initial commit

base/decoder_first.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3071377562292da4d34bf9d0ddcfe168fd10c3b81d4689d25c207179d2d58578
+size 243

base/decoder_first.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0fbe1879a296cf22a0441826a3028ae2ec63bfc8e9ff019132681d2a93610324
+size 453

base/decoder_first.mlmodelc/metadata.json

@@ -0,0 +1,106 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16)",
+        "shortDescription" : "",
+        "shape" : "[]",
+        "name" : "dummy",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.writeState" : 14,
+      "Shape" : 12,
+      "Ios18.linear" : 12,
+      "Identity" : 1,
+      "Ios18.gather" : 12,
+      "Ios18.concat" : 12,
+      "Ios18.sliceUpdate" : 14,
+      "Ios18.cast" : 24,
+      "Ios18.expandDims" : 12,
+      "Ios18.readState" : 14
+    },
+    "computePrecision" : "Mixed (Float16, Int16, Int32, UInt16)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 6 × 1 × 448 × 512)",
+        "shortDescription" : "",
+        "shape" : "[6, 1, 448, 512]",
+        "name" : "k_cache1",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 6 × 1 × 448 × 512)",
+        "shortDescription" : "",
+        "shape" : "[6, 1, 448, 512]",
+        "name" : "v_cache1",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 6 × 1 × 1500 × 512)",
+        "shortDescription" : "",
+        "shape" : "[6, 1, 1500, 512]",
+        "name" : "k_cache2",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 6 × 1 × 1500 × 512)",
+        "shortDescription" : "",
+        "shape" : "[6, 1, 1500, 512]",
+        "name" : "v_cache2",
+        "type" : "State"
+      }
+    ],
+    "availability" : {
+      "macOS" : "15.0",
+      "tvOS" : "18.0",
+      "visionOS" : "2.0",
+      "watchOS" : "11.0",
+      "iOS" : "18.0",
+      "macCatalyst" : "18.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.4.1",
+      "com.github.apple.coremltools.version" : "8.0"
+    },
+    "inputSchema" : [
+      {
+        "dataType" : "Float16",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1 × 1...1500 × 512",
+        "shapeRange" : "[[1, 1], [1, 1500], [512, 512]]",
+        "formattedType" : "MultiArray (Float16 1 × 1 × 512)",
+        "type" : "MultiArray",
+        "shape" : "[1, 1, 512]",
+        "name" : "audio_data",
+        "shortDescription" : ""
+      }
+    ],
+    "generatedClassName" : "decoder_first",
+    "method" : "predict"
+  }
+]

