| 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>(state<tensor<fp16, [4, 1, 448, 384]>> k_cache1, state<tensor<fp16, [4, 1, 1500, 384]>> k_cache2, tensor<fp16, [1, ?]> offset_mask, tensor<int32, [1, ?]> token_data, state<tensor<fp16, [4, 1, 448, 384]>> v_cache1, state<tensor<fp16, [4, 1, 1500, 384]>> v_cache2) [FlexibleShapeInformation = tuple<tuple<string, dict<string, tensor<int32, [?]>>>, tuple<string, dict<string, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"offset_mask", [1, 1]}, {"token_data", [1, 1]}}), ("RangeDims", {{"offset_mask", [[1, 1], [1, 448]]}, {"token_data", [[1, 1], [1, 448]]}})))] { | |
| tensor<int32, [2]> var_22_shape_cast_fp16 = shape(x = offset_mask)[name = string("op_22_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_22_shape_cast_fp16_to_int16_dtype_0 = const()[name = string("op_22_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, [2]> var_22_shape_cast_fp16_to_int16 = cast(dtype = var_22_shape_cast_fp16_to_int16_dtype_0, x = var_22_shape_cast_fp16)[name = string("cast_58")]; | |
| 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_22_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, [2]> var_26_shape = shape(x = token_data)[name = string("op_26_shape")]; | |
| 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_26_shape_to_uint16_dtype_0 = const()[name = string("op_26_shape_to_uint16_dtype_0"), val = string("uint16")]; | |
| uint16 select_1_to_uint16 = const()[name = string("select_1_to_uint16"), val = uint16(1)]; | |
| tensor<uint16, [2]> var_26_shape_to_uint16 = cast(dtype = var_26_shape_to_uint16_dtype_0, x = var_26_shape)[name = string("cast_56")]; | |
| 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_26_shape_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")]; | |
| 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_55")]; | |
| 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_57")]; | |
| int32 offset = sub(x = gather_0_cast_uint16_to_int32, y = gather_1_cast_uint16_to_int32)[name = string("offset")]; | |
| int32 var_42_axis_0 = const()[name = string("op_42_axis_0"), val = int32(0)]; | |
| int32 var_42_batch_dims_0 = const()[name = string("op_42_batch_dims_0"), val = int32(0)]; | |
| bool var_42_validate_indices_0 = const()[name = string("op_42_validate_indices_0"), val = bool(false)]; | |
| tensor<fp16, [51865, 384]> token_embedding_weight_to_fp16 = const()[name = string("token_embedding_weight_to_fp16"), val = tensor<fp16, [51865, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; | |
| tensor<fp16, [1, ?, 384]> var_42_cast_fp16 = gather(axis = var_42_axis_0, batch_dims = var_42_batch_dims_0, indices = token_data, validate_indices = var_42_validate_indices_0, x = token_embedding_weight_to_fp16)[name = string("op_42_cast_fp16")]; | |
| int32 concat_0_values1_0 = const()[name = string("concat_0_values1_0"), val = int32(0)]; | |
| int32 concat_0_axis_0 = const()[name = string("concat_0_axis_0"), val = int32(0)]; | |
| bool concat_0_interleave_0 = const()[name = string("concat_0_interleave_0"), val = bool(false)]; | |
| tensor<int32, [2]> concat_0 = concat(axis = concat_0_axis_0, interleave = concat_0_interleave_0, values = (offset, concat_0_values1_0))[name = string("concat_0")]; | |
| int32 concat_1_values1_0 = const()[name = string("concat_1_values1_0"), val = int32(384)]; | |
| int32 concat_1_axis_0 = const()[name = string("concat_1_axis_0"), val = int32(0)]; | |
| bool concat_1_interleave_0 = const()[name = string("concat_1_interleave_0"), val = bool(false)]; | |
| tensor<int32, [2]> concat_1 = concat(axis = concat_1_axis_0, interleave = concat_1_interleave_0, values = (gather_0_cast_uint16_to_int32, concat_1_values1_0))[name = string("concat_1")]; | |
| tensor<bool, [2]> var_45_end_mask_0 = const()[name = string("op_45_end_mask_0"), val = tensor<bool, [2]>([false, true])]; | |
| tensor<fp16, [448, 384]> positional_embedding_to_fp16 = const()[name = string("positional_embedding_to_fp16"), val = tensor<fp16, [448, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39832448)))]; | |
| tensor<fp16, [?, ?]> var_45_cast_fp16 = slice_by_index(begin = concat_0, end = concat_1, end_mask = var_45_end_mask_0, x = positional_embedding_to_fp16)[name = string("op_45_cast_fp16")]; | |
| tensor<fp16, [1, ?, 384]> x_3_cast_fp16 = add(x = var_42_cast_fp16, y = var_45_cast_fp16)[name = string("x_3_cast_fp16")]; | |
| tensor<fp16, [4, 1, 448, 384]> read_state_0 = read_state(input = k_cache1)[name = string("read_state_0")]; | |
| tensor<int32, [4]> k_cache_1_begin_0 = const()[name = string("k_cache_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [4]> k_cache_1_end_0 = const()[name = string("k_cache_1_end_0"), val = tensor<int32, [4]>([1, 1, 448, 384])]; | |
| tensor<bool, [4]> k_cache_1_end_mask_0 = const()[name = string("k_cache_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> k_cache_1_squeeze_mask_0 = const()[name = string("k_cache_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [1, 448, 384]> k_cache_1_cast_fp16 = slice_by_index(begin = k_cache_1_begin_0, end = k_cache_1_end_0, end_mask = k_cache_1_end_mask_0, squeeze_mask = k_cache_1_squeeze_mask_0, x = read_state_0)[name = string("k_cache_1_cast_fp16")]; | |
| tensor<fp16, [4, 1, 448, 384]> read_state_1 = read_state(input = v_cache1)[name = string("read_state_1")]; | |
| tensor<int32, [4]> v_cache_1_begin_0 = const()[name = string("v_cache_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [4]> v_cache_1_end_0 = const()[name = string("v_cache_1_end_0"), val = tensor<int32, [4]>([1, 1, 448, 384])]; | |
| tensor<bool, [4]> v_cache_1_end_mask_0 = const()[name = string("v_cache_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> v_cache_1_squeeze_mask_0 = const()[name = string("v_cache_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [1, 448, 384]> v_cache_1_cast_fp16 = slice_by_index(begin = v_cache_1_begin_0, end = v_cache_1_end_0, end_mask = v_cache_1_end_mask_0, squeeze_mask = v_cache_1_squeeze_mask_0, x = read_state_1)[name = string("v_cache_1_cast_fp16")]; | |
| tensor<fp16, [4, 1, 1500, 384]> read_state_2 = read_state(input = k_cache2)[name = string("read_state_2")]; | |
| tensor<int32, [4]> k_cache_3_begin_0 = const()[name = string("k_cache_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [4]> k_cache_3_end_0 = const()[name = string("k_cache_3_end_0"), val = tensor<int32, [4]>([1, 1, 1500, 384])]; | |
| tensor<bool, [4]> k_cache_3_end_mask_0 = const()[name = string("k_cache_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> k_cache_3_squeeze_mask_0 = const()[name = string("k_cache_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [1, 1500, 384]> k_cache_3_cast_fp16 = slice_by_index(begin = k_cache_3_begin_0, end = k_cache_3_end_0, end_mask = k_cache_3_end_mask_0, squeeze_mask = k_cache_3_squeeze_mask_0, x = read_state_2)[name = string("k_cache_3_cast_fp16")]; | |
| tensor<fp16, [4, 1, 1500, 384]> read_state_3 = read_state(input = v_cache2)[name = string("read_state_3")]; | |
| tensor<int32, [4]> v_cache_3_begin_0 = const()[name = string("v_cache_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| tensor<int32, [4]> v_cache_3_end_0 = const()[name = string("v_cache_3_end_0"), val = tensor<int32, [4]>([1, 1, 1500, 384])]; | |
| tensor<bool, [4]> v_cache_3_end_mask_0 = const()[name = string("v_cache_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> v_cache_3_squeeze_mask_0 = const()[name = string("v_cache_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [1, 1500, 384]> v_cache_3_cast_fp16 = slice_by_index(begin = v_cache_3_begin_0, end = v_cache_3_end_0, end_mask = v_cache_3_end_mask_0, squeeze_mask = v_cache_3_squeeze_mask_0, x = read_state_3)[name = string("v_cache_3_cast_fp16")]; | |
| int32 var_67 = const()[name = string("op_67"), val = int32(-1)]; | |
| tensor<int32, [1]> var_85_axes_0 = const()[name = string("op_85_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [384]> blocks_0_attn_ln_weight_to_fp16 = const()[name = string("blocks_0_attn_ln_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40176576)))]; | |
| tensor<fp16, [384]> blocks_0_attn_ln_bias_to_fp16 = const()[name = string("blocks_0_attn_ln_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40177408)))]; | |
| fp16 var_73_to_fp16 = const()[name = string("op_73_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [1, ?, 384]> var_85_cast_fp16 = layer_norm(axes = var_85_axes_0, beta = blocks_0_attn_ln_bias_to_fp16, epsilon = var_73_to_fp16, gamma = blocks_0_attn_ln_weight_to_fp16, x = x_3_cast_fp16)[name = string("op_85_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_96_to_fp16 = const()[name = string("op_96_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40178240)))]; | |
| tensor<fp16, [384]> var_97_to_fp16 = const()[name = string("op_97_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40473216)))]; | |
| tensor<fp16, [1, ?, 384]> linear_0_cast_fp16 = linear(bias = var_97_to_fp16, weight = var_96_to_fp16, x = var_85_cast_fp16)[name = string("linear_0_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_100_to_fp16 = const()[name = string("op_100_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40474048)))]; | |
| tensor<fp16, [384]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40769024)))]; | |
| tensor<fp16, [1, ?, 384]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_100_to_fp16, x = var_85_cast_fp16)[name = string("linear_1_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_104_to_fp16 = const()[name = string("op_104_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40769856)))]; | |
| tensor<fp16, [384]> var_105_to_fp16 = const()[name = string("op_105_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41064832)))]; | |
| tensor<fp16, [1, ?, 384]> linear_2_cast_fp16 = linear(bias = var_105_to_fp16, weight = var_104_to_fp16, x = var_85_cast_fp16)[name = string("linear_2_cast_fp16")]; | |
| tensor<int32, [3]> var_107_shape_cast_fp16 = shape(x = linear_0_cast_fp16)[name = string("op_107_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_107_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_107_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_107_shape_cast_fp16_to_uint16 = cast(dtype = var_107_shape_cast_fp16_to_uint16_dtype_0, x = var_107_shape_cast_fp16)[name = string("cast_54")]; | |
| 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_107_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")]; | |
| 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_53")]; | |
| int32 end_step_3 = add(x = offset, y = gather_2_cast_uint16_to_int32)[name = string("end_step_3")]; | |
| tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, [1]> expand_dims_1_axes_0 = const()[name = string("expand_dims_1_axes_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, [1]> expand_dims_1 = expand_dims(axes = expand_dims_1_axes_0, x = offset)[name = string("expand_dims_1")]; | |
| tensor<int32, [1]> expand_dims_2 = const()[name = string("expand_dims_2"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, [1]> expand_dims_3_axes_0 = const()[name = string("expand_dims_3_axes_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, [1]> expand_dims_3 = expand_dims(axes = expand_dims_3_axes_0, x = end_step_3)[name = string("expand_dims_3")]; | |
| tensor<int32, [1]> concat_4_values0_0 = const()[name = string("concat_4_values0_0"), val = tensor<int32, [1]>([0])]; | |
| int32 concat_4_axis_0 = const()[name = string("concat_4_axis_0"), val = int32(0)]; | |
| bool concat_4_interleave_0 = const()[name = string("concat_4_interleave_0"), val = bool(false)]; | |
| tensor<int32, [4]> concat_4 = concat(axis = concat_4_axis_0, interleave = concat_4_interleave_0, values = (concat_4_values0_0, expand_dims_0, expand_dims_1, expand_dims_2))[name = string("concat_4")]; | |
| tensor<int32, [1]> concat_5_values0_0 = const()[name = string("concat_5_values0_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, [1]> concat_5_values1_0 = const()[name = string("concat_5_values1_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, [1]> concat_5_values3_0 = const()[name = string("concat_5_values3_0"), val = tensor<int32, [1]>([0])]; | |
| int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)]; | |
| bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)]; | |
| tensor<int32, [4]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (concat_5_values0_0, concat_5_values1_0, expand_dims_3, concat_5_values3_0))[name = string("concat_5")]; | |
| 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]>([false, true, false, 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]>([true, false, false, false])]; | |
| tensor<fp16, [4, 1, 448, 384]> k_cache1_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_4, begin_mask = k_cache1_internal_tensor_assign_1_begin_mask_0, end = concat_5, 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 = linear_1_cast_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_8_write_state")]; | |
| tensor<fp16, [4, 1, 448, 384]> coreml_update_state_8 = read_state(input = k_cache1)[name = string("coreml_update_state_8")]; | |
| 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]>([false, true, false, 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]>([true, false, false, false])]; | |
| tensor<fp16, [4, 1, 448, 384]> v_cache1_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_4, begin_mask = v_cache1_internal_tensor_assign_1_begin_mask_0, end = concat_5, 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 = linear_2_cast_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_9_write_state")]; | |
| tensor<fp16, [4, 1, 448, 384]> coreml_update_state_9 = read_state(input = v_cache1)[name = string("coreml_update_state_9")]; | |
| int32 concat_10_values0_0 = const()[name = string("concat_10_values0_0"), val = int32(1)]; | |
| int32 concat_10_values2_0 = const()[name = string("concat_10_values2_0"), val = int32(384)]; | |
| int32 concat_10_axis_0 = const()[name = string("concat_10_axis_0"), val = int32(0)]; | |
| bool concat_10_interleave_0 = const()[name = string("concat_10_interleave_0"), val = bool(false)]; | |
| tensor<int32, [3]> concat_10 = concat(axis = concat_10_axis_0, interleave = concat_10_interleave_0, values = (concat_10_values0_0, end_step_3, concat_10_values2_0))[name = string("concat_10")]; | |
| tensor<int32, [3]> var_123_begin_0 = const()[name = string("op_123_begin_0"), val = tensor<int32, [3]>([0, 0, 0])]; | |
