
+            tensor<bool, [1, 32768, 1, 128]> probs_greater = greater(x = probs_cumsum, y = random_number_scaled)[name = string("probs_greater")];
+            string probs_greater_int32_dtype_0 = const()[name = string("probs_greater_int32_dtype_0"), val = string("int32")];
+            int32 sampled_index_axis_0 = const()[name = string("sampled_index_axis_0"), val = int32(1)];
+            bool sampled_index_keep_dims_0 = const()[name = string("sampled_index_keep_dims_0"), val = bool(true)];
+            string sampled_index_output_dtype_0 = const()[name = string("sampled_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 32768, 1, 128]> probs_greater_int32 = cast(dtype = probs_greater_int32_dtype_0, x = probs_greater)[name = string("cast_254")];
+            tensor<int32, [1, 1, 1, 128]> sampled_index = reduce_argmax(axis = sampled_index_axis_0, keep_dims = sampled_index_keep_dims_0, output_dtype = sampled_index_output_dtype_0, x = probs_greater_int32)[name = string("sampled_index")];
+            int32 sampled_index_probability_axis_0 = const()[name = string("sampled_index_probability_axis_0"), val = int32(1)];
+            bool sampled_index_probability_validate_indices_0 = const()[name = string("sampled_index_probability_validate_indices_0"), val = bool(false)];
+            tensor<fp32, [1, 1, 1, 128]> sampled_index_probability = gather_along_axis(axis = sampled_index_probability_axis_0, indices = sampled_index, validate_indices = sampled_index_probability_validate_indices_0, x = probs_fp32)[name = string("sampled_index_probability")];
+            int32 max_logit_index_axis_0 = const()[name = string("max_logit_index_axis_0"), val = int32(1)];
+            bool max_logit_index_keep_dims_0 = const()[name = string("max_logit_index_keep_dims_0"), val = bool(true)];
+            string max_logit_index_output_dtype_0 = const()[name = string("max_logit_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 1, 1, 128]> max_logit_index = reduce_argmax(axis = max_logit_index_axis_0, keep_dims = max_logit_index_keep_dims_0, output_dtype = max_logit_index_output_dtype_0, x = logits_max_logits_chunks)[name = string("max_logit_index")];
+            string indices_chunk_0_int32_dtype_0 = const()[name = string("indices_chunk_0_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_1_int32_dtype_0 = const()[name = string("indices_chunk_1_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_2_int32_dtype_0 = const()[name = string("indices_chunk_2_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_3_int32_dtype_0 = const()[name = string("indices_chunk_3_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_4_int32_dtype_0 = const()[name = string("indices_chunk_4_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_5_int32_dtype_0 = const()[name = string("indices_chunk_5_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_6_int32_dtype_0 = const()[name = string("indices_chunk_6_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_7_int32_dtype_0 = const()[name = string("indices_chunk_7_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_8_int32_dtype_0 = const()[name = string("indices_chunk_8_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_9_int32_dtype_0 = const()[name = string("indices_chunk_9_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_10_int32_dtype_0 = const()[name = string("indices_chunk_10_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_11_int32_dtype_0 = const()[name = string("indices_chunk_11_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_12_int32_dtype_0 = const()[name = string("indices_chunk_12_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_13_int32_dtype_0 = const()[name = string("indices_chunk_13_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_14_int32_dtype_0 = const()[name = string("indices_chunk_14_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_15_int32_dtype_0 = const()[name = string("indices_chunk_15_int32_dtype_0"), val = string("int32")];
+            int32 indices_axis_0 = const()[name = string("indices_axis_0"), val = int32(1)];
+            bool indices_interleave_0 = const()[name = string("indices_interleave_0"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_15_int32 = cast(dtype = indices_chunk_15_int32_dtype_0, x = logits_chunk_15_argmax)[name = string("cast_238")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_14_int32 = cast(dtype = indices_chunk_14_int32_dtype_0, x = logits_chunk_14_argmax)[name = string("cast_239")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_13_int32 = cast(dtype = indices_chunk_13_int32_dtype_0, x = logits_chunk_13_argmax)[name = string("cast_240")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_12_int32 = cast(dtype = indices_chunk_12_int32_dtype_0, x = logits_chunk_12_argmax)[name = string("cast_241")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_11_int32 = cast(dtype = indices_chunk_11_int32_dtype_0, x = logits_chunk_11_argmax)[name = string("cast_242")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_10_int32 = cast(dtype = indices_chunk_10_int32_dtype_0, x = logits_chunk_10_argmax)[name = string("cast_243")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_9_int32 = cast(dtype = indices_chunk_9_int32_dtype_0, x = logits_chunk_9_argmax)[name = string("cast_244")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_8_int32 = cast(dtype = indices_chunk_8_int32_dtype_0, x = logits_chunk_8_argmax)[name = string("cast_245")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_7_int32 = cast(dtype = indices_chunk_7_int32_dtype_0, x = logits_chunk_7_argmax)[name = string("cast_246")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_6_int32 = cast(dtype = indices_chunk_6_int32_dtype_0, x = logits_chunk_6_argmax)[name = string("cast_247")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_5_int32 = cast(dtype = indices_chunk_5_int32_dtype_0, x = logits_chunk_5_argmax)[name = string("cast_248")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_4_int32 = cast(dtype = indices_chunk_4_int32_dtype_0, x = logits_chunk_4_argmax)[name = string("cast_249")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_3_int32 = cast(dtype = indices_chunk_3_int32_dtype_0, x = logits_chunk_3_argmax)[name = string("cast_250")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_2_int32 = cast(dtype = indices_chunk_2_int32_dtype_0, x = logits_chunk_2_argmax)[name = string("cast_251")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_1_int32 = cast(dtype = indices_chunk_1_int32_dtype_0, x = logits_chunk_1_argmax)[name = string("cast_252")];
+            tensor<int32, [1, 1, 1, 128]> indices_chunk_0_int32 = cast(dtype = indices_chunk_0_int32_dtype_0, x = logits_chunk_0_argmax)[name = string("cast_253")];
+            tensor<int32, [1, 16, 1, 128]> indices = concat(axis = indices_axis_0, interleave = indices_interleave_0, values = (indices_chunk_0_int32, indices_chunk_1_int32, indices_chunk_2_int32, indices_chunk_3_int32, indices_chunk_4_int32, indices_chunk_5_int32, indices_chunk_6_int32, indices_chunk_7_int32, indices_chunk_8_int32, indices_chunk_9_int32, indices_chunk_10_int32, indices_chunk_11_int32, indices_chunk_12_int32, indices_chunk_13_int32, indices_chunk_14_int32, indices_chunk_15_int32))[name = string("indices")];
+            int32 argmax_chunks_axis_0 = const()[name = string("argmax_chunks_axis_0"), val = int32(1)];
+            bool argmax_chunks_validate_indices_0 = const()[name = string("argmax_chunks_validate_indices_0"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 128]> argmax_chunks = gather_along_axis(axis = argmax_chunks_axis_0, indices = max_logit_index, validate_indices = argmax_chunks_validate_indices_0, x = indices)[name = string("argmax_chunks")];
+            int32 mul_0_x_0 = const()[name = string("mul_0_x_0"), val = int32(2048)];
+            tensor<int32, [1, 1, 1, 128]> mul_0 = mul(x = mul_0_x_0, y = max_logit_index)[name = string("mul_0")];
+            tensor<int32, [1, 1, 1, 128]> argmax = add(x = argmax_chunks, y = mul_0)[name = string("argmax")];
+        } -> (sampled_index, sampled_index_probability, argmax, max_prob);
+    func min_p_length_16<ios18>(tensor<fp16, [1, 2048, 1, 16]> hidden_states, tensor<fp16, [1]> p, tensor<fp32, [16]> random_number, tensor<fp16, [1]> temp) {
+            tensor<fp16, [1, 2048, 1, 16]> final_norm_rmsnorm_abs = abs(x = hidden_states)[name = string("final_norm_rmsnorm_abs")];
+            tensor<int32, [1]> final_norm_rmsnorm_maxval_axes_0 = const()[name = string("final_norm_rmsnorm_maxval_axes_0"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_maxval_keep_dims_0 = const()[name = string("final_norm_rmsnorm_maxval_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> final_norm_rmsnorm_maxval = reduce_max(axes = final_norm_rmsnorm_maxval_axes_0, keep_dims = final_norm_rmsnorm_maxval_keep_dims_0, x = final_norm_rmsnorm_abs)[name = string("final_norm_rmsnorm_maxval")];
+            fp16 final_norm_rmsnorm_maxval_clipped_alpha_0 = const()[name = string("final_norm_rmsnorm_maxval_clipped_alpha_0"), val = fp16(0x1p-14)];
+            fp16 final_norm_rmsnorm_maxval_clipped_beta_0 = const()[name = string("final_norm_rmsnorm_maxval_clipped_beta_0"), val = fp16(inf)];
+            tensor<fp16, [1, 1, 1, 16]> final_norm_rmsnorm_maxval_clipped = clip(alpha = final_norm_rmsnorm_maxval_clipped_alpha_0, beta = final_norm_rmsnorm_maxval_clipped_beta_0, x = final_norm_rmsnorm_maxval)[name = string("final_norm_rmsnorm_maxval_clipped")];
+            tensor<fp16, [1, 2048, 1, 16]> final_norm_rmsnorm_scaled = real_div(x = hidden_states, y = final_norm_rmsnorm_maxval_clipped)[name = string("final_norm_rmsnorm_scaled")];
+            tensor<int32, [1]> final_norm_rmsnorm_squared_sum_axes_0 = const()[name = string("final_norm_rmsnorm_squared_sum_axes_0"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_squared_sum_keep_dims_0 = const()[name = string("final_norm_rmsnorm_squared_sum_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> final_norm_rmsnorm_squared_sum = reduce_sum_square(axes = final_norm_rmsnorm_squared_sum_axes_0, keep_dims = final_norm_rmsnorm_squared_sum_keep_dims_0, x = final_norm_rmsnorm_scaled)[name = string("final_norm_rmsnorm_squared_sum")];
+            fp16 final_norm_rmsnorm_rsqrt_epsilon_0 = const()[name = string("final_norm_rmsnorm_rsqrt_epsilon_0"), val = fp16(0x1p-14)];
+            tensor<fp16, [1, 1, 1, 16]> final_norm_rmsnorm_rsqrt = rsqrt(epsilon = final_norm_rmsnorm_rsqrt_epsilon_0, x = final_norm_rmsnorm_squared_sum)[name = string("final_norm_rmsnorm_rsqrt")];
+            fp16 final_norm_rmsnorm_dim_scaled_y_0 = const()[name = string("final_norm_rmsnorm_dim_scaled_y_0"), val = fp16(0x1.6ap+5)];
+            tensor<fp16, [1, 2048, 1, 16]> final_norm_rmsnorm_dim_scaled = mul(x = final_norm_rmsnorm_scaled, y = final_norm_rmsnorm_dim_scaled_y_0)[name = string("final_norm_rmsnorm_dim_scaled")];
+            tensor<fp16, [1, 2048, 1, 16]> final_norm_rmsnorm_normalized = mul(x = final_norm_rmsnorm_dim_scaled, y = final_norm_rmsnorm_rsqrt)[name = string("final_norm_rmsnorm_normalized")];
+            tensor<fp16, [2048, 1, 1]> final_norm_rmsnorm_y_0 = const()[name = string("final_norm_rmsnorm_y_0"), val = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 2048, 1, 16]> final_norm_rmsnorm = mul(x = final_norm_rmsnorm_normalized, y = final_norm_rmsnorm_y_0)[name = string("final_norm_rmsnorm")];
+            fp16 temp_inverse_epsilon_0 = const()[name = string("temp_inverse_epsilon_0"), val = fp16(0x0p+0)];
+            tensor<fp16, [1]> temp_inverse = inverse(epsilon = temp_inverse_epsilon_0, x = temp)[name = string("temp_inverse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_0_weight_0 = const()[name = string("logits_chunk_0_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4224)))];
+            tensor<int32, [2]> logits_chunk_0_strides_0 = const()[name = string("logits_chunk_0_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_0_pad_type_0 = const()[name = string("logits_chunk_0_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_0_pad_0 = const()[name = string("logits_chunk_0_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_0_dilations_0 = const()[name = string("logits_chunk_0_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_0_groups_0 = const()[name = string("logits_chunk_0_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_0 = conv(dilations = logits_chunk_0_dilations_0, groups = logits_chunk_0_groups_0, pad = logits_chunk_0_pad_0, pad_type = logits_chunk_0_pad_type_0, strides = logits_chunk_0_strides_0, weight = logits_chunk_0_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_0")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_0_mul = mul(x = logits_chunk_0, y = temp_inverse)[name = string("logits_chunk_0_mul")];
+            tensor<int32, [1]> logits_chunk_0_max_axes_0 = const()[name = string("logits_chunk_0_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_0_max_keep_dims_0 = const()[name = string("logits_chunk_0_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_0_max = reduce_max(axes = logits_chunk_0_max_axes_0, keep_dims = logits_chunk_0_max_keep_dims_0, x = logits_chunk_0_mul)[name = string("logits_chunk_0_max")];
+            int32 logits_chunk_0_argmax_axis_0 = const()[name = string("logits_chunk_0_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_0_argmax_keep_dims_0 = const()[name = string("logits_chunk_0_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_0_argmax_output_dtype_0 = const()[name = string("logits_chunk_0_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> logits_chunk_0_argmax = reduce_argmax(axis = logits_chunk_0_argmax_axis_0, keep_dims = logits_chunk_0_argmax_keep_dims_0, output_dtype = logits_chunk_0_argmax_output_dtype_0, x = logits_chunk_0_mul)[name = string("logits_chunk_0_argmax")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_0_sub = sub(x = logits_chunk_0_mul, y = logits_chunk_0_max)[name = string("logits_chunk_0_sub")];
+            tensor<int32, [1]> logits_chunk_0_lse_sub_axes_0 = const()[name = string("logits_chunk_0_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_0_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_0_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_0_lse_sub = reduce_log_sum_exp(axes = logits_chunk_0_lse_sub_axes_0, keep_dims = logits_chunk_0_lse_sub_keep_dims_0, x = logits_chunk_0_sub)[name = string("logits_chunk_0_lse_sub")];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_0_lse = add(x = logits_chunk_0_lse_sub, y = logits_chunk_0_max)[name = string("logits_chunk_0_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_1_weight_0 = const()[name = string("logits_chunk_1_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8392896)))];
+            tensor<int32, [2]> logits_chunk_1_strides_0 = const()[name = string("logits_chunk_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_1_pad_type_0 = const()[name = string("logits_chunk_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_1_pad_0 = const()[name = string("logits_chunk_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_1_dilations_0 = const()[name = string("logits_chunk_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_1_groups_0 = const()[name = string("logits_chunk_1_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_1 = conv(dilations = logits_chunk_1_dilations_0, groups = logits_chunk_1_groups_0, pad = logits_chunk_1_pad_0, pad_type = logits_chunk_1_pad_type_0, strides = logits_chunk_1_strides_0, weight = logits_chunk_1_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_1")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_1_mul = mul(x = logits_chunk_1, y = temp_inverse)[name = string("logits_chunk_1_mul")];
+            tensor<int32, [1]> logits_chunk_1_max_axes_0 = const()[name = string("logits_chunk_1_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_1_max_keep_dims_0 = const()[name = string("logits_chunk_1_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_1_max = reduce_max(axes = logits_chunk_1_max_axes_0, keep_dims = logits_chunk_1_max_keep_dims_0, x = logits_chunk_1_mul)[name = string("logits_chunk_1_max")];
+            int32 logits_chunk_1_argmax_axis_0 = const()[name = string("logits_chunk_1_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_1_argmax_keep_dims_0 = const()[name = string("logits_chunk_1_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_1_argmax_output_dtype_0 = const()[name = string("logits_chunk_1_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> logits_chunk_1_argmax = reduce_argmax(axis = logits_chunk_1_argmax_axis_0, keep_dims = logits_chunk_1_argmax_keep_dims_0, output_dtype = logits_chunk_1_argmax_output_dtype_0, x = logits_chunk_1_mul)[name = string("logits_chunk_1_argmax")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_1_sub = sub(x = logits_chunk_1_mul, y = logits_chunk_1_max)[name = string("logits_chunk_1_sub")];
+            tensor<int32, [1]> logits_chunk_1_lse_sub_axes_0 = const()[name = string("logits_chunk_1_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_1_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_1_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_1_lse_sub = reduce_log_sum_exp(axes = logits_chunk_1_lse_sub_axes_0, keep_dims = logits_chunk_1_lse_sub_keep_dims_0, x = logits_chunk_1_sub)[name = string("logits_chunk_1_lse_sub")];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_1_lse = add(x = logits_chunk_1_lse_sub, y = logits_chunk_1_max)[name = string("logits_chunk_1_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_2_weight_0 = const()[name = string("logits_chunk_2_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16781568)))];
+            tensor<int32, [2]> logits_chunk_2_strides_0 = const()[name = string("logits_chunk_2_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_2_pad_type_0 = const()[name = string("logits_chunk_2_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_2_pad_0 = const()[name = string("logits_chunk_2_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_2_dilations_0 = const()[name = string("logits_chunk_2_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_2_groups_0 = const()[name = string("logits_chunk_2_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_2 = conv(dilations = logits_chunk_2_dilations_0, groups = logits_chunk_2_groups_0, pad = logits_chunk_2_pad_0, pad_type = logits_chunk_2_pad_type_0, strides = logits_chunk_2_strides_0, weight = logits_chunk_2_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_2")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_2_mul = mul(x = logits_chunk_2, y = temp_inverse)[name = string("logits_chunk_2_mul")];
+            tensor<int32, [1]> logits_chunk_2_max_axes_0 = const()[name = string("logits_chunk_2_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_2_max_keep_dims_0 = const()[name = string("logits_chunk_2_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_2_max = reduce_max(axes = logits_chunk_2_max_axes_0, keep_dims = logits_chunk_2_max_keep_dims_0, x = logits_chunk_2_mul)[name = string("logits_chunk_2_max")];
+            int32 logits_chunk_2_argmax_axis_0 = const()[name = string("logits_chunk_2_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_2_argmax_keep_dims_0 = const()[name = string("logits_chunk_2_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_2_argmax_output_dtype_0 = const()[name = string("logits_chunk_2_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> logits_chunk_2_argmax = reduce_argmax(axis = logits_chunk_2_argmax_axis_0, keep_dims = logits_chunk_2_argmax_keep_dims_0, output_dtype = logits_chunk_2_argmax_output_dtype_0, x = logits_chunk_2_mul)[name = string("logits_chunk_2_argmax")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_2_sub = sub(x = logits_chunk_2_mul, y = logits_chunk_2_max)[name = string("logits_chunk_2_sub")];
+            tensor<int32, [1]> logits_chunk_2_lse_sub_axes_0 = const()[name = string("logits_chunk_2_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_2_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_2_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_2_lse_sub = reduce_log_sum_exp(axes = logits_chunk_2_lse_sub_axes_0, keep_dims = logits_chunk_2_lse_sub_keep_dims_0, x = logits_chunk_2_sub)[name = string("logits_chunk_2_lse_sub")];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_2_lse = add(x = logits_chunk_2_lse_sub, y = logits_chunk_2_max)[name = string("logits_chunk_2_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_3_weight_0 = const()[name = string("logits_chunk_3_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25170240)))];
+            tensor<int32, [2]> logits_chunk_3_strides_0 = const()[name = string("logits_chunk_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_3_pad_type_0 = const()[name = string("logits_chunk_3_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_3_pad_0 = const()[name = string("logits_chunk_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_3_dilations_0 = const()[name = string("logits_chunk_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_3_groups_0 = const()[name = string("logits_chunk_3_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_3 = conv(dilations = logits_chunk_3_dilations_0, groups = logits_chunk_3_groups_0, pad = logits_chunk_3_pad_0, pad_type = logits_chunk_3_pad_type_0, strides = logits_chunk_3_strides_0, weight = logits_chunk_3_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_3")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_3_mul = mul(x = logits_chunk_3, y = temp_inverse)[name = string("logits_chunk_3_mul")];
+            tensor<int32, [1]> logits_chunk_3_max_axes_0 = const()[name = string("logits_chunk_3_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_3_max_keep_dims_0 = const()[name = string("logits_chunk_3_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_3_max = reduce_max(axes = logits_chunk_3_max_axes_0, keep_dims = logits_chunk_3_max_keep_dims_0, x = logits_chunk_3_mul)[name = string("logits_chunk_3_max")];
+            int32 logits_chunk_3_argmax_axis_0 = const()[name = string("logits_chunk_3_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_3_argmax_keep_dims_0 = const()[name = string("logits_chunk_3_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_3_argmax_output_dtype_0 = const()[name = string("logits_chunk_3_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> logits_chunk_3_argmax = reduce_argmax(axis = logits_chunk_3_argmax_axis_0, keep_dims = logits_chunk_3_argmax_keep_dims_0, output_dtype = logits_chunk_3_argmax_output_dtype_0, x = logits_chunk_3_mul)[name = string("logits_chunk_3_argmax")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_3_sub = sub(x = logits_chunk_3_mul, y = logits_chunk_3_max)[name = string("logits_chunk_3_sub")];
+            tensor<int32, [1]> logits_chunk_3_lse_sub_axes_0 = const()[name = string("logits_chunk_3_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_3_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_3_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_3_lse_sub = reduce_log_sum_exp(axes = logits_chunk_3_lse_sub_axes_0, keep_dims = logits_chunk_3_lse_sub_keep_dims_0, x = logits_chunk_3_sub)[name = string("logits_chunk_3_lse_sub")];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_3_lse = add(x = logits_chunk_3_lse_sub, y = logits_chunk_3_max)[name = string("logits_chunk_3_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_4_weight_0 = const()[name = string("logits_chunk_4_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33558912)))];
+            tensor<int32, [2]> logits_chunk_4_strides_0 = const()[name = string("logits_chunk_4_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_4_pad_type_0 = const()[name = string("logits_chunk_4_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_4_pad_0 = const()[name = string("logits_chunk_4_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_4_dilations_0 = const()[name = string("logits_chunk_4_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_4_groups_0 = const()[name = string("logits_chunk_4_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_4 = conv(dilations = logits_chunk_4_dilations_0, groups = logits_chunk_4_groups_0, pad = logits_chunk_4_pad_0, pad_type = logits_chunk_4_pad_type_0, strides = logits_chunk_4_strides_0, weight = logits_chunk_4_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_4")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_4_mul = mul(x = logits_chunk_4, y = temp_inverse)[name = string("logits_chunk_4_mul")];
+            tensor<int32, [1]> logits_chunk_4_max_axes_0 = const()[name = string("logits_chunk_4_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_4_max_keep_dims_0 = const()[name = string("logits_chunk_4_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_4_max = reduce_max(axes = logits_chunk_4_max_axes_0, keep_dims = logits_chunk_4_max_keep_dims_0, x = logits_chunk_4_mul)[name = string("logits_chunk_4_max")];
+            int32 logits_chunk_4_argmax_axis_0 = const()[name = string("logits_chunk_4_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_4_argmax_keep_dims_0 = const()[name = string("logits_chunk_4_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_4_argmax_output_dtype_0 = const()[name = string("logits_chunk_4_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> logits_chunk_4_argmax = reduce_argmax(axis = logits_chunk_4_argmax_axis_0, keep_dims = logits_chunk_4_argmax_keep_dims_0, output_dtype = logits_chunk_4_argmax_output_dtype_0, x = logits_chunk_4_mul)[name = string("logits_chunk_4_argmax")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_4_sub = sub(x = logits_chunk_4_mul, y = logits_chunk_4_max)[name = string("logits_chunk_4_sub")];
+            tensor<int32, [1]> logits_chunk_4_lse_sub_axes_0 = const()[name = string("logits_chunk_4_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_4_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_4_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_4_lse_sub = reduce_log_sum_exp(axes = logits_chunk_4_lse_sub_axes_0, keep_dims = logits_chunk_4_lse_sub_keep_dims_0, x = logits_chunk_4_sub)[name = string("logits_chunk_4_lse_sub")];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_4_lse = add(x = logits_chunk_4_lse_sub, y = logits_chunk_4_max)[name = string("logits_chunk_4_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_5_weight_0 = const()[name = string("logits_chunk_5_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41947584)))];
+            tensor<int32, [2]> logits_chunk_5_strides_0 = const()[name = string("logits_chunk_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_5_pad_type_0 = const()[name = string("logits_chunk_5_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_5_pad_0 = const()[name = string("logits_chunk_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_5_dilations_0 = const()[name = string("logits_chunk_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_5_groups_0 = const()[name = string("logits_chunk_5_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_5 = conv(dilations = logits_chunk_5_dilations_0, groups = logits_chunk_5_groups_0, pad = logits_chunk_5_pad_0, pad_type = logits_chunk_5_pad_type_0, strides = logits_chunk_5_strides_0, weight = logits_chunk_5_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_5")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_5_mul = mul(x = logits_chunk_5, y = temp_inverse)[name = string("logits_chunk_5_mul")];
+            tensor<int32, [1]> logits_chunk_5_max_axes_0 = const()[name = string("logits_chunk_5_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_5_max_keep_dims_0 = const()[name = string("logits_chunk_5_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_5_max = reduce_max(axes = logits_chunk_5_max_axes_0, keep_dims = logits_chunk_5_max_keep_dims_0, x = logits_chunk_5_mul)[name = string("logits_chunk_5_max")];
+            int32 logits_chunk_5_argmax_axis_0 = const()[name = string("logits_chunk_5_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_5_argmax_keep_dims_0 = const()[name = string("logits_chunk_5_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_5_argmax_output_dtype_0 = const()[name = string("logits_chunk_5_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> logits_chunk_5_argmax = reduce_argmax(axis = logits_chunk_5_argmax_axis_0, keep_dims = logits_chunk_5_argmax_keep_dims_0, output_dtype = logits_chunk_5_argmax_output_dtype_0, x = logits_chunk_5_mul)[name = string("logits_chunk_5_argmax")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_5_sub = sub(x = logits_chunk_5_mul, y = logits_chunk_5_max)[name = string("logits_chunk_5_sub")];
+            tensor<int32, [1]> logits_chunk_5_lse_sub_axes_0 = const()[name = string("logits_chunk_5_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_5_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_5_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_5_lse_sub = reduce_log_sum_exp(axes = logits_chunk_5_lse_sub_axes_0, keep_dims = logits_chunk_5_lse_sub_keep_dims_0, x = logits_chunk_5_sub)[name = string("logits_chunk_5_lse_sub")];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_5_lse = add(x = logits_chunk_5_lse_sub, y = logits_chunk_5_max)[name = string("logits_chunk_5_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_6_weight_0 = const()[name = string("logits_chunk_6_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50336256)))];
+            tensor<int32, [2]> logits_chunk_6_strides_0 = const()[name = string("logits_chunk_6_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_6_pad_type_0 = const()[name = string("logits_chunk_6_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_6_pad_0 = const()[name = string("logits_chunk_6_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_6_dilations_0 = const()[name = string("logits_chunk_6_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_6_groups_0 = const()[name = string("logits_chunk_6_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_6 = conv(dilations = logits_chunk_6_dilations_0, groups = logits_chunk_6_groups_0, pad = logits_chunk_6_pad_0, pad_type = logits_chunk_6_pad_type_0, strides = logits_chunk_6_strides_0, weight = logits_chunk_6_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_6")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_6_mul = mul(x = logits_chunk_6, y = temp_inverse)[name = string("logits_chunk_6_mul")];
+            tensor<int32, [1]> logits_chunk_6_max_axes_0 = const()[name = string("logits_chunk_6_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_6_max_keep_dims_0 = const()[name = string("logits_chunk_6_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_6_max = reduce_max(axes = logits_chunk_6_max_axes_0, keep_dims = logits_chunk_6_max_keep_dims_0, x = logits_chunk_6_mul)[name = string("logits_chunk_6_max")];
+            int32 logits_chunk_6_argmax_axis_0 = const()[name = string("logits_chunk_6_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_6_argmax_keep_dims_0 = const()[name = string("logits_chunk_6_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_6_argmax_output_dtype_0 = const()[name = string("logits_chunk_6_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> logits_chunk_6_argmax = reduce_argmax(axis = logits_chunk_6_argmax_axis_0, keep_dims = logits_chunk_6_argmax_keep_dims_0, output_dtype = logits_chunk_6_argmax_output_dtype_0, x = logits_chunk_6_mul)[name = string("logits_chunk_6_argmax")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_6_sub = sub(x = logits_chunk_6_mul, y = logits_chunk_6_max)[name = string("logits_chunk_6_sub")];
+            tensor<int32, [1]> logits_chunk_6_lse_sub_axes_0 = const()[name = string("logits_chunk_6_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_6_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_6_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_6_lse_sub = reduce_log_sum_exp(axes = logits_chunk_6_lse_sub_axes_0, keep_dims = logits_chunk_6_lse_sub_keep_dims_0, x = logits_chunk_6_sub)[name = string("logits_chunk_6_lse_sub")];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_6_lse = add(x = logits_chunk_6_lse_sub, y = logits_chunk_6_max)[name = string("logits_chunk_6_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_7_weight_0 = const()[name = string("logits_chunk_7_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58724928)))];
+            tensor<int32, [2]> logits_chunk_7_strides_0 = const()[name = string("logits_chunk_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_7_pad_type_0 = const()[name = string("logits_chunk_7_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_7_pad_0 = const()[name = string("logits_chunk_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_7_dilations_0 = const()[name = string("logits_chunk_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_7_groups_0 = const()[name = string("logits_chunk_7_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_7 = conv(dilations = logits_chunk_7_dilations_0, groups = logits_chunk_7_groups_0, pad = logits_chunk_7_pad_0, pad_type = logits_chunk_7_pad_type_0, strides = logits_chunk_7_strides_0, weight = logits_chunk_7_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_7")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_7_mul = mul(x = logits_chunk_7, y = temp_inverse)[name = string("logits_chunk_7_mul")];
+            tensor<int32, [1]> logits_chunk_7_max_axes_0 = const()[name = string("logits_chunk_7_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_7_max_keep_dims_0 = const()[name = string("logits_chunk_7_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_7_max = reduce_max(axes = logits_chunk_7_max_axes_0, keep_dims = logits_chunk_7_max_keep_dims_0, x = logits_chunk_7_mul)[name = string("logits_chunk_7_max")];
+            int32 logits_chunk_7_argmax_axis_0 = const()[name = string("logits_chunk_7_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_7_argmax_keep_dims_0 = const()[name = string("logits_chunk_7_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_7_argmax_output_dtype_0 = const()[name = string("logits_chunk_7_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> logits_chunk_7_argmax = reduce_argmax(axis = logits_chunk_7_argmax_axis_0, keep_dims = logits_chunk_7_argmax_keep_dims_0, output_dtype = logits_chunk_7_argmax_output_dtype_0, x = logits_chunk_7_mul)[name = string("logits_chunk_7_argmax")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_7_sub = sub(x = logits_chunk_7_mul, y = logits_chunk_7_max)[name = string("logits_chunk_7_sub")];
+            tensor<int32, [1]> logits_chunk_7_lse_sub_axes_0 = const()[name = string("logits_chunk_7_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_7_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_7_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_7_lse_sub = reduce_log_sum_exp(axes = logits_chunk_7_lse_sub_axes_0, keep_dims = logits_chunk_7_lse_sub_keep_dims_0, x = logits_chunk_7_sub)[name = string("logits_chunk_7_lse_sub")];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_7_lse = add(x = logits_chunk_7_lse_sub, y = logits_chunk_7_max)[name = string("logits_chunk_7_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_8_weight_0 = const()[name = string("logits_chunk_8_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67113600)))];
+            tensor<int32, [2]> logits_chunk_8_strides_0 = const()[name = string("logits_chunk_8_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_8_pad_type_0 = const()[name = string("logits_chunk_8_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_8_pad_0 = const()[name = string("logits_chunk_8_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_8_dilations_0 = const()[name = string("logits_chunk_8_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_8_groups_0 = const()[name = string("logits_chunk_8_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_8 = conv(dilations = logits_chunk_8_dilations_0, groups = logits_chunk_8_groups_0, pad = logits_chunk_8_pad_0, pad_type = logits_chunk_8_pad_type_0, strides = logits_chunk_8_strides_0, weight = logits_chunk_8_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_8")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_8_mul = mul(x = logits_chunk_8, y = temp_inverse)[name = string("logits_chunk_8_mul")];
+            tensor<int32, [1]> logits_chunk_8_max_axes_0 = const()[name = string("logits_chunk_8_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_8_max_keep_dims_0 = const()[name = string("logits_chunk_8_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_8_max = reduce_max(axes = logits_chunk_8_max_axes_0, keep_dims = logits_chunk_8_max_keep_dims_0, x = logits_chunk_8_mul)[name = string("logits_chunk_8_max")];
+            int32 logits_chunk_8_argmax_axis_0 = const()[name = string("logits_chunk_8_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_8_argmax_keep_dims_0 = const()[name = string("logits_chunk_8_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_8_argmax_output_dtype_0 = const()[name = string("logits_chunk_8_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> logits_chunk_8_argmax = reduce_argmax(axis = logits_chunk_8_argmax_axis_0, keep_dims = logits_chunk_8_argmax_keep_dims_0, output_dtype = logits_chunk_8_argmax_output_dtype_0, x = logits_chunk_8_mul)[name = string("logits_chunk_8_argmax")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_8_sub = sub(x = logits_chunk_8_mul, y = logits_chunk_8_max)[name = string("logits_chunk_8_sub")];
+            tensor<int32, [1]> logits_chunk_8_lse_sub_axes_0 = const()[name = string("logits_chunk_8_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_8_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_8_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_8_lse_sub = reduce_log_sum_exp(axes = logits_chunk_8_lse_sub_axes_0, keep_dims = logits_chunk_8_lse_sub_keep_dims_0, x = logits_chunk_8_sub)[name = string("logits_chunk_8_lse_sub")];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_8_lse = add(x = logits_chunk_8_lse_sub, y = logits_chunk_8_max)[name = string("logits_chunk_8_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_9_weight_0 = const()[name = string("logits_chunk_9_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75502272)))];
+            tensor<int32, [2]> logits_chunk_9_strides_0 = const()[name = string("logits_chunk_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_9_pad_type_0 = const()[name = string("logits_chunk_9_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_9_pad_0 = const()[name = string("logits_chunk_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_9_dilations_0 = const()[name = string("logits_chunk_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_9_groups_0 = const()[name = string("logits_chunk_9_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_9 = conv(dilations = logits_chunk_9_dilations_0, groups = logits_chunk_9_groups_0, pad = logits_chunk_9_pad_0, pad_type = logits_chunk_9_pad_type_0, strides = logits_chunk_9_strides_0, weight = logits_chunk_9_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_9")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_9_mul = mul(x = logits_chunk_9, y = temp_inverse)[name = string("logits_chunk_9_mul")];
+            tensor<int32, [1]> logits_chunk_9_max_axes_0 = const()[name = string("logits_chunk_9_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_9_max_keep_dims_0 = const()[name = string("logits_chunk_9_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_9_max = reduce_max(axes = logits_chunk_9_max_axes_0, keep_dims = logits_chunk_9_max_keep_dims_0, x = logits_chunk_9_mul)[name = string("logits_chunk_9_max")];
+            int32 logits_chunk_9_argmax_axis_0 = const()[name = string("logits_chunk_9_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_9_argmax_keep_dims_0 = const()[name = string("logits_chunk_9_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_9_argmax_output_dtype_0 = const()[name = string("logits_chunk_9_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> logits_chunk_9_argmax = reduce_argmax(axis = logits_chunk_9_argmax_axis_0, keep_dims = logits_chunk_9_argmax_keep_dims_0, output_dtype = logits_chunk_9_argmax_output_dtype_0, x = logits_chunk_9_mul)[name = string("logits_chunk_9_argmax")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_9_sub = sub(x = logits_chunk_9_mul, y = logits_chunk_9_max)[name = string("logits_chunk_9_sub")];
+            tensor<int32, [1]> logits_chunk_9_lse_sub_axes_0 = const()[name = string("logits_chunk_9_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_9_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_9_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_9_lse_sub = reduce_log_sum_exp(axes = logits_chunk_9_lse_sub_axes_0, keep_dims = logits_chunk_9_lse_sub_keep_dims_0, x = logits_chunk_9_sub)[name = string("logits_chunk_9_lse_sub")];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_9_lse = add(x = logits_chunk_9_lse_sub, y = logits_chunk_9_max)[name = string("logits_chunk_9_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_10_weight_0 = const()[name = string("logits_chunk_10_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83890944)))];
+            tensor<int32, [2]> logits_chunk_10_strides_0 = const()[name = string("logits_chunk_10_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_10_pad_type_0 = const()[name = string("logits_chunk_10_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_10_pad_0 = const()[name = string("logits_chunk_10_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_10_dilations_0 = const()[name = string("logits_chunk_10_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_10_groups_0 = const()[name = string("logits_chunk_10_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_10 = conv(dilations = logits_chunk_10_dilations_0, groups = logits_chunk_10_groups_0, pad = logits_chunk_10_pad_0, pad_type = logits_chunk_10_pad_type_0, strides = logits_chunk_10_strides_0, weight = logits_chunk_10_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_10")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_10_mul = mul(x = logits_chunk_10, y = temp_inverse)[name = string("logits_chunk_10_mul")];
+            tensor<int32, [1]> logits_chunk_10_max_axes_0 = const()[name = string("logits_chunk_10_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_10_max_keep_dims_0 = const()[name = string("logits_chunk_10_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_10_max = reduce_max(axes = logits_chunk_10_max_axes_0, keep_dims = logits_chunk_10_max_keep_dims_0, x = logits_chunk_10_mul)[name = string("logits_chunk_10_max")];
+            int32 logits_chunk_10_argmax_axis_0 = const()[name = string("logits_chunk_10_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_10_argmax_keep_dims_0 = const()[name = string("logits_chunk_10_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_10_argmax_output_dtype_0 = const()[name = string("logits_chunk_10_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> logits_chunk_10_argmax = reduce_argmax(axis = logits_chunk_10_argmax_axis_0, keep_dims = logits_chunk_10_argmax_keep_dims_0, output_dtype = logits_chunk_10_argmax_output_dtype_0, x = logits_chunk_10_mul)[name = string("logits_chunk_10_argmax")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_10_sub = sub(x = logits_chunk_10_mul, y = logits_chunk_10_max)[name = string("logits_chunk_10_sub")];
+            tensor<int32, [1]> logits_chunk_10_lse_sub_axes_0 = const()[name = string("logits_chunk_10_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_10_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_10_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_10_lse_sub = reduce_log_sum_exp(axes = logits_chunk_10_lse_sub_axes_0, keep_dims = logits_chunk_10_lse_sub_keep_dims_0, x = logits_chunk_10_sub)[name = string("logits_chunk_10_lse_sub")];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_10_lse = add(x = logits_chunk_10_lse_sub, y = logits_chunk_10_max)[name = string("logits_chunk_10_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_11_weight_0 = const()[name = string("logits_chunk_11_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92279616)))];
+            tensor<int32, [2]> logits_chunk_11_strides_0 = const()[name = string("logits_chunk_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_11_pad_type_0 = const()[name = string("logits_chunk_11_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_11_pad_0 = const()[name = string("logits_chunk_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_11_dilations_0 = const()[name = string("logits_chunk_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_11_groups_0 = const()[name = string("logits_chunk_11_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_11 = conv(dilations = logits_chunk_11_dilations_0, groups = logits_chunk_11_groups_0, pad = logits_chunk_11_pad_0, pad_type = logits_chunk_11_pad_type_0, strides = logits_chunk_11_strides_0, weight = logits_chunk_11_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_11")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_11_mul = mul(x = logits_chunk_11, y = temp_inverse)[name = string("logits_chunk_11_mul")];
+            tensor<int32, [1]> logits_chunk_11_max_axes_0 = const()[name = string("logits_chunk_11_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_11_max_keep_dims_0 = const()[name = string("logits_chunk_11_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_11_max = reduce_max(axes = logits_chunk_11_max_axes_0, keep_dims = logits_chunk_11_max_keep_dims_0, x = logits_chunk_11_mul)[name = string("logits_chunk_11_max")];
+            int32 logits_chunk_11_argmax_axis_0 = const()[name = string("logits_chunk_11_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_11_argmax_keep_dims_0 = const()[name = string("logits_chunk_11_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_11_argmax_output_dtype_0 = const()[name = string("logits_chunk_11_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> logits_chunk_11_argmax = reduce_argmax(axis = logits_chunk_11_argmax_axis_0, keep_dims = logits_chunk_11_argmax_keep_dims_0, output_dtype = logits_chunk_11_argmax_output_dtype_0, x = logits_chunk_11_mul)[name = string("logits_chunk_11_argmax")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_11_sub = sub(x = logits_chunk_11_mul, y = logits_chunk_11_max)[name = string("logits_chunk_11_sub")];
+            tensor<int32, [1]> logits_chunk_11_lse_sub_axes_0 = const()[name = string("logits_chunk_11_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_11_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_11_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_11_lse_sub = reduce_log_sum_exp(axes = logits_chunk_11_lse_sub_axes_0, keep_dims = logits_chunk_11_lse_sub_keep_dims_0, x = logits_chunk_11_sub)[name = string("logits_chunk_11_lse_sub")];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_11_lse = add(x = logits_chunk_11_lse_sub, y = logits_chunk_11_max)[name = string("logits_chunk_11_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_12_weight_0 = const()[name = string("logits_chunk_12_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100668288)))];
+            tensor<int32, [2]> logits_chunk_12_strides_0 = const()[name = string("logits_chunk_12_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_12_pad_type_0 = const()[name = string("logits_chunk_12_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_12_pad_0 = const()[name = string("logits_chunk_12_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_12_dilations_0 = const()[name = string("logits_chunk_12_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_12_groups_0 = const()[name = string("logits_chunk_12_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_12 = conv(dilations = logits_chunk_12_dilations_0, groups = logits_chunk_12_groups_0, pad = logits_chunk_12_pad_0, pad_type = logits_chunk_12_pad_type_0, strides = logits_chunk_12_strides_0, weight = logits_chunk_12_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_12")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_12_mul = mul(x = logits_chunk_12, y = temp_inverse)[name = string("logits_chunk_12_mul")];
+            tensor<int32, [1]> logits_chunk_12_max_axes_0 = const()[name = string("logits_chunk_12_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_12_max_keep_dims_0 = const()[name = string("logits_chunk_12_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_12_max = reduce_max(axes = logits_chunk_12_max_axes_0, keep_dims = logits_chunk_12_max_keep_dims_0, x = logits_chunk_12_mul)[name = string("logits_chunk_12_max")];
+            int32 logits_chunk_12_argmax_axis_0 = const()[name = string("logits_chunk_12_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_12_argmax_keep_dims_0 = const()[name = string("logits_chunk_12_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_12_argmax_output_dtype_0 = const()[name = string("logits_chunk_12_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> logits_chunk_12_argmax = reduce_argmax(axis = logits_chunk_12_argmax_axis_0, keep_dims = logits_chunk_12_argmax_keep_dims_0, output_dtype = logits_chunk_12_argmax_output_dtype_0, x = logits_chunk_12_mul)[name = string("logits_chunk_12_argmax")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_12_sub = sub(x = logits_chunk_12_mul, y = logits_chunk_12_max)[name = string("logits_chunk_12_sub")];
+            tensor<int32, [1]> logits_chunk_12_lse_sub_axes_0 = const()[name = string("logits_chunk_12_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_12_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_12_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_12_lse_sub = reduce_log_sum_exp(axes = logits_chunk_12_lse_sub_axes_0, keep_dims = logits_chunk_12_lse_sub_keep_dims_0, x = logits_chunk_12_sub)[name = string("logits_chunk_12_lse_sub")];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_12_lse = add(x = logits_chunk_12_lse_sub, y = logits_chunk_12_max)[name = string("logits_chunk_12_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_13_weight_0 = const()[name = string("logits_chunk_13_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109056960)))];