base/decoder_first.mlmodelc/model.mil

@@ -0,0 +1,369 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3400.43.1"}, {"coremlc-version", "3400.58.2"}, {"coremltools-component-torch", "2.4.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.0"}})]
+{
+    func main<ios18>(tensor<fp16, [1, ?, 512]> audio_data, state<tensor<fp16, [6, 1, 448, 512]>> k_cache1, state<tensor<fp16, [6, 1, 1500, 512]>> k_cache2, state<tensor<fp16, [6, 1, 448, 512]>> v_cache1, state<tensor<fp16, [6, 1, 1500, 512]>> v_cache2) [FlexibleShapeInformation = tuple<tuple<string, dict<string, tensor<int32, [?]>>>, tuple<string, dict<string, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"audio_data", [1, 1, 512]}}), ("RangeDims", {{"audio_data", [[1, 1], [1, 1500], [512, 512]]}})))] {
+            tensor<fp16, [1, ?, 512]> dummy = identity(x = audio_data)[name = string("identity_0")];
+            tensor<fp16, [6, 1, 448, 512]> read_state_0 = read_state(input = k_cache1)[name = string("read_state_0")];
+            tensor<int32, [4]> concat_0 = const()[name = string("concat_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> concat_1 = const()[name = string("concat_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> k_cache1_internal_tensor_assign_1_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> k_cache1_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> k_cache1_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<bool, [4]> k_cache1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [6, 1, 448, 512]> const_0_to_fp16 = const()[name = string("const_0_to_fp16"), val = tensor<fp16, [6, 1, 448, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [6, 1, 448, 512]> k_cache1_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_0, begin_mask = k_cache1_internal_tensor_assign_1_begin_mask_0, end = concat_1, end_mask = k_cache1_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_1_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_1_stride_0, update = const_0_to_fp16, x = read_state_0)[name = string("k_cache1_internal_tensor_assign_1_cast_fp16")];
+            write_state(data = k_cache1_internal_tensor_assign_1_cast_fp16, input = k_cache1)[name = string("coreml_update_state_14_write_state")];
+            tensor<fp16, [6, 1, 448, 512]> read_state_1 = read_state(input = v_cache1)[name = string("read_state_1")];
+            tensor<int32, [4]> concat_2 = const()[name = string("concat_2"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> concat_3 = const()[name = string("concat_3"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> v_cache1_internal_tensor_assign_1_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> v_cache1_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> v_cache1_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<bool, [4]> v_cache1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<fp16, [6, 1, 448, 512]> v_cache1_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_2, begin_mask = v_cache1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = v_cache1_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_1_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_1_stride_0, update = const_0_to_fp16, x = read_state_1)[name = string("v_cache1_internal_tensor_assign_1_cast_fp16")];
+            write_state(data = v_cache1_internal_tensor_assign_1_cast_fp16, input = v_cache1)[name = string("coreml_update_state_15_write_state")];
+            tensor<fp16, [6, 1, 1500, 512]> read_state_2 = read_state(input = k_cache2)[name = string("read_state_2")];
+            tensor<fp16, [6, 1, 1500, 512]> read_state_3 = read_state(input = v_cache2)[name = string("read_state_3")];
+            tensor<fp16, [512, 512]> var_79_to_fp16 = const()[name = string("op_79_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2752640)))];
+            tensor<fp16, [512]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3276992)))];
+            tensor<fp16, [1, ?, 512]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = var_79_to_fp16, x = audio_data)[name = string("linear_0_cast_fp16")];
+            tensor<fp16, [512, 512]> var_83_to_fp16 = const()[name = string("op_83_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3278080)))];
+            tensor<fp16, [512]> var_84_to_fp16 = const()[name = string("op_84_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3802432)))];
+            tensor<fp16, [1, ?, 512]> linear_1_cast_fp16 = linear(bias = var_84_to_fp16, weight = var_83_to_fp16, x = audio_data)[name = string("linear_1_cast_fp16")];
+            tensor<int32, [3]> var_86_shape_cast_fp16 = shape(x = linear_0_cast_fp16)[name = string("op_86_shape_cast_fp16")];
+            int32 gather_0_axis_0 = const()[name = string("gather_0_axis_0"), val = int32(0)];
+            int32 gather_0_batch_dims_0 = const()[name = string("gather_0_batch_dims_0"), val = int32(0)];
+            bool gather_0_validate_indices_0 = const()[name = string("gather_0_validate_indices_0"), val = bool(false)];
+            string var_86_shape_cast_fp16_to_int16_dtype_0 = const()[name = string("op_86_shape_cast_fp16_to_int16_dtype_0"), val = string("int16")];
+            uint16 select_0_to_uint16 = const()[name = string("select_0_to_uint16"), val = uint16(1)];
+            tensor<int16, [3]> var_86_shape_cast_fp16_to_int16 = cast(dtype = var_86_shape_cast_fp16_to_int16_dtype_0, x = var_86_shape_cast_fp16)[name = string("cast_43")];
+            int16 gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = select_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_86_shape_cast_fp16_to_int16)[name = string("gather_0_cast_uint16")];
+            string gather_0_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_0_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            tensor<int32, [1]> expand_dims_11_axes_0 = const()[name = string("expand_dims_11_axes_0"), val = tensor<int32, [1]>([0])];
+            int32 gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = string("cast_42")];
+            tensor<int32, [1]> expand_dims_11 = expand_dims(axes = expand_dims_11_axes_0, x = gather_0_cast_uint16_to_int32)[name = string("expand_dims_11")];
+            tensor<int32, [4]> concat_5 = const()[name = string("concat_5"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [1]> concat_6_values0_0 = const()[name = string("concat_6_values0_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_6_values3_0 = const()[name = string("concat_6_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
+            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (concat_6_values0_0, concat_6_values1_0, expand_dims_11, concat_6_values3_0))[name = string("concat_6")];
+            tensor<int32, [4]> k_cache2_internal_tensor_assign_1_stride_0 = const()[name = string("k_cache2_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> k_cache2_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_cache2_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> k_cache2_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_cache2_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> k_cache2_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_cache2_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [6, 1, 1500, 512]> k_cache2_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = k_cache2_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = k_cache2_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_cache2_internal_tensor_assign_1_squeeze_mask_0, stride = k_cache2_internal_tensor_assign_1_stride_0, update = linear_0_cast_fp16, x = read_state_2)[name = string("k_cache2_internal_tensor_assign_1_cast_fp16")];
+            write_state(data = k_cache2_internal_tensor_assign_1_cast_fp16, input = k_cache2)[name = string("coreml_update_state_16_write_state")];
+            tensor<fp16, [6, 1, 1500, 512]> coreml_update_state_16 = read_state(input = k_cache2)[name = string("coreml_update_state_16")];
+            tensor<int32, [3]> var_91_shape_cast_fp16 = shape(x = linear_1_cast_fp16)[name = string("op_91_shape_cast_fp16")];
+            int32 gather_1_axis_0 = const()[name = string("gather_1_axis_0"), val = int32(0)];
+            int32 gather_1_batch_dims_0 = const()[name = string("gather_1_batch_dims_0"), val = int32(0)];
+            bool gather_1_validate_indices_0 = const()[name = string("gather_1_validate_indices_0"), val = bool(false)];
+            string var_91_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_91_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
+            uint16 select_1_to_uint16 = const()[name = string("select_1_to_uint16"), val = uint16(1)];
+            tensor<uint16, [3]> var_91_shape_cast_fp16_to_uint16 = cast(dtype = var_91_shape_cast_fp16_to_uint16_dtype_0, x = var_91_shape_cast_fp16)[name = string("cast_41")];
+            uint16 gather_1_cast_uint16 = gather(axis = gather_1_axis_0, batch_dims = gather_1_batch_dims_0, indices = select_1_to_uint16, validate_indices = gather_1_validate_indices_0, x = var_91_shape_cast_fp16_to_uint16)[name = string("gather_1_cast_uint16")];
+            string gather_1_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_1_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            tensor<int32, [1]> expand_dims_15_axes_0 = const()[name = string("expand_dims_15_axes_0"), val = tensor<int32, [1]>([0])];
+            int32 gather_1_cast_uint16_to_int32 = cast(dtype = gather_1_cast_uint16_to_int32_dtype_0, x = gather_1_cast_uint16)[name = string("cast_40")];
+            tensor<int32, [1]> expand_dims_15 = expand_dims(axes = expand_dims_15_axes_0, x = gather_1_cast_uint16_to_int32)[name = string("expand_dims_15")];
+            tensor<int32, [4]> concat_8 = const()[name = string("concat_8"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [1]> concat_9_values0_0 = const()[name = string("concat_9_values0_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_9_values1_0 = const()[name = string("concat_9_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_9_values3_0 = const()[name = string("concat_9_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_9_axis_0 = const()[name = string("concat_9_axis_0"), val = int32(0)];
+            bool concat_9_interleave_0 = const()[name = string("concat_9_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_9 = concat(axis = concat_9_axis_0, interleave = concat_9_interleave_0, values = (concat_9_values0_0, concat_9_values1_0, expand_dims_15, concat_9_values3_0))[name = string("concat_9")];
+            tensor<int32, [4]> v_cache2_internal_tensor_assign_1_stride_0 = const()[name = string("v_cache2_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> v_cache2_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_cache2_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> v_cache2_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_cache2_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> v_cache2_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_cache2_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [6, 1, 1500, 512]> v_cache2_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_8, begin_mask = v_cache2_internal_tensor_assign_1_begin_mask_0, end = concat_9, end_mask = v_cache2_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_cache2_internal_tensor_assign_1_squeeze_mask_0, stride = v_cache2_internal_tensor_assign_1_stride_0, update = linear_1_cast_fp16, x = read_state_3)[name = string("v_cache2_internal_tensor_assign_1_cast_fp16")];
+            write_state(data = v_cache2_internal_tensor_assign_1_cast_fp16, input = v_cache2)[name = string("coreml_update_state_17_write_state")];
+            tensor<fp16, [6, 1, 1500, 512]> coreml_update_state_17 = read_state(input = v_cache2)[name = string("coreml_update_state_17")];
+            tensor<fp16, [512, 512]> var_113_to_fp16 = const()[name = string("op_113_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3803520)))];
+            tensor<fp16, [1, ?, 512]> linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = var_113_to_fp16, x = audio_data)[name = string("linear_2_cast_fp16")];
+            tensor<fp16, [512, 512]> var_117_to_fp16 = const()[name = string("op_117_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4327872)))];
+            tensor<fp16, [512]> var_118_to_fp16 = const()[name = string("op_118_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4852224)))];
+            tensor<fp16, [1, ?, 512]> linear_3_cast_fp16 = linear(bias = var_118_to_fp16, weight = var_117_to_fp16, x = audio_data)[name = string("linear_3_cast_fp16")];
+            tensor<int32, [3]> var_120_shape_cast_fp16 = shape(x = linear_2_cast_fp16)[name = string("op_120_shape_cast_fp16")];
+            int32 gather_2_axis_0 = const()[name = string("gather_2_axis_0"), val = int32(0)];
+            int32 gather_2_batch_dims_0 = const()[name = string("gather_2_batch_dims_0"), val = int32(0)];
+            bool gather_2_validate_indices_0 = const()[name = string("gather_2_validate_indices_0"), val = bool(false)];
+            string var_120_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_120_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
+            uint16 select_2_to_uint16 = const()[name = string("select_2_to_uint16"), val = uint16(1)];
+            tensor<uint16, [3]> var_120_shape_cast_fp16_to_uint16 = cast(dtype = var_120_shape_cast_fp16_to_uint16_dtype_0, x = var_120_shape_cast_fp16)[name = string("cast_39")];
+            uint16 gather_2_cast_uint16 = gather(axis = gather_2_axis_0, batch_dims = gather_2_batch_dims_0, indices = select_2_to_uint16, validate_indices = gather_2_validate_indices_0, x = var_120_shape_cast_fp16_to_uint16)[name = string("gather_2_cast_uint16")];
+            string gather_2_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_2_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            tensor<int32, [1]> expand_dims_19_axes_0 = const()[name = string("expand_dims_19_axes_0"), val = tensor<int32, [1]>([0])];
+            int32 gather_2_cast_uint16_to_int32 = cast(dtype = gather_2_cast_uint16_to_int32_dtype_0, x = gather_2_cast_uint16)[name = string("cast_38")];
+            tensor<int32, [1]> expand_dims_19 = expand_dims(axes = expand_dims_19_axes_0, x = gather_2_cast_uint16_to_int32)[name = string("expand_dims_19")];
+            tensor<int32, [4]> concat_11 = const()[name = string("concat_11"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [1]> concat_12_values0_0 = const()[name = string("concat_12_values0_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_12_values1_0 = const()[name = string("concat_12_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_12_values3_0 = const()[name = string("concat_12_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_12_axis_0 = const()[name = string("concat_12_axis_0"), val = int32(0)];
+            bool concat_12_interleave_0 = const()[name = string("concat_12_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_12 = concat(axis = concat_12_axis_0, interleave = concat_12_interleave_0, values = (concat_12_values0_0, concat_12_values1_0, expand_dims_19, concat_12_values3_0))[name = string("concat_12")];
+            tensor<int32, [4]> k_cache2_internal_tensor_assign_2_stride_0 = const()[name = string("k_cache2_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> k_cache2_internal_tensor_assign_2_begin_mask_0 = const()[name = string("k_cache2_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> k_cache2_internal_tensor_assign_2_end_mask_0 = const()[name = string("k_cache2_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> k_cache2_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("k_cache2_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [6, 1, 1500, 512]> k_cache2_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_11, begin_mask = k_cache2_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = k_cache2_internal_tensor_assign_2_end_mask_0, squeeze_mask = k_cache2_internal_tensor_assign_2_squeeze_mask_0, stride = k_cache2_internal_tensor_assign_2_stride_0, update = linear_2_cast_fp16, x = coreml_update_state_16)[name = string("k_cache2_internal_tensor_assign_2_cast_fp16")];
+            write_state(data = k_cache2_internal_tensor_assign_2_cast_fp16, input = k_cache2)[name = string("coreml_update_state_18_write_state")];
+            tensor<fp16, [6, 1, 1500, 512]> coreml_update_state_18 = read_state(input = k_cache2)[name = string("coreml_update_state_18")];
+            tensor<int32, [3]> var_125_shape_cast_fp16 = shape(x = linear_3_cast_fp16)[name = string("op_125_shape_cast_fp16")];
+            int32 gather_3_axis_0 = const()[name = string("gather_3_axis_0"), val = int32(0)];
+            int32 gather_3_batch_dims_0 = const()[name = string("gather_3_batch_dims_0"), val = int32(0)];
+            bool gather_3_validate_indices_0 = const()[name = string("gather_3_validate_indices_0"), val = bool(false)];
+            string var_125_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_125_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
+            uint16 select_3_to_uint16 = const()[name = string("select_3_to_uint16"), val = uint16(1)];
+            tensor<uint16, [3]> var_125_shape_cast_fp16_to_uint16 = cast(dtype = var_125_shape_cast_fp16_to_uint16_dtype_0, x = var_125_shape_cast_fp16)[name = string("cast_37")];
+            uint16 gather_3_cast_uint16 = gather(axis = gather_3_axis_0, batch_dims = gather_3_batch_dims_0, indices = select_3_to_uint16, validate_indices = gather_3_validate_indices_0, x = var_125_shape_cast_fp16_to_uint16)[name = string("gather_3_cast_uint16")];
+            string gather_3_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_3_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            tensor<int32, [1]> expand_dims_23_axes_0 = const()[name = string("expand_dims_23_axes_0"), val = tensor<int32, [1]>([0])];
+            int32 gather_3_cast_uint16_to_int32 = cast(dtype = gather_3_cast_uint16_to_int32_dtype_0, x = gather_3_cast_uint16)[name = string("cast_36")];
+            tensor<int32, [1]> expand_dims_23 = expand_dims(axes = expand_dims_23_axes_0, x = gather_3_cast_uint16_to_int32)[name = string("expand_dims_23")];
+            tensor<int32, [4]> concat_14 = const()[name = string("concat_14"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [1]> concat_15_values0_0 = const()[name = string("concat_15_values0_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_15_values1_0 = const()[name = string("concat_15_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_15_values3_0 = const()[name = string("concat_15_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_15_axis_0 = const()[name = string("concat_15_axis_0"), val = int32(0)];
+            bool concat_15_interleave_0 = const()[name = string("concat_15_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_15 = concat(axis = concat_15_axis_0, interleave = concat_15_interleave_0, values = (concat_15_values0_0, concat_15_values1_0, expand_dims_23, concat_15_values3_0))[name = string("concat_15")];
+            tensor<int32, [4]> v_cache2_internal_tensor_assign_2_stride_0 = const()[name = string("v_cache2_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> v_cache2_internal_tensor_assign_2_begin_mask_0 = const()[name = string("v_cache2_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> v_cache2_internal_tensor_assign_2_end_mask_0 = const()[name = string("v_cache2_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> v_cache2_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("v_cache2_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [6, 1, 1500, 512]> v_cache2_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_14, begin_mask = v_cache2_internal_tensor_assign_2_begin_mask_0, end = concat_15, end_mask = v_cache2_internal_tensor_assign_2_end_mask_0, squeeze_mask = v_cache2_internal_tensor_assign_2_squeeze_mask_0, stride = v_cache2_internal_tensor_assign_2_stride_0, update = linear_3_cast_fp16, x = coreml_update_state_17)[name = string("v_cache2_internal_tensor_assign_2_cast_fp16")];
+            write_state(data = v_cache2_internal_tensor_assign_2_cast_fp16, input = v_cache2)[name = string("coreml_update_state_19_write_state")];
+            tensor<fp16, [6, 1, 1500, 512]> coreml_update_state_19 = read_state(input = v_cache2)[name = string("coreml_update_state_19")];
+            tensor<fp16, [512, 512]> var_147_to_fp16 = const()[name = string("op_147_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4853312)))];
+            tensor<fp16, [1, ?, 512]> linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = var_147_to_fp16, x = audio_data)[name = string("linear_4_cast_fp16")];
+            tensor<fp16, [512, 512]> var_151_to_fp16 = const()[name = string("op_151_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5377664)))];
+            tensor<fp16, [512]> var_152_to_fp16 = const()[name = string("op_152_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5902016)))];
+            tensor<fp16, [1, ?, 512]> linear_5_cast_fp16 = linear(bias = var_152_to_fp16, weight = var_151_to_fp16, x = audio_data)[name = string("linear_5_cast_fp16")];
+            tensor<int32, [3]> var_154_shape_cast_fp16 = shape(x = linear_4_cast_fp16)[name = string("op_154_shape_cast_fp16")];
+            int32 gather_4_axis_0 = const()[name = string("gather_4_axis_0"), val = int32(0)];
+            int32 gather_4_batch_dims_0 = const()[name = string("gather_4_batch_dims_0"), val = int32(0)];
+            bool gather_4_validate_indices_0 = const()[name = string("gather_4_validate_indices_0"), val = bool(false)];
+            string var_154_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_154_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
+            uint16 select_4_to_uint16 = const()[name = string("select_4_to_uint16"), val = uint16(1)];
+            tensor<uint16, [3]> var_154_shape_cast_fp16_to_uint16 = cast(dtype = var_154_shape_cast_fp16_to_uint16_dtype_0, x = var_154_shape_cast_fp16)[name = string("cast_35")];
+            uint16 gather_4_cast_uint16 = gather(axis = gather_4_axis_0, batch_dims = gather_4_batch_dims_0, indices = select_4_to_uint16, validate_indices = gather_4_validate_indices_0, x = var_154_shape_cast_fp16_to_uint16)[name = string("gather_4_cast_uint16")];
+            string gather_4_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_4_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            tensor<int32, [1]> expand_dims_27_axes_0 = const()[name = string("expand_dims_27_axes_0"), val = tensor<int32, [1]>([0])];
+            int32 gather_4_cast_uint16_to_int32 = cast(dtype = gather_4_cast_uint16_to_int32_dtype_0, x = gather_4_cast_uint16)[name = string("cast_34")];
+            tensor<int32, [1]> expand_dims_27 = expand_dims(axes = expand_dims_27_axes_0, x = gather_4_cast_uint16_to_int32)[name = string("expand_dims_27")];
+            tensor<int32, [4]> concat_17 = const()[name = string("concat_17"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [1]> concat_18_values0_0 = const()[name = string("concat_18_values0_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_18_values1_0 = const()[name = string("concat_18_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_18_values3_0 = const()[name = string("concat_18_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
+            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (concat_18_values0_0, concat_18_values1_0, expand_dims_27, concat_18_values3_0))[name = string("concat_18")];
+            tensor<int32, [4]> k_cache2_internal_tensor_assign_3_stride_0 = const()[name = string("k_cache2_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> k_cache2_internal_tensor_assign_3_begin_mask_0 = const()[name = string("k_cache2_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> k_cache2_internal_tensor_assign_3_end_mask_0 = const()[name = string("k_cache2_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> k_cache2_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("k_cache2_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [6, 1, 1500, 512]> k_cache2_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_17, begin_mask = k_cache2_internal_tensor_assign_3_begin_mask_0, end = concat_18, end_mask = k_cache2_internal_tensor_assign_3_end_mask_0, squeeze_mask = k_cache2_internal_tensor_assign_3_squeeze_mask_0, stride = k_cache2_internal_tensor_assign_3_stride_0, update = linear_4_cast_fp16, x = coreml_update_state_18)[name = string("k_cache2_internal_tensor_assign_3_cast_fp16")];
+            write_state(data = k_cache2_internal_tensor_assign_3_cast_fp16, input = k_cache2)[name = string("coreml_update_state_20_write_state")];
+            tensor<fp16, [6, 1, 1500, 512]> coreml_update_state_20 = read_state(input = k_cache2)[name = string("coreml_update_state_20")];
+            tensor<int32, [3]> var_159_shape_cast_fp16 = shape(x = linear_5_cast_fp16)[name = string("op_159_shape_cast_fp16")];
+            int32 gather_5_axis_0 = const()[name = string("gather_5_axis_0"), val = int32(0)];
+            int32 gather_5_batch_dims_0 = const()[name = string("gather_5_batch_dims_0"), val = int32(0)];
+            bool gather_5_validate_indices_0 = const()[name = string("gather_5_validate_indices_0"), val = bool(false)];
+            string var_159_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_159_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
+            uint16 select_5_to_uint16 = const()[name = string("select_5_to_uint16"), val = uint16(1)];
+            tensor<uint16, [3]> var_159_shape_cast_fp16_to_uint16 = cast(dtype = var_159_shape_cast_fp16_to_uint16_dtype_0, x = var_159_shape_cast_fp16)[name = string("cast_33")];
+            uint16 gather_5_cast_uint16 = gather(axis = gather_5_axis_0, batch_dims = gather_5_batch_dims_0, indices = select_5_to_uint16, validate_indices = gather_5_validate_indices_0, x = var_159_shape_cast_fp16_to_uint16)[name = string("gather_5_cast_uint16")];
+            string gather_5_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_5_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            tensor<int32, [1]> expand_dims_31_axes_0 = const()[name = string("expand_dims_31_axes_0"), val = tensor<int32, [1]>([0])];
+            int32 gather_5_cast_uint16_to_int32 = cast(dtype = gather_5_cast_uint16_to_int32_dtype_0, x = gather_5_cast_uint16)[name = string("cast_32")];
+            tensor<int32, [1]> expand_dims_31 = expand_dims(axes = expand_dims_31_axes_0, x = gather_5_cast_uint16_to_int32)[name = string("expand_dims_31")];
+            tensor<int32, [4]> concat_20 = const()[name = string("concat_20"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [1]> concat_21_values0_0 = const()[name = string("concat_21_values0_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_21_values1_0 = const()[name = string("concat_21_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_21_values3_0 = const()[name = string("concat_21_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_21_axis_0 = const()[name = string("concat_21_axis_0"), val = int32(0)];