| tensor<bool, [3]> var_123_end_mask_0 = const()[name = string("op_123_end_mask_0"), val = tensor<bool, [3]>([true, false, true])]; | |
| tensor<fp16, [1, ?, 384]> var_123_cast_fp16 = slice_by_index(begin = var_123_begin_0, end = concat_10, end_mask = var_123_end_mask_0, x = k_cache_1_cast_fp16)[name = string("op_123_cast_fp16")]; | |
| tensor<int32, [3]> var_126_begin_0 = const()[name = string("op_126_begin_0"), val = tensor<int32, [3]>([0, 0, 0])]; | |
| tensor<bool, [3]> var_126_end_mask_0 = const()[name = string("op_126_end_mask_0"), val = tensor<bool, [3]>([true, false, true])]; | |
| tensor<fp16, [1, ?, 384]> var_126_cast_fp16 = slice_by_index(begin = var_126_begin_0, end = concat_10, end_mask = var_126_end_mask_0, x = v_cache_1_cast_fp16)[name = string("op_126_cast_fp16")]; | |
| tensor<int32, [4]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [4]>([1, -1, 6, 64])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_136_cast_fp16 = reshape(shape = concat_12x, x = linear_0_cast_fp16)[name = string("op_136_cast_fp16")]; | |
| tensor<fp16, [1, 1, 1, 1]> const_20_to_fp16 = const()[name = string("const_20_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])]; | |
| tensor<fp16, [1, ?, 6, 64]> q_3_cast_fp16 = mul(x = var_136_cast_fp16, y = const_20_to_fp16)[name = string("q_3_cast_fp16")]; | |
| tensor<int32, [4]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [4]>([1, -1, 6, 64])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_143_cast_fp16 = reshape(shape = concat_13x, x = var_123_cast_fp16)[name = string("op_143_cast_fp16")]; | |
| tensor<fp16, [1, 1, 1, 1]> const_21_to_fp16 = const()[name = string("const_21_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])]; | |
| tensor<fp16, [1, ?, 6, 64]> k_5_cast_fp16 = mul(x = var_143_cast_fp16, y = const_21_to_fp16)[name = string("k_5_cast_fp16")]; | |
| tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([1, -1, 6, 64])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_150_cast_fp16 = reshape(shape = concat_14x, x = var_126_cast_fp16)[name = string("op_150_cast_fp16")]; | |
| tensor<int32, [4]> var_151 = const()[name = string("op_151"), 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_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; | |
| tensor<int32, [4]> transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; | |
| tensor<fp16, [1, 6, 64, ?]> transpose_34 = transpose(perm = transpose_34_perm_0, x = k_5_cast_fp16)[name = string("transpose_78")]; | |
| tensor<fp16, [1, 6, ?, 64]> transpose_33 = transpose(perm = transpose_33_perm_0, x = q_3_cast_fp16)[name = string("transpose_79")]; | |
| tensor<fp16, [1, 6, ?, ?]> qk_1_cast_fp16 = matmul(transpose_x = qk_1_transpose_x_0, transpose_y = qk_1_transpose_y_0, x = transpose_33, y = transpose_34)[name = string("qk_1_cast_fp16")]; | |
| int32 concat_15_values1_0 = const()[name = string("concat_15_values1_0"), val = int32(448)]; | |
| 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, [2]> concat_15 = concat(axis = concat_15_axis_0, interleave = concat_15_interleave_0, values = (gather_2_cast_uint16_to_int32, concat_15_values1_0))[name = string("concat_15")]; | |
| tensor<int32, [2]> var_154_begin_0 = const()[name = string("op_154_begin_0"), val = tensor<int32, [2]>([0, 0])]; | |
| tensor<bool, [2]> var_154_end_mask_0 = const()[name = string("op_154_end_mask_0"), val = tensor<bool, [2]>([false, true])]; | |
| tensor<fp16, [448, 448]> mask_to_fp16 = const()[name = string("mask_to_fp16"), val = tensor<fp16, [448, 448]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41065664)))]; | |
| tensor<fp16, [?, 448]> var_154_cast_fp16 = slice_by_index(begin = var_154_begin_0, end = concat_15, end_mask = var_154_end_mask_0, x = mask_to_fp16)[name = string("op_154_cast_fp16")]; | |
| int32 concat_16_values0_0 = const()[name = string("concat_16_values0_0"), val = int32(0)]; | |
| int32 concat_16_axis_0 = const()[name = string("concat_16_axis_0"), val = int32(0)]; | |
| bool concat_16_interleave_0 = const()[name = string("concat_16_interleave_0"), val = bool(false)]; | |
| tensor<int32, [2]> concat_16 = concat(axis = concat_16_axis_0, interleave = concat_16_interleave_0, values = (concat_16_values0_0, gather_2_cast_uint16_to_int32))[name = string("concat_16")]; | |
| tensor<int32, [2]> var_155_begin_0 = const()[name = string("op_155_begin_0"), val = tensor<int32, [2]>([0, 0])]; | |
| tensor<bool, [2]> var_155_end_mask_0 = const()[name = string("op_155_end_mask_0"), val = tensor<bool, [2]>([true, false])]; | |
| tensor<fp16, [?, ?]> var_155_cast_fp16 = slice_by_index(begin = var_155_begin_0, end = concat_16, end_mask = var_155_end_mask_0, x = var_154_cast_fp16)[name = string("op_155_cast_fp16")]; | |
| tensor<fp16, [1, 6, ?, ?]> qk_3_cast_fp16 = add(x = qk_1_cast_fp16, y = var_155_cast_fp16)[name = string("qk_3_cast_fp16")]; | |
| tensor<fp16, [1, 6, ?, ?]> var_158_cast_fp16 = softmax(axis = var_67, x = qk_3_cast_fp16)[name = string("op_158_cast_fp16")]; | |
| bool var_160_transpose_x_0 = const()[name = string("op_160_transpose_x_0"), val = bool(false)]; | |
| bool var_160_transpose_y_0 = const()[name = string("op_160_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [1, 6, ?, 64]> v_5_cast_fp16 = transpose(perm = var_151, x = var_150_cast_fp16)[name = string("transpose_80")]; | |
| tensor<fp16, [1, 6, ?, 64]> var_160_cast_fp16 = matmul(transpose_x = var_160_transpose_x_0, transpose_y = var_160_transpose_y_0, x = var_158_cast_fp16, y = v_5_cast_fp16)[name = string("op_160_cast_fp16")]; | |
| tensor<int32, [4]> var_161 = const()[name = string("op_161"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> concat_17x = const()[name = string("concat_17x"), val = tensor<int32, [3]>([1, -1, 384])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_162_cast_fp16 = transpose(perm = var_161, x = var_160_cast_fp16)[name = string("transpose_77")]; | |
| tensor<fp16, [1, ?, 384]> x_7_cast_fp16 = reshape(shape = concat_17x, x = var_162_cast_fp16)[name = string("x_7_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_166_to_fp16 = const()[name = string("op_166_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41467136)))]; | |
| tensor<fp16, [384]> var_167_to_fp16 = const()[name = string("op_167_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41762112)))]; | |
| tensor<fp16, [1, ?, 384]> linear_3_cast_fp16 = linear(bias = var_167_to_fp16, weight = var_166_to_fp16, x = x_7_cast_fp16)[name = string("linear_3_cast_fp16")]; | |
| tensor<fp16, [1, ?, 384]> x_9_cast_fp16 = add(x = x_3_cast_fp16, y = linear_3_cast_fp16)[name = string("x_9_cast_fp16")]; | |
| tensor<int32, [1]> var_174_axes_0 = const()[name = string("op_174_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [384]> blocks_0_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_0_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41762944)))]; | |
| tensor<fp16, [384]> blocks_0_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_0_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41763776)))]; | |
| tensor<fp16, [1, ?, 384]> var_174_cast_fp16 = layer_norm(axes = var_174_axes_0, beta = blocks_0_cross_attn_ln_bias_to_fp16, epsilon = var_73_to_fp16, gamma = blocks_0_cross_attn_ln_weight_to_fp16, x = x_9_cast_fp16)[name = string("op_174_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_183_to_fp16 = const()[name = string("op_183_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41764608)))]; | |
| tensor<fp16, [384]> var_184_to_fp16 = const()[name = string("op_184_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42059584)))]; | |
| tensor<fp16, [1, ?, 384]> linear_4_cast_fp16 = linear(bias = var_184_to_fp16, weight = var_183_to_fp16, x = var_174_cast_fp16)[name = string("linear_4_cast_fp16")]; | |
| tensor<int32, [3]> concat_18 = const()[name = string("concat_18"), val = tensor<int32, [3]>([0, 0, 0])]; | |
| tensor<int32, [3]> concat_19 = const()[name = string("concat_19"), val = tensor<int32, [3]>([0, 1500, 0])]; | |
| tensor<int32, [3]> k_7_internal_tensor_assign_1_stride_0 = const()[name = string("k_7_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])]; | |
| tensor<bool, [3]> k_7_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_7_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])]; | |
| tensor<bool, [3]> k_7_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_7_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])]; | |
| tensor<bool, [3]> k_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])]; | |
| tensor<fp16, [1, 1500, 384]> k_7_to_fp16 = const()[name = string("k_7_to_fp16"), val = tensor<fp16, [1, 1500, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42060416)))]; | |
| tensor<fp16, [1, 1500, 384]> k_7_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_18, begin_mask = k_7_internal_tensor_assign_1_begin_mask_0, end = concat_19, end_mask = k_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_7_internal_tensor_assign_1_squeeze_mask_0, stride = k_7_internal_tensor_assign_1_stride_0, update = k_cache_3_cast_fp16, x = k_7_to_fp16)[name = string("k_7_internal_tensor_assign_1_cast_fp16")]; | |
| tensor<int32, [3]> concat_20 = const()[name = string("concat_20"), val = tensor<int32, [3]>([0, 0, 0])]; | |
| tensor<int32, [3]> concat_21 = const()[name = string("concat_21"), val = tensor<int32, [3]>([0, 1500, 0])]; | |
| tensor<int32, [3]> v_7_internal_tensor_assign_1_stride_0 = const()[name = string("v_7_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])]; | |
| tensor<bool, [3]> v_7_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_7_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])]; | |
| tensor<bool, [3]> v_7_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_7_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])]; | |
| tensor<bool, [3]> v_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])]; | |
| tensor<fp16, [1, 1500, 384]> v_7_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_20, begin_mask = v_7_internal_tensor_assign_1_begin_mask_0, end = concat_21, end_mask = v_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_7_internal_tensor_assign_1_squeeze_mask_0, stride = v_7_internal_tensor_assign_1_stride_0, update = v_cache_3_cast_fp16, x = k_7_to_fp16)[name = string("v_7_internal_tensor_assign_1_cast_fp16")]; | |
| tensor<int32, [4]> concat_22x = const()[name = string("concat_22x"), val = tensor<int32, [4]>([1, -1, 6, 64])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_204_cast_fp16 = reshape(shape = concat_22x, x = linear_4_cast_fp16)[name = string("op_204_cast_fp16")]; | |
| tensor<fp16, [1, 1, 1, 1]> const_22_to_fp16 = const()[name = string("const_22_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])]; | |
| tensor<fp16, [1, ?, 6, 64]> q_7_cast_fp16 = mul(x = var_204_cast_fp16, y = const_22_to_fp16)[name = string("q_7_cast_fp16")]; | |
| tensor<int32, [4]> var_210 = const()[name = string("op_210"), val = tensor<int32, [4]>([1, 1500, 6, -1])]; | |
| tensor<fp16, [1, 1500, 6, 64]> var_211_cast_fp16 = reshape(shape = var_210, x = k_7_internal_tensor_assign_1_cast_fp16)[name = string("op_211_cast_fp16")]; | |
| tensor<fp16, [1, 1, 1, 1]> const_23_to_fp16 = const()[name = string("const_23_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])]; | |
| tensor<fp16, [1, 1500, 6, 64]> k_9_cast_fp16 = mul(x = var_211_cast_fp16, y = const_23_to_fp16)[name = string("k_9_cast_fp16")]; | |
| tensor<int32, [4]> var_217 = const()[name = string("op_217"), val = tensor<int32, [4]>([1, 1500, 6, -1])]; | |
| tensor<fp16, [1, 1500, 6, 64]> var_218_cast_fp16 = reshape(shape = var_217, x = v_7_internal_tensor_assign_1_cast_fp16)[name = string("op_218_cast_fp16")]; | |
| tensor<int32, [4]> var_219 = const()[name = string("op_219"), 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_35_perm_0 = const()[name = string("transpose_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; | |
| tensor<int32, [4]> transpose_36_perm_0 = const()[name = string("transpose_36_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; | |
| tensor<fp16, [1, 6, 64, 1500]> transpose_36 = transpose(perm = transpose_36_perm_0, x = k_9_cast_fp16)[name = string("transpose_74")]; | |
| tensor<fp16, [1, 6, ?, 64]> transpose_35 = transpose(perm = transpose_35_perm_0, x = q_7_cast_fp16)[name = string("transpose_75")]; | |
| tensor<fp16, [1, 6, ?, 1500]> qk_5_cast_fp16 = matmul(transpose_x = qk_5_transpose_x_0, transpose_y = qk_5_transpose_y_0, x = transpose_35, y = transpose_36)[name = string("qk_5_cast_fp16")]; | |
| tensor<fp16, [1, 6, ?, 1500]> var_223_cast_fp16 = softmax(axis = var_67, x = qk_5_cast_fp16)[name = string("op_223_cast_fp16")]; | |
| bool var_225_transpose_x_0 = const()[name = string("op_225_transpose_x_0"), val = bool(false)]; | |
| bool var_225_transpose_y_0 = const()[name = string("op_225_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [1, 6, 1500, 64]> v_9_cast_fp16 = transpose(perm = var_219, x = var_218_cast_fp16)[name = string("transpose_76")]; | |
| tensor<fp16, [1, 6, ?, 64]> var_225_cast_fp16 = matmul(transpose_x = var_225_transpose_x_0, transpose_y = var_225_transpose_y_0, x = var_223_cast_fp16, y = v_9_cast_fp16)[name = string("op_225_cast_fp16")]; | |
| tensor<int32, [4]> var_226 = const()[name = string("op_226"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> concat_23x = const()[name = string("concat_23x"), val = tensor<int32, [3]>([1, -1, 384])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_227_cast_fp16 = transpose(perm = var_226, x = var_225_cast_fp16)[name = string("transpose_73")]; | |
| tensor<fp16, [1, ?, 384]> x_13_cast_fp16 = reshape(shape = concat_23x, x = var_227_cast_fp16)[name = string("x_13_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_231_to_fp16 = const()[name = string("op_231_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43212480)))]; | |
| tensor<fp16, [384]> var_232_to_fp16 = const()[name = string("op_232_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43507456)))]; | |
| tensor<fp16, [1, ?, 384]> linear_5_cast_fp16 = linear(bias = var_232_to_fp16, weight = var_231_to_fp16, x = x_13_cast_fp16)[name = string("linear_5_cast_fp16")]; | |
| tensor<fp16, [1, ?, 384]> x_15_cast_fp16 = add(x = x_9_cast_fp16, y = linear_5_cast_fp16)[name = string("x_15_cast_fp16")]; | |
| tensor<int32, [1]> var_239_axes_0 = const()[name = string("op_239_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [384]> blocks_0_mlp_ln_weight_to_fp16 = const()[name = string("blocks_0_mlp_ln_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43508288)))]; | |