+            tensor<int32, [2]> logits_chunk_13_strides_0 = const()[name = string("logits_chunk_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_13_pad_type_0 = const()[name = string("logits_chunk_13_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_13_pad_0 = const()[name = string("logits_chunk_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_13_dilations_0 = const()[name = string("logits_chunk_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_13_groups_0 = const()[name = string("logits_chunk_13_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_13 = conv(dilations = logits_chunk_13_dilations_0, groups = logits_chunk_13_groups_0, pad = logits_chunk_13_pad_0, pad_type = logits_chunk_13_pad_type_0, strides = logits_chunk_13_strides_0, weight = logits_chunk_13_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_13")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_13_mul = mul(x = logits_chunk_13, y = temp_inverse)[name = string("logits_chunk_13_mul")];
+            tensor<int32, [1]> logits_chunk_13_max_axes_0 = const()[name = string("logits_chunk_13_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_13_max_keep_dims_0 = const()[name = string("logits_chunk_13_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_13_max = reduce_max(axes = logits_chunk_13_max_axes_0, keep_dims = logits_chunk_13_max_keep_dims_0, x = logits_chunk_13_mul)[name = string("logits_chunk_13_max")];
+            int32 logits_chunk_13_argmax_axis_0 = const()[name = string("logits_chunk_13_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_13_argmax_keep_dims_0 = const()[name = string("logits_chunk_13_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_13_argmax_output_dtype_0 = const()[name = string("logits_chunk_13_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> logits_chunk_13_argmax = reduce_argmax(axis = logits_chunk_13_argmax_axis_0, keep_dims = logits_chunk_13_argmax_keep_dims_0, output_dtype = logits_chunk_13_argmax_output_dtype_0, x = logits_chunk_13_mul)[name = string("logits_chunk_13_argmax")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_13_sub = sub(x = logits_chunk_13_mul, y = logits_chunk_13_max)[name = string("logits_chunk_13_sub")];
+            tensor<int32, [1]> logits_chunk_13_lse_sub_axes_0 = const()[name = string("logits_chunk_13_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_13_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_13_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_13_lse_sub = reduce_log_sum_exp(axes = logits_chunk_13_lse_sub_axes_0, keep_dims = logits_chunk_13_lse_sub_keep_dims_0, x = logits_chunk_13_sub)[name = string("logits_chunk_13_lse_sub")];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_13_lse = add(x = logits_chunk_13_lse_sub, y = logits_chunk_13_max)[name = string("logits_chunk_13_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_14_weight_0 = const()[name = string("logits_chunk_14_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117445632)))];
+            tensor<int32, [2]> logits_chunk_14_strides_0 = const()[name = string("logits_chunk_14_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_14_pad_type_0 = const()[name = string("logits_chunk_14_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_14_pad_0 = const()[name = string("logits_chunk_14_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_14_dilations_0 = const()[name = string("logits_chunk_14_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_14_groups_0 = const()[name = string("logits_chunk_14_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_14 = conv(dilations = logits_chunk_14_dilations_0, groups = logits_chunk_14_groups_0, pad = logits_chunk_14_pad_0, pad_type = logits_chunk_14_pad_type_0, strides = logits_chunk_14_strides_0, weight = logits_chunk_14_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_14")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_14_mul = mul(x = logits_chunk_14, y = temp_inverse)[name = string("logits_chunk_14_mul")];
+            tensor<int32, [1]> logits_chunk_14_max_axes_0 = const()[name = string("logits_chunk_14_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_14_max_keep_dims_0 = const()[name = string("logits_chunk_14_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_14_max = reduce_max(axes = logits_chunk_14_max_axes_0, keep_dims = logits_chunk_14_max_keep_dims_0, x = logits_chunk_14_mul)[name = string("logits_chunk_14_max")];
+            int32 logits_chunk_14_argmax_axis_0 = const()[name = string("logits_chunk_14_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_14_argmax_keep_dims_0 = const()[name = string("logits_chunk_14_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_14_argmax_output_dtype_0 = const()[name = string("logits_chunk_14_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> logits_chunk_14_argmax = reduce_argmax(axis = logits_chunk_14_argmax_axis_0, keep_dims = logits_chunk_14_argmax_keep_dims_0, output_dtype = logits_chunk_14_argmax_output_dtype_0, x = logits_chunk_14_mul)[name = string("logits_chunk_14_argmax")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_14_sub = sub(x = logits_chunk_14_mul, y = logits_chunk_14_max)[name = string("logits_chunk_14_sub")];
+            tensor<int32, [1]> logits_chunk_14_lse_sub_axes_0 = const()[name = string("logits_chunk_14_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_14_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_14_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_14_lse_sub = reduce_log_sum_exp(axes = logits_chunk_14_lse_sub_axes_0, keep_dims = logits_chunk_14_lse_sub_keep_dims_0, x = logits_chunk_14_sub)[name = string("logits_chunk_14_lse_sub")];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_14_lse = add(x = logits_chunk_14_lse_sub, y = logits_chunk_14_max)[name = string("logits_chunk_14_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_15_weight_0 = const()[name = string("logits_chunk_15_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125834304)))];
+            tensor<int32, [2]> logits_chunk_15_strides_0 = const()[name = string("logits_chunk_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_15_pad_type_0 = const()[name = string("logits_chunk_15_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_15_pad_0 = const()[name = string("logits_chunk_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_15_dilations_0 = const()[name = string("logits_chunk_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_15_groups_0 = const()[name = string("logits_chunk_15_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_15 = conv(dilations = logits_chunk_15_dilations_0, groups = logits_chunk_15_groups_0, pad = logits_chunk_15_pad_0, pad_type = logits_chunk_15_pad_type_0, strides = logits_chunk_15_strides_0, weight = logits_chunk_15_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_15")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_15_mul = mul(x = logits_chunk_15, y = temp_inverse)[name = string("logits_chunk_15_mul")];
+            tensor<int32, [1]> logits_chunk_15_max_axes_0 = const()[name = string("logits_chunk_15_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_15_max_keep_dims_0 = const()[name = string("logits_chunk_15_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_15_max = reduce_max(axes = logits_chunk_15_max_axes_0, keep_dims = logits_chunk_15_max_keep_dims_0, x = logits_chunk_15_mul)[name = string("logits_chunk_15_max")];
+            int32 logits_chunk_15_argmax_axis_0 = const()[name = string("logits_chunk_15_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_15_argmax_keep_dims_0 = const()[name = string("logits_chunk_15_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_15_argmax_output_dtype_0 = const()[name = string("logits_chunk_15_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 16]> logits_chunk_15_argmax = reduce_argmax(axis = logits_chunk_15_argmax_axis_0, keep_dims = logits_chunk_15_argmax_keep_dims_0, output_dtype = logits_chunk_15_argmax_output_dtype_0, x = logits_chunk_15_mul)[name = string("logits_chunk_15_argmax")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_15_sub = sub(x = logits_chunk_15_mul, y = logits_chunk_15_max)[name = string("logits_chunk_15_sub")];
+            tensor<int32, [1]> logits_chunk_15_lse_sub_axes_0 = const()[name = string("logits_chunk_15_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_15_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_15_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_15_lse_sub = reduce_log_sum_exp(axes = logits_chunk_15_lse_sub_axes_0, keep_dims = logits_chunk_15_lse_sub_keep_dims_0, x = logits_chunk_15_sub)[name = string("logits_chunk_15_lse_sub")];
+            tensor<fp16, [1, 1, 1, 16]> logits_chunk_15_lse = add(x = logits_chunk_15_lse_sub, y = logits_chunk_15_max)[name = string("logits_chunk_15_lse")];
+            int32 logits_lses_axis_0 = const()[name = string("logits_lses_axis_0"), val = int32(1)];
+            bool logits_lses_interleave_0 = const()[name = string("logits_lses_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 16]> logits_lses = concat(axis = logits_lses_axis_0, interleave = logits_lses_interleave_0, values = (logits_chunk_0_lse, logits_chunk_1_lse, logits_chunk_2_lse, logits_chunk_3_lse, logits_chunk_4_lse, logits_chunk_5_lse, logits_chunk_6_lse, logits_chunk_7_lse, logits_chunk_8_lse, logits_chunk_9_lse, logits_chunk_10_lse, logits_chunk_11_lse, logits_chunk_12_lse, logits_chunk_13_lse, logits_chunk_14_lse, logits_chunk_15_lse))[name = string("logits_lses")];
+            tensor<int32, [1]> logits_lses_max_axes_0 = const()[name = string("logits_lses_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_lses_max_keep_dims_0 = const()[name = string("logits_lses_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_lses_max = reduce_max(axes = logits_lses_max_axes_0, keep_dims = logits_lses_max_keep_dims_0, x = logits_lses)[name = string("logits_lses_max")];
+            tensor<fp16, [1, 16, 1, 16]> logits_lses_sub = sub(x = logits_lses, y = logits_lses_max)[name = string("logits_lses_sub")];
+            tensor<int32, [1]> logits_lses_logsumexp_axes_0 = const()[name = string("logits_lses_logsumexp_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_lses_logsumexp_keep_dims_0 = const()[name = string("logits_lses_logsumexp_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_lses_logsumexp = reduce_log_sum_exp(axes = logits_lses_logsumexp_axes_0, keep_dims = logits_lses_logsumexp_keep_dims_0, x = logits_lses_sub)[name = string("logits_lses_logsumexp")];
+            tensor<fp16, [1, 1, 1, 16]> logits_lse = add(x = logits_lses_logsumexp, y = logits_lses_max)[name = string("logits_lse")];
+            int32 logits_max_logits_chunks_axis_0 = const()[name = string("logits_max_logits_chunks_axis_0"), val = int32(1)];
+            bool logits_max_logits_chunks_interleave_0 = const()[name = string("logits_max_logits_chunks_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 16]> logits_max_logits_chunks = concat(axis = logits_max_logits_chunks_axis_0, interleave = logits_max_logits_chunks_interleave_0, values = (logits_chunk_0_max, logits_chunk_1_max, logits_chunk_2_max, logits_chunk_3_max, logits_chunk_4_max, logits_chunk_5_max, logits_chunk_6_max, logits_chunk_7_max, logits_chunk_8_max, logits_chunk_9_max, logits_chunk_10_max, logits_chunk_11_max, logits_chunk_12_max, logits_chunk_13_max, logits_chunk_14_max, logits_chunk_15_max))[name = string("logits_max_logits_chunks")];
+            tensor<int32, [1]> logits_max_logit_axes_0 = const()[name = string("logits_max_logit_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_max_logit_keep_dims_0 = const()[name = string("logits_max_logit_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 16]> logits_max_logit = reduce_max(axes = logits_max_logit_axes_0, keep_dims = logits_max_logit_keep_dims_0, x = logits_max_logits_chunks)[name = string("logits_max_logit")];
+            tensor<fp16, [1, 1, 1, 16]> logits_max_logit_sub = sub(x = logits_max_logit, y = logits_lse)[name = string("logits_max_logit_sub")];
+            tensor<fp16, [1, 1, 1, 16]> max_prob = exp(x = logits_max_logit_sub)[name = string("max_prob")];
+            tensor<fp16, [1, 1, 1, 16]> min_p_thresh = mul(x = max_prob, y = p)[name = string("min_p_thresh")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_0_sub_1 = sub(x = logits_chunk_0_mul, y = logits_lse)[name = string("logits_chunk_0_sub_1")];
+            tensor<fp16, [1, 2048, 1, 16]> probs_chunk_0 = exp(x = logits_chunk_0_sub_1)[name = string("probs_chunk_0")];
+            tensor<bool, [1, 2048, 1, 16]> mask_probs_chunk_0 = greater_equal(x = probs_chunk_0, y = min_p_thresh)[name = string("mask_probs_chunk_0")];
+            string mask_chunk_0_fp16_dtype_0 = const()[name = string("mask_chunk_0_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 16]> mask_chunk_0_fp16 = cast(dtype = mask_chunk_0_fp16_dtype_0, x = mask_probs_chunk_0)[name = string("cast_135")];
+            tensor<fp16, [1, 2048, 1, 16]> masked_probs_chunk_0 = select(a = probs_chunk_0, b = mask_chunk_0_fp16, cond = mask_probs_chunk_0)[name = string("masked_probs_chunk_0")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_1_sub_1 = sub(x = logits_chunk_1_mul, y = logits_lse)[name = string("logits_chunk_1_sub_1")];
+            tensor<fp16, [1, 2048, 1, 16]> probs_chunk_1 = exp(x = logits_chunk_1_sub_1)[name = string("probs_chunk_1")];
+            tensor<bool, [1, 2048, 1, 16]> mask_probs_chunk_1 = greater_equal(x = probs_chunk_1, y = min_p_thresh)[name = string("mask_probs_chunk_1")];
+            string mask_chunk_1_fp16_dtype_0 = const()[name = string("mask_chunk_1_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 16]> mask_chunk_1_fp16 = cast(dtype = mask_chunk_1_fp16_dtype_0, x = mask_probs_chunk_1)[name = string("cast_134")];
+            tensor<fp16, [1, 2048, 1, 16]> masked_probs_chunk_1 = select(a = probs_chunk_1, b = mask_chunk_1_fp16, cond = mask_probs_chunk_1)[name = string("masked_probs_chunk_1")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_2_sub_1 = sub(x = logits_chunk_2_mul, y = logits_lse)[name = string("logits_chunk_2_sub_1")];
+            tensor<fp16, [1, 2048, 1, 16]> probs_chunk_2 = exp(x = logits_chunk_2_sub_1)[name = string("probs_chunk_2")];
+            tensor<bool, [1, 2048, 1, 16]> mask_probs_chunk_2 = greater_equal(x = probs_chunk_2, y = min_p_thresh)[name = string("mask_probs_chunk_2")];
+            string mask_chunk_2_fp16_dtype_0 = const()[name = string("mask_chunk_2_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 16]> mask_chunk_2_fp16 = cast(dtype = mask_chunk_2_fp16_dtype_0, x = mask_probs_chunk_2)[name = string("cast_133")];
+            tensor<fp16, [1, 2048, 1, 16]> masked_probs_chunk_2 = select(a = probs_chunk_2, b = mask_chunk_2_fp16, cond = mask_probs_chunk_2)[name = string("masked_probs_chunk_2")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_3_sub_1 = sub(x = logits_chunk_3_mul, y = logits_lse)[name = string("logits_chunk_3_sub_1")];
+            tensor<fp16, [1, 2048, 1, 16]> probs_chunk_3 = exp(x = logits_chunk_3_sub_1)[name = string("probs_chunk_3")];
+            tensor<bool, [1, 2048, 1, 16]> mask_probs_chunk_3 = greater_equal(x = probs_chunk_3, y = min_p_thresh)[name = string("mask_probs_chunk_3")];
+            string mask_chunk_3_fp16_dtype_0 = const()[name = string("mask_chunk_3_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 16]> mask_chunk_3_fp16 = cast(dtype = mask_chunk_3_fp16_dtype_0, x = mask_probs_chunk_3)[name = string("cast_132")];
+            tensor<fp16, [1, 2048, 1, 16]> masked_probs_chunk_3 = select(a = probs_chunk_3, b = mask_chunk_3_fp16, cond = mask_probs_chunk_3)[name = string("masked_probs_chunk_3")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_4_sub_1 = sub(x = logits_chunk_4_mul, y = logits_lse)[name = string("logits_chunk_4_sub_1")];
+            tensor<fp16, [1, 2048, 1, 16]> probs_chunk_4 = exp(x = logits_chunk_4_sub_1)[name = string("probs_chunk_4")];
+            tensor<bool, [1, 2048, 1, 16]> mask_probs_chunk_4 = greater_equal(x = probs_chunk_4, y = min_p_thresh)[name = string("mask_probs_chunk_4")];
+            string mask_chunk_4_fp16_dtype_0 = const()[name = string("mask_chunk_4_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 16]> mask_chunk_4_fp16 = cast(dtype = mask_chunk_4_fp16_dtype_0, x = mask_probs_chunk_4)[name = string("cast_131")];
+            tensor<fp16, [1, 2048, 1, 16]> masked_probs_chunk_4 = select(a = probs_chunk_4, b = mask_chunk_4_fp16, cond = mask_probs_chunk_4)[name = string("masked_probs_chunk_4")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_5_sub_1 = sub(x = logits_chunk_5_mul, y = logits_lse)[name = string("logits_chunk_5_sub_1")];
+            tensor<fp16, [1, 2048, 1, 16]> probs_chunk_5 = exp(x = logits_chunk_5_sub_1)[name = string("probs_chunk_5")];
+            tensor<bool, [1, 2048, 1, 16]> mask_probs_chunk_5 = greater_equal(x = probs_chunk_5, y = min_p_thresh)[name = string("mask_probs_chunk_5")];
+            string mask_chunk_5_fp16_dtype_0 = const()[name = string("mask_chunk_5_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 16]> mask_chunk_5_fp16 = cast(dtype = mask_chunk_5_fp16_dtype_0, x = mask_probs_chunk_5)[name = string("cast_130")];
+            tensor<fp16, [1, 2048, 1, 16]> masked_probs_chunk_5 = select(a = probs_chunk_5, b = mask_chunk_5_fp16, cond = mask_probs_chunk_5)[name = string("masked_probs_chunk_5")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_6_sub_1 = sub(x = logits_chunk_6_mul, y = logits_lse)[name = string("logits_chunk_6_sub_1")];
+            tensor<fp16, [1, 2048, 1, 16]> probs_chunk_6 = exp(x = logits_chunk_6_sub_1)[name = string("probs_chunk_6")];
+            tensor<bool, [1, 2048, 1, 16]> mask_probs_chunk_6 = greater_equal(x = probs_chunk_6, y = min_p_thresh)[name = string("mask_probs_chunk_6")];
+            string mask_chunk_6_fp16_dtype_0 = const()[name = string("mask_chunk_6_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 16]> mask_chunk_6_fp16 = cast(dtype = mask_chunk_6_fp16_dtype_0, x = mask_probs_chunk_6)[name = string("cast_129")];
+            tensor<fp16, [1, 2048, 1, 16]> masked_probs_chunk_6 = select(a = probs_chunk_6, b = mask_chunk_6_fp16, cond = mask_probs_chunk_6)[name = string("masked_probs_chunk_6")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_7_sub_1 = sub(x = logits_chunk_7_mul, y = logits_lse)[name = string("logits_chunk_7_sub_1")];
+            tensor<fp16, [1, 2048, 1, 16]> probs_chunk_7 = exp(x = logits_chunk_7_sub_1)[name = string("probs_chunk_7")];
+            tensor<bool, [1, 2048, 1, 16]> mask_probs_chunk_7 = greater_equal(x = probs_chunk_7, y = min_p_thresh)[name = string("mask_probs_chunk_7")];
+            string mask_chunk_7_fp16_dtype_0 = const()[name = string("mask_chunk_7_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 16]> mask_chunk_7_fp16 = cast(dtype = mask_chunk_7_fp16_dtype_0, x = mask_probs_chunk_7)[name = string("cast_128")];
+            tensor<fp16, [1, 2048, 1, 16]> masked_probs_chunk_7 = select(a = probs_chunk_7, b = mask_chunk_7_fp16, cond = mask_probs_chunk_7)[name = string("masked_probs_chunk_7")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_8_sub_1 = sub(x = logits_chunk_8_mul, y = logits_lse)[name = string("logits_chunk_8_sub_1")];
+            tensor<fp16, [1, 2048, 1, 16]> probs_chunk_8 = exp(x = logits_chunk_8_sub_1)[name = string("probs_chunk_8")];
+            tensor<bool, [1, 2048, 1, 16]> mask_probs_chunk_8 = greater_equal(x = probs_chunk_8, y = min_p_thresh)[name = string("mask_probs_chunk_8")];
+            string mask_chunk_8_fp16_dtype_0 = const()[name = string("mask_chunk_8_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 16]> mask_chunk_8_fp16 = cast(dtype = mask_chunk_8_fp16_dtype_0, x = mask_probs_chunk_8)[name = string("cast_127")];
+            tensor<fp16, [1, 2048, 1, 16]> masked_probs_chunk_8 = select(a = probs_chunk_8, b = mask_chunk_8_fp16, cond = mask_probs_chunk_8)[name = string("masked_probs_chunk_8")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_9_sub_1 = sub(x = logits_chunk_9_mul, y = logits_lse)[name = string("logits_chunk_9_sub_1")];
+            tensor<fp16, [1, 2048, 1, 16]> probs_chunk_9 = exp(x = logits_chunk_9_sub_1)[name = string("probs_chunk_9")];
+            tensor<bool, [1, 2048, 1, 16]> mask_probs_chunk_9 = greater_equal(x = probs_chunk_9, y = min_p_thresh)[name = string("mask_probs_chunk_9")];
+            string mask_chunk_9_fp16_dtype_0 = const()[name = string("mask_chunk_9_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 16]> mask_chunk_9_fp16 = cast(dtype = mask_chunk_9_fp16_dtype_0, x = mask_probs_chunk_9)[name = string("cast_126")];
+            tensor<fp16, [1, 2048, 1, 16]> masked_probs_chunk_9 = select(a = probs_chunk_9, b = mask_chunk_9_fp16, cond = mask_probs_chunk_9)[name = string("masked_probs_chunk_9")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_10_sub_1 = sub(x = logits_chunk_10_mul, y = logits_lse)[name = string("logits_chunk_10_sub_1")];
+            tensor<fp16, [1, 2048, 1, 16]> probs_chunk_10 = exp(x = logits_chunk_10_sub_1)[name = string("probs_chunk_10")];
+            tensor<bool, [1, 2048, 1, 16]> mask_probs_chunk_10 = greater_equal(x = probs_chunk_10, y = min_p_thresh)[name = string("mask_probs_chunk_10")];
+            string mask_chunk_10_fp16_dtype_0 = const()[name = string("mask_chunk_10_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 16]> mask_chunk_10_fp16 = cast(dtype = mask_chunk_10_fp16_dtype_0, x = mask_probs_chunk_10)[name = string("cast_125")];
+            tensor<fp16, [1, 2048, 1, 16]> masked_probs_chunk_10 = select(a = probs_chunk_10, b = mask_chunk_10_fp16, cond = mask_probs_chunk_10)[name = string("masked_probs_chunk_10")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_11_sub_1 = sub(x = logits_chunk_11_mul, y = logits_lse)[name = string("logits_chunk_11_sub_1")];
+            tensor<fp16, [1, 2048, 1, 16]> probs_chunk_11 = exp(x = logits_chunk_11_sub_1)[name = string("probs_chunk_11")];
+            tensor<bool, [1, 2048, 1, 16]> mask_probs_chunk_11 = greater_equal(x = probs_chunk_11, y = min_p_thresh)[name = string("mask_probs_chunk_11")];
+            string mask_chunk_11_fp16_dtype_0 = const()[name = string("mask_chunk_11_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 16]> mask_chunk_11_fp16 = cast(dtype = mask_chunk_11_fp16_dtype_0, x = mask_probs_chunk_11)[name = string("cast_124")];
+            tensor<fp16, [1, 2048, 1, 16]> masked_probs_chunk_11 = select(a = probs_chunk_11, b = mask_chunk_11_fp16, cond = mask_probs_chunk_11)[name = string("masked_probs_chunk_11")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_12_sub_1 = sub(x = logits_chunk_12_mul, y = logits_lse)[name = string("logits_chunk_12_sub_1")];
+            tensor<fp16, [1, 2048, 1, 16]> probs_chunk_12 = exp(x = logits_chunk_12_sub_1)[name = string("probs_chunk_12")];
+            tensor<bool, [1, 2048, 1, 16]> mask_probs_chunk_12 = greater_equal(x = probs_chunk_12, y = min_p_thresh)[name = string("mask_probs_chunk_12")];
+            string mask_chunk_12_fp16_dtype_0 = const()[name = string("mask_chunk_12_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 16]> mask_chunk_12_fp16 = cast(dtype = mask_chunk_12_fp16_dtype_0, x = mask_probs_chunk_12)[name = string("cast_123")];
+            tensor<fp16, [1, 2048, 1, 16]> masked_probs_chunk_12 = select(a = probs_chunk_12, b = mask_chunk_12_fp16, cond = mask_probs_chunk_12)[name = string("masked_probs_chunk_12")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_13_sub_1 = sub(x = logits_chunk_13_mul, y = logits_lse)[name = string("logits_chunk_13_sub_1")];
+            tensor<fp16, [1, 2048, 1, 16]> probs_chunk_13 = exp(x = logits_chunk_13_sub_1)[name = string("probs_chunk_13")];
+            tensor<bool, [1, 2048, 1, 16]> mask_probs_chunk_13 = greater_equal(x = probs_chunk_13, y = min_p_thresh)[name = string("mask_probs_chunk_13")];
+            string mask_chunk_13_fp16_dtype_0 = const()[name = string("mask_chunk_13_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 16]> mask_chunk_13_fp16 = cast(dtype = mask_chunk_13_fp16_dtype_0, x = mask_probs_chunk_13)[name = string("cast_122")];
+            tensor<fp16, [1, 2048, 1, 16]> masked_probs_chunk_13 = select(a = probs_chunk_13, b = mask_chunk_13_fp16, cond = mask_probs_chunk_13)[name = string("masked_probs_chunk_13")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_14_sub_1 = sub(x = logits_chunk_14_mul, y = logits_lse)[name = string("logits_chunk_14_sub_1")];
+            tensor<fp16, [1, 2048, 1, 16]> probs_chunk_14 = exp(x = logits_chunk_14_sub_1)[name = string("probs_chunk_14")];
+            tensor<bool, [1, 2048, 1, 16]> mask_probs_chunk_14 = greater_equal(x = probs_chunk_14, y = min_p_thresh)[name = string("mask_probs_chunk_14")];
+            string mask_chunk_14_fp16_dtype_0 = const()[name = string("mask_chunk_14_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 16]> mask_chunk_14_fp16 = cast(dtype = mask_chunk_14_fp16_dtype_0, x = mask_probs_chunk_14)[name = string("cast_121")];
+            tensor<fp16, [1, 2048, 1, 16]> masked_probs_chunk_14 = select(a = probs_chunk_14, b = mask_chunk_14_fp16, cond = mask_probs_chunk_14)[name = string("masked_probs_chunk_14")];
+            tensor<fp16, [1, 2048, 1, 16]> logits_chunk_15_sub_1 = sub(x = logits_chunk_15_mul, y = logits_lse)[name = string("logits_chunk_15_sub_1")];
+            tensor<fp16, [1, 2048, 1, 16]> probs_chunk_15 = exp(x = logits_chunk_15_sub_1)[name = string("probs_chunk_15")];
+            tensor<bool, [1, 2048, 1, 16]> mask_probs_chunk_15 = greater_equal(x = probs_chunk_15, y = min_p_thresh)[name = string("mask_probs_chunk_15")];
+            string mask_chunk_15_fp16_dtype_0 = const()[name = string("mask_chunk_15_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 16]> mask_chunk_15_fp16 = cast(dtype = mask_chunk_15_fp16_dtype_0, x = mask_probs_chunk_15)[name = string("cast_120")];
+            tensor<fp16, [1, 2048, 1, 16]> masked_probs_chunk_15 = select(a = probs_chunk_15, b = mask_chunk_15_fp16, cond = mask_probs_chunk_15)[name = string("masked_probs_chunk_15")];
+            int32 probs_axis_0 = const()[name = string("probs_axis_0"), val = int32(1)];
+            bool probs_interleave_0 = const()[name = string("probs_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 32768, 1, 16]> probs = concat(axis = probs_axis_0, interleave = probs_interleave_0, values = (masked_probs_chunk_0, masked_probs_chunk_1, masked_probs_chunk_2, masked_probs_chunk_3, masked_probs_chunk_4, masked_probs_chunk_5, masked_probs_chunk_6, masked_probs_chunk_7, masked_probs_chunk_8, masked_probs_chunk_9, masked_probs_chunk_10, masked_probs_chunk_11, masked_probs_chunk_12, masked_probs_chunk_13, masked_probs_chunk_14, masked_probs_chunk_15))[name = string("probs")];
+            string probs_fp32_dtype_0 = const()[name = string("probs_fp32_dtype_0"), val = string("fp32")];
+            int32 probs_cumsum_axis_0 = const()[name = string("probs_cumsum_axis_0"), val = int32(1)];
+            bool probs_cumsum_exclusive_0 = const()[name = string("probs_cumsum_exclusive_0"), val = bool(false)];
+            bool probs_cumsum_reverse_0 = const()[name = string("probs_cumsum_reverse_0"), val = bool(false)];
+            tensor<fp32, [1, 32768, 1, 16]> probs_fp32 = cast(dtype = probs_fp32_dtype_0, x = probs)[name = string("cast_119")];
+            tensor<fp32, [1, 32768, 1, 16]> probs_cumsum = cumsum(axis = probs_cumsum_axis_0, exclusive = probs_cumsum_exclusive_0, reverse = probs_cumsum_reverse_0, x = probs_fp32)[name = string("probs_cumsum")];
+            tensor<int32, [1]> probs_sum_indices_0 = const()[name = string("probs_sum_indices_0"), val = tensor<int32, [1]>([32767])];
+            int32 probs_sum_axis_0 = const()[name = string("probs_sum_axis_0"), val = int32(1)];
+            int32 probs_sum_batch_dims_0 = const()[name = string("probs_sum_batch_dims_0"), val = int32(0)];
+            bool probs_sum_validate_indices_0 = const()[name = string("probs_sum_validate_indices_0"), val = bool(false)];
+            tensor<fp32, [1, 1, 1, 16]> probs_sum = gather(axis = probs_sum_axis_0, batch_dims = probs_sum_batch_dims_0, indices = probs_sum_indices_0, validate_indices = probs_sum_validate_indices_0, x = probs_cumsum)[name = string("probs_sum")];
+            tensor<fp32, [1, 1, 1, 16]> random_number_scaled = mul(x = random_number, y = probs_sum)[name = string("random_number_scaled")];
+            tensor<bool, [1, 32768, 1, 16]> probs_greater = greater(x = probs_cumsum, y = random_number_scaled)[name = string("probs_greater")];
+            string probs_greater_int32_dtype_0 = const()[name = string("probs_greater_int32_dtype_0"), val = string("int32")];
+            int32 sampled_index_axis_0 = const()[name = string("sampled_index_axis_0"), val = int32(1)];
+            bool sampled_index_keep_dims_0 = const()[name = string("sampled_index_keep_dims_0"), val = bool(true)];
+            string sampled_index_output_dtype_0 = const()[name = string("sampled_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 32768, 1, 16]> probs_greater_int32 = cast(dtype = probs_greater_int32_dtype_0, x = probs_greater)[name = string("cast_118")];
+            tensor<int32, [1, 1, 1, 16]> sampled_index = reduce_argmax(axis = sampled_index_axis_0, keep_dims = sampled_index_keep_dims_0, output_dtype = sampled_index_output_dtype_0, x = probs_greater_int32)[name = string("sampled_index")];
+            int32 sampled_index_probability_axis_0 = const()[name = string("sampled_index_probability_axis_0"), val = int32(1)];
+            bool sampled_index_probability_validate_indices_0 = const()[name = string("sampled_index_probability_validate_indices_0"), val = bool(false)];
+            tensor<fp32, [1, 1, 1, 16]> sampled_index_probability = gather_along_axis(axis = sampled_index_probability_axis_0, indices = sampled_index, validate_indices = sampled_index_probability_validate_indices_0, x = probs_fp32)[name = string("sampled_index_probability")];
+            int32 max_logit_index_axis_0 = const()[name = string("max_logit_index_axis_0"), val = int32(1)];
+            bool max_logit_index_keep_dims_0 = const()[name = string("max_logit_index_keep_dims_0"), val = bool(true)];
+            string max_logit_index_output_dtype_0 = const()[name = string("max_logit_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 1, 1, 16]> max_logit_index = reduce_argmax(axis = max_logit_index_axis_0, keep_dims = max_logit_index_keep_dims_0, output_dtype = max_logit_index_output_dtype_0, x = logits_max_logits_chunks)[name = string("max_logit_index")];
+            string indices_chunk_0_int32_dtype_0 = const()[name = string("indices_chunk_0_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_1_int32_dtype_0 = const()[name = string("indices_chunk_1_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_2_int32_dtype_0 = const()[name = string("indices_chunk_2_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_3_int32_dtype_0 = const()[name = string("indices_chunk_3_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_4_int32_dtype_0 = const()[name = string("indices_chunk_4_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_5_int32_dtype_0 = const()[name = string("indices_chunk_5_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_6_int32_dtype_0 = const()[name = string("indices_chunk_6_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_7_int32_dtype_0 = const()[name = string("indices_chunk_7_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_8_int32_dtype_0 = const()[name = string("indices_chunk_8_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_9_int32_dtype_0 = const()[name = string("indices_chunk_9_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_10_int32_dtype_0 = const()[name = string("indices_chunk_10_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_11_int32_dtype_0 = const()[name = string("indices_chunk_11_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_12_int32_dtype_0 = const()[name = string("indices_chunk_12_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_13_int32_dtype_0 = const()[name = string("indices_chunk_13_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_14_int32_dtype_0 = const()[name = string("indices_chunk_14_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_15_int32_dtype_0 = const()[name = string("indices_chunk_15_int32_dtype_0"), val = string("int32")];
+            int32 indices_axis_0 = const()[name = string("indices_axis_0"), val = int32(1)];
+            bool indices_interleave_0 = const()[name = string("indices_interleave_0"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_15_int32 = cast(dtype = indices_chunk_15_int32_dtype_0, x = logits_chunk_15_argmax)[name = string("cast_102")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_14_int32 = cast(dtype = indices_chunk_14_int32_dtype_0, x = logits_chunk_14_argmax)[name = string("cast_103")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_13_int32 = cast(dtype = indices_chunk_13_int32_dtype_0, x = logits_chunk_13_argmax)[name = string("cast_104")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_12_int32 = cast(dtype = indices_chunk_12_int32_dtype_0, x = logits_chunk_12_argmax)[name = string("cast_105")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_11_int32 = cast(dtype = indices_chunk_11_int32_dtype_0, x = logits_chunk_11_argmax)[name = string("cast_106")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_10_int32 = cast(dtype = indices_chunk_10_int32_dtype_0, x = logits_chunk_10_argmax)[name = string("cast_107")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_9_int32 = cast(dtype = indices_chunk_9_int32_dtype_0, x = logits_chunk_9_argmax)[name = string("cast_108")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_8_int32 = cast(dtype = indices_chunk_8_int32_dtype_0, x = logits_chunk_8_argmax)[name = string("cast_109")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_7_int32 = cast(dtype = indices_chunk_7_int32_dtype_0, x = logits_chunk_7_argmax)[name = string("cast_110")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_6_int32 = cast(dtype = indices_chunk_6_int32_dtype_0, x = logits_chunk_6_argmax)[name = string("cast_111")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_5_int32 = cast(dtype = indices_chunk_5_int32_dtype_0, x = logits_chunk_5_argmax)[name = string("cast_112")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_4_int32 = cast(dtype = indices_chunk_4_int32_dtype_0, x = logits_chunk_4_argmax)[name = string("cast_113")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_3_int32 = cast(dtype = indices_chunk_3_int32_dtype_0, x = logits_chunk_3_argmax)[name = string("cast_114")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_2_int32 = cast(dtype = indices_chunk_2_int32_dtype_0, x = logits_chunk_2_argmax)[name = string("cast_115")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_1_int32 = cast(dtype = indices_chunk_1_int32_dtype_0, x = logits_chunk_1_argmax)[name = string("cast_116")];
+            tensor<int32, [1, 1, 1, 16]> indices_chunk_0_int32 = cast(dtype = indices_chunk_0_int32_dtype_0, x = logits_chunk_0_argmax)[name = string("cast_117")];
+            tensor<int32, [1, 16, 1, 16]> indices = concat(axis = indices_axis_0, interleave = indices_interleave_0, values = (indices_chunk_0_int32, indices_chunk_1_int32, indices_chunk_2_int32, indices_chunk_3_int32, indices_chunk_4_int32, indices_chunk_5_int32, indices_chunk_6_int32, indices_chunk_7_int32, indices_chunk_8_int32, indices_chunk_9_int32, indices_chunk_10_int32, indices_chunk_11_int32, indices_chunk_12_int32, indices_chunk_13_int32, indices_chunk_14_int32, indices_chunk_15_int32))[name = string("indices")];
+            int32 argmax_chunks_axis_0 = const()[name = string("argmax_chunks_axis_0"), val = int32(1)];
+            bool argmax_chunks_validate_indices_0 = const()[name = string("argmax_chunks_validate_indices_0"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 16]> argmax_chunks = gather_along_axis(axis = argmax_chunks_axis_0, indices = max_logit_index, validate_indices = argmax_chunks_validate_indices_0, x = indices)[name = string("argmax_chunks")];
+            int32 mul_0_x_0 = const()[name = string("mul_0_x_0"), val = int32(2048)];
+            tensor<int32, [1, 1, 1, 16]> mul_0 = mul(x = mul_0_x_0, y = max_logit_index)[name = string("mul_0")];
+            tensor<int32, [1, 1, 1, 16]> argmax = add(x = argmax_chunks, y = mul_0)[name = string("argmax")];
+        } -> (sampled_index, sampled_index_probability, argmax, max_prob);
+    func min_p_length_32<ios18>(tensor<fp16, [1, 2048, 1, 32]> hidden_states, tensor<fp16, [1]> p, tensor<fp32, [32]> random_number, tensor<fp16, [1]> temp) {
+            tensor<fp16, [1, 2048, 1, 32]> final_norm_rmsnorm_abs = abs(x = hidden_states)[name = string("final_norm_rmsnorm_abs")];
+            tensor<int32, [1]> final_norm_rmsnorm_maxval_axes_0 = const()[name = string("final_norm_rmsnorm_maxval_axes_0"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_maxval_keep_dims_0 = const()[name = string("final_norm_rmsnorm_maxval_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> final_norm_rmsnorm_maxval = reduce_max(axes = final_norm_rmsnorm_maxval_axes_0, keep_dims = final_norm_rmsnorm_maxval_keep_dims_0, x = final_norm_rmsnorm_abs)[name = string("final_norm_rmsnorm_maxval")];
+            fp16 final_norm_rmsnorm_maxval_clipped_alpha_0 = const()[name = string("final_norm_rmsnorm_maxval_clipped_alpha_0"), val = fp16(0x1p-14)];
+            fp16 final_norm_rmsnorm_maxval_clipped_beta_0 = const()[name = string("final_norm_rmsnorm_maxval_clipped_beta_0"), val = fp16(inf)];
+            tensor<fp16, [1, 1, 1, 32]> final_norm_rmsnorm_maxval_clipped = clip(alpha = final_norm_rmsnorm_maxval_clipped_alpha_0, beta = final_norm_rmsnorm_maxval_clipped_beta_0, x = final_norm_rmsnorm_maxval)[name = string("final_norm_rmsnorm_maxval_clipped")];
+            tensor<fp16, [1, 2048, 1, 32]> final_norm_rmsnorm_scaled = real_div(x = hidden_states, y = final_norm_rmsnorm_maxval_clipped)[name = string("final_norm_rmsnorm_scaled")];
+            tensor<int32, [1]> final_norm_rmsnorm_squared_sum_axes_0 = const()[name = string("final_norm_rmsnorm_squared_sum_axes_0"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_squared_sum_keep_dims_0 = const()[name = string("final_norm_rmsnorm_squared_sum_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> final_norm_rmsnorm_squared_sum = reduce_sum_square(axes = final_norm_rmsnorm_squared_sum_axes_0, keep_dims = final_norm_rmsnorm_squared_sum_keep_dims_0, x = final_norm_rmsnorm_scaled)[name = string("final_norm_rmsnorm_squared_sum")];
+            fp16 final_norm_rmsnorm_rsqrt_epsilon_0 = const()[name = string("final_norm_rmsnorm_rsqrt_epsilon_0"), val = fp16(0x1p-14)];
+            tensor<fp16, [1, 1, 1, 32]> final_norm_rmsnorm_rsqrt = rsqrt(epsilon = final_norm_rmsnorm_rsqrt_epsilon_0, x = final_norm_rmsnorm_squared_sum)[name = string("final_norm_rmsnorm_rsqrt")];
+            fp16 final_norm_rmsnorm_dim_scaled_y_0 = const()[name = string("final_norm_rmsnorm_dim_scaled_y_0"), val = fp16(0x1.6ap+5)];
+            tensor<fp16, [1, 2048, 1, 32]> final_norm_rmsnorm_dim_scaled = mul(x = final_norm_rmsnorm_scaled, y = final_norm_rmsnorm_dim_scaled_y_0)[name = string("final_norm_rmsnorm_dim_scaled")];
+            tensor<fp16, [1, 2048, 1, 32]> final_norm_rmsnorm_normalized = mul(x = final_norm_rmsnorm_dim_scaled, y = final_norm_rmsnorm_rsqrt)[name = string("final_norm_rmsnorm_normalized")];
+            tensor<fp16, [2048, 1, 1]> final_norm_rmsnorm_y_0 = const()[name = string("final_norm_rmsnorm_y_0"), val = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 2048, 1, 32]> final_norm_rmsnorm = mul(x = final_norm_rmsnorm_normalized, y = final_norm_rmsnorm_y_0)[name = string("final_norm_rmsnorm")];
+            fp16 temp_inverse_epsilon_0 = const()[name = string("temp_inverse_epsilon_0"), val = fp16(0x0p+0)];
+            tensor<fp16, [1]> temp_inverse = inverse(epsilon = temp_inverse_epsilon_0, x = temp)[name = string("temp_inverse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_0_weight_0 = const()[name = string("logits_chunk_0_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4224)))];
+            tensor<int32, [2]> logits_chunk_0_strides_0 = const()[name = string("logits_chunk_0_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_0_pad_type_0 = const()[name = string("logits_chunk_0_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_0_pad_0 = const()[name = string("logits_chunk_0_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_0_dilations_0 = const()[name = string("logits_chunk_0_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_0_groups_0 = const()[name = string("logits_chunk_0_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_0 = conv(dilations = logits_chunk_0_dilations_0, groups = logits_chunk_0_groups_0, pad = logits_chunk_0_pad_0, pad_type = logits_chunk_0_pad_type_0, strides = logits_chunk_0_strides_0, weight = logits_chunk_0_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_0")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_0_mul = mul(x = logits_chunk_0, y = temp_inverse)[name = string("logits_chunk_0_mul")];
+            tensor<int32, [1]> logits_chunk_0_max_axes_0 = const()[name = string("logits_chunk_0_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_0_max_keep_dims_0 = const()[name = string("logits_chunk_0_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_0_max = reduce_max(axes = logits_chunk_0_max_axes_0, keep_dims = logits_chunk_0_max_keep_dims_0, x = logits_chunk_0_mul)[name = string("logits_chunk_0_max")];
+            int32 logits_chunk_0_argmax_axis_0 = const()[name = string("logits_chunk_0_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_0_argmax_keep_dims_0 = const()[name = string("logits_chunk_0_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_0_argmax_output_dtype_0 = const()[name = string("logits_chunk_0_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> logits_chunk_0_argmax = reduce_argmax(axis = logits_chunk_0_argmax_axis_0, keep_dims = logits_chunk_0_argmax_keep_dims_0, output_dtype = logits_chunk_0_argmax_output_dtype_0, x = logits_chunk_0_mul)[name = string("logits_chunk_0_argmax")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_0_sub = sub(x = logits_chunk_0_mul, y = logits_chunk_0_max)[name = string("logits_chunk_0_sub")];
+            tensor<int32, [1]> logits_chunk_0_lse_sub_axes_0 = const()[name = string("logits_chunk_0_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_0_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_0_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_0_lse_sub = reduce_log_sum_exp(axes = logits_chunk_0_lse_sub_axes_0, keep_dims = logits_chunk_0_lse_sub_keep_dims_0, x = logits_chunk_0_sub)[name = string("logits_chunk_0_lse_sub")];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_0_lse = add(x = logits_chunk_0_lse_sub, y = logits_chunk_0_max)[name = string("logits_chunk_0_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_1_weight_0 = const()[name = string("logits_chunk_1_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8392896)))];
+            tensor<int32, [2]> logits_chunk_1_strides_0 = const()[name = string("logits_chunk_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_1_pad_type_0 = const()[name = string("logits_chunk_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_1_pad_0 = const()[name = string("logits_chunk_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_1_dilations_0 = const()[name = string("logits_chunk_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_1_groups_0 = const()[name = string("logits_chunk_1_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_1 = conv(dilations = logits_chunk_1_dilations_0, groups = logits_chunk_1_groups_0, pad = logits_chunk_1_pad_0, pad_type = logits_chunk_1_pad_type_0, strides = logits_chunk_1_strides_0, weight = logits_chunk_1_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_1")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_1_mul = mul(x = logits_chunk_1, y = temp_inverse)[name = string("logits_chunk_1_mul")];
+            tensor<int32, [1]> logits_chunk_1_max_axes_0 = const()[name = string("logits_chunk_1_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_1_max_keep_dims_0 = const()[name = string("logits_chunk_1_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_1_max = reduce_max(axes = logits_chunk_1_max_axes_0, keep_dims = logits_chunk_1_max_keep_dims_0, x = logits_chunk_1_mul)[name = string("logits_chunk_1_max")];
+            int32 logits_chunk_1_argmax_axis_0 = const()[name = string("logits_chunk_1_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_1_argmax_keep_dims_0 = const()[name = string("logits_chunk_1_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_1_argmax_output_dtype_0 = const()[name = string("logits_chunk_1_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> logits_chunk_1_argmax = reduce_argmax(axis = logits_chunk_1_argmax_axis_0, keep_dims = logits_chunk_1_argmax_keep_dims_0, output_dtype = logits_chunk_1_argmax_output_dtype_0, x = logits_chunk_1_mul)[name = string("logits_chunk_1_argmax")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_1_sub = sub(x = logits_chunk_1_mul, y = logits_chunk_1_max)[name = string("logits_chunk_1_sub")];
+            tensor<int32, [1]> logits_chunk_1_lse_sub_axes_0 = const()[name = string("logits_chunk_1_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_1_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_1_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_1_lse_sub = reduce_log_sum_exp(axes = logits_chunk_1_lse_sub_axes_0, keep_dims = logits_chunk_1_lse_sub_keep_dims_0, x = logits_chunk_1_sub)[name = string("logits_chunk_1_lse_sub")];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_1_lse = add(x = logits_chunk_1_lse_sub, y = logits_chunk_1_max)[name = string("logits_chunk_1_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_2_weight_0 = const()[name = string("logits_chunk_2_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16781568)))];
+            tensor<int32, [2]> logits_chunk_2_strides_0 = const()[name = string("logits_chunk_2_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_2_pad_type_0 = const()[name = string("logits_chunk_2_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_2_pad_0 = const()[name = string("logits_chunk_2_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_2_dilations_0 = const()[name = string("logits_chunk_2_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_2_groups_0 = const()[name = string("logits_chunk_2_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_2 = conv(dilations = logits_chunk_2_dilations_0, groups = logits_chunk_2_groups_0, pad = logits_chunk_2_pad_0, pad_type = logits_chunk_2_pad_type_0, strides = logits_chunk_2_strides_0, weight = logits_chunk_2_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_2")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_2_mul = mul(x = logits_chunk_2, y = temp_inverse)[name = string("logits_chunk_2_mul")];
+            tensor<int32, [1]> logits_chunk_2_max_axes_0 = const()[name = string("logits_chunk_2_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_2_max_keep_dims_0 = const()[name = string("logits_chunk_2_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_2_max = reduce_max(axes = logits_chunk_2_max_axes_0, keep_dims = logits_chunk_2_max_keep_dims_0, x = logits_chunk_2_mul)[name = string("logits_chunk_2_max")];
+            int32 logits_chunk_2_argmax_axis_0 = const()[name = string("logits_chunk_2_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_2_argmax_keep_dims_0 = const()[name = string("logits_chunk_2_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_2_argmax_output_dtype_0 = const()[name = string("logits_chunk_2_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> logits_chunk_2_argmax = reduce_argmax(axis = logits_chunk_2_argmax_axis_0, keep_dims = logits_chunk_2_argmax_keep_dims_0, output_dtype = logits_chunk_2_argmax_output_dtype_0, x = logits_chunk_2_mul)[name = string("logits_chunk_2_argmax")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_2_sub = sub(x = logits_chunk_2_mul, y = logits_chunk_2_max)[name = string("logits_chunk_2_sub")];
+            tensor<int32, [1]> logits_chunk_2_lse_sub_axes_0 = const()[name = string("logits_chunk_2_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_2_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_2_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_2_lse_sub = reduce_log_sum_exp(axes = logits_chunk_2_lse_sub_axes_0, keep_dims = logits_chunk_2_lse_sub_keep_dims_0, x = logits_chunk_2_sub)[name = string("logits_chunk_2_lse_sub")];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_2_lse = add(x = logits_chunk_2_lse_sub, y = logits_chunk_2_max)[name = string("logits_chunk_2_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_3_weight_0 = const()[name = string("logits_chunk_3_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25170240)))];
+            tensor<int32, [2]> logits_chunk_3_strides_0 = const()[name = string("logits_chunk_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_3_pad_type_0 = const()[name = string("logits_chunk_3_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_3_pad_0 = const()[name = string("logits_chunk_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_3_dilations_0 = const()[name = string("logits_chunk_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_3_groups_0 = const()[name = string("logits_chunk_3_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_3 = conv(dilations = logits_chunk_3_dilations_0, groups = logits_chunk_3_groups_0, pad = logits_chunk_3_pad_0, pad_type = logits_chunk_3_pad_type_0, strides = logits_chunk_3_strides_0, weight = logits_chunk_3_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_3")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_3_mul = mul(x = logits_chunk_3, y = temp_inverse)[name = string("logits_chunk_3_mul")];