+            bool concat_21_interleave_0 = const()[name = string("concat_21_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_21 = concat(axis = concat_21_axis_0, interleave = concat_21_interleave_0, values = (concat_21_values0_0, concat_21_values1_0, expand_dims_31, concat_21_values3_0))[name = string("concat_21")];
+            tensor<int32, [4]> v_cache2_internal_tensor_assign_3_stride_0 = const()[name = string("v_cache2_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> v_cache2_internal_tensor_assign_3_begin_mask_0 = const()[name = string("v_cache2_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> v_cache2_internal_tensor_assign_3_end_mask_0 = const()[name = string("v_cache2_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> v_cache2_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("v_cache2_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [6, 1, 1500, 512]> v_cache2_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_20, begin_mask = v_cache2_internal_tensor_assign_3_begin_mask_0, end = concat_21, end_mask = v_cache2_internal_tensor_assign_3_end_mask_0, squeeze_mask = v_cache2_internal_tensor_assign_3_squeeze_mask_0, stride = v_cache2_internal_tensor_assign_3_stride_0, update = linear_5_cast_fp16, x = coreml_update_state_19)[name = string("v_cache2_internal_tensor_assign_3_cast_fp16")];
+            write_state(data = v_cache2_internal_tensor_assign_3_cast_fp16, input = v_cache2)[name = string("coreml_update_state_21_write_state")];
+            tensor<fp16, [6, 1, 1500, 512]> coreml_update_state_21 = read_state(input = v_cache2)[name = string("coreml_update_state_21")];
+            tensor<fp16, [512, 512]> var_181_to_fp16 = const()[name = string("op_181_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5903104)))];
+            tensor<fp16, [1, ?, 512]> linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = var_181_to_fp16, x = audio_data)[name = string("linear_6_cast_fp16")];
+            tensor<fp16, [512, 512]> var_185_to_fp16 = const()[name = string("op_185_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6427456)))];
+            tensor<fp16, [512]> var_186_to_fp16 = const()[name = string("op_186_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6951808)))];
+            tensor<fp16, [1, ?, 512]> linear_7_cast_fp16 = linear(bias = var_186_to_fp16, weight = var_185_to_fp16, x = audio_data)[name = string("linear_7_cast_fp16")];
+            tensor<int32, [3]> var_188_shape_cast_fp16 = shape(x = linear_6_cast_fp16)[name = string("op_188_shape_cast_fp16")];
+            int32 gather_6_axis_0 = const()[name = string("gather_6_axis_0"), val = int32(0)];
+            int32 gather_6_batch_dims_0 = const()[name = string("gather_6_batch_dims_0"), val = int32(0)];
+            bool gather_6_validate_indices_0 = const()[name = string("gather_6_validate_indices_0"), val = bool(false)];
+            string var_188_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_188_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
+            uint16 select_6_to_uint16 = const()[name = string("select_6_to_uint16"), val = uint16(1)];
+            tensor<uint16, [3]> var_188_shape_cast_fp16_to_uint16 = cast(dtype = var_188_shape_cast_fp16_to_uint16_dtype_0, x = var_188_shape_cast_fp16)[name = string("cast_31")];
+            uint16 gather_6_cast_uint16 = gather(axis = gather_6_axis_0, batch_dims = gather_6_batch_dims_0, indices = select_6_to_uint16, validate_indices = gather_6_validate_indices_0, x = var_188_shape_cast_fp16_to_uint16)[name = string("gather_6_cast_uint16")];
+            string gather_6_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_6_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            tensor<int32, [1]> expand_dims_35_axes_0 = const()[name = string("expand_dims_35_axes_0"), val = tensor<int32, [1]>([0])];
+            int32 gather_6_cast_uint16_to_int32 = cast(dtype = gather_6_cast_uint16_to_int32_dtype_0, x = gather_6_cast_uint16)[name = string("cast_30")];
+            tensor<int32, [1]> expand_dims_35 = expand_dims(axes = expand_dims_35_axes_0, x = gather_6_cast_uint16_to_int32)[name = string("expand_dims_35")];
+            tensor<int32, [4]> concat_23 = const()[name = string("concat_23"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [1]> concat_24_values0_0 = const()[name = string("concat_24_values0_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_24_values1_0 = const()[name = string("concat_24_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_24_values3_0 = const()[name = string("concat_24_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_24_axis_0 = const()[name = string("concat_24_axis_0"), val = int32(0)];
+            bool concat_24_interleave_0 = const()[name = string("concat_24_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_24 = concat(axis = concat_24_axis_0, interleave = concat_24_interleave_0, values = (concat_24_values0_0, concat_24_values1_0, expand_dims_35, concat_24_values3_0))[name = string("concat_24")];
+            tensor<int32, [4]> k_cache2_internal_tensor_assign_4_stride_0 = const()[name = string("k_cache2_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> k_cache2_internal_tensor_assign_4_begin_mask_0 = const()[name = string("k_cache2_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> k_cache2_internal_tensor_assign_4_end_mask_0 = const()[name = string("k_cache2_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> k_cache2_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("k_cache2_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [6, 1, 1500, 512]> k_cache2_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_23, begin_mask = k_cache2_internal_tensor_assign_4_begin_mask_0, end = concat_24, end_mask = k_cache2_internal_tensor_assign_4_end_mask_0, squeeze_mask = k_cache2_internal_tensor_assign_4_squeeze_mask_0, stride = k_cache2_internal_tensor_assign_4_stride_0, update = linear_6_cast_fp16, x = coreml_update_state_20)[name = string("k_cache2_internal_tensor_assign_4_cast_fp16")];
+            write_state(data = k_cache2_internal_tensor_assign_4_cast_fp16, input = k_cache2)[name = string("coreml_update_state_22_write_state")];
+            tensor<fp16, [6, 1, 1500, 512]> coreml_update_state_22 = read_state(input = k_cache2)[name = string("coreml_update_state_22")];
+            tensor<int32, [3]> var_193_shape_cast_fp16 = shape(x = linear_7_cast_fp16)[name = string("op_193_shape_cast_fp16")];
+            int32 gather_7_axis_0 = const()[name = string("gather_7_axis_0"), val = int32(0)];
+            int32 gather_7_batch_dims_0 = const()[name = string("gather_7_batch_dims_0"), val = int32(0)];
+            bool gather_7_validate_indices_0 = const()[name = string("gather_7_validate_indices_0"), val = bool(false)];
+            string var_193_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_193_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
+            uint16 select_7_to_uint16 = const()[name = string("select_7_to_uint16"), val = uint16(1)];
+            tensor<uint16, [3]> var_193_shape_cast_fp16_to_uint16 = cast(dtype = var_193_shape_cast_fp16_to_uint16_dtype_0, x = var_193_shape_cast_fp16)[name = string("cast_29")];
+            uint16 gather_7_cast_uint16 = gather(axis = gather_7_axis_0, batch_dims = gather_7_batch_dims_0, indices = select_7_to_uint16, validate_indices = gather_7_validate_indices_0, x = var_193_shape_cast_fp16_to_uint16)[name = string("gather_7_cast_uint16")];
+            string gather_7_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_7_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            tensor<int32, [1]> expand_dims_39_axes_0 = const()[name = string("expand_dims_39_axes_0"), val = tensor<int32, [1]>([0])];
+            int32 gather_7_cast_uint16_to_int32 = cast(dtype = gather_7_cast_uint16_to_int32_dtype_0, x = gather_7_cast_uint16)[name = string("cast_28")];
+            tensor<int32, [1]> expand_dims_39 = expand_dims(axes = expand_dims_39_axes_0, x = gather_7_cast_uint16_to_int32)[name = string("expand_dims_39")];
+            tensor<int32, [4]> concat_26 = const()[name = string("concat_26"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [1]> concat_27_values0_0 = const()[name = string("concat_27_values0_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_27_values1_0 = const()[name = string("concat_27_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_27_values3_0 = const()[name = string("concat_27_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_27_axis_0 = const()[name = string("concat_27_axis_0"), val = int32(0)];
+            bool concat_27_interleave_0 = const()[name = string("concat_27_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_27 = concat(axis = concat_27_axis_0, interleave = concat_27_interleave_0, values = (concat_27_values0_0, concat_27_values1_0, expand_dims_39, concat_27_values3_0))[name = string("concat_27")];
+            tensor<int32, [4]> v_cache2_internal_tensor_assign_4_stride_0 = const()[name = string("v_cache2_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> v_cache2_internal_tensor_assign_4_begin_mask_0 = const()[name = string("v_cache2_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> v_cache2_internal_tensor_assign_4_end_mask_0 = const()[name = string("v_cache2_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> v_cache2_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("v_cache2_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [6, 1, 1500, 512]> v_cache2_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_26, begin_mask = v_cache2_internal_tensor_assign_4_begin_mask_0, end = concat_27, end_mask = v_cache2_internal_tensor_assign_4_end_mask_0, squeeze_mask = v_cache2_internal_tensor_assign_4_squeeze_mask_0, stride = v_cache2_internal_tensor_assign_4_stride_0, update = linear_7_cast_fp16, x = coreml_update_state_21)[name = string("v_cache2_internal_tensor_assign_4_cast_fp16")];
+            write_state(data = v_cache2_internal_tensor_assign_4_cast_fp16, input = v_cache2)[name = string("coreml_update_state_23_write_state")];
+            tensor<fp16, [6, 1, 1500, 512]> coreml_update_state_23 = read_state(input = v_cache2)[name = string("coreml_update_state_23")];
+            tensor<fp16, [512, 512]> var_215_to_fp16 = const()[name = string("op_215_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6952896)))];
+            tensor<fp16, [1, ?, 512]> linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = var_215_to_fp16, x = audio_data)[name = string("linear_8_cast_fp16")];
+            tensor<fp16, [512, 512]> var_219_to_fp16 = const()[name = string("op_219_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7477248)))];
+            tensor<fp16, [512]> var_220_to_fp16 = const()[name = string("op_220_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8001600)))];
+            tensor<fp16, [1, ?, 512]> linear_9_cast_fp16 = linear(bias = var_220_to_fp16, weight = var_219_to_fp16, x = audio_data)[name = string("linear_9_cast_fp16")];
+            tensor<int32, [3]> var_222_shape_cast_fp16 = shape(x = linear_8_cast_fp16)[name = string("op_222_shape_cast_fp16")];
+            int32 gather_8_axis_0 = const()[name = string("gather_8_axis_0"), val = int32(0)];
+            int32 gather_8_batch_dims_0 = const()[name = string("gather_8_batch_dims_0"), val = int32(0)];
+            bool gather_8_validate_indices_0 = const()[name = string("gather_8_validate_indices_0"), val = bool(false)];
+            string var_222_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_222_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
+            uint16 select_8_to_uint16 = const()[name = string("select_8_to_uint16"), val = uint16(1)];
+            tensor<uint16, [3]> var_222_shape_cast_fp16_to_uint16 = cast(dtype = var_222_shape_cast_fp16_to_uint16_dtype_0, x = var_222_shape_cast_fp16)[name = string("cast_27")];
+            uint16 gather_8_cast_uint16 = gather(axis = gather_8_axis_0, batch_dims = gather_8_batch_dims_0, indices = select_8_to_uint16, validate_indices = gather_8_validate_indices_0, x = var_222_shape_cast_fp16_to_uint16)[name = string("gather_8_cast_uint16")];
+            string gather_8_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_8_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            tensor<int32, [1]> expand_dims_43_axes_0 = const()[name = string("expand_dims_43_axes_0"), val = tensor<int32, [1]>([0])];
+            int32 gather_8_cast_uint16_to_int32 = cast(dtype = gather_8_cast_uint16_to_int32_dtype_0, x = gather_8_cast_uint16)[name = string("cast_26")];
+            tensor<int32, [1]> expand_dims_43 = expand_dims(axes = expand_dims_43_axes_0, x = gather_8_cast_uint16_to_int32)[name = string("expand_dims_43")];
+            tensor<int32, [4]> concat_29 = const()[name = string("concat_29"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [1]> concat_30_values0_0 = const()[name = string("concat_30_values0_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_30_values1_0 = const()[name = string("concat_30_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_30_values3_0 = const()[name = string("concat_30_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_30_axis_0 = const()[name = string("concat_30_axis_0"), val = int32(0)];
+            bool concat_30_interleave_0 = const()[name = string("concat_30_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (concat_30_values0_0, concat_30_values1_0, expand_dims_43, concat_30_values3_0))[name = string("concat_30")];
+            tensor<int32, [4]> k_cache2_internal_tensor_assign_5_stride_0 = const()[name = string("k_cache2_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> k_cache2_internal_tensor_assign_5_begin_mask_0 = const()[name = string("k_cache2_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> k_cache2_internal_tensor_assign_5_end_mask_0 = const()[name = string("k_cache2_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> k_cache2_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("k_cache2_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [6, 1, 1500, 512]> k_cache2_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_29, begin_mask = k_cache2_internal_tensor_assign_5_begin_mask_0, end = concat_30, end_mask = k_cache2_internal_tensor_assign_5_end_mask_0, squeeze_mask = k_cache2_internal_tensor_assign_5_squeeze_mask_0, stride = k_cache2_internal_tensor_assign_5_stride_0, update = linear_8_cast_fp16, x = coreml_update_state_22)[name = string("k_cache2_internal_tensor_assign_5_cast_fp16")];
+            write_state(data = k_cache2_internal_tensor_assign_5_cast_fp16, input = k_cache2)[name = string("coreml_update_state_24_write_state")];
+            tensor<fp16, [6, 1, 1500, 512]> coreml_update_state_24 = read_state(input = k_cache2)[name = string("coreml_update_state_24")];
+            tensor<int32, [3]> var_227_shape_cast_fp16 = shape(x = linear_9_cast_fp16)[name = string("op_227_shape_cast_fp16")];
+            int32 gather_9_axis_0 = const()[name = string("gather_9_axis_0"), val = int32(0)];
+            int32 gather_9_batch_dims_0 = const()[name = string("gather_9_batch_dims_0"), val = int32(0)];
+            bool gather_9_validate_indices_0 = const()[name = string("gather_9_validate_indices_0"), val = bool(false)];
+            string var_227_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_227_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
+            uint16 select_9_to_uint16 = const()[name = string("select_9_to_uint16"), val = uint16(1)];
+            tensor<uint16, [3]> var_227_shape_cast_fp16_to_uint16 = cast(dtype = var_227_shape_cast_fp16_to_uint16_dtype_0, x = var_227_shape_cast_fp16)[name = string("cast_25")];
+            uint16 gather_9_cast_uint16 = gather(axis = gather_9_axis_0, batch_dims = gather_9_batch_dims_0, indices = select_9_to_uint16, validate_indices = gather_9_validate_indices_0, x = var_227_shape_cast_fp16_to_uint16)[name = string("gather_9_cast_uint16")];
+            string gather_9_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_9_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            tensor<int32, [1]> expand_dims_47_axes_0 = const()[name = string("expand_dims_47_axes_0"), val = tensor<int32, [1]>([0])];
+            int32 gather_9_cast_uint16_to_int32 = cast(dtype = gather_9_cast_uint16_to_int32_dtype_0, x = gather_9_cast_uint16)[name = string("cast_24")];
+            tensor<int32, [1]> expand_dims_47 = expand_dims(axes = expand_dims_47_axes_0, x = gather_9_cast_uint16_to_int32)[name = string("expand_dims_47")];
+            tensor<int32, [4]> concat_32 = const()[name = string("concat_32"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [1]> concat_33_values0_0 = const()[name = string("concat_33_values0_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_33_values1_0 = const()[name = string("concat_33_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_33_values3_0 = const()[name = string("concat_33_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_33_axis_0 = const()[name = string("concat_33_axis_0"), val = int32(0)];
+            bool concat_33_interleave_0 = const()[name = string("concat_33_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_33 = concat(axis = concat_33_axis_0, interleave = concat_33_interleave_0, values = (concat_33_values0_0, concat_33_values1_0, expand_dims_47, concat_33_values3_0))[name = string("concat_33")];
+            tensor<int32, [4]> v_cache2_internal_tensor_assign_5_stride_0 = const()[name = string("v_cache2_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> v_cache2_internal_tensor_assign_5_begin_mask_0 = const()[name = string("v_cache2_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> v_cache2_internal_tensor_assign_5_end_mask_0 = const()[name = string("v_cache2_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> v_cache2_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("v_cache2_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [6, 1, 1500, 512]> v_cache2_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_32, begin_mask = v_cache2_internal_tensor_assign_5_begin_mask_0, end = concat_33, end_mask = v_cache2_internal_tensor_assign_5_end_mask_0, squeeze_mask = v_cache2_internal_tensor_assign_5_squeeze_mask_0, stride = v_cache2_internal_tensor_assign_5_stride_0, update = linear_9_cast_fp16, x = coreml_update_state_23)[name = string("v_cache2_internal_tensor_assign_5_cast_fp16")];
+            write_state(data = v_cache2_internal_tensor_assign_5_cast_fp16, input = v_cache2)[name = string("coreml_update_state_25_write_state")];
+            tensor<fp16, [6, 1, 1500, 512]> coreml_update_state_25 = read_state(input = v_cache2)[name = string("coreml_update_state_25")];
+            tensor<fp16, [512, 512]> var_249_to_fp16 = const()[name = string("op_249_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8002688)))];
+            tensor<fp16, [1, ?, 512]> linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = var_249_to_fp16, x = audio_data)[name = string("linear_10_cast_fp16")];
+            tensor<fp16, [512, 512]> var_253_to_fp16 = const()[name = string("op_253_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8527040)))];
+            tensor<fp16, [512]> var_254_to_fp16 = const()[name = string("op_254_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9051392)))];
+            tensor<fp16, [1, ?, 512]> linear_11_cast_fp16 = linear(bias = var_254_to_fp16, weight = var_253_to_fp16, x = audio_data)[name = string("linear_11_cast_fp16")];
+            tensor<int32, [3]> var_256_shape_cast_fp16 = shape(x = linear_10_cast_fp16)[name = string("op_256_shape_cast_fp16")];
+            int32 gather_10_axis_0 = const()[name = string("gather_10_axis_0"), val = int32(0)];
+            int32 gather_10_batch_dims_0 = const()[name = string("gather_10_batch_dims_0"), val = int32(0)];
+            bool gather_10_validate_indices_0 = const()[name = string("gather_10_validate_indices_0"), val = bool(false)];
+            string var_256_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_256_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
+            uint16 select_10_to_uint16 = const()[name = string("select_10_to_uint16"), val = uint16(1)];
+            tensor<uint16, [3]> var_256_shape_cast_fp16_to_uint16 = cast(dtype = var_256_shape_cast_fp16_to_uint16_dtype_0, x = var_256_shape_cast_fp16)[name = string("cast_23")];
+            uint16 gather_10_cast_uint16 = gather(axis = gather_10_axis_0, batch_dims = gather_10_batch_dims_0, indices = select_10_to_uint16, validate_indices = gather_10_validate_indices_0, x = var_256_shape_cast_fp16_to_uint16)[name = string("gather_10_cast_uint16")];
+            string gather_10_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_10_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            tensor<int32, [1]> expand_dims_51_axes_0 = const()[name = string("expand_dims_51_axes_0"), val = tensor<int32, [1]>([0])];
+            int32 gather_10_cast_uint16_to_int32 = cast(dtype = gather_10_cast_uint16_to_int32_dtype_0, x = gather_10_cast_uint16)[name = string("cast_22")];
+            tensor<int32, [1]> expand_dims_51 = expand_dims(axes = expand_dims_51_axes_0, x = gather_10_cast_uint16_to_int32)[name = string("expand_dims_51")];
+            tensor<int32, [4]> concat_35 = const()[name = string("concat_35"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [1]> concat_36_values0_0 = const()[name = string("concat_36_values0_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_36_values1_0 = const()[name = string("concat_36_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_36_values3_0 = const()[name = string("concat_36_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_36_axis_0 = const()[name = string("concat_36_axis_0"), val = int32(0)];
+            bool concat_36_interleave_0 = const()[name = string("concat_36_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_36 = concat(axis = concat_36_axis_0, interleave = concat_36_interleave_0, values = (concat_36_values0_0, concat_36_values1_0, expand_dims_51, concat_36_values3_0))[name = string("concat_36")];
+            tensor<int32, [4]> k_cache2_internal_tensor_assign_6_stride_0 = const()[name = string("k_cache2_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> k_cache2_internal_tensor_assign_6_begin_mask_0 = const()[name = string("k_cache2_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> k_cache2_internal_tensor_assign_6_end_mask_0 = const()[name = string("k_cache2_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> k_cache2_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("k_cache2_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [6, 1, 1500, 512]> k_cache2_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_35, begin_mask = k_cache2_internal_tensor_assign_6_begin_mask_0, end = concat_36, end_mask = k_cache2_internal_tensor_assign_6_end_mask_0, squeeze_mask = k_cache2_internal_tensor_assign_6_squeeze_mask_0, stride = k_cache2_internal_tensor_assign_6_stride_0, update = linear_10_cast_fp16, x = coreml_update_state_24)[name = string("k_cache2_internal_tensor_assign_6_cast_fp16")];
+            write_state(data = k_cache2_internal_tensor_assign_6_cast_fp16, input = k_cache2)[name = string("coreml_update_state_26_write_state")];
+            tensor<int32, [3]> var_261_shape_cast_fp16 = shape(x = linear_11_cast_fp16)[name = string("op_261_shape_cast_fp16")];
+            int32 gather_11_axis_0 = const()[name = string("gather_11_axis_0"), val = int32(0)];
+            int32 gather_11_batch_dims_0 = const()[name = string("gather_11_batch_dims_0"), val = int32(0)];
+            bool gather_11_validate_indices_0 = const()[name = string("gather_11_validate_indices_0"), val = bool(false)];
+            string var_261_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_261_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
+            uint16 select_11_to_uint16 = const()[name = string("select_11_to_uint16"), val = uint16(1)];
+            tensor<uint16, [3]> var_261_shape_cast_fp16_to_uint16 = cast(dtype = var_261_shape_cast_fp16_to_uint16_dtype_0, x = var_261_shape_cast_fp16)[name = string("cast_21")];
+            uint16 gather_11_cast_uint16 = gather(axis = gather_11_axis_0, batch_dims = gather_11_batch_dims_0, indices = select_11_to_uint16, validate_indices = gather_11_validate_indices_0, x = var_261_shape_cast_fp16_to_uint16)[name = string("gather_11_cast_uint16")];
+            string gather_11_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_11_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            tensor<int32, [1]> expand_dims_55_axes_0 = const()[name = string("expand_dims_55_axes_0"), val = tensor<int32, [1]>([0])];
+            int32 gather_11_cast_uint16_to_int32 = cast(dtype = gather_11_cast_uint16_to_int32_dtype_0, x = gather_11_cast_uint16)[name = string("cast_20")];
+            tensor<int32, [1]> expand_dims_55 = expand_dims(axes = expand_dims_55_axes_0, x = gather_11_cast_uint16_to_int32)[name = string("expand_dims_55")];
+            tensor<int32, [4]> concat_38 = const()[name = string("concat_38"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [1]> concat_39_values0_0 = const()[name = string("concat_39_values0_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_39_values1_0 = const()[name = string("concat_39_values1_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> concat_39_values3_0 = const()[name = string("concat_39_values3_0"), val = tensor<int32, [1]>([0])];
+            int32 concat_39_axis_0 = const()[name = string("concat_39_axis_0"), val = int32(0)];
+            bool concat_39_interleave_0 = const()[name = string("concat_39_interleave_0"), val = bool(false)];
+            tensor<int32, [4]> concat_39 = concat(axis = concat_39_axis_0, interleave = concat_39_interleave_0, values = (concat_39_values0_0, concat_39_values1_0, expand_dims_55, concat_39_values3_0))[name = string("concat_39")];
+            tensor<int32, [4]> v_cache2_internal_tensor_assign_6_stride_0 = const()[name = string("v_cache2_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
+            tensor<bool, [4]> v_cache2_internal_tensor_assign_6_begin_mask_0 = const()[name = string("v_cache2_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
+            tensor<bool, [4]> v_cache2_internal_tensor_assign_6_end_mask_0 = const()[name = string("v_cache2_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])];
+            tensor<bool, [4]> v_cache2_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("v_cache2_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [6, 1, 1500, 512]> v_cache2_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_38, begin_mask = v_cache2_internal_tensor_assign_6_begin_mask_0, end = concat_39, end_mask = v_cache2_internal_tensor_assign_6_end_mask_0, squeeze_mask = v_cache2_internal_tensor_assign_6_squeeze_mask_0, stride = v_cache2_internal_tensor_assign_6_stride_0, update = linear_11_cast_fp16, x = coreml_update_state_25)[name = string("v_cache2_internal_tensor_assign_6_cast_fp16")];
+            write_state(data = v_cache2_internal_tensor_assign_6_cast_fp16, input = v_cache2)[name = string("coreml_update_state_27_write_state")];
+        } -> (dummy);
+}