| tensor<fp16, [384]> blocks_0_mlp_ln_bias_to_fp16 = const()[name = string("blocks_0_mlp_ln_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43509120)))]; | |
| tensor<fp16, [1, ?, 384]> var_239_cast_fp16 = layer_norm(axes = var_239_axes_0, beta = blocks_0_mlp_ln_bias_to_fp16, epsilon = var_73_to_fp16, gamma = blocks_0_mlp_ln_weight_to_fp16, x = x_15_cast_fp16)[name = string("op_239_cast_fp16")]; | |
| tensor<fp16, [1536, 384]> var_248_to_fp16 = const()[name = string("op_248_to_fp16"), val = tensor<fp16, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43509952)))]; | |
| tensor<fp16, [1536]> var_249_to_fp16 = const()[name = string("op_249_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44689664)))]; | |
| tensor<fp16, [1, ?, 1536]> linear_6_cast_fp16 = linear(bias = var_249_to_fp16, weight = var_248_to_fp16, x = var_239_cast_fp16)[name = string("linear_6_cast_fp16")]; | |
| string x_19_mode_0 = const()[name = string("x_19_mode_0"), val = string("EXACT")]; | |
| tensor<fp16, [1, ?, 1536]> x_19_cast_fp16 = gelu(mode = x_19_mode_0, x = linear_6_cast_fp16)[name = string("x_19_cast_fp16")]; | |
| tensor<fp16, [384, 1536]> var_254_to_fp16 = const()[name = string("op_254_to_fp16"), val = tensor<fp16, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44692800)))]; | |
| tensor<fp16, [384]> var_255_to_fp16 = const()[name = string("op_255_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45872512)))]; | |
| tensor<fp16, [1, ?, 384]> linear_7_cast_fp16 = linear(bias = var_255_to_fp16, weight = var_254_to_fp16, x = x_19_cast_fp16)[name = string("linear_7_cast_fp16")]; | |
| tensor<fp16, [1, ?, 384]> x_21_cast_fp16 = add(x = x_15_cast_fp16, y = linear_7_cast_fp16)[name = string("x_21_cast_fp16")]; | |
| tensor<int32, [4]> k_cache_5_begin_0 = const()[name = string("k_cache_5_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])]; | |
| tensor<int32, [4]> k_cache_5_end_0 = const()[name = string("k_cache_5_end_0"), val = tensor<int32, [4]>([2, 1, 448, 384])]; | |
| tensor<bool, [4]> k_cache_5_end_mask_0 = const()[name = string("k_cache_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> k_cache_5_squeeze_mask_0 = const()[name = string("k_cache_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [1, 448, 384]> k_cache_5_cast_fp16 = slice_by_index(begin = k_cache_5_begin_0, end = k_cache_5_end_0, end_mask = k_cache_5_end_mask_0, squeeze_mask = k_cache_5_squeeze_mask_0, x = coreml_update_state_8)[name = string("k_cache_5_cast_fp16")]; | |
| tensor<int32, [4]> v_cache_5_begin_0 = const()[name = string("v_cache_5_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])]; | |
| tensor<int32, [4]> v_cache_5_end_0 = const()[name = string("v_cache_5_end_0"), val = tensor<int32, [4]>([2, 1, 448, 384])]; | |
| tensor<bool, [4]> v_cache_5_end_mask_0 = const()[name = string("v_cache_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> v_cache_5_squeeze_mask_0 = const()[name = string("v_cache_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [1, 448, 384]> v_cache_5_cast_fp16 = slice_by_index(begin = v_cache_5_begin_0, end = v_cache_5_end_0, end_mask = v_cache_5_end_mask_0, squeeze_mask = v_cache_5_squeeze_mask_0, x = coreml_update_state_9)[name = string("v_cache_5_cast_fp16")]; | |
| tensor<int32, [4]> k_cache_7_begin_0 = const()[name = string("k_cache_7_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])]; | |
| tensor<int32, [4]> k_cache_7_end_0 = const()[name = string("k_cache_7_end_0"), val = tensor<int32, [4]>([2, 1, 1500, 384])]; | |
| tensor<bool, [4]> k_cache_7_end_mask_0 = const()[name = string("k_cache_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> k_cache_7_squeeze_mask_0 = const()[name = string("k_cache_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [1, 1500, 384]> k_cache_7_cast_fp16 = slice_by_index(begin = k_cache_7_begin_0, end = k_cache_7_end_0, end_mask = k_cache_7_end_mask_0, squeeze_mask = k_cache_7_squeeze_mask_0, x = read_state_2)[name = string("k_cache_7_cast_fp16")]; | |
| tensor<int32, [4]> v_cache_7_begin_0 = const()[name = string("v_cache_7_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])]; | |
| tensor<int32, [4]> v_cache_7_end_0 = const()[name = string("v_cache_7_end_0"), val = tensor<int32, [4]>([2, 1, 1500, 384])]; | |
| tensor<bool, [4]> v_cache_7_end_mask_0 = const()[name = string("v_cache_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> v_cache_7_squeeze_mask_0 = const()[name = string("v_cache_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [1, 1500, 384]> v_cache_7_cast_fp16 = slice_by_index(begin = v_cache_7_begin_0, end = v_cache_7_end_0, end_mask = v_cache_7_end_mask_0, squeeze_mask = v_cache_7_squeeze_mask_0, x = read_state_3)[name = string("v_cache_7_cast_fp16")]; | |
| int32 var_277 = const()[name = string("op_277"), val = int32(-1)]; | |
| tensor<int32, [1]> var_295_axes_0 = const()[name = string("op_295_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [384]> blocks_1_attn_ln_weight_to_fp16 = const()[name = string("blocks_1_attn_ln_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45873344)))]; | |
| tensor<fp16, [384]> blocks_1_attn_ln_bias_to_fp16 = const()[name = string("blocks_1_attn_ln_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45874176)))]; | |
| fp16 var_283_to_fp16 = const()[name = string("op_283_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [1, ?, 384]> var_295_cast_fp16 = layer_norm(axes = var_295_axes_0, beta = blocks_1_attn_ln_bias_to_fp16, epsilon = var_283_to_fp16, gamma = blocks_1_attn_ln_weight_to_fp16, x = x_21_cast_fp16)[name = string("op_295_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_306_to_fp16 = const()[name = string("op_306_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45875008)))]; | |
| tensor<fp16, [384]> var_307_to_fp16 = const()[name = string("op_307_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46169984)))]; | |
| tensor<fp16, [1, ?, 384]> linear_8_cast_fp16 = linear(bias = var_307_to_fp16, weight = var_306_to_fp16, x = var_295_cast_fp16)[name = string("linear_8_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_310_to_fp16 = const()[name = string("op_310_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46170816)))]; | |
| tensor<fp16, [1, ?, 384]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_310_to_fp16, x = var_295_cast_fp16)[name = string("linear_9_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_314_to_fp16 = const()[name = string("op_314_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46465792)))]; | |
| tensor<fp16, [384]> var_315_to_fp16 = const()[name = string("op_315_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46760768)))]; | |
| tensor<fp16, [1, ?, 384]> linear_10_cast_fp16 = linear(bias = var_315_to_fp16, weight = var_314_to_fp16, x = var_295_cast_fp16)[name = string("linear_10_cast_fp16")]; | |
| tensor<int32, [3]> var_317_shape_cast_fp16 = shape(x = linear_8_cast_fp16)[name = string("op_317_shape_cast_fp16")]; | |
| int32 gather_14_axis_0 = const()[name = string("gather_14_axis_0"), val = int32(0)]; | |
| int32 gather_14_batch_dims_0 = const()[name = string("gather_14_batch_dims_0"), val = int32(0)]; | |
| bool gather_14_validate_indices_0 = const()[name = string("gather_14_validate_indices_0"), val = bool(false)]; | |
| string var_317_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_317_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")]; | |
| uint16 select_14_to_uint16 = const()[name = string("select_14_to_uint16"), val = uint16(1)]; | |
| tensor<uint16, [3]> var_317_shape_cast_fp16_to_uint16 = cast(dtype = var_317_shape_cast_fp16_to_uint16_dtype_0, x = var_317_shape_cast_fp16)[name = string("cast_52")]; | |
| uint16 gather_14_cast_uint16 = gather(axis = gather_14_axis_0, batch_dims = gather_14_batch_dims_0, indices = select_14_to_uint16, validate_indices = gather_14_validate_indices_0, x = var_317_shape_cast_fp16_to_uint16)[name = string("gather_14_cast_uint16")]; | |
| string gather_14_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_14_cast_uint16_to_int32_dtype_0"), val = string("int32")]; | |
| int32 gather_14_cast_uint16_to_int32 = cast(dtype = gather_14_cast_uint16_to_int32_dtype_0, x = gather_14_cast_uint16)[name = string("cast_51")]; | |
| int32 end_step_5 = add(x = offset, y = gather_14_cast_uint16_to_int32)[name = string("end_step_5")]; | |
| tensor<int32, [1]> expand_dims_16 = const()[name = string("expand_dims_16"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, [1]> expand_dims_18 = const()[name = string("expand_dims_18"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, [1]> expand_dims_19_axes_0 = const()[name = string("expand_dims_19_axes_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, [1]> expand_dims_19 = expand_dims(axes = expand_dims_19_axes_0, x = end_step_5)[name = string("expand_dims_19")]; | |
| tensor<int32, [1]> concat_26_values0_0 = const()[name = string("concat_26_values0_0"), val = tensor<int32, [1]>([1])]; | |
| int32 concat_26_axis_0 = const()[name = string("concat_26_axis_0"), val = int32(0)]; | |
| bool concat_26_interleave_0 = const()[name = string("concat_26_interleave_0"), val = bool(false)]; | |
| tensor<int32, [4]> concat_26 = concat(axis = concat_26_axis_0, interleave = concat_26_interleave_0, values = (concat_26_values0_0, expand_dims_16, expand_dims_1, expand_dims_18))[name = string("concat_26")]; | |
| 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_19, concat_27_values3_0))[name = string("concat_27")]; | |
| tensor<int32, [4]> k_cache1_internal_tensor_assign_2_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])]; | |
| tensor<bool, [4]> k_cache1_internal_tensor_assign_2_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])]; | |
| tensor<bool, [4]> k_cache1_internal_tensor_assign_2_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])]; | |
| tensor<bool, [4]> k_cache1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [4, 1, 448, 384]> k_cache1_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_26, begin_mask = k_cache1_internal_tensor_assign_2_begin_mask_0, end = concat_27, end_mask = k_cache1_internal_tensor_assign_2_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_2_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_2_stride_0, update = linear_9_cast_fp16, x = coreml_update_state_8)[name = string("k_cache1_internal_tensor_assign_2_cast_fp16")]; | |
| write_state(data = k_cache1_internal_tensor_assign_2_cast_fp16, input = k_cache1)[name = string("coreml_update_state_10_write_state")]; | |
| tensor<fp16, [4, 1, 448, 384]> coreml_update_state_10 = read_state(input = k_cache1)[name = string("coreml_update_state_10")]; | |
| tensor<int32, [4]> v_cache1_internal_tensor_assign_2_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])]; | |
| tensor<bool, [4]> v_cache1_internal_tensor_assign_2_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])]; | |
| tensor<bool, [4]> v_cache1_internal_tensor_assign_2_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])]; | |
| tensor<bool, [4]> v_cache1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [4, 1, 448, 384]> v_cache1_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_26, begin_mask = v_cache1_internal_tensor_assign_2_begin_mask_0, end = concat_27, end_mask = v_cache1_internal_tensor_assign_2_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_2_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_2_stride_0, update = linear_10_cast_fp16, x = coreml_update_state_9)[name = string("v_cache1_internal_tensor_assign_2_cast_fp16")]; | |
| write_state(data = v_cache1_internal_tensor_assign_2_cast_fp16, input = v_cache1)[name = string("coreml_update_state_11_write_state")]; | |
| tensor<fp16, [4, 1, 448, 384]> coreml_update_state_11 = read_state(input = v_cache1)[name = string("coreml_update_state_11")]; | |
| int32 concat_32_values0_0 = const()[name = string("concat_32_values0_0"), val = int32(1)]; | |
| int32 concat_32_values2_0 = const()[name = string("concat_32_values2_0"), val = int32(384)]; | |
| int32 concat_32_axis_0 = const()[name = string("concat_32_axis_0"), val = int32(0)]; | |
| bool concat_32_interleave_0 = const()[name = string("concat_32_interleave_0"), val = bool(false)]; | |
| tensor<int32, [3]> concat_32 = concat(axis = concat_32_axis_0, interleave = concat_32_interleave_0, values = (concat_32_values0_0, end_step_5, concat_32_values2_0))[name = string("concat_32")]; | |
| tensor<int32, [3]> var_333_begin_0 = const()[name = string("op_333_begin_0"), val = tensor<int32, [3]>([0, 0, 0])]; | |
| tensor<bool, [3]> var_333_end_mask_0 = const()[name = string("op_333_end_mask_0"), val = tensor<bool, [3]>([true, false, true])]; | |
| tensor<fp16, [1, ?, 384]> var_333_cast_fp16 = slice_by_index(begin = var_333_begin_0, end = concat_32, end_mask = var_333_end_mask_0, x = k_cache_5_cast_fp16)[name = string("op_333_cast_fp16")]; | |
| tensor<int32, [3]> var_336_begin_0 = const()[name = string("op_336_begin_0"), val = tensor<int32, [3]>([0, 0, 0])]; | |
| tensor<bool, [3]> var_336_end_mask_0 = const()[name = string("op_336_end_mask_0"), val = tensor<bool, [3]>([true, false, true])]; | |
| tensor<fp16, [1, ?, 384]> var_336_cast_fp16 = slice_by_index(begin = var_336_begin_0, end = concat_32, end_mask = var_336_end_mask_0, x = v_cache_5_cast_fp16)[name = string("op_336_cast_fp16")]; | |
| tensor<int32, [4]> concat_34x = const()[name = string("concat_34x"), val = tensor<int32, [4]>([1, -1, 6, 64])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_346_cast_fp16 = reshape(shape = concat_34x, x = linear_8_cast_fp16)[name = string("op_346_cast_fp16")]; | |
| tensor<fp16, [1, 1, 1, 1]> const_24_to_fp16 = const()[name = string("const_24_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])]; | |
| tensor<fp16, [1, ?, 6, 64]> q_11_cast_fp16 = mul(x = var_346_cast_fp16, y = const_24_to_fp16)[name = string("q_11_cast_fp16")]; | |
| tensor<int32, [4]> concat_35x = const()[name = string("concat_35x"), val = tensor<int32, [4]>([1, -1, 6, 64])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_353_cast_fp16 = reshape(shape = concat_35x, x = var_333_cast_fp16)[name = string("op_353_cast_fp16")]; | |
| tensor<fp16, [1, 1, 1, 1]> const_25_to_fp16 = const()[name = string("const_25_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])]; | |
| tensor<fp16, [1, ?, 6, 64]> k_15_cast_fp16 = mul(x = var_353_cast_fp16, y = const_25_to_fp16)[name = string("k_15_cast_fp16")]; | |