+            tensor<int32, [1]> logits_chunk_3_max_axes_0 = const()[name = string("logits_chunk_3_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_3_max_keep_dims_0 = const()[name = string("logits_chunk_3_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_3_max = reduce_max(axes = logits_chunk_3_max_axes_0, keep_dims = logits_chunk_3_max_keep_dims_0, x = logits_chunk_3_mul)[name = string("logits_chunk_3_max")];
+            int32 logits_chunk_3_argmax_axis_0 = const()[name = string("logits_chunk_3_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_3_argmax_keep_dims_0 = const()[name = string("logits_chunk_3_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_3_argmax_output_dtype_0 = const()[name = string("logits_chunk_3_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> logits_chunk_3_argmax = reduce_argmax(axis = logits_chunk_3_argmax_axis_0, keep_dims = logits_chunk_3_argmax_keep_dims_0, output_dtype = logits_chunk_3_argmax_output_dtype_0, x = logits_chunk_3_mul)[name = string("logits_chunk_3_argmax")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_3_sub = sub(x = logits_chunk_3_mul, y = logits_chunk_3_max)[name = string("logits_chunk_3_sub")];
+            tensor<int32, [1]> logits_chunk_3_lse_sub_axes_0 = const()[name = string("logits_chunk_3_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_3_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_3_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_3_lse_sub = reduce_log_sum_exp(axes = logits_chunk_3_lse_sub_axes_0, keep_dims = logits_chunk_3_lse_sub_keep_dims_0, x = logits_chunk_3_sub)[name = string("logits_chunk_3_lse_sub")];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_3_lse = add(x = logits_chunk_3_lse_sub, y = logits_chunk_3_max)[name = string("logits_chunk_3_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_4_weight_0 = const()[name = string("logits_chunk_4_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33558912)))];
+            tensor<int32, [2]> logits_chunk_4_strides_0 = const()[name = string("logits_chunk_4_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_4_pad_type_0 = const()[name = string("logits_chunk_4_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_4_pad_0 = const()[name = string("logits_chunk_4_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_4_dilations_0 = const()[name = string("logits_chunk_4_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_4_groups_0 = const()[name = string("logits_chunk_4_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_4 = conv(dilations = logits_chunk_4_dilations_0, groups = logits_chunk_4_groups_0, pad = logits_chunk_4_pad_0, pad_type = logits_chunk_4_pad_type_0, strides = logits_chunk_4_strides_0, weight = logits_chunk_4_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_4")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_4_mul = mul(x = logits_chunk_4, y = temp_inverse)[name = string("logits_chunk_4_mul")];
+            tensor<int32, [1]> logits_chunk_4_max_axes_0 = const()[name = string("logits_chunk_4_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_4_max_keep_dims_0 = const()[name = string("logits_chunk_4_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_4_max = reduce_max(axes = logits_chunk_4_max_axes_0, keep_dims = logits_chunk_4_max_keep_dims_0, x = logits_chunk_4_mul)[name = string("logits_chunk_4_max")];
+            int32 logits_chunk_4_argmax_axis_0 = const()[name = string("logits_chunk_4_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_4_argmax_keep_dims_0 = const()[name = string("logits_chunk_4_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_4_argmax_output_dtype_0 = const()[name = string("logits_chunk_4_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> logits_chunk_4_argmax = reduce_argmax(axis = logits_chunk_4_argmax_axis_0, keep_dims = logits_chunk_4_argmax_keep_dims_0, output_dtype = logits_chunk_4_argmax_output_dtype_0, x = logits_chunk_4_mul)[name = string("logits_chunk_4_argmax")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_4_sub = sub(x = logits_chunk_4_mul, y = logits_chunk_4_max)[name = string("logits_chunk_4_sub")];
+            tensor<int32, [1]> logits_chunk_4_lse_sub_axes_0 = const()[name = string("logits_chunk_4_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_4_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_4_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_4_lse_sub = reduce_log_sum_exp(axes = logits_chunk_4_lse_sub_axes_0, keep_dims = logits_chunk_4_lse_sub_keep_dims_0, x = logits_chunk_4_sub)[name = string("logits_chunk_4_lse_sub")];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_4_lse = add(x = logits_chunk_4_lse_sub, y = logits_chunk_4_max)[name = string("logits_chunk_4_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_5_weight_0 = const()[name = string("logits_chunk_5_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41947584)))];
+            tensor<int32, [2]> logits_chunk_5_strides_0 = const()[name = string("logits_chunk_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_5_pad_type_0 = const()[name = string("logits_chunk_5_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_5_pad_0 = const()[name = string("logits_chunk_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_5_dilations_0 = const()[name = string("logits_chunk_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_5_groups_0 = const()[name = string("logits_chunk_5_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_5 = conv(dilations = logits_chunk_5_dilations_0, groups = logits_chunk_5_groups_0, pad = logits_chunk_5_pad_0, pad_type = logits_chunk_5_pad_type_0, strides = logits_chunk_5_strides_0, weight = logits_chunk_5_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_5")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_5_mul = mul(x = logits_chunk_5, y = temp_inverse)[name = string("logits_chunk_5_mul")];
+            tensor<int32, [1]> logits_chunk_5_max_axes_0 = const()[name = string("logits_chunk_5_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_5_max_keep_dims_0 = const()[name = string("logits_chunk_5_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_5_max = reduce_max(axes = logits_chunk_5_max_axes_0, keep_dims = logits_chunk_5_max_keep_dims_0, x = logits_chunk_5_mul)[name = string("logits_chunk_5_max")];
+            int32 logits_chunk_5_argmax_axis_0 = const()[name = string("logits_chunk_5_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_5_argmax_keep_dims_0 = const()[name = string("logits_chunk_5_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_5_argmax_output_dtype_0 = const()[name = string("logits_chunk_5_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> logits_chunk_5_argmax = reduce_argmax(axis = logits_chunk_5_argmax_axis_0, keep_dims = logits_chunk_5_argmax_keep_dims_0, output_dtype = logits_chunk_5_argmax_output_dtype_0, x = logits_chunk_5_mul)[name = string("logits_chunk_5_argmax")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_5_sub = sub(x = logits_chunk_5_mul, y = logits_chunk_5_max)[name = string("logits_chunk_5_sub")];
+            tensor<int32, [1]> logits_chunk_5_lse_sub_axes_0 = const()[name = string("logits_chunk_5_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_5_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_5_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_5_lse_sub = reduce_log_sum_exp(axes = logits_chunk_5_lse_sub_axes_0, keep_dims = logits_chunk_5_lse_sub_keep_dims_0, x = logits_chunk_5_sub)[name = string("logits_chunk_5_lse_sub")];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_5_lse = add(x = logits_chunk_5_lse_sub, y = logits_chunk_5_max)[name = string("logits_chunk_5_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_6_weight_0 = const()[name = string("logits_chunk_6_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50336256)))];
+            tensor<int32, [2]> logits_chunk_6_strides_0 = const()[name = string("logits_chunk_6_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_6_pad_type_0 = const()[name = string("logits_chunk_6_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_6_pad_0 = const()[name = string("logits_chunk_6_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_6_dilations_0 = const()[name = string("logits_chunk_6_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_6_groups_0 = const()[name = string("logits_chunk_6_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_6 = conv(dilations = logits_chunk_6_dilations_0, groups = logits_chunk_6_groups_0, pad = logits_chunk_6_pad_0, pad_type = logits_chunk_6_pad_type_0, strides = logits_chunk_6_strides_0, weight = logits_chunk_6_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_6")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_6_mul = mul(x = logits_chunk_6, y = temp_inverse)[name = string("logits_chunk_6_mul")];
+            tensor<int32, [1]> logits_chunk_6_max_axes_0 = const()[name = string("logits_chunk_6_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_6_max_keep_dims_0 = const()[name = string("logits_chunk_6_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_6_max = reduce_max(axes = logits_chunk_6_max_axes_0, keep_dims = logits_chunk_6_max_keep_dims_0, x = logits_chunk_6_mul)[name = string("logits_chunk_6_max")];
+            int32 logits_chunk_6_argmax_axis_0 = const()[name = string("logits_chunk_6_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_6_argmax_keep_dims_0 = const()[name = string("logits_chunk_6_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_6_argmax_output_dtype_0 = const()[name = string("logits_chunk_6_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> logits_chunk_6_argmax = reduce_argmax(axis = logits_chunk_6_argmax_axis_0, keep_dims = logits_chunk_6_argmax_keep_dims_0, output_dtype = logits_chunk_6_argmax_output_dtype_0, x = logits_chunk_6_mul)[name = string("logits_chunk_6_argmax")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_6_sub = sub(x = logits_chunk_6_mul, y = logits_chunk_6_max)[name = string("logits_chunk_6_sub")];
+            tensor<int32, [1]> logits_chunk_6_lse_sub_axes_0 = const()[name = string("logits_chunk_6_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_6_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_6_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_6_lse_sub = reduce_log_sum_exp(axes = logits_chunk_6_lse_sub_axes_0, keep_dims = logits_chunk_6_lse_sub_keep_dims_0, x = logits_chunk_6_sub)[name = string("logits_chunk_6_lse_sub")];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_6_lse = add(x = logits_chunk_6_lse_sub, y = logits_chunk_6_max)[name = string("logits_chunk_6_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_7_weight_0 = const()[name = string("logits_chunk_7_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58724928)))];
+            tensor<int32, [2]> logits_chunk_7_strides_0 = const()[name = string("logits_chunk_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_7_pad_type_0 = const()[name = string("logits_chunk_7_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_7_pad_0 = const()[name = string("logits_chunk_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_7_dilations_0 = const()[name = string("logits_chunk_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_7_groups_0 = const()[name = string("logits_chunk_7_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_7 = conv(dilations = logits_chunk_7_dilations_0, groups = logits_chunk_7_groups_0, pad = logits_chunk_7_pad_0, pad_type = logits_chunk_7_pad_type_0, strides = logits_chunk_7_strides_0, weight = logits_chunk_7_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_7")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_7_mul = mul(x = logits_chunk_7, y = temp_inverse)[name = string("logits_chunk_7_mul")];
+            tensor<int32, [1]> logits_chunk_7_max_axes_0 = const()[name = string("logits_chunk_7_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_7_max_keep_dims_0 = const()[name = string("logits_chunk_7_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_7_max = reduce_max(axes = logits_chunk_7_max_axes_0, keep_dims = logits_chunk_7_max_keep_dims_0, x = logits_chunk_7_mul)[name = string("logits_chunk_7_max")];
+            int32 logits_chunk_7_argmax_axis_0 = const()[name = string("logits_chunk_7_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_7_argmax_keep_dims_0 = const()[name = string("logits_chunk_7_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_7_argmax_output_dtype_0 = const()[name = string("logits_chunk_7_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> logits_chunk_7_argmax = reduce_argmax(axis = logits_chunk_7_argmax_axis_0, keep_dims = logits_chunk_7_argmax_keep_dims_0, output_dtype = logits_chunk_7_argmax_output_dtype_0, x = logits_chunk_7_mul)[name = string("logits_chunk_7_argmax")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_7_sub = sub(x = logits_chunk_7_mul, y = logits_chunk_7_max)[name = string("logits_chunk_7_sub")];
+            tensor<int32, [1]> logits_chunk_7_lse_sub_axes_0 = const()[name = string("logits_chunk_7_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_7_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_7_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_7_lse_sub = reduce_log_sum_exp(axes = logits_chunk_7_lse_sub_axes_0, keep_dims = logits_chunk_7_lse_sub_keep_dims_0, x = logits_chunk_7_sub)[name = string("logits_chunk_7_lse_sub")];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_7_lse = add(x = logits_chunk_7_lse_sub, y = logits_chunk_7_max)[name = string("logits_chunk_7_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_8_weight_0 = const()[name = string("logits_chunk_8_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67113600)))];
+            tensor<int32, [2]> logits_chunk_8_strides_0 = const()[name = string("logits_chunk_8_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_8_pad_type_0 = const()[name = string("logits_chunk_8_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_8_pad_0 = const()[name = string("logits_chunk_8_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_8_dilations_0 = const()[name = string("logits_chunk_8_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_8_groups_0 = const()[name = string("logits_chunk_8_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_8 = conv(dilations = logits_chunk_8_dilations_0, groups = logits_chunk_8_groups_0, pad = logits_chunk_8_pad_0, pad_type = logits_chunk_8_pad_type_0, strides = logits_chunk_8_strides_0, weight = logits_chunk_8_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_8")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_8_mul = mul(x = logits_chunk_8, y = temp_inverse)[name = string("logits_chunk_8_mul")];
+            tensor<int32, [1]> logits_chunk_8_max_axes_0 = const()[name = string("logits_chunk_8_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_8_max_keep_dims_0 = const()[name = string("logits_chunk_8_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_8_max = reduce_max(axes = logits_chunk_8_max_axes_0, keep_dims = logits_chunk_8_max_keep_dims_0, x = logits_chunk_8_mul)[name = string("logits_chunk_8_max")];
+            int32 logits_chunk_8_argmax_axis_0 = const()[name = string("logits_chunk_8_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_8_argmax_keep_dims_0 = const()[name = string("logits_chunk_8_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_8_argmax_output_dtype_0 = const()[name = string("logits_chunk_8_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> logits_chunk_8_argmax = reduce_argmax(axis = logits_chunk_8_argmax_axis_0, keep_dims = logits_chunk_8_argmax_keep_dims_0, output_dtype = logits_chunk_8_argmax_output_dtype_0, x = logits_chunk_8_mul)[name = string("logits_chunk_8_argmax")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_8_sub = sub(x = logits_chunk_8_mul, y = logits_chunk_8_max)[name = string("logits_chunk_8_sub")];
+            tensor<int32, [1]> logits_chunk_8_lse_sub_axes_0 = const()[name = string("logits_chunk_8_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_8_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_8_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_8_lse_sub = reduce_log_sum_exp(axes = logits_chunk_8_lse_sub_axes_0, keep_dims = logits_chunk_8_lse_sub_keep_dims_0, x = logits_chunk_8_sub)[name = string("logits_chunk_8_lse_sub")];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_8_lse = add(x = logits_chunk_8_lse_sub, y = logits_chunk_8_max)[name = string("logits_chunk_8_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_9_weight_0 = const()[name = string("logits_chunk_9_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75502272)))];
+            tensor<int32, [2]> logits_chunk_9_strides_0 = const()[name = string("logits_chunk_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_9_pad_type_0 = const()[name = string("logits_chunk_9_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_9_pad_0 = const()[name = string("logits_chunk_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_9_dilations_0 = const()[name = string("logits_chunk_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_9_groups_0 = const()[name = string("logits_chunk_9_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_9 = conv(dilations = logits_chunk_9_dilations_0, groups = logits_chunk_9_groups_0, pad = logits_chunk_9_pad_0, pad_type = logits_chunk_9_pad_type_0, strides = logits_chunk_9_strides_0, weight = logits_chunk_9_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_9")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_9_mul = mul(x = logits_chunk_9, y = temp_inverse)[name = string("logits_chunk_9_mul")];
+            tensor<int32, [1]> logits_chunk_9_max_axes_0 = const()[name = string("logits_chunk_9_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_9_max_keep_dims_0 = const()[name = string("logits_chunk_9_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_9_max = reduce_max(axes = logits_chunk_9_max_axes_0, keep_dims = logits_chunk_9_max_keep_dims_0, x = logits_chunk_9_mul)[name = string("logits_chunk_9_max")];
+            int32 logits_chunk_9_argmax_axis_0 = const()[name = string("logits_chunk_9_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_9_argmax_keep_dims_0 = const()[name = string("logits_chunk_9_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_9_argmax_output_dtype_0 = const()[name = string("logits_chunk_9_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> logits_chunk_9_argmax = reduce_argmax(axis = logits_chunk_9_argmax_axis_0, keep_dims = logits_chunk_9_argmax_keep_dims_0, output_dtype = logits_chunk_9_argmax_output_dtype_0, x = logits_chunk_9_mul)[name = string("logits_chunk_9_argmax")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_9_sub = sub(x = logits_chunk_9_mul, y = logits_chunk_9_max)[name = string("logits_chunk_9_sub")];
+            tensor<int32, [1]> logits_chunk_9_lse_sub_axes_0 = const()[name = string("logits_chunk_9_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_9_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_9_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_9_lse_sub = reduce_log_sum_exp(axes = logits_chunk_9_lse_sub_axes_0, keep_dims = logits_chunk_9_lse_sub_keep_dims_0, x = logits_chunk_9_sub)[name = string("logits_chunk_9_lse_sub")];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_9_lse = add(x = logits_chunk_9_lse_sub, y = logits_chunk_9_max)[name = string("logits_chunk_9_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_10_weight_0 = const()[name = string("logits_chunk_10_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83890944)))];
+            tensor<int32, [2]> logits_chunk_10_strides_0 = const()[name = string("logits_chunk_10_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_10_pad_type_0 = const()[name = string("logits_chunk_10_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_10_pad_0 = const()[name = string("logits_chunk_10_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_10_dilations_0 = const()[name = string("logits_chunk_10_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_10_groups_0 = const()[name = string("logits_chunk_10_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_10 = conv(dilations = logits_chunk_10_dilations_0, groups = logits_chunk_10_groups_0, pad = logits_chunk_10_pad_0, pad_type = logits_chunk_10_pad_type_0, strides = logits_chunk_10_strides_0, weight = logits_chunk_10_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_10")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_10_mul = mul(x = logits_chunk_10, y = temp_inverse)[name = string("logits_chunk_10_mul")];
+            tensor<int32, [1]> logits_chunk_10_max_axes_0 = const()[name = string("logits_chunk_10_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_10_max_keep_dims_0 = const()[name = string("logits_chunk_10_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_10_max = reduce_max(axes = logits_chunk_10_max_axes_0, keep_dims = logits_chunk_10_max_keep_dims_0, x = logits_chunk_10_mul)[name = string("logits_chunk_10_max")];
+            int32 logits_chunk_10_argmax_axis_0 = const()[name = string("logits_chunk_10_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_10_argmax_keep_dims_0 = const()[name = string("logits_chunk_10_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_10_argmax_output_dtype_0 = const()[name = string("logits_chunk_10_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> logits_chunk_10_argmax = reduce_argmax(axis = logits_chunk_10_argmax_axis_0, keep_dims = logits_chunk_10_argmax_keep_dims_0, output_dtype = logits_chunk_10_argmax_output_dtype_0, x = logits_chunk_10_mul)[name = string("logits_chunk_10_argmax")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_10_sub = sub(x = logits_chunk_10_mul, y = logits_chunk_10_max)[name = string("logits_chunk_10_sub")];
+            tensor<int32, [1]> logits_chunk_10_lse_sub_axes_0 = const()[name = string("logits_chunk_10_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_10_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_10_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_10_lse_sub = reduce_log_sum_exp(axes = logits_chunk_10_lse_sub_axes_0, keep_dims = logits_chunk_10_lse_sub_keep_dims_0, x = logits_chunk_10_sub)[name = string("logits_chunk_10_lse_sub")];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_10_lse = add(x = logits_chunk_10_lse_sub, y = logits_chunk_10_max)[name = string("logits_chunk_10_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_11_weight_0 = const()[name = string("logits_chunk_11_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92279616)))];
+            tensor<int32, [2]> logits_chunk_11_strides_0 = const()[name = string("logits_chunk_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_11_pad_type_0 = const()[name = string("logits_chunk_11_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_11_pad_0 = const()[name = string("logits_chunk_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_11_dilations_0 = const()[name = string("logits_chunk_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_11_groups_0 = const()[name = string("logits_chunk_11_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_11 = conv(dilations = logits_chunk_11_dilations_0, groups = logits_chunk_11_groups_0, pad = logits_chunk_11_pad_0, pad_type = logits_chunk_11_pad_type_0, strides = logits_chunk_11_strides_0, weight = logits_chunk_11_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_11")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_11_mul = mul(x = logits_chunk_11, y = temp_inverse)[name = string("logits_chunk_11_mul")];
+            tensor<int32, [1]> logits_chunk_11_max_axes_0 = const()[name = string("logits_chunk_11_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_11_max_keep_dims_0 = const()[name = string("logits_chunk_11_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_11_max = reduce_max(axes = logits_chunk_11_max_axes_0, keep_dims = logits_chunk_11_max_keep_dims_0, x = logits_chunk_11_mul)[name = string("logits_chunk_11_max")];
+            int32 logits_chunk_11_argmax_axis_0 = const()[name = string("logits_chunk_11_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_11_argmax_keep_dims_0 = const()[name = string("logits_chunk_11_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_11_argmax_output_dtype_0 = const()[name = string("logits_chunk_11_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> logits_chunk_11_argmax = reduce_argmax(axis = logits_chunk_11_argmax_axis_0, keep_dims = logits_chunk_11_argmax_keep_dims_0, output_dtype = logits_chunk_11_argmax_output_dtype_0, x = logits_chunk_11_mul)[name = string("logits_chunk_11_argmax")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_11_sub = sub(x = logits_chunk_11_mul, y = logits_chunk_11_max)[name = string("logits_chunk_11_sub")];
+            tensor<int32, [1]> logits_chunk_11_lse_sub_axes_0 = const()[name = string("logits_chunk_11_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_11_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_11_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_11_lse_sub = reduce_log_sum_exp(axes = logits_chunk_11_lse_sub_axes_0, keep_dims = logits_chunk_11_lse_sub_keep_dims_0, x = logits_chunk_11_sub)[name = string("logits_chunk_11_lse_sub")];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_11_lse = add(x = logits_chunk_11_lse_sub, y = logits_chunk_11_max)[name = string("logits_chunk_11_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_12_weight_0 = const()[name = string("logits_chunk_12_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100668288)))];
+            tensor<int32, [2]> logits_chunk_12_strides_0 = const()[name = string("logits_chunk_12_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_12_pad_type_0 = const()[name = string("logits_chunk_12_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_12_pad_0 = const()[name = string("logits_chunk_12_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_12_dilations_0 = const()[name = string("logits_chunk_12_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_12_groups_0 = const()[name = string("logits_chunk_12_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_12 = conv(dilations = logits_chunk_12_dilations_0, groups = logits_chunk_12_groups_0, pad = logits_chunk_12_pad_0, pad_type = logits_chunk_12_pad_type_0, strides = logits_chunk_12_strides_0, weight = logits_chunk_12_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_12")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_12_mul = mul(x = logits_chunk_12, y = temp_inverse)[name = string("logits_chunk_12_mul")];
+            tensor<int32, [1]> logits_chunk_12_max_axes_0 = const()[name = string("logits_chunk_12_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_12_max_keep_dims_0 = const()[name = string("logits_chunk_12_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_12_max = reduce_max(axes = logits_chunk_12_max_axes_0, keep_dims = logits_chunk_12_max_keep_dims_0, x = logits_chunk_12_mul)[name = string("logits_chunk_12_max")];
+            int32 logits_chunk_12_argmax_axis_0 = const()[name = string("logits_chunk_12_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_12_argmax_keep_dims_0 = const()[name = string("logits_chunk_12_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_12_argmax_output_dtype_0 = const()[name = string("logits_chunk_12_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> logits_chunk_12_argmax = reduce_argmax(axis = logits_chunk_12_argmax_axis_0, keep_dims = logits_chunk_12_argmax_keep_dims_0, output_dtype = logits_chunk_12_argmax_output_dtype_0, x = logits_chunk_12_mul)[name = string("logits_chunk_12_argmax")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_12_sub = sub(x = logits_chunk_12_mul, y = logits_chunk_12_max)[name = string("logits_chunk_12_sub")];
+            tensor<int32, [1]> logits_chunk_12_lse_sub_axes_0 = const()[name = string("logits_chunk_12_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_12_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_12_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_12_lse_sub = reduce_log_sum_exp(axes = logits_chunk_12_lse_sub_axes_0, keep_dims = logits_chunk_12_lse_sub_keep_dims_0, x = logits_chunk_12_sub)[name = string("logits_chunk_12_lse_sub")];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_12_lse = add(x = logits_chunk_12_lse_sub, y = logits_chunk_12_max)[name = string("logits_chunk_12_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_13_weight_0 = const()[name = string("logits_chunk_13_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109056960)))];
+            tensor<int32, [2]> logits_chunk_13_strides_0 = const()[name = string("logits_chunk_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_13_pad_type_0 = const()[name = string("logits_chunk_13_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_13_pad_0 = const()[name = string("logits_chunk_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_13_dilations_0 = const()[name = string("logits_chunk_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_13_groups_0 = const()[name = string("logits_chunk_13_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_13 = conv(dilations = logits_chunk_13_dilations_0, groups = logits_chunk_13_groups_0, pad = logits_chunk_13_pad_0, pad_type = logits_chunk_13_pad_type_0, strides = logits_chunk_13_strides_0, weight = logits_chunk_13_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_13")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_13_mul = mul(x = logits_chunk_13, y = temp_inverse)[name = string("logits_chunk_13_mul")];
+            tensor<int32, [1]> logits_chunk_13_max_axes_0 = const()[name = string("logits_chunk_13_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_13_max_keep_dims_0 = const()[name = string("logits_chunk_13_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_13_max = reduce_max(axes = logits_chunk_13_max_axes_0, keep_dims = logits_chunk_13_max_keep_dims_0, x = logits_chunk_13_mul)[name = string("logits_chunk_13_max")];
+            int32 logits_chunk_13_argmax_axis_0 = const()[name = string("logits_chunk_13_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_13_argmax_keep_dims_0 = const()[name = string("logits_chunk_13_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_13_argmax_output_dtype_0 = const()[name = string("logits_chunk_13_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> logits_chunk_13_argmax = reduce_argmax(axis = logits_chunk_13_argmax_axis_0, keep_dims = logits_chunk_13_argmax_keep_dims_0, output_dtype = logits_chunk_13_argmax_output_dtype_0, x = logits_chunk_13_mul)[name = string("logits_chunk_13_argmax")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_13_sub = sub(x = logits_chunk_13_mul, y = logits_chunk_13_max)[name = string("logits_chunk_13_sub")];
+            tensor<int32, [1]> logits_chunk_13_lse_sub_axes_0 = const()[name = string("logits_chunk_13_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_13_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_13_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_13_lse_sub = reduce_log_sum_exp(axes = logits_chunk_13_lse_sub_axes_0, keep_dims = logits_chunk_13_lse_sub_keep_dims_0, x = logits_chunk_13_sub)[name = string("logits_chunk_13_lse_sub")];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_13_lse = add(x = logits_chunk_13_lse_sub, y = logits_chunk_13_max)[name = string("logits_chunk_13_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_14_weight_0 = const()[name = string("logits_chunk_14_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117445632)))];
+            tensor<int32, [2]> logits_chunk_14_strides_0 = const()[name = string("logits_chunk_14_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_14_pad_type_0 = const()[name = string("logits_chunk_14_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_14_pad_0 = const()[name = string("logits_chunk_14_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_14_dilations_0 = const()[name = string("logits_chunk_14_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_14_groups_0 = const()[name = string("logits_chunk_14_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_14 = conv(dilations = logits_chunk_14_dilations_0, groups = logits_chunk_14_groups_0, pad = logits_chunk_14_pad_0, pad_type = logits_chunk_14_pad_type_0, strides = logits_chunk_14_strides_0, weight = logits_chunk_14_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_14")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_14_mul = mul(x = logits_chunk_14, y = temp_inverse)[name = string("logits_chunk_14_mul")];
+            tensor<int32, [1]> logits_chunk_14_max_axes_0 = const()[name = string("logits_chunk_14_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_14_max_keep_dims_0 = const()[name = string("logits_chunk_14_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_14_max = reduce_max(axes = logits_chunk_14_max_axes_0, keep_dims = logits_chunk_14_max_keep_dims_0, x = logits_chunk_14_mul)[name = string("logits_chunk_14_max")];
+            int32 logits_chunk_14_argmax_axis_0 = const()[name = string("logits_chunk_14_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_14_argmax_keep_dims_0 = const()[name = string("logits_chunk_14_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_14_argmax_output_dtype_0 = const()[name = string("logits_chunk_14_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> logits_chunk_14_argmax = reduce_argmax(axis = logits_chunk_14_argmax_axis_0, keep_dims = logits_chunk_14_argmax_keep_dims_0, output_dtype = logits_chunk_14_argmax_output_dtype_0, x = logits_chunk_14_mul)[name = string("logits_chunk_14_argmax")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_14_sub = sub(x = logits_chunk_14_mul, y = logits_chunk_14_max)[name = string("logits_chunk_14_sub")];
+            tensor<int32, [1]> logits_chunk_14_lse_sub_axes_0 = const()[name = string("logits_chunk_14_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_14_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_14_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_14_lse_sub = reduce_log_sum_exp(axes = logits_chunk_14_lse_sub_axes_0, keep_dims = logits_chunk_14_lse_sub_keep_dims_0, x = logits_chunk_14_sub)[name = string("logits_chunk_14_lse_sub")];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_14_lse = add(x = logits_chunk_14_lse_sub, y = logits_chunk_14_max)[name = string("logits_chunk_14_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_15_weight_0 = const()[name = string("logits_chunk_15_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125834304)))];
+            tensor<int32, [2]> logits_chunk_15_strides_0 = const()[name = string("logits_chunk_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_15_pad_type_0 = const()[name = string("logits_chunk_15_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_15_pad_0 = const()[name = string("logits_chunk_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_15_dilations_0 = const()[name = string("logits_chunk_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_15_groups_0 = const()[name = string("logits_chunk_15_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_15 = conv(dilations = logits_chunk_15_dilations_0, groups = logits_chunk_15_groups_0, pad = logits_chunk_15_pad_0, pad_type = logits_chunk_15_pad_type_0, strides = logits_chunk_15_strides_0, weight = logits_chunk_15_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_15")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_15_mul = mul(x = logits_chunk_15, y = temp_inverse)[name = string("logits_chunk_15_mul")];
+            tensor<int32, [1]> logits_chunk_15_max_axes_0 = const()[name = string("logits_chunk_15_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_15_max_keep_dims_0 = const()[name = string("logits_chunk_15_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_15_max = reduce_max(axes = logits_chunk_15_max_axes_0, keep_dims = logits_chunk_15_max_keep_dims_0, x = logits_chunk_15_mul)[name = string("logits_chunk_15_max")];
+            int32 logits_chunk_15_argmax_axis_0 = const()[name = string("logits_chunk_15_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_15_argmax_keep_dims_0 = const()[name = string("logits_chunk_15_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_15_argmax_output_dtype_0 = const()[name = string("logits_chunk_15_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 32]> logits_chunk_15_argmax = reduce_argmax(axis = logits_chunk_15_argmax_axis_0, keep_dims = logits_chunk_15_argmax_keep_dims_0, output_dtype = logits_chunk_15_argmax_output_dtype_0, x = logits_chunk_15_mul)[name = string("logits_chunk_15_argmax")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_15_sub = sub(x = logits_chunk_15_mul, y = logits_chunk_15_max)[name = string("logits_chunk_15_sub")];
+            tensor<int32, [1]> logits_chunk_15_lse_sub_axes_0 = const()[name = string("logits_chunk_15_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_15_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_15_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_15_lse_sub = reduce_log_sum_exp(axes = logits_chunk_15_lse_sub_axes_0, keep_dims = logits_chunk_15_lse_sub_keep_dims_0, x = logits_chunk_15_sub)[name = string("logits_chunk_15_lse_sub")];
+            tensor<fp16, [1, 1, 1, 32]> logits_chunk_15_lse = add(x = logits_chunk_15_lse_sub, y = logits_chunk_15_max)[name = string("logits_chunk_15_lse")];
+            int32 logits_lses_axis_0 = const()[name = string("logits_lses_axis_0"), val = int32(1)];
+            bool logits_lses_interleave_0 = const()[name = string("logits_lses_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 32]> logits_lses = concat(axis = logits_lses_axis_0, interleave = logits_lses_interleave_0, values = (logits_chunk_0_lse, logits_chunk_1_lse, logits_chunk_2_lse, logits_chunk_3_lse, logits_chunk_4_lse, logits_chunk_5_lse, logits_chunk_6_lse, logits_chunk_7_lse, logits_chunk_8_lse, logits_chunk_9_lse, logits_chunk_10_lse, logits_chunk_11_lse, logits_chunk_12_lse, logits_chunk_13_lse, logits_chunk_14_lse, logits_chunk_15_lse))[name = string("logits_lses")];
+            tensor<int32, [1]> logits_lses_max_axes_0 = const()[name = string("logits_lses_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_lses_max_keep_dims_0 = const()[name = string("logits_lses_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_lses_max = reduce_max(axes = logits_lses_max_axes_0, keep_dims = logits_lses_max_keep_dims_0, x = logits_lses)[name = string("logits_lses_max")];
+            tensor<fp16, [1, 16, 1, 32]> logits_lses_sub = sub(x = logits_lses, y = logits_lses_max)[name = string("logits_lses_sub")];
+            tensor<int32, [1]> logits_lses_logsumexp_axes_0 = const()[name = string("logits_lses_logsumexp_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_lses_logsumexp_keep_dims_0 = const()[name = string("logits_lses_logsumexp_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_lses_logsumexp = reduce_log_sum_exp(axes = logits_lses_logsumexp_axes_0, keep_dims = logits_lses_logsumexp_keep_dims_0, x = logits_lses_sub)[name = string("logits_lses_logsumexp")];
+            tensor<fp16, [1, 1, 1, 32]> logits_lse = add(x = logits_lses_logsumexp, y = logits_lses_max)[name = string("logits_lse")];
+            int32 logits_max_logits_chunks_axis_0 = const()[name = string("logits_max_logits_chunks_axis_0"), val = int32(1)];
+            bool logits_max_logits_chunks_interleave_0 = const()[name = string("logits_max_logits_chunks_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 32]> logits_max_logits_chunks = concat(axis = logits_max_logits_chunks_axis_0, interleave = logits_max_logits_chunks_interleave_0, values = (logits_chunk_0_max, logits_chunk_1_max, logits_chunk_2_max, logits_chunk_3_max, logits_chunk_4_max, logits_chunk_5_max, logits_chunk_6_max, logits_chunk_7_max, logits_chunk_8_max, logits_chunk_9_max, logits_chunk_10_max, logits_chunk_11_max, logits_chunk_12_max, logits_chunk_13_max, logits_chunk_14_max, logits_chunk_15_max))[name = string("logits_max_logits_chunks")];
+            tensor<int32, [1]> logits_max_logit_axes_0 = const()[name = string("logits_max_logit_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_max_logit_keep_dims_0 = const()[name = string("logits_max_logit_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 32]> logits_max_logit = reduce_max(axes = logits_max_logit_axes_0, keep_dims = logits_max_logit_keep_dims_0, x = logits_max_logits_chunks)[name = string("logits_max_logit")];
+            tensor<fp16, [1, 1, 1, 32]> logits_max_logit_sub = sub(x = logits_max_logit, y = logits_lse)[name = string("logits_max_logit_sub")];
+            tensor<fp16, [1, 1, 1, 32]> max_prob = exp(x = logits_max_logit_sub)[name = string("max_prob")];
+            tensor<fp16, [1, 1, 1, 32]> min_p_thresh = mul(x = max_prob, y = p)[name = string("min_p_thresh")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_0_sub_1 = sub(x = logits_chunk_0_mul, y = logits_lse)[name = string("logits_chunk_0_sub_1")];
+            tensor<fp16, [1, 2048, 1, 32]> probs_chunk_0 = exp(x = logits_chunk_0_sub_1)[name = string("probs_chunk_0")];
+            tensor<bool, [1, 2048, 1, 32]> mask_probs_chunk_0 = greater_equal(x = probs_chunk_0, y = min_p_thresh)[name = string("mask_probs_chunk_0")];
+            string mask_chunk_0_fp16_dtype_0 = const()[name = string("mask_chunk_0_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 32]> mask_chunk_0_fp16 = cast(dtype = mask_chunk_0_fp16_dtype_0, x = mask_probs_chunk_0)[name = string("cast_169")];
+            tensor<fp16, [1, 2048, 1, 32]> masked_probs_chunk_0 = select(a = probs_chunk_0, b = mask_chunk_0_fp16, cond = mask_probs_chunk_0)[name = string("masked_probs_chunk_0")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_1_sub_1 = sub(x = logits_chunk_1_mul, y = logits_lse)[name = string("logits_chunk_1_sub_1")];
+            tensor<fp16, [1, 2048, 1, 32]> probs_chunk_1 = exp(x = logits_chunk_1_sub_1)[name = string("probs_chunk_1")];
+            tensor<bool, [1, 2048, 1, 32]> mask_probs_chunk_1 = greater_equal(x = probs_chunk_1, y = min_p_thresh)[name = string("mask_probs_chunk_1")];
+            string mask_chunk_1_fp16_dtype_0 = const()[name = string("mask_chunk_1_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 32]> mask_chunk_1_fp16 = cast(dtype = mask_chunk_1_fp16_dtype_0, x = mask_probs_chunk_1)[name = string("cast_168")];
+            tensor<fp16, [1, 2048, 1, 32]> masked_probs_chunk_1 = select(a = probs_chunk_1, b = mask_chunk_1_fp16, cond = mask_probs_chunk_1)[name = string("masked_probs_chunk_1")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_2_sub_1 = sub(x = logits_chunk_2_mul, y = logits_lse)[name = string("logits_chunk_2_sub_1")];
+            tensor<fp16, [1, 2048, 1, 32]> probs_chunk_2 = exp(x = logits_chunk_2_sub_1)[name = string("probs_chunk_2")];
+            tensor<bool, [1, 2048, 1, 32]> mask_probs_chunk_2 = greater_equal(x = probs_chunk_2, y = min_p_thresh)[name = string("mask_probs_chunk_2")];
+            string mask_chunk_2_fp16_dtype_0 = const()[name = string("mask_chunk_2_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 32]> mask_chunk_2_fp16 = cast(dtype = mask_chunk_2_fp16_dtype_0, x = mask_probs_chunk_2)[name = string("cast_167")];
+            tensor<fp16, [1, 2048, 1, 32]> masked_probs_chunk_2 = select(a = probs_chunk_2, b = mask_chunk_2_fp16, cond = mask_probs_chunk_2)[name = string("masked_probs_chunk_2")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_3_sub_1 = sub(x = logits_chunk_3_mul, y = logits_lse)[name = string("logits_chunk_3_sub_1")];
+            tensor<fp16, [1, 2048, 1, 32]> probs_chunk_3 = exp(x = logits_chunk_3_sub_1)[name = string("probs_chunk_3")];
+            tensor<bool, [1, 2048, 1, 32]> mask_probs_chunk_3 = greater_equal(x = probs_chunk_3, y = min_p_thresh)[name = string("mask_probs_chunk_3")];
+            string mask_chunk_3_fp16_dtype_0 = const()[name = string("mask_chunk_3_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 32]> mask_chunk_3_fp16 = cast(dtype = mask_chunk_3_fp16_dtype_0, x = mask_probs_chunk_3)[name = string("cast_166")];
+            tensor<fp16, [1, 2048, 1, 32]> masked_probs_chunk_3 = select(a = probs_chunk_3, b = mask_chunk_3_fp16, cond = mask_probs_chunk_3)[name = string("masked_probs_chunk_3")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_4_sub_1 = sub(x = logits_chunk_4_mul, y = logits_lse)[name = string("logits_chunk_4_sub_1")];
+            tensor<fp16, [1, 2048, 1, 32]> probs_chunk_4 = exp(x = logits_chunk_4_sub_1)[name = string("probs_chunk_4")];
+            tensor<bool, [1, 2048, 1, 32]> mask_probs_chunk_4 = greater_equal(x = probs_chunk_4, y = min_p_thresh)[name = string("mask_probs_chunk_4")];
+            string mask_chunk_4_fp16_dtype_0 = const()[name = string("mask_chunk_4_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 32]> mask_chunk_4_fp16 = cast(dtype = mask_chunk_4_fp16_dtype_0, x = mask_probs_chunk_4)[name = string("cast_165")];
+            tensor<fp16, [1, 2048, 1, 32]> masked_probs_chunk_4 = select(a = probs_chunk_4, b = mask_chunk_4_fp16, cond = mask_probs_chunk_4)[name = string("masked_probs_chunk_4")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_5_sub_1 = sub(x = logits_chunk_5_mul, y = logits_lse)[name = string("logits_chunk_5_sub_1")];
+            tensor<fp16, [1, 2048, 1, 32]> probs_chunk_5 = exp(x = logits_chunk_5_sub_1)[name = string("probs_chunk_5")];
+            tensor<bool, [1, 2048, 1, 32]> mask_probs_chunk_5 = greater_equal(x = probs_chunk_5, y = min_p_thresh)[name = string("mask_probs_chunk_5")];
+            string mask_chunk_5_fp16_dtype_0 = const()[name = string("mask_chunk_5_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 32]> mask_chunk_5_fp16 = cast(dtype = mask_chunk_5_fp16_dtype_0, x = mask_probs_chunk_5)[name = string("cast_164")];
+            tensor<fp16, [1, 2048, 1, 32]> masked_probs_chunk_5 = select(a = probs_chunk_5, b = mask_chunk_5_fp16, cond = mask_probs_chunk_5)[name = string("masked_probs_chunk_5")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_6_sub_1 = sub(x = logits_chunk_6_mul, y = logits_lse)[name = string("logits_chunk_6_sub_1")];
+            tensor<fp16, [1, 2048, 1, 32]> probs_chunk_6 = exp(x = logits_chunk_6_sub_1)[name = string("probs_chunk_6")];
+            tensor<bool, [1, 2048, 1, 32]> mask_probs_chunk_6 = greater_equal(x = probs_chunk_6, y = min_p_thresh)[name = string("mask_probs_chunk_6")];
+            string mask_chunk_6_fp16_dtype_0 = const()[name = string("mask_chunk_6_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 32]> mask_chunk_6_fp16 = cast(dtype = mask_chunk_6_fp16_dtype_0, x = mask_probs_chunk_6)[name = string("cast_163")];
+            tensor<fp16, [1, 2048, 1, 32]> masked_probs_chunk_6 = select(a = probs_chunk_6, b = mask_chunk_6_fp16, cond = mask_probs_chunk_6)[name = string("masked_probs_chunk_6")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_7_sub_1 = sub(x = logits_chunk_7_mul, y = logits_lse)[name = string("logits_chunk_7_sub_1")];
+            tensor<fp16, [1, 2048, 1, 32]> probs_chunk_7 = exp(x = logits_chunk_7_sub_1)[name = string("probs_chunk_7")];
+            tensor<bool, [1, 2048, 1, 32]> mask_probs_chunk_7 = greater_equal(x = probs_chunk_7, y = min_p_thresh)[name = string("mask_probs_chunk_7")];
+            string mask_chunk_7_fp16_dtype_0 = const()[name = string("mask_chunk_7_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 32]> mask_chunk_7_fp16 = cast(dtype = mask_chunk_7_fp16_dtype_0, x = mask_probs_chunk_7)[name = string("cast_162")];
+            tensor<fp16, [1, 2048, 1, 32]> masked_probs_chunk_7 = select(a = probs_chunk_7, b = mask_chunk_7_fp16, cond = mask_probs_chunk_7)[name = string("masked_probs_chunk_7")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_8_sub_1 = sub(x = logits_chunk_8_mul, y = logits_lse)[name = string("logits_chunk_8_sub_1")];
+            tensor<fp16, [1, 2048, 1, 32]> probs_chunk_8 = exp(x = logits_chunk_8_sub_1)[name = string("probs_chunk_8")];
+            tensor<bool, [1, 2048, 1, 32]> mask_probs_chunk_8 = greater_equal(x = probs_chunk_8, y = min_p_thresh)[name = string("mask_probs_chunk_8")];
+            string mask_chunk_8_fp16_dtype_0 = const()[name = string("mask_chunk_8_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 32]> mask_chunk_8_fp16 = cast(dtype = mask_chunk_8_fp16_dtype_0, x = mask_probs_chunk_8)[name = string("cast_161")];
+            tensor<fp16, [1, 2048, 1, 32]> masked_probs_chunk_8 = select(a = probs_chunk_8, b = mask_chunk_8_fp16, cond = mask_probs_chunk_8)[name = string("masked_probs_chunk_8")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_9_sub_1 = sub(x = logits_chunk_9_mul, y = logits_lse)[name = string("logits_chunk_9_sub_1")];
+            tensor<fp16, [1, 2048, 1, 32]> probs_chunk_9 = exp(x = logits_chunk_9_sub_1)[name = string("probs_chunk_9")];
+            tensor<bool, [1, 2048, 1, 32]> mask_probs_chunk_9 = greater_equal(x = probs_chunk_9, y = min_p_thresh)[name = string("mask_probs_chunk_9")];
+            string mask_chunk_9_fp16_dtype_0 = const()[name = string("mask_chunk_9_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 32]> mask_chunk_9_fp16 = cast(dtype = mask_chunk_9_fp16_dtype_0, x = mask_probs_chunk_9)[name = string("cast_160")];
+            tensor<fp16, [1, 2048, 1, 32]> masked_probs_chunk_9 = select(a = probs_chunk_9, b = mask_chunk_9_fp16, cond = mask_probs_chunk_9)[name = string("masked_probs_chunk_9")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_10_sub_1 = sub(x = logits_chunk_10_mul, y = logits_lse)[name = string("logits_chunk_10_sub_1")];
+            tensor<fp16, [1, 2048, 1, 32]> probs_chunk_10 = exp(x = logits_chunk_10_sub_1)[name = string("probs_chunk_10")];
+            tensor<bool, [1, 2048, 1, 32]> mask_probs_chunk_10 = greater_equal(x = probs_chunk_10, y = min_p_thresh)[name = string("mask_probs_chunk_10")];
+            string mask_chunk_10_fp16_dtype_0 = const()[name = string("mask_chunk_10_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 32]> mask_chunk_10_fp16 = cast(dtype = mask_chunk_10_fp16_dtype_0, x = mask_probs_chunk_10)[name = string("cast_159")];
+            tensor<fp16, [1, 2048, 1, 32]> masked_probs_chunk_10 = select(a = probs_chunk_10, b = mask_chunk_10_fp16, cond = mask_probs_chunk_10)[name = string("masked_probs_chunk_10")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_11_sub_1 = sub(x = logits_chunk_11_mul, y = logits_lse)[name = string("logits_chunk_11_sub_1")];