base/decoder_first.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4fdbcff86cdfe9e0b8842ad4bc1af8ebbf22082b1d0342a8304023f63dd3663f
+size 9052480

base/decoder_second.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:68a3b3426587d83e56d3286cc0b733c9b8a5bff6b1ad6f9e1789a3cb55164455
+size 243

base/decoder_second.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a6c0272d581c200e0ab4f29c687e4a7b49152e241cea335fa6faa6a430a460b6
+size 487

base/decoder_second.mlmodelc/metadata.json

@@ -0,0 +1,127 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16)",
+        "shortDescription" : "",
+        "shape" : "[]",
+        "name" : "logits",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.linear" : 49,
+      "Ios18.readState" : 14,
+      "Ios18.expandDims" : 7,
+      "Ios18.sub" : 1,
+      "Ios18.matmul" : 24,
+      "Ios18.gelu" : 6,
+      "Ios18.gather" : 9,
+      "Ios18.concat" : 32,
+      "Shape" : 8,
+      "Ios18.add" : 31,
+      "Ios18.sliceUpdate" : 24,
+      "Ios18.sliceByIndex" : 49,
+      "Ios18.layerNorm" : 19,
+      "Ios18.cast" : 16,
+      "Ios18.transpose" : 48,
+      "Ios18.writeState" : 12,
+      "Ios18.reshape" : 48,
+      "Ios18.softmax" : 12,
+      "Ios18.mul" : 24
+    },
+    "computePrecision" : "Mixed (Float16, Int16, Int32, UInt16)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 6 × 1 × 448 × 512)",
+        "shortDescription" : "",
+        "shape" : "[6, 1, 448, 512]",
+        "name" : "k_cache1",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 6 × 1 × 448 × 512)",
+        "shortDescription" : "",
+        "shape" : "[6, 1, 448, 512]",
+        "name" : "v_cache1",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 6 × 1 × 1500 × 512)",
+        "shortDescription" : "",
+        "shape" : "[6, 1, 1500, 512]",
+        "name" : "k_cache2",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 6 × 1 × 1500 × 512)",
+        "shortDescription" : "",
+        "shape" : "[6, 1, 1500, 512]",
+        "name" : "v_cache2",
+        "type" : "State"
+      }
+    ],
+    "availability" : {
+      "macOS" : "15.0",
+      "tvOS" : "18.0",
+      "visionOS" : "2.0",
+      "watchOS" : "11.0",
+      "iOS" : "18.0",
+      "macCatalyst" : "18.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.4.1",
+      "com.github.apple.coremltools.version" : "8.0"
+    },
+    "inputSchema" : [
+      {
+        "dataType" : "Int32",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1 × 1...448",
+        "shapeRange" : "[[1, 1], [1, 448]]",
+        "formattedType" : "MultiArray (Int32 1 × 1)",
+        "type" : "MultiArray",
+        "shape" : "[1, 1]",
+        "name" : "token_data",
+        "shortDescription" : ""
+      },
+      {
+        "dataType" : "Float16",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1 × 1...448",
+        "shapeRange" : "[[1, 1], [1, 448]]",
+        "formattedType" : "MultiArray (Float16 1 × 1)",
+        "type" : "MultiArray",
+        "shape" : "[1, 1]",
+        "name" : "offset_mask",
+        "shortDescription" : ""
+      }
+    ],
+    "generatedClassName" : "decoder_second",
+    "method" : "predict"
+  }
+]

base/decoder_second.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94338bcd9475d6d8848699ee40dd6fac40d1e597c1e28d124454a7bf37bff672
+size 99759858

base/encoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:405dc318443c493222a32916d66c5d908d7cc1d250f73e9a192d5b734a8494ed
+size 243

base/encoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7b04d5884b5a2d983f52a0e557aff2e7d3dff78b2a9f9d496a5280546bacfaff
+size 318

base/encoder.mlmodelc/metadata.json

@@ -0,0 +1,69 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 1500 × 512)",
+        "shortDescription" : "",
+        "shape" : "[1, 1500, 512]",
+        "name" : "output",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.mul" : 12,
+      "Ios18.softmax" : 6,
+      "Ios18.linear" : 36,
+      "Ios18.gelu" : 8,
+      "Ios18.layerNorm" : 13,
+      "Ios18.transpose" : 25,
+      "Ios18.matmul" : 12,
+      "Ios18.conv" : 2,
+      "Ios18.add" : 13,
+      "Ios18.reshape" : 24
+    },
+    "computePrecision" : "Mixed (Float16, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "15.0",
+      "tvOS" : "18.0",
+      "visionOS" : "2.0",
+      "watchOS" : "11.0",
+      "iOS" : "18.0",
+      "macCatalyst" : "18.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.4.1",
+      "com.github.apple.coremltools.version" : "8.0"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 80 × 3000)",
+        "shortDescription" : "",
+        "shape" : "[1, 80, 3000]",
+        "name" : "logmel_data",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "encoder",
+    "method" : "predict"
+  }
+]