| tensor<int32, [4]> concat_36x = const()[name = string("concat_36x"), val = tensor<int32, [4]>([1, -1, 6, 64])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_360_cast_fp16 = reshape(shape = concat_36x, x = var_336_cast_fp16)[name = string("op_360_cast_fp16")]; | |
| tensor<int32, [4]> var_361 = const()[name = string("op_361"), 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_37_perm_0 = const()[name = string("transpose_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; | |
| tensor<int32, [4]> transpose_38_perm_0 = const()[name = string("transpose_38_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; | |
| tensor<fp16, [1, 6, 64, ?]> transpose_38 = transpose(perm = transpose_38_perm_0, x = k_15_cast_fp16)[name = string("transpose_70")]; | |
| tensor<fp16, [1, 6, ?, 64]> transpose_37 = transpose(perm = transpose_37_perm_0, x = q_11_cast_fp16)[name = string("transpose_71")]; | |
| tensor<fp16, [1, 6, ?, ?]> qk_7_cast_fp16 = matmul(transpose_x = qk_7_transpose_x_0, transpose_y = qk_7_transpose_y_0, x = transpose_37, y = transpose_38)[name = string("qk_7_cast_fp16")]; | |
| int32 concat_37_values1_0 = const()[name = string("concat_37_values1_0"), val = int32(448)]; | |
| int32 concat_37_axis_0 = const()[name = string("concat_37_axis_0"), val = int32(0)]; | |
| bool concat_37_interleave_0 = const()[name = string("concat_37_interleave_0"), val = bool(false)]; | |
| tensor<int32, [2]> concat_37 = concat(axis = concat_37_axis_0, interleave = concat_37_interleave_0, values = (gather_14_cast_uint16_to_int32, concat_37_values1_0))[name = string("concat_37")]; | |
| tensor<int32, [2]> var_364_begin_0 = const()[name = string("op_364_begin_0"), val = tensor<int32, [2]>([0, 0])]; | |
| tensor<bool, [2]> var_364_end_mask_0 = const()[name = string("op_364_end_mask_0"), val = tensor<bool, [2]>([false, true])]; | |
| tensor<fp16, [?, 448]> var_364_cast_fp16 = slice_by_index(begin = var_364_begin_0, end = concat_37, end_mask = var_364_end_mask_0, x = mask_to_fp16)[name = string("op_364_cast_fp16")]; | |
| int32 concat_38_values0_0 = const()[name = string("concat_38_values0_0"), val = int32(0)]; | |
| int32 concat_38_axis_0 = const()[name = string("concat_38_axis_0"), val = int32(0)]; | |
| bool concat_38_interleave_0 = const()[name = string("concat_38_interleave_0"), val = bool(false)]; | |
| tensor<int32, [2]> concat_38 = concat(axis = concat_38_axis_0, interleave = concat_38_interleave_0, values = (concat_38_values0_0, gather_14_cast_uint16_to_int32))[name = string("concat_38")]; | |
| tensor<int32, [2]> var_365_begin_0 = const()[name = string("op_365_begin_0"), val = tensor<int32, [2]>([0, 0])]; | |
| tensor<bool, [2]> var_365_end_mask_0 = const()[name = string("op_365_end_mask_0"), val = tensor<bool, [2]>([true, false])]; | |
| tensor<fp16, [?, ?]> var_365_cast_fp16 = slice_by_index(begin = var_365_begin_0, end = concat_38, end_mask = var_365_end_mask_0, x = var_364_cast_fp16)[name = string("op_365_cast_fp16")]; | |
| tensor<fp16, [1, 6, ?, ?]> qk_9_cast_fp16 = add(x = qk_7_cast_fp16, y = var_365_cast_fp16)[name = string("qk_9_cast_fp16")]; | |
| tensor<fp16, [1, 6, ?, ?]> var_368_cast_fp16 = softmax(axis = var_277, x = qk_9_cast_fp16)[name = string("op_368_cast_fp16")]; | |
| bool var_370_transpose_x_0 = const()[name = string("op_370_transpose_x_0"), val = bool(false)]; | |
| bool var_370_transpose_y_0 = const()[name = string("op_370_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [1, 6, ?, 64]> v_15_cast_fp16 = transpose(perm = var_361, x = var_360_cast_fp16)[name = string("transpose_72")]; | |
| tensor<fp16, [1, 6, ?, 64]> var_370_cast_fp16 = matmul(transpose_x = var_370_transpose_x_0, transpose_y = var_370_transpose_y_0, x = var_368_cast_fp16, y = v_15_cast_fp16)[name = string("op_370_cast_fp16")]; | |
| tensor<int32, [4]> var_371 = const()[name = string("op_371"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> concat_39x = const()[name = string("concat_39x"), val = tensor<int32, [3]>([1, -1, 384])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_372_cast_fp16 = transpose(perm = var_371, x = var_370_cast_fp16)[name = string("transpose_69")]; | |
| tensor<fp16, [1, ?, 384]> x_25_cast_fp16 = reshape(shape = concat_39x, x = var_372_cast_fp16)[name = string("x_25_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_376_to_fp16 = const()[name = string("op_376_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46761600)))]; | |
| tensor<fp16, [384]> var_377_to_fp16 = const()[name = string("op_377_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47056576)))]; | |
| tensor<fp16, [1, ?, 384]> linear_11_cast_fp16 = linear(bias = var_377_to_fp16, weight = var_376_to_fp16, x = x_25_cast_fp16)[name = string("linear_11_cast_fp16")]; | |
| tensor<fp16, [1, ?, 384]> x_27_cast_fp16 = add(x = x_21_cast_fp16, y = linear_11_cast_fp16)[name = string("x_27_cast_fp16")]; | |
| tensor<int32, [1]> var_384_axes_0 = const()[name = string("op_384_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [384]> blocks_1_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_1_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47057408)))]; | |
| tensor<fp16, [384]> blocks_1_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_1_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47058240)))]; | |
| tensor<fp16, [1, ?, 384]> var_384_cast_fp16 = layer_norm(axes = var_384_axes_0, beta = blocks_1_cross_attn_ln_bias_to_fp16, epsilon = var_283_to_fp16, gamma = blocks_1_cross_attn_ln_weight_to_fp16, x = x_27_cast_fp16)[name = string("op_384_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_393_to_fp16 = const()[name = string("op_393_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47059072)))]; | |
| tensor<fp16, [384]> var_394_to_fp16 = const()[name = string("op_394_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47354048)))]; | |
| tensor<fp16, [1, ?, 384]> linear_12_cast_fp16 = linear(bias = var_394_to_fp16, weight = var_393_to_fp16, x = var_384_cast_fp16)[name = string("linear_12_cast_fp16")]; | |
| tensor<int32, [3]> concat_40 = const()[name = string("concat_40"), val = tensor<int32, [3]>([0, 0, 0])]; | |
| tensor<int32, [3]> concat_41 = const()[name = string("concat_41"), val = tensor<int32, [3]>([0, 1500, 0])]; | |
| tensor<int32, [3]> k_17_internal_tensor_assign_1_stride_0 = const()[name = string("k_17_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])]; | |
| tensor<bool, [3]> k_17_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_17_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])]; | |
| tensor<bool, [3]> k_17_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_17_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])]; | |
| tensor<bool, [3]> k_17_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_17_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])]; | |
| tensor<fp16, [1, 1500, 384]> k_17_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_40, begin_mask = k_17_internal_tensor_assign_1_begin_mask_0, end = concat_41, end_mask = k_17_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_17_internal_tensor_assign_1_squeeze_mask_0, stride = k_17_internal_tensor_assign_1_stride_0, update = k_cache_7_cast_fp16, x = k_7_to_fp16)[name = string("k_17_internal_tensor_assign_1_cast_fp16")]; | |
| tensor<int32, [3]> concat_42 = const()[name = string("concat_42"), val = tensor<int32, [3]>([0, 0, 0])]; | |
| tensor<int32, [3]> concat_43 = const()[name = string("concat_43"), val = tensor<int32, [3]>([0, 1500, 0])]; | |
| tensor<int32, [3]> v_17_internal_tensor_assign_1_stride_0 = const()[name = string("v_17_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])]; | |
| tensor<bool, [3]> v_17_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_17_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])]; | |
| tensor<bool, [3]> v_17_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_17_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])]; | |
| tensor<bool, [3]> v_17_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_17_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])]; | |
| tensor<fp16, [1, 1500, 384]> v_17_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_42, begin_mask = v_17_internal_tensor_assign_1_begin_mask_0, end = concat_43, end_mask = v_17_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_17_internal_tensor_assign_1_squeeze_mask_0, stride = v_17_internal_tensor_assign_1_stride_0, update = v_cache_7_cast_fp16, x = k_7_to_fp16)[name = string("v_17_internal_tensor_assign_1_cast_fp16")]; | |
| tensor<int32, [4]> concat_44x = const()[name = string("concat_44x"), val = tensor<int32, [4]>([1, -1, 6, 64])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_414_cast_fp16 = reshape(shape = concat_44x, x = linear_12_cast_fp16)[name = string("op_414_cast_fp16")]; | |
| tensor<fp16, [1, 1, 1, 1]> const_26_to_fp16 = const()[name = string("const_26_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])]; | |
| tensor<fp16, [1, ?, 6, 64]> q_15_cast_fp16 = mul(x = var_414_cast_fp16, y = const_26_to_fp16)[name = string("q_15_cast_fp16")]; | |
| tensor<int32, [4]> var_420 = const()[name = string("op_420"), val = tensor<int32, [4]>([1, 1500, 6, -1])]; | |
| tensor<fp16, [1, 1500, 6, 64]> var_421_cast_fp16 = reshape(shape = var_420, x = k_17_internal_tensor_assign_1_cast_fp16)[name = string("op_421_cast_fp16")]; | |
| tensor<fp16, [1, 1, 1, 1]> const_27_to_fp16 = const()[name = string("const_27_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])]; | |
| tensor<fp16, [1, 1500, 6, 64]> k_19_cast_fp16 = mul(x = var_421_cast_fp16, y = const_27_to_fp16)[name = string("k_19_cast_fp16")]; | |
| tensor<int32, [4]> var_427 = const()[name = string("op_427"), val = tensor<int32, [4]>([1, 1500, 6, -1])]; | |
| tensor<fp16, [1, 1500, 6, 64]> var_428_cast_fp16 = reshape(shape = var_427, x = v_17_internal_tensor_assign_1_cast_fp16)[name = string("op_428_cast_fp16")]; | |
| tensor<int32, [4]> var_429 = const()[name = string("op_429"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| bool qk_11_transpose_x_0 = const()[name = string("qk_11_transpose_x_0"), val = bool(false)]; | |
| bool qk_11_transpose_y_0 = const()[name = string("qk_11_transpose_y_0"), val = bool(false)]; | |
| tensor<int32, [4]> transpose_39_perm_0 = const()[name = string("transpose_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; | |
| tensor<int32, [4]> transpose_40_perm_0 = const()[name = string("transpose_40_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; | |
| tensor<fp16, [1, 6, 64, 1500]> transpose_40 = transpose(perm = transpose_40_perm_0, x = k_19_cast_fp16)[name = string("transpose_66")]; | |
| tensor<fp16, [1, 6, ?, 64]> transpose_39 = transpose(perm = transpose_39_perm_0, x = q_15_cast_fp16)[name = string("transpose_67")]; | |
| tensor<fp16, [1, 6, ?, 1500]> qk_11_cast_fp16 = matmul(transpose_x = qk_11_transpose_x_0, transpose_y = qk_11_transpose_y_0, x = transpose_39, y = transpose_40)[name = string("qk_11_cast_fp16")]; | |
| tensor<fp16, [1, 6, ?, 1500]> var_433_cast_fp16 = softmax(axis = var_277, x = qk_11_cast_fp16)[name = string("op_433_cast_fp16")]; | |
| bool var_435_transpose_x_0 = const()[name = string("op_435_transpose_x_0"), val = bool(false)]; | |
| bool var_435_transpose_y_0 = const()[name = string("op_435_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [1, 6, 1500, 64]> v_19_cast_fp16 = transpose(perm = var_429, x = var_428_cast_fp16)[name = string("transpose_68")]; | |
| tensor<fp16, [1, 6, ?, 64]> var_435_cast_fp16 = matmul(transpose_x = var_435_transpose_x_0, transpose_y = var_435_transpose_y_0, x = var_433_cast_fp16, y = v_19_cast_fp16)[name = string("op_435_cast_fp16")]; | |
| tensor<int32, [4]> var_436 = const()[name = string("op_436"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> concat_45x = const()[name = string("concat_45x"), val = tensor<int32, [3]>([1, -1, 384])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_437_cast_fp16 = transpose(perm = var_436, x = var_435_cast_fp16)[name = string("transpose_65")]; | |
| tensor<fp16, [1, ?, 384]> x_31_cast_fp16 = reshape(shape = concat_45x, x = var_437_cast_fp16)[name = string("x_31_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_441_to_fp16 = const()[name = string("op_441_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47354880)))]; | |
| tensor<fp16, [384]> var_442_to_fp16 = const()[name = string("op_442_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47649856)))]; | |
| tensor<fp16, [1, ?, 384]> linear_13_cast_fp16 = linear(bias = var_442_to_fp16, weight = var_441_to_fp16, x = x_31_cast_fp16)[name = string("linear_13_cast_fp16")]; | |
| tensor<fp16, [1, ?, 384]> x_33_cast_fp16 = add(x = x_27_cast_fp16, y = linear_13_cast_fp16)[name = string("x_33_cast_fp16")]; | |
| tensor<int32, [1]> var_449_axes_0 = const()[name = string("op_449_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [384]> blocks_1_mlp_ln_weight_to_fp16 = const()[name = string("blocks_1_mlp_ln_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47650688)))]; | |
| tensor<fp16, [384]> blocks_1_mlp_ln_bias_to_fp16 = const()[name = string("blocks_1_mlp_ln_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47651520)))]; | |
| tensor<fp16, [1, ?, 384]> var_449_cast_fp16 = layer_norm(axes = var_449_axes_0, beta = blocks_1_mlp_ln_bias_to_fp16, epsilon = var_283_to_fp16, gamma = blocks_1_mlp_ln_weight_to_fp16, x = x_33_cast_fp16)[name = string("op_449_cast_fp16")]; | |
| tensor<fp16, [1536, 384]> var_458_to_fp16 = const()[name = string("op_458_to_fp16"), val = tensor<fp16, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47652352)))]; | |
| tensor<fp16, [1536]> var_459_to_fp16 = const()[name = string("op_459_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48832064)))]; | |
| tensor<fp16, [1, ?, 1536]> linear_14_cast_fp16 = linear(bias = var_459_to_fp16, weight = var_458_to_fp16, x = var_449_cast_fp16)[name = string("linear_14_cast_fp16")]; | |
| string x_37_mode_0 = const()[name = string("x_37_mode_0"), val = string("EXACT")]; | |
| tensor<fp16, [1, ?, 1536]> x_37_cast_fp16 = gelu(mode = x_37_mode_0, x = linear_14_cast_fp16)[name = string("x_37_cast_fp16")]; | |
| tensor<fp16, [384, 1536]> var_464_to_fp16 = const()[name = string("op_464_to_fp16"), val = tensor<fp16, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48835200)))]; | |
| tensor<fp16, [384]> var_465_to_fp16 = const()[name = string("op_465_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50014912)))]; | |