+            tensor<fp16, [1, 2048, 1, 32]> probs_chunk_11 = exp(x = logits_chunk_11_sub_1)[name = string("probs_chunk_11")];
+            tensor<bool, [1, 2048, 1, 32]> mask_probs_chunk_11 = greater_equal(x = probs_chunk_11, y = min_p_thresh)[name = string("mask_probs_chunk_11")];
+            string mask_chunk_11_fp16_dtype_0 = const()[name = string("mask_chunk_11_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 32]> mask_chunk_11_fp16 = cast(dtype = mask_chunk_11_fp16_dtype_0, x = mask_probs_chunk_11)[name = string("cast_158")];
+            tensor<fp16, [1, 2048, 1, 32]> masked_probs_chunk_11 = select(a = probs_chunk_11, b = mask_chunk_11_fp16, cond = mask_probs_chunk_11)[name = string("masked_probs_chunk_11")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_12_sub_1 = sub(x = logits_chunk_12_mul, y = logits_lse)[name = string("logits_chunk_12_sub_1")];
+            tensor<fp16, [1, 2048, 1, 32]> probs_chunk_12 = exp(x = logits_chunk_12_sub_1)[name = string("probs_chunk_12")];
+            tensor<bool, [1, 2048, 1, 32]> mask_probs_chunk_12 = greater_equal(x = probs_chunk_12, y = min_p_thresh)[name = string("mask_probs_chunk_12")];
+            string mask_chunk_12_fp16_dtype_0 = const()[name = string("mask_chunk_12_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 32]> mask_chunk_12_fp16 = cast(dtype = mask_chunk_12_fp16_dtype_0, x = mask_probs_chunk_12)[name = string("cast_157")];
+            tensor<fp16, [1, 2048, 1, 32]> masked_probs_chunk_12 = select(a = probs_chunk_12, b = mask_chunk_12_fp16, cond = mask_probs_chunk_12)[name = string("masked_probs_chunk_12")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_13_sub_1 = sub(x = logits_chunk_13_mul, y = logits_lse)[name = string("logits_chunk_13_sub_1")];
+            tensor<fp16, [1, 2048, 1, 32]> probs_chunk_13 = exp(x = logits_chunk_13_sub_1)[name = string("probs_chunk_13")];
+            tensor<bool, [1, 2048, 1, 32]> mask_probs_chunk_13 = greater_equal(x = probs_chunk_13, y = min_p_thresh)[name = string("mask_probs_chunk_13")];
+            string mask_chunk_13_fp16_dtype_0 = const()[name = string("mask_chunk_13_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 32]> mask_chunk_13_fp16 = cast(dtype = mask_chunk_13_fp16_dtype_0, x = mask_probs_chunk_13)[name = string("cast_156")];
+            tensor<fp16, [1, 2048, 1, 32]> masked_probs_chunk_13 = select(a = probs_chunk_13, b = mask_chunk_13_fp16, cond = mask_probs_chunk_13)[name = string("masked_probs_chunk_13")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_14_sub_1 = sub(x = logits_chunk_14_mul, y = logits_lse)[name = string("logits_chunk_14_sub_1")];
+            tensor<fp16, [1, 2048, 1, 32]> probs_chunk_14 = exp(x = logits_chunk_14_sub_1)[name = string("probs_chunk_14")];
+            tensor<bool, [1, 2048, 1, 32]> mask_probs_chunk_14 = greater_equal(x = probs_chunk_14, y = min_p_thresh)[name = string("mask_probs_chunk_14")];
+            string mask_chunk_14_fp16_dtype_0 = const()[name = string("mask_chunk_14_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 32]> mask_chunk_14_fp16 = cast(dtype = mask_chunk_14_fp16_dtype_0, x = mask_probs_chunk_14)[name = string("cast_155")];
+            tensor<fp16, [1, 2048, 1, 32]> masked_probs_chunk_14 = select(a = probs_chunk_14, b = mask_chunk_14_fp16, cond = mask_probs_chunk_14)[name = string("masked_probs_chunk_14")];
+            tensor<fp16, [1, 2048, 1, 32]> logits_chunk_15_sub_1 = sub(x = logits_chunk_15_mul, y = logits_lse)[name = string("logits_chunk_15_sub_1")];
+            tensor<fp16, [1, 2048, 1, 32]> probs_chunk_15 = exp(x = logits_chunk_15_sub_1)[name = string("probs_chunk_15")];
+            tensor<bool, [1, 2048, 1, 32]> mask_probs_chunk_15 = greater_equal(x = probs_chunk_15, y = min_p_thresh)[name = string("mask_probs_chunk_15")];
+            string mask_chunk_15_fp16_dtype_0 = const()[name = string("mask_chunk_15_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 32]> mask_chunk_15_fp16 = cast(dtype = mask_chunk_15_fp16_dtype_0, x = mask_probs_chunk_15)[name = string("cast_154")];
+            tensor<fp16, [1, 2048, 1, 32]> masked_probs_chunk_15 = select(a = probs_chunk_15, b = mask_chunk_15_fp16, cond = mask_probs_chunk_15)[name = string("masked_probs_chunk_15")];
+            int32 probs_axis_0 = const()[name = string("probs_axis_0"), val = int32(1)];
+            bool probs_interleave_0 = const()[name = string("probs_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 32768, 1, 32]> probs = concat(axis = probs_axis_0, interleave = probs_interleave_0, values = (masked_probs_chunk_0, masked_probs_chunk_1, masked_probs_chunk_2, masked_probs_chunk_3, masked_probs_chunk_4, masked_probs_chunk_5, masked_probs_chunk_6, masked_probs_chunk_7, masked_probs_chunk_8, masked_probs_chunk_9, masked_probs_chunk_10, masked_probs_chunk_11, masked_probs_chunk_12, masked_probs_chunk_13, masked_probs_chunk_14, masked_probs_chunk_15))[name = string("probs")];
+            string probs_fp32_dtype_0 = const()[name = string("probs_fp32_dtype_0"), val = string("fp32")];
+            int32 probs_cumsum_axis_0 = const()[name = string("probs_cumsum_axis_0"), val = int32(1)];
+            bool probs_cumsum_exclusive_0 = const()[name = string("probs_cumsum_exclusive_0"), val = bool(false)];
+            bool probs_cumsum_reverse_0 = const()[name = string("probs_cumsum_reverse_0"), val = bool(false)];
+            tensor<fp32, [1, 32768, 1, 32]> probs_fp32 = cast(dtype = probs_fp32_dtype_0, x = probs)[name = string("cast_153")];
+            tensor<fp32, [1, 32768, 1, 32]> probs_cumsum = cumsum(axis = probs_cumsum_axis_0, exclusive = probs_cumsum_exclusive_0, reverse = probs_cumsum_reverse_0, x = probs_fp32)[name = string("probs_cumsum")];
+            tensor<int32, [1]> probs_sum_indices_0 = const()[name = string("probs_sum_indices_0"), val = tensor<int32, [1]>([32767])];
+            int32 probs_sum_axis_0 = const()[name = string("probs_sum_axis_0"), val = int32(1)];
+            int32 probs_sum_batch_dims_0 = const()[name = string("probs_sum_batch_dims_0"), val = int32(0)];
+            bool probs_sum_validate_indices_0 = const()[name = string("probs_sum_validate_indices_0"), val = bool(false)];
+            tensor<fp32, [1, 1, 1, 32]> probs_sum = gather(axis = probs_sum_axis_0, batch_dims = probs_sum_batch_dims_0, indices = probs_sum_indices_0, validate_indices = probs_sum_validate_indices_0, x = probs_cumsum)[name = string("probs_sum")];
+            tensor<fp32, [1, 1, 1, 32]> random_number_scaled = mul(x = random_number, y = probs_sum)[name = string("random_number_scaled")];
+            tensor<bool, [1, 32768, 1, 32]> probs_greater = greater(x = probs_cumsum, y = random_number_scaled)[name = string("probs_greater")];
+            string probs_greater_int32_dtype_0 = const()[name = string("probs_greater_int32_dtype_0"), val = string("int32")];
+            int32 sampled_index_axis_0 = const()[name = string("sampled_index_axis_0"), val = int32(1)];
+            bool sampled_index_keep_dims_0 = const()[name = string("sampled_index_keep_dims_0"), val = bool(true)];
+            string sampled_index_output_dtype_0 = const()[name = string("sampled_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 32768, 1, 32]> probs_greater_int32 = cast(dtype = probs_greater_int32_dtype_0, x = probs_greater)[name = string("cast_152")];
+            tensor<int32, [1, 1, 1, 32]> sampled_index = reduce_argmax(axis = sampled_index_axis_0, keep_dims = sampled_index_keep_dims_0, output_dtype = sampled_index_output_dtype_0, x = probs_greater_int32)[name = string("sampled_index")];
+            int32 sampled_index_probability_axis_0 = const()[name = string("sampled_index_probability_axis_0"), val = int32(1)];
+            bool sampled_index_probability_validate_indices_0 = const()[name = string("sampled_index_probability_validate_indices_0"), val = bool(false)];
+            tensor<fp32, [1, 1, 1, 32]> sampled_index_probability = gather_along_axis(axis = sampled_index_probability_axis_0, indices = sampled_index, validate_indices = sampled_index_probability_validate_indices_0, x = probs_fp32)[name = string("sampled_index_probability")];
+            int32 max_logit_index_axis_0 = const()[name = string("max_logit_index_axis_0"), val = int32(1)];
+            bool max_logit_index_keep_dims_0 = const()[name = string("max_logit_index_keep_dims_0"), val = bool(true)];
+            string max_logit_index_output_dtype_0 = const()[name = string("max_logit_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 1, 1, 32]> max_logit_index = reduce_argmax(axis = max_logit_index_axis_0, keep_dims = max_logit_index_keep_dims_0, output_dtype = max_logit_index_output_dtype_0, x = logits_max_logits_chunks)[name = string("max_logit_index")];
+            string indices_chunk_0_int32_dtype_0 = const()[name = string("indices_chunk_0_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_1_int32_dtype_0 = const()[name = string("indices_chunk_1_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_2_int32_dtype_0 = const()[name = string("indices_chunk_2_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_3_int32_dtype_0 = const()[name = string("indices_chunk_3_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_4_int32_dtype_0 = const()[name = string("indices_chunk_4_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_5_int32_dtype_0 = const()[name = string("indices_chunk_5_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_6_int32_dtype_0 = const()[name = string("indices_chunk_6_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_7_int32_dtype_0 = const()[name = string("indices_chunk_7_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_8_int32_dtype_0 = const()[name = string("indices_chunk_8_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_9_int32_dtype_0 = const()[name = string("indices_chunk_9_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_10_int32_dtype_0 = const()[name = string("indices_chunk_10_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_11_int32_dtype_0 = const()[name = string("indices_chunk_11_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_12_int32_dtype_0 = const()[name = string("indices_chunk_12_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_13_int32_dtype_0 = const()[name = string("indices_chunk_13_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_14_int32_dtype_0 = const()[name = string("indices_chunk_14_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_15_int32_dtype_0 = const()[name = string("indices_chunk_15_int32_dtype_0"), val = string("int32")];
+            int32 indices_axis_0 = const()[name = string("indices_axis_0"), val = int32(1)];
+            bool indices_interleave_0 = const()[name = string("indices_interleave_0"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_15_int32 = cast(dtype = indices_chunk_15_int32_dtype_0, x = logits_chunk_15_argmax)[name = string("cast_136")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_14_int32 = cast(dtype = indices_chunk_14_int32_dtype_0, x = logits_chunk_14_argmax)[name = string("cast_137")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_13_int32 = cast(dtype = indices_chunk_13_int32_dtype_0, x = logits_chunk_13_argmax)[name = string("cast_138")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_12_int32 = cast(dtype = indices_chunk_12_int32_dtype_0, x = logits_chunk_12_argmax)[name = string("cast_139")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_11_int32 = cast(dtype = indices_chunk_11_int32_dtype_0, x = logits_chunk_11_argmax)[name = string("cast_140")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_10_int32 = cast(dtype = indices_chunk_10_int32_dtype_0, x = logits_chunk_10_argmax)[name = string("cast_141")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_9_int32 = cast(dtype = indices_chunk_9_int32_dtype_0, x = logits_chunk_9_argmax)[name = string("cast_142")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_8_int32 = cast(dtype = indices_chunk_8_int32_dtype_0, x = logits_chunk_8_argmax)[name = string("cast_143")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_7_int32 = cast(dtype = indices_chunk_7_int32_dtype_0, x = logits_chunk_7_argmax)[name = string("cast_144")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_6_int32 = cast(dtype = indices_chunk_6_int32_dtype_0, x = logits_chunk_6_argmax)[name = string("cast_145")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_5_int32 = cast(dtype = indices_chunk_5_int32_dtype_0, x = logits_chunk_5_argmax)[name = string("cast_146")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_4_int32 = cast(dtype = indices_chunk_4_int32_dtype_0, x = logits_chunk_4_argmax)[name = string("cast_147")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_3_int32 = cast(dtype = indices_chunk_3_int32_dtype_0, x = logits_chunk_3_argmax)[name = string("cast_148")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_2_int32 = cast(dtype = indices_chunk_2_int32_dtype_0, x = logits_chunk_2_argmax)[name = string("cast_149")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_1_int32 = cast(dtype = indices_chunk_1_int32_dtype_0, x = logits_chunk_1_argmax)[name = string("cast_150")];
+            tensor<int32, [1, 1, 1, 32]> indices_chunk_0_int32 = cast(dtype = indices_chunk_0_int32_dtype_0, x = logits_chunk_0_argmax)[name = string("cast_151")];
+            tensor<int32, [1, 16, 1, 32]> indices = concat(axis = indices_axis_0, interleave = indices_interleave_0, values = (indices_chunk_0_int32, indices_chunk_1_int32, indices_chunk_2_int32, indices_chunk_3_int32, indices_chunk_4_int32, indices_chunk_5_int32, indices_chunk_6_int32, indices_chunk_7_int32, indices_chunk_8_int32, indices_chunk_9_int32, indices_chunk_10_int32, indices_chunk_11_int32, indices_chunk_12_int32, indices_chunk_13_int32, indices_chunk_14_int32, indices_chunk_15_int32))[name = string("indices")];
+            int32 argmax_chunks_axis_0 = const()[name = string("argmax_chunks_axis_0"), val = int32(1)];
+            bool argmax_chunks_validate_indices_0 = const()[name = string("argmax_chunks_validate_indices_0"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 32]> argmax_chunks = gather_along_axis(axis = argmax_chunks_axis_0, indices = max_logit_index, validate_indices = argmax_chunks_validate_indices_0, x = indices)[name = string("argmax_chunks")];
+            int32 mul_0_x_0 = const()[name = string("mul_0_x_0"), val = int32(2048)];
+            tensor<int32, [1, 1, 1, 32]> mul_0 = mul(x = mul_0_x_0, y = max_logit_index)[name = string("mul_0")];
+            tensor<int32, [1, 1, 1, 32]> argmax = add(x = argmax_chunks, y = mul_0)[name = string("argmax")];
+        } -> (sampled_index, sampled_index_probability, argmax, max_prob);
+    func min_p_length_48<ios18>(tensor<fp16, [1, 2048, 1, 48]> hidden_states, tensor<fp16, [1]> p, tensor<fp32, [48]> random_number, tensor<fp16, [1]> temp) {
+            tensor<fp16, [1, 2048, 1, 48]> final_norm_rmsnorm_abs = abs(x = hidden_states)[name = string("final_norm_rmsnorm_abs")];
+            tensor<int32, [1]> final_norm_rmsnorm_maxval_axes_0 = const()[name = string("final_norm_rmsnorm_maxval_axes_0"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_maxval_keep_dims_0 = const()[name = string("final_norm_rmsnorm_maxval_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> final_norm_rmsnorm_maxval = reduce_max(axes = final_norm_rmsnorm_maxval_axes_0, keep_dims = final_norm_rmsnorm_maxval_keep_dims_0, x = final_norm_rmsnorm_abs)[name = string("final_norm_rmsnorm_maxval")];
+            fp16 final_norm_rmsnorm_maxval_clipped_alpha_0 = const()[name = string("final_norm_rmsnorm_maxval_clipped_alpha_0"), val = fp16(0x1p-14)];
+            fp16 final_norm_rmsnorm_maxval_clipped_beta_0 = const()[name = string("final_norm_rmsnorm_maxval_clipped_beta_0"), val = fp16(inf)];
+            tensor<fp16, [1, 1, 1, 48]> final_norm_rmsnorm_maxval_clipped = clip(alpha = final_norm_rmsnorm_maxval_clipped_alpha_0, beta = final_norm_rmsnorm_maxval_clipped_beta_0, x = final_norm_rmsnorm_maxval)[name = string("final_norm_rmsnorm_maxval_clipped")];
+            tensor<fp16, [1, 2048, 1, 48]> final_norm_rmsnorm_scaled = real_div(x = hidden_states, y = final_norm_rmsnorm_maxval_clipped)[name = string("final_norm_rmsnorm_scaled")];
+            tensor<int32, [1]> final_norm_rmsnorm_squared_sum_axes_0 = const()[name = string("final_norm_rmsnorm_squared_sum_axes_0"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_squared_sum_keep_dims_0 = const()[name = string("final_norm_rmsnorm_squared_sum_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> final_norm_rmsnorm_squared_sum = reduce_sum_square(axes = final_norm_rmsnorm_squared_sum_axes_0, keep_dims = final_norm_rmsnorm_squared_sum_keep_dims_0, x = final_norm_rmsnorm_scaled)[name = string("final_norm_rmsnorm_squared_sum")];
+            fp16 final_norm_rmsnorm_rsqrt_epsilon_0 = const()[name = string("final_norm_rmsnorm_rsqrt_epsilon_0"), val = fp16(0x1p-14)];
+            tensor<fp16, [1, 1, 1, 48]> final_norm_rmsnorm_rsqrt = rsqrt(epsilon = final_norm_rmsnorm_rsqrt_epsilon_0, x = final_norm_rmsnorm_squared_sum)[name = string("final_norm_rmsnorm_rsqrt")];
+            fp16 final_norm_rmsnorm_dim_scaled_y_0 = const()[name = string("final_norm_rmsnorm_dim_scaled_y_0"), val = fp16(0x1.6ap+5)];
+            tensor<fp16, [1, 2048, 1, 48]> final_norm_rmsnorm_dim_scaled = mul(x = final_norm_rmsnorm_scaled, y = final_norm_rmsnorm_dim_scaled_y_0)[name = string("final_norm_rmsnorm_dim_scaled")];
+            tensor<fp16, [1, 2048, 1, 48]> final_norm_rmsnorm_normalized = mul(x = final_norm_rmsnorm_dim_scaled, y = final_norm_rmsnorm_rsqrt)[name = string("final_norm_rmsnorm_normalized")];
+            tensor<fp16, [2048, 1, 1]> final_norm_rmsnorm_y_0 = const()[name = string("final_norm_rmsnorm_y_0"), val = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 2048, 1, 48]> final_norm_rmsnorm = mul(x = final_norm_rmsnorm_normalized, y = final_norm_rmsnorm_y_0)[name = string("final_norm_rmsnorm")];
+            fp16 temp_inverse_epsilon_0 = const()[name = string("temp_inverse_epsilon_0"), val = fp16(0x0p+0)];
+            tensor<fp16, [1]> temp_inverse = inverse(epsilon = temp_inverse_epsilon_0, x = temp)[name = string("temp_inverse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_0_weight_0 = const()[name = string("logits_chunk_0_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4224)))];
+            tensor<int32, [2]> logits_chunk_0_strides_0 = const()[name = string("logits_chunk_0_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_0_pad_type_0 = const()[name = string("logits_chunk_0_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_0_pad_0 = const()[name = string("logits_chunk_0_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_0_dilations_0 = const()[name = string("logits_chunk_0_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_0_groups_0 = const()[name = string("logits_chunk_0_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_0 = conv(dilations = logits_chunk_0_dilations_0, groups = logits_chunk_0_groups_0, pad = logits_chunk_0_pad_0, pad_type = logits_chunk_0_pad_type_0, strides = logits_chunk_0_strides_0, weight = logits_chunk_0_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_0")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_0_mul = mul(x = logits_chunk_0, y = temp_inverse)[name = string("logits_chunk_0_mul")];
+            tensor<int32, [1]> logits_chunk_0_max_axes_0 = const()[name = string("logits_chunk_0_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_0_max_keep_dims_0 = const()[name = string("logits_chunk_0_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_0_max = reduce_max(axes = logits_chunk_0_max_axes_0, keep_dims = logits_chunk_0_max_keep_dims_0, x = logits_chunk_0_mul)[name = string("logits_chunk_0_max")];
+            int32 logits_chunk_0_argmax_axis_0 = const()[name = string("logits_chunk_0_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_0_argmax_keep_dims_0 = const()[name = string("logits_chunk_0_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_0_argmax_output_dtype_0 = const()[name = string("logits_chunk_0_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> logits_chunk_0_argmax = reduce_argmax(axis = logits_chunk_0_argmax_axis_0, keep_dims = logits_chunk_0_argmax_keep_dims_0, output_dtype = logits_chunk_0_argmax_output_dtype_0, x = logits_chunk_0_mul)[name = string("logits_chunk_0_argmax")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_0_sub = sub(x = logits_chunk_0_mul, y = logits_chunk_0_max)[name = string("logits_chunk_0_sub")];
+            tensor<int32, [1]> logits_chunk_0_lse_sub_axes_0 = const()[name = string("logits_chunk_0_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_0_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_0_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_0_lse_sub = reduce_log_sum_exp(axes = logits_chunk_0_lse_sub_axes_0, keep_dims = logits_chunk_0_lse_sub_keep_dims_0, x = logits_chunk_0_sub)[name = string("logits_chunk_0_lse_sub")];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_0_lse = add(x = logits_chunk_0_lse_sub, y = logits_chunk_0_max)[name = string("logits_chunk_0_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_1_weight_0 = const()[name = string("logits_chunk_1_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8392896)))];
+            tensor<int32, [2]> logits_chunk_1_strides_0 = const()[name = string("logits_chunk_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_1_pad_type_0 = const()[name = string("logits_chunk_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_1_pad_0 = const()[name = string("logits_chunk_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_1_dilations_0 = const()[name = string("logits_chunk_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_1_groups_0 = const()[name = string("logits_chunk_1_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_1 = conv(dilations = logits_chunk_1_dilations_0, groups = logits_chunk_1_groups_0, pad = logits_chunk_1_pad_0, pad_type = logits_chunk_1_pad_type_0, strides = logits_chunk_1_strides_0, weight = logits_chunk_1_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_1")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_1_mul = mul(x = logits_chunk_1, y = temp_inverse)[name = string("logits_chunk_1_mul")];
+            tensor<int32, [1]> logits_chunk_1_max_axes_0 = const()[name = string("logits_chunk_1_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_1_max_keep_dims_0 = const()[name = string("logits_chunk_1_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_1_max = reduce_max(axes = logits_chunk_1_max_axes_0, keep_dims = logits_chunk_1_max_keep_dims_0, x = logits_chunk_1_mul)[name = string("logits_chunk_1_max")];
+            int32 logits_chunk_1_argmax_axis_0 = const()[name = string("logits_chunk_1_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_1_argmax_keep_dims_0 = const()[name = string("logits_chunk_1_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_1_argmax_output_dtype_0 = const()[name = string("logits_chunk_1_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> logits_chunk_1_argmax = reduce_argmax(axis = logits_chunk_1_argmax_axis_0, keep_dims = logits_chunk_1_argmax_keep_dims_0, output_dtype = logits_chunk_1_argmax_output_dtype_0, x = logits_chunk_1_mul)[name = string("logits_chunk_1_argmax")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_1_sub = sub(x = logits_chunk_1_mul, y = logits_chunk_1_max)[name = string("logits_chunk_1_sub")];
+            tensor<int32, [1]> logits_chunk_1_lse_sub_axes_0 = const()[name = string("logits_chunk_1_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_1_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_1_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_1_lse_sub = reduce_log_sum_exp(axes = logits_chunk_1_lse_sub_axes_0, keep_dims = logits_chunk_1_lse_sub_keep_dims_0, x = logits_chunk_1_sub)[name = string("logits_chunk_1_lse_sub")];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_1_lse = add(x = logits_chunk_1_lse_sub, y = logits_chunk_1_max)[name = string("logits_chunk_1_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_2_weight_0 = const()[name = string("logits_chunk_2_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16781568)))];
+            tensor<int32, [2]> logits_chunk_2_strides_0 = const()[name = string("logits_chunk_2_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_2_pad_type_0 = const()[name = string("logits_chunk_2_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_2_pad_0 = const()[name = string("logits_chunk_2_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_2_dilations_0 = const()[name = string("logits_chunk_2_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_2_groups_0 = const()[name = string("logits_chunk_2_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_2 = conv(dilations = logits_chunk_2_dilations_0, groups = logits_chunk_2_groups_0, pad = logits_chunk_2_pad_0, pad_type = logits_chunk_2_pad_type_0, strides = logits_chunk_2_strides_0, weight = logits_chunk_2_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_2")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_2_mul = mul(x = logits_chunk_2, y = temp_inverse)[name = string("logits_chunk_2_mul")];
+            tensor<int32, [1]> logits_chunk_2_max_axes_0 = const()[name = string("logits_chunk_2_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_2_max_keep_dims_0 = const()[name = string("logits_chunk_2_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_2_max = reduce_max(axes = logits_chunk_2_max_axes_0, keep_dims = logits_chunk_2_max_keep_dims_0, x = logits_chunk_2_mul)[name = string("logits_chunk_2_max")];
+            int32 logits_chunk_2_argmax_axis_0 = const()[name = string("logits_chunk_2_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_2_argmax_keep_dims_0 = const()[name = string("logits_chunk_2_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_2_argmax_output_dtype_0 = const()[name = string("logits_chunk_2_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> logits_chunk_2_argmax = reduce_argmax(axis = logits_chunk_2_argmax_axis_0, keep_dims = logits_chunk_2_argmax_keep_dims_0, output_dtype = logits_chunk_2_argmax_output_dtype_0, x = logits_chunk_2_mul)[name = string("logits_chunk_2_argmax")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_2_sub = sub(x = logits_chunk_2_mul, y = logits_chunk_2_max)[name = string("logits_chunk_2_sub")];
+            tensor<int32, [1]> logits_chunk_2_lse_sub_axes_0 = const()[name = string("logits_chunk_2_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_2_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_2_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_2_lse_sub = reduce_log_sum_exp(axes = logits_chunk_2_lse_sub_axes_0, keep_dims = logits_chunk_2_lse_sub_keep_dims_0, x = logits_chunk_2_sub)[name = string("logits_chunk_2_lse_sub")];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_2_lse = add(x = logits_chunk_2_lse_sub, y = logits_chunk_2_max)[name = string("logits_chunk_2_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_3_weight_0 = const()[name = string("logits_chunk_3_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25170240)))];
+            tensor<int32, [2]> logits_chunk_3_strides_0 = const()[name = string("logits_chunk_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_3_pad_type_0 = const()[name = string("logits_chunk_3_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_3_pad_0 = const()[name = string("logits_chunk_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_3_dilations_0 = const()[name = string("logits_chunk_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_3_groups_0 = const()[name = string("logits_chunk_3_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_3 = conv(dilations = logits_chunk_3_dilations_0, groups = logits_chunk_3_groups_0, pad = logits_chunk_3_pad_0, pad_type = logits_chunk_3_pad_type_0, strides = logits_chunk_3_strides_0, weight = logits_chunk_3_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_3")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_3_mul = mul(x = logits_chunk_3, y = temp_inverse)[name = string("logits_chunk_3_mul")];
+            tensor<int32, [1]> logits_chunk_3_max_axes_0 = const()[name = string("logits_chunk_3_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_3_max_keep_dims_0 = const()[name = string("logits_chunk_3_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_3_max = reduce_max(axes = logits_chunk_3_max_axes_0, keep_dims = logits_chunk_3_max_keep_dims_0, x = logits_chunk_3_mul)[name = string("logits_chunk_3_max")];
+            int32 logits_chunk_3_argmax_axis_0 = const()[name = string("logits_chunk_3_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_3_argmax_keep_dims_0 = const()[name = string("logits_chunk_3_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_3_argmax_output_dtype_0 = const()[name = string("logits_chunk_3_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> logits_chunk_3_argmax = reduce_argmax(axis = logits_chunk_3_argmax_axis_0, keep_dims = logits_chunk_3_argmax_keep_dims_0, output_dtype = logits_chunk_3_argmax_output_dtype_0, x = logits_chunk_3_mul)[name = string("logits_chunk_3_argmax")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_3_sub = sub(x = logits_chunk_3_mul, y = logits_chunk_3_max)[name = string("logits_chunk_3_sub")];
+            tensor<int32, [1]> logits_chunk_3_lse_sub_axes_0 = const()[name = string("logits_chunk_3_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_3_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_3_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_3_lse_sub = reduce_log_sum_exp(axes = logits_chunk_3_lse_sub_axes_0, keep_dims = logits_chunk_3_lse_sub_keep_dims_0, x = logits_chunk_3_sub)[name = string("logits_chunk_3_lse_sub")];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_3_lse = add(x = logits_chunk_3_lse_sub, y = logits_chunk_3_max)[name = string("logits_chunk_3_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_4_weight_0 = const()[name = string("logits_chunk_4_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33558912)))];
+            tensor<int32, [2]> logits_chunk_4_strides_0 = const()[name = string("logits_chunk_4_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_4_pad_type_0 = const()[name = string("logits_chunk_4_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_4_pad_0 = const()[name = string("logits_chunk_4_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_4_dilations_0 = const()[name = string("logits_chunk_4_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_4_groups_0 = const()[name = string("logits_chunk_4_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_4 = conv(dilations = logits_chunk_4_dilations_0, groups = logits_chunk_4_groups_0, pad = logits_chunk_4_pad_0, pad_type = logits_chunk_4_pad_type_0, strides = logits_chunk_4_strides_0, weight = logits_chunk_4_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_4")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_4_mul = mul(x = logits_chunk_4, y = temp_inverse)[name = string("logits_chunk_4_mul")];
+            tensor<int32, [1]> logits_chunk_4_max_axes_0 = const()[name = string("logits_chunk_4_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_4_max_keep_dims_0 = const()[name = string("logits_chunk_4_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_4_max = reduce_max(axes = logits_chunk_4_max_axes_0, keep_dims = logits_chunk_4_max_keep_dims_0, x = logits_chunk_4_mul)[name = string("logits_chunk_4_max")];
+            int32 logits_chunk_4_argmax_axis_0 = const()[name = string("logits_chunk_4_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_4_argmax_keep_dims_0 = const()[name = string("logits_chunk_4_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_4_argmax_output_dtype_0 = const()[name = string("logits_chunk_4_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> logits_chunk_4_argmax = reduce_argmax(axis = logits_chunk_4_argmax_axis_0, keep_dims = logits_chunk_4_argmax_keep_dims_0, output_dtype = logits_chunk_4_argmax_output_dtype_0, x = logits_chunk_4_mul)[name = string("logits_chunk_4_argmax")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_4_sub = sub(x = logits_chunk_4_mul, y = logits_chunk_4_max)[name = string("logits_chunk_4_sub")];
+            tensor<int32, [1]> logits_chunk_4_lse_sub_axes_0 = const()[name = string("logits_chunk_4_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_4_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_4_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_4_lse_sub = reduce_log_sum_exp(axes = logits_chunk_4_lse_sub_axes_0, keep_dims = logits_chunk_4_lse_sub_keep_dims_0, x = logits_chunk_4_sub)[name = string("logits_chunk_4_lse_sub")];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_4_lse = add(x = logits_chunk_4_lse_sub, y = logits_chunk_4_max)[name = string("logits_chunk_4_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_5_weight_0 = const()[name = string("logits_chunk_5_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41947584)))];
+            tensor<int32, [2]> logits_chunk_5_strides_0 = const()[name = string("logits_chunk_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_5_pad_type_0 = const()[name = string("logits_chunk_5_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_5_pad_0 = const()[name = string("logits_chunk_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_5_dilations_0 = const()[name = string("logits_chunk_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_5_groups_0 = const()[name = string("logits_chunk_5_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_5 = conv(dilations = logits_chunk_5_dilations_0, groups = logits_chunk_5_groups_0, pad = logits_chunk_5_pad_0, pad_type = logits_chunk_5_pad_type_0, strides = logits_chunk_5_strides_0, weight = logits_chunk_5_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_5")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_5_mul = mul(x = logits_chunk_5, y = temp_inverse)[name = string("logits_chunk_5_mul")];
+            tensor<int32, [1]> logits_chunk_5_max_axes_0 = const()[name = string("logits_chunk_5_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_5_max_keep_dims_0 = const()[name = string("logits_chunk_5_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_5_max = reduce_max(axes = logits_chunk_5_max_axes_0, keep_dims = logits_chunk_5_max_keep_dims_0, x = logits_chunk_5_mul)[name = string("logits_chunk_5_max")];
+            int32 logits_chunk_5_argmax_axis_0 = const()[name = string("logits_chunk_5_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_5_argmax_keep_dims_0 = const()[name = string("logits_chunk_5_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_5_argmax_output_dtype_0 = const()[name = string("logits_chunk_5_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> logits_chunk_5_argmax = reduce_argmax(axis = logits_chunk_5_argmax_axis_0, keep_dims = logits_chunk_5_argmax_keep_dims_0, output_dtype = logits_chunk_5_argmax_output_dtype_0, x = logits_chunk_5_mul)[name = string("logits_chunk_5_argmax")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_5_sub = sub(x = logits_chunk_5_mul, y = logits_chunk_5_max)[name = string("logits_chunk_5_sub")];
+            tensor<int32, [1]> logits_chunk_5_lse_sub_axes_0 = const()[name = string("logits_chunk_5_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_5_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_5_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_5_lse_sub = reduce_log_sum_exp(axes = logits_chunk_5_lse_sub_axes_0, keep_dims = logits_chunk_5_lse_sub_keep_dims_0, x = logits_chunk_5_sub)[name = string("logits_chunk_5_lse_sub")];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_5_lse = add(x = logits_chunk_5_lse_sub, y = logits_chunk_5_max)[name = string("logits_chunk_5_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_6_weight_0 = const()[name = string("logits_chunk_6_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50336256)))];
+            tensor<int32, [2]> logits_chunk_6_strides_0 = const()[name = string("logits_chunk_6_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_6_pad_type_0 = const()[name = string("logits_chunk_6_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_6_pad_0 = const()[name = string("logits_chunk_6_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_6_dilations_0 = const()[name = string("logits_chunk_6_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_6_groups_0 = const()[name = string("logits_chunk_6_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_6 = conv(dilations = logits_chunk_6_dilations_0, groups = logits_chunk_6_groups_0, pad = logits_chunk_6_pad_0, pad_type = logits_chunk_6_pad_type_0, strides = logits_chunk_6_strides_0, weight = logits_chunk_6_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_6")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_6_mul = mul(x = logits_chunk_6, y = temp_inverse)[name = string("logits_chunk_6_mul")];
+            tensor<int32, [1]> logits_chunk_6_max_axes_0 = const()[name = string("logits_chunk_6_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_6_max_keep_dims_0 = const()[name = string("logits_chunk_6_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_6_max = reduce_max(axes = logits_chunk_6_max_axes_0, keep_dims = logits_chunk_6_max_keep_dims_0, x = logits_chunk_6_mul)[name = string("logits_chunk_6_max")];
+            int32 logits_chunk_6_argmax_axis_0 = const()[name = string("logits_chunk_6_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_6_argmax_keep_dims_0 = const()[name = string("logits_chunk_6_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_6_argmax_output_dtype_0 = const()[name = string("logits_chunk_6_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> logits_chunk_6_argmax = reduce_argmax(axis = logits_chunk_6_argmax_axis_0, keep_dims = logits_chunk_6_argmax_keep_dims_0, output_dtype = logits_chunk_6_argmax_output_dtype_0, x = logits_chunk_6_mul)[name = string("logits_chunk_6_argmax")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_6_sub = sub(x = logits_chunk_6_mul, y = logits_chunk_6_max)[name = string("logits_chunk_6_sub")];
+            tensor<int32, [1]> logits_chunk_6_lse_sub_axes_0 = const()[name = string("logits_chunk_6_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_6_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_6_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_6_lse_sub = reduce_log_sum_exp(axes = logits_chunk_6_lse_sub_axes_0, keep_dims = logits_chunk_6_lse_sub_keep_dims_0, x = logits_chunk_6_sub)[name = string("logits_chunk_6_lse_sub")];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_6_lse = add(x = logits_chunk_6_lse_sub, y = logits_chunk_6_max)[name = string("logits_chunk_6_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_7_weight_0 = const()[name = string("logits_chunk_7_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58724928)))];
+            tensor<int32, [2]> logits_chunk_7_strides_0 = const()[name = string("logits_chunk_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_7_pad_type_0 = const()[name = string("logits_chunk_7_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_7_pad_0 = const()[name = string("logits_chunk_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_7_dilations_0 = const()[name = string("logits_chunk_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_7_groups_0 = const()[name = string("logits_chunk_7_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_7 = conv(dilations = logits_chunk_7_dilations_0, groups = logits_chunk_7_groups_0, pad = logits_chunk_7_pad_0, pad_type = logits_chunk_7_pad_type_0, strides = logits_chunk_7_strides_0, weight = logits_chunk_7_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_7")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_7_mul = mul(x = logits_chunk_7, y = temp_inverse)[name = string("logits_chunk_7_mul")];
+            tensor<int32, [1]> logits_chunk_7_max_axes_0 = const()[name = string("logits_chunk_7_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_7_max_keep_dims_0 = const()[name = string("logits_chunk_7_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_7_max = reduce_max(axes = logits_chunk_7_max_axes_0, keep_dims = logits_chunk_7_max_keep_dims_0, x = logits_chunk_7_mul)[name = string("logits_chunk_7_max")];
+            int32 logits_chunk_7_argmax_axis_0 = const()[name = string("logits_chunk_7_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_7_argmax_keep_dims_0 = const()[name = string("logits_chunk_7_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_7_argmax_output_dtype_0 = const()[name = string("logits_chunk_7_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> logits_chunk_7_argmax = reduce_argmax(axis = logits_chunk_7_argmax_axis_0, keep_dims = logits_chunk_7_argmax_keep_dims_0, output_dtype = logits_chunk_7_argmax_output_dtype_0, x = logits_chunk_7_mul)[name = string("logits_chunk_7_argmax")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_7_sub = sub(x = logits_chunk_7_mul, y = logits_chunk_7_max)[name = string("logits_chunk_7_sub")];
+            tensor<int32, [1]> logits_chunk_7_lse_sub_axes_0 = const()[name = string("logits_chunk_7_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_7_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_7_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_7_lse_sub = reduce_log_sum_exp(axes = logits_chunk_7_lse_sub_axes_0, keep_dims = logits_chunk_7_lse_sub_keep_dims_0, x = logits_chunk_7_sub)[name = string("logits_chunk_7_lse_sub")];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_7_lse = add(x = logits_chunk_7_lse_sub, y = logits_chunk_7_max)[name = string("logits_chunk_7_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_8_weight_0 = const()[name = string("logits_chunk_8_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67113600)))];
+            tensor<int32, [2]> logits_chunk_8_strides_0 = const()[name = string("logits_chunk_8_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_8_pad_type_0 = const()[name = string("logits_chunk_8_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_8_pad_0 = const()[name = string("logits_chunk_8_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_8_dilations_0 = const()[name = string("logits_chunk_8_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_8_groups_0 = const()[name = string("logits_chunk_8_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_8 = conv(dilations = logits_chunk_8_dilations_0, groups = logits_chunk_8_groups_0, pad = logits_chunk_8_pad_0, pad_type = logits_chunk_8_pad_type_0, strides = logits_chunk_8_strides_0, weight = logits_chunk_8_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_8")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_8_mul = mul(x = logits_chunk_8, y = temp_inverse)[name = string("logits_chunk_8_mul")];
+            tensor<int32, [1]> logits_chunk_8_max_axes_0 = const()[name = string("logits_chunk_8_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_8_max_keep_dims_0 = const()[name = string("logits_chunk_8_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_8_max = reduce_max(axes = logits_chunk_8_max_axes_0, keep_dims = logits_chunk_8_max_keep_dims_0, x = logits_chunk_8_mul)[name = string("logits_chunk_8_max")];
+            int32 logits_chunk_8_argmax_axis_0 = const()[name = string("logits_chunk_8_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_8_argmax_keep_dims_0 = const()[name = string("logits_chunk_8_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_8_argmax_output_dtype_0 = const()[name = string("logits_chunk_8_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> logits_chunk_8_argmax = reduce_argmax(axis = logits_chunk_8_argmax_axis_0, keep_dims = logits_chunk_8_argmax_keep_dims_0, output_dtype = logits_chunk_8_argmax_output_dtype_0, x = logits_chunk_8_mul)[name = string("logits_chunk_8_argmax")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_8_sub = sub(x = logits_chunk_8_mul, y = logits_chunk_8_max)[name = string("logits_chunk_8_sub")];
+            tensor<int32, [1]> logits_chunk_8_lse_sub_axes_0 = const()[name = string("logits_chunk_8_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_8_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_8_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_8_lse_sub = reduce_log_sum_exp(axes = logits_chunk_8_lse_sub_axes_0, keep_dims = logits_chunk_8_lse_sub_keep_dims_0, x = logits_chunk_8_sub)[name = string("logits_chunk_8_lse_sub")];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_8_lse = add(x = logits_chunk_8_lse_sub, y = logits_chunk_8_max)[name = string("logits_chunk_8_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_9_weight_0 = const()[name = string("logits_chunk_9_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75502272)))];
+            tensor<int32, [2]> logits_chunk_9_strides_0 = const()[name = string("logits_chunk_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_9_pad_type_0 = const()[name = string("logits_chunk_9_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_9_pad_0 = const()[name = string("logits_chunk_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_9_dilations_0 = const()[name = string("logits_chunk_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_9_groups_0 = const()[name = string("logits_chunk_9_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_9 = conv(dilations = logits_chunk_9_dilations_0, groups = logits_chunk_9_groups_0, pad = logits_chunk_9_pad_0, pad_type = logits_chunk_9_pad_type_0, strides = logits_chunk_9_strides_0, weight = logits_chunk_9_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_9")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_9_mul = mul(x = logits_chunk_9, y = temp_inverse)[name = string("logits_chunk_9_mul")];
+            tensor<int32, [1]> logits_chunk_9_max_axes_0 = const()[name = string("logits_chunk_9_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_9_max_keep_dims_0 = const()[name = string("logits_chunk_9_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_9_max = reduce_max(axes = logits_chunk_9_max_axes_0, keep_dims = logits_chunk_9_max_keep_dims_0, x = logits_chunk_9_mul)[name = string("logits_chunk_9_max")];
+            int32 logits_chunk_9_argmax_axis_0 = const()[name = string("logits_chunk_9_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_9_argmax_keep_dims_0 = const()[name = string("logits_chunk_9_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_9_argmax_output_dtype_0 = const()[name = string("logits_chunk_9_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> logits_chunk_9_argmax = reduce_argmax(axis = logits_chunk_9_argmax_axis_0, keep_dims = logits_chunk_9_argmax_keep_dims_0, output_dtype = logits_chunk_9_argmax_output_dtype_0, x = logits_chunk_9_mul)[name = string("logits_chunk_9_argmax")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_9_sub = sub(x = logits_chunk_9_mul, y = logits_chunk_9_max)[name = string("logits_chunk_9_sub")];
+            tensor<int32, [1]> logits_chunk_9_lse_sub_axes_0 = const()[name = string("logits_chunk_9_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_9_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_9_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_9_lse_sub = reduce_log_sum_exp(axes = logits_chunk_9_lse_sub_axes_0, keep_dims = logits_chunk_9_lse_sub_keep_dims_0, x = logits_chunk_9_sub)[name = string("logits_chunk_9_lse_sub")];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_9_lse = add(x = logits_chunk_9_lse_sub, y = logits_chunk_9_max)[name = string("logits_chunk_9_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_10_weight_0 = const()[name = string("logits_chunk_10_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83890944)))];
+            tensor<int32, [2]> logits_chunk_10_strides_0 = const()[name = string("logits_chunk_10_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_10_pad_type_0 = const()[name = string("logits_chunk_10_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_10_pad_0 = const()[name = string("logits_chunk_10_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_10_dilations_0 = const()[name = string("logits_chunk_10_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_10_groups_0 = const()[name = string("logits_chunk_10_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_10 = conv(dilations = logits_chunk_10_dilations_0, groups = logits_chunk_10_groups_0, pad = logits_chunk_10_pad_0, pad_type = logits_chunk_10_pad_type_0, strides = logits_chunk_10_strides_0, weight = logits_chunk_10_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_10")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_10_mul = mul(x = logits_chunk_10, y = temp_inverse)[name = string("logits_chunk_10_mul")];
+            tensor<int32, [1]> logits_chunk_10_max_axes_0 = const()[name = string("logits_chunk_10_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_10_max_keep_dims_0 = const()[name = string("logits_chunk_10_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_10_max = reduce_max(axes = logits_chunk_10_max_axes_0, keep_dims = logits_chunk_10_max_keep_dims_0, x = logits_chunk_10_mul)[name = string("logits_chunk_10_max")];
+            int32 logits_chunk_10_argmax_axis_0 = const()[name = string("logits_chunk_10_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_10_argmax_keep_dims_0 = const()[name = string("logits_chunk_10_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_10_argmax_output_dtype_0 = const()[name = string("logits_chunk_10_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> logits_chunk_10_argmax = reduce_argmax(axis = logits_chunk_10_argmax_axis_0, keep_dims = logits_chunk_10_argmax_keep_dims_0, output_dtype = logits_chunk_10_argmax_output_dtype_0, x = logits_chunk_10_mul)[name = string("logits_chunk_10_argmax")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_10_sub = sub(x = logits_chunk_10_mul, y = logits_chunk_10_max)[name = string("logits_chunk_10_sub")];