base/encoder.mlmodelc/model.mil

@@ -0,0 +1,384 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3400.43.1"}, {"coremlc-version", "3400.58.2"}, {"coremltools-component-torch", "2.4.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.0"}})]
+{
+    func main<ios18>(tensor<fp16, [1, 80, 3000]> logmel_data) {
+            string var_32_pad_type_0 = const()[name = string("op_32_pad_type_0"), val = string("custom")];
+            tensor<int32, [2]> var_32_pad_0 = const()[name = string("op_32_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> var_32_strides_0 = const()[name = string("op_32_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> var_32_dilations_0 = const()[name = string("op_32_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 var_32_groups_0 = const()[name = string("op_32_groups_0"), val = int32(1)];
+            tensor<fp16, [512, 80, 3]> weight_3_to_fp16 = const()[name = string("weight_3_to_fp16"), val = tensor<fp16, [512, 80, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [512]> bias_3_to_fp16 = const()[name = string("bias_3_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245888)))];
+            tensor<fp16, [1, 512, 3000]> var_32_cast_fp16 = conv(bias = bias_3_to_fp16, dilations = var_32_dilations_0, groups = var_32_groups_0, pad = var_32_pad_0, pad_type = var_32_pad_type_0, strides = var_32_strides_0, weight = weight_3_to_fp16, x = logmel_data)[name = string("op_32_cast_fp16")];
+            string input_1_mode_0 = const()[name = string("input_1_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 512, 3000]> input_1_cast_fp16 = gelu(mode = input_1_mode_0, x = var_32_cast_fp16)[name = string("input_1_cast_fp16")];
+            string var_50_pad_type_0 = const()[name = string("op_50_pad_type_0"), val = string("custom")];
+            tensor<int32, [2]> var_50_pad_0 = const()[name = string("op_50_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> var_50_strides_0 = const()[name = string("op_50_strides_0"), val = tensor<int32, [1]>([2])];
+            tensor<int32, [1]> var_50_dilations_0 = const()[name = string("op_50_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 var_50_groups_0 = const()[name = string("op_50_groups_0"), val = int32(1)];
+            tensor<fp16, [512, 512, 3]> weight_7_to_fp16 = const()[name = string("weight_7_to_fp16"), val = tensor<fp16, [512, 512, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246976)))];
+            tensor<fp16, [512]> bias_7_to_fp16 = const()[name = string("bias_7_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1819904)))];
+            tensor<fp16, [1, 512, 1500]> var_50_cast_fp16 = conv(bias = bias_7_to_fp16, dilations = var_50_dilations_0, groups = var_50_groups_0, pad = var_50_pad_0, pad_type = var_50_pad_type_0, strides = var_50_strides_0, weight = weight_7_to_fp16, x = input_1_cast_fp16)[name = string("op_50_cast_fp16")];
+            string x_3_mode_0 = const()[name = string("x_3_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 512, 1500]> x_3_cast_fp16 = gelu(mode = x_3_mode_0, x = var_50_cast_fp16)[name = string("x_3_cast_fp16")];
+            tensor<int32, [3]> var_56 = const()[name = string("op_56"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1500, 512]> positional_embedding_to_fp16 = const()[name = string("positional_embedding_to_fp16"), val = tensor<fp16, [1500, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1820992)))];
+            tensor<fp16, [1, 1500, 512]> x_5_cast_fp16 = transpose(perm = var_56, x = x_3_cast_fp16)[name = string("transpose_60")];
+            tensor<fp16, [1, 1500, 512]> var_59_cast_fp16 = add(x = x_5_cast_fp16, y = positional_embedding_to_fp16)[name = string("op_59_cast_fp16")];
+            int32 var_72 = const()[name = string("op_72"), val = int32(-1)];
+            tensor<int32, [1]> var_88_axes_0 = const()[name = string("op_88_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> blocks_0_attn_ln_weight_to_fp16 = const()[name = string("blocks_0_attn_ln_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3357056)))];
+            tensor<fp16, [512]> blocks_0_attn_ln_bias_to_fp16 = const()[name = string("blocks_0_attn_ln_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3358144)))];
+            fp16 var_78_to_fp16 = const()[name = string("op_78_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1500, 512]> var_88_cast_fp16 = layer_norm(axes = var_88_axes_0, beta = blocks_0_attn_ln_bias_to_fp16, epsilon = var_78_to_fp16, gamma = blocks_0_attn_ln_weight_to_fp16, x = var_59_cast_fp16)[name = string("op_88_cast_fp16")];
+            tensor<fp16, [512, 512]> var_99_to_fp16 = const()[name = string("op_99_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3359232)))];
+            tensor<fp16, [512]> var_100_to_fp16 = const()[name = string("op_100_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3883584)))];
+            tensor<fp16, [1, 1500, 512]> linear_0_cast_fp16 = linear(bias = var_100_to_fp16, weight = var_99_to_fp16, x = var_88_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<fp16, [512, 512]> var_103_to_fp16 = const()[name = string("op_103_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3884672)))];
+            tensor<fp16, [512]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4409024)))];
+            tensor<fp16, [1, 1500, 512]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_103_to_fp16, x = var_88_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<fp16, [512, 512]> var_107_to_fp16 = const()[name = string("op_107_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4410112)))];
+            tensor<fp16, [512]> var_108_to_fp16 = const()[name = string("op_108_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4934464)))];
+            tensor<fp16, [1, 1500, 512]> linear_2_cast_fp16 = linear(bias = var_108_to_fp16, weight = var_107_to_fp16, x = var_88_cast_fp16)[name = string("linear_2_cast_fp16")];
+            tensor<int32, [4]> var_116 = const()[name = string("op_116"), val = tensor<int32, [4]>([1, 1500, 8, -1])];
+            tensor<fp16, [1, 1500, 8, 64]> var_117_cast_fp16 = reshape(shape = var_116, x = linear_0_cast_fp16)[name = string("op_117_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 1]> const_42_to_fp16 = const()[name = string("const_42_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
+            tensor<fp16, [1, 1500, 8, 64]> q_3_cast_fp16 = mul(x = var_117_cast_fp16, y = const_42_to_fp16)[name = string("q_3_cast_fp16")];
+            tensor<int32, [4]> var_123 = const()[name = string("op_123"), val = tensor<int32, [4]>([1, 1500, 8, -1])];
+            tensor<fp16, [1, 1500, 8, 64]> var_124_cast_fp16 = reshape(shape = var_123, x = linear_1_cast_fp16)[name = string("op_124_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 1]> const_43_to_fp16 = const()[name = string("const_43_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
+            tensor<fp16, [1, 1500, 8, 64]> k_3_cast_fp16 = mul(x = var_124_cast_fp16, y = const_43_to_fp16)[name = string("k_3_cast_fp16")];
+            tensor<int32, [4]> var_130 = const()[name = string("op_130"), val = tensor<int32, [4]>([1, 1500, 8, -1])];
+            tensor<fp16, [1, 1500, 8, 64]> var_131_cast_fp16 = reshape(shape = var_130, x = linear_2_cast_fp16)[name = string("op_131_cast_fp16")];
+            tensor<int32, [4]> var_132 = const()[name = string("op_132"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            bool qk_1_transpose_x_0 = const()[name = string("qk_1_transpose_x_0"), val = bool(false)];
+            bool qk_1_transpose_y_0 = const()[name = string("qk_1_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_24_perm_0 = const()[name = string("transpose_24_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_25_perm_0 = const()[name = string("transpose_25_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 1500]> transpose_25 = transpose(perm = transpose_25_perm_0, x = k_3_cast_fp16)[name = string("transpose_57")];
+            tensor<fp16, [1, 8, 1500, 64]> transpose_24 = transpose(perm = transpose_24_perm_0, x = q_3_cast_fp16)[name = string("transpose_58")];
+            tensor<fp16, [1, 8, 1500, 1500]> qk_1_cast_fp16 = matmul(transpose_x = qk_1_transpose_x_0, transpose_y = qk_1_transpose_y_0, x = transpose_24, y = transpose_25)[name = string("qk_1_cast_fp16")];
+            tensor<fp16, [1, 8, 1500, 1500]> var_136_cast_fp16 = softmax(axis = var_72, x = qk_1_cast_fp16)[name = string("op_136_cast_fp16")];
+            bool var_138_transpose_x_0 = const()[name = string("op_138_transpose_x_0"), val = bool(false)];
+            bool var_138_transpose_y_0 = const()[name = string("op_138_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 1500, 64]> v_3_cast_fp16 = transpose(perm = var_132, x = var_131_cast_fp16)[name = string("transpose_59")];
+            tensor<fp16, [1, 8, 1500, 64]> var_138_cast_fp16 = matmul(transpose_x = var_138_transpose_x_0, transpose_y = var_138_transpose_y_0, x = var_136_cast_fp16, y = v_3_cast_fp16)[name = string("op_138_cast_fp16")];
+            tensor<int32, [4]> var_139 = const()[name = string("op_139"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> concat_0 = const()[name = string("concat_0"), val = tensor<int32, [3]>([1, 1500, 512])];
+            tensor<fp16, [1, 1500, 8, 64]> var_140_cast_fp16 = transpose(perm = var_139, x = var_138_cast_fp16)[name = string("transpose_56")];
+            tensor<fp16, [1, 1500, 512]> x_11_cast_fp16 = reshape(shape = concat_0, x = var_140_cast_fp16)[name = string("x_11_cast_fp16")];
+            tensor<fp16, [512, 512]> var_144_to_fp16 = const()[name = string("op_144_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4935552)))];
+            tensor<fp16, [512]> var_145_to_fp16 = const()[name = string("op_145_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5459904)))];
+            tensor<fp16, [1, 1500, 512]> linear_3_cast_fp16 = linear(bias = var_145_to_fp16, weight = var_144_to_fp16, x = x_11_cast_fp16)[name = string("linear_3_cast_fp16")];
+            tensor<fp16, [1, 1500, 512]> x_13_cast_fp16 = add(x = var_59_cast_fp16, y = linear_3_cast_fp16)[name = string("x_13_cast_fp16")];
+            tensor<int32, [1]> var_152_axes_0 = const()[name = string("op_152_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> blocks_0_mlp_ln_weight_to_fp16 = const()[name = string("blocks_0_mlp_ln_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5460992)))];
+            tensor<fp16, [512]> blocks_0_mlp_ln_bias_to_fp16 = const()[name = string("blocks_0_mlp_ln_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5462080)))];
+            tensor<fp16, [1, 1500, 512]> var_152_cast_fp16 = layer_norm(axes = var_152_axes_0, beta = blocks_0_mlp_ln_bias_to_fp16, epsilon = var_78_to_fp16, gamma = blocks_0_mlp_ln_weight_to_fp16, x = x_13_cast_fp16)[name = string("op_152_cast_fp16")];
+            tensor<fp16, [2048, 512]> var_161_to_fp16 = const()[name = string("op_161_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5463168)))];
+            tensor<fp16, [2048]> var_162_to_fp16 = const()[name = string("op_162_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7560384)))];
+            tensor<fp16, [1, 1500, 2048]> linear_4_cast_fp16 = linear(bias = var_162_to_fp16, weight = var_161_to_fp16, x = var_152_cast_fp16)[name = string("linear_4_cast_fp16")];
+            string x_17_mode_0 = const()[name = string("x_17_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 1500, 2048]> x_17_cast_fp16 = gelu(mode = x_17_mode_0, x = linear_4_cast_fp16)[name = string("x_17_cast_fp16")];
+            tensor<fp16, [512, 2048]> var_167_to_fp16 = const()[name = string("op_167_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7564544)))];
+            tensor<fp16, [512]> var_168_to_fp16 = const()[name = string("op_168_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9661760)))];
+            tensor<fp16, [1, 1500, 512]> linear_5_cast_fp16 = linear(bias = var_168_to_fp16, weight = var_167_to_fp16, x = x_17_cast_fp16)[name = string("linear_5_cast_fp16")];
+            tensor<fp16, [1, 1500, 512]> x_19_cast_fp16 = add(x = x_13_cast_fp16, y = linear_5_cast_fp16)[name = string("x_19_cast_fp16")];
+            int32 var_178 = const()[name = string("op_178"), val = int32(-1)];
+            tensor<int32, [1]> var_194_axes_0 = const()[name = string("op_194_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> blocks_1_attn_ln_weight_to_fp16 = const()[name = string("blocks_1_attn_ln_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9662848)))];
+            tensor<fp16, [512]> blocks_1_attn_ln_bias_to_fp16 = const()[name = string("blocks_1_attn_ln_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9663936)))];
+            fp16 var_184_to_fp16 = const()[name = string("op_184_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1500, 512]> var_194_cast_fp16 = layer_norm(axes = var_194_axes_0, beta = blocks_1_attn_ln_bias_to_fp16, epsilon = var_184_to_fp16, gamma = blocks_1_attn_ln_weight_to_fp16, x = x_19_cast_fp16)[name = string("op_194_cast_fp16")];
+            tensor<fp16, [512, 512]> var_205_to_fp16 = const()[name = string("op_205_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9665024)))];
+            tensor<fp16, [512]> var_206_to_fp16 = const()[name = string("op_206_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10189376)))];
+            tensor<fp16, [1, 1500, 512]> linear_6_cast_fp16 = linear(bias = var_206_to_fp16, weight = var_205_to_fp16, x = var_194_cast_fp16)[name = string("linear_6_cast_fp16")];
+            tensor<fp16, [512, 512]> var_209_to_fp16 = const()[name = string("op_209_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10190464)))];
+            tensor<fp16, [1, 1500, 512]> linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_209_to_fp16, x = var_194_cast_fp16)[name = string("linear_7_cast_fp16")];
+            tensor<fp16, [512, 512]> var_213_to_fp16 = const()[name = string("op_213_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10714816)))];
+            tensor<fp16, [512]> var_214_to_fp16 = const()[name = string("op_214_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11239168)))];
+            tensor<fp16, [1, 1500, 512]> linear_8_cast_fp16 = linear(bias = var_214_to_fp16, weight = var_213_to_fp16, x = var_194_cast_fp16)[name = string("linear_8_cast_fp16")];
+            tensor<int32, [4]> var_222 = const()[name = string("op_222"), val = tensor<int32, [4]>([1, 1500, 8, -1])];
+            tensor<fp16, [1, 1500, 8, 64]> var_223_cast_fp16 = reshape(shape = var_222, x = linear_6_cast_fp16)[name = string("op_223_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 1]> const_44_to_fp16 = const()[name = string("const_44_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
+            tensor<fp16, [1, 1500, 8, 64]> q_7_cast_fp16 = mul(x = var_223_cast_fp16, y = const_44_to_fp16)[name = string("q_7_cast_fp16")];
+            tensor<int32, [4]> var_229 = const()[name = string("op_229"), val = tensor<int32, [4]>([1, 1500, 8, -1])];
+            tensor<fp16, [1, 1500, 8, 64]> var_230_cast_fp16 = reshape(shape = var_229, x = linear_7_cast_fp16)[name = string("op_230_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 1]> const_45_to_fp16 = const()[name = string("const_45_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
+            tensor<fp16, [1, 1500, 8, 64]> k_7_cast_fp16 = mul(x = var_230_cast_fp16, y = const_45_to_fp16)[name = string("k_7_cast_fp16")];
+            tensor<int32, [4]> var_236 = const()[name = string("op_236"), val = tensor<int32, [4]>([1, 1500, 8, -1])];
+            tensor<fp16, [1, 1500, 8, 64]> var_237_cast_fp16 = reshape(shape = var_236, x = linear_8_cast_fp16)[name = string("op_237_cast_fp16")];
+            tensor<int32, [4]> var_238 = const()[name = string("op_238"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            bool qk_3_transpose_x_0 = const()[name = string("qk_3_transpose_x_0"), val = bool(false)];
+            bool qk_3_transpose_y_0 = const()[name = string("qk_3_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_26_perm_0 = const()[name = string("transpose_26_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_27_perm_0 = const()[name = string("transpose_27_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 1500]> transpose_27 = transpose(perm = transpose_27_perm_0, x = k_7_cast_fp16)[name = string("transpose_53")];
+            tensor<fp16, [1, 8, 1500, 64]> transpose_26 = transpose(perm = transpose_26_perm_0, x = q_7_cast_fp16)[name = string("transpose_54")];
+            tensor<fp16, [1, 8, 1500, 1500]> qk_3_cast_fp16 = matmul(transpose_x = qk_3_transpose_x_0, transpose_y = qk_3_transpose_y_0, x = transpose_26, y = transpose_27)[name = string("qk_3_cast_fp16")];
+            tensor<fp16, [1, 8, 1500, 1500]> var_242_cast_fp16 = softmax(axis = var_178, x = qk_3_cast_fp16)[name = string("op_242_cast_fp16")];
+            bool var_244_transpose_x_0 = const()[name = string("op_244_transpose_x_0"), val = bool(false)];
+            bool var_244_transpose_y_0 = const()[name = string("op_244_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 1500, 64]> v_7_cast_fp16 = transpose(perm = var_238, x = var_237_cast_fp16)[name = string("transpose_55")];
+            tensor<fp16, [1, 8, 1500, 64]> var_244_cast_fp16 = matmul(transpose_x = var_244_transpose_x_0, transpose_y = var_244_transpose_y_0, x = var_242_cast_fp16, y = v_7_cast_fp16)[name = string("op_244_cast_fp16")];
+            tensor<int32, [4]> var_245 = const()[name = string("op_245"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> concat_1 = const()[name = string("concat_1"), val = tensor<int32, [3]>([1, 1500, 512])];
+            tensor<fp16, [1, 1500, 8, 64]> var_246_cast_fp16 = transpose(perm = var_245, x = var_244_cast_fp16)[name = string("transpose_52")];
+            tensor<fp16, [1, 1500, 512]> x_23_cast_fp16 = reshape(shape = concat_1, x = var_246_cast_fp16)[name = string("x_23_cast_fp16")];
+            tensor<fp16, [512, 512]> var_250_to_fp16 = const()[name = string("op_250_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11240256)))];
+            tensor<fp16, [512]> var_251_to_fp16 = const()[name = string("op_251_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11764608)))];
+            tensor<fp16, [1, 1500, 512]> linear_9_cast_fp16 = linear(bias = var_251_to_fp16, weight = var_250_to_fp16, x = x_23_cast_fp16)[name = string("linear_9_cast_fp16")];
+            tensor<fp16, [1, 1500, 512]> x_25_cast_fp16 = add(x = x_19_cast_fp16, y = linear_9_cast_fp16)[name = string("x_25_cast_fp16")];
+            tensor<int32, [1]> var_258_axes_0 = const()[name = string("op_258_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> blocks_1_mlp_ln_weight_to_fp16 = const()[name = string("blocks_1_mlp_ln_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11765696)))];
+            tensor<fp16, [512]> blocks_1_mlp_ln_bias_to_fp16 = const()[name = string("blocks_1_mlp_ln_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11766784)))];
+            tensor<fp16, [1, 1500, 512]> var_258_cast_fp16 = layer_norm(axes = var_258_axes_0, beta = blocks_1_mlp_ln_bias_to_fp16, epsilon = var_184_to_fp16, gamma = blocks_1_mlp_ln_weight_to_fp16, x = x_25_cast_fp16)[name = string("op_258_cast_fp16")];
+            tensor<fp16, [2048, 512]> var_267_to_fp16 = const()[name = string("op_267_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11767872)))];
+            tensor<fp16, [2048]> var_268_to_fp16 = const()[name = string("op_268_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13865088)))];
+            tensor<fp16, [1, 1500, 2048]> linear_10_cast_fp16 = linear(bias = var_268_to_fp16, weight = var_267_to_fp16, x = var_258_cast_fp16)[name = string("linear_10_cast_fp16")];
+            string x_29_mode_0 = const()[name = string("x_29_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 1500, 2048]> x_29_cast_fp16 = gelu(mode = x_29_mode_0, x = linear_10_cast_fp16)[name = string("x_29_cast_fp16")];
+            tensor<fp16, [512, 2048]> var_273_to_fp16 = const()[name = string("op_273_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13869248)))];
+            tensor<fp16, [512]> var_274_to_fp16 = const()[name = string("op_274_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15966464)))];
+            tensor<fp16, [1, 1500, 512]> linear_11_cast_fp16 = linear(bias = var_274_to_fp16, weight = var_273_to_fp16, x = x_29_cast_fp16)[name = string("linear_11_cast_fp16")];
+            tensor<fp16, [1, 1500, 512]> x_31_cast_fp16 = add(x = x_25_cast_fp16, y = linear_11_cast_fp16)[name = string("x_31_cast_fp16")];
+            int32 var_284 = const()[name = string("op_284"), val = int32(-1)];
+            tensor<int32, [1]> var_300_axes_0 = const()[name = string("op_300_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> blocks_2_attn_ln_weight_to_fp16 = const()[name = string("blocks_2_attn_ln_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15967552)))];
+            tensor<fp16, [512]> blocks_2_attn_ln_bias_to_fp16 = const()[name = string("blocks_2_attn_ln_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15968640)))];
+            fp16 var_290_to_fp16 = const()[name = string("op_290_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1500, 512]> var_300_cast_fp16 = layer_norm(axes = var_300_axes_0, beta = blocks_2_attn_ln_bias_to_fp16, epsilon = var_290_to_fp16, gamma = blocks_2_attn_ln_weight_to_fp16, x = x_31_cast_fp16)[name = string("op_300_cast_fp16")];
+            tensor<fp16, [512, 512]> var_311_to_fp16 = const()[name = string("op_311_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15969728)))];
+            tensor<fp16, [512]> var_312_to_fp16 = const()[name = string("op_312_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16494080)))];
+            tensor<fp16, [1, 1500, 512]> linear_12_cast_fp16 = linear(bias = var_312_to_fp16, weight = var_311_to_fp16, x = var_300_cast_fp16)[name = string("linear_12_cast_fp16")];
+            tensor<fp16, [512, 512]> var_315_to_fp16 = const()[name = string("op_315_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16495168)))];
+            tensor<fp16, [1, 1500, 512]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_315_to_fp16, x = var_300_cast_fp16)[name = string("linear_13_cast_fp16")];
+            tensor<fp16, [512, 512]> var_319_to_fp16 = const()[name = string("op_319_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17019520)))];
+            tensor<fp16, [512]> var_320_to_fp16 = const()[name = string("op_320_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17543872)))];
+            tensor<fp16, [1, 1500, 512]> linear_14_cast_fp16 = linear(bias = var_320_to_fp16, weight = var_319_to_fp16, x = var_300_cast_fp16)[name = string("linear_14_cast_fp16")];
+            tensor<int32, [4]> var_328 = const()[name = string("op_328"), val = tensor<int32, [4]>([1, 1500, 8, -1])];
+            tensor<fp16, [1, 1500, 8, 64]> var_329_cast_fp16 = reshape(shape = var_328, x = linear_12_cast_fp16)[name = string("op_329_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 1]> const_46_to_fp16 = const()[name = string("const_46_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
+            tensor<fp16, [1, 1500, 8, 64]> q_11_cast_fp16 = mul(x = var_329_cast_fp16, y = const_46_to_fp16)[name = string("q_11_cast_fp16")];
+            tensor<int32, [4]> var_335 = const()[name = string("op_335"), val = tensor<int32, [4]>([1, 1500, 8, -1])];
+            tensor<fp16, [1, 1500, 8, 64]> var_336_cast_fp16 = reshape(shape = var_335, x = linear_13_cast_fp16)[name = string("op_336_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 1]> const_47_to_fp16 = const()[name = string("const_47_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
+            tensor<fp16, [1, 1500, 8, 64]> k_11_cast_fp16 = mul(x = var_336_cast_fp16, y = const_47_to_fp16)[name = string("k_11_cast_fp16")];
+            tensor<int32, [4]> var_342 = const()[name = string("op_342"), val = tensor<int32, [4]>([1, 1500, 8, -1])];
+            tensor<fp16, [1, 1500, 8, 64]> var_343_cast_fp16 = reshape(shape = var_342, x = linear_14_cast_fp16)[name = string("op_343_cast_fp16")];
+            tensor<int32, [4]> var_344 = const()[name = string("op_344"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            bool qk_5_transpose_x_0 = const()[name = string("qk_5_transpose_x_0"), val = bool(false)];
+            bool qk_5_transpose_y_0 = const()[name = string("qk_5_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_28_perm_0 = const()[name = string("transpose_28_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_29_perm_0 = const()[name = string("transpose_29_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 1500]> transpose_29 = transpose(perm = transpose_29_perm_0, x = k_11_cast_fp16)[name = string("transpose_49")];
+            tensor<fp16, [1, 8, 1500, 64]> transpose_28 = transpose(perm = transpose_28_perm_0, x = q_11_cast_fp16)[name = string("transpose_50")];
+            tensor<fp16, [1, 8, 1500, 1500]> qk_5_cast_fp16 = matmul(transpose_x = qk_5_transpose_x_0, transpose_y = qk_5_transpose_y_0, x = transpose_28, y = transpose_29)[name = string("qk_5_cast_fp16")];
+            tensor<fp16, [1, 8, 1500, 1500]> var_348_cast_fp16 = softmax(axis = var_284, x = qk_5_cast_fp16)[name = string("op_348_cast_fp16")];
+            bool var_350_transpose_x_0 = const()[name = string("op_350_transpose_x_0"), val = bool(false)];
+            bool var_350_transpose_y_0 = const()[name = string("op_350_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 1500, 64]> v_11_cast_fp16 = transpose(perm = var_344, x = var_343_cast_fp16)[name = string("transpose_51")];
+            tensor<fp16, [1, 8, 1500, 64]> var_350_cast_fp16 = matmul(transpose_x = var_350_transpose_x_0, transpose_y = var_350_transpose_y_0, x = var_348_cast_fp16, y = v_11_cast_fp16)[name = string("op_350_cast_fp16")];
+            tensor<int32, [4]> var_351 = const()[name = string("op_351"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> concat_2 = const()[name = string("concat_2"), val = tensor<int32, [3]>([1, 1500, 512])];
+            tensor<fp16, [1, 1500, 8, 64]> var_352_cast_fp16 = transpose(perm = var_351, x = var_350_cast_fp16)[name = string("transpose_48")];
+            tensor<fp16, [1, 1500, 512]> x_35_cast_fp16 = reshape(shape = concat_2, x = var_352_cast_fp16)[name = string("x_35_cast_fp16")];
+            tensor<fp16, [512, 512]> var_356_to_fp16 = const()[name = string("op_356_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17544960)))];
+            tensor<fp16, [512]> var_357_to_fp16 = const()[name = string("op_357_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18069312)))];
+            tensor<fp16, [1, 1500, 512]> linear_15_cast_fp16 = linear(bias = var_357_to_fp16, weight = var_356_to_fp16, x = x_35_cast_fp16)[name = string("linear_15_cast_fp16")];
+            tensor<fp16, [1, 1500, 512]> x_37_cast_fp16 = add(x = x_31_cast_fp16, y = linear_15_cast_fp16)[name = string("x_37_cast_fp16")];
+            tensor<int32, [1]> var_364_axes_0 = const()[name = string("op_364_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> blocks_2_mlp_ln_weight_to_fp16 = const()[name = string("blocks_2_mlp_ln_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18070400)))];
+            tensor<fp16, [512]> blocks_2_mlp_ln_bias_to_fp16 = const()[name = string("blocks_2_mlp_ln_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18071488)))];