| tensor<fp16, [1, ?, 384]> linear_15_cast_fp16 = linear(bias = var_465_to_fp16, weight = var_464_to_fp16, x = x_37_cast_fp16)[name = string("linear_15_cast_fp16")]; | |
| tensor<fp16, [1, ?, 384]> x_39_cast_fp16 = add(x = x_33_cast_fp16, y = linear_15_cast_fp16)[name = string("x_39_cast_fp16")]; | |
| tensor<int32, [4]> k_cache_9_begin_0 = const()[name = string("k_cache_9_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])]; | |
| tensor<int32, [4]> k_cache_9_end_0 = const()[name = string("k_cache_9_end_0"), val = tensor<int32, [4]>([3, 1, 448, 384])]; | |
| tensor<bool, [4]> k_cache_9_end_mask_0 = const()[name = string("k_cache_9_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> k_cache_9_squeeze_mask_0 = const()[name = string("k_cache_9_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [1, 448, 384]> k_cache_9_cast_fp16 = slice_by_index(begin = k_cache_9_begin_0, end = k_cache_9_end_0, end_mask = k_cache_9_end_mask_0, squeeze_mask = k_cache_9_squeeze_mask_0, x = coreml_update_state_10)[name = string("k_cache_9_cast_fp16")]; | |
| tensor<int32, [4]> v_cache_9_begin_0 = const()[name = string("v_cache_9_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])]; | |
| tensor<int32, [4]> v_cache_9_end_0 = const()[name = string("v_cache_9_end_0"), val = tensor<int32, [4]>([3, 1, 448, 384])]; | |
| tensor<bool, [4]> v_cache_9_end_mask_0 = const()[name = string("v_cache_9_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> v_cache_9_squeeze_mask_0 = const()[name = string("v_cache_9_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [1, 448, 384]> v_cache_9_cast_fp16 = slice_by_index(begin = v_cache_9_begin_0, end = v_cache_9_end_0, end_mask = v_cache_9_end_mask_0, squeeze_mask = v_cache_9_squeeze_mask_0, x = coreml_update_state_11)[name = string("v_cache_9_cast_fp16")]; | |
| tensor<int32, [4]> k_cache_11_begin_0 = const()[name = string("k_cache_11_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])]; | |
| tensor<int32, [4]> k_cache_11_end_0 = const()[name = string("k_cache_11_end_0"), val = tensor<int32, [4]>([3, 1, 1500, 384])]; | |
| tensor<bool, [4]> k_cache_11_end_mask_0 = const()[name = string("k_cache_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> k_cache_11_squeeze_mask_0 = const()[name = string("k_cache_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [1, 1500, 384]> k_cache_11_cast_fp16 = slice_by_index(begin = k_cache_11_begin_0, end = k_cache_11_end_0, end_mask = k_cache_11_end_mask_0, squeeze_mask = k_cache_11_squeeze_mask_0, x = read_state_2)[name = string("k_cache_11_cast_fp16")]; | |
| tensor<int32, [4]> v_cache_11_begin_0 = const()[name = string("v_cache_11_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])]; | |
| tensor<int32, [4]> v_cache_11_end_0 = const()[name = string("v_cache_11_end_0"), val = tensor<int32, [4]>([3, 1, 1500, 384])]; | |
| tensor<bool, [4]> v_cache_11_end_mask_0 = const()[name = string("v_cache_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> v_cache_11_squeeze_mask_0 = const()[name = string("v_cache_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [1, 1500, 384]> v_cache_11_cast_fp16 = slice_by_index(begin = v_cache_11_begin_0, end = v_cache_11_end_0, end_mask = v_cache_11_end_mask_0, squeeze_mask = v_cache_11_squeeze_mask_0, x = read_state_3)[name = string("v_cache_11_cast_fp16")]; | |
| int32 var_487 = const()[name = string("op_487"), val = int32(-1)]; | |
| tensor<int32, [1]> var_505_axes_0 = const()[name = string("op_505_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [384]> blocks_2_attn_ln_weight_to_fp16 = const()[name = string("blocks_2_attn_ln_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50015744)))]; | |
| tensor<fp16, [384]> blocks_2_attn_ln_bias_to_fp16 = const()[name = string("blocks_2_attn_ln_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50016576)))]; | |
| fp16 var_493_to_fp16 = const()[name = string("op_493_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [1, ?, 384]> var_505_cast_fp16 = layer_norm(axes = var_505_axes_0, beta = blocks_2_attn_ln_bias_to_fp16, epsilon = var_493_to_fp16, gamma = blocks_2_attn_ln_weight_to_fp16, x = x_39_cast_fp16)[name = string("op_505_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_516_to_fp16 = const()[name = string("op_516_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50017408)))]; | |
| tensor<fp16, [384]> var_517_to_fp16 = const()[name = string("op_517_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50312384)))]; | |
| tensor<fp16, [1, ?, 384]> linear_16_cast_fp16 = linear(bias = var_517_to_fp16, weight = var_516_to_fp16, x = var_505_cast_fp16)[name = string("linear_16_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_520_to_fp16 = const()[name = string("op_520_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50313216)))]; | |
| tensor<fp16, [1, ?, 384]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_520_to_fp16, x = var_505_cast_fp16)[name = string("linear_17_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_524_to_fp16 = const()[name = string("op_524_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50608192)))]; | |
| tensor<fp16, [384]> var_525_to_fp16 = const()[name = string("op_525_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50903168)))]; | |
| tensor<fp16, [1, ?, 384]> linear_18_cast_fp16 = linear(bias = var_525_to_fp16, weight = var_524_to_fp16, x = var_505_cast_fp16)[name = string("linear_18_cast_fp16")]; | |
| tensor<int32, [3]> var_527_shape_cast_fp16 = shape(x = linear_16_cast_fp16)[name = string("op_527_shape_cast_fp16")]; | |
| int32 gather_26_axis_0 = const()[name = string("gather_26_axis_0"), val = int32(0)]; | |
| int32 gather_26_batch_dims_0 = const()[name = string("gather_26_batch_dims_0"), val = int32(0)]; | |
| bool gather_26_validate_indices_0 = const()[name = string("gather_26_validate_indices_0"), val = bool(false)]; | |
| string var_527_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_527_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")]; | |
| uint16 select_26_to_uint16 = const()[name = string("select_26_to_uint16"), val = uint16(1)]; | |
| tensor<uint16, [3]> var_527_shape_cast_fp16_to_uint16 = cast(dtype = var_527_shape_cast_fp16_to_uint16_dtype_0, x = var_527_shape_cast_fp16)[name = string("cast_50")]; | |
| uint16 gather_26_cast_uint16 = gather(axis = gather_26_axis_0, batch_dims = gather_26_batch_dims_0, indices = select_26_to_uint16, validate_indices = gather_26_validate_indices_0, x = var_527_shape_cast_fp16_to_uint16)[name = string("gather_26_cast_uint16")]; | |
| string gather_26_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_26_cast_uint16_to_int32_dtype_0"), val = string("int32")]; | |
| int32 gather_26_cast_uint16_to_int32 = cast(dtype = gather_26_cast_uint16_to_int32_dtype_0, x = gather_26_cast_uint16)[name = string("cast_49")]; | |
| int32 end_step_7 = add(x = offset, y = gather_26_cast_uint16_to_int32)[name = string("end_step_7")]; | |
| tensor<int32, [1]> expand_dims_32 = const()[name = string("expand_dims_32"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, [1]> expand_dims_34 = const()[name = string("expand_dims_34"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, [1]> expand_dims_35_axes_0 = const()[name = string("expand_dims_35_axes_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, [1]> expand_dims_35 = expand_dims(axes = expand_dims_35_axes_0, x = end_step_7)[name = string("expand_dims_35")]; | |
| tensor<int32, [1]> concat_48_values0_0 = const()[name = string("concat_48_values0_0"), val = tensor<int32, [1]>([2])]; | |
| int32 concat_48_axis_0 = const()[name = string("concat_48_axis_0"), val = int32(0)]; | |
| bool concat_48_interleave_0 = const()[name = string("concat_48_interleave_0"), val = bool(false)]; | |
| tensor<int32, [4]> concat_48 = concat(axis = concat_48_axis_0, interleave = concat_48_interleave_0, values = (concat_48_values0_0, expand_dims_32, expand_dims_1, expand_dims_34))[name = string("concat_48")]; | |
| tensor<int32, [1]> concat_49_values0_0 = const()[name = string("concat_49_values0_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, [1]> concat_49_values1_0 = const()[name = string("concat_49_values1_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, [1]> concat_49_values3_0 = const()[name = string("concat_49_values3_0"), val = tensor<int32, [1]>([0])]; | |
| int32 concat_49_axis_0 = const()[name = string("concat_49_axis_0"), val = int32(0)]; | |
| bool concat_49_interleave_0 = const()[name = string("concat_49_interleave_0"), val = bool(false)]; | |
| tensor<int32, [4]> concat_49 = concat(axis = concat_49_axis_0, interleave = concat_49_interleave_0, values = (concat_49_values0_0, concat_49_values1_0, expand_dims_35, concat_49_values3_0))[name = string("concat_49")]; | |
| tensor<int32, [4]> k_cache1_internal_tensor_assign_3_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])]; | |
| tensor<bool, [4]> k_cache1_internal_tensor_assign_3_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])]; | |
| tensor<bool, [4]> k_cache1_internal_tensor_assign_3_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])]; | |
| tensor<bool, [4]> k_cache1_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [4, 1, 448, 384]> k_cache1_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_48, begin_mask = k_cache1_internal_tensor_assign_3_begin_mask_0, end = concat_49, end_mask = k_cache1_internal_tensor_assign_3_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_3_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_3_stride_0, update = linear_17_cast_fp16, x = coreml_update_state_10)[name = string("k_cache1_internal_tensor_assign_3_cast_fp16")]; | |
| write_state(data = k_cache1_internal_tensor_assign_3_cast_fp16, input = k_cache1)[name = string("coreml_update_state_12_write_state")]; | |
| tensor<fp16, [4, 1, 448, 384]> coreml_update_state_12 = read_state(input = k_cache1)[name = string("coreml_update_state_12")]; | |
| tensor<int32, [4]> v_cache1_internal_tensor_assign_3_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])]; | |
| tensor<bool, [4]> v_cache1_internal_tensor_assign_3_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])]; | |
| tensor<bool, [4]> v_cache1_internal_tensor_assign_3_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])]; | |
| tensor<bool, [4]> v_cache1_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [4, 1, 448, 384]> v_cache1_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_48, begin_mask = v_cache1_internal_tensor_assign_3_begin_mask_0, end = concat_49, end_mask = v_cache1_internal_tensor_assign_3_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_3_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_3_stride_0, update = linear_18_cast_fp16, x = coreml_update_state_11)[name = string("v_cache1_internal_tensor_assign_3_cast_fp16")]; | |
| write_state(data = v_cache1_internal_tensor_assign_3_cast_fp16, input = v_cache1)[name = string("coreml_update_state_13_write_state")]; | |
| tensor<fp16, [4, 1, 448, 384]> coreml_update_state_13 = read_state(input = v_cache1)[name = string("coreml_update_state_13")]; | |
| int32 concat_54_values0_0 = const()[name = string("concat_54_values0_0"), val = int32(1)]; | |
| int32 concat_54_values2_0 = const()[name = string("concat_54_values2_0"), val = int32(384)]; | |
| int32 concat_54_axis_0 = const()[name = string("concat_54_axis_0"), val = int32(0)]; | |
| bool concat_54_interleave_0 = const()[name = string("concat_54_interleave_0"), val = bool(false)]; | |
| tensor<int32, [3]> concat_54 = concat(axis = concat_54_axis_0, interleave = concat_54_interleave_0, values = (concat_54_values0_0, end_step_7, concat_54_values2_0))[name = string("concat_54")]; | |
| tensor<int32, [3]> var_543_begin_0 = const()[name = string("op_543_begin_0"), val = tensor<int32, [3]>([0, 0, 0])]; | |
| tensor<bool, [3]> var_543_end_mask_0 = const()[name = string("op_543_end_mask_0"), val = tensor<bool, [3]>([true, false, true])]; | |
| tensor<fp16, [1, ?, 384]> var_543_cast_fp16 = slice_by_index(begin = var_543_begin_0, end = concat_54, end_mask = var_543_end_mask_0, x = k_cache_9_cast_fp16)[name = string("op_543_cast_fp16")]; | |
| tensor<int32, [3]> var_546_begin_0 = const()[name = string("op_546_begin_0"), val = tensor<int32, [3]>([0, 0, 0])]; | |
| tensor<bool, [3]> var_546_end_mask_0 = const()[name = string("op_546_end_mask_0"), val = tensor<bool, [3]>([true, false, true])]; | |
| tensor<fp16, [1, ?, 384]> var_546_cast_fp16 = slice_by_index(begin = var_546_begin_0, end = concat_54, end_mask = var_546_end_mask_0, x = v_cache_9_cast_fp16)[name = string("op_546_cast_fp16")]; | |
| tensor<int32, [4]> concat_56x = const()[name = string("concat_56x"), val = tensor<int32, [4]>([1, -1, 6, 64])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_556_cast_fp16 = reshape(shape = concat_56x, x = linear_16_cast_fp16)[name = string("op_556_cast_fp16")]; | |
| tensor<fp16, [1, 1, 1, 1]> const_28_to_fp16 = const()[name = string("const_28_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])]; | |
| tensor<fp16, [1, ?, 6, 64]> q_19_cast_fp16 = mul(x = var_556_cast_fp16, y = const_28_to_fp16)[name = string("q_19_cast_fp16")]; | |
| tensor<int32, [4]> concat_57x = const()[name = string("concat_57x"), val = tensor<int32, [4]>([1, -1, 6, 64])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_563_cast_fp16 = reshape(shape = concat_57x, x = var_543_cast_fp16)[name = string("op_563_cast_fp16")]; | |
| tensor<fp16, [1, 1, 1, 1]> const_29_to_fp16 = const()[name = string("const_29_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])]; | |
| tensor<fp16, [1, ?, 6, 64]> k_25_cast_fp16 = mul(x = var_563_cast_fp16, y = const_29_to_fp16)[name = string("k_25_cast_fp16")]; | |
| tensor<int32, [4]> concat_58x = const()[name = string("concat_58x"), val = tensor<int32, [4]>([1, -1, 6, 64])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_570_cast_fp16 = reshape(shape = concat_58x, x = var_546_cast_fp16)[name = string("op_570_cast_fp16")]; | |
| tensor<int32, [4]> var_571 = const()[name = string("op_571"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| bool qk_13_transpose_x_0 = const()[name = string("qk_13_transpose_x_0"), val = bool(false)]; | |
| bool qk_13_transpose_y_0 = const()[name = string("qk_13_transpose_y_0"), val = bool(false)]; | |
| tensor<int32, [4]> transpose_41_perm_0 = const()[name = string("transpose_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; | |
| tensor<int32, [4]> transpose_42_perm_0 = const()[name = string("transpose_42_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; | |