+            tensor<int32, [1]> logits_chunk_10_lse_sub_axes_0 = const()[name = string("logits_chunk_10_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_10_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_10_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_10_lse_sub = reduce_log_sum_exp(axes = logits_chunk_10_lse_sub_axes_0, keep_dims = logits_chunk_10_lse_sub_keep_dims_0, x = logits_chunk_10_sub)[name = string("logits_chunk_10_lse_sub")];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_10_lse = add(x = logits_chunk_10_lse_sub, y = logits_chunk_10_max)[name = string("logits_chunk_10_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_11_weight_0 = const()[name = string("logits_chunk_11_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92279616)))];
+            tensor<int32, [2]> logits_chunk_11_strides_0 = const()[name = string("logits_chunk_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_11_pad_type_0 = const()[name = string("logits_chunk_11_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_11_pad_0 = const()[name = string("logits_chunk_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_11_dilations_0 = const()[name = string("logits_chunk_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_11_groups_0 = const()[name = string("logits_chunk_11_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_11 = conv(dilations = logits_chunk_11_dilations_0, groups = logits_chunk_11_groups_0, pad = logits_chunk_11_pad_0, pad_type = logits_chunk_11_pad_type_0, strides = logits_chunk_11_strides_0, weight = logits_chunk_11_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_11")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_11_mul = mul(x = logits_chunk_11, y = temp_inverse)[name = string("logits_chunk_11_mul")];
+            tensor<int32, [1]> logits_chunk_11_max_axes_0 = const()[name = string("logits_chunk_11_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_11_max_keep_dims_0 = const()[name = string("logits_chunk_11_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_11_max = reduce_max(axes = logits_chunk_11_max_axes_0, keep_dims = logits_chunk_11_max_keep_dims_0, x = logits_chunk_11_mul)[name = string("logits_chunk_11_max")];
+            int32 logits_chunk_11_argmax_axis_0 = const()[name = string("logits_chunk_11_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_11_argmax_keep_dims_0 = const()[name = string("logits_chunk_11_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_11_argmax_output_dtype_0 = const()[name = string("logits_chunk_11_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> logits_chunk_11_argmax = reduce_argmax(axis = logits_chunk_11_argmax_axis_0, keep_dims = logits_chunk_11_argmax_keep_dims_0, output_dtype = logits_chunk_11_argmax_output_dtype_0, x = logits_chunk_11_mul)[name = string("logits_chunk_11_argmax")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_11_sub = sub(x = logits_chunk_11_mul, y = logits_chunk_11_max)[name = string("logits_chunk_11_sub")];
+            tensor<int32, [1]> logits_chunk_11_lse_sub_axes_0 = const()[name = string("logits_chunk_11_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_11_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_11_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_11_lse_sub = reduce_log_sum_exp(axes = logits_chunk_11_lse_sub_axes_0, keep_dims = logits_chunk_11_lse_sub_keep_dims_0, x = logits_chunk_11_sub)[name = string("logits_chunk_11_lse_sub")];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_11_lse = add(x = logits_chunk_11_lse_sub, y = logits_chunk_11_max)[name = string("logits_chunk_11_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_12_weight_0 = const()[name = string("logits_chunk_12_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100668288)))];
+            tensor<int32, [2]> logits_chunk_12_strides_0 = const()[name = string("logits_chunk_12_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_12_pad_type_0 = const()[name = string("logits_chunk_12_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_12_pad_0 = const()[name = string("logits_chunk_12_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_12_dilations_0 = const()[name = string("logits_chunk_12_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_12_groups_0 = const()[name = string("logits_chunk_12_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_12 = conv(dilations = logits_chunk_12_dilations_0, groups = logits_chunk_12_groups_0, pad = logits_chunk_12_pad_0, pad_type = logits_chunk_12_pad_type_0, strides = logits_chunk_12_strides_0, weight = logits_chunk_12_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_12")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_12_mul = mul(x = logits_chunk_12, y = temp_inverse)[name = string("logits_chunk_12_mul")];
+            tensor<int32, [1]> logits_chunk_12_max_axes_0 = const()[name = string("logits_chunk_12_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_12_max_keep_dims_0 = const()[name = string("logits_chunk_12_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_12_max = reduce_max(axes = logits_chunk_12_max_axes_0, keep_dims = logits_chunk_12_max_keep_dims_0, x = logits_chunk_12_mul)[name = string("logits_chunk_12_max")];
+            int32 logits_chunk_12_argmax_axis_0 = const()[name = string("logits_chunk_12_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_12_argmax_keep_dims_0 = const()[name = string("logits_chunk_12_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_12_argmax_output_dtype_0 = const()[name = string("logits_chunk_12_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> logits_chunk_12_argmax = reduce_argmax(axis = logits_chunk_12_argmax_axis_0, keep_dims = logits_chunk_12_argmax_keep_dims_0, output_dtype = logits_chunk_12_argmax_output_dtype_0, x = logits_chunk_12_mul)[name = string("logits_chunk_12_argmax")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_12_sub = sub(x = logits_chunk_12_mul, y = logits_chunk_12_max)[name = string("logits_chunk_12_sub")];
+            tensor<int32, [1]> logits_chunk_12_lse_sub_axes_0 = const()[name = string("logits_chunk_12_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_12_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_12_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_12_lse_sub = reduce_log_sum_exp(axes = logits_chunk_12_lse_sub_axes_0, keep_dims = logits_chunk_12_lse_sub_keep_dims_0, x = logits_chunk_12_sub)[name = string("logits_chunk_12_lse_sub")];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_12_lse = add(x = logits_chunk_12_lse_sub, y = logits_chunk_12_max)[name = string("logits_chunk_12_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_13_weight_0 = const()[name = string("logits_chunk_13_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109056960)))];
+            tensor<int32, [2]> logits_chunk_13_strides_0 = const()[name = string("logits_chunk_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_13_pad_type_0 = const()[name = string("logits_chunk_13_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_13_pad_0 = const()[name = string("logits_chunk_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_13_dilations_0 = const()[name = string("logits_chunk_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_13_groups_0 = const()[name = string("logits_chunk_13_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_13 = conv(dilations = logits_chunk_13_dilations_0, groups = logits_chunk_13_groups_0, pad = logits_chunk_13_pad_0, pad_type = logits_chunk_13_pad_type_0, strides = logits_chunk_13_strides_0, weight = logits_chunk_13_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_13")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_13_mul = mul(x = logits_chunk_13, y = temp_inverse)[name = string("logits_chunk_13_mul")];
+            tensor<int32, [1]> logits_chunk_13_max_axes_0 = const()[name = string("logits_chunk_13_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_13_max_keep_dims_0 = const()[name = string("logits_chunk_13_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_13_max = reduce_max(axes = logits_chunk_13_max_axes_0, keep_dims = logits_chunk_13_max_keep_dims_0, x = logits_chunk_13_mul)[name = string("logits_chunk_13_max")];
+            int32 logits_chunk_13_argmax_axis_0 = const()[name = string("logits_chunk_13_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_13_argmax_keep_dims_0 = const()[name = string("logits_chunk_13_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_13_argmax_output_dtype_0 = const()[name = string("logits_chunk_13_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> logits_chunk_13_argmax = reduce_argmax(axis = logits_chunk_13_argmax_axis_0, keep_dims = logits_chunk_13_argmax_keep_dims_0, output_dtype = logits_chunk_13_argmax_output_dtype_0, x = logits_chunk_13_mul)[name = string("logits_chunk_13_argmax")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_13_sub = sub(x = logits_chunk_13_mul, y = logits_chunk_13_max)[name = string("logits_chunk_13_sub")];
+            tensor<int32, [1]> logits_chunk_13_lse_sub_axes_0 = const()[name = string("logits_chunk_13_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_13_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_13_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_13_lse_sub = reduce_log_sum_exp(axes = logits_chunk_13_lse_sub_axes_0, keep_dims = logits_chunk_13_lse_sub_keep_dims_0, x = logits_chunk_13_sub)[name = string("logits_chunk_13_lse_sub")];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_13_lse = add(x = logits_chunk_13_lse_sub, y = logits_chunk_13_max)[name = string("logits_chunk_13_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_14_weight_0 = const()[name = string("logits_chunk_14_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117445632)))];
+            tensor<int32, [2]> logits_chunk_14_strides_0 = const()[name = string("logits_chunk_14_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_14_pad_type_0 = const()[name = string("logits_chunk_14_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_14_pad_0 = const()[name = string("logits_chunk_14_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_14_dilations_0 = const()[name = string("logits_chunk_14_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_14_groups_0 = const()[name = string("logits_chunk_14_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_14 = conv(dilations = logits_chunk_14_dilations_0, groups = logits_chunk_14_groups_0, pad = logits_chunk_14_pad_0, pad_type = logits_chunk_14_pad_type_0, strides = logits_chunk_14_strides_0, weight = logits_chunk_14_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_14")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_14_mul = mul(x = logits_chunk_14, y = temp_inverse)[name = string("logits_chunk_14_mul")];
+            tensor<int32, [1]> logits_chunk_14_max_axes_0 = const()[name = string("logits_chunk_14_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_14_max_keep_dims_0 = const()[name = string("logits_chunk_14_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_14_max = reduce_max(axes = logits_chunk_14_max_axes_0, keep_dims = logits_chunk_14_max_keep_dims_0, x = logits_chunk_14_mul)[name = string("logits_chunk_14_max")];
+            int32 logits_chunk_14_argmax_axis_0 = const()[name = string("logits_chunk_14_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_14_argmax_keep_dims_0 = const()[name = string("logits_chunk_14_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_14_argmax_output_dtype_0 = const()[name = string("logits_chunk_14_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> logits_chunk_14_argmax = reduce_argmax(axis = logits_chunk_14_argmax_axis_0, keep_dims = logits_chunk_14_argmax_keep_dims_0, output_dtype = logits_chunk_14_argmax_output_dtype_0, x = logits_chunk_14_mul)[name = string("logits_chunk_14_argmax")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_14_sub = sub(x = logits_chunk_14_mul, y = logits_chunk_14_max)[name = string("logits_chunk_14_sub")];
+            tensor<int32, [1]> logits_chunk_14_lse_sub_axes_0 = const()[name = string("logits_chunk_14_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_14_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_14_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_14_lse_sub = reduce_log_sum_exp(axes = logits_chunk_14_lse_sub_axes_0, keep_dims = logits_chunk_14_lse_sub_keep_dims_0, x = logits_chunk_14_sub)[name = string("logits_chunk_14_lse_sub")];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_14_lse = add(x = logits_chunk_14_lse_sub, y = logits_chunk_14_max)[name = string("logits_chunk_14_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_15_weight_0 = const()[name = string("logits_chunk_15_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125834304)))];
+            tensor<int32, [2]> logits_chunk_15_strides_0 = const()[name = string("logits_chunk_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_15_pad_type_0 = const()[name = string("logits_chunk_15_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_15_pad_0 = const()[name = string("logits_chunk_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_15_dilations_0 = const()[name = string("logits_chunk_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_15_groups_0 = const()[name = string("logits_chunk_15_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_15 = conv(dilations = logits_chunk_15_dilations_0, groups = logits_chunk_15_groups_0, pad = logits_chunk_15_pad_0, pad_type = logits_chunk_15_pad_type_0, strides = logits_chunk_15_strides_0, weight = logits_chunk_15_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_15")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_15_mul = mul(x = logits_chunk_15, y = temp_inverse)[name = string("logits_chunk_15_mul")];
+            tensor<int32, [1]> logits_chunk_15_max_axes_0 = const()[name = string("logits_chunk_15_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_15_max_keep_dims_0 = const()[name = string("logits_chunk_15_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_15_max = reduce_max(axes = logits_chunk_15_max_axes_0, keep_dims = logits_chunk_15_max_keep_dims_0, x = logits_chunk_15_mul)[name = string("logits_chunk_15_max")];
+            int32 logits_chunk_15_argmax_axis_0 = const()[name = string("logits_chunk_15_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_15_argmax_keep_dims_0 = const()[name = string("logits_chunk_15_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_15_argmax_output_dtype_0 = const()[name = string("logits_chunk_15_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 48]> logits_chunk_15_argmax = reduce_argmax(axis = logits_chunk_15_argmax_axis_0, keep_dims = logits_chunk_15_argmax_keep_dims_0, output_dtype = logits_chunk_15_argmax_output_dtype_0, x = logits_chunk_15_mul)[name = string("logits_chunk_15_argmax")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_15_sub = sub(x = logits_chunk_15_mul, y = logits_chunk_15_max)[name = string("logits_chunk_15_sub")];
+            tensor<int32, [1]> logits_chunk_15_lse_sub_axes_0 = const()[name = string("logits_chunk_15_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_15_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_15_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_15_lse_sub = reduce_log_sum_exp(axes = logits_chunk_15_lse_sub_axes_0, keep_dims = logits_chunk_15_lse_sub_keep_dims_0, x = logits_chunk_15_sub)[name = string("logits_chunk_15_lse_sub")];
+            tensor<fp16, [1, 1, 1, 48]> logits_chunk_15_lse = add(x = logits_chunk_15_lse_sub, y = logits_chunk_15_max)[name = string("logits_chunk_15_lse")];
+            int32 logits_lses_axis_0 = const()[name = string("logits_lses_axis_0"), val = int32(1)];
+            bool logits_lses_interleave_0 = const()[name = string("logits_lses_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 48]> logits_lses = concat(axis = logits_lses_axis_0, interleave = logits_lses_interleave_0, values = (logits_chunk_0_lse, logits_chunk_1_lse, logits_chunk_2_lse, logits_chunk_3_lse, logits_chunk_4_lse, logits_chunk_5_lse, logits_chunk_6_lse, logits_chunk_7_lse, logits_chunk_8_lse, logits_chunk_9_lse, logits_chunk_10_lse, logits_chunk_11_lse, logits_chunk_12_lse, logits_chunk_13_lse, logits_chunk_14_lse, logits_chunk_15_lse))[name = string("logits_lses")];
+            tensor<int32, [1]> logits_lses_max_axes_0 = const()[name = string("logits_lses_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_lses_max_keep_dims_0 = const()[name = string("logits_lses_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_lses_max = reduce_max(axes = logits_lses_max_axes_0, keep_dims = logits_lses_max_keep_dims_0, x = logits_lses)[name = string("logits_lses_max")];
+            tensor<fp16, [1, 16, 1, 48]> logits_lses_sub = sub(x = logits_lses, y = logits_lses_max)[name = string("logits_lses_sub")];
+            tensor<int32, [1]> logits_lses_logsumexp_axes_0 = const()[name = string("logits_lses_logsumexp_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_lses_logsumexp_keep_dims_0 = const()[name = string("logits_lses_logsumexp_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_lses_logsumexp = reduce_log_sum_exp(axes = logits_lses_logsumexp_axes_0, keep_dims = logits_lses_logsumexp_keep_dims_0, x = logits_lses_sub)[name = string("logits_lses_logsumexp")];
+            tensor<fp16, [1, 1, 1, 48]> logits_lse = add(x = logits_lses_logsumexp, y = logits_lses_max)[name = string("logits_lse")];
+            int32 logits_max_logits_chunks_axis_0 = const()[name = string("logits_max_logits_chunks_axis_0"), val = int32(1)];
+            bool logits_max_logits_chunks_interleave_0 = const()[name = string("logits_max_logits_chunks_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 48]> logits_max_logits_chunks = concat(axis = logits_max_logits_chunks_axis_0, interleave = logits_max_logits_chunks_interleave_0, values = (logits_chunk_0_max, logits_chunk_1_max, logits_chunk_2_max, logits_chunk_3_max, logits_chunk_4_max, logits_chunk_5_max, logits_chunk_6_max, logits_chunk_7_max, logits_chunk_8_max, logits_chunk_9_max, logits_chunk_10_max, logits_chunk_11_max, logits_chunk_12_max, logits_chunk_13_max, logits_chunk_14_max, logits_chunk_15_max))[name = string("logits_max_logits_chunks")];
+            tensor<int32, [1]> logits_max_logit_axes_0 = const()[name = string("logits_max_logit_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_max_logit_keep_dims_0 = const()[name = string("logits_max_logit_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 48]> logits_max_logit = reduce_max(axes = logits_max_logit_axes_0, keep_dims = logits_max_logit_keep_dims_0, x = logits_max_logits_chunks)[name = string("logits_max_logit")];
+            tensor<fp16, [1, 1, 1, 48]> logits_max_logit_sub = sub(x = logits_max_logit, y = logits_lse)[name = string("logits_max_logit_sub")];
+            tensor<fp16, [1, 1, 1, 48]> max_prob = exp(x = logits_max_logit_sub)[name = string("max_prob")];
+            tensor<fp16, [1, 1, 1, 48]> min_p_thresh = mul(x = max_prob, y = p)[name = string("min_p_thresh")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_0_sub_1 = sub(x = logits_chunk_0_mul, y = logits_lse)[name = string("logits_chunk_0_sub_1")];
+            tensor<fp16, [1, 2048, 1, 48]> probs_chunk_0 = exp(x = logits_chunk_0_sub_1)[name = string("probs_chunk_0")];
+            tensor<bool, [1, 2048, 1, 48]> mask_probs_chunk_0 = greater_equal(x = probs_chunk_0, y = min_p_thresh)[name = string("mask_probs_chunk_0")];
+            string mask_chunk_0_fp16_dtype_0 = const()[name = string("mask_chunk_0_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 48]> mask_chunk_0_fp16 = cast(dtype = mask_chunk_0_fp16_dtype_0, x = mask_probs_chunk_0)[name = string("cast_203")];
+            tensor<fp16, [1, 2048, 1, 48]> masked_probs_chunk_0 = select(a = probs_chunk_0, b = mask_chunk_0_fp16, cond = mask_probs_chunk_0)[name = string("masked_probs_chunk_0")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_1_sub_1 = sub(x = logits_chunk_1_mul, y = logits_lse)[name = string("logits_chunk_1_sub_1")];
+            tensor<fp16, [1, 2048, 1, 48]> probs_chunk_1 = exp(x = logits_chunk_1_sub_1)[name = string("probs_chunk_1")];
+            tensor<bool, [1, 2048, 1, 48]> mask_probs_chunk_1 = greater_equal(x = probs_chunk_1, y = min_p_thresh)[name = string("mask_probs_chunk_1")];
+            string mask_chunk_1_fp16_dtype_0 = const()[name = string("mask_chunk_1_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 48]> mask_chunk_1_fp16 = cast(dtype = mask_chunk_1_fp16_dtype_0, x = mask_probs_chunk_1)[name = string("cast_202")];
+            tensor<fp16, [1, 2048, 1, 48]> masked_probs_chunk_1 = select(a = probs_chunk_1, b = mask_chunk_1_fp16, cond = mask_probs_chunk_1)[name = string("masked_probs_chunk_1")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_2_sub_1 = sub(x = logits_chunk_2_mul, y = logits_lse)[name = string("logits_chunk_2_sub_1")];
+            tensor<fp16, [1, 2048, 1, 48]> probs_chunk_2 = exp(x = logits_chunk_2_sub_1)[name = string("probs_chunk_2")];
+            tensor<bool, [1, 2048, 1, 48]> mask_probs_chunk_2 = greater_equal(x = probs_chunk_2, y = min_p_thresh)[name = string("mask_probs_chunk_2")];
+            string mask_chunk_2_fp16_dtype_0 = const()[name = string("mask_chunk_2_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 48]> mask_chunk_2_fp16 = cast(dtype = mask_chunk_2_fp16_dtype_0, x = mask_probs_chunk_2)[name = string("cast_201")];
+            tensor<fp16, [1, 2048, 1, 48]> masked_probs_chunk_2 = select(a = probs_chunk_2, b = mask_chunk_2_fp16, cond = mask_probs_chunk_2)[name = string("masked_probs_chunk_2")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_3_sub_1 = sub(x = logits_chunk_3_mul, y = logits_lse)[name = string("logits_chunk_3_sub_1")];
+            tensor<fp16, [1, 2048, 1, 48]> probs_chunk_3 = exp(x = logits_chunk_3_sub_1)[name = string("probs_chunk_3")];
+            tensor<bool, [1, 2048, 1, 48]> mask_probs_chunk_3 = greater_equal(x = probs_chunk_3, y = min_p_thresh)[name = string("mask_probs_chunk_3")];
+            string mask_chunk_3_fp16_dtype_0 = const()[name = string("mask_chunk_3_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 48]> mask_chunk_3_fp16 = cast(dtype = mask_chunk_3_fp16_dtype_0, x = mask_probs_chunk_3)[name = string("cast_200")];
+            tensor<fp16, [1, 2048, 1, 48]> masked_probs_chunk_3 = select(a = probs_chunk_3, b = mask_chunk_3_fp16, cond = mask_probs_chunk_3)[name = string("masked_probs_chunk_3")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_4_sub_1 = sub(x = logits_chunk_4_mul, y = logits_lse)[name = string("logits_chunk_4_sub_1")];
+            tensor<fp16, [1, 2048, 1, 48]> probs_chunk_4 = exp(x = logits_chunk_4_sub_1)[name = string("probs_chunk_4")];
+            tensor<bool, [1, 2048, 1, 48]> mask_probs_chunk_4 = greater_equal(x = probs_chunk_4, y = min_p_thresh)[name = string("mask_probs_chunk_4")];
+            string mask_chunk_4_fp16_dtype_0 = const()[name = string("mask_chunk_4_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 48]> mask_chunk_4_fp16 = cast(dtype = mask_chunk_4_fp16_dtype_0, x = mask_probs_chunk_4)[name = string("cast_199")];
+            tensor<fp16, [1, 2048, 1, 48]> masked_probs_chunk_4 = select(a = probs_chunk_4, b = mask_chunk_4_fp16, cond = mask_probs_chunk_4)[name = string("masked_probs_chunk_4")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_5_sub_1 = sub(x = logits_chunk_5_mul, y = logits_lse)[name = string("logits_chunk_5_sub_1")];
+            tensor<fp16, [1, 2048, 1, 48]> probs_chunk_5 = exp(x = logits_chunk_5_sub_1)[name = string("probs_chunk_5")];
+            tensor<bool, [1, 2048, 1, 48]> mask_probs_chunk_5 = greater_equal(x = probs_chunk_5, y = min_p_thresh)[name = string("mask_probs_chunk_5")];
+            string mask_chunk_5_fp16_dtype_0 = const()[name = string("mask_chunk_5_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 48]> mask_chunk_5_fp16 = cast(dtype = mask_chunk_5_fp16_dtype_0, x = mask_probs_chunk_5)[name = string("cast_198")];
+            tensor<fp16, [1, 2048, 1, 48]> masked_probs_chunk_5 = select(a = probs_chunk_5, b = mask_chunk_5_fp16, cond = mask_probs_chunk_5)[name = string("masked_probs_chunk_5")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_6_sub_1 = sub(x = logits_chunk_6_mul, y = logits_lse)[name = string("logits_chunk_6_sub_1")];
+            tensor<fp16, [1, 2048, 1, 48]> probs_chunk_6 = exp(x = logits_chunk_6_sub_1)[name = string("probs_chunk_6")];
+            tensor<bool, [1, 2048, 1, 48]> mask_probs_chunk_6 = greater_equal(x = probs_chunk_6, y = min_p_thresh)[name = string("mask_probs_chunk_6")];
+            string mask_chunk_6_fp16_dtype_0 = const()[name = string("mask_chunk_6_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 48]> mask_chunk_6_fp16 = cast(dtype = mask_chunk_6_fp16_dtype_0, x = mask_probs_chunk_6)[name = string("cast_197")];
+            tensor<fp16, [1, 2048, 1, 48]> masked_probs_chunk_6 = select(a = probs_chunk_6, b = mask_chunk_6_fp16, cond = mask_probs_chunk_6)[name = string("masked_probs_chunk_6")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_7_sub_1 = sub(x = logits_chunk_7_mul, y = logits_lse)[name = string("logits_chunk_7_sub_1")];
+            tensor<fp16, [1, 2048, 1, 48]> probs_chunk_7 = exp(x = logits_chunk_7_sub_1)[name = string("probs_chunk_7")];
+            tensor<bool, [1, 2048, 1, 48]> mask_probs_chunk_7 = greater_equal(x = probs_chunk_7, y = min_p_thresh)[name = string("mask_probs_chunk_7")];
+            string mask_chunk_7_fp16_dtype_0 = const()[name = string("mask_chunk_7_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 48]> mask_chunk_7_fp16 = cast(dtype = mask_chunk_7_fp16_dtype_0, x = mask_probs_chunk_7)[name = string("cast_196")];
+            tensor<fp16, [1, 2048, 1, 48]> masked_probs_chunk_7 = select(a = probs_chunk_7, b = mask_chunk_7_fp16, cond = mask_probs_chunk_7)[name = string("masked_probs_chunk_7")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_8_sub_1 = sub(x = logits_chunk_8_mul, y = logits_lse)[name = string("logits_chunk_8_sub_1")];
+            tensor<fp16, [1, 2048, 1, 48]> probs_chunk_8 = exp(x = logits_chunk_8_sub_1)[name = string("probs_chunk_8")];
+            tensor<bool, [1, 2048, 1, 48]> mask_probs_chunk_8 = greater_equal(x = probs_chunk_8, y = min_p_thresh)[name = string("mask_probs_chunk_8")];
+            string mask_chunk_8_fp16_dtype_0 = const()[name = string("mask_chunk_8_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 48]> mask_chunk_8_fp16 = cast(dtype = mask_chunk_8_fp16_dtype_0, x = mask_probs_chunk_8)[name = string("cast_195")];
+            tensor<fp16, [1, 2048, 1, 48]> masked_probs_chunk_8 = select(a = probs_chunk_8, b = mask_chunk_8_fp16, cond = mask_probs_chunk_8)[name = string("masked_probs_chunk_8")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_9_sub_1 = sub(x = logits_chunk_9_mul, y = logits_lse)[name = string("logits_chunk_9_sub_1")];
+            tensor<fp16, [1, 2048, 1, 48]> probs_chunk_9 = exp(x = logits_chunk_9_sub_1)[name = string("probs_chunk_9")];
+            tensor<bool, [1, 2048, 1, 48]> mask_probs_chunk_9 = greater_equal(x = probs_chunk_9, y = min_p_thresh)[name = string("mask_probs_chunk_9")];
+            string mask_chunk_9_fp16_dtype_0 = const()[name = string("mask_chunk_9_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 48]> mask_chunk_9_fp16 = cast(dtype = mask_chunk_9_fp16_dtype_0, x = mask_probs_chunk_9)[name = string("cast_194")];
+            tensor<fp16, [1, 2048, 1, 48]> masked_probs_chunk_9 = select(a = probs_chunk_9, b = mask_chunk_9_fp16, cond = mask_probs_chunk_9)[name = string("masked_probs_chunk_9")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_10_sub_1 = sub(x = logits_chunk_10_mul, y = logits_lse)[name = string("logits_chunk_10_sub_1")];
+            tensor<fp16, [1, 2048, 1, 48]> probs_chunk_10 = exp(x = logits_chunk_10_sub_1)[name = string("probs_chunk_10")];
+            tensor<bool, [1, 2048, 1, 48]> mask_probs_chunk_10 = greater_equal(x = probs_chunk_10, y = min_p_thresh)[name = string("mask_probs_chunk_10")];
+            string mask_chunk_10_fp16_dtype_0 = const()[name = string("mask_chunk_10_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 48]> mask_chunk_10_fp16 = cast(dtype = mask_chunk_10_fp16_dtype_0, x = mask_probs_chunk_10)[name = string("cast_193")];
+            tensor<fp16, [1, 2048, 1, 48]> masked_probs_chunk_10 = select(a = probs_chunk_10, b = mask_chunk_10_fp16, cond = mask_probs_chunk_10)[name = string("masked_probs_chunk_10")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_11_sub_1 = sub(x = logits_chunk_11_mul, y = logits_lse)[name = string("logits_chunk_11_sub_1")];
+            tensor<fp16, [1, 2048, 1, 48]> probs_chunk_11 = exp(x = logits_chunk_11_sub_1)[name = string("probs_chunk_11")];
+            tensor<bool, [1, 2048, 1, 48]> mask_probs_chunk_11 = greater_equal(x = probs_chunk_11, y = min_p_thresh)[name = string("mask_probs_chunk_11")];
+            string mask_chunk_11_fp16_dtype_0 = const()[name = string("mask_chunk_11_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 48]> mask_chunk_11_fp16 = cast(dtype = mask_chunk_11_fp16_dtype_0, x = mask_probs_chunk_11)[name = string("cast_192")];
+            tensor<fp16, [1, 2048, 1, 48]> masked_probs_chunk_11 = select(a = probs_chunk_11, b = mask_chunk_11_fp16, cond = mask_probs_chunk_11)[name = string("masked_probs_chunk_11")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_12_sub_1 = sub(x = logits_chunk_12_mul, y = logits_lse)[name = string("logits_chunk_12_sub_1")];
+            tensor<fp16, [1, 2048, 1, 48]> probs_chunk_12 = exp(x = logits_chunk_12_sub_1)[name = string("probs_chunk_12")];
+            tensor<bool, [1, 2048, 1, 48]> mask_probs_chunk_12 = greater_equal(x = probs_chunk_12, y = min_p_thresh)[name = string("mask_probs_chunk_12")];
+            string mask_chunk_12_fp16_dtype_0 = const()[name = string("mask_chunk_12_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 48]> mask_chunk_12_fp16 = cast(dtype = mask_chunk_12_fp16_dtype_0, x = mask_probs_chunk_12)[name = string("cast_191")];
+            tensor<fp16, [1, 2048, 1, 48]> masked_probs_chunk_12 = select(a = probs_chunk_12, b = mask_chunk_12_fp16, cond = mask_probs_chunk_12)[name = string("masked_probs_chunk_12")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_13_sub_1 = sub(x = logits_chunk_13_mul, y = logits_lse)[name = string("logits_chunk_13_sub_1")];
+            tensor<fp16, [1, 2048, 1, 48]> probs_chunk_13 = exp(x = logits_chunk_13_sub_1)[name = string("probs_chunk_13")];
+            tensor<bool, [1, 2048, 1, 48]> mask_probs_chunk_13 = greater_equal(x = probs_chunk_13, y = min_p_thresh)[name = string("mask_probs_chunk_13")];
+            string mask_chunk_13_fp16_dtype_0 = const()[name = string("mask_chunk_13_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 48]> mask_chunk_13_fp16 = cast(dtype = mask_chunk_13_fp16_dtype_0, x = mask_probs_chunk_13)[name = string("cast_190")];
+            tensor<fp16, [1, 2048, 1, 48]> masked_probs_chunk_13 = select(a = probs_chunk_13, b = mask_chunk_13_fp16, cond = mask_probs_chunk_13)[name = string("masked_probs_chunk_13")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_14_sub_1 = sub(x = logits_chunk_14_mul, y = logits_lse)[name = string("logits_chunk_14_sub_1")];
+            tensor<fp16, [1, 2048, 1, 48]> probs_chunk_14 = exp(x = logits_chunk_14_sub_1)[name = string("probs_chunk_14")];
+            tensor<bool, [1, 2048, 1, 48]> mask_probs_chunk_14 = greater_equal(x = probs_chunk_14, y = min_p_thresh)[name = string("mask_probs_chunk_14")];
+            string mask_chunk_14_fp16_dtype_0 = const()[name = string("mask_chunk_14_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 48]> mask_chunk_14_fp16 = cast(dtype = mask_chunk_14_fp16_dtype_0, x = mask_probs_chunk_14)[name = string("cast_189")];
+            tensor<fp16, [1, 2048, 1, 48]> masked_probs_chunk_14 = select(a = probs_chunk_14, b = mask_chunk_14_fp16, cond = mask_probs_chunk_14)[name = string("masked_probs_chunk_14")];
+            tensor<fp16, [1, 2048, 1, 48]> logits_chunk_15_sub_1 = sub(x = logits_chunk_15_mul, y = logits_lse)[name = string("logits_chunk_15_sub_1")];
+            tensor<fp16, [1, 2048, 1, 48]> probs_chunk_15 = exp(x = logits_chunk_15_sub_1)[name = string("probs_chunk_15")];
+            tensor<bool, [1, 2048, 1, 48]> mask_probs_chunk_15 = greater_equal(x = probs_chunk_15, y = min_p_thresh)[name = string("mask_probs_chunk_15")];
+            string mask_chunk_15_fp16_dtype_0 = const()[name = string("mask_chunk_15_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 48]> mask_chunk_15_fp16 = cast(dtype = mask_chunk_15_fp16_dtype_0, x = mask_probs_chunk_15)[name = string("cast_188")];
+            tensor<fp16, [1, 2048, 1, 48]> masked_probs_chunk_15 = select(a = probs_chunk_15, b = mask_chunk_15_fp16, cond = mask_probs_chunk_15)[name = string("masked_probs_chunk_15")];
+            int32 probs_axis_0 = const()[name = string("probs_axis_0"), val = int32(1)];
+            bool probs_interleave_0 = const()[name = string("probs_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 32768, 1, 48]> probs = concat(axis = probs_axis_0, interleave = probs_interleave_0, values = (masked_probs_chunk_0, masked_probs_chunk_1, masked_probs_chunk_2, masked_probs_chunk_3, masked_probs_chunk_4, masked_probs_chunk_5, masked_probs_chunk_6, masked_probs_chunk_7, masked_probs_chunk_8, masked_probs_chunk_9, masked_probs_chunk_10, masked_probs_chunk_11, masked_probs_chunk_12, masked_probs_chunk_13, masked_probs_chunk_14, masked_probs_chunk_15))[name = string("probs")];
+            string probs_fp32_dtype_0 = const()[name = string("probs_fp32_dtype_0"), val = string("fp32")];
+            int32 probs_cumsum_axis_0 = const()[name = string("probs_cumsum_axis_0"), val = int32(1)];
+            bool probs_cumsum_exclusive_0 = const()[name = string("probs_cumsum_exclusive_0"), val = bool(false)];
+            bool probs_cumsum_reverse_0 = const()[name = string("probs_cumsum_reverse_0"), val = bool(false)];
+            tensor<fp32, [1, 32768, 1, 48]> probs_fp32 = cast(dtype = probs_fp32_dtype_0, x = probs)[name = string("cast_187")];
+            tensor<fp32, [1, 32768, 1, 48]> probs_cumsum = cumsum(axis = probs_cumsum_axis_0, exclusive = probs_cumsum_exclusive_0, reverse = probs_cumsum_reverse_0, x = probs_fp32)[name = string("probs_cumsum")];
+            tensor<int32, [1]> probs_sum_indices_0 = const()[name = string("probs_sum_indices_0"), val = tensor<int32, [1]>([32767])];
+            int32 probs_sum_axis_0 = const()[name = string("probs_sum_axis_0"), val = int32(1)];
+            int32 probs_sum_batch_dims_0 = const()[name = string("probs_sum_batch_dims_0"), val = int32(0)];
+            bool probs_sum_validate_indices_0 = const()[name = string("probs_sum_validate_indices_0"), val = bool(false)];
+            tensor<fp32, [1, 1, 1, 48]> probs_sum = gather(axis = probs_sum_axis_0, batch_dims = probs_sum_batch_dims_0, indices = probs_sum_indices_0, validate_indices = probs_sum_validate_indices_0, x = probs_cumsum)[name = string("probs_sum")];
+            tensor<fp32, [1, 1, 1, 48]> random_number_scaled = mul(x = random_number, y = probs_sum)[name = string("random_number_scaled")];
+            tensor<bool, [1, 32768, 1, 48]> probs_greater = greater(x = probs_cumsum, y = random_number_scaled)[name = string("probs_greater")];
+            string probs_greater_int32_dtype_0 = const()[name = string("probs_greater_int32_dtype_0"), val = string("int32")];
+            int32 sampled_index_axis_0 = const()[name = string("sampled_index_axis_0"), val = int32(1)];
+            bool sampled_index_keep_dims_0 = const()[name = string("sampled_index_keep_dims_0"), val = bool(true)];
+            string sampled_index_output_dtype_0 = const()[name = string("sampled_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 32768, 1, 48]> probs_greater_int32 = cast(dtype = probs_greater_int32_dtype_0, x = probs_greater)[name = string("cast_186")];
+            tensor<int32, [1, 1, 1, 48]> sampled_index = reduce_argmax(axis = sampled_index_axis_0, keep_dims = sampled_index_keep_dims_0, output_dtype = sampled_index_output_dtype_0, x = probs_greater_int32)[name = string("sampled_index")];
+            int32 sampled_index_probability_axis_0 = const()[name = string("sampled_index_probability_axis_0"), val = int32(1)];
+            bool sampled_index_probability_validate_indices_0 = const()[name = string("sampled_index_probability_validate_indices_0"), val = bool(false)];
+            tensor<fp32, [1, 1, 1, 48]> sampled_index_probability = gather_along_axis(axis = sampled_index_probability_axis_0, indices = sampled_index, validate_indices = sampled_index_probability_validate_indices_0, x = probs_fp32)[name = string("sampled_index_probability")];
+            int32 max_logit_index_axis_0 = const()[name = string("max_logit_index_axis_0"), val = int32(1)];
+            bool max_logit_index_keep_dims_0 = const()[name = string("max_logit_index_keep_dims_0"), val = bool(true)];
+            string max_logit_index_output_dtype_0 = const()[name = string("max_logit_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 1, 1, 48]> max_logit_index = reduce_argmax(axis = max_logit_index_axis_0, keep_dims = max_logit_index_keep_dims_0, output_dtype = max_logit_index_output_dtype_0, x = logits_max_logits_chunks)[name = string("max_logit_index")];
+            string indices_chunk_0_int32_dtype_0 = const()[name = string("indices_chunk_0_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_1_int32_dtype_0 = const()[name = string("indices_chunk_1_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_2_int32_dtype_0 = const()[name = string("indices_chunk_2_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_3_int32_dtype_0 = const()[name = string("indices_chunk_3_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_4_int32_dtype_0 = const()[name = string("indices_chunk_4_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_5_int32_dtype_0 = const()[name = string("indices_chunk_5_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_6_int32_dtype_0 = const()[name = string("indices_chunk_6_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_7_int32_dtype_0 = const()[name = string("indices_chunk_7_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_8_int32_dtype_0 = const()[name = string("indices_chunk_8_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_9_int32_dtype_0 = const()[name = string("indices_chunk_9_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_10_int32_dtype_0 = const()[name = string("indices_chunk_10_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_11_int32_dtype_0 = const()[name = string("indices_chunk_11_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_12_int32_dtype_0 = const()[name = string("indices_chunk_12_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_13_int32_dtype_0 = const()[name = string("indices_chunk_13_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_14_int32_dtype_0 = const()[name = string("indices_chunk_14_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_15_int32_dtype_0 = const()[name = string("indices_chunk_15_int32_dtype_0"), val = string("int32")];
+            int32 indices_axis_0 = const()[name = string("indices_axis_0"), val = int32(1)];
+            bool indices_interleave_0 = const()[name = string("indices_interleave_0"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_15_int32 = cast(dtype = indices_chunk_15_int32_dtype_0, x = logits_chunk_15_argmax)[name = string("cast_170")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_14_int32 = cast(dtype = indices_chunk_14_int32_dtype_0, x = logits_chunk_14_argmax)[name = string("cast_171")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_13_int32 = cast(dtype = indices_chunk_13_int32_dtype_0, x = logits_chunk_13_argmax)[name = string("cast_172")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_12_int32 = cast(dtype = indices_chunk_12_int32_dtype_0, x = logits_chunk_12_argmax)[name = string("cast_173")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_11_int32 = cast(dtype = indices_chunk_11_int32_dtype_0, x = logits_chunk_11_argmax)[name = string("cast_174")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_10_int32 = cast(dtype = indices_chunk_10_int32_dtype_0, x = logits_chunk_10_argmax)[name = string("cast_175")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_9_int32 = cast(dtype = indices_chunk_9_int32_dtype_0, x = logits_chunk_9_argmax)[name = string("cast_176")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_8_int32 = cast(dtype = indices_chunk_8_int32_dtype_0, x = logits_chunk_8_argmax)[name = string("cast_177")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_7_int32 = cast(dtype = indices_chunk_7_int32_dtype_0, x = logits_chunk_7_argmax)[name = string("cast_178")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_6_int32 = cast(dtype = indices_chunk_6_int32_dtype_0, x = logits_chunk_6_argmax)[name = string("cast_179")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_5_int32 = cast(dtype = indices_chunk_5_int32_dtype_0, x = logits_chunk_5_argmax)[name = string("cast_180")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_4_int32 = cast(dtype = indices_chunk_4_int32_dtype_0, x = logits_chunk_4_argmax)[name = string("cast_181")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_3_int32 = cast(dtype = indices_chunk_3_int32_dtype_0, x = logits_chunk_3_argmax)[name = string("cast_182")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_2_int32 = cast(dtype = indices_chunk_2_int32_dtype_0, x = logits_chunk_2_argmax)[name = string("cast_183")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_1_int32 = cast(dtype = indices_chunk_1_int32_dtype_0, x = logits_chunk_1_argmax)[name = string("cast_184")];
+            tensor<int32, [1, 1, 1, 48]> indices_chunk_0_int32 = cast(dtype = indices_chunk_0_int32_dtype_0, x = logits_chunk_0_argmax)[name = string("cast_185")];
+            tensor<int32, [1, 16, 1, 48]> indices = concat(axis = indices_axis_0, interleave = indices_interleave_0, values = (indices_chunk_0_int32, indices_chunk_1_int32, indices_chunk_2_int32, indices_chunk_3_int32, indices_chunk_4_int32, indices_chunk_5_int32, indices_chunk_6_int32, indices_chunk_7_int32, indices_chunk_8_int32, indices_chunk_9_int32, indices_chunk_10_int32, indices_chunk_11_int32, indices_chunk_12_int32, indices_chunk_13_int32, indices_chunk_14_int32, indices_chunk_15_int32))[name = string("indices")];
+            int32 argmax_chunks_axis_0 = const()[name = string("argmax_chunks_axis_0"), val = int32(1)];
+            bool argmax_chunks_validate_indices_0 = const()[name = string("argmax_chunks_validate_indices_0"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 48]> argmax_chunks = gather_along_axis(axis = argmax_chunks_axis_0, indices = max_logit_index, validate_indices = argmax_chunks_validate_indices_0, x = indices)[name = string("argmax_chunks")];
+            int32 mul_0_x_0 = const()[name = string("mul_0_x_0"), val = int32(2048)];
+            tensor<int32, [1, 1, 1, 48]> mul_0 = mul(x = mul_0_x_0, y = max_logit_index)[name = string("mul_0")];
+            tensor<int32, [1, 1, 1, 48]> argmax = add(x = argmax_chunks, y = mul_0)[name = string("argmax")];
+        } -> (sampled_index, sampled_index_probability, argmax, max_prob);
+    func min_p_length_64<ios18>(tensor<fp16, [1, 2048, 1, 64]> hidden_states, tensor<fp16, [1]> p, tensor<fp32, [64]> random_number, tensor<fp16, [1]> temp) {
+            tensor<fp16, [1, 2048, 1, 64]> final_norm_rmsnorm_abs = abs(x = hidden_states)[name = string("final_norm_rmsnorm_abs")];
+            tensor<int32, [1]> final_norm_rmsnorm_maxval_axes_0 = const()[name = string("final_norm_rmsnorm_maxval_axes_0"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_maxval_keep_dims_0 = const()[name = string("final_norm_rmsnorm_maxval_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> final_norm_rmsnorm_maxval = reduce_max(axes = final_norm_rmsnorm_maxval_axes_0, keep_dims = final_norm_rmsnorm_maxval_keep_dims_0, x = final_norm_rmsnorm_abs)[name = string("final_norm_rmsnorm_maxval")];
+            fp16 final_norm_rmsnorm_maxval_clipped_alpha_0 = const()[name = string("final_norm_rmsnorm_maxval_clipped_alpha_0"), val = fp16(0x1p-14)];
+            fp16 final_norm_rmsnorm_maxval_clipped_beta_0 = const()[name = string("final_norm_rmsnorm_maxval_clipped_beta_0"), val = fp16(inf)];
+            tensor<fp16, [1, 1, 1, 64]> final_norm_rmsnorm_maxval_clipped = clip(alpha = final_norm_rmsnorm_maxval_clipped_alpha_0, beta = final_norm_rmsnorm_maxval_clipped_beta_0, x = final_norm_rmsnorm_maxval)[name = string("final_norm_rmsnorm_maxval_clipped")];
+            tensor<fp16, [1, 2048, 1, 64]> final_norm_rmsnorm_scaled = real_div(x = hidden_states, y = final_norm_rmsnorm_maxval_clipped)[name = string("final_norm_rmsnorm_scaled")];
+            tensor<int32, [1]> final_norm_rmsnorm_squared_sum_axes_0 = const()[name = string("final_norm_rmsnorm_squared_sum_axes_0"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_squared_sum_keep_dims_0 = const()[name = string("final_norm_rmsnorm_squared_sum_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> final_norm_rmsnorm_squared_sum = reduce_sum_square(axes = final_norm_rmsnorm_squared_sum_axes_0, keep_dims = final_norm_rmsnorm_squared_sum_keep_dims_0, x = final_norm_rmsnorm_scaled)[name = string("final_norm_rmsnorm_squared_sum")];
+            fp16 final_norm_rmsnorm_rsqrt_epsilon_0 = const()[name = string("final_norm_rmsnorm_rsqrt_epsilon_0"), val = fp16(0x1p-14)];
+            tensor<fp16, [1, 1, 1, 64]> final_norm_rmsnorm_rsqrt = rsqrt(epsilon = final_norm_rmsnorm_rsqrt_epsilon_0, x = final_norm_rmsnorm_squared_sum)[name = string("final_norm_rmsnorm_rsqrt")];
+            fp16 final_norm_rmsnorm_dim_scaled_y_0 = const()[name = string("final_norm_rmsnorm_dim_scaled_y_0"), val = fp16(0x1.6ap+5)];
+            tensor<fp16, [1, 2048, 1, 64]> final_norm_rmsnorm_dim_scaled = mul(x = final_norm_rmsnorm_scaled, y = final_norm_rmsnorm_dim_scaled_y_0)[name = string("final_norm_rmsnorm_dim_scaled")];
+            tensor<fp16, [1, 2048, 1, 64]> final_norm_rmsnorm_normalized = mul(x = final_norm_rmsnorm_dim_scaled, y = final_norm_rmsnorm_rsqrt)[name = string("final_norm_rmsnorm_normalized")];
+            tensor<fp16, [2048, 1, 1]> final_norm_rmsnorm_y_0 = const()[name = string("final_norm_rmsnorm_y_0"), val = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 2048, 1, 64]> final_norm_rmsnorm = mul(x = final_norm_rmsnorm_normalized, y = final_norm_rmsnorm_y_0)[name = string("final_norm_rmsnorm")];
+            fp16 temp_inverse_epsilon_0 = const()[name = string("temp_inverse_epsilon_0"), val = fp16(0x0p+0)];
+            tensor<fp16, [1]> temp_inverse = inverse(epsilon = temp_inverse_epsilon_0, x = temp)[name = string("temp_inverse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_0_weight_0 = const()[name = string("logits_chunk_0_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4224)))];
+            tensor<int32, [2]> logits_chunk_0_strides_0 = const()[name = string("logits_chunk_0_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_0_pad_type_0 = const()[name = string("logits_chunk_0_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_0_pad_0 = const()[name = string("logits_chunk_0_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_0_dilations_0 = const()[name = string("logits_chunk_0_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_0_groups_0 = const()[name = string("logits_chunk_0_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_0 = conv(dilations = logits_chunk_0_dilations_0, groups = logits_chunk_0_groups_0, pad = logits_chunk_0_pad_0, pad_type = logits_chunk_0_pad_type_0, strides = logits_chunk_0_strides_0, weight = logits_chunk_0_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_0")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_0_mul = mul(x = logits_chunk_0, y = temp_inverse)[name = string("logits_chunk_0_mul")];
+            tensor<int32, [1]> logits_chunk_0_max_axes_0 = const()[name = string("logits_chunk_0_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_0_max_keep_dims_0 = const()[name = string("logits_chunk_0_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_0_max = reduce_max(axes = logits_chunk_0_max_axes_0, keep_dims = logits_chunk_0_max_keep_dims_0, x = logits_chunk_0_mul)[name = string("logits_chunk_0_max")];
+            int32 logits_chunk_0_argmax_axis_0 = const()[name = string("logits_chunk_0_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_0_argmax_keep_dims_0 = const()[name = string("logits_chunk_0_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_0_argmax_output_dtype_0 = const()[name = string("logits_chunk_0_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> logits_chunk_0_argmax = reduce_argmax(axis = logits_chunk_0_argmax_axis_0, keep_dims = logits_chunk_0_argmax_keep_dims_0, output_dtype = logits_chunk_0_argmax_output_dtype_0, x = logits_chunk_0_mul)[name = string("logits_chunk_0_argmax")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_0_sub = sub(x = logits_chunk_0_mul, y = logits_chunk_0_max)[name = string("logits_chunk_0_sub")];
+            tensor<int32, [1]> logits_chunk_0_lse_sub_axes_0 = const()[name = string("logits_chunk_0_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_0_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_0_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_0_lse_sub = reduce_log_sum_exp(axes = logits_chunk_0_lse_sub_axes_0, keep_dims = logits_chunk_0_lse_sub_keep_dims_0, x = logits_chunk_0_sub)[name = string("logits_chunk_0_lse_sub")];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_0_lse = add(x = logits_chunk_0_lse_sub, y = logits_chunk_0_max)[name = string("logits_chunk_0_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_1_weight_0 = const()[name = string("logits_chunk_1_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8392896)))];