+            tensor<fp16, [1, 1500, 512]> var_364_cast_fp16 = layer_norm(axes = var_364_axes_0, beta = blocks_2_mlp_ln_bias_to_fp16, epsilon = var_290_to_fp16, gamma = blocks_2_mlp_ln_weight_to_fp16, x = x_37_cast_fp16)[name = string("op_364_cast_fp16")];
+            tensor<fp16, [2048, 512]> var_373_to_fp16 = const()[name = string("op_373_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18072576)))];
+            tensor<fp16, [2048]> var_374_to_fp16 = const()[name = string("op_374_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20169792)))];
+            tensor<fp16, [1, 1500, 2048]> linear_16_cast_fp16 = linear(bias = var_374_to_fp16, weight = var_373_to_fp16, x = var_364_cast_fp16)[name = string("linear_16_cast_fp16")];
+            string x_41_mode_0 = const()[name = string("x_41_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 1500, 2048]> x_41_cast_fp16 = gelu(mode = x_41_mode_0, x = linear_16_cast_fp16)[name = string("x_41_cast_fp16")];
+            tensor<fp16, [512, 2048]> var_379_to_fp16 = const()[name = string("op_379_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20173952)))];
+            tensor<fp16, [512]> var_380_to_fp16 = const()[name = string("op_380_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22271168)))];
+            tensor<fp16, [1, 1500, 512]> linear_17_cast_fp16 = linear(bias = var_380_to_fp16, weight = var_379_to_fp16, x = x_41_cast_fp16)[name = string("linear_17_cast_fp16")];
+            tensor<fp16, [1, 1500, 512]> x_43_cast_fp16 = add(x = x_37_cast_fp16, y = linear_17_cast_fp16)[name = string("x_43_cast_fp16")];
+            int32 var_390 = const()[name = string("op_390"), val = int32(-1)];
+            tensor<int32, [1]> var_406_axes_0 = const()[name = string("op_406_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> blocks_3_attn_ln_weight_to_fp16 = const()[name = string("blocks_3_attn_ln_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22272256)))];
+            tensor<fp16, [512]> blocks_3_attn_ln_bias_to_fp16 = const()[name = string("blocks_3_attn_ln_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22273344)))];
+            fp16 var_396_to_fp16 = const()[name = string("op_396_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1500, 512]> var_406_cast_fp16 = layer_norm(axes = var_406_axes_0, beta = blocks_3_attn_ln_bias_to_fp16, epsilon = var_396_to_fp16, gamma = blocks_3_attn_ln_weight_to_fp16, x = x_43_cast_fp16)[name = string("op_406_cast_fp16")];
+            tensor<fp16, [512, 512]> var_417_to_fp16 = const()[name = string("op_417_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22274432)))];
+            tensor<fp16, [512]> var_418_to_fp16 = const()[name = string("op_418_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22798784)))];
+            tensor<fp16, [1, 1500, 512]> linear_18_cast_fp16 = linear(bias = var_418_to_fp16, weight = var_417_to_fp16, x = var_406_cast_fp16)[name = string("linear_18_cast_fp16")];
+            tensor<fp16, [512, 512]> var_421_to_fp16 = const()[name = string("op_421_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22799872)))];
+            tensor<fp16, [1, 1500, 512]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_421_to_fp16, x = var_406_cast_fp16)[name = string("linear_19_cast_fp16")];
+            tensor<fp16, [512, 512]> var_425_to_fp16 = const()[name = string("op_425_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23324224)))];
+            tensor<fp16, [512]> var_426_to_fp16 = const()[name = string("op_426_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23848576)))];
+            tensor<fp16, [1, 1500, 512]> linear_20_cast_fp16 = linear(bias = var_426_to_fp16, weight = var_425_to_fp16, x = var_406_cast_fp16)[name = string("linear_20_cast_fp16")];
+            tensor<int32, [4]> var_434 = const()[name = string("op_434"), val = tensor<int32, [4]>([1, 1500, 8, -1])];
+            tensor<fp16, [1, 1500, 8, 64]> var_435_cast_fp16 = reshape(shape = var_434, x = linear_18_cast_fp16)[name = string("op_435_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 1]> const_48_to_fp16 = const()[name = string("const_48_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
+            tensor<fp16, [1, 1500, 8, 64]> q_15_cast_fp16 = mul(x = var_435_cast_fp16, y = const_48_to_fp16)[name = string("q_15_cast_fp16")];
+            tensor<int32, [4]> var_441 = const()[name = string("op_441"), val = tensor<int32, [4]>([1, 1500, 8, -1])];
+            tensor<fp16, [1, 1500, 8, 64]> var_442_cast_fp16 = reshape(shape = var_441, x = linear_19_cast_fp16)[name = string("op_442_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 1]> const_49_to_fp16 = const()[name = string("const_49_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
+            tensor<fp16, [1, 1500, 8, 64]> k_15_cast_fp16 = mul(x = var_442_cast_fp16, y = const_49_to_fp16)[name = string("k_15_cast_fp16")];
+            tensor<int32, [4]> var_448 = const()[name = string("op_448"), val = tensor<int32, [4]>([1, 1500, 8, -1])];
+            tensor<fp16, [1, 1500, 8, 64]> var_449_cast_fp16 = reshape(shape = var_448, x = linear_20_cast_fp16)[name = string("op_449_cast_fp16")];
+            tensor<int32, [4]> var_450 = const()[name = string("op_450"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            bool qk_7_transpose_x_0 = const()[name = string("qk_7_transpose_x_0"), val = bool(false)];
+            bool qk_7_transpose_y_0 = const()[name = string("qk_7_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_30_perm_0 = const()[name = string("transpose_30_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_31_perm_0 = const()[name = string("transpose_31_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 1500]> transpose_31 = transpose(perm = transpose_31_perm_0, x = k_15_cast_fp16)[name = string("transpose_45")];
+            tensor<fp16, [1, 8, 1500, 64]> transpose_30 = transpose(perm = transpose_30_perm_0, x = q_15_cast_fp16)[name = string("transpose_46")];
+            tensor<fp16, [1, 8, 1500, 1500]> qk_7_cast_fp16 = matmul(transpose_x = qk_7_transpose_x_0, transpose_y = qk_7_transpose_y_0, x = transpose_30, y = transpose_31)[name = string("qk_7_cast_fp16")];
+            tensor<fp16, [1, 8, 1500, 1500]> var_454_cast_fp16 = softmax(axis = var_390, x = qk_7_cast_fp16)[name = string("op_454_cast_fp16")];
+            bool var_456_transpose_x_0 = const()[name = string("op_456_transpose_x_0"), val = bool(false)];
+            bool var_456_transpose_y_0 = const()[name = string("op_456_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 1500, 64]> v_15_cast_fp16 = transpose(perm = var_450, x = var_449_cast_fp16)[name = string("transpose_47")];
+            tensor<fp16, [1, 8, 1500, 64]> var_456_cast_fp16 = matmul(transpose_x = var_456_transpose_x_0, transpose_y = var_456_transpose_y_0, x = var_454_cast_fp16, y = v_15_cast_fp16)[name = string("op_456_cast_fp16")];
+            tensor<int32, [4]> var_457 = const()[name = string("op_457"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> concat_3 = const()[name = string("concat_3"), val = tensor<int32, [3]>([1, 1500, 512])];
+            tensor<fp16, [1, 1500, 8, 64]> var_458_cast_fp16 = transpose(perm = var_457, x = var_456_cast_fp16)[name = string("transpose_44")];
+            tensor<fp16, [1, 1500, 512]> x_47_cast_fp16 = reshape(shape = concat_3, x = var_458_cast_fp16)[name = string("x_47_cast_fp16")];
+            tensor<fp16, [512, 512]> var_462_to_fp16 = const()[name = string("op_462_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23849664)))];
+            tensor<fp16, [512]> var_463_to_fp16 = const()[name = string("op_463_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24374016)))];
+            tensor<fp16, [1, 1500, 512]> linear_21_cast_fp16 = linear(bias = var_463_to_fp16, weight = var_462_to_fp16, x = x_47_cast_fp16)[name = string("linear_21_cast_fp16")];
+            tensor<fp16, [1, 1500, 512]> x_49_cast_fp16 = add(x = x_43_cast_fp16, y = linear_21_cast_fp16)[name = string("x_49_cast_fp16")];
+            tensor<int32, [1]> var_470_axes_0 = const()[name = string("op_470_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> blocks_3_mlp_ln_weight_to_fp16 = const()[name = string("blocks_3_mlp_ln_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24375104)))];
+            tensor<fp16, [512]> blocks_3_mlp_ln_bias_to_fp16 = const()[name = string("blocks_3_mlp_ln_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24376192)))];
+            tensor<fp16, [1, 1500, 512]> var_470_cast_fp16 = layer_norm(axes = var_470_axes_0, beta = blocks_3_mlp_ln_bias_to_fp16, epsilon = var_396_to_fp16, gamma = blocks_3_mlp_ln_weight_to_fp16, x = x_49_cast_fp16)[name = string("op_470_cast_fp16")];
+            tensor<fp16, [2048, 512]> var_479_to_fp16 = const()[name = string("op_479_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24377280)))];
+            tensor<fp16, [2048]> var_480_to_fp16 = const()[name = string("op_480_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26474496)))];
+            tensor<fp16, [1, 1500, 2048]> linear_22_cast_fp16 = linear(bias = var_480_to_fp16, weight = var_479_to_fp16, x = var_470_cast_fp16)[name = string("linear_22_cast_fp16")];
+            string x_53_mode_0 = const()[name = string("x_53_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 1500, 2048]> x_53_cast_fp16 = gelu(mode = x_53_mode_0, x = linear_22_cast_fp16)[name = string("x_53_cast_fp16")];
+            tensor<fp16, [512, 2048]> var_485_to_fp16 = const()[name = string("op_485_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26478656)))];
+            tensor<fp16, [512]> var_486_to_fp16 = const()[name = string("op_486_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28575872)))];
+            tensor<fp16, [1, 1500, 512]> linear_23_cast_fp16 = linear(bias = var_486_to_fp16, weight = var_485_to_fp16, x = x_53_cast_fp16)[name = string("linear_23_cast_fp16")];
+            tensor<fp16, [1, 1500, 512]> x_55_cast_fp16 = add(x = x_49_cast_fp16, y = linear_23_cast_fp16)[name = string("x_55_cast_fp16")];
+            int32 var_496 = const()[name = string("op_496"), val = int32(-1)];
+            tensor<int32, [1]> var_512_axes_0 = const()[name = string("op_512_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> blocks_4_attn_ln_weight_to_fp16 = const()[name = string("blocks_4_attn_ln_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28576960)))];
+            tensor<fp16, [512]> blocks_4_attn_ln_bias_to_fp16 = const()[name = string("blocks_4_attn_ln_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28578048)))];
+            fp16 var_502_to_fp16 = const()[name = string("op_502_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1500, 512]> var_512_cast_fp16 = layer_norm(axes = var_512_axes_0, beta = blocks_4_attn_ln_bias_to_fp16, epsilon = var_502_to_fp16, gamma = blocks_4_attn_ln_weight_to_fp16, x = x_55_cast_fp16)[name = string("op_512_cast_fp16")];
+            tensor<fp16, [512, 512]> var_523_to_fp16 = const()[name = string("op_523_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28579136)))];
+            tensor<fp16, [512]> var_524_to_fp16 = const()[name = string("op_524_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29103488)))];
+            tensor<fp16, [1, 1500, 512]> linear_24_cast_fp16 = linear(bias = var_524_to_fp16, weight = var_523_to_fp16, x = var_512_cast_fp16)[name = string("linear_24_cast_fp16")];
+            tensor<fp16, [512, 512]> var_527_to_fp16 = const()[name = string("op_527_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29104576)))];
+            tensor<fp16, [1, 1500, 512]> linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_527_to_fp16, x = var_512_cast_fp16)[name = string("linear_25_cast_fp16")];
+            tensor<fp16, [512, 512]> var_531_to_fp16 = const()[name = string("op_531_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29628928)))];
+            tensor<fp16, [512]> var_532_to_fp16 = const()[name = string("op_532_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30153280)))];
+            tensor<fp16, [1, 1500, 512]> linear_26_cast_fp16 = linear(bias = var_532_to_fp16, weight = var_531_to_fp16, x = var_512_cast_fp16)[name = string("linear_26_cast_fp16")];
+            tensor<int32, [4]> var_540 = const()[name = string("op_540"), val = tensor<int32, [4]>([1, 1500, 8, -1])];
+            tensor<fp16, [1, 1500, 8, 64]> var_541_cast_fp16 = reshape(shape = var_540, x = linear_24_cast_fp16)[name = string("op_541_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 1]> const_50_to_fp16 = const()[name = string("const_50_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
+            tensor<fp16, [1, 1500, 8, 64]> q_19_cast_fp16 = mul(x = var_541_cast_fp16, y = const_50_to_fp16)[name = string("q_19_cast_fp16")];
+            tensor<int32, [4]> var_547 = const()[name = string("op_547"), val = tensor<int32, [4]>([1, 1500, 8, -1])];
+            tensor<fp16, [1, 1500, 8, 64]> var_548_cast_fp16 = reshape(shape = var_547, x = linear_25_cast_fp16)[name = string("op_548_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 1]> const_51_to_fp16 = const()[name = string("const_51_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
+            tensor<fp16, [1, 1500, 8, 64]> k_19_cast_fp16 = mul(x = var_548_cast_fp16, y = const_51_to_fp16)[name = string("k_19_cast_fp16")];
+            tensor<int32, [4]> var_554 = const()[name = string("op_554"), val = tensor<int32, [4]>([1, 1500, 8, -1])];
+            tensor<fp16, [1, 1500, 8, 64]> var_555_cast_fp16 = reshape(shape = var_554, x = linear_26_cast_fp16)[name = string("op_555_cast_fp16")];
+            tensor<int32, [4]> var_556 = const()[name = string("op_556"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            bool qk_9_transpose_x_0 = const()[name = string("qk_9_transpose_x_0"), val = bool(false)];
+            bool qk_9_transpose_y_0 = const()[name = string("qk_9_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_32_perm_0 = const()[name = string("transpose_32_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 1500]> transpose_33 = transpose(perm = transpose_33_perm_0, x = k_19_cast_fp16)[name = string("transpose_41")];
+            tensor<fp16, [1, 8, 1500, 64]> transpose_32 = transpose(perm = transpose_32_perm_0, x = q_19_cast_fp16)[name = string("transpose_42")];
+            tensor<fp16, [1, 8, 1500, 1500]> qk_9_cast_fp16 = matmul(transpose_x = qk_9_transpose_x_0, transpose_y = qk_9_transpose_y_0, x = transpose_32, y = transpose_33)[name = string("qk_9_cast_fp16")];
+            tensor<fp16, [1, 8, 1500, 1500]> var_560_cast_fp16 = softmax(axis = var_496, x = qk_9_cast_fp16)[name = string("op_560_cast_fp16")];
+            bool var_562_transpose_x_0 = const()[name = string("op_562_transpose_x_0"), val = bool(false)];
+            bool var_562_transpose_y_0 = const()[name = string("op_562_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 1500, 64]> v_19_cast_fp16 = transpose(perm = var_556, x = var_555_cast_fp16)[name = string("transpose_43")];
+            tensor<fp16, [1, 8, 1500, 64]> var_562_cast_fp16 = matmul(transpose_x = var_562_transpose_x_0, transpose_y = var_562_transpose_y_0, x = var_560_cast_fp16, y = v_19_cast_fp16)[name = string("op_562_cast_fp16")];
+            tensor<int32, [4]> var_563 = const()[name = string("op_563"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> concat_4 = const()[name = string("concat_4"), val = tensor<int32, [3]>([1, 1500, 512])];
+            tensor<fp16, [1, 1500, 8, 64]> var_564_cast_fp16 = transpose(perm = var_563, x = var_562_cast_fp16)[name = string("transpose_40")];
+            tensor<fp16, [1, 1500, 512]> x_59_cast_fp16 = reshape(shape = concat_4, x = var_564_cast_fp16)[name = string("x_59_cast_fp16")];
+            tensor<fp16, [512, 512]> var_568_to_fp16 = const()[name = string("op_568_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30154368)))];
+            tensor<fp16, [512]> var_569_to_fp16 = const()[name = string("op_569_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30678720)))];
+            tensor<fp16, [1, 1500, 512]> linear_27_cast_fp16 = linear(bias = var_569_to_fp16, weight = var_568_to_fp16, x = x_59_cast_fp16)[name = string("linear_27_cast_fp16")];
+            tensor<fp16, [1, 1500, 512]> x_61_cast_fp16 = add(x = x_55_cast_fp16, y = linear_27_cast_fp16)[name = string("x_61_cast_fp16")];
+            tensor<int32, [1]> var_576_axes_0 = const()[name = string("op_576_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> blocks_4_mlp_ln_weight_to_fp16 = const()[name = string("blocks_4_mlp_ln_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30679808)))];
+            tensor<fp16, [512]> blocks_4_mlp_ln_bias_to_fp16 = const()[name = string("blocks_4_mlp_ln_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30680896)))];
+            tensor<fp16, [1, 1500, 512]> var_576_cast_fp16 = layer_norm(axes = var_576_axes_0, beta = blocks_4_mlp_ln_bias_to_fp16, epsilon = var_502_to_fp16, gamma = blocks_4_mlp_ln_weight_to_fp16, x = x_61_cast_fp16)[name = string("op_576_cast_fp16")];
+            tensor<fp16, [2048, 512]> var_585_to_fp16 = const()[name = string("op_585_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30681984)))];
+            tensor<fp16, [2048]> var_586_to_fp16 = const()[name = string("op_586_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32779200)))];
+            tensor<fp16, [1, 1500, 2048]> linear_28_cast_fp16 = linear(bias = var_586_to_fp16, weight = var_585_to_fp16, x = var_576_cast_fp16)[name = string("linear_28_cast_fp16")];
+            string x_65_mode_0 = const()[name = string("x_65_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 1500, 2048]> x_65_cast_fp16 = gelu(mode = x_65_mode_0, x = linear_28_cast_fp16)[name = string("x_65_cast_fp16")];
+            tensor<fp16, [512, 2048]> var_591_to_fp16 = const()[name = string("op_591_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32783360)))];
+            tensor<fp16, [512]> var_592_to_fp16 = const()[name = string("op_592_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34880576)))];
+            tensor<fp16, [1, 1500, 512]> linear_29_cast_fp16 = linear(bias = var_592_to_fp16, weight = var_591_to_fp16, x = x_65_cast_fp16)[name = string("linear_29_cast_fp16")];
+            tensor<fp16, [1, 1500, 512]> x_67_cast_fp16 = add(x = x_61_cast_fp16, y = linear_29_cast_fp16)[name = string("x_67_cast_fp16")];
+            int32 var_602 = const()[name = string("op_602"), val = int32(-1)];
+            tensor<int32, [1]> var_618_axes_0 = const()[name = string("op_618_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> blocks_5_attn_ln_weight_to_fp16 = const()[name = string("blocks_5_attn_ln_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34881664)))];
+            tensor<fp16, [512]> blocks_5_attn_ln_bias_to_fp16 = const()[name = string("blocks_5_attn_ln_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34882752)))];
+            fp16 var_608_to_fp16 = const()[name = string("op_608_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1500, 512]> var_618_cast_fp16 = layer_norm(axes = var_618_axes_0, beta = blocks_5_attn_ln_bias_to_fp16, epsilon = var_608_to_fp16, gamma = blocks_5_attn_ln_weight_to_fp16, x = x_67_cast_fp16)[name = string("op_618_cast_fp16")];
+            tensor<fp16, [512, 512]> var_629_to_fp16 = const()[name = string("op_629_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34883840)))];
+            tensor<fp16, [512]> var_630_to_fp16 = const()[name = string("op_630_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35408192)))];
+            tensor<fp16, [1, 1500, 512]> linear_30_cast_fp16 = linear(bias = var_630_to_fp16, weight = var_629_to_fp16, x = var_618_cast_fp16)[name = string("linear_30_cast_fp16")];
+            tensor<fp16, [512, 512]> var_633_to_fp16 = const()[name = string("op_633_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35409280)))];
+            tensor<fp16, [1, 1500, 512]> linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_633_to_fp16, x = var_618_cast_fp16)[name = string("linear_31_cast_fp16")];
+            tensor<fp16, [512, 512]> var_637_to_fp16 = const()[name = string("op_637_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35933632)))];
+            tensor<fp16, [512]> var_638_to_fp16 = const()[name = string("op_638_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36457984)))];
+            tensor<fp16, [1, 1500, 512]> linear_32_cast_fp16 = linear(bias = var_638_to_fp16, weight = var_637_to_fp16, x = var_618_cast_fp16)[name = string("linear_32_cast_fp16")];
+            tensor<int32, [4]> var_646 = const()[name = string("op_646"), val = tensor<int32, [4]>([1, 1500, 8, -1])];
+            tensor<fp16, [1, 1500, 8, 64]> var_647_cast_fp16 = reshape(shape = var_646, x = linear_30_cast_fp16)[name = string("op_647_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 1]> const_52_to_fp16 = const()[name = string("const_52_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
+            tensor<fp16, [1, 1500, 8, 64]> q_cast_fp16 = mul(x = var_647_cast_fp16, y = const_52_to_fp16)[name = string("q_cast_fp16")];
+            tensor<int32, [4]> var_653 = const()[name = string("op_653"), val = tensor<int32, [4]>([1, 1500, 8, -1])];
+            tensor<fp16, [1, 1500, 8, 64]> var_654_cast_fp16 = reshape(shape = var_653, x = linear_31_cast_fp16)[name = string("op_654_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 1]> const_53_to_fp16 = const()[name = string("const_53_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])];
+            tensor<fp16, [1, 1500, 8, 64]> k_cast_fp16 = mul(x = var_654_cast_fp16, y = const_53_to_fp16)[name = string("k_cast_fp16")];
+            tensor<int32, [4]> var_660 = const()[name = string("op_660"), val = tensor<int32, [4]>([1, 1500, 8, -1])];
+            tensor<fp16, [1, 1500, 8, 64]> var_661_cast_fp16 = reshape(shape = var_660, x = linear_32_cast_fp16)[name = string("op_661_cast_fp16")];
+            tensor<int32, [4]> var_662 = const()[name = string("op_662"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            bool qk_transpose_x_0 = const()[name = string("qk_transpose_x_0"), val = bool(false)];
+            bool qk_transpose_y_0 = const()[name = string("qk_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_35_perm_0 = const()[name = string("transpose_35_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 1500]> transpose_35 = transpose(perm = transpose_35_perm_0, x = k_cast_fp16)[name = string("transpose_37")];
+            tensor<fp16, [1, 8, 1500, 64]> transpose_34 = transpose(perm = transpose_34_perm_0, x = q_cast_fp16)[name = string("transpose_38")];
+            tensor<fp16, [1, 8, 1500, 1500]> qk_cast_fp16 = matmul(transpose_x = qk_transpose_x_0, transpose_y = qk_transpose_y_0, x = transpose_34, y = transpose_35)[name = string("qk_cast_fp16")];
+            tensor<fp16, [1, 8, 1500, 1500]> var_666_cast_fp16 = softmax(axis = var_602, x = qk_cast_fp16)[name = string("op_666_cast_fp16")];
+            bool var_668_transpose_x_0 = const()[name = string("op_668_transpose_x_0"), val = bool(false)];
+            bool var_668_transpose_y_0 = const()[name = string("op_668_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 1500, 64]> v_cast_fp16 = transpose(perm = var_662, x = var_661_cast_fp16)[name = string("transpose_39")];
+            tensor<fp16, [1, 8, 1500, 64]> var_668_cast_fp16 = matmul(transpose_x = var_668_transpose_x_0, transpose_y = var_668_transpose_y_0, x = var_666_cast_fp16, y = v_cast_fp16)[name = string("op_668_cast_fp16")];
+            tensor<int32, [4]> var_669 = const()[name = string("op_669"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> concat_5 = const()[name = string("concat_5"), val = tensor<int32, [3]>([1, 1500, 512])];
+            tensor<fp16, [1, 1500, 8, 64]> var_670_cast_fp16 = transpose(perm = var_669, x = var_668_cast_fp16)[name = string("transpose_36")];
+            tensor<fp16, [1, 1500, 512]> x_71_cast_fp16 = reshape(shape = concat_5, x = var_670_cast_fp16)[name = string("x_71_cast_fp16")];
+            tensor<fp16, [512, 512]> var_674_to_fp16 = const()[name = string("op_674_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36459072)))];
+            tensor<fp16, [512]> var_675_to_fp16 = const()[name = string("op_675_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36983424)))];
+            tensor<fp16, [1, 1500, 512]> linear_33_cast_fp16 = linear(bias = var_675_to_fp16, weight = var_674_to_fp16, x = x_71_cast_fp16)[name = string("linear_33_cast_fp16")];
+            tensor<fp16, [1, 1500, 512]> x_73_cast_fp16 = add(x = x_67_cast_fp16, y = linear_33_cast_fp16)[name = string("x_73_cast_fp16")];
+            tensor<int32, [1]> var_682_axes_0 = const()[name = string("op_682_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> blocks_5_mlp_ln_weight_to_fp16 = const()[name = string("blocks_5_mlp_ln_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36984512)))];
+            tensor<fp16, [512]> blocks_5_mlp_ln_bias_to_fp16 = const()[name = string("blocks_5_mlp_ln_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36985600)))];
+            tensor<fp16, [1, 1500, 512]> var_682_cast_fp16 = layer_norm(axes = var_682_axes_0, beta = blocks_5_mlp_ln_bias_to_fp16, epsilon = var_608_to_fp16, gamma = blocks_5_mlp_ln_weight_to_fp16, x = x_73_cast_fp16)[name = string("op_682_cast_fp16")];
+            tensor<fp16, [2048, 512]> var_691_to_fp16 = const()[name = string("op_691_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36986688)))];
+            tensor<fp16, [2048]> var_692_to_fp16 = const()[name = string("op_692_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39083904)))];
+            tensor<fp16, [1, 1500, 2048]> linear_34_cast_fp16 = linear(bias = var_692_to_fp16, weight = var_691_to_fp16, x = var_682_cast_fp16)[name = string("linear_34_cast_fp16")];
+            string x_77_mode_0 = const()[name = string("x_77_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 1500, 2048]> x_77_cast_fp16 = gelu(mode = x_77_mode_0, x = linear_34_cast_fp16)[name = string("x_77_cast_fp16")];
+            tensor<fp16, [512, 2048]> var_697_to_fp16 = const()[name = string("op_697_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39088064)))];
+            tensor<fp16, [512]> var_698_to_fp16 = const()[name = string("op_698_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41185280)))];
+            tensor<fp16, [1, 1500, 512]> linear_35_cast_fp16 = linear(bias = var_698_to_fp16, weight = var_697_to_fp16, x = x_77_cast_fp16)[name = string("linear_35_cast_fp16")];
+            tensor<fp16, [1, 1500, 512]> x_cast_fp16 = add(x = x_73_cast_fp16, y = linear_35_cast_fp16)[name = string("x_cast_fp16")];
+            tensor<int32, [1]> var_711_axes_0 = const()[name = string("op_711_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> ln_post_weight_to_fp16 = const()[name = string("ln_post_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41186368)))];
+            tensor<fp16, [512]> ln_post_bias_to_fp16 = const()[name = string("ln_post_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41187456)))];
+            fp16 var_702_to_fp16 = const()[name = string("op_702_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1500, 512]> output = layer_norm(axes = var_711_axes_0, beta = ln_post_bias_to_fp16, epsilon = var_702_to_fp16, gamma = ln_post_weight_to_fp16, x = x_cast_fp16)[name = string("op_711_cast_fp16")];
+        } -> (output);
+}