| tensor<fp16, [1, 6, 64, ?]> transpose_42 = transpose(perm = transpose_42_perm_0, x = k_25_cast_fp16)[name = string("transpose_62")]; | |
| tensor<fp16, [1, 6, ?, 64]> transpose_41 = transpose(perm = transpose_41_perm_0, x = q_19_cast_fp16)[name = string("transpose_63")]; | |
| tensor<fp16, [1, 6, ?, ?]> qk_13_cast_fp16 = matmul(transpose_x = qk_13_transpose_x_0, transpose_y = qk_13_transpose_y_0, x = transpose_41, y = transpose_42)[name = string("qk_13_cast_fp16")]; | |
| int32 concat_59_values1_0 = const()[name = string("concat_59_values1_0"), val = int32(448)]; | |
| int32 concat_59_axis_0 = const()[name = string("concat_59_axis_0"), val = int32(0)]; | |
| bool concat_59_interleave_0 = const()[name = string("concat_59_interleave_0"), val = bool(false)]; | |
| tensor<int32, [2]> concat_59 = concat(axis = concat_59_axis_0, interleave = concat_59_interleave_0, values = (gather_26_cast_uint16_to_int32, concat_59_values1_0))[name = string("concat_59")]; | |
| tensor<int32, [2]> var_574_begin_0 = const()[name = string("op_574_begin_0"), val = tensor<int32, [2]>([0, 0])]; | |
| tensor<bool, [2]> var_574_end_mask_0 = const()[name = string("op_574_end_mask_0"), val = tensor<bool, [2]>([false, true])]; | |
| tensor<fp16, [?, 448]> var_574_cast_fp16 = slice_by_index(begin = var_574_begin_0, end = concat_59, end_mask = var_574_end_mask_0, x = mask_to_fp16)[name = string("op_574_cast_fp16")]; | |
| int32 concat_60_values0_0 = const()[name = string("concat_60_values0_0"), val = int32(0)]; | |
| int32 concat_60_axis_0 = const()[name = string("concat_60_axis_0"), val = int32(0)]; | |
| bool concat_60_interleave_0 = const()[name = string("concat_60_interleave_0"), val = bool(false)]; | |
| tensor<int32, [2]> concat_60 = concat(axis = concat_60_axis_0, interleave = concat_60_interleave_0, values = (concat_60_values0_0, gather_26_cast_uint16_to_int32))[name = string("concat_60")]; | |
| tensor<int32, [2]> var_575_begin_0 = const()[name = string("op_575_begin_0"), val = tensor<int32, [2]>([0, 0])]; | |
| tensor<bool, [2]> var_575_end_mask_0 = const()[name = string("op_575_end_mask_0"), val = tensor<bool, [2]>([true, false])]; | |
| tensor<fp16, [?, ?]> var_575_cast_fp16 = slice_by_index(begin = var_575_begin_0, end = concat_60, end_mask = var_575_end_mask_0, x = var_574_cast_fp16)[name = string("op_575_cast_fp16")]; | |
| tensor<fp16, [1, 6, ?, ?]> qk_15_cast_fp16 = add(x = qk_13_cast_fp16, y = var_575_cast_fp16)[name = string("qk_15_cast_fp16")]; | |
| tensor<fp16, [1, 6, ?, ?]> var_578_cast_fp16 = softmax(axis = var_487, x = qk_15_cast_fp16)[name = string("op_578_cast_fp16")]; | |
| bool var_580_transpose_x_0 = const()[name = string("op_580_transpose_x_0"), val = bool(false)]; | |
| bool var_580_transpose_y_0 = const()[name = string("op_580_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [1, 6, ?, 64]> v_25_cast_fp16 = transpose(perm = var_571, x = var_570_cast_fp16)[name = string("transpose_64")]; | |
| tensor<fp16, [1, 6, ?, 64]> var_580_cast_fp16 = matmul(transpose_x = var_580_transpose_x_0, transpose_y = var_580_transpose_y_0, x = var_578_cast_fp16, y = v_25_cast_fp16)[name = string("op_580_cast_fp16")]; | |
| tensor<int32, [4]> var_581 = const()[name = string("op_581"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> concat_61x = const()[name = string("concat_61x"), val = tensor<int32, [3]>([1, -1, 384])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_582_cast_fp16 = transpose(perm = var_581, x = var_580_cast_fp16)[name = string("transpose_61")]; | |
| tensor<fp16, [1, ?, 384]> x_43_cast_fp16 = reshape(shape = concat_61x, x = var_582_cast_fp16)[name = string("x_43_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_586_to_fp16 = const()[name = string("op_586_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50904000)))]; | |
| tensor<fp16, [384]> var_587_to_fp16 = const()[name = string("op_587_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51198976)))]; | |
| tensor<fp16, [1, ?, 384]> linear_19_cast_fp16 = linear(bias = var_587_to_fp16, weight = var_586_to_fp16, x = x_43_cast_fp16)[name = string("linear_19_cast_fp16")]; | |
| tensor<fp16, [1, ?, 384]> x_45_cast_fp16 = add(x = x_39_cast_fp16, y = linear_19_cast_fp16)[name = string("x_45_cast_fp16")]; | |
| tensor<int32, [1]> var_594_axes_0 = const()[name = string("op_594_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [384]> blocks_2_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_2_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51199808)))]; | |
| tensor<fp16, [384]> blocks_2_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_2_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51200640)))]; | |
| tensor<fp16, [1, ?, 384]> var_594_cast_fp16 = layer_norm(axes = var_594_axes_0, beta = blocks_2_cross_attn_ln_bias_to_fp16, epsilon = var_493_to_fp16, gamma = blocks_2_cross_attn_ln_weight_to_fp16, x = x_45_cast_fp16)[name = string("op_594_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_603_to_fp16 = const()[name = string("op_603_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51201472)))]; | |
| tensor<fp16, [384]> var_604_to_fp16 = const()[name = string("op_604_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51496448)))]; | |
| tensor<fp16, [1, ?, 384]> linear_20_cast_fp16 = linear(bias = var_604_to_fp16, weight = var_603_to_fp16, x = var_594_cast_fp16)[name = string("linear_20_cast_fp16")]; | |
| tensor<int32, [3]> concat_62 = const()[name = string("concat_62"), val = tensor<int32, [3]>([0, 0, 0])]; | |
| tensor<int32, [3]> concat_63 = const()[name = string("concat_63"), val = tensor<int32, [3]>([0, 1500, 0])]; | |
| tensor<int32, [3]> k_27_internal_tensor_assign_1_stride_0 = const()[name = string("k_27_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])]; | |
| tensor<bool, [3]> k_27_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_27_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])]; | |
| tensor<bool, [3]> k_27_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_27_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])]; | |
| tensor<bool, [3]> k_27_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_27_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])]; | |
| tensor<fp16, [1, 1500, 384]> k_27_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_62, begin_mask = k_27_internal_tensor_assign_1_begin_mask_0, end = concat_63, end_mask = k_27_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_27_internal_tensor_assign_1_squeeze_mask_0, stride = k_27_internal_tensor_assign_1_stride_0, update = k_cache_11_cast_fp16, x = k_7_to_fp16)[name = string("k_27_internal_tensor_assign_1_cast_fp16")]; | |
| tensor<int32, [3]> concat_64 = const()[name = string("concat_64"), val = tensor<int32, [3]>([0, 0, 0])]; | |
| tensor<int32, [3]> concat_65 = const()[name = string("concat_65"), val = tensor<int32, [3]>([0, 1500, 0])]; | |
| tensor<int32, [3]> v_27_internal_tensor_assign_1_stride_0 = const()[name = string("v_27_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])]; | |
| tensor<bool, [3]> v_27_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_27_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])]; | |
| tensor<bool, [3]> v_27_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_27_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])]; | |
| tensor<bool, [3]> v_27_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_27_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])]; | |
| tensor<fp16, [1, 1500, 384]> v_27_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_64, begin_mask = v_27_internal_tensor_assign_1_begin_mask_0, end = concat_65, end_mask = v_27_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_27_internal_tensor_assign_1_squeeze_mask_0, stride = v_27_internal_tensor_assign_1_stride_0, update = v_cache_11_cast_fp16, x = k_7_to_fp16)[name = string("v_27_internal_tensor_assign_1_cast_fp16")]; | |
| tensor<int32, [4]> concat_66x = const()[name = string("concat_66x"), val = tensor<int32, [4]>([1, -1, 6, 64])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_624_cast_fp16 = reshape(shape = concat_66x, x = linear_20_cast_fp16)[name = string("op_624_cast_fp16")]; | |
| tensor<fp16, [1, 1, 1, 1]> const_30_to_fp16 = const()[name = string("const_30_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])]; | |
| tensor<fp16, [1, ?, 6, 64]> q_23_cast_fp16 = mul(x = var_624_cast_fp16, y = const_30_to_fp16)[name = string("q_23_cast_fp16")]; | |
| tensor<int32, [4]> var_630 = const()[name = string("op_630"), val = tensor<int32, [4]>([1, 1500, 6, -1])]; | |
| tensor<fp16, [1, 1500, 6, 64]> var_631_cast_fp16 = reshape(shape = var_630, x = k_27_internal_tensor_assign_1_cast_fp16)[name = string("op_631_cast_fp16")]; | |
| tensor<fp16, [1, 1, 1, 1]> const_31_to_fp16 = const()[name = string("const_31_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])]; | |
| tensor<fp16, [1, 1500, 6, 64]> k_29_cast_fp16 = mul(x = var_631_cast_fp16, y = const_31_to_fp16)[name = string("k_29_cast_fp16")]; | |
| tensor<int32, [4]> var_637 = const()[name = string("op_637"), val = tensor<int32, [4]>([1, 1500, 6, -1])]; | |
| tensor<fp16, [1, 1500, 6, 64]> var_638_cast_fp16 = reshape(shape = var_637, x = v_27_internal_tensor_assign_1_cast_fp16)[name = string("op_638_cast_fp16")]; | |
| tensor<int32, [4]> var_639 = const()[name = string("op_639"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| bool qk_17_transpose_x_0 = const()[name = string("qk_17_transpose_x_0"), val = bool(false)]; | |
| bool qk_17_transpose_y_0 = const()[name = string("qk_17_transpose_y_0"), val = bool(false)]; | |
| tensor<int32, [4]> transpose_43_perm_0 = const()[name = string("transpose_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; | |
| tensor<int32, [4]> transpose_44_perm_0 = const()[name = string("transpose_44_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; | |
| tensor<fp16, [1, 6, 64, 1500]> transpose_44 = transpose(perm = transpose_44_perm_0, x = k_29_cast_fp16)[name = string("transpose_58")]; | |
| tensor<fp16, [1, 6, ?, 64]> transpose_43 = transpose(perm = transpose_43_perm_0, x = q_23_cast_fp16)[name = string("transpose_59")]; | |
| tensor<fp16, [1, 6, ?, 1500]> qk_17_cast_fp16 = matmul(transpose_x = qk_17_transpose_x_0, transpose_y = qk_17_transpose_y_0, x = transpose_43, y = transpose_44)[name = string("qk_17_cast_fp16")]; | |
| tensor<fp16, [1, 6, ?, 1500]> var_643_cast_fp16 = softmax(axis = var_487, x = qk_17_cast_fp16)[name = string("op_643_cast_fp16")]; | |
| bool var_645_transpose_x_0 = const()[name = string("op_645_transpose_x_0"), val = bool(false)]; | |
| bool var_645_transpose_y_0 = const()[name = string("op_645_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [1, 6, 1500, 64]> v_29_cast_fp16 = transpose(perm = var_639, x = var_638_cast_fp16)[name = string("transpose_60")]; | |
| tensor<fp16, [1, 6, ?, 64]> var_645_cast_fp16 = matmul(transpose_x = var_645_transpose_x_0, transpose_y = var_645_transpose_y_0, x = var_643_cast_fp16, y = v_29_cast_fp16)[name = string("op_645_cast_fp16")]; | |
| tensor<int32, [4]> var_646 = const()[name = string("op_646"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> concat_67x = const()[name = string("concat_67x"), val = tensor<int32, [3]>([1, -1, 384])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_647_cast_fp16 = transpose(perm = var_646, x = var_645_cast_fp16)[name = string("transpose_57")]; | |
| tensor<fp16, [1, ?, 384]> x_49_cast_fp16 = reshape(shape = concat_67x, x = var_647_cast_fp16)[name = string("x_49_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_651_to_fp16 = const()[name = string("op_651_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51497280)))]; | |
| tensor<fp16, [384]> var_652_to_fp16 = const()[name = string("op_652_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51792256)))]; | |
| tensor<fp16, [1, ?, 384]> linear_21_cast_fp16 = linear(bias = var_652_to_fp16, weight = var_651_to_fp16, x = x_49_cast_fp16)[name = string("linear_21_cast_fp16")]; | |
| tensor<fp16, [1, ?, 384]> x_51_cast_fp16 = add(x = x_45_cast_fp16, y = linear_21_cast_fp16)[name = string("x_51_cast_fp16")]; | |
| tensor<int32, [1]> var_659_axes_0 = const()[name = string("op_659_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [384]> blocks_2_mlp_ln_weight_to_fp16 = const()[name = string("blocks_2_mlp_ln_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51793088)))]; | |
| tensor<fp16, [384]> blocks_2_mlp_ln_bias_to_fp16 = const()[name = string("blocks_2_mlp_ln_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51793920)))]; | |
| tensor<fp16, [1, ?, 384]> var_659_cast_fp16 = layer_norm(axes = var_659_axes_0, beta = blocks_2_mlp_ln_bias_to_fp16, epsilon = var_493_to_fp16, gamma = blocks_2_mlp_ln_weight_to_fp16, x = x_51_cast_fp16)[name = string("op_659_cast_fp16")]; | |
| tensor<fp16, [1536, 384]> var_668_to_fp16 = const()[name = string("op_668_to_fp16"), val = tensor<fp16, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51794752)))]; | |
| tensor<fp16, [1536]> var_669_to_fp16 = const()[name = string("op_669_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52974464)))]; | |
| tensor<fp16, [1, ?, 1536]> linear_22_cast_fp16 = linear(bias = var_669_to_fp16, weight = var_668_to_fp16, x = var_659_cast_fp16)[name = string("linear_22_cast_fp16")]; | |
| string x_55_mode_0 = const()[name = string("x_55_mode_0"), val = string("EXACT")]; | |
| tensor<fp16, [1, ?, 1536]> x_55_cast_fp16 = gelu(mode = x_55_mode_0, x = linear_22_cast_fp16)[name = string("x_55_cast_fp16")]; | |
| tensor<fp16, [384, 1536]> var_674_to_fp16 = const()[name = string("op_674_to_fp16"), val = tensor<fp16, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52977600)))]; | |
| tensor<fp16, [384]> var_675_to_fp16 = const()[name = string("op_675_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54157312)))]; | |
| tensor<fp16, [1, ?, 384]> linear_23_cast_fp16 = linear(bias = var_675_to_fp16, weight = var_674_to_fp16, x = x_55_cast_fp16)[name = string("linear_23_cast_fp16")]; | |
| tensor<fp16, [1, ?, 384]> x_57_cast_fp16 = add(x = x_51_cast_fp16, y = linear_23_cast_fp16)[name = string("x_57_cast_fp16")]; | |
| tensor<int32, [4]> k_cache_13_begin_0 = const()[name = string("k_cache_13_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])]; | |
| tensor<int32, [4]> k_cache_13_end_0 = const()[name = string("k_cache_13_end_0"), val = tensor<int32, [4]>([4, 1, 448, 384])]; | |