+            tensor<int32, [2]> logits_chunk_1_strides_0 = const()[name = string("logits_chunk_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_1_pad_type_0 = const()[name = string("logits_chunk_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_1_pad_0 = const()[name = string("logits_chunk_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_1_dilations_0 = const()[name = string("logits_chunk_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_1_groups_0 = const()[name = string("logits_chunk_1_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_1 = conv(dilations = logits_chunk_1_dilations_0, groups = logits_chunk_1_groups_0, pad = logits_chunk_1_pad_0, pad_type = logits_chunk_1_pad_type_0, strides = logits_chunk_1_strides_0, weight = logits_chunk_1_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_1")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_1_mul = mul(x = logits_chunk_1, y = temp_inverse)[name = string("logits_chunk_1_mul")];
+            tensor<int32, [1]> logits_chunk_1_max_axes_0 = const()[name = string("logits_chunk_1_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_1_max_keep_dims_0 = const()[name = string("logits_chunk_1_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_1_max = reduce_max(axes = logits_chunk_1_max_axes_0, keep_dims = logits_chunk_1_max_keep_dims_0, x = logits_chunk_1_mul)[name = string("logits_chunk_1_max")];
+            int32 logits_chunk_1_argmax_axis_0 = const()[name = string("logits_chunk_1_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_1_argmax_keep_dims_0 = const()[name = string("logits_chunk_1_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_1_argmax_output_dtype_0 = const()[name = string("logits_chunk_1_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> logits_chunk_1_argmax = reduce_argmax(axis = logits_chunk_1_argmax_axis_0, keep_dims = logits_chunk_1_argmax_keep_dims_0, output_dtype = logits_chunk_1_argmax_output_dtype_0, x = logits_chunk_1_mul)[name = string("logits_chunk_1_argmax")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_1_sub = sub(x = logits_chunk_1_mul, y = logits_chunk_1_max)[name = string("logits_chunk_1_sub")];
+            tensor<int32, [1]> logits_chunk_1_lse_sub_axes_0 = const()[name = string("logits_chunk_1_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_1_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_1_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_1_lse_sub = reduce_log_sum_exp(axes = logits_chunk_1_lse_sub_axes_0, keep_dims = logits_chunk_1_lse_sub_keep_dims_0, x = logits_chunk_1_sub)[name = string("logits_chunk_1_lse_sub")];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_1_lse = add(x = logits_chunk_1_lse_sub, y = logits_chunk_1_max)[name = string("logits_chunk_1_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_2_weight_0 = const()[name = string("logits_chunk_2_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16781568)))];
+            tensor<int32, [2]> logits_chunk_2_strides_0 = const()[name = string("logits_chunk_2_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_2_pad_type_0 = const()[name = string("logits_chunk_2_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_2_pad_0 = const()[name = string("logits_chunk_2_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_2_dilations_0 = const()[name = string("logits_chunk_2_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_2_groups_0 = const()[name = string("logits_chunk_2_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_2 = conv(dilations = logits_chunk_2_dilations_0, groups = logits_chunk_2_groups_0, pad = logits_chunk_2_pad_0, pad_type = logits_chunk_2_pad_type_0, strides = logits_chunk_2_strides_0, weight = logits_chunk_2_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_2")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_2_mul = mul(x = logits_chunk_2, y = temp_inverse)[name = string("logits_chunk_2_mul")];
+            tensor<int32, [1]> logits_chunk_2_max_axes_0 = const()[name = string("logits_chunk_2_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_2_max_keep_dims_0 = const()[name = string("logits_chunk_2_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_2_max = reduce_max(axes = logits_chunk_2_max_axes_0, keep_dims = logits_chunk_2_max_keep_dims_0, x = logits_chunk_2_mul)[name = string("logits_chunk_2_max")];
+            int32 logits_chunk_2_argmax_axis_0 = const()[name = string("logits_chunk_2_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_2_argmax_keep_dims_0 = const()[name = string("logits_chunk_2_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_2_argmax_output_dtype_0 = const()[name = string("logits_chunk_2_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> logits_chunk_2_argmax = reduce_argmax(axis = logits_chunk_2_argmax_axis_0, keep_dims = logits_chunk_2_argmax_keep_dims_0, output_dtype = logits_chunk_2_argmax_output_dtype_0, x = logits_chunk_2_mul)[name = string("logits_chunk_2_argmax")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_2_sub = sub(x = logits_chunk_2_mul, y = logits_chunk_2_max)[name = string("logits_chunk_2_sub")];
+            tensor<int32, [1]> logits_chunk_2_lse_sub_axes_0 = const()[name = string("logits_chunk_2_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_2_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_2_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_2_lse_sub = reduce_log_sum_exp(axes = logits_chunk_2_lse_sub_axes_0, keep_dims = logits_chunk_2_lse_sub_keep_dims_0, x = logits_chunk_2_sub)[name = string("logits_chunk_2_lse_sub")];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_2_lse = add(x = logits_chunk_2_lse_sub, y = logits_chunk_2_max)[name = string("logits_chunk_2_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_3_weight_0 = const()[name = string("logits_chunk_3_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25170240)))];
+            tensor<int32, [2]> logits_chunk_3_strides_0 = const()[name = string("logits_chunk_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_3_pad_type_0 = const()[name = string("logits_chunk_3_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_3_pad_0 = const()[name = string("logits_chunk_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_3_dilations_0 = const()[name = string("logits_chunk_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_3_groups_0 = const()[name = string("logits_chunk_3_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_3 = conv(dilations = logits_chunk_3_dilations_0, groups = logits_chunk_3_groups_0, pad = logits_chunk_3_pad_0, pad_type = logits_chunk_3_pad_type_0, strides = logits_chunk_3_strides_0, weight = logits_chunk_3_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_3")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_3_mul = mul(x = logits_chunk_3, y = temp_inverse)[name = string("logits_chunk_3_mul")];
+            tensor<int32, [1]> logits_chunk_3_max_axes_0 = const()[name = string("logits_chunk_3_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_3_max_keep_dims_0 = const()[name = string("logits_chunk_3_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_3_max = reduce_max(axes = logits_chunk_3_max_axes_0, keep_dims = logits_chunk_3_max_keep_dims_0, x = logits_chunk_3_mul)[name = string("logits_chunk_3_max")];
+            int32 logits_chunk_3_argmax_axis_0 = const()[name = string("logits_chunk_3_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_3_argmax_keep_dims_0 = const()[name = string("logits_chunk_3_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_3_argmax_output_dtype_0 = const()[name = string("logits_chunk_3_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> logits_chunk_3_argmax = reduce_argmax(axis = logits_chunk_3_argmax_axis_0, keep_dims = logits_chunk_3_argmax_keep_dims_0, output_dtype = logits_chunk_3_argmax_output_dtype_0, x = logits_chunk_3_mul)[name = string("logits_chunk_3_argmax")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_3_sub = sub(x = logits_chunk_3_mul, y = logits_chunk_3_max)[name = string("logits_chunk_3_sub")];
+            tensor<int32, [1]> logits_chunk_3_lse_sub_axes_0 = const()[name = string("logits_chunk_3_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_3_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_3_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_3_lse_sub = reduce_log_sum_exp(axes = logits_chunk_3_lse_sub_axes_0, keep_dims = logits_chunk_3_lse_sub_keep_dims_0, x = logits_chunk_3_sub)[name = string("logits_chunk_3_lse_sub")];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_3_lse = add(x = logits_chunk_3_lse_sub, y = logits_chunk_3_max)[name = string("logits_chunk_3_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_4_weight_0 = const()[name = string("logits_chunk_4_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33558912)))];
+            tensor<int32, [2]> logits_chunk_4_strides_0 = const()[name = string("logits_chunk_4_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_4_pad_type_0 = const()[name = string("logits_chunk_4_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_4_pad_0 = const()[name = string("logits_chunk_4_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_4_dilations_0 = const()[name = string("logits_chunk_4_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_4_groups_0 = const()[name = string("logits_chunk_4_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_4 = conv(dilations = logits_chunk_4_dilations_0, groups = logits_chunk_4_groups_0, pad = logits_chunk_4_pad_0, pad_type = logits_chunk_4_pad_type_0, strides = logits_chunk_4_strides_0, weight = logits_chunk_4_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_4")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_4_mul = mul(x = logits_chunk_4, y = temp_inverse)[name = string("logits_chunk_4_mul")];
+            tensor<int32, [1]> logits_chunk_4_max_axes_0 = const()[name = string("logits_chunk_4_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_4_max_keep_dims_0 = const()[name = string("logits_chunk_4_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_4_max = reduce_max(axes = logits_chunk_4_max_axes_0, keep_dims = logits_chunk_4_max_keep_dims_0, x = logits_chunk_4_mul)[name = string("logits_chunk_4_max")];
+            int32 logits_chunk_4_argmax_axis_0 = const()[name = string("logits_chunk_4_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_4_argmax_keep_dims_0 = const()[name = string("logits_chunk_4_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_4_argmax_output_dtype_0 = const()[name = string("logits_chunk_4_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> logits_chunk_4_argmax = reduce_argmax(axis = logits_chunk_4_argmax_axis_0, keep_dims = logits_chunk_4_argmax_keep_dims_0, output_dtype = logits_chunk_4_argmax_output_dtype_0, x = logits_chunk_4_mul)[name = string("logits_chunk_4_argmax")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_4_sub = sub(x = logits_chunk_4_mul, y = logits_chunk_4_max)[name = string("logits_chunk_4_sub")];
+            tensor<int32, [1]> logits_chunk_4_lse_sub_axes_0 = const()[name = string("logits_chunk_4_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_4_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_4_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_4_lse_sub = reduce_log_sum_exp(axes = logits_chunk_4_lse_sub_axes_0, keep_dims = logits_chunk_4_lse_sub_keep_dims_0, x = logits_chunk_4_sub)[name = string("logits_chunk_4_lse_sub")];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_4_lse = add(x = logits_chunk_4_lse_sub, y = logits_chunk_4_max)[name = string("logits_chunk_4_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_5_weight_0 = const()[name = string("logits_chunk_5_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41947584)))];
+            tensor<int32, [2]> logits_chunk_5_strides_0 = const()[name = string("logits_chunk_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_5_pad_type_0 = const()[name = string("logits_chunk_5_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_5_pad_0 = const()[name = string("logits_chunk_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_5_dilations_0 = const()[name = string("logits_chunk_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_5_groups_0 = const()[name = string("logits_chunk_5_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_5 = conv(dilations = logits_chunk_5_dilations_0, groups = logits_chunk_5_groups_0, pad = logits_chunk_5_pad_0, pad_type = logits_chunk_5_pad_type_0, strides = logits_chunk_5_strides_0, weight = logits_chunk_5_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_5")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_5_mul = mul(x = logits_chunk_5, y = temp_inverse)[name = string("logits_chunk_5_mul")];
+            tensor<int32, [1]> logits_chunk_5_max_axes_0 = const()[name = string("logits_chunk_5_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_5_max_keep_dims_0 = const()[name = string("logits_chunk_5_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_5_max = reduce_max(axes = logits_chunk_5_max_axes_0, keep_dims = logits_chunk_5_max_keep_dims_0, x = logits_chunk_5_mul)[name = string("logits_chunk_5_max")];
+            int32 logits_chunk_5_argmax_axis_0 = const()[name = string("logits_chunk_5_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_5_argmax_keep_dims_0 = const()[name = string("logits_chunk_5_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_5_argmax_output_dtype_0 = const()[name = string("logits_chunk_5_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> logits_chunk_5_argmax = reduce_argmax(axis = logits_chunk_5_argmax_axis_0, keep_dims = logits_chunk_5_argmax_keep_dims_0, output_dtype = logits_chunk_5_argmax_output_dtype_0, x = logits_chunk_5_mul)[name = string("logits_chunk_5_argmax")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_5_sub = sub(x = logits_chunk_5_mul, y = logits_chunk_5_max)[name = string("logits_chunk_5_sub")];
+            tensor<int32, [1]> logits_chunk_5_lse_sub_axes_0 = const()[name = string("logits_chunk_5_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_5_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_5_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_5_lse_sub = reduce_log_sum_exp(axes = logits_chunk_5_lse_sub_axes_0, keep_dims = logits_chunk_5_lse_sub_keep_dims_0, x = logits_chunk_5_sub)[name = string("logits_chunk_5_lse_sub")];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_5_lse = add(x = logits_chunk_5_lse_sub, y = logits_chunk_5_max)[name = string("logits_chunk_5_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_6_weight_0 = const()[name = string("logits_chunk_6_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50336256)))];
+            tensor<int32, [2]> logits_chunk_6_strides_0 = const()[name = string("logits_chunk_6_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_6_pad_type_0 = const()[name = string("logits_chunk_6_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_6_pad_0 = const()[name = string("logits_chunk_6_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_6_dilations_0 = const()[name = string("logits_chunk_6_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_6_groups_0 = const()[name = string("logits_chunk_6_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_6 = conv(dilations = logits_chunk_6_dilations_0, groups = logits_chunk_6_groups_0, pad = logits_chunk_6_pad_0, pad_type = logits_chunk_6_pad_type_0, strides = logits_chunk_6_strides_0, weight = logits_chunk_6_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_6")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_6_mul = mul(x = logits_chunk_6, y = temp_inverse)[name = string("logits_chunk_6_mul")];
+            tensor<int32, [1]> logits_chunk_6_max_axes_0 = const()[name = string("logits_chunk_6_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_6_max_keep_dims_0 = const()[name = string("logits_chunk_6_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_6_max = reduce_max(axes = logits_chunk_6_max_axes_0, keep_dims = logits_chunk_6_max_keep_dims_0, x = logits_chunk_6_mul)[name = string("logits_chunk_6_max")];
+            int32 logits_chunk_6_argmax_axis_0 = const()[name = string("logits_chunk_6_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_6_argmax_keep_dims_0 = const()[name = string("logits_chunk_6_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_6_argmax_output_dtype_0 = const()[name = string("logits_chunk_6_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> logits_chunk_6_argmax = reduce_argmax(axis = logits_chunk_6_argmax_axis_0, keep_dims = logits_chunk_6_argmax_keep_dims_0, output_dtype = logits_chunk_6_argmax_output_dtype_0, x = logits_chunk_6_mul)[name = string("logits_chunk_6_argmax")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_6_sub = sub(x = logits_chunk_6_mul, y = logits_chunk_6_max)[name = string("logits_chunk_6_sub")];
+            tensor<int32, [1]> logits_chunk_6_lse_sub_axes_0 = const()[name = string("logits_chunk_6_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_6_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_6_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_6_lse_sub = reduce_log_sum_exp(axes = logits_chunk_6_lse_sub_axes_0, keep_dims = logits_chunk_6_lse_sub_keep_dims_0, x = logits_chunk_6_sub)[name = string("logits_chunk_6_lse_sub")];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_6_lse = add(x = logits_chunk_6_lse_sub, y = logits_chunk_6_max)[name = string("logits_chunk_6_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_7_weight_0 = const()[name = string("logits_chunk_7_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58724928)))];
+            tensor<int32, [2]> logits_chunk_7_strides_0 = const()[name = string("logits_chunk_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_7_pad_type_0 = const()[name = string("logits_chunk_7_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_7_pad_0 = const()[name = string("logits_chunk_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_7_dilations_0 = const()[name = string("logits_chunk_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_7_groups_0 = const()[name = string("logits_chunk_7_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_7 = conv(dilations = logits_chunk_7_dilations_0, groups = logits_chunk_7_groups_0, pad = logits_chunk_7_pad_0, pad_type = logits_chunk_7_pad_type_0, strides = logits_chunk_7_strides_0, weight = logits_chunk_7_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_7")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_7_mul = mul(x = logits_chunk_7, y = temp_inverse)[name = string("logits_chunk_7_mul")];
+            tensor<int32, [1]> logits_chunk_7_max_axes_0 = const()[name = string("logits_chunk_7_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_7_max_keep_dims_0 = const()[name = string("logits_chunk_7_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_7_max = reduce_max(axes = logits_chunk_7_max_axes_0, keep_dims = logits_chunk_7_max_keep_dims_0, x = logits_chunk_7_mul)[name = string("logits_chunk_7_max")];
+            int32 logits_chunk_7_argmax_axis_0 = const()[name = string("logits_chunk_7_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_7_argmax_keep_dims_0 = const()[name = string("logits_chunk_7_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_7_argmax_output_dtype_0 = const()[name = string("logits_chunk_7_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> logits_chunk_7_argmax = reduce_argmax(axis = logits_chunk_7_argmax_axis_0, keep_dims = logits_chunk_7_argmax_keep_dims_0, output_dtype = logits_chunk_7_argmax_output_dtype_0, x = logits_chunk_7_mul)[name = string("logits_chunk_7_argmax")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_7_sub = sub(x = logits_chunk_7_mul, y = logits_chunk_7_max)[name = string("logits_chunk_7_sub")];
+            tensor<int32, [1]> logits_chunk_7_lse_sub_axes_0 = const()[name = string("logits_chunk_7_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_7_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_7_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_7_lse_sub = reduce_log_sum_exp(axes = logits_chunk_7_lse_sub_axes_0, keep_dims = logits_chunk_7_lse_sub_keep_dims_0, x = logits_chunk_7_sub)[name = string("logits_chunk_7_lse_sub")];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_7_lse = add(x = logits_chunk_7_lse_sub, y = logits_chunk_7_max)[name = string("logits_chunk_7_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_8_weight_0 = const()[name = string("logits_chunk_8_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67113600)))];
+            tensor<int32, [2]> logits_chunk_8_strides_0 = const()[name = string("logits_chunk_8_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_8_pad_type_0 = const()[name = string("logits_chunk_8_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_8_pad_0 = const()[name = string("logits_chunk_8_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_8_dilations_0 = const()[name = string("logits_chunk_8_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_8_groups_0 = const()[name = string("logits_chunk_8_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_8 = conv(dilations = logits_chunk_8_dilations_0, groups = logits_chunk_8_groups_0, pad = logits_chunk_8_pad_0, pad_type = logits_chunk_8_pad_type_0, strides = logits_chunk_8_strides_0, weight = logits_chunk_8_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_8")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_8_mul = mul(x = logits_chunk_8, y = temp_inverse)[name = string("logits_chunk_8_mul")];
+            tensor<int32, [1]> logits_chunk_8_max_axes_0 = const()[name = string("logits_chunk_8_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_8_max_keep_dims_0 = const()[name = string("logits_chunk_8_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_8_max = reduce_max(axes = logits_chunk_8_max_axes_0, keep_dims = logits_chunk_8_max_keep_dims_0, x = logits_chunk_8_mul)[name = string("logits_chunk_8_max")];
+            int32 logits_chunk_8_argmax_axis_0 = const()[name = string("logits_chunk_8_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_8_argmax_keep_dims_0 = const()[name = string("logits_chunk_8_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_8_argmax_output_dtype_0 = const()[name = string("logits_chunk_8_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> logits_chunk_8_argmax = reduce_argmax(axis = logits_chunk_8_argmax_axis_0, keep_dims = logits_chunk_8_argmax_keep_dims_0, output_dtype = logits_chunk_8_argmax_output_dtype_0, x = logits_chunk_8_mul)[name = string("logits_chunk_8_argmax")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_8_sub = sub(x = logits_chunk_8_mul, y = logits_chunk_8_max)[name = string("logits_chunk_8_sub")];
+            tensor<int32, [1]> logits_chunk_8_lse_sub_axes_0 = const()[name = string("logits_chunk_8_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_8_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_8_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_8_lse_sub = reduce_log_sum_exp(axes = logits_chunk_8_lse_sub_axes_0, keep_dims = logits_chunk_8_lse_sub_keep_dims_0, x = logits_chunk_8_sub)[name = string("logits_chunk_8_lse_sub")];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_8_lse = add(x = logits_chunk_8_lse_sub, y = logits_chunk_8_max)[name = string("logits_chunk_8_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_9_weight_0 = const()[name = string("logits_chunk_9_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75502272)))];
+            tensor<int32, [2]> logits_chunk_9_strides_0 = const()[name = string("logits_chunk_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_9_pad_type_0 = const()[name = string("logits_chunk_9_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_9_pad_0 = const()[name = string("logits_chunk_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_9_dilations_0 = const()[name = string("logits_chunk_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_9_groups_0 = const()[name = string("logits_chunk_9_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_9 = conv(dilations = logits_chunk_9_dilations_0, groups = logits_chunk_9_groups_0, pad = logits_chunk_9_pad_0, pad_type = logits_chunk_9_pad_type_0, strides = logits_chunk_9_strides_0, weight = logits_chunk_9_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_9")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_9_mul = mul(x = logits_chunk_9, y = temp_inverse)[name = string("logits_chunk_9_mul")];
+            tensor<int32, [1]> logits_chunk_9_max_axes_0 = const()[name = string("logits_chunk_9_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_9_max_keep_dims_0 = const()[name = string("logits_chunk_9_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_9_max = reduce_max(axes = logits_chunk_9_max_axes_0, keep_dims = logits_chunk_9_max_keep_dims_0, x = logits_chunk_9_mul)[name = string("logits_chunk_9_max")];
+            int32 logits_chunk_9_argmax_axis_0 = const()[name = string("logits_chunk_9_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_9_argmax_keep_dims_0 = const()[name = string("logits_chunk_9_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_9_argmax_output_dtype_0 = const()[name = string("logits_chunk_9_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> logits_chunk_9_argmax = reduce_argmax(axis = logits_chunk_9_argmax_axis_0, keep_dims = logits_chunk_9_argmax_keep_dims_0, output_dtype = logits_chunk_9_argmax_output_dtype_0, x = logits_chunk_9_mul)[name = string("logits_chunk_9_argmax")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_9_sub = sub(x = logits_chunk_9_mul, y = logits_chunk_9_max)[name = string("logits_chunk_9_sub")];
+            tensor<int32, [1]> logits_chunk_9_lse_sub_axes_0 = const()[name = string("logits_chunk_9_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_9_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_9_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_9_lse_sub = reduce_log_sum_exp(axes = logits_chunk_9_lse_sub_axes_0, keep_dims = logits_chunk_9_lse_sub_keep_dims_0, x = logits_chunk_9_sub)[name = string("logits_chunk_9_lse_sub")];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_9_lse = add(x = logits_chunk_9_lse_sub, y = logits_chunk_9_max)[name = string("logits_chunk_9_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_10_weight_0 = const()[name = string("logits_chunk_10_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83890944)))];
+            tensor<int32, [2]> logits_chunk_10_strides_0 = const()[name = string("logits_chunk_10_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_10_pad_type_0 = const()[name = string("logits_chunk_10_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_10_pad_0 = const()[name = string("logits_chunk_10_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_10_dilations_0 = const()[name = string("logits_chunk_10_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_10_groups_0 = const()[name = string("logits_chunk_10_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_10 = conv(dilations = logits_chunk_10_dilations_0, groups = logits_chunk_10_groups_0, pad = logits_chunk_10_pad_0, pad_type = logits_chunk_10_pad_type_0, strides = logits_chunk_10_strides_0, weight = logits_chunk_10_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_10")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_10_mul = mul(x = logits_chunk_10, y = temp_inverse)[name = string("logits_chunk_10_mul")];
+            tensor<int32, [1]> logits_chunk_10_max_axes_0 = const()[name = string("logits_chunk_10_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_10_max_keep_dims_0 = const()[name = string("logits_chunk_10_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_10_max = reduce_max(axes = logits_chunk_10_max_axes_0, keep_dims = logits_chunk_10_max_keep_dims_0, x = logits_chunk_10_mul)[name = string("logits_chunk_10_max")];
+            int32 logits_chunk_10_argmax_axis_0 = const()[name = string("logits_chunk_10_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_10_argmax_keep_dims_0 = const()[name = string("logits_chunk_10_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_10_argmax_output_dtype_0 = const()[name = string("logits_chunk_10_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> logits_chunk_10_argmax = reduce_argmax(axis = logits_chunk_10_argmax_axis_0, keep_dims = logits_chunk_10_argmax_keep_dims_0, output_dtype = logits_chunk_10_argmax_output_dtype_0, x = logits_chunk_10_mul)[name = string("logits_chunk_10_argmax")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_10_sub = sub(x = logits_chunk_10_mul, y = logits_chunk_10_max)[name = string("logits_chunk_10_sub")];
+            tensor<int32, [1]> logits_chunk_10_lse_sub_axes_0 = const()[name = string("logits_chunk_10_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_10_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_10_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_10_lse_sub = reduce_log_sum_exp(axes = logits_chunk_10_lse_sub_axes_0, keep_dims = logits_chunk_10_lse_sub_keep_dims_0, x = logits_chunk_10_sub)[name = string("logits_chunk_10_lse_sub")];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_10_lse = add(x = logits_chunk_10_lse_sub, y = logits_chunk_10_max)[name = string("logits_chunk_10_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_11_weight_0 = const()[name = string("logits_chunk_11_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92279616)))];
+            tensor<int32, [2]> logits_chunk_11_strides_0 = const()[name = string("logits_chunk_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_11_pad_type_0 = const()[name = string("logits_chunk_11_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_11_pad_0 = const()[name = string("logits_chunk_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_11_dilations_0 = const()[name = string("logits_chunk_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_11_groups_0 = const()[name = string("logits_chunk_11_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_11 = conv(dilations = logits_chunk_11_dilations_0, groups = logits_chunk_11_groups_0, pad = logits_chunk_11_pad_0, pad_type = logits_chunk_11_pad_type_0, strides = logits_chunk_11_strides_0, weight = logits_chunk_11_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_11")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_11_mul = mul(x = logits_chunk_11, y = temp_inverse)[name = string("logits_chunk_11_mul")];
+            tensor<int32, [1]> logits_chunk_11_max_axes_0 = const()[name = string("logits_chunk_11_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_11_max_keep_dims_0 = const()[name = string("logits_chunk_11_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_11_max = reduce_max(axes = logits_chunk_11_max_axes_0, keep_dims = logits_chunk_11_max_keep_dims_0, x = logits_chunk_11_mul)[name = string("logits_chunk_11_max")];
+            int32 logits_chunk_11_argmax_axis_0 = const()[name = string("logits_chunk_11_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_11_argmax_keep_dims_0 = const()[name = string("logits_chunk_11_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_11_argmax_output_dtype_0 = const()[name = string("logits_chunk_11_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> logits_chunk_11_argmax = reduce_argmax(axis = logits_chunk_11_argmax_axis_0, keep_dims = logits_chunk_11_argmax_keep_dims_0, output_dtype = logits_chunk_11_argmax_output_dtype_0, x = logits_chunk_11_mul)[name = string("logits_chunk_11_argmax")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_11_sub = sub(x = logits_chunk_11_mul, y = logits_chunk_11_max)[name = string("logits_chunk_11_sub")];
+            tensor<int32, [1]> logits_chunk_11_lse_sub_axes_0 = const()[name = string("logits_chunk_11_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_11_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_11_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_11_lse_sub = reduce_log_sum_exp(axes = logits_chunk_11_lse_sub_axes_0, keep_dims = logits_chunk_11_lse_sub_keep_dims_0, x = logits_chunk_11_sub)[name = string("logits_chunk_11_lse_sub")];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_11_lse = add(x = logits_chunk_11_lse_sub, y = logits_chunk_11_max)[name = string("logits_chunk_11_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_12_weight_0 = const()[name = string("logits_chunk_12_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100668288)))];
+            tensor<int32, [2]> logits_chunk_12_strides_0 = const()[name = string("logits_chunk_12_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_12_pad_type_0 = const()[name = string("logits_chunk_12_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_12_pad_0 = const()[name = string("logits_chunk_12_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_12_dilations_0 = const()[name = string("logits_chunk_12_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_12_groups_0 = const()[name = string("logits_chunk_12_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_12 = conv(dilations = logits_chunk_12_dilations_0, groups = logits_chunk_12_groups_0, pad = logits_chunk_12_pad_0, pad_type = logits_chunk_12_pad_type_0, strides = logits_chunk_12_strides_0, weight = logits_chunk_12_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_12")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_12_mul = mul(x = logits_chunk_12, y = temp_inverse)[name = string("logits_chunk_12_mul")];
+            tensor<int32, [1]> logits_chunk_12_max_axes_0 = const()[name = string("logits_chunk_12_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_12_max_keep_dims_0 = const()[name = string("logits_chunk_12_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_12_max = reduce_max(axes = logits_chunk_12_max_axes_0, keep_dims = logits_chunk_12_max_keep_dims_0, x = logits_chunk_12_mul)[name = string("logits_chunk_12_max")];
+            int32 logits_chunk_12_argmax_axis_0 = const()[name = string("logits_chunk_12_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_12_argmax_keep_dims_0 = const()[name = string("logits_chunk_12_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_12_argmax_output_dtype_0 = const()[name = string("logits_chunk_12_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> logits_chunk_12_argmax = reduce_argmax(axis = logits_chunk_12_argmax_axis_0, keep_dims = logits_chunk_12_argmax_keep_dims_0, output_dtype = logits_chunk_12_argmax_output_dtype_0, x = logits_chunk_12_mul)[name = string("logits_chunk_12_argmax")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_12_sub = sub(x = logits_chunk_12_mul, y = logits_chunk_12_max)[name = string("logits_chunk_12_sub")];
+            tensor<int32, [1]> logits_chunk_12_lse_sub_axes_0 = const()[name = string("logits_chunk_12_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_12_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_12_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_12_lse_sub = reduce_log_sum_exp(axes = logits_chunk_12_lse_sub_axes_0, keep_dims = logits_chunk_12_lse_sub_keep_dims_0, x = logits_chunk_12_sub)[name = string("logits_chunk_12_lse_sub")];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_12_lse = add(x = logits_chunk_12_lse_sub, y = logits_chunk_12_max)[name = string("logits_chunk_12_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_13_weight_0 = const()[name = string("logits_chunk_13_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109056960)))];
+            tensor<int32, [2]> logits_chunk_13_strides_0 = const()[name = string("logits_chunk_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_13_pad_type_0 = const()[name = string("logits_chunk_13_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_13_pad_0 = const()[name = string("logits_chunk_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_13_dilations_0 = const()[name = string("logits_chunk_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_13_groups_0 = const()[name = string("logits_chunk_13_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_13 = conv(dilations = logits_chunk_13_dilations_0, groups = logits_chunk_13_groups_0, pad = logits_chunk_13_pad_0, pad_type = logits_chunk_13_pad_type_0, strides = logits_chunk_13_strides_0, weight = logits_chunk_13_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_13")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_13_mul = mul(x = logits_chunk_13, y = temp_inverse)[name = string("logits_chunk_13_mul")];
+            tensor<int32, [1]> logits_chunk_13_max_axes_0 = const()[name = string("logits_chunk_13_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_13_max_keep_dims_0 = const()[name = string("logits_chunk_13_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_13_max = reduce_max(axes = logits_chunk_13_max_axes_0, keep_dims = logits_chunk_13_max_keep_dims_0, x = logits_chunk_13_mul)[name = string("logits_chunk_13_max")];
+            int32 logits_chunk_13_argmax_axis_0 = const()[name = string("logits_chunk_13_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_13_argmax_keep_dims_0 = const()[name = string("logits_chunk_13_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_13_argmax_output_dtype_0 = const()[name = string("logits_chunk_13_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> logits_chunk_13_argmax = reduce_argmax(axis = logits_chunk_13_argmax_axis_0, keep_dims = logits_chunk_13_argmax_keep_dims_0, output_dtype = logits_chunk_13_argmax_output_dtype_0, x = logits_chunk_13_mul)[name = string("logits_chunk_13_argmax")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_13_sub = sub(x = logits_chunk_13_mul, y = logits_chunk_13_max)[name = string("logits_chunk_13_sub")];
+            tensor<int32, [1]> logits_chunk_13_lse_sub_axes_0 = const()[name = string("logits_chunk_13_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_13_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_13_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_13_lse_sub = reduce_log_sum_exp(axes = logits_chunk_13_lse_sub_axes_0, keep_dims = logits_chunk_13_lse_sub_keep_dims_0, x = logits_chunk_13_sub)[name = string("logits_chunk_13_lse_sub")];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_13_lse = add(x = logits_chunk_13_lse_sub, y = logits_chunk_13_max)[name = string("logits_chunk_13_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_14_weight_0 = const()[name = string("logits_chunk_14_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117445632)))];
+            tensor<int32, [2]> logits_chunk_14_strides_0 = const()[name = string("logits_chunk_14_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_14_pad_type_0 = const()[name = string("logits_chunk_14_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_14_pad_0 = const()[name = string("logits_chunk_14_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_14_dilations_0 = const()[name = string("logits_chunk_14_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_14_groups_0 = const()[name = string("logits_chunk_14_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_14 = conv(dilations = logits_chunk_14_dilations_0, groups = logits_chunk_14_groups_0, pad = logits_chunk_14_pad_0, pad_type = logits_chunk_14_pad_type_0, strides = logits_chunk_14_strides_0, weight = logits_chunk_14_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_14")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_14_mul = mul(x = logits_chunk_14, y = temp_inverse)[name = string("logits_chunk_14_mul")];
+            tensor<int32, [1]> logits_chunk_14_max_axes_0 = const()[name = string("logits_chunk_14_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_14_max_keep_dims_0 = const()[name = string("logits_chunk_14_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_14_max = reduce_max(axes = logits_chunk_14_max_axes_0, keep_dims = logits_chunk_14_max_keep_dims_0, x = logits_chunk_14_mul)[name = string("logits_chunk_14_max")];
+            int32 logits_chunk_14_argmax_axis_0 = const()[name = string("logits_chunk_14_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_14_argmax_keep_dims_0 = const()[name = string("logits_chunk_14_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_14_argmax_output_dtype_0 = const()[name = string("logits_chunk_14_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> logits_chunk_14_argmax = reduce_argmax(axis = logits_chunk_14_argmax_axis_0, keep_dims = logits_chunk_14_argmax_keep_dims_0, output_dtype = logits_chunk_14_argmax_output_dtype_0, x = logits_chunk_14_mul)[name = string("logits_chunk_14_argmax")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_14_sub = sub(x = logits_chunk_14_mul, y = logits_chunk_14_max)[name = string("logits_chunk_14_sub")];
+            tensor<int32, [1]> logits_chunk_14_lse_sub_axes_0 = const()[name = string("logits_chunk_14_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_14_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_14_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_14_lse_sub = reduce_log_sum_exp(axes = logits_chunk_14_lse_sub_axes_0, keep_dims = logits_chunk_14_lse_sub_keep_dims_0, x = logits_chunk_14_sub)[name = string("logits_chunk_14_lse_sub")];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_14_lse = add(x = logits_chunk_14_lse_sub, y = logits_chunk_14_max)[name = string("logits_chunk_14_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_15_weight_0 = const()[name = string("logits_chunk_15_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125834304)))];
+            tensor<int32, [2]> logits_chunk_15_strides_0 = const()[name = string("logits_chunk_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_15_pad_type_0 = const()[name = string("logits_chunk_15_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_15_pad_0 = const()[name = string("logits_chunk_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_15_dilations_0 = const()[name = string("logits_chunk_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_15_groups_0 = const()[name = string("logits_chunk_15_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_15 = conv(dilations = logits_chunk_15_dilations_0, groups = logits_chunk_15_groups_0, pad = logits_chunk_15_pad_0, pad_type = logits_chunk_15_pad_type_0, strides = logits_chunk_15_strides_0, weight = logits_chunk_15_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_15")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_15_mul = mul(x = logits_chunk_15, y = temp_inverse)[name = string("logits_chunk_15_mul")];
+            tensor<int32, [1]> logits_chunk_15_max_axes_0 = const()[name = string("logits_chunk_15_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_15_max_keep_dims_0 = const()[name = string("logits_chunk_15_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_15_max = reduce_max(axes = logits_chunk_15_max_axes_0, keep_dims = logits_chunk_15_max_keep_dims_0, x = logits_chunk_15_mul)[name = string("logits_chunk_15_max")];
+            int32 logits_chunk_15_argmax_axis_0 = const()[name = string("logits_chunk_15_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_15_argmax_keep_dims_0 = const()[name = string("logits_chunk_15_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_15_argmax_output_dtype_0 = const()[name = string("logits_chunk_15_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 64]> logits_chunk_15_argmax = reduce_argmax(axis = logits_chunk_15_argmax_axis_0, keep_dims = logits_chunk_15_argmax_keep_dims_0, output_dtype = logits_chunk_15_argmax_output_dtype_0, x = logits_chunk_15_mul)[name = string("logits_chunk_15_argmax")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_15_sub = sub(x = logits_chunk_15_mul, y = logits_chunk_15_max)[name = string("logits_chunk_15_sub")];
+            tensor<int32, [1]> logits_chunk_15_lse_sub_axes_0 = const()[name = string("logits_chunk_15_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_15_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_15_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_15_lse_sub = reduce_log_sum_exp(axes = logits_chunk_15_lse_sub_axes_0, keep_dims = logits_chunk_15_lse_sub_keep_dims_0, x = logits_chunk_15_sub)[name = string("logits_chunk_15_lse_sub")];
+            tensor<fp16, [1, 1, 1, 64]> logits_chunk_15_lse = add(x = logits_chunk_15_lse_sub, y = logits_chunk_15_max)[name = string("logits_chunk_15_lse")];
+            int32 logits_lses_axis_0 = const()[name = string("logits_lses_axis_0"), val = int32(1)];
+            bool logits_lses_interleave_0 = const()[name = string("logits_lses_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 64]> logits_lses = concat(axis = logits_lses_axis_0, interleave = logits_lses_interleave_0, values = (logits_chunk_0_lse, logits_chunk_1_lse, logits_chunk_2_lse, logits_chunk_3_lse, logits_chunk_4_lse, logits_chunk_5_lse, logits_chunk_6_lse, logits_chunk_7_lse, logits_chunk_8_lse, logits_chunk_9_lse, logits_chunk_10_lse, logits_chunk_11_lse, logits_chunk_12_lse, logits_chunk_13_lse, logits_chunk_14_lse, logits_chunk_15_lse))[name = string("logits_lses")];
+            tensor<int32, [1]> logits_lses_max_axes_0 = const()[name = string("logits_lses_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_lses_max_keep_dims_0 = const()[name = string("logits_lses_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_lses_max = reduce_max(axes = logits_lses_max_axes_0, keep_dims = logits_lses_max_keep_dims_0, x = logits_lses)[name = string("logits_lses_max")];
+            tensor<fp16, [1, 16, 1, 64]> logits_lses_sub = sub(x = logits_lses, y = logits_lses_max)[name = string("logits_lses_sub")];
+            tensor<int32, [1]> logits_lses_logsumexp_axes_0 = const()[name = string("logits_lses_logsumexp_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_lses_logsumexp_keep_dims_0 = const()[name = string("logits_lses_logsumexp_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_lses_logsumexp = reduce_log_sum_exp(axes = logits_lses_logsumexp_axes_0, keep_dims = logits_lses_logsumexp_keep_dims_0, x = logits_lses_sub)[name = string("logits_lses_logsumexp")];
+            tensor<fp16, [1, 1, 1, 64]> logits_lse = add(x = logits_lses_logsumexp, y = logits_lses_max)[name = string("logits_lse")];
+            int32 logits_max_logits_chunks_axis_0 = const()[name = string("logits_max_logits_chunks_axis_0"), val = int32(1)];
+            bool logits_max_logits_chunks_interleave_0 = const()[name = string("logits_max_logits_chunks_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 64]> logits_max_logits_chunks = concat(axis = logits_max_logits_chunks_axis_0, interleave = logits_max_logits_chunks_interleave_0, values = (logits_chunk_0_max, logits_chunk_1_max, logits_chunk_2_max, logits_chunk_3_max, logits_chunk_4_max, logits_chunk_5_max, logits_chunk_6_max, logits_chunk_7_max, logits_chunk_8_max, logits_chunk_9_max, logits_chunk_10_max, logits_chunk_11_max, logits_chunk_12_max, logits_chunk_13_max, logits_chunk_14_max, logits_chunk_15_max))[name = string("logits_max_logits_chunks")];
+            tensor<int32, [1]> logits_max_logit_axes_0 = const()[name = string("logits_max_logit_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_max_logit_keep_dims_0 = const()[name = string("logits_max_logit_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 64]> logits_max_logit = reduce_max(axes = logits_max_logit_axes_0, keep_dims = logits_max_logit_keep_dims_0, x = logits_max_logits_chunks)[name = string("logits_max_logit")];
+            tensor<fp16, [1, 1, 1, 64]> logits_max_logit_sub = sub(x = logits_max_logit, y = logits_lse)[name = string("logits_max_logit_sub")];
+            tensor<fp16, [1, 1, 1, 64]> max_prob = exp(x = logits_max_logit_sub)[name = string("max_prob")];
+            tensor<fp16, [1, 1, 1, 64]> min_p_thresh = mul(x = max_prob, y = p)[name = string("min_p_thresh")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_0_sub_1 = sub(x = logits_chunk_0_mul, y = logits_lse)[name = string("logits_chunk_0_sub_1")];
+            tensor<fp16, [1, 2048, 1, 64]> probs_chunk_0 = exp(x = logits_chunk_0_sub_1)[name = string("probs_chunk_0")];
+            tensor<bool, [1, 2048, 1, 64]> mask_probs_chunk_0 = greater_equal(x = probs_chunk_0, y = min_p_thresh)[name = string("mask_probs_chunk_0")];
+            string mask_chunk_0_fp16_dtype_0 = const()[name = string("mask_chunk_0_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 64]> mask_chunk_0_fp16 = cast(dtype = mask_chunk_0_fp16_dtype_0, x = mask_probs_chunk_0)[name = string("cast_237")];
+            tensor<fp16, [1, 2048, 1, 64]> masked_probs_chunk_0 = select(a = probs_chunk_0, b = mask_chunk_0_fp16, cond = mask_probs_chunk_0)[name = string("masked_probs_chunk_0")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_1_sub_1 = sub(x = logits_chunk_1_mul, y = logits_lse)[name = string("logits_chunk_1_sub_1")];
+            tensor<fp16, [1, 2048, 1, 64]> probs_chunk_1 = exp(x = logits_chunk_1_sub_1)[name = string("probs_chunk_1")];
+            tensor<bool, [1, 2048, 1, 64]> mask_probs_chunk_1 = greater_equal(x = probs_chunk_1, y = min_p_thresh)[name = string("mask_probs_chunk_1")];
+            string mask_chunk_1_fp16_dtype_0 = const()[name = string("mask_chunk_1_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 64]> mask_chunk_1_fp16 = cast(dtype = mask_chunk_1_fp16_dtype_0, x = mask_probs_chunk_1)[name = string("cast_236")];
+            tensor<fp16, [1, 2048, 1, 64]> masked_probs_chunk_1 = select(a = probs_chunk_1, b = mask_chunk_1_fp16, cond = mask_probs_chunk_1)[name = string("masked_probs_chunk_1")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_2_sub_1 = sub(x = logits_chunk_2_mul, y = logits_lse)[name = string("logits_chunk_2_sub_1")];
+            tensor<fp16, [1, 2048, 1, 64]> probs_chunk_2 = exp(x = logits_chunk_2_sub_1)[name = string("probs_chunk_2")];
+            tensor<bool, [1, 2048, 1, 64]> mask_probs_chunk_2 = greater_equal(x = probs_chunk_2, y = min_p_thresh)[name = string("mask_probs_chunk_2")];
+            string mask_chunk_2_fp16_dtype_0 = const()[name = string("mask_chunk_2_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 64]> mask_chunk_2_fp16 = cast(dtype = mask_chunk_2_fp16_dtype_0, x = mask_probs_chunk_2)[name = string("cast_235")];
+            tensor<fp16, [1, 2048, 1, 64]> masked_probs_chunk_2 = select(a = probs_chunk_2, b = mask_chunk_2_fp16, cond = mask_probs_chunk_2)[name = string("masked_probs_chunk_2")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_3_sub_1 = sub(x = logits_chunk_3_mul, y = logits_lse)[name = string("logits_chunk_3_sub_1")];
+            tensor<fp16, [1, 2048, 1, 64]> probs_chunk_3 = exp(x = logits_chunk_3_sub_1)[name = string("probs_chunk_3")];