base/encoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c45bee989219532c4cec616d439c51f280ac9d7b04f7847c4b7d7daba1d47523
+size 41188544

base/model_dims.json

@@ -0,0 +1,12 @@
+{
+  "n_mels": 80,
+  "n_audio_ctx": 1500,
+  "n_audio_state": 512,
+  "n_audio_head": 8,
+  "n_audio_layer": 6,
+  "n_vocab": 51865,
+  "n_text_ctx": 448,
+  "n_text_state": 512,
+  "n_text_head": 8,
+  "n_text_layer": 6
+}

compile_model.sh

@@ -0,0 +1,34 @@
+#!/bin/bash
+
+for d in work/*
+do
+    echo $d
+    pushd $d >/dev/null
+
+    if [ -d encoder ]; then
+        xcrun coremlcompiler compile encoder/chunked_pipeline.mlpackage .
+        rm -rf encoder.mlmodelc
+        mv chunked_pipeline.mlmodelc encoder.mlmodelc
+    else
+        xcrun coremlcompiler compile encoder.mlpackage .
+    fi
+    xcrun coremlcompiler compile decoder_first.mlpackage .
+    xcrun coremlcompiler compile decoder_second.mlpackage .
+
+    popd >/dev/null
+done
+
+mkdir -p output
+for d in work/*
+do
+    out=${d/work/output}
+    mkdir -p $out
+    mv $d/*.mlmodelc $d/model_dims.json $out/
+done
+
+mkdir -p index
+for d in output/*
+do
+    model=${d##*/}
+    (cd $d && find * -type f) > index/$model
+done

index/base

@@ -0,0 +1,16 @@
+decoder_first.mlmodelc/weights/weight.bin
+decoder_first.mlmodelc/metadata.json
+decoder_first.mlmodelc/model.mil
+decoder_first.mlmodelc/coremldata.bin
+decoder_first.mlmodelc/analytics/coremldata.bin
+decoder_second.mlmodelc/weights/weight.bin
+decoder_second.mlmodelc/metadata.json
+decoder_second.mlmodelc/model.mil
+decoder_second.mlmodelc/coremldata.bin
+decoder_second.mlmodelc/analytics/coremldata.bin
+encoder.mlmodelc/weights/weight.bin
+encoder.mlmodelc/metadata.json
+encoder.mlmodelc/model.mil
+encoder.mlmodelc/coremldata.bin
+encoder.mlmodelc/analytics/coremldata.bin
+model_dims.json