| tensor<bool, [4]> k_cache_13_end_mask_0 = const()[name = string("k_cache_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> k_cache_13_squeeze_mask_0 = const()[name = string("k_cache_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [1, 448, 384]> k_cache_13_cast_fp16 = slice_by_index(begin = k_cache_13_begin_0, end = k_cache_13_end_0, end_mask = k_cache_13_end_mask_0, squeeze_mask = k_cache_13_squeeze_mask_0, x = coreml_update_state_12)[name = string("k_cache_13_cast_fp16")]; | |
| tensor<int32, [4]> v_cache_13_begin_0 = const()[name = string("v_cache_13_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])]; | |
| tensor<int32, [4]> v_cache_13_end_0 = const()[name = string("v_cache_13_end_0"), val = tensor<int32, [4]>([4, 1, 448, 384])]; | |
| tensor<bool, [4]> v_cache_13_end_mask_0 = const()[name = string("v_cache_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> v_cache_13_squeeze_mask_0 = const()[name = string("v_cache_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [1, 448, 384]> v_cache_13_cast_fp16 = slice_by_index(begin = v_cache_13_begin_0, end = v_cache_13_end_0, end_mask = v_cache_13_end_mask_0, squeeze_mask = v_cache_13_squeeze_mask_0, x = coreml_update_state_13)[name = string("v_cache_13_cast_fp16")]; | |
| tensor<int32, [4]> k_cache_begin_0 = const()[name = string("k_cache_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])]; | |
| tensor<int32, [4]> k_cache_end_0 = const()[name = string("k_cache_end_0"), val = tensor<int32, [4]>([4, 1, 1500, 384])]; | |
| tensor<bool, [4]> k_cache_end_mask_0 = const()[name = string("k_cache_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> k_cache_squeeze_mask_0 = const()[name = string("k_cache_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [1, 1500, 384]> k_cache_cast_fp16 = slice_by_index(begin = k_cache_begin_0, end = k_cache_end_0, end_mask = k_cache_end_mask_0, squeeze_mask = k_cache_squeeze_mask_0, x = read_state_2)[name = string("k_cache_cast_fp16")]; | |
| tensor<int32, [4]> v_cache_begin_0 = const()[name = string("v_cache_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])]; | |
| tensor<int32, [4]> v_cache_end_0 = const()[name = string("v_cache_end_0"), val = tensor<int32, [4]>([4, 1, 1500, 384])]; | |
| tensor<bool, [4]> v_cache_end_mask_0 = const()[name = string("v_cache_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; | |
| tensor<bool, [4]> v_cache_squeeze_mask_0 = const()[name = string("v_cache_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [1, 1500, 384]> v_cache_cast_fp16 = slice_by_index(begin = v_cache_begin_0, end = v_cache_end_0, end_mask = v_cache_end_mask_0, squeeze_mask = v_cache_squeeze_mask_0, x = read_state_3)[name = string("v_cache_cast_fp16")]; | |
| int32 var_697 = const()[name = string("op_697"), val = int32(-1)]; | |
| tensor<int32, [1]> var_715_axes_0 = const()[name = string("op_715_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [384]> blocks_3_attn_ln_weight_to_fp16 = const()[name = string("blocks_3_attn_ln_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54158144)))]; | |
| tensor<fp16, [384]> blocks_3_attn_ln_bias_to_fp16 = const()[name = string("blocks_3_attn_ln_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54158976)))]; | |
| fp16 var_703_to_fp16 = const()[name = string("op_703_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [1, ?, 384]> var_715_cast_fp16 = layer_norm(axes = var_715_axes_0, beta = blocks_3_attn_ln_bias_to_fp16, epsilon = var_703_to_fp16, gamma = blocks_3_attn_ln_weight_to_fp16, x = x_57_cast_fp16)[name = string("op_715_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_726_to_fp16 = const()[name = string("op_726_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54159808)))]; | |
| tensor<fp16, [384]> var_727_to_fp16 = const()[name = string("op_727_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54454784)))]; | |
| tensor<fp16, [1, ?, 384]> linear_24_cast_fp16 = linear(bias = var_727_to_fp16, weight = var_726_to_fp16, x = var_715_cast_fp16)[name = string("linear_24_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_730_to_fp16 = const()[name = string("op_730_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54455616)))]; | |
| tensor<fp16, [1, ?, 384]> linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_730_to_fp16, x = var_715_cast_fp16)[name = string("linear_25_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_734_to_fp16 = const()[name = string("op_734_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54750592)))]; | |
| tensor<fp16, [384]> var_735_to_fp16 = const()[name = string("op_735_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55045568)))]; | |
| tensor<fp16, [1, ?, 384]> linear_26_cast_fp16 = linear(bias = var_735_to_fp16, weight = var_734_to_fp16, x = var_715_cast_fp16)[name = string("linear_26_cast_fp16")]; | |
| tensor<int32, [3]> var_737_shape_cast_fp16 = shape(x = linear_24_cast_fp16)[name = string("op_737_shape_cast_fp16")]; | |
| int32 gather_38_axis_0 = const()[name = string("gather_38_axis_0"), val = int32(0)]; | |
| int32 gather_38_batch_dims_0 = const()[name = string("gather_38_batch_dims_0"), val = int32(0)]; | |
| bool gather_38_validate_indices_0 = const()[name = string("gather_38_validate_indices_0"), val = bool(false)]; | |
| string var_737_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_737_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")]; | |
| uint16 select_38_to_uint16 = const()[name = string("select_38_to_uint16"), val = uint16(1)]; | |
| tensor<uint16, [3]> var_737_shape_cast_fp16_to_uint16 = cast(dtype = var_737_shape_cast_fp16_to_uint16_dtype_0, x = var_737_shape_cast_fp16)[name = string("cast_48")]; | |
| uint16 gather_38_cast_uint16 = gather(axis = gather_38_axis_0, batch_dims = gather_38_batch_dims_0, indices = select_38_to_uint16, validate_indices = gather_38_validate_indices_0, x = var_737_shape_cast_fp16_to_uint16)[name = string("gather_38_cast_uint16")]; | |
| string gather_38_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_38_cast_uint16_to_int32_dtype_0"), val = string("int32")]; | |
| int32 gather_38_cast_uint16_to_int32 = cast(dtype = gather_38_cast_uint16_to_int32_dtype_0, x = gather_38_cast_uint16)[name = string("cast_47")]; | |
| int32 end_step = add(x = offset, y = gather_38_cast_uint16_to_int32)[name = string("end_step")]; | |
| tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, [1]> expand_dims_50 = const()[name = string("expand_dims_50"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, [1]> expand_dims_51_axes_0 = const()[name = string("expand_dims_51_axes_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, [1]> expand_dims_51 = expand_dims(axes = expand_dims_51_axes_0, x = end_step)[name = string("expand_dims_51")]; | |
| tensor<int32, [1]> concat_70_values0_0 = const()[name = string("concat_70_values0_0"), val = tensor<int32, [1]>([3])]; | |
| int32 concat_70_axis_0 = const()[name = string("concat_70_axis_0"), val = int32(0)]; | |
| bool concat_70_interleave_0 = const()[name = string("concat_70_interleave_0"), val = bool(false)]; | |
| tensor<int32, [4]> concat_70 = concat(axis = concat_70_axis_0, interleave = concat_70_interleave_0, values = (concat_70_values0_0, expand_dims_48, expand_dims_1, expand_dims_50))[name = string("concat_70")]; | |
| tensor<int32, [1]> concat_71_values0_0 = const()[name = string("concat_71_values0_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, [1]> concat_71_values1_0 = const()[name = string("concat_71_values1_0"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, [1]> concat_71_values3_0 = const()[name = string("concat_71_values3_0"), val = tensor<int32, [1]>([0])]; | |
| int32 concat_71_axis_0 = const()[name = string("concat_71_axis_0"), val = int32(0)]; | |
| bool concat_71_interleave_0 = const()[name = string("concat_71_interleave_0"), val = bool(false)]; | |
| tensor<int32, [4]> concat_71 = concat(axis = concat_71_axis_0, interleave = concat_71_interleave_0, values = (concat_71_values0_0, concat_71_values1_0, expand_dims_51, concat_71_values3_0))[name = string("concat_71")]; | |
| tensor<int32, [4]> k_cache1_internal_tensor_assign_4_stride_0 = const()[name = string("k_cache1_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])]; | |
| tensor<bool, [4]> k_cache1_internal_tensor_assign_4_begin_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])]; | |
| tensor<bool, [4]> k_cache1_internal_tensor_assign_4_end_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])]; | |
| tensor<bool, [4]> k_cache1_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("k_cache1_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [4, 1, 448, 384]> k_cache1_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_70, begin_mask = k_cache1_internal_tensor_assign_4_begin_mask_0, end = concat_71, end_mask = k_cache1_internal_tensor_assign_4_end_mask_0, squeeze_mask = k_cache1_internal_tensor_assign_4_squeeze_mask_0, stride = k_cache1_internal_tensor_assign_4_stride_0, update = linear_25_cast_fp16, x = coreml_update_state_12)[name = string("k_cache1_internal_tensor_assign_4_cast_fp16")]; | |
| write_state(data = k_cache1_internal_tensor_assign_4_cast_fp16, input = k_cache1)[name = string("coreml_update_state_14_write_state")]; | |
| tensor<int32, [4]> v_cache1_internal_tensor_assign_4_stride_0 = const()[name = string("v_cache1_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])]; | |
| tensor<bool, [4]> v_cache1_internal_tensor_assign_4_begin_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, false, false, false])]; | |
| tensor<bool, [4]> v_cache1_internal_tensor_assign_4_end_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, false, true])]; | |
| tensor<bool, [4]> v_cache1_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("v_cache1_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; | |
| tensor<fp16, [4, 1, 448, 384]> v_cache1_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_70, begin_mask = v_cache1_internal_tensor_assign_4_begin_mask_0, end = concat_71, end_mask = v_cache1_internal_tensor_assign_4_end_mask_0, squeeze_mask = v_cache1_internal_tensor_assign_4_squeeze_mask_0, stride = v_cache1_internal_tensor_assign_4_stride_0, update = linear_26_cast_fp16, x = coreml_update_state_13)[name = string("v_cache1_internal_tensor_assign_4_cast_fp16")]; | |
| write_state(data = v_cache1_internal_tensor_assign_4_cast_fp16, input = v_cache1)[name = string("coreml_update_state_15_write_state")]; | |
| int32 concat_76_values0_0 = const()[name = string("concat_76_values0_0"), val = int32(1)]; | |
| int32 concat_76_values2_0 = const()[name = string("concat_76_values2_0"), val = int32(384)]; | |
| int32 concat_76_axis_0 = const()[name = string("concat_76_axis_0"), val = int32(0)]; | |
| bool concat_76_interleave_0 = const()[name = string("concat_76_interleave_0"), val = bool(false)]; | |
| tensor<int32, [3]> concat_76 = concat(axis = concat_76_axis_0, interleave = concat_76_interleave_0, values = (concat_76_values0_0, end_step, concat_76_values2_0))[name = string("concat_76")]; | |
| tensor<int32, [3]> var_753_begin_0 = const()[name = string("op_753_begin_0"), val = tensor<int32, [3]>([0, 0, 0])]; | |
| tensor<bool, [3]> var_753_end_mask_0 = const()[name = string("op_753_end_mask_0"), val = tensor<bool, [3]>([true, false, true])]; | |
| tensor<fp16, [1, ?, 384]> var_753_cast_fp16 = slice_by_index(begin = var_753_begin_0, end = concat_76, end_mask = var_753_end_mask_0, x = k_cache_13_cast_fp16)[name = string("op_753_cast_fp16")]; | |
| tensor<int32, [3]> var_756_begin_0 = const()[name = string("op_756_begin_0"), val = tensor<int32, [3]>([0, 0, 0])]; | |
| tensor<bool, [3]> var_756_end_mask_0 = const()[name = string("op_756_end_mask_0"), val = tensor<bool, [3]>([true, false, true])]; | |
| tensor<fp16, [1, ?, 384]> var_756_cast_fp16 = slice_by_index(begin = var_756_begin_0, end = concat_76, end_mask = var_756_end_mask_0, x = v_cache_13_cast_fp16)[name = string("op_756_cast_fp16")]; | |
| tensor<int32, [4]> concat_78x = const()[name = string("concat_78x"), val = tensor<int32, [4]>([1, -1, 6, 64])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_766_cast_fp16 = reshape(shape = concat_78x, x = linear_24_cast_fp16)[name = string("op_766_cast_fp16")]; | |
| tensor<fp16, [1, 1, 1, 1]> const_32_to_fp16 = const()[name = string("const_32_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])]; | |
| tensor<fp16, [1, ?, 6, 64]> q_27_cast_fp16 = mul(x = var_766_cast_fp16, y = const_32_to_fp16)[name = string("q_27_cast_fp16")]; | |
| tensor<int32, [4]> concat_79x = const()[name = string("concat_79x"), val = tensor<int32, [4]>([1, -1, 6, 64])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_773_cast_fp16 = reshape(shape = concat_79x, x = var_753_cast_fp16)[name = string("op_773_cast_fp16")]; | |
| tensor<fp16, [1, 1, 1, 1]> const_33_to_fp16 = const()[name = string("const_33_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])]; | |
| tensor<fp16, [1, ?, 6, 64]> k_35_cast_fp16 = mul(x = var_773_cast_fp16, y = const_33_to_fp16)[name = string("k_35_cast_fp16")]; | |
| tensor<int32, [4]> concat_80x = const()[name = string("concat_80x"), val = tensor<int32, [4]>([1, -1, 6, 64])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_780_cast_fp16 = reshape(shape = concat_80x, x = var_756_cast_fp16)[name = string("op_780_cast_fp16")]; | |
| tensor<int32, [4]> var_781 = const()[name = string("op_781"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| bool qk_19_transpose_x_0 = const()[name = string("qk_19_transpose_x_0"), val = bool(false)]; | |
| bool qk_19_transpose_y_0 = const()[name = string("qk_19_transpose_y_0"), val = bool(false)]; | |
| tensor<int32, [4]> transpose_45_perm_0 = const()[name = string("transpose_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; | |
| tensor<int32, [4]> transpose_46_perm_0 = const()[name = string("transpose_46_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; | |
| tensor<fp16, [1, 6, 64, ?]> transpose_46 = transpose(perm = transpose_46_perm_0, x = k_35_cast_fp16)[name = string("transpose_54")]; | |
| tensor<fp16, [1, 6, ?, 64]> transpose_45 = transpose(perm = transpose_45_perm_0, x = q_27_cast_fp16)[name = string("transpose_55")]; | |
| tensor<fp16, [1, 6, ?, ?]> qk_19_cast_fp16 = matmul(transpose_x = qk_19_transpose_x_0, transpose_y = qk_19_transpose_y_0, x = transpose_45, y = transpose_46)[name = string("qk_19_cast_fp16")]; | |
| int32 concat_81_values1_0 = const()[name = string("concat_81_values1_0"), val = int32(448)]; | |
| int32 concat_81_axis_0 = const()[name = string("concat_81_axis_0"), val = int32(0)]; | |