+            tensor<bool, [1, 2048, 1, 64]> mask_probs_chunk_3 = greater_equal(x = probs_chunk_3, y = min_p_thresh)[name = string("mask_probs_chunk_3")];
+            string mask_chunk_3_fp16_dtype_0 = const()[name = string("mask_chunk_3_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 64]> mask_chunk_3_fp16 = cast(dtype = mask_chunk_3_fp16_dtype_0, x = mask_probs_chunk_3)[name = string("cast_234")];
+            tensor<fp16, [1, 2048, 1, 64]> masked_probs_chunk_3 = select(a = probs_chunk_3, b = mask_chunk_3_fp16, cond = mask_probs_chunk_3)[name = string("masked_probs_chunk_3")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_4_sub_1 = sub(x = logits_chunk_4_mul, y = logits_lse)[name = string("logits_chunk_4_sub_1")];
+            tensor<fp16, [1, 2048, 1, 64]> probs_chunk_4 = exp(x = logits_chunk_4_sub_1)[name = string("probs_chunk_4")];
+            tensor<bool, [1, 2048, 1, 64]> mask_probs_chunk_4 = greater_equal(x = probs_chunk_4, y = min_p_thresh)[name = string("mask_probs_chunk_4")];
+            string mask_chunk_4_fp16_dtype_0 = const()[name = string("mask_chunk_4_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 64]> mask_chunk_4_fp16 = cast(dtype = mask_chunk_4_fp16_dtype_0, x = mask_probs_chunk_4)[name = string("cast_233")];
+            tensor<fp16, [1, 2048, 1, 64]> masked_probs_chunk_4 = select(a = probs_chunk_4, b = mask_chunk_4_fp16, cond = mask_probs_chunk_4)[name = string("masked_probs_chunk_4")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_5_sub_1 = sub(x = logits_chunk_5_mul, y = logits_lse)[name = string("logits_chunk_5_sub_1")];
+            tensor<fp16, [1, 2048, 1, 64]> probs_chunk_5 = exp(x = logits_chunk_5_sub_1)[name = string("probs_chunk_5")];
+            tensor<bool, [1, 2048, 1, 64]> mask_probs_chunk_5 = greater_equal(x = probs_chunk_5, y = min_p_thresh)[name = string("mask_probs_chunk_5")];
+            string mask_chunk_5_fp16_dtype_0 = const()[name = string("mask_chunk_5_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 64]> mask_chunk_5_fp16 = cast(dtype = mask_chunk_5_fp16_dtype_0, x = mask_probs_chunk_5)[name = string("cast_232")];
+            tensor<fp16, [1, 2048, 1, 64]> masked_probs_chunk_5 = select(a = probs_chunk_5, b = mask_chunk_5_fp16, cond = mask_probs_chunk_5)[name = string("masked_probs_chunk_5")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_6_sub_1 = sub(x = logits_chunk_6_mul, y = logits_lse)[name = string("logits_chunk_6_sub_1")];
+            tensor<fp16, [1, 2048, 1, 64]> probs_chunk_6 = exp(x = logits_chunk_6_sub_1)[name = string("probs_chunk_6")];
+            tensor<bool, [1, 2048, 1, 64]> mask_probs_chunk_6 = greater_equal(x = probs_chunk_6, y = min_p_thresh)[name = string("mask_probs_chunk_6")];
+            string mask_chunk_6_fp16_dtype_0 = const()[name = string("mask_chunk_6_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 64]> mask_chunk_6_fp16 = cast(dtype = mask_chunk_6_fp16_dtype_0, x = mask_probs_chunk_6)[name = string("cast_231")];
+            tensor<fp16, [1, 2048, 1, 64]> masked_probs_chunk_6 = select(a = probs_chunk_6, b = mask_chunk_6_fp16, cond = mask_probs_chunk_6)[name = string("masked_probs_chunk_6")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_7_sub_1 = sub(x = logits_chunk_7_mul, y = logits_lse)[name = string("logits_chunk_7_sub_1")];
+            tensor<fp16, [1, 2048, 1, 64]> probs_chunk_7 = exp(x = logits_chunk_7_sub_1)[name = string("probs_chunk_7")];
+            tensor<bool, [1, 2048, 1, 64]> mask_probs_chunk_7 = greater_equal(x = probs_chunk_7, y = min_p_thresh)[name = string("mask_probs_chunk_7")];
+            string mask_chunk_7_fp16_dtype_0 = const()[name = string("mask_chunk_7_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 64]> mask_chunk_7_fp16 = cast(dtype = mask_chunk_7_fp16_dtype_0, x = mask_probs_chunk_7)[name = string("cast_230")];
+            tensor<fp16, [1, 2048, 1, 64]> masked_probs_chunk_7 = select(a = probs_chunk_7, b = mask_chunk_7_fp16, cond = mask_probs_chunk_7)[name = string("masked_probs_chunk_7")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_8_sub_1 = sub(x = logits_chunk_8_mul, y = logits_lse)[name = string("logits_chunk_8_sub_1")];
+            tensor<fp16, [1, 2048, 1, 64]> probs_chunk_8 = exp(x = logits_chunk_8_sub_1)[name = string("probs_chunk_8")];
+            tensor<bool, [1, 2048, 1, 64]> mask_probs_chunk_8 = greater_equal(x = probs_chunk_8, y = min_p_thresh)[name = string("mask_probs_chunk_8")];
+            string mask_chunk_8_fp16_dtype_0 = const()[name = string("mask_chunk_8_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 64]> mask_chunk_8_fp16 = cast(dtype = mask_chunk_8_fp16_dtype_0, x = mask_probs_chunk_8)[name = string("cast_229")];
+            tensor<fp16, [1, 2048, 1, 64]> masked_probs_chunk_8 = select(a = probs_chunk_8, b = mask_chunk_8_fp16, cond = mask_probs_chunk_8)[name = string("masked_probs_chunk_8")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_9_sub_1 = sub(x = logits_chunk_9_mul, y = logits_lse)[name = string("logits_chunk_9_sub_1")];
+            tensor<fp16, [1, 2048, 1, 64]> probs_chunk_9 = exp(x = logits_chunk_9_sub_1)[name = string("probs_chunk_9")];
+            tensor<bool, [1, 2048, 1, 64]> mask_probs_chunk_9 = greater_equal(x = probs_chunk_9, y = min_p_thresh)[name = string("mask_probs_chunk_9")];
+            string mask_chunk_9_fp16_dtype_0 = const()[name = string("mask_chunk_9_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 64]> mask_chunk_9_fp16 = cast(dtype = mask_chunk_9_fp16_dtype_0, x = mask_probs_chunk_9)[name = string("cast_228")];
+            tensor<fp16, [1, 2048, 1, 64]> masked_probs_chunk_9 = select(a = probs_chunk_9, b = mask_chunk_9_fp16, cond = mask_probs_chunk_9)[name = string("masked_probs_chunk_9")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_10_sub_1 = sub(x = logits_chunk_10_mul, y = logits_lse)[name = string("logits_chunk_10_sub_1")];
+            tensor<fp16, [1, 2048, 1, 64]> probs_chunk_10 = exp(x = logits_chunk_10_sub_1)[name = string("probs_chunk_10")];
+            tensor<bool, [1, 2048, 1, 64]> mask_probs_chunk_10 = greater_equal(x = probs_chunk_10, y = min_p_thresh)[name = string("mask_probs_chunk_10")];
+            string mask_chunk_10_fp16_dtype_0 = const()[name = string("mask_chunk_10_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 64]> mask_chunk_10_fp16 = cast(dtype = mask_chunk_10_fp16_dtype_0, x = mask_probs_chunk_10)[name = string("cast_227")];
+            tensor<fp16, [1, 2048, 1, 64]> masked_probs_chunk_10 = select(a = probs_chunk_10, b = mask_chunk_10_fp16, cond = mask_probs_chunk_10)[name = string("masked_probs_chunk_10")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_11_sub_1 = sub(x = logits_chunk_11_mul, y = logits_lse)[name = string("logits_chunk_11_sub_1")];
+            tensor<fp16, [1, 2048, 1, 64]> probs_chunk_11 = exp(x = logits_chunk_11_sub_1)[name = string("probs_chunk_11")];
+            tensor<bool, [1, 2048, 1, 64]> mask_probs_chunk_11 = greater_equal(x = probs_chunk_11, y = min_p_thresh)[name = string("mask_probs_chunk_11")];
+            string mask_chunk_11_fp16_dtype_0 = const()[name = string("mask_chunk_11_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 64]> mask_chunk_11_fp16 = cast(dtype = mask_chunk_11_fp16_dtype_0, x = mask_probs_chunk_11)[name = string("cast_226")];
+            tensor<fp16, [1, 2048, 1, 64]> masked_probs_chunk_11 = select(a = probs_chunk_11, b = mask_chunk_11_fp16, cond = mask_probs_chunk_11)[name = string("masked_probs_chunk_11")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_12_sub_1 = sub(x = logits_chunk_12_mul, y = logits_lse)[name = string("logits_chunk_12_sub_1")];
+            tensor<fp16, [1, 2048, 1, 64]> probs_chunk_12 = exp(x = logits_chunk_12_sub_1)[name = string("probs_chunk_12")];
+            tensor<bool, [1, 2048, 1, 64]> mask_probs_chunk_12 = greater_equal(x = probs_chunk_12, y = min_p_thresh)[name = string("mask_probs_chunk_12")];
+            string mask_chunk_12_fp16_dtype_0 = const()[name = string("mask_chunk_12_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 64]> mask_chunk_12_fp16 = cast(dtype = mask_chunk_12_fp16_dtype_0, x = mask_probs_chunk_12)[name = string("cast_225")];
+            tensor<fp16, [1, 2048, 1, 64]> masked_probs_chunk_12 = select(a = probs_chunk_12, b = mask_chunk_12_fp16, cond = mask_probs_chunk_12)[name = string("masked_probs_chunk_12")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_13_sub_1 = sub(x = logits_chunk_13_mul, y = logits_lse)[name = string("logits_chunk_13_sub_1")];
+            tensor<fp16, [1, 2048, 1, 64]> probs_chunk_13 = exp(x = logits_chunk_13_sub_1)[name = string("probs_chunk_13")];
+            tensor<bool, [1, 2048, 1, 64]> mask_probs_chunk_13 = greater_equal(x = probs_chunk_13, y = min_p_thresh)[name = string("mask_probs_chunk_13")];
+            string mask_chunk_13_fp16_dtype_0 = const()[name = string("mask_chunk_13_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 64]> mask_chunk_13_fp16 = cast(dtype = mask_chunk_13_fp16_dtype_0, x = mask_probs_chunk_13)[name = string("cast_224")];
+            tensor<fp16, [1, 2048, 1, 64]> masked_probs_chunk_13 = select(a = probs_chunk_13, b = mask_chunk_13_fp16, cond = mask_probs_chunk_13)[name = string("masked_probs_chunk_13")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_14_sub_1 = sub(x = logits_chunk_14_mul, y = logits_lse)[name = string("logits_chunk_14_sub_1")];
+            tensor<fp16, [1, 2048, 1, 64]> probs_chunk_14 = exp(x = logits_chunk_14_sub_1)[name = string("probs_chunk_14")];
+            tensor<bool, [1, 2048, 1, 64]> mask_probs_chunk_14 = greater_equal(x = probs_chunk_14, y = min_p_thresh)[name = string("mask_probs_chunk_14")];
+            string mask_chunk_14_fp16_dtype_0 = const()[name = string("mask_chunk_14_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 64]> mask_chunk_14_fp16 = cast(dtype = mask_chunk_14_fp16_dtype_0, x = mask_probs_chunk_14)[name = string("cast_223")];
+            tensor<fp16, [1, 2048, 1, 64]> masked_probs_chunk_14 = select(a = probs_chunk_14, b = mask_chunk_14_fp16, cond = mask_probs_chunk_14)[name = string("masked_probs_chunk_14")];
+            tensor<fp16, [1, 2048, 1, 64]> logits_chunk_15_sub_1 = sub(x = logits_chunk_15_mul, y = logits_lse)[name = string("logits_chunk_15_sub_1")];
+            tensor<fp16, [1, 2048, 1, 64]> probs_chunk_15 = exp(x = logits_chunk_15_sub_1)[name = string("probs_chunk_15")];
+            tensor<bool, [1, 2048, 1, 64]> mask_probs_chunk_15 = greater_equal(x = probs_chunk_15, y = min_p_thresh)[name = string("mask_probs_chunk_15")];
+            string mask_chunk_15_fp16_dtype_0 = const()[name = string("mask_chunk_15_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 64]> mask_chunk_15_fp16 = cast(dtype = mask_chunk_15_fp16_dtype_0, x = mask_probs_chunk_15)[name = string("cast_222")];
+            tensor<fp16, [1, 2048, 1, 64]> masked_probs_chunk_15 = select(a = probs_chunk_15, b = mask_chunk_15_fp16, cond = mask_probs_chunk_15)[name = string("masked_probs_chunk_15")];
+            int32 probs_axis_0 = const()[name = string("probs_axis_0"), val = int32(1)];
+            bool probs_interleave_0 = const()[name = string("probs_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 32768, 1, 64]> probs = concat(axis = probs_axis_0, interleave = probs_interleave_0, values = (masked_probs_chunk_0, masked_probs_chunk_1, masked_probs_chunk_2, masked_probs_chunk_3, masked_probs_chunk_4, masked_probs_chunk_5, masked_probs_chunk_6, masked_probs_chunk_7, masked_probs_chunk_8, masked_probs_chunk_9, masked_probs_chunk_10, masked_probs_chunk_11, masked_probs_chunk_12, masked_probs_chunk_13, masked_probs_chunk_14, masked_probs_chunk_15))[name = string("probs")];
+            string probs_fp32_dtype_0 = const()[name = string("probs_fp32_dtype_0"), val = string("fp32")];
+            int32 probs_cumsum_axis_0 = const()[name = string("probs_cumsum_axis_0"), val = int32(1)];
+            bool probs_cumsum_exclusive_0 = const()[name = string("probs_cumsum_exclusive_0"), val = bool(false)];
+            bool probs_cumsum_reverse_0 = const()[name = string("probs_cumsum_reverse_0"), val = bool(false)];
+            tensor<fp32, [1, 32768, 1, 64]> probs_fp32 = cast(dtype = probs_fp32_dtype_0, x = probs)[name = string("cast_221")];
+            tensor<fp32, [1, 32768, 1, 64]> probs_cumsum = cumsum(axis = probs_cumsum_axis_0, exclusive = probs_cumsum_exclusive_0, reverse = probs_cumsum_reverse_0, x = probs_fp32)[name = string("probs_cumsum")];
+            tensor<int32, [1]> probs_sum_indices_0 = const()[name = string("probs_sum_indices_0"), val = tensor<int32, [1]>([32767])];
+            int32 probs_sum_axis_0 = const()[name = string("probs_sum_axis_0"), val = int32(1)];
+            int32 probs_sum_batch_dims_0 = const()[name = string("probs_sum_batch_dims_0"), val = int32(0)];
+            bool probs_sum_validate_indices_0 = const()[name = string("probs_sum_validate_indices_0"), val = bool(false)];
+            tensor<fp32, [1, 1, 1, 64]> probs_sum = gather(axis = probs_sum_axis_0, batch_dims = probs_sum_batch_dims_0, indices = probs_sum_indices_0, validate_indices = probs_sum_validate_indices_0, x = probs_cumsum)[name = string("probs_sum")];
+            tensor<fp32, [1, 1, 1, 64]> random_number_scaled = mul(x = random_number, y = probs_sum)[name = string("random_number_scaled")];
+            tensor<bool, [1, 32768, 1, 64]> probs_greater = greater(x = probs_cumsum, y = random_number_scaled)[name = string("probs_greater")];
+            string probs_greater_int32_dtype_0 = const()[name = string("probs_greater_int32_dtype_0"), val = string("int32")];
+            int32 sampled_index_axis_0 = const()[name = string("sampled_index_axis_0"), val = int32(1)];
+            bool sampled_index_keep_dims_0 = const()[name = string("sampled_index_keep_dims_0"), val = bool(true)];
+            string sampled_index_output_dtype_0 = const()[name = string("sampled_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 32768, 1, 64]> probs_greater_int32 = cast(dtype = probs_greater_int32_dtype_0, x = probs_greater)[name = string("cast_220")];
+            tensor<int32, [1, 1, 1, 64]> sampled_index = reduce_argmax(axis = sampled_index_axis_0, keep_dims = sampled_index_keep_dims_0, output_dtype = sampled_index_output_dtype_0, x = probs_greater_int32)[name = string("sampled_index")];
+            int32 sampled_index_probability_axis_0 = const()[name = string("sampled_index_probability_axis_0"), val = int32(1)];
+            bool sampled_index_probability_validate_indices_0 = const()[name = string("sampled_index_probability_validate_indices_0"), val = bool(false)];
+            tensor<fp32, [1, 1, 1, 64]> sampled_index_probability = gather_along_axis(axis = sampled_index_probability_axis_0, indices = sampled_index, validate_indices = sampled_index_probability_validate_indices_0, x = probs_fp32)[name = string("sampled_index_probability")];
+            int32 max_logit_index_axis_0 = const()[name = string("max_logit_index_axis_0"), val = int32(1)];
+            bool max_logit_index_keep_dims_0 = const()[name = string("max_logit_index_keep_dims_0"), val = bool(true)];
+            string max_logit_index_output_dtype_0 = const()[name = string("max_logit_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 1, 1, 64]> max_logit_index = reduce_argmax(axis = max_logit_index_axis_0, keep_dims = max_logit_index_keep_dims_0, output_dtype = max_logit_index_output_dtype_0, x = logits_max_logits_chunks)[name = string("max_logit_index")];
+            string indices_chunk_0_int32_dtype_0 = const()[name = string("indices_chunk_0_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_1_int32_dtype_0 = const()[name = string("indices_chunk_1_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_2_int32_dtype_0 = const()[name = string("indices_chunk_2_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_3_int32_dtype_0 = const()[name = string("indices_chunk_3_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_4_int32_dtype_0 = const()[name = string("indices_chunk_4_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_5_int32_dtype_0 = const()[name = string("indices_chunk_5_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_6_int32_dtype_0 = const()[name = string("indices_chunk_6_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_7_int32_dtype_0 = const()[name = string("indices_chunk_7_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_8_int32_dtype_0 = const()[name = string("indices_chunk_8_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_9_int32_dtype_0 = const()[name = string("indices_chunk_9_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_10_int32_dtype_0 = const()[name = string("indices_chunk_10_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_11_int32_dtype_0 = const()[name = string("indices_chunk_11_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_12_int32_dtype_0 = const()[name = string("indices_chunk_12_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_13_int32_dtype_0 = const()[name = string("indices_chunk_13_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_14_int32_dtype_0 = const()[name = string("indices_chunk_14_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_15_int32_dtype_0 = const()[name = string("indices_chunk_15_int32_dtype_0"), val = string("int32")];
+            int32 indices_axis_0 = const()[name = string("indices_axis_0"), val = int32(1)];
+            bool indices_interleave_0 = const()[name = string("indices_interleave_0"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_15_int32 = cast(dtype = indices_chunk_15_int32_dtype_0, x = logits_chunk_15_argmax)[name = string("cast_204")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_14_int32 = cast(dtype = indices_chunk_14_int32_dtype_0, x = logits_chunk_14_argmax)[name = string("cast_205")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_13_int32 = cast(dtype = indices_chunk_13_int32_dtype_0, x = logits_chunk_13_argmax)[name = string("cast_206")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_12_int32 = cast(dtype = indices_chunk_12_int32_dtype_0, x = logits_chunk_12_argmax)[name = string("cast_207")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_11_int32 = cast(dtype = indices_chunk_11_int32_dtype_0, x = logits_chunk_11_argmax)[name = string("cast_208")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_10_int32 = cast(dtype = indices_chunk_10_int32_dtype_0, x = logits_chunk_10_argmax)[name = string("cast_209")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_9_int32 = cast(dtype = indices_chunk_9_int32_dtype_0, x = logits_chunk_9_argmax)[name = string("cast_210")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_8_int32 = cast(dtype = indices_chunk_8_int32_dtype_0, x = logits_chunk_8_argmax)[name = string("cast_211")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_7_int32 = cast(dtype = indices_chunk_7_int32_dtype_0, x = logits_chunk_7_argmax)[name = string("cast_212")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_6_int32 = cast(dtype = indices_chunk_6_int32_dtype_0, x = logits_chunk_6_argmax)[name = string("cast_213")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_5_int32 = cast(dtype = indices_chunk_5_int32_dtype_0, x = logits_chunk_5_argmax)[name = string("cast_214")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_4_int32 = cast(dtype = indices_chunk_4_int32_dtype_0, x = logits_chunk_4_argmax)[name = string("cast_215")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_3_int32 = cast(dtype = indices_chunk_3_int32_dtype_0, x = logits_chunk_3_argmax)[name = string("cast_216")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_2_int32 = cast(dtype = indices_chunk_2_int32_dtype_0, x = logits_chunk_2_argmax)[name = string("cast_217")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_1_int32 = cast(dtype = indices_chunk_1_int32_dtype_0, x = logits_chunk_1_argmax)[name = string("cast_218")];
+            tensor<int32, [1, 1, 1, 64]> indices_chunk_0_int32 = cast(dtype = indices_chunk_0_int32_dtype_0, x = logits_chunk_0_argmax)[name = string("cast_219")];
+            tensor<int32, [1, 16, 1, 64]> indices = concat(axis = indices_axis_0, interleave = indices_interleave_0, values = (indices_chunk_0_int32, indices_chunk_1_int32, indices_chunk_2_int32, indices_chunk_3_int32, indices_chunk_4_int32, indices_chunk_5_int32, indices_chunk_6_int32, indices_chunk_7_int32, indices_chunk_8_int32, indices_chunk_9_int32, indices_chunk_10_int32, indices_chunk_11_int32, indices_chunk_12_int32, indices_chunk_13_int32, indices_chunk_14_int32, indices_chunk_15_int32))[name = string("indices")];
+            int32 argmax_chunks_axis_0 = const()[name = string("argmax_chunks_axis_0"), val = int32(1)];
+            bool argmax_chunks_validate_indices_0 = const()[name = string("argmax_chunks_validate_indices_0"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 64]> argmax_chunks = gather_along_axis(axis = argmax_chunks_axis_0, indices = max_logit_index, validate_indices = argmax_chunks_validate_indices_0, x = indices)[name = string("argmax_chunks")];
+            int32 mul_0_x_0 = const()[name = string("mul_0_x_0"), val = int32(2048)];
+            tensor<int32, [1, 1, 1, 64]> mul_0 = mul(x = mul_0_x_0, y = max_logit_index)[name = string("mul_0")];
+            tensor<int32, [1, 1, 1, 64]> argmax = add(x = argmax_chunks, y = mul_0)[name = string("argmax")];
+        } -> (sampled_index, sampled_index_probability, argmax, max_prob);
+    func min_p_length_8<ios18>(tensor<fp16, [1, 2048, 1, 8]> hidden_states, tensor<fp16, [1]> p, tensor<fp32, [8]> random_number, tensor<fp16, [1]> temp) {
+            tensor<fp16, [1, 2048, 1, 8]> final_norm_rmsnorm_abs = abs(x = hidden_states)[name = string("final_norm_rmsnorm_abs")];
+            tensor<int32, [1]> final_norm_rmsnorm_maxval_axes_0 = const()[name = string("final_norm_rmsnorm_maxval_axes_0"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_maxval_keep_dims_0 = const()[name = string("final_norm_rmsnorm_maxval_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> final_norm_rmsnorm_maxval = reduce_max(axes = final_norm_rmsnorm_maxval_axes_0, keep_dims = final_norm_rmsnorm_maxval_keep_dims_0, x = final_norm_rmsnorm_abs)[name = string("final_norm_rmsnorm_maxval")];
+            fp16 final_norm_rmsnorm_maxval_clipped_alpha_0 = const()[name = string("final_norm_rmsnorm_maxval_clipped_alpha_0"), val = fp16(0x1p-14)];
+            fp16 final_norm_rmsnorm_maxval_clipped_beta_0 = const()[name = string("final_norm_rmsnorm_maxval_clipped_beta_0"), val = fp16(inf)];
+            tensor<fp16, [1, 1, 1, 8]> final_norm_rmsnorm_maxval_clipped = clip(alpha = final_norm_rmsnorm_maxval_clipped_alpha_0, beta = final_norm_rmsnorm_maxval_clipped_beta_0, x = final_norm_rmsnorm_maxval)[name = string("final_norm_rmsnorm_maxval_clipped")];
+            tensor<fp16, [1, 2048, 1, 8]> final_norm_rmsnorm_scaled = real_div(x = hidden_states, y = final_norm_rmsnorm_maxval_clipped)[name = string("final_norm_rmsnorm_scaled")];
+            tensor<int32, [1]> final_norm_rmsnorm_squared_sum_axes_0 = const()[name = string("final_norm_rmsnorm_squared_sum_axes_0"), val = tensor<int32, [1]>([1])];
+            bool final_norm_rmsnorm_squared_sum_keep_dims_0 = const()[name = string("final_norm_rmsnorm_squared_sum_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> final_norm_rmsnorm_squared_sum = reduce_sum_square(axes = final_norm_rmsnorm_squared_sum_axes_0, keep_dims = final_norm_rmsnorm_squared_sum_keep_dims_0, x = final_norm_rmsnorm_scaled)[name = string("final_norm_rmsnorm_squared_sum")];
+            fp16 final_norm_rmsnorm_rsqrt_epsilon_0 = const()[name = string("final_norm_rmsnorm_rsqrt_epsilon_0"), val = fp16(0x1p-14)];
+            tensor<fp16, [1, 1, 1, 8]> final_norm_rmsnorm_rsqrt = rsqrt(epsilon = final_norm_rmsnorm_rsqrt_epsilon_0, x = final_norm_rmsnorm_squared_sum)[name = string("final_norm_rmsnorm_rsqrt")];
+            fp16 final_norm_rmsnorm_dim_scaled_y_0 = const()[name = string("final_norm_rmsnorm_dim_scaled_y_0"), val = fp16(0x1.6ap+5)];
+            tensor<fp16, [1, 2048, 1, 8]> final_norm_rmsnorm_dim_scaled = mul(x = final_norm_rmsnorm_scaled, y = final_norm_rmsnorm_dim_scaled_y_0)[name = string("final_norm_rmsnorm_dim_scaled")];
+            tensor<fp16, [1, 2048, 1, 8]> final_norm_rmsnorm_normalized = mul(x = final_norm_rmsnorm_dim_scaled, y = final_norm_rmsnorm_rsqrt)[name = string("final_norm_rmsnorm_normalized")];
+            tensor<fp16, [2048, 1, 1]> final_norm_rmsnorm_y_0 = const()[name = string("final_norm_rmsnorm_y_0"), val = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 2048, 1, 8]> final_norm_rmsnorm = mul(x = final_norm_rmsnorm_normalized, y = final_norm_rmsnorm_y_0)[name = string("final_norm_rmsnorm")];
+            fp16 temp_inverse_epsilon_0 = const()[name = string("temp_inverse_epsilon_0"), val = fp16(0x0p+0)];
+            tensor<fp16, [1]> temp_inverse = inverse(epsilon = temp_inverse_epsilon_0, x = temp)[name = string("temp_inverse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_0_weight_0 = const()[name = string("logits_chunk_0_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4224)))];
+            tensor<int32, [2]> logits_chunk_0_strides_0 = const()[name = string("logits_chunk_0_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_0_pad_type_0 = const()[name = string("logits_chunk_0_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_0_pad_0 = const()[name = string("logits_chunk_0_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_0_dilations_0 = const()[name = string("logits_chunk_0_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_0_groups_0 = const()[name = string("logits_chunk_0_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_0 = conv(dilations = logits_chunk_0_dilations_0, groups = logits_chunk_0_groups_0, pad = logits_chunk_0_pad_0, pad_type = logits_chunk_0_pad_type_0, strides = logits_chunk_0_strides_0, weight = logits_chunk_0_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_0")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_0_mul = mul(x = logits_chunk_0, y = temp_inverse)[name = string("logits_chunk_0_mul")];
+            tensor<int32, [1]> logits_chunk_0_max_axes_0 = const()[name = string("logits_chunk_0_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_0_max_keep_dims_0 = const()[name = string("logits_chunk_0_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_0_max = reduce_max(axes = logits_chunk_0_max_axes_0, keep_dims = logits_chunk_0_max_keep_dims_0, x = logits_chunk_0_mul)[name = string("logits_chunk_0_max")];
+            int32 logits_chunk_0_argmax_axis_0 = const()[name = string("logits_chunk_0_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_0_argmax_keep_dims_0 = const()[name = string("logits_chunk_0_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_0_argmax_output_dtype_0 = const()[name = string("logits_chunk_0_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> logits_chunk_0_argmax = reduce_argmax(axis = logits_chunk_0_argmax_axis_0, keep_dims = logits_chunk_0_argmax_keep_dims_0, output_dtype = logits_chunk_0_argmax_output_dtype_0, x = logits_chunk_0_mul)[name = string("logits_chunk_0_argmax")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_0_sub = sub(x = logits_chunk_0_mul, y = logits_chunk_0_max)[name = string("logits_chunk_0_sub")];
+            tensor<int32, [1]> logits_chunk_0_lse_sub_axes_0 = const()[name = string("logits_chunk_0_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_0_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_0_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_0_lse_sub = reduce_log_sum_exp(axes = logits_chunk_0_lse_sub_axes_0, keep_dims = logits_chunk_0_lse_sub_keep_dims_0, x = logits_chunk_0_sub)[name = string("logits_chunk_0_lse_sub")];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_0_lse = add(x = logits_chunk_0_lse_sub, y = logits_chunk_0_max)[name = string("logits_chunk_0_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_1_weight_0 = const()[name = string("logits_chunk_1_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8392896)))];
+            tensor<int32, [2]> logits_chunk_1_strides_0 = const()[name = string("logits_chunk_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_1_pad_type_0 = const()[name = string("logits_chunk_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_1_pad_0 = const()[name = string("logits_chunk_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_1_dilations_0 = const()[name = string("logits_chunk_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_1_groups_0 = const()[name = string("logits_chunk_1_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_1 = conv(dilations = logits_chunk_1_dilations_0, groups = logits_chunk_1_groups_0, pad = logits_chunk_1_pad_0, pad_type = logits_chunk_1_pad_type_0, strides = logits_chunk_1_strides_0, weight = logits_chunk_1_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_1")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_1_mul = mul(x = logits_chunk_1, y = temp_inverse)[name = string("logits_chunk_1_mul")];
+            tensor<int32, [1]> logits_chunk_1_max_axes_0 = const()[name = string("logits_chunk_1_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_1_max_keep_dims_0 = const()[name = string("logits_chunk_1_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_1_max = reduce_max(axes = logits_chunk_1_max_axes_0, keep_dims = logits_chunk_1_max_keep_dims_0, x = logits_chunk_1_mul)[name = string("logits_chunk_1_max")];
+            int32 logits_chunk_1_argmax_axis_0 = const()[name = string("logits_chunk_1_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_1_argmax_keep_dims_0 = const()[name = string("logits_chunk_1_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_1_argmax_output_dtype_0 = const()[name = string("logits_chunk_1_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> logits_chunk_1_argmax = reduce_argmax(axis = logits_chunk_1_argmax_axis_0, keep_dims = logits_chunk_1_argmax_keep_dims_0, output_dtype = logits_chunk_1_argmax_output_dtype_0, x = logits_chunk_1_mul)[name = string("logits_chunk_1_argmax")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_1_sub = sub(x = logits_chunk_1_mul, y = logits_chunk_1_max)[name = string("logits_chunk_1_sub")];
+            tensor<int32, [1]> logits_chunk_1_lse_sub_axes_0 = const()[name = string("logits_chunk_1_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_1_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_1_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_1_lse_sub = reduce_log_sum_exp(axes = logits_chunk_1_lse_sub_axes_0, keep_dims = logits_chunk_1_lse_sub_keep_dims_0, x = logits_chunk_1_sub)[name = string("logits_chunk_1_lse_sub")];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_1_lse = add(x = logits_chunk_1_lse_sub, y = logits_chunk_1_max)[name = string("logits_chunk_1_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_2_weight_0 = const()[name = string("logits_chunk_2_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16781568)))];
+            tensor<int32, [2]> logits_chunk_2_strides_0 = const()[name = string("logits_chunk_2_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_2_pad_type_0 = const()[name = string("logits_chunk_2_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_2_pad_0 = const()[name = string("logits_chunk_2_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_2_dilations_0 = const()[name = string("logits_chunk_2_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_2_groups_0 = const()[name = string("logits_chunk_2_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_2 = conv(dilations = logits_chunk_2_dilations_0, groups = logits_chunk_2_groups_0, pad = logits_chunk_2_pad_0, pad_type = logits_chunk_2_pad_type_0, strides = logits_chunk_2_strides_0, weight = logits_chunk_2_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_2")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_2_mul = mul(x = logits_chunk_2, y = temp_inverse)[name = string("logits_chunk_2_mul")];
+            tensor<int32, [1]> logits_chunk_2_max_axes_0 = const()[name = string("logits_chunk_2_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_2_max_keep_dims_0 = const()[name = string("logits_chunk_2_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_2_max = reduce_max(axes = logits_chunk_2_max_axes_0, keep_dims = logits_chunk_2_max_keep_dims_0, x = logits_chunk_2_mul)[name = string("logits_chunk_2_max")];
+            int32 logits_chunk_2_argmax_axis_0 = const()[name = string("logits_chunk_2_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_2_argmax_keep_dims_0 = const()[name = string("logits_chunk_2_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_2_argmax_output_dtype_0 = const()[name = string("logits_chunk_2_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> logits_chunk_2_argmax = reduce_argmax(axis = logits_chunk_2_argmax_axis_0, keep_dims = logits_chunk_2_argmax_keep_dims_0, output_dtype = logits_chunk_2_argmax_output_dtype_0, x = logits_chunk_2_mul)[name = string("logits_chunk_2_argmax")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_2_sub = sub(x = logits_chunk_2_mul, y = logits_chunk_2_max)[name = string("logits_chunk_2_sub")];
+            tensor<int32, [1]> logits_chunk_2_lse_sub_axes_0 = const()[name = string("logits_chunk_2_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_2_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_2_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_2_lse_sub = reduce_log_sum_exp(axes = logits_chunk_2_lse_sub_axes_0, keep_dims = logits_chunk_2_lse_sub_keep_dims_0, x = logits_chunk_2_sub)[name = string("logits_chunk_2_lse_sub")];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_2_lse = add(x = logits_chunk_2_lse_sub, y = logits_chunk_2_max)[name = string("logits_chunk_2_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_3_weight_0 = const()[name = string("logits_chunk_3_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25170240)))];
+            tensor<int32, [2]> logits_chunk_3_strides_0 = const()[name = string("logits_chunk_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_3_pad_type_0 = const()[name = string("logits_chunk_3_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_3_pad_0 = const()[name = string("logits_chunk_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_3_dilations_0 = const()[name = string("logits_chunk_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_3_groups_0 = const()[name = string("logits_chunk_3_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_3 = conv(dilations = logits_chunk_3_dilations_0, groups = logits_chunk_3_groups_0, pad = logits_chunk_3_pad_0, pad_type = logits_chunk_3_pad_type_0, strides = logits_chunk_3_strides_0, weight = logits_chunk_3_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_3")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_3_mul = mul(x = logits_chunk_3, y = temp_inverse)[name = string("logits_chunk_3_mul")];
+            tensor<int32, [1]> logits_chunk_3_max_axes_0 = const()[name = string("logits_chunk_3_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_3_max_keep_dims_0 = const()[name = string("logits_chunk_3_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_3_max = reduce_max(axes = logits_chunk_3_max_axes_0, keep_dims = logits_chunk_3_max_keep_dims_0, x = logits_chunk_3_mul)[name = string("logits_chunk_3_max")];
+            int32 logits_chunk_3_argmax_axis_0 = const()[name = string("logits_chunk_3_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_3_argmax_keep_dims_0 = const()[name = string("logits_chunk_3_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_3_argmax_output_dtype_0 = const()[name = string("logits_chunk_3_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> logits_chunk_3_argmax = reduce_argmax(axis = logits_chunk_3_argmax_axis_0, keep_dims = logits_chunk_3_argmax_keep_dims_0, output_dtype = logits_chunk_3_argmax_output_dtype_0, x = logits_chunk_3_mul)[name = string("logits_chunk_3_argmax")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_3_sub = sub(x = logits_chunk_3_mul, y = logits_chunk_3_max)[name = string("logits_chunk_3_sub")];
+            tensor<int32, [1]> logits_chunk_3_lse_sub_axes_0 = const()[name = string("logits_chunk_3_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_3_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_3_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_3_lse_sub = reduce_log_sum_exp(axes = logits_chunk_3_lse_sub_axes_0, keep_dims = logits_chunk_3_lse_sub_keep_dims_0, x = logits_chunk_3_sub)[name = string("logits_chunk_3_lse_sub")];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_3_lse = add(x = logits_chunk_3_lse_sub, y = logits_chunk_3_max)[name = string("logits_chunk_3_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_4_weight_0 = const()[name = string("logits_chunk_4_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33558912)))];
+            tensor<int32, [2]> logits_chunk_4_strides_0 = const()[name = string("logits_chunk_4_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_4_pad_type_0 = const()[name = string("logits_chunk_4_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_4_pad_0 = const()[name = string("logits_chunk_4_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_4_dilations_0 = const()[name = string("logits_chunk_4_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_4_groups_0 = const()[name = string("logits_chunk_4_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_4 = conv(dilations = logits_chunk_4_dilations_0, groups = logits_chunk_4_groups_0, pad = logits_chunk_4_pad_0, pad_type = logits_chunk_4_pad_type_0, strides = logits_chunk_4_strides_0, weight = logits_chunk_4_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_4")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_4_mul = mul(x = logits_chunk_4, y = temp_inverse)[name = string("logits_chunk_4_mul")];
+            tensor<int32, [1]> logits_chunk_4_max_axes_0 = const()[name = string("logits_chunk_4_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_4_max_keep_dims_0 = const()[name = string("logits_chunk_4_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_4_max = reduce_max(axes = logits_chunk_4_max_axes_0, keep_dims = logits_chunk_4_max_keep_dims_0, x = logits_chunk_4_mul)[name = string("logits_chunk_4_max")];
+            int32 logits_chunk_4_argmax_axis_0 = const()[name = string("logits_chunk_4_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_4_argmax_keep_dims_0 = const()[name = string("logits_chunk_4_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_4_argmax_output_dtype_0 = const()[name = string("logits_chunk_4_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> logits_chunk_4_argmax = reduce_argmax(axis = logits_chunk_4_argmax_axis_0, keep_dims = logits_chunk_4_argmax_keep_dims_0, output_dtype = logits_chunk_4_argmax_output_dtype_0, x = logits_chunk_4_mul)[name = string("logits_chunk_4_argmax")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_4_sub = sub(x = logits_chunk_4_mul, y = logits_chunk_4_max)[name = string("logits_chunk_4_sub")];
+            tensor<int32, [1]> logits_chunk_4_lse_sub_axes_0 = const()[name = string("logits_chunk_4_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_4_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_4_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_4_lse_sub = reduce_log_sum_exp(axes = logits_chunk_4_lse_sub_axes_0, keep_dims = logits_chunk_4_lse_sub_keep_dims_0, x = logits_chunk_4_sub)[name = string("logits_chunk_4_lse_sub")];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_4_lse = add(x = logits_chunk_4_lse_sub, y = logits_chunk_4_max)[name = string("logits_chunk_4_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_5_weight_0 = const()[name = string("logits_chunk_5_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41947584)))];
+            tensor<int32, [2]> logits_chunk_5_strides_0 = const()[name = string("logits_chunk_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_5_pad_type_0 = const()[name = string("logits_chunk_5_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_5_pad_0 = const()[name = string("logits_chunk_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_5_dilations_0 = const()[name = string("logits_chunk_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_5_groups_0 = const()[name = string("logits_chunk_5_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_5 = conv(dilations = logits_chunk_5_dilations_0, groups = logits_chunk_5_groups_0, pad = logits_chunk_5_pad_0, pad_type = logits_chunk_5_pad_type_0, strides = logits_chunk_5_strides_0, weight = logits_chunk_5_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_5")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_5_mul = mul(x = logits_chunk_5, y = temp_inverse)[name = string("logits_chunk_5_mul")];
+            tensor<int32, [1]> logits_chunk_5_max_axes_0 = const()[name = string("logits_chunk_5_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_5_max_keep_dims_0 = const()[name = string("logits_chunk_5_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_5_max = reduce_max(axes = logits_chunk_5_max_axes_0, keep_dims = logits_chunk_5_max_keep_dims_0, x = logits_chunk_5_mul)[name = string("logits_chunk_5_max")];
+            int32 logits_chunk_5_argmax_axis_0 = const()[name = string("logits_chunk_5_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_5_argmax_keep_dims_0 = const()[name = string("logits_chunk_5_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_5_argmax_output_dtype_0 = const()[name = string("logits_chunk_5_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> logits_chunk_5_argmax = reduce_argmax(axis = logits_chunk_5_argmax_axis_0, keep_dims = logits_chunk_5_argmax_keep_dims_0, output_dtype = logits_chunk_5_argmax_output_dtype_0, x = logits_chunk_5_mul)[name = string("logits_chunk_5_argmax")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_5_sub = sub(x = logits_chunk_5_mul, y = logits_chunk_5_max)[name = string("logits_chunk_5_sub")];
+            tensor<int32, [1]> logits_chunk_5_lse_sub_axes_0 = const()[name = string("logits_chunk_5_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_5_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_5_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_5_lse_sub = reduce_log_sum_exp(axes = logits_chunk_5_lse_sub_axes_0, keep_dims = logits_chunk_5_lse_sub_keep_dims_0, x = logits_chunk_5_sub)[name = string("logits_chunk_5_lse_sub")];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_5_lse = add(x = logits_chunk_5_lse_sub, y = logits_chunk_5_max)[name = string("logits_chunk_5_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_6_weight_0 = const()[name = string("logits_chunk_6_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50336256)))];
+            tensor<int32, [2]> logits_chunk_6_strides_0 = const()[name = string("logits_chunk_6_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_6_pad_type_0 = const()[name = string("logits_chunk_6_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_6_pad_0 = const()[name = string("logits_chunk_6_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_6_dilations_0 = const()[name = string("logits_chunk_6_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_6_groups_0 = const()[name = string("logits_chunk_6_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_6 = conv(dilations = logits_chunk_6_dilations_0, groups = logits_chunk_6_groups_0, pad = logits_chunk_6_pad_0, pad_type = logits_chunk_6_pad_type_0, strides = logits_chunk_6_strides_0, weight = logits_chunk_6_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_6")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_6_mul = mul(x = logits_chunk_6, y = temp_inverse)[name = string("logits_chunk_6_mul")];
+            tensor<int32, [1]> logits_chunk_6_max_axes_0 = const()[name = string("logits_chunk_6_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_6_max_keep_dims_0 = const()[name = string("logits_chunk_6_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_6_max = reduce_max(axes = logits_chunk_6_max_axes_0, keep_dims = logits_chunk_6_max_keep_dims_0, x = logits_chunk_6_mul)[name = string("logits_chunk_6_max")];
+            int32 logits_chunk_6_argmax_axis_0 = const()[name = string("logits_chunk_6_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_6_argmax_keep_dims_0 = const()[name = string("logits_chunk_6_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_6_argmax_output_dtype_0 = const()[name = string("logits_chunk_6_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> logits_chunk_6_argmax = reduce_argmax(axis = logits_chunk_6_argmax_axis_0, keep_dims = logits_chunk_6_argmax_keep_dims_0, output_dtype = logits_chunk_6_argmax_output_dtype_0, x = logits_chunk_6_mul)[name = string("logits_chunk_6_argmax")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_6_sub = sub(x = logits_chunk_6_mul, y = logits_chunk_6_max)[name = string("logits_chunk_6_sub")];
+            tensor<int32, [1]> logits_chunk_6_lse_sub_axes_0 = const()[name = string("logits_chunk_6_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_6_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_6_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_6_lse_sub = reduce_log_sum_exp(axes = logits_chunk_6_lse_sub_axes_0, keep_dims = logits_chunk_6_lse_sub_keep_dims_0, x = logits_chunk_6_sub)[name = string("logits_chunk_6_lse_sub")];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_6_lse = add(x = logits_chunk_6_lse_sub, y = logits_chunk_6_max)[name = string("logits_chunk_6_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_7_weight_0 = const()[name = string("logits_chunk_7_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58724928)))];
+            tensor<int32, [2]> logits_chunk_7_strides_0 = const()[name = string("logits_chunk_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_7_pad_type_0 = const()[name = string("logits_chunk_7_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_7_pad_0 = const()[name = string("logits_chunk_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_7_dilations_0 = const()[name = string("logits_chunk_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_7_groups_0 = const()[name = string("logits_chunk_7_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_7 = conv(dilations = logits_chunk_7_dilations_0, groups = logits_chunk_7_groups_0, pad = logits_chunk_7_pad_0, pad_type = logits_chunk_7_pad_type_0, strides = logits_chunk_7_strides_0, weight = logits_chunk_7_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_7")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_7_mul = mul(x = logits_chunk_7, y = temp_inverse)[name = string("logits_chunk_7_mul")];
+            tensor<int32, [1]> logits_chunk_7_max_axes_0 = const()[name = string("logits_chunk_7_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_7_max_keep_dims_0 = const()[name = string("logits_chunk_7_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_7_max = reduce_max(axes = logits_chunk_7_max_axes_0, keep_dims = logits_chunk_7_max_keep_dims_0, x = logits_chunk_7_mul)[name = string("logits_chunk_7_max")];
+            int32 logits_chunk_7_argmax_axis_0 = const()[name = string("logits_chunk_7_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_7_argmax_keep_dims_0 = const()[name = string("logits_chunk_7_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_7_argmax_output_dtype_0 = const()[name = string("logits_chunk_7_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> logits_chunk_7_argmax = reduce_argmax(axis = logits_chunk_7_argmax_axis_0, keep_dims = logits_chunk_7_argmax_keep_dims_0, output_dtype = logits_chunk_7_argmax_output_dtype_0, x = logits_chunk_7_mul)[name = string("logits_chunk_7_argmax")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_7_sub = sub(x = logits_chunk_7_mul, y = logits_chunk_7_max)[name = string("logits_chunk_7_sub")];
+            tensor<int32, [1]> logits_chunk_7_lse_sub_axes_0 = const()[name = string("logits_chunk_7_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_7_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_7_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_7_lse_sub = reduce_log_sum_exp(axes = logits_chunk_7_lse_sub_axes_0, keep_dims = logits_chunk_7_lse_sub_keep_dims_0, x = logits_chunk_7_sub)[name = string("logits_chunk_7_lse_sub")];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_7_lse = add(x = logits_chunk_7_lse_sub, y = logits_chunk_7_max)[name = string("logits_chunk_7_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_8_weight_0 = const()[name = string("logits_chunk_8_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67113600)))];
+            tensor<int32, [2]> logits_chunk_8_strides_0 = const()[name = string("logits_chunk_8_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_8_pad_type_0 = const()[name = string("logits_chunk_8_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_8_pad_0 = const()[name = string("logits_chunk_8_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_8_dilations_0 = const()[name = string("logits_chunk_8_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_8_groups_0 = const()[name = string("logits_chunk_8_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_8 = conv(dilations = logits_chunk_8_dilations_0, groups = logits_chunk_8_groups_0, pad = logits_chunk_8_pad_0, pad_type = logits_chunk_8_pad_type_0, strides = logits_chunk_8_strides_0, weight = logits_chunk_8_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_8")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_8_mul = mul(x = logits_chunk_8, y = temp_inverse)[name = string("logits_chunk_8_mul")];
+            tensor<int32, [1]> logits_chunk_8_max_axes_0 = const()[name = string("logits_chunk_8_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_8_max_keep_dims_0 = const()[name = string("logits_chunk_8_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_8_max = reduce_max(axes = logits_chunk_8_max_axes_0, keep_dims = logits_chunk_8_max_keep_dims_0, x = logits_chunk_8_mul)[name = string("logits_chunk_8_max")];