index/large-v2

@@ -0,0 +1,22 @@
+decoder_first.mlmodelc/weights/weight.bin
+decoder_first.mlmodelc/metadata.json
+decoder_first.mlmodelc/model.mil
+decoder_first.mlmodelc/coremldata.bin
+decoder_first.mlmodelc/analytics/coremldata.bin
+decoder_second.mlmodelc/weights/weight.bin
+decoder_second.mlmodelc/metadata.json
+decoder_second.mlmodelc/model.mil
+decoder_second.mlmodelc/coremldata.bin
+decoder_second.mlmodelc/analytics/coremldata.bin
+encoder.mlmodelc/metadata.json
+encoder.mlmodelc/model0/weights/0-weight.bin
+encoder.mlmodelc/model0/model.mil
+encoder.mlmodelc/model0/coremldata.bin
+encoder.mlmodelc/model0/analytics/coremldata.bin
+encoder.mlmodelc/model1/weights/1-weight.bin
+encoder.mlmodelc/model1/model.mil
+encoder.mlmodelc/model1/coremldata.bin
+encoder.mlmodelc/model1/analytics/coremldata.bin
+encoder.mlmodelc/coremldata.bin
+encoder.mlmodelc/analytics/coremldata.bin
+model_dims.json

index/large-v3

@@ -0,0 +1,22 @@
+decoder_first.mlmodelc/weights/weight.bin
+decoder_first.mlmodelc/metadata.json
+decoder_first.mlmodelc/model.mil
+decoder_first.mlmodelc/coremldata.bin
+decoder_first.mlmodelc/analytics/coremldata.bin
+decoder_second.mlmodelc/weights/weight.bin
+decoder_second.mlmodelc/metadata.json
+decoder_second.mlmodelc/model.mil
+decoder_second.mlmodelc/coremldata.bin
+decoder_second.mlmodelc/analytics/coremldata.bin
+encoder.mlmodelc/metadata.json
+encoder.mlmodelc/model0/weights/0-weight.bin
+encoder.mlmodelc/model0/model.mil
+encoder.mlmodelc/model0/coremldata.bin
+encoder.mlmodelc/model0/analytics/coremldata.bin
+encoder.mlmodelc/model1/weights/1-weight.bin
+encoder.mlmodelc/model1/model.mil
+encoder.mlmodelc/model1/coremldata.bin
+encoder.mlmodelc/model1/analytics/coremldata.bin
+encoder.mlmodelc/coremldata.bin
+encoder.mlmodelc/analytics/coremldata.bin
+model_dims.json

index/medium

@@ -0,0 +1,16 @@
+decoder_first.mlmodelc/weights/weight.bin
+decoder_first.mlmodelc/metadata.json
+decoder_first.mlmodelc/model.mil
+decoder_first.mlmodelc/coremldata.bin
+decoder_first.mlmodelc/analytics/coremldata.bin
+decoder_second.mlmodelc/weights/weight.bin
+decoder_second.mlmodelc/metadata.json
+decoder_second.mlmodelc/model.mil
+decoder_second.mlmodelc/coremldata.bin
+decoder_second.mlmodelc/analytics/coremldata.bin
+encoder.mlmodelc/weights/weight.bin
+encoder.mlmodelc/metadata.json
+encoder.mlmodelc/model.mil
+encoder.mlmodelc/coremldata.bin
+encoder.mlmodelc/analytics/coremldata.bin
+model_dims.json

index/small

@@ -0,0 +1,16 @@
+decoder_first.mlmodelc/weights/weight.bin
+decoder_first.mlmodelc/metadata.json
+decoder_first.mlmodelc/model.mil
+decoder_first.mlmodelc/coremldata.bin
+decoder_first.mlmodelc/analytics/coremldata.bin
+decoder_second.mlmodelc/weights/weight.bin
+decoder_second.mlmodelc/metadata.json
+decoder_second.mlmodelc/model.mil
+decoder_second.mlmodelc/coremldata.bin
+decoder_second.mlmodelc/analytics/coremldata.bin
+encoder.mlmodelc/weights/weight.bin
+encoder.mlmodelc/metadata.json
+encoder.mlmodelc/model.mil
+encoder.mlmodelc/coremldata.bin
+encoder.mlmodelc/analytics/coremldata.bin
+model_dims.json

index/tiny

@@ -0,0 +1,16 @@
+decoder_first.mlmodelc/weights/weight.bin
+decoder_first.mlmodelc/metadata.json
+decoder_first.mlmodelc/model.mil
+decoder_first.mlmodelc/coremldata.bin
+decoder_first.mlmodelc/analytics/coremldata.bin
+decoder_second.mlmodelc/weights/weight.bin
+decoder_second.mlmodelc/metadata.json
+decoder_second.mlmodelc/model.mil
+decoder_second.mlmodelc/coremldata.bin
+decoder_second.mlmodelc/analytics/coremldata.bin
+encoder.mlmodelc/weights/weight.bin
+encoder.mlmodelc/metadata.json
+encoder.mlmodelc/model.mil
+encoder.mlmodelc/coremldata.bin
+encoder.mlmodelc/analytics/coremldata.bin
+model_dims.json

large-v2/decoder_first.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a564dfd67cfcb3c0ee8cd9f7ef9f303fbfc561e635709bd3a46c5870571079de
+size 243

large-v2/decoder_first.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6214be9e110a102836fb1fdb960a2fb564e60f5d9e3d1e25a9b7f978309480e
+size 453

large-v2/decoder_first.mlmodelc/metadata.json

@@ -0,0 +1,106 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16)",
+        "shortDescription" : "",
+        "shape" : "[]",
+        "name" : "dummy",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.writeState" : 66,
+      "Shape" : 64,
+      "Ios18.linear" : 64,
+      "Identity" : 1,
+      "Ios18.gather" : 64,
+      "Ios18.concat" : 64,
+      "Ios18.sliceUpdate" : 66,
+      "Ios18.cast" : 128,
+      "Ios18.expandDims" : 64,
+      "Ios18.readState" : 66
+    },
+    "computePrecision" : "Mixed (Float16, Int16, Int32, UInt16)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 32 × 1 × 448 × 1280)",
+        "shortDescription" : "",
+        "shape" : "[32, 1, 448, 1280]",
+        "name" : "k_cache1",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 32 × 1 × 448 × 1280)",
+        "shortDescription" : "",
+        "shape" : "[32, 1, 448, 1280]",
+        "name" : "v_cache1",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 32 × 1 × 1500 × 1280)",
+        "shortDescription" : "",
+        "shape" : "[32, 1, 1500, 1280]",
+        "name" : "k_cache2",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 32 × 1 × 1500 × 1280)",
+        "shortDescription" : "",
+        "shape" : "[32, 1, 1500, 1280]",
+        "name" : "v_cache2",
+        "type" : "State"
+      }
+    ],
+    "availability" : {
+      "macOS" : "15.0",
+      "tvOS" : "18.0",
+      "visionOS" : "2.0",
+      "watchOS" : "11.0",
+      "iOS" : "18.0",
+      "macCatalyst" : "18.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.4.1",
+      "com.github.apple.coremltools.version" : "8.0"
+    },
+    "inputSchema" : [
+      {
+        "dataType" : "Float16",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1 × 1...1500 × 1280",
+        "shapeRange" : "[[1, 1], [1, 1500], [1280, 1280]]",
+        "formattedType" : "MultiArray (Float16 1 × 1 × 1280)",
+        "type" : "MultiArray",
+        "shape" : "[1, 1, 1280]",
+        "name" : "audio_data",
+        "shortDescription" : ""
+      }
+    ],
+    "generatedClassName" : "decoder_first",
+    "method" : "predict"
+  }
+]

large-v2/decoder_first.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ac46c34d51832dd11fbc34c772a9a35a5fb4cace68406b7044dd4ba652dca1c
+size 246506176

large-v2/decoder_second.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1afd3cf2ab05eb2b7268afb62f418b5df01b6b5d60d746bdeec2b5ad8d760f65
+size 243

large-v2/decoder_second.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5601244df54c60a16c26b761742867d06c6ef440ab8b0776ce5f6d1b4875c95
+size 487

large-v2/decoder_second.mlmodelc/metadata.json

@@ -0,0 +1,127 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16)",
+        "shortDescription" : "",
+        "shape" : "[]",
+        "name" : "logits",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.linear" : 257,
+      "Ios18.readState" : 66,
+      "Ios18.expandDims" : 33,
+      "Ios18.sub" : 1,
+      "Ios18.matmul" : 128,
+      "Ios18.gelu" : 32,
+      "Ios18.gather" : 35,
+      "Ios18.concat" : 162,
+      "Shape" : 34,
+      "Ios18.add" : 161,
+      "Ios18.sliceUpdate" : 128,
+      "Ios18.sliceByIndex" : 257,
+      "Ios18.layerNorm" : 97,
+      "Ios18.cast" : 68,
+      "Ios18.transpose" : 256,
+      "Ios18.writeState" : 64,
+      "Ios18.reshape" : 256,
+      "Ios18.softmax" : 64,
+      "Ios18.mul" : 128
+    },
+    "computePrecision" : "Mixed (Float16, Int16, Int32, UInt16)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 32 × 1 × 448 × 1280)",
+        "shortDescription" : "",
+        "shape" : "[32, 1, 448, 1280]",
+        "name" : "k_cache1",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 32 × 1 × 448 × 1280)",
+        "shortDescription" : "",
+        "shape" : "[32, 1, 448, 1280]",
+        "name" : "v_cache1",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 32 × 1 × 1500 × 1280)",
+        "shortDescription" : "",
+        "shape" : "[32, 1, 1500, 1280]",
+        "name" : "k_cache2",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 32 × 1 × 1500 × 1280)",
+        "shortDescription" : "",
+        "shape" : "[32, 1, 1500, 1280]",
+        "name" : "v_cache2",
+        "type" : "State"
+      }
+    ],
+    "availability" : {
+      "macOS" : "15.0",
+      "tvOS" : "18.0",
+      "visionOS" : "2.0",
+      "watchOS" : "11.0",
+      "iOS" : "18.0",
+      "macCatalyst" : "18.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.4.1",
+      "com.github.apple.coremltools.version" : "8.0"
+    },
+    "inputSchema" : [
+      {
+        "dataType" : "Int32",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1 × 1...448",
+        "shapeRange" : "[[1, 1], [1, 448]]",
+        "formattedType" : "MultiArray (Int32 1 × 1)",
+        "type" : "MultiArray",
+        "shape" : "[1, 1]",
+        "name" : "token_data",
+        "shortDescription" : ""
+      },
+      {
+        "dataType" : "Float16",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1 × 1...448",
+        "shapeRange" : "[[1, 1], [1, 448]]",
+        "formattedType" : "MultiArray (Float16 1 × 1)",
+        "type" : "MultiArray",
+        "shape" : "[1, 1]",
+        "name" : "offset_mask",
+        "shortDescription" : ""
+      }
+    ],
+    "generatedClassName" : "decoder_second",
+    "method" : "predict"
+  }
+]

large-v2/decoder_second.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf6ddb1f8892bfddf9f96b11c6b596934c1fe6c01839f81c06d8d2e094f19533
+size 1607634802

large-v2/encoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:abf4666de503da73cee5379b4dfa17d4a3f06bff3c0c8e310d0a0e1cf2554f87
+size 202

large-v2/encoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e2ae6d3a42a2ca10398635e0b210846dcbc24a31184c93f9302694163bcadaf
+size 196

large-v2/encoder.mlmodelc/metadata.json

@@ -0,0 +1,76 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 1500 × 1280)",
+        "shortDescription" : "",
+        "shape" : "[1, 1500, 1280]",
+        "name" : "output",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.add" : 65,
+      "Ios18.reshape" : 128,
+      "Ios18.linear" : 192,
+      "Ios18.gelu" : 34,
+      "Ios18.matmul" : 64,
+      "Ios18.transpose" : 129,
+      "Ios18.layerNorm" : 65,
+      "Ios18.conv" : 2,
+      "Ios18.cast" : 4,
+      "Ios18.softmax" : 32,
+      "Ios18.mul" : 64
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "15.0",
+      "tvOS" : "18.0",
+      "visionOS" : "2.0",
+      "watchOS" : "11.0",
+      "iOS" : "18.0",
+      "macCatalyst" : "18.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_pipeline",
+      "structure" : [
+        {
+          "name" : "MLModelType_mlProgram"
+        },
+        {
+          "name" : "MLModelType_mlProgram"
+        }
+      ]
+    },
+    "userDefinedMetadata" : {
+
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 80 × 3000)",
+        "shortDescription" : "",
+        "shape" : "[1, 80, 3000]",
+        "name" : "logmel_data",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "chunked_pipeline",
+    "method" : "predict"
+  }
+]

large-v2/encoder.mlmodelc/model0/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a8281049b2a65a3be541cfd9f949e84b8fe1c5251ce90e46da1626fed54e58a
+size 108

large-v2/encoder.mlmodelc/model0/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a30736ebb8641d231fc84aa2d3d05770adb9603bdca174d439416450827b75a
+size 200

large-v2/encoder.mlmodelc/model0/weights/0-weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0d6315a62c6344e1bf4ac88f7f7c8408cc886645c98e8989f249229fd9e9c70
+size 643945408

large-v2/encoder.mlmodelc/model1/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a8281049b2a65a3be541cfd9f949e84b8fe1c5251ce90e46da1626fed54e58a
+size 108

large-v2/encoder.mlmodelc/model1/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70195139816248a2b1fbef695f96decb60b35af6f364f84a7d2293a3d0a09e11
+size 196

large-v2/encoder.mlmodelc/model1/weights/1-weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b18deffd43b1f394f0f9d6434ef3e042c9e0424f8b590891a5cb0c21e4951163
+size 629660416

large-v2/model_dims.json

@@ -0,0 +1,12 @@
+{
+  "n_mels": 80,
+  "n_audio_ctx": 1500,
+  "n_audio_state": 1280,
+  "n_audio_head": 20,
+  "n_audio_layer": 32,
+  "n_vocab": 51865,
+  "n_text_ctx": 448,
+  "n_text_state": 1280,
+  "n_text_head": 20,
+  "n_text_layer": 32
+}

large-v3/decoder_first.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a564dfd67cfcb3c0ee8cd9f7ef9f303fbfc561e635709bd3a46c5870571079de
+size 243

large-v3/decoder_first.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6214be9e110a102836fb1fdb960a2fb564e60f5d9e3d1e25a9b7f978309480e
+size 453

large-v3/decoder_first.mlmodelc/metadata.json

@@ -0,0 +1,106 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16)",
+        "shortDescription" : "",
+        "shape" : "[]",
+        "name" : "dummy",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.writeState" : 66,
+      "Shape" : 64,
+      "Ios18.linear" : 64,
+      "Identity" : 1,
+      "Ios18.gather" : 64,
+      "Ios18.concat" : 64,
+      "Ios18.sliceUpdate" : 66,
+      "Ios18.cast" : 128,
+      "Ios18.expandDims" : 64,
+      "Ios18.readState" : 66
+    },
+    "computePrecision" : "Mixed (Float16, Int16, Int32, UInt16)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 32 × 1 × 448 × 1280)",
+        "shortDescription" : "",
+        "shape" : "[32, 1, 448, 1280]",
+        "name" : "k_cache1",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 32 × 1 × 448 × 1280)",
+        "shortDescription" : "",
+        "shape" : "[32, 1, 448, 1280]",
+        "name" : "v_cache1",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 32 × 1 × 1500 × 1280)",
+        "shortDescription" : "",
+        "shape" : "[32, 1, 1500, 1280]",
+        "name" : "k_cache2",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 32 × 1 × 1500 × 1280)",
+        "shortDescription" : "",
+        "shape" : "[32, 1, 1500, 1280]",
+        "name" : "v_cache2",
+        "type" : "State"
+      }
+    ],
+    "availability" : {
+      "macOS" : "15.0",
+      "tvOS" : "18.0",
+      "visionOS" : "2.0",
+      "watchOS" : "11.0",
+      "iOS" : "18.0",
+      "macCatalyst" : "18.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.4.1",
+      "com.github.apple.coremltools.version" : "8.0"
+    },
+    "inputSchema" : [
+      {
+        "dataType" : "Float16",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1 × 1...1500 × 1280",
+        "shapeRange" : "[[1, 1], [1, 1500], [1280, 1280]]",
+        "formattedType" : "MultiArray (Float16 1 × 1 × 1280)",
+        "type" : "MultiArray",
+        "shape" : "[1, 1, 1280]",
+        "name" : "audio_data",
+        "shortDescription" : ""
+      }
+    ],
+    "generatedClassName" : "decoder_first",
+    "method" : "predict"
+  }
+]

large-v3/decoder_first.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:346ed969f2a1ddb144c4add194c7b2a9a7d7b4a2e536d1e4a2afbfe5a4f62818
+size 246506176