| bool concat_81_interleave_0 = const()[name = string("concat_81_interleave_0"), val = bool(false)]; | |
| tensor<int32, [2]> concat_81 = concat(axis = concat_81_axis_0, interleave = concat_81_interleave_0, values = (gather_38_cast_uint16_to_int32, concat_81_values1_0))[name = string("concat_81")]; | |
| tensor<int32, [2]> var_784_begin_0 = const()[name = string("op_784_begin_0"), val = tensor<int32, [2]>([0, 0])]; | |
| tensor<bool, [2]> var_784_end_mask_0 = const()[name = string("op_784_end_mask_0"), val = tensor<bool, [2]>([false, true])]; | |
| tensor<fp16, [?, 448]> var_784_cast_fp16 = slice_by_index(begin = var_784_begin_0, end = concat_81, end_mask = var_784_end_mask_0, x = mask_to_fp16)[name = string("op_784_cast_fp16")]; | |
| int32 concat_82_values0_0 = const()[name = string("concat_82_values0_0"), val = int32(0)]; | |
| int32 concat_82_axis_0 = const()[name = string("concat_82_axis_0"), val = int32(0)]; | |
| bool concat_82_interleave_0 = const()[name = string("concat_82_interleave_0"), val = bool(false)]; | |
| tensor<int32, [2]> concat_82 = concat(axis = concat_82_axis_0, interleave = concat_82_interleave_0, values = (concat_82_values0_0, gather_38_cast_uint16_to_int32))[name = string("concat_82")]; | |
| tensor<int32, [2]> var_785_begin_0 = const()[name = string("op_785_begin_0"), val = tensor<int32, [2]>([0, 0])]; | |
| tensor<bool, [2]> var_785_end_mask_0 = const()[name = string("op_785_end_mask_0"), val = tensor<bool, [2]>([true, false])]; | |
| tensor<fp16, [?, ?]> var_785_cast_fp16 = slice_by_index(begin = var_785_begin_0, end = concat_82, end_mask = var_785_end_mask_0, x = var_784_cast_fp16)[name = string("op_785_cast_fp16")]; | |
| tensor<fp16, [1, 6, ?, ?]> qk_21_cast_fp16 = add(x = qk_19_cast_fp16, y = var_785_cast_fp16)[name = string("qk_21_cast_fp16")]; | |
| tensor<fp16, [1, 6, ?, ?]> var_788_cast_fp16 = softmax(axis = var_697, x = qk_21_cast_fp16)[name = string("op_788_cast_fp16")]; | |
| bool var_790_transpose_x_0 = const()[name = string("op_790_transpose_x_0"), val = bool(false)]; | |
| bool var_790_transpose_y_0 = const()[name = string("op_790_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [1, 6, ?, 64]> v_35_cast_fp16 = transpose(perm = var_781, x = var_780_cast_fp16)[name = string("transpose_56")]; | |
| tensor<fp16, [1, 6, ?, 64]> var_790_cast_fp16 = matmul(transpose_x = var_790_transpose_x_0, transpose_y = var_790_transpose_y_0, x = var_788_cast_fp16, y = v_35_cast_fp16)[name = string("op_790_cast_fp16")]; | |
| tensor<int32, [4]> var_791 = const()[name = string("op_791"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> concat_83x = const()[name = string("concat_83x"), val = tensor<int32, [3]>([1, -1, 384])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_792_cast_fp16 = transpose(perm = var_791, x = var_790_cast_fp16)[name = string("transpose_53")]; | |
| tensor<fp16, [1, ?, 384]> x_61_cast_fp16 = reshape(shape = concat_83x, x = var_792_cast_fp16)[name = string("x_61_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_796_to_fp16 = const()[name = string("op_796_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55046400)))]; | |
| tensor<fp16, [384]> var_797_to_fp16 = const()[name = string("op_797_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55341376)))]; | |
| tensor<fp16, [1, ?, 384]> linear_27_cast_fp16 = linear(bias = var_797_to_fp16, weight = var_796_to_fp16, x = x_61_cast_fp16)[name = string("linear_27_cast_fp16")]; | |
| tensor<fp16, [1, ?, 384]> x_63_cast_fp16 = add(x = x_57_cast_fp16, y = linear_27_cast_fp16)[name = string("x_63_cast_fp16")]; | |
| tensor<int32, [1]> var_804_axes_0 = const()[name = string("op_804_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [384]> blocks_3_cross_attn_ln_weight_to_fp16 = const()[name = string("blocks_3_cross_attn_ln_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55342208)))]; | |
| tensor<fp16, [384]> blocks_3_cross_attn_ln_bias_to_fp16 = const()[name = string("blocks_3_cross_attn_ln_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55343040)))]; | |
| tensor<fp16, [1, ?, 384]> var_804_cast_fp16 = layer_norm(axes = var_804_axes_0, beta = blocks_3_cross_attn_ln_bias_to_fp16, epsilon = var_703_to_fp16, gamma = blocks_3_cross_attn_ln_weight_to_fp16, x = x_63_cast_fp16)[name = string("op_804_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_813_to_fp16 = const()[name = string("op_813_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55343872)))]; | |
| tensor<fp16, [384]> var_814_to_fp16 = const()[name = string("op_814_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55638848)))]; | |
| tensor<fp16, [1, ?, 384]> linear_28_cast_fp16 = linear(bias = var_814_to_fp16, weight = var_813_to_fp16, x = var_804_cast_fp16)[name = string("linear_28_cast_fp16")]; | |
| tensor<int32, [3]> concat_84 = const()[name = string("concat_84"), val = tensor<int32, [3]>([0, 0, 0])]; | |
| tensor<int32, [3]> concat_85 = const()[name = string("concat_85"), val = tensor<int32, [3]>([0, 1500, 0])]; | |
| tensor<int32, [3]> k_37_internal_tensor_assign_1_stride_0 = const()[name = string("k_37_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])]; | |
| tensor<bool, [3]> k_37_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_37_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])]; | |
| tensor<bool, [3]> k_37_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_37_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])]; | |
| tensor<bool, [3]> k_37_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_37_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])]; | |
| tensor<fp16, [1, 1500, 384]> k_37_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_84, begin_mask = k_37_internal_tensor_assign_1_begin_mask_0, end = concat_85, end_mask = k_37_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_37_internal_tensor_assign_1_squeeze_mask_0, stride = k_37_internal_tensor_assign_1_stride_0, update = k_cache_cast_fp16, x = k_7_to_fp16)[name = string("k_37_internal_tensor_assign_1_cast_fp16")]; | |
| tensor<int32, [3]> concat_86 = const()[name = string("concat_86"), val = tensor<int32, [3]>([0, 0, 0])]; | |
| tensor<int32, [3]> concat_87 = const()[name = string("concat_87"), val = tensor<int32, [3]>([0, 1500, 0])]; | |
| tensor<int32, [3]> v_37_internal_tensor_assign_1_stride_0 = const()[name = string("v_37_internal_tensor_assign_1_stride_0"), val = tensor<int32, [3]>([1, 1, 1])]; | |
| tensor<bool, [3]> v_37_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_37_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [3]>([false, false, false])]; | |
| tensor<bool, [3]> v_37_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_37_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [3]>([true, false, true])]; | |
| tensor<bool, [3]> v_37_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_37_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, false])]; | |
| tensor<fp16, [1, 1500, 384]> v_37_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_86, begin_mask = v_37_internal_tensor_assign_1_begin_mask_0, end = concat_87, end_mask = v_37_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_37_internal_tensor_assign_1_squeeze_mask_0, stride = v_37_internal_tensor_assign_1_stride_0, update = v_cache_cast_fp16, x = k_7_to_fp16)[name = string("v_37_internal_tensor_assign_1_cast_fp16")]; | |
| tensor<int32, [4]> concat_88x = const()[name = string("concat_88x"), val = tensor<int32, [4]>([1, -1, 6, 64])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_834_cast_fp16 = reshape(shape = concat_88x, x = linear_28_cast_fp16)[name = string("op_834_cast_fp16")]; | |
| tensor<fp16, [1, 1, 1, 1]> const_34_to_fp16 = const()[name = string("const_34_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])]; | |
| tensor<fp16, [1, ?, 6, 64]> q_cast_fp16 = mul(x = var_834_cast_fp16, y = const_34_to_fp16)[name = string("q_cast_fp16")]; | |
| tensor<int32, [4]> var_840 = const()[name = string("op_840"), val = tensor<int32, [4]>([1, 1500, 6, -1])]; | |
| tensor<fp16, [1, 1500, 6, 64]> var_841_cast_fp16 = reshape(shape = var_840, x = k_37_internal_tensor_assign_1_cast_fp16)[name = string("op_841_cast_fp16")]; | |
| tensor<fp16, [1, 1, 1, 1]> const_35_to_fp16 = const()[name = string("const_35_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.6ap-2]]]])]; | |
| tensor<fp16, [1, 1500, 6, 64]> k_cast_fp16 = mul(x = var_841_cast_fp16, y = const_35_to_fp16)[name = string("k_cast_fp16")]; | |
| tensor<int32, [4]> var_847 = const()[name = string("op_847"), val = tensor<int32, [4]>([1, 1500, 6, -1])]; | |
| tensor<fp16, [1, 1500, 6, 64]> var_848_cast_fp16 = reshape(shape = var_847, x = v_37_internal_tensor_assign_1_cast_fp16)[name = string("op_848_cast_fp16")]; | |
| tensor<int32, [4]> var_849 = const()[name = string("op_849"), 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_47_perm_0 = const()[name = string("transpose_47_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; | |
| tensor<int32, [4]> transpose_48_perm_0 = const()[name = string("transpose_48_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; | |
| tensor<fp16, [1, 6, 64, 1500]> transpose_48 = transpose(perm = transpose_48_perm_0, x = k_cast_fp16)[name = string("transpose_50")]; | |
| tensor<fp16, [1, 6, ?, 64]> transpose_47 = transpose(perm = transpose_47_perm_0, x = q_cast_fp16)[name = string("transpose_51")]; | |
| tensor<fp16, [1, 6, ?, 1500]> qk_cast_fp16 = matmul(transpose_x = qk_transpose_x_0, transpose_y = qk_transpose_y_0, x = transpose_47, y = transpose_48)[name = string("qk_cast_fp16")]; | |
| tensor<fp16, [1, 6, ?, 1500]> var_853_cast_fp16 = softmax(axis = var_697, x = qk_cast_fp16)[name = string("op_853_cast_fp16")]; | |
| bool var_855_transpose_x_0 = const()[name = string("op_855_transpose_x_0"), val = bool(false)]; | |
| bool var_855_transpose_y_0 = const()[name = string("op_855_transpose_y_0"), val = bool(false)]; | |
| tensor<fp16, [1, 6, 1500, 64]> v_cast_fp16 = transpose(perm = var_849, x = var_848_cast_fp16)[name = string("transpose_52")]; | |
| tensor<fp16, [1, 6, ?, 64]> var_855_cast_fp16 = matmul(transpose_x = var_855_transpose_x_0, transpose_y = var_855_transpose_y_0, x = var_853_cast_fp16, y = v_cast_fp16)[name = string("op_855_cast_fp16")]; | |
| tensor<int32, [4]> var_856 = const()[name = string("op_856"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> concat_89x = const()[name = string("concat_89x"), val = tensor<int32, [3]>([1, -1, 384])]; | |
| tensor<fp16, [1, ?, 6, 64]> var_857_cast_fp16 = transpose(perm = var_856, x = var_855_cast_fp16)[name = string("transpose_49")]; | |
| tensor<fp16, [1, ?, 384]> x_67_cast_fp16 = reshape(shape = concat_89x, x = var_857_cast_fp16)[name = string("x_67_cast_fp16")]; | |
| tensor<fp16, [384, 384]> var_861_to_fp16 = const()[name = string("op_861_to_fp16"), val = tensor<fp16, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55639680)))]; | |
| tensor<fp16, [384]> var_862_to_fp16 = const()[name = string("op_862_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55934656)))]; | |
| tensor<fp16, [1, ?, 384]> linear_29_cast_fp16 = linear(bias = var_862_to_fp16, weight = var_861_to_fp16, x = x_67_cast_fp16)[name = string("linear_29_cast_fp16")]; | |
| tensor<fp16, [1, ?, 384]> x_69_cast_fp16 = add(x = x_63_cast_fp16, y = linear_29_cast_fp16)[name = string("x_69_cast_fp16")]; | |
| tensor<int32, [1]> var_869_axes_0 = const()[name = string("op_869_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [384]> blocks_3_mlp_ln_weight_to_fp16 = const()[name = string("blocks_3_mlp_ln_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55935488)))]; | |
| tensor<fp16, [384]> blocks_3_mlp_ln_bias_to_fp16 = const()[name = string("blocks_3_mlp_ln_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55936320)))]; | |
| tensor<fp16, [1, ?, 384]> var_869_cast_fp16 = layer_norm(axes = var_869_axes_0, beta = blocks_3_mlp_ln_bias_to_fp16, epsilon = var_703_to_fp16, gamma = blocks_3_mlp_ln_weight_to_fp16, x = x_69_cast_fp16)[name = string("op_869_cast_fp16")]; | |
| tensor<fp16, [1536, 384]> var_878_to_fp16 = const()[name = string("op_878_to_fp16"), val = tensor<fp16, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55937152)))]; | |
| tensor<fp16, [1536]> var_879_to_fp16 = const()[name = string("op_879_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57116864)))]; | |
| tensor<fp16, [1, ?, 1536]> linear_30_cast_fp16 = linear(bias = var_879_to_fp16, weight = var_878_to_fp16, x = var_869_cast_fp16)[name = string("linear_30_cast_fp16")]; | |
| string x_73_mode_0 = const()[name = string("x_73_mode_0"), val = string("EXACT")]; | |
| tensor<fp16, [1, ?, 1536]> x_73_cast_fp16 = gelu(mode = x_73_mode_0, x = linear_30_cast_fp16)[name = string("x_73_cast_fp16")]; | |
| tensor<fp16, [384, 1536]> var_884_to_fp16 = const()[name = string("op_884_to_fp16"), val = tensor<fp16, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57120000)))]; | |
| tensor<fp16, [384]> var_885_to_fp16 = const()[name = string("op_885_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58299712)))]; | |
| tensor<fp16, [1, ?, 384]> linear_31_cast_fp16 = linear(bias = var_885_to_fp16, weight = var_884_to_fp16, x = x_73_cast_fp16)[name = string("linear_31_cast_fp16")]; | |
| tensor<fp16, [1, ?, 384]> x_75_cast_fp16 = add(x = x_69_cast_fp16, y = linear_31_cast_fp16)[name = string("x_75_cast_fp16")]; | |
| tensor<int32, [1]> var_898_axes_0 = const()[name = string("op_898_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [384]> ln_weight_to_fp16 = const()[name = string("ln_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58300544)))]; | |
| tensor<fp16, [384]> ln_bias_to_fp16 = const()[name = string("ln_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58301376)))]; | |
| fp16 var_889_to_fp16 = const()[name = string("op_889_to_fp16"), val = fp16(0x1.5p-17)]; | |
| tensor<fp16, [1, ?, 384]> var_898_cast_fp16 = layer_norm(axes = var_898_axes_0, beta = ln_bias_to_fp16, epsilon = var_889_to_fp16, gamma = ln_weight_to_fp16, x = x_75_cast_fp16)[name = string("op_898_cast_fp16")]; | |
| tensor<fp16, [51865]> var_908_bias_0_to_fp16 = const()[name = string("op_908_bias_0_to_fp16"), val = tensor<fp16, [51865]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58302208)))]; | |
| tensor<fp16, [1, ?, 51865]> logits = linear(bias = var_908_bias_0_to_fp16, weight = token_embedding_weight_to_fp16, x = var_898_cast_fp16)[name = string("op_908_cast_fp16")]; | |
| } -> (logits); | |
| } |