+            int32 logits_chunk_8_argmax_axis_0 = const()[name = string("logits_chunk_8_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_8_argmax_keep_dims_0 = const()[name = string("logits_chunk_8_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_8_argmax_output_dtype_0 = const()[name = string("logits_chunk_8_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> logits_chunk_8_argmax = reduce_argmax(axis = logits_chunk_8_argmax_axis_0, keep_dims = logits_chunk_8_argmax_keep_dims_0, output_dtype = logits_chunk_8_argmax_output_dtype_0, x = logits_chunk_8_mul)[name = string("logits_chunk_8_argmax")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_8_sub = sub(x = logits_chunk_8_mul, y = logits_chunk_8_max)[name = string("logits_chunk_8_sub")];
+            tensor<int32, [1]> logits_chunk_8_lse_sub_axes_0 = const()[name = string("logits_chunk_8_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_8_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_8_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_8_lse_sub = reduce_log_sum_exp(axes = logits_chunk_8_lse_sub_axes_0, keep_dims = logits_chunk_8_lse_sub_keep_dims_0, x = logits_chunk_8_sub)[name = string("logits_chunk_8_lse_sub")];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_8_lse = add(x = logits_chunk_8_lse_sub, y = logits_chunk_8_max)[name = string("logits_chunk_8_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_9_weight_0 = const()[name = string("logits_chunk_9_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75502272)))];
+            tensor<int32, [2]> logits_chunk_9_strides_0 = const()[name = string("logits_chunk_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_9_pad_type_0 = const()[name = string("logits_chunk_9_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_9_pad_0 = const()[name = string("logits_chunk_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_9_dilations_0 = const()[name = string("logits_chunk_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_9_groups_0 = const()[name = string("logits_chunk_9_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_9 = conv(dilations = logits_chunk_9_dilations_0, groups = logits_chunk_9_groups_0, pad = logits_chunk_9_pad_0, pad_type = logits_chunk_9_pad_type_0, strides = logits_chunk_9_strides_0, weight = logits_chunk_9_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_9")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_9_mul = mul(x = logits_chunk_9, y = temp_inverse)[name = string("logits_chunk_9_mul")];
+            tensor<int32, [1]> logits_chunk_9_max_axes_0 = const()[name = string("logits_chunk_9_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_9_max_keep_dims_0 = const()[name = string("logits_chunk_9_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_9_max = reduce_max(axes = logits_chunk_9_max_axes_0, keep_dims = logits_chunk_9_max_keep_dims_0, x = logits_chunk_9_mul)[name = string("logits_chunk_9_max")];
+            int32 logits_chunk_9_argmax_axis_0 = const()[name = string("logits_chunk_9_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_9_argmax_keep_dims_0 = const()[name = string("logits_chunk_9_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_9_argmax_output_dtype_0 = const()[name = string("logits_chunk_9_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> logits_chunk_9_argmax = reduce_argmax(axis = logits_chunk_9_argmax_axis_0, keep_dims = logits_chunk_9_argmax_keep_dims_0, output_dtype = logits_chunk_9_argmax_output_dtype_0, x = logits_chunk_9_mul)[name = string("logits_chunk_9_argmax")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_9_sub = sub(x = logits_chunk_9_mul, y = logits_chunk_9_max)[name = string("logits_chunk_9_sub")];
+            tensor<int32, [1]> logits_chunk_9_lse_sub_axes_0 = const()[name = string("logits_chunk_9_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_9_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_9_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_9_lse_sub = reduce_log_sum_exp(axes = logits_chunk_9_lse_sub_axes_0, keep_dims = logits_chunk_9_lse_sub_keep_dims_0, x = logits_chunk_9_sub)[name = string("logits_chunk_9_lse_sub")];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_9_lse = add(x = logits_chunk_9_lse_sub, y = logits_chunk_9_max)[name = string("logits_chunk_9_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_10_weight_0 = const()[name = string("logits_chunk_10_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83890944)))];
+            tensor<int32, [2]> logits_chunk_10_strides_0 = const()[name = string("logits_chunk_10_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_10_pad_type_0 = const()[name = string("logits_chunk_10_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_10_pad_0 = const()[name = string("logits_chunk_10_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_10_dilations_0 = const()[name = string("logits_chunk_10_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_10_groups_0 = const()[name = string("logits_chunk_10_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_10 = conv(dilations = logits_chunk_10_dilations_0, groups = logits_chunk_10_groups_0, pad = logits_chunk_10_pad_0, pad_type = logits_chunk_10_pad_type_0, strides = logits_chunk_10_strides_0, weight = logits_chunk_10_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_10")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_10_mul = mul(x = logits_chunk_10, y = temp_inverse)[name = string("logits_chunk_10_mul")];
+            tensor<int32, [1]> logits_chunk_10_max_axes_0 = const()[name = string("logits_chunk_10_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_10_max_keep_dims_0 = const()[name = string("logits_chunk_10_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_10_max = reduce_max(axes = logits_chunk_10_max_axes_0, keep_dims = logits_chunk_10_max_keep_dims_0, x = logits_chunk_10_mul)[name = string("logits_chunk_10_max")];
+            int32 logits_chunk_10_argmax_axis_0 = const()[name = string("logits_chunk_10_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_10_argmax_keep_dims_0 = const()[name = string("logits_chunk_10_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_10_argmax_output_dtype_0 = const()[name = string("logits_chunk_10_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> logits_chunk_10_argmax = reduce_argmax(axis = logits_chunk_10_argmax_axis_0, keep_dims = logits_chunk_10_argmax_keep_dims_0, output_dtype = logits_chunk_10_argmax_output_dtype_0, x = logits_chunk_10_mul)[name = string("logits_chunk_10_argmax")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_10_sub = sub(x = logits_chunk_10_mul, y = logits_chunk_10_max)[name = string("logits_chunk_10_sub")];
+            tensor<int32, [1]> logits_chunk_10_lse_sub_axes_0 = const()[name = string("logits_chunk_10_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_10_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_10_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_10_lse_sub = reduce_log_sum_exp(axes = logits_chunk_10_lse_sub_axes_0, keep_dims = logits_chunk_10_lse_sub_keep_dims_0, x = logits_chunk_10_sub)[name = string("logits_chunk_10_lse_sub")];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_10_lse = add(x = logits_chunk_10_lse_sub, y = logits_chunk_10_max)[name = string("logits_chunk_10_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_11_weight_0 = const()[name = string("logits_chunk_11_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92279616)))];
+            tensor<int32, [2]> logits_chunk_11_strides_0 = const()[name = string("logits_chunk_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_11_pad_type_0 = const()[name = string("logits_chunk_11_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_11_pad_0 = const()[name = string("logits_chunk_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_11_dilations_0 = const()[name = string("logits_chunk_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_11_groups_0 = const()[name = string("logits_chunk_11_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_11 = conv(dilations = logits_chunk_11_dilations_0, groups = logits_chunk_11_groups_0, pad = logits_chunk_11_pad_0, pad_type = logits_chunk_11_pad_type_0, strides = logits_chunk_11_strides_0, weight = logits_chunk_11_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_11")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_11_mul = mul(x = logits_chunk_11, y = temp_inverse)[name = string("logits_chunk_11_mul")];
+            tensor<int32, [1]> logits_chunk_11_max_axes_0 = const()[name = string("logits_chunk_11_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_11_max_keep_dims_0 = const()[name = string("logits_chunk_11_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_11_max = reduce_max(axes = logits_chunk_11_max_axes_0, keep_dims = logits_chunk_11_max_keep_dims_0, x = logits_chunk_11_mul)[name = string("logits_chunk_11_max")];
+            int32 logits_chunk_11_argmax_axis_0 = const()[name = string("logits_chunk_11_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_11_argmax_keep_dims_0 = const()[name = string("logits_chunk_11_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_11_argmax_output_dtype_0 = const()[name = string("logits_chunk_11_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> logits_chunk_11_argmax = reduce_argmax(axis = logits_chunk_11_argmax_axis_0, keep_dims = logits_chunk_11_argmax_keep_dims_0, output_dtype = logits_chunk_11_argmax_output_dtype_0, x = logits_chunk_11_mul)[name = string("logits_chunk_11_argmax")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_11_sub = sub(x = logits_chunk_11_mul, y = logits_chunk_11_max)[name = string("logits_chunk_11_sub")];
+            tensor<int32, [1]> logits_chunk_11_lse_sub_axes_0 = const()[name = string("logits_chunk_11_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_11_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_11_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_11_lse_sub = reduce_log_sum_exp(axes = logits_chunk_11_lse_sub_axes_0, keep_dims = logits_chunk_11_lse_sub_keep_dims_0, x = logits_chunk_11_sub)[name = string("logits_chunk_11_lse_sub")];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_11_lse = add(x = logits_chunk_11_lse_sub, y = logits_chunk_11_max)[name = string("logits_chunk_11_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_12_weight_0 = const()[name = string("logits_chunk_12_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100668288)))];
+            tensor<int32, [2]> logits_chunk_12_strides_0 = const()[name = string("logits_chunk_12_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_12_pad_type_0 = const()[name = string("logits_chunk_12_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_12_pad_0 = const()[name = string("logits_chunk_12_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_12_dilations_0 = const()[name = string("logits_chunk_12_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_12_groups_0 = const()[name = string("logits_chunk_12_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_12 = conv(dilations = logits_chunk_12_dilations_0, groups = logits_chunk_12_groups_0, pad = logits_chunk_12_pad_0, pad_type = logits_chunk_12_pad_type_0, strides = logits_chunk_12_strides_0, weight = logits_chunk_12_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_12")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_12_mul = mul(x = logits_chunk_12, y = temp_inverse)[name = string("logits_chunk_12_mul")];
+            tensor<int32, [1]> logits_chunk_12_max_axes_0 = const()[name = string("logits_chunk_12_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_12_max_keep_dims_0 = const()[name = string("logits_chunk_12_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_12_max = reduce_max(axes = logits_chunk_12_max_axes_0, keep_dims = logits_chunk_12_max_keep_dims_0, x = logits_chunk_12_mul)[name = string("logits_chunk_12_max")];
+            int32 logits_chunk_12_argmax_axis_0 = const()[name = string("logits_chunk_12_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_12_argmax_keep_dims_0 = const()[name = string("logits_chunk_12_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_12_argmax_output_dtype_0 = const()[name = string("logits_chunk_12_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> logits_chunk_12_argmax = reduce_argmax(axis = logits_chunk_12_argmax_axis_0, keep_dims = logits_chunk_12_argmax_keep_dims_0, output_dtype = logits_chunk_12_argmax_output_dtype_0, x = logits_chunk_12_mul)[name = string("logits_chunk_12_argmax")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_12_sub = sub(x = logits_chunk_12_mul, y = logits_chunk_12_max)[name = string("logits_chunk_12_sub")];
+            tensor<int32, [1]> logits_chunk_12_lse_sub_axes_0 = const()[name = string("logits_chunk_12_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_12_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_12_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_12_lse_sub = reduce_log_sum_exp(axes = logits_chunk_12_lse_sub_axes_0, keep_dims = logits_chunk_12_lse_sub_keep_dims_0, x = logits_chunk_12_sub)[name = string("logits_chunk_12_lse_sub")];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_12_lse = add(x = logits_chunk_12_lse_sub, y = logits_chunk_12_max)[name = string("logits_chunk_12_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_13_weight_0 = const()[name = string("logits_chunk_13_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109056960)))];
+            tensor<int32, [2]> logits_chunk_13_strides_0 = const()[name = string("logits_chunk_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_13_pad_type_0 = const()[name = string("logits_chunk_13_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_13_pad_0 = const()[name = string("logits_chunk_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_13_dilations_0 = const()[name = string("logits_chunk_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_13_groups_0 = const()[name = string("logits_chunk_13_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_13 = conv(dilations = logits_chunk_13_dilations_0, groups = logits_chunk_13_groups_0, pad = logits_chunk_13_pad_0, pad_type = logits_chunk_13_pad_type_0, strides = logits_chunk_13_strides_0, weight = logits_chunk_13_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_13")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_13_mul = mul(x = logits_chunk_13, y = temp_inverse)[name = string("logits_chunk_13_mul")];
+            tensor<int32, [1]> logits_chunk_13_max_axes_0 = const()[name = string("logits_chunk_13_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_13_max_keep_dims_0 = const()[name = string("logits_chunk_13_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_13_max = reduce_max(axes = logits_chunk_13_max_axes_0, keep_dims = logits_chunk_13_max_keep_dims_0, x = logits_chunk_13_mul)[name = string("logits_chunk_13_max")];
+            int32 logits_chunk_13_argmax_axis_0 = const()[name = string("logits_chunk_13_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_13_argmax_keep_dims_0 = const()[name = string("logits_chunk_13_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_13_argmax_output_dtype_0 = const()[name = string("logits_chunk_13_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> logits_chunk_13_argmax = reduce_argmax(axis = logits_chunk_13_argmax_axis_0, keep_dims = logits_chunk_13_argmax_keep_dims_0, output_dtype = logits_chunk_13_argmax_output_dtype_0, x = logits_chunk_13_mul)[name = string("logits_chunk_13_argmax")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_13_sub = sub(x = logits_chunk_13_mul, y = logits_chunk_13_max)[name = string("logits_chunk_13_sub")];
+            tensor<int32, [1]> logits_chunk_13_lse_sub_axes_0 = const()[name = string("logits_chunk_13_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_13_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_13_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_13_lse_sub = reduce_log_sum_exp(axes = logits_chunk_13_lse_sub_axes_0, keep_dims = logits_chunk_13_lse_sub_keep_dims_0, x = logits_chunk_13_sub)[name = string("logits_chunk_13_lse_sub")];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_13_lse = add(x = logits_chunk_13_lse_sub, y = logits_chunk_13_max)[name = string("logits_chunk_13_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_14_weight_0 = const()[name = string("logits_chunk_14_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117445632)))];
+            tensor<int32, [2]> logits_chunk_14_strides_0 = const()[name = string("logits_chunk_14_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_14_pad_type_0 = const()[name = string("logits_chunk_14_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_14_pad_0 = const()[name = string("logits_chunk_14_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_14_dilations_0 = const()[name = string("logits_chunk_14_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_14_groups_0 = const()[name = string("logits_chunk_14_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_14 = conv(dilations = logits_chunk_14_dilations_0, groups = logits_chunk_14_groups_0, pad = logits_chunk_14_pad_0, pad_type = logits_chunk_14_pad_type_0, strides = logits_chunk_14_strides_0, weight = logits_chunk_14_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_14")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_14_mul = mul(x = logits_chunk_14, y = temp_inverse)[name = string("logits_chunk_14_mul")];
+            tensor<int32, [1]> logits_chunk_14_max_axes_0 = const()[name = string("logits_chunk_14_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_14_max_keep_dims_0 = const()[name = string("logits_chunk_14_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_14_max = reduce_max(axes = logits_chunk_14_max_axes_0, keep_dims = logits_chunk_14_max_keep_dims_0, x = logits_chunk_14_mul)[name = string("logits_chunk_14_max")];
+            int32 logits_chunk_14_argmax_axis_0 = const()[name = string("logits_chunk_14_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_14_argmax_keep_dims_0 = const()[name = string("logits_chunk_14_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_14_argmax_output_dtype_0 = const()[name = string("logits_chunk_14_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> logits_chunk_14_argmax = reduce_argmax(axis = logits_chunk_14_argmax_axis_0, keep_dims = logits_chunk_14_argmax_keep_dims_0, output_dtype = logits_chunk_14_argmax_output_dtype_0, x = logits_chunk_14_mul)[name = string("logits_chunk_14_argmax")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_14_sub = sub(x = logits_chunk_14_mul, y = logits_chunk_14_max)[name = string("logits_chunk_14_sub")];
+            tensor<int32, [1]> logits_chunk_14_lse_sub_axes_0 = const()[name = string("logits_chunk_14_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_14_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_14_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_14_lse_sub = reduce_log_sum_exp(axes = logits_chunk_14_lse_sub_axes_0, keep_dims = logits_chunk_14_lse_sub_keep_dims_0, x = logits_chunk_14_sub)[name = string("logits_chunk_14_lse_sub")];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_14_lse = add(x = logits_chunk_14_lse_sub, y = logits_chunk_14_max)[name = string("logits_chunk_14_lse")];
+            tensor<fp16, [2048, 2048, 1, 1]> logits_chunk_15_weight_0 = const()[name = string("logits_chunk_15_weight_0"), val = tensor<fp16, [2048, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125834304)))];
+            tensor<int32, [2]> logits_chunk_15_strides_0 = const()[name = string("logits_chunk_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            string logits_chunk_15_pad_type_0 = const()[name = string("logits_chunk_15_pad_type_0"), val = string("valid")];
+            tensor<int32, [4]> logits_chunk_15_pad_0 = const()[name = string("logits_chunk_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> logits_chunk_15_dilations_0 = const()[name = string("logits_chunk_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 logits_chunk_15_groups_0 = const()[name = string("logits_chunk_15_groups_0"), val = int32(1)];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_15 = conv(dilations = logits_chunk_15_dilations_0, groups = logits_chunk_15_groups_0, pad = logits_chunk_15_pad_0, pad_type = logits_chunk_15_pad_type_0, strides = logits_chunk_15_strides_0, weight = logits_chunk_15_weight_0, x = final_norm_rmsnorm)[name = string("logits_chunk_15")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_15_mul = mul(x = logits_chunk_15, y = temp_inverse)[name = string("logits_chunk_15_mul")];
+            tensor<int32, [1]> logits_chunk_15_max_axes_0 = const()[name = string("logits_chunk_15_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_15_max_keep_dims_0 = const()[name = string("logits_chunk_15_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_15_max = reduce_max(axes = logits_chunk_15_max_axes_0, keep_dims = logits_chunk_15_max_keep_dims_0, x = logits_chunk_15_mul)[name = string("logits_chunk_15_max")];
+            int32 logits_chunk_15_argmax_axis_0 = const()[name = string("logits_chunk_15_argmax_axis_0"), val = int32(1)];
+            bool logits_chunk_15_argmax_keep_dims_0 = const()[name = string("logits_chunk_15_argmax_keep_dims_0"), val = bool(true)];
+            string logits_chunk_15_argmax_output_dtype_0 = const()[name = string("logits_chunk_15_argmax_output_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 1, 1, 8]> logits_chunk_15_argmax = reduce_argmax(axis = logits_chunk_15_argmax_axis_0, keep_dims = logits_chunk_15_argmax_keep_dims_0, output_dtype = logits_chunk_15_argmax_output_dtype_0, x = logits_chunk_15_mul)[name = string("logits_chunk_15_argmax")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_15_sub = sub(x = logits_chunk_15_mul, y = logits_chunk_15_max)[name = string("logits_chunk_15_sub")];
+            tensor<int32, [1]> logits_chunk_15_lse_sub_axes_0 = const()[name = string("logits_chunk_15_lse_sub_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_chunk_15_lse_sub_keep_dims_0 = const()[name = string("logits_chunk_15_lse_sub_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_15_lse_sub = reduce_log_sum_exp(axes = logits_chunk_15_lse_sub_axes_0, keep_dims = logits_chunk_15_lse_sub_keep_dims_0, x = logits_chunk_15_sub)[name = string("logits_chunk_15_lse_sub")];
+            tensor<fp16, [1, 1, 1, 8]> logits_chunk_15_lse = add(x = logits_chunk_15_lse_sub, y = logits_chunk_15_max)[name = string("logits_chunk_15_lse")];
+            int32 logits_lses_axis_0 = const()[name = string("logits_lses_axis_0"), val = int32(1)];
+            bool logits_lses_interleave_0 = const()[name = string("logits_lses_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 8]> logits_lses = concat(axis = logits_lses_axis_0, interleave = logits_lses_interleave_0, values = (logits_chunk_0_lse, logits_chunk_1_lse, logits_chunk_2_lse, logits_chunk_3_lse, logits_chunk_4_lse, logits_chunk_5_lse, logits_chunk_6_lse, logits_chunk_7_lse, logits_chunk_8_lse, logits_chunk_9_lse, logits_chunk_10_lse, logits_chunk_11_lse, logits_chunk_12_lse, logits_chunk_13_lse, logits_chunk_14_lse, logits_chunk_15_lse))[name = string("logits_lses")];
+            tensor<int32, [1]> logits_lses_max_axes_0 = const()[name = string("logits_lses_max_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_lses_max_keep_dims_0 = const()[name = string("logits_lses_max_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_lses_max = reduce_max(axes = logits_lses_max_axes_0, keep_dims = logits_lses_max_keep_dims_0, x = logits_lses)[name = string("logits_lses_max")];
+            tensor<fp16, [1, 16, 1, 8]> logits_lses_sub = sub(x = logits_lses, y = logits_lses_max)[name = string("logits_lses_sub")];
+            tensor<int32, [1]> logits_lses_logsumexp_axes_0 = const()[name = string("logits_lses_logsumexp_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_lses_logsumexp_keep_dims_0 = const()[name = string("logits_lses_logsumexp_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_lses_logsumexp = reduce_log_sum_exp(axes = logits_lses_logsumexp_axes_0, keep_dims = logits_lses_logsumexp_keep_dims_0, x = logits_lses_sub)[name = string("logits_lses_logsumexp")];
+            tensor<fp16, [1, 1, 1, 8]> logits_lse = add(x = logits_lses_logsumexp, y = logits_lses_max)[name = string("logits_lse")];
+            int32 logits_max_logits_chunks_axis_0 = const()[name = string("logits_max_logits_chunks_axis_0"), val = int32(1)];
+            bool logits_max_logits_chunks_interleave_0 = const()[name = string("logits_max_logits_chunks_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 16, 1, 8]> logits_max_logits_chunks = concat(axis = logits_max_logits_chunks_axis_0, interleave = logits_max_logits_chunks_interleave_0, values = (logits_chunk_0_max, logits_chunk_1_max, logits_chunk_2_max, logits_chunk_3_max, logits_chunk_4_max, logits_chunk_5_max, logits_chunk_6_max, logits_chunk_7_max, logits_chunk_8_max, logits_chunk_9_max, logits_chunk_10_max, logits_chunk_11_max, logits_chunk_12_max, logits_chunk_13_max, logits_chunk_14_max, logits_chunk_15_max))[name = string("logits_max_logits_chunks")];
+            tensor<int32, [1]> logits_max_logit_axes_0 = const()[name = string("logits_max_logit_axes_0"), val = tensor<int32, [1]>([1])];
+            bool logits_max_logit_keep_dims_0 = const()[name = string("logits_max_logit_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 1, 1, 8]> logits_max_logit = reduce_max(axes = logits_max_logit_axes_0, keep_dims = logits_max_logit_keep_dims_0, x = logits_max_logits_chunks)[name = string("logits_max_logit")];
+            tensor<fp16, [1, 1, 1, 8]> logits_max_logit_sub = sub(x = logits_max_logit, y = logits_lse)[name = string("logits_max_logit_sub")];
+            tensor<fp16, [1, 1, 1, 8]> max_prob = exp(x = logits_max_logit_sub)[name = string("max_prob")];
+            tensor<fp16, [1, 1, 1, 8]> min_p_thresh = mul(x = max_prob, y = p)[name = string("min_p_thresh")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_0_sub_1 = sub(x = logits_chunk_0_mul, y = logits_lse)[name = string("logits_chunk_0_sub_1")];
+            tensor<fp16, [1, 2048, 1, 8]> probs_chunk_0 = exp(x = logits_chunk_0_sub_1)[name = string("probs_chunk_0")];
+            tensor<bool, [1, 2048, 1, 8]> mask_probs_chunk_0 = greater_equal(x = probs_chunk_0, y = min_p_thresh)[name = string("mask_probs_chunk_0")];
+            string mask_chunk_0_fp16_dtype_0 = const()[name = string("mask_chunk_0_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 8]> mask_chunk_0_fp16 = cast(dtype = mask_chunk_0_fp16_dtype_0, x = mask_probs_chunk_0)[name = string("cast_101")];
+            tensor<fp16, [1, 2048, 1, 8]> masked_probs_chunk_0 = select(a = probs_chunk_0, b = mask_chunk_0_fp16, cond = mask_probs_chunk_0)[name = string("masked_probs_chunk_0")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_1_sub_1 = sub(x = logits_chunk_1_mul, y = logits_lse)[name = string("logits_chunk_1_sub_1")];
+            tensor<fp16, [1, 2048, 1, 8]> probs_chunk_1 = exp(x = logits_chunk_1_sub_1)[name = string("probs_chunk_1")];
+            tensor<bool, [1, 2048, 1, 8]> mask_probs_chunk_1 = greater_equal(x = probs_chunk_1, y = min_p_thresh)[name = string("mask_probs_chunk_1")];
+            string mask_chunk_1_fp16_dtype_0 = const()[name = string("mask_chunk_1_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 8]> mask_chunk_1_fp16 = cast(dtype = mask_chunk_1_fp16_dtype_0, x = mask_probs_chunk_1)[name = string("cast_100")];
+            tensor<fp16, [1, 2048, 1, 8]> masked_probs_chunk_1 = select(a = probs_chunk_1, b = mask_chunk_1_fp16, cond = mask_probs_chunk_1)[name = string("masked_probs_chunk_1")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_2_sub_1 = sub(x = logits_chunk_2_mul, y = logits_lse)[name = string("logits_chunk_2_sub_1")];
+            tensor<fp16, [1, 2048, 1, 8]> probs_chunk_2 = exp(x = logits_chunk_2_sub_1)[name = string("probs_chunk_2")];
+            tensor<bool, [1, 2048, 1, 8]> mask_probs_chunk_2 = greater_equal(x = probs_chunk_2, y = min_p_thresh)[name = string("mask_probs_chunk_2")];
+            string mask_chunk_2_fp16_dtype_0 = const()[name = string("mask_chunk_2_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 8]> mask_chunk_2_fp16 = cast(dtype = mask_chunk_2_fp16_dtype_0, x = mask_probs_chunk_2)[name = string("cast_99")];
+            tensor<fp16, [1, 2048, 1, 8]> masked_probs_chunk_2 = select(a = probs_chunk_2, b = mask_chunk_2_fp16, cond = mask_probs_chunk_2)[name = string("masked_probs_chunk_2")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_3_sub_1 = sub(x = logits_chunk_3_mul, y = logits_lse)[name = string("logits_chunk_3_sub_1")];
+            tensor<fp16, [1, 2048, 1, 8]> probs_chunk_3 = exp(x = logits_chunk_3_sub_1)[name = string("probs_chunk_3")];
+            tensor<bool, [1, 2048, 1, 8]> mask_probs_chunk_3 = greater_equal(x = probs_chunk_3, y = min_p_thresh)[name = string("mask_probs_chunk_3")];
+            string mask_chunk_3_fp16_dtype_0 = const()[name = string("mask_chunk_3_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 8]> mask_chunk_3_fp16 = cast(dtype = mask_chunk_3_fp16_dtype_0, x = mask_probs_chunk_3)[name = string("cast_98")];
+            tensor<fp16, [1, 2048, 1, 8]> masked_probs_chunk_3 = select(a = probs_chunk_3, b = mask_chunk_3_fp16, cond = mask_probs_chunk_3)[name = string("masked_probs_chunk_3")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_4_sub_1 = sub(x = logits_chunk_4_mul, y = logits_lse)[name = string("logits_chunk_4_sub_1")];
+            tensor<fp16, [1, 2048, 1, 8]> probs_chunk_4 = exp(x = logits_chunk_4_sub_1)[name = string("probs_chunk_4")];
+            tensor<bool, [1, 2048, 1, 8]> mask_probs_chunk_4 = greater_equal(x = probs_chunk_4, y = min_p_thresh)[name = string("mask_probs_chunk_4")];
+            string mask_chunk_4_fp16_dtype_0 = const()[name = string("mask_chunk_4_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 8]> mask_chunk_4_fp16 = cast(dtype = mask_chunk_4_fp16_dtype_0, x = mask_probs_chunk_4)[name = string("cast_97")];
+            tensor<fp16, [1, 2048, 1, 8]> masked_probs_chunk_4 = select(a = probs_chunk_4, b = mask_chunk_4_fp16, cond = mask_probs_chunk_4)[name = string("masked_probs_chunk_4")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_5_sub_1 = sub(x = logits_chunk_5_mul, y = logits_lse)[name = string("logits_chunk_5_sub_1")];
+            tensor<fp16, [1, 2048, 1, 8]> probs_chunk_5 = exp(x = logits_chunk_5_sub_1)[name = string("probs_chunk_5")];
+            tensor<bool, [1, 2048, 1, 8]> mask_probs_chunk_5 = greater_equal(x = probs_chunk_5, y = min_p_thresh)[name = string("mask_probs_chunk_5")];
+            string mask_chunk_5_fp16_dtype_0 = const()[name = string("mask_chunk_5_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 8]> mask_chunk_5_fp16 = cast(dtype = mask_chunk_5_fp16_dtype_0, x = mask_probs_chunk_5)[name = string("cast_96")];
+            tensor<fp16, [1, 2048, 1, 8]> masked_probs_chunk_5 = select(a = probs_chunk_5, b = mask_chunk_5_fp16, cond = mask_probs_chunk_5)[name = string("masked_probs_chunk_5")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_6_sub_1 = sub(x = logits_chunk_6_mul, y = logits_lse)[name = string("logits_chunk_6_sub_1")];
+            tensor<fp16, [1, 2048, 1, 8]> probs_chunk_6 = exp(x = logits_chunk_6_sub_1)[name = string("probs_chunk_6")];
+            tensor<bool, [1, 2048, 1, 8]> mask_probs_chunk_6 = greater_equal(x = probs_chunk_6, y = min_p_thresh)[name = string("mask_probs_chunk_6")];
+            string mask_chunk_6_fp16_dtype_0 = const()[name = string("mask_chunk_6_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 8]> mask_chunk_6_fp16 = cast(dtype = mask_chunk_6_fp16_dtype_0, x = mask_probs_chunk_6)[name = string("cast_95")];
+            tensor<fp16, [1, 2048, 1, 8]> masked_probs_chunk_6 = select(a = probs_chunk_6, b = mask_chunk_6_fp16, cond = mask_probs_chunk_6)[name = string("masked_probs_chunk_6")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_7_sub_1 = sub(x = logits_chunk_7_mul, y = logits_lse)[name = string("logits_chunk_7_sub_1")];
+            tensor<fp16, [1, 2048, 1, 8]> probs_chunk_7 = exp(x = logits_chunk_7_sub_1)[name = string("probs_chunk_7")];
+            tensor<bool, [1, 2048, 1, 8]> mask_probs_chunk_7 = greater_equal(x = probs_chunk_7, y = min_p_thresh)[name = string("mask_probs_chunk_7")];
+            string mask_chunk_7_fp16_dtype_0 = const()[name = string("mask_chunk_7_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 8]> mask_chunk_7_fp16 = cast(dtype = mask_chunk_7_fp16_dtype_0, x = mask_probs_chunk_7)[name = string("cast_94")];
+            tensor<fp16, [1, 2048, 1, 8]> masked_probs_chunk_7 = select(a = probs_chunk_7, b = mask_chunk_7_fp16, cond = mask_probs_chunk_7)[name = string("masked_probs_chunk_7")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_8_sub_1 = sub(x = logits_chunk_8_mul, y = logits_lse)[name = string("logits_chunk_8_sub_1")];
+            tensor<fp16, [1, 2048, 1, 8]> probs_chunk_8 = exp(x = logits_chunk_8_sub_1)[name = string("probs_chunk_8")];
+            tensor<bool, [1, 2048, 1, 8]> mask_probs_chunk_8 = greater_equal(x = probs_chunk_8, y = min_p_thresh)[name = string("mask_probs_chunk_8")];
+            string mask_chunk_8_fp16_dtype_0 = const()[name = string("mask_chunk_8_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 8]> mask_chunk_8_fp16 = cast(dtype = mask_chunk_8_fp16_dtype_0, x = mask_probs_chunk_8)[name = string("cast_93")];
+            tensor<fp16, [1, 2048, 1, 8]> masked_probs_chunk_8 = select(a = probs_chunk_8, b = mask_chunk_8_fp16, cond = mask_probs_chunk_8)[name = string("masked_probs_chunk_8")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_9_sub_1 = sub(x = logits_chunk_9_mul, y = logits_lse)[name = string("logits_chunk_9_sub_1")];
+            tensor<fp16, [1, 2048, 1, 8]> probs_chunk_9 = exp(x = logits_chunk_9_sub_1)[name = string("probs_chunk_9")];
+            tensor<bool, [1, 2048, 1, 8]> mask_probs_chunk_9 = greater_equal(x = probs_chunk_9, y = min_p_thresh)[name = string("mask_probs_chunk_9")];
+            string mask_chunk_9_fp16_dtype_0 = const()[name = string("mask_chunk_9_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 8]> mask_chunk_9_fp16 = cast(dtype = mask_chunk_9_fp16_dtype_0, x = mask_probs_chunk_9)[name = string("cast_92")];
+            tensor<fp16, [1, 2048, 1, 8]> masked_probs_chunk_9 = select(a = probs_chunk_9, b = mask_chunk_9_fp16, cond = mask_probs_chunk_9)[name = string("masked_probs_chunk_9")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_10_sub_1 = sub(x = logits_chunk_10_mul, y = logits_lse)[name = string("logits_chunk_10_sub_1")];
+            tensor<fp16, [1, 2048, 1, 8]> probs_chunk_10 = exp(x = logits_chunk_10_sub_1)[name = string("probs_chunk_10")];
+            tensor<bool, [1, 2048, 1, 8]> mask_probs_chunk_10 = greater_equal(x = probs_chunk_10, y = min_p_thresh)[name = string("mask_probs_chunk_10")];
+            string mask_chunk_10_fp16_dtype_0 = const()[name = string("mask_chunk_10_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 8]> mask_chunk_10_fp16 = cast(dtype = mask_chunk_10_fp16_dtype_0, x = mask_probs_chunk_10)[name = string("cast_91")];
+            tensor<fp16, [1, 2048, 1, 8]> masked_probs_chunk_10 = select(a = probs_chunk_10, b = mask_chunk_10_fp16, cond = mask_probs_chunk_10)[name = string("masked_probs_chunk_10")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_11_sub_1 = sub(x = logits_chunk_11_mul, y = logits_lse)[name = string("logits_chunk_11_sub_1")];
+            tensor<fp16, [1, 2048, 1, 8]> probs_chunk_11 = exp(x = logits_chunk_11_sub_1)[name = string("probs_chunk_11")];
+            tensor<bool, [1, 2048, 1, 8]> mask_probs_chunk_11 = greater_equal(x = probs_chunk_11, y = min_p_thresh)[name = string("mask_probs_chunk_11")];
+            string mask_chunk_11_fp16_dtype_0 = const()[name = string("mask_chunk_11_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 8]> mask_chunk_11_fp16 = cast(dtype = mask_chunk_11_fp16_dtype_0, x = mask_probs_chunk_11)[name = string("cast_90")];
+            tensor<fp16, [1, 2048, 1, 8]> masked_probs_chunk_11 = select(a = probs_chunk_11, b = mask_chunk_11_fp16, cond = mask_probs_chunk_11)[name = string("masked_probs_chunk_11")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_12_sub_1 = sub(x = logits_chunk_12_mul, y = logits_lse)[name = string("logits_chunk_12_sub_1")];
+            tensor<fp16, [1, 2048, 1, 8]> probs_chunk_12 = exp(x = logits_chunk_12_sub_1)[name = string("probs_chunk_12")];
+            tensor<bool, [1, 2048, 1, 8]> mask_probs_chunk_12 = greater_equal(x = probs_chunk_12, y = min_p_thresh)[name = string("mask_probs_chunk_12")];
+            string mask_chunk_12_fp16_dtype_0 = const()[name = string("mask_chunk_12_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 8]> mask_chunk_12_fp16 = cast(dtype = mask_chunk_12_fp16_dtype_0, x = mask_probs_chunk_12)[name = string("cast_89")];
+            tensor<fp16, [1, 2048, 1, 8]> masked_probs_chunk_12 = select(a = probs_chunk_12, b = mask_chunk_12_fp16, cond = mask_probs_chunk_12)[name = string("masked_probs_chunk_12")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_13_sub_1 = sub(x = logits_chunk_13_mul, y = logits_lse)[name = string("logits_chunk_13_sub_1")];
+            tensor<fp16, [1, 2048, 1, 8]> probs_chunk_13 = exp(x = logits_chunk_13_sub_1)[name = string("probs_chunk_13")];
+            tensor<bool, [1, 2048, 1, 8]> mask_probs_chunk_13 = greater_equal(x = probs_chunk_13, y = min_p_thresh)[name = string("mask_probs_chunk_13")];
+            string mask_chunk_13_fp16_dtype_0 = const()[name = string("mask_chunk_13_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 8]> mask_chunk_13_fp16 = cast(dtype = mask_chunk_13_fp16_dtype_0, x = mask_probs_chunk_13)[name = string("cast_88")];
+            tensor<fp16, [1, 2048, 1, 8]> masked_probs_chunk_13 = select(a = probs_chunk_13, b = mask_chunk_13_fp16, cond = mask_probs_chunk_13)[name = string("masked_probs_chunk_13")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_14_sub_1 = sub(x = logits_chunk_14_mul, y = logits_lse)[name = string("logits_chunk_14_sub_1")];
+            tensor<fp16, [1, 2048, 1, 8]> probs_chunk_14 = exp(x = logits_chunk_14_sub_1)[name = string("probs_chunk_14")];
+            tensor<bool, [1, 2048, 1, 8]> mask_probs_chunk_14 = greater_equal(x = probs_chunk_14, y = min_p_thresh)[name = string("mask_probs_chunk_14")];
+            string mask_chunk_14_fp16_dtype_0 = const()[name = string("mask_chunk_14_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 8]> mask_chunk_14_fp16 = cast(dtype = mask_chunk_14_fp16_dtype_0, x = mask_probs_chunk_14)[name = string("cast_87")];
+            tensor<fp16, [1, 2048, 1, 8]> masked_probs_chunk_14 = select(a = probs_chunk_14, b = mask_chunk_14_fp16, cond = mask_probs_chunk_14)[name = string("masked_probs_chunk_14")];
+            tensor<fp16, [1, 2048, 1, 8]> logits_chunk_15_sub_1 = sub(x = logits_chunk_15_mul, y = logits_lse)[name = string("logits_chunk_15_sub_1")];
+            tensor<fp16, [1, 2048, 1, 8]> probs_chunk_15 = exp(x = logits_chunk_15_sub_1)[name = string("probs_chunk_15")];
+            tensor<bool, [1, 2048, 1, 8]> mask_probs_chunk_15 = greater_equal(x = probs_chunk_15, y = min_p_thresh)[name = string("mask_probs_chunk_15")];
+            string mask_chunk_15_fp16_dtype_0 = const()[name = string("mask_chunk_15_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 2048, 1, 8]> mask_chunk_15_fp16 = cast(dtype = mask_chunk_15_fp16_dtype_0, x = mask_probs_chunk_15)[name = string("cast_86")];
+            tensor<fp16, [1, 2048, 1, 8]> masked_probs_chunk_15 = select(a = probs_chunk_15, b = mask_chunk_15_fp16, cond = mask_probs_chunk_15)[name = string("masked_probs_chunk_15")];
+            int32 probs_axis_0 = const()[name = string("probs_axis_0"), val = int32(1)];
+            bool probs_interleave_0 = const()[name = string("probs_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 32768, 1, 8]> probs = concat(axis = probs_axis_0, interleave = probs_interleave_0, values = (masked_probs_chunk_0, masked_probs_chunk_1, masked_probs_chunk_2, masked_probs_chunk_3, masked_probs_chunk_4, masked_probs_chunk_5, masked_probs_chunk_6, masked_probs_chunk_7, masked_probs_chunk_8, masked_probs_chunk_9, masked_probs_chunk_10, masked_probs_chunk_11, masked_probs_chunk_12, masked_probs_chunk_13, masked_probs_chunk_14, masked_probs_chunk_15))[name = string("probs")];
+            string probs_fp32_dtype_0 = const()[name = string("probs_fp32_dtype_0"), val = string("fp32")];
+            int32 probs_cumsum_axis_0 = const()[name = string("probs_cumsum_axis_0"), val = int32(1)];
+            bool probs_cumsum_exclusive_0 = const()[name = string("probs_cumsum_exclusive_0"), val = bool(false)];
+            bool probs_cumsum_reverse_0 = const()[name = string("probs_cumsum_reverse_0"), val = bool(false)];
+            tensor<fp32, [1, 32768, 1, 8]> probs_fp32 = cast(dtype = probs_fp32_dtype_0, x = probs)[name = string("cast_85")];
+            tensor<fp32, [1, 32768, 1, 8]> probs_cumsum = cumsum(axis = probs_cumsum_axis_0, exclusive = probs_cumsum_exclusive_0, reverse = probs_cumsum_reverse_0, x = probs_fp32)[name = string("probs_cumsum")];
+            tensor<int32, [1]> probs_sum_indices_0 = const()[name = string("probs_sum_indices_0"), val = tensor<int32, [1]>([32767])];
+            int32 probs_sum_axis_0 = const()[name = string("probs_sum_axis_0"), val = int32(1)];
+            int32 probs_sum_batch_dims_0 = const()[name = string("probs_sum_batch_dims_0"), val = int32(0)];
+            bool probs_sum_validate_indices_0 = const()[name = string("probs_sum_validate_indices_0"), val = bool(false)];
+            tensor<fp32, [1, 1, 1, 8]> probs_sum = gather(axis = probs_sum_axis_0, batch_dims = probs_sum_batch_dims_0, indices = probs_sum_indices_0, validate_indices = probs_sum_validate_indices_0, x = probs_cumsum)[name = string("probs_sum")];
+            tensor<fp32, [1, 1, 1, 8]> random_number_scaled = mul(x = random_number, y = probs_sum)[name = string("random_number_scaled")];
+            tensor<bool, [1, 32768, 1, 8]> probs_greater = greater(x = probs_cumsum, y = random_number_scaled)[name = string("probs_greater")];
+            string probs_greater_int32_dtype_0 = const()[name = string("probs_greater_int32_dtype_0"), val = string("int32")];
+            int32 sampled_index_axis_0 = const()[name = string("sampled_index_axis_0"), val = int32(1)];
+            bool sampled_index_keep_dims_0 = const()[name = string("sampled_index_keep_dims_0"), val = bool(true)];
+            string sampled_index_output_dtype_0 = const()[name = string("sampled_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 32768, 1, 8]> probs_greater_int32 = cast(dtype = probs_greater_int32_dtype_0, x = probs_greater)[name = string("cast_84")];
+            tensor<int32, [1, 1, 1, 8]> sampled_index = reduce_argmax(axis = sampled_index_axis_0, keep_dims = sampled_index_keep_dims_0, output_dtype = sampled_index_output_dtype_0, x = probs_greater_int32)[name = string("sampled_index")];
+            int32 sampled_index_probability_axis_0 = const()[name = string("sampled_index_probability_axis_0"), val = int32(1)];
+            bool sampled_index_probability_validate_indices_0 = const()[name = string("sampled_index_probability_validate_indices_0"), val = bool(false)];
+            tensor<fp32, [1, 1, 1, 8]> sampled_index_probability = gather_along_axis(axis = sampled_index_probability_axis_0, indices = sampled_index, validate_indices = sampled_index_probability_validate_indices_0, x = probs_fp32)[name = string("sampled_index_probability")];
+            int32 max_logit_index_axis_0 = const()[name = string("max_logit_index_axis_0"), val = int32(1)];
+            bool max_logit_index_keep_dims_0 = const()[name = string("max_logit_index_keep_dims_0"), val = bool(true)];
+            string max_logit_index_output_dtype_0 = const()[name = string("max_logit_index_output_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 1, 1, 8]> max_logit_index = reduce_argmax(axis = max_logit_index_axis_0, keep_dims = max_logit_index_keep_dims_0, output_dtype = max_logit_index_output_dtype_0, x = logits_max_logits_chunks)[name = string("max_logit_index")];
+            string indices_chunk_0_int32_dtype_0 = const()[name = string("indices_chunk_0_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_1_int32_dtype_0 = const()[name = string("indices_chunk_1_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_2_int32_dtype_0 = const()[name = string("indices_chunk_2_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_3_int32_dtype_0 = const()[name = string("indices_chunk_3_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_4_int32_dtype_0 = const()[name = string("indices_chunk_4_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_5_int32_dtype_0 = const()[name = string("indices_chunk_5_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_6_int32_dtype_0 = const()[name = string("indices_chunk_6_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_7_int32_dtype_0 = const()[name = string("indices_chunk_7_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_8_int32_dtype_0 = const()[name = string("indices_chunk_8_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_9_int32_dtype_0 = const()[name = string("indices_chunk_9_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_10_int32_dtype_0 = const()[name = string("indices_chunk_10_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_11_int32_dtype_0 = const()[name = string("indices_chunk_11_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_12_int32_dtype_0 = const()[name = string("indices_chunk_12_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_13_int32_dtype_0 = const()[name = string("indices_chunk_13_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_14_int32_dtype_0 = const()[name = string("indices_chunk_14_int32_dtype_0"), val = string("int32")];
+            string indices_chunk_15_int32_dtype_0 = const()[name = string("indices_chunk_15_int32_dtype_0"), val = string("int32")];
+            int32 indices_axis_0 = const()[name = string("indices_axis_0"), val = int32(1)];
+            bool indices_interleave_0 = const()[name = string("indices_interleave_0"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_15_int32 = cast(dtype = indices_chunk_15_int32_dtype_0, x = logits_chunk_15_argmax)[name = string("cast_68")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_14_int32 = cast(dtype = indices_chunk_14_int32_dtype_0, x = logits_chunk_14_argmax)[name = string("cast_69")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_13_int32 = cast(dtype = indices_chunk_13_int32_dtype_0, x = logits_chunk_13_argmax)[name = string("cast_70")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_12_int32 = cast(dtype = indices_chunk_12_int32_dtype_0, x = logits_chunk_12_argmax)[name = string("cast_71")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_11_int32 = cast(dtype = indices_chunk_11_int32_dtype_0, x = logits_chunk_11_argmax)[name = string("cast_72")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_10_int32 = cast(dtype = indices_chunk_10_int32_dtype_0, x = logits_chunk_10_argmax)[name = string("cast_73")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_9_int32 = cast(dtype = indices_chunk_9_int32_dtype_0, x = logits_chunk_9_argmax)[name = string("cast_74")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_8_int32 = cast(dtype = indices_chunk_8_int32_dtype_0, x = logits_chunk_8_argmax)[name = string("cast_75")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_7_int32 = cast(dtype = indices_chunk_7_int32_dtype_0, x = logits_chunk_7_argmax)[name = string("cast_76")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_6_int32 = cast(dtype = indices_chunk_6_int32_dtype_0, x = logits_chunk_6_argmax)[name = string("cast_77")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_5_int32 = cast(dtype = indices_chunk_5_int32_dtype_0, x = logits_chunk_5_argmax)[name = string("cast_78")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_4_int32 = cast(dtype = indices_chunk_4_int32_dtype_0, x = logits_chunk_4_argmax)[name = string("cast_79")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_3_int32 = cast(dtype = indices_chunk_3_int32_dtype_0, x = logits_chunk_3_argmax)[name = string("cast_80")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_2_int32 = cast(dtype = indices_chunk_2_int32_dtype_0, x = logits_chunk_2_argmax)[name = string("cast_81")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_1_int32 = cast(dtype = indices_chunk_1_int32_dtype_0, x = logits_chunk_1_argmax)[name = string("cast_82")];
+            tensor<int32, [1, 1, 1, 8]> indices_chunk_0_int32 = cast(dtype = indices_chunk_0_int32_dtype_0, x = logits_chunk_0_argmax)[name = string("cast_83")];
+            tensor<int32, [1, 16, 1, 8]> indices = concat(axis = indices_axis_0, interleave = indices_interleave_0, values = (indices_chunk_0_int32, indices_chunk_1_int32, indices_chunk_2_int32, indices_chunk_3_int32, indices_chunk_4_int32, indices_chunk_5_int32, indices_chunk_6_int32, indices_chunk_7_int32, indices_chunk_8_int32, indices_chunk_9_int32, indices_chunk_10_int32, indices_chunk_11_int32, indices_chunk_12_int32, indices_chunk_13_int32, indices_chunk_14_int32, indices_chunk_15_int32))[name = string("indices")];
+            int32 argmax_chunks_axis_0 = const()[name = string("argmax_chunks_axis_0"), val = int32(1)];
+            bool argmax_chunks_validate_indices_0 = const()[name = string("argmax_chunks_validate_indices_0"), val = bool(false)];
+            tensor<int32, [1, 1, 1, 8]> argmax_chunks = gather_along_axis(axis = argmax_chunks_axis_0, indices = max_logit_index, validate_indices = argmax_chunks_validate_indices_0, x = indices)[name = string("argmax_chunks")];
+            int32 mul_0_x_0 = const()[name = string("mul_0_x_0"), val = int32(2048)];
+            tensor<int32, [1, 1, 1, 8]> mul_0 = mul(x = mul_0_x_0, y = max_logit_index)[name = string("mul_0")];
+            tensor<int32, [1, 1, 1, 8]> argmax = add(x = argmax_chunks, y = mul_0)[name = string("argmax")];
+        } -> (sampled_index, sampled_index_probability, argmax, max_prob);
+}
\ No newline at end of file