program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3404.16.1"}, {"coremlc-version", "3404.23.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.2"}})] { func main(tensor input_ids_1) { tensor cast_2_dtype_0 = const()[name = tensor("cast_2_dtype_0"), val = tensor("int32")]; tensor inputs_embeds_axis_0 = const()[name = tensor("inputs_embeds_axis_0"), val = tensor(0)]; tensor inputs_embeds_batch_dims_0 = const()[name = tensor("inputs_embeds_batch_dims_0"), val = tensor(0)]; tensor inputs_embeds_validate_indices_0 = const()[name = tensor("inputs_embeds_validate_indices_0"), val = tensor(false)]; tensor model_transformer_wte_weight_to_fp16 = const()[name = tensor("model_transformer_wte_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; tensor input_ids_1_to_int32 = cast(dtype = cast_2_dtype_0, x = input_ids_1)[name = tensor("cast_32")]; tensor inputs_embeds_cast_fp16 = gather(axis = inputs_embeds_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = input_ids_1_to_int32, validate_indices = inputs_embeds_validate_indices_0, x = model_transformer_wte_weight_to_fp16)[name = tensor("inputs_embeds_cast_fp16")]; tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201216)))]; tensor input_3_cast_fp16 = add(x = inputs_embeds_cast_fp16, y = const_3_to_fp16)[name = tensor("input_3_cast_fp16")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; tensor model_transformer_h_0_ln_1_weight_to_fp16 = const()[name = tensor("model_transformer_h_0_ln_1_weight_to_fp16"), val = tensor([0x1p+0, 0x1p+0])]; tensor model_transformer_h_0_ln_1_bias_to_fp16 = const()[name = tensor("model_transformer_h_0_ln_1_bias_to_fp16"), val = tensor([0x0p+0, 0x0p+0])]; tensor var_19_to_fp16 = const()[name = tensor("op_19_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_1_cast_fp16 = layer_norm(axes = x_1_axes_0, beta = model_transformer_h_0_ln_1_bias_to_fp16, epsilon = var_19_to_fp16, gamma = model_transformer_h_0_ln_1_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("x_1_cast_fp16")]; tensor var_88 = const()[name = tensor("op_88"), val = tensor([-1, 2])]; tensor var_89_cast_fp16 = reshape(shape = var_88, x = x_1_cast_fp16)[name = tensor("op_89_cast_fp16")]; tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(false)]; tensor model_transformer_h_0_attn_c_attn_weight_to_fp16 = const()[name = tensor("model_transformer_h_0_attn_c_attn_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201344)))]; tensor matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = var_89_cast_fp16, y = model_transformer_h_0_attn_c_attn_weight_to_fp16)[name = tensor("matmul_0_cast_fp16")]; tensor var_91 = const()[name = tensor("op_91"), val = tensor([1, 5, 6])]; tensor var_92_cast_fp16 = reshape(shape = var_91, x = matmul_0_cast_fp16)[name = tensor("op_92_cast_fp16")]; tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([2, 2, 2])]; tensor var_93_axis_0 = const()[name = tensor("op_93_axis_0"), val = tensor(2)]; tensor var_93_cast_fp16_0, tensor var_93_cast_fp16_1, tensor var_93_cast_fp16_2 = split(axis = var_93_axis_0, split_sizes = tile_0, x = var_92_cast_fp16)[name = tensor("op_93_cast_fp16")]; tensor var_101 = const()[name = tensor("op_101"), val = tensor([1, 5, -1, 1])]; tensor var_102_cast_fp16 = reshape(shape = var_101, x = var_93_cast_fp16_0)[name = tensor("op_102_cast_fp16")]; tensor var_104 = const()[name = tensor("op_104"), val = tensor([1, 5, -1, 1])]; tensor var_105_cast_fp16 = reshape(shape = var_104, x = var_93_cast_fp16_1)[name = tensor("op_105_cast_fp16")]; tensor var_107 = const()[name = tensor("op_107"), val = tensor([1, 5, -1, 1])]; tensor var_108_cast_fp16 = reshape(shape = var_107, x = var_93_cast_fp16_2)[name = tensor("op_108_cast_fp16")]; tensor value_1_perm_0 = const()[name = tensor("value_1_perm_0"), val = tensor([0, 2, 1, 3])]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; tensor transpose_8_perm_0 = const()[name = tensor("transpose_8_perm_0"), val = tensor([0, 2, -3, -1])]; tensor transpose_9_perm_0 = const()[name = tensor("transpose_9_perm_0"), val = tensor([0, 2, -3, -1])]; tensor transpose_9 = transpose(perm = transpose_9_perm_0, x = var_105_cast_fp16)[name = tensor("transpose_17")]; tensor transpose_8 = transpose(perm = transpose_8_perm_0, x = var_102_cast_fp16)[name = tensor("transpose_18")]; tensor matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = transpose_8, y = transpose_9)[name = tensor("matmul_1_cast_fp16")]; tensor var_64_to_fp16 = const()[name = tensor("op_64_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201472)))]; tensor add_0_cast_fp16 = add(x = matmul_1_cast_fp16, y = var_64_to_fp16)[name = tensor("add_0_cast_fp16")]; tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; tensor softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = add_0_cast_fp16)[name = tensor("softmax_0_cast_fp16")]; tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor value_1_cast_fp16 = transpose(perm = value_1_perm_0, x = var_108_cast_fp16)[name = tensor("transpose_19")]; tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0_cast_fp16, y = value_1_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; tensor var_119_perm_0 = const()[name = tensor("op_119_perm_0"), val = tensor([0, 2, 1, 3])]; tensor var_131 = const()[name = tensor("op_131"), val = tensor([-1, 2])]; tensor var_119_cast_fp16 = transpose(perm = var_119_perm_0, x = attn_output_1_cast_fp16)[name = tensor("transpose_16")]; tensor var_132_cast_fp16 = reshape(shape = var_131, x = var_119_cast_fp16)[name = tensor("op_132_cast_fp16")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; tensor model_transformer_h_0_attn_c_proj_weight_to_fp16 = const()[name = tensor("model_transformer_h_0_attn_c_proj_weight_to_fp16"), val = tensor([[-0x1.54cp-10, -0x1.3d8p-10], [-0x1.d4p-5, -0x1.8p-6]])]; tensor matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = var_132_cast_fp16, y = model_transformer_h_0_attn_c_proj_weight_to_fp16)[name = tensor("matmul_2_cast_fp16")]; tensor var_134 = const()[name = tensor("op_134"), val = tensor([1, 5, 2])]; tensor input_5_cast_fp16 = reshape(shape = var_134, x = matmul_2_cast_fp16)[name = tensor("input_5_cast_fp16")]; tensor input_7_cast_fp16 = add(x = input_5_cast_fp16, y = input_3_cast_fp16)[name = tensor("input_7_cast_fp16")]; tensor x_9_axes_0 = const()[name = tensor("x_9_axes_0"), val = tensor([-1])]; tensor model_transformer_h_0_ln_2_weight_to_fp16 = const()[name = tensor("model_transformer_h_0_ln_2_weight_to_fp16"), val = tensor([0x1p+0, 0x1p+0])]; tensor model_transformer_h_0_ln_2_bias_to_fp16 = const()[name = tensor("model_transformer_h_0_ln_2_bias_to_fp16"), val = tensor([0x0p+0, 0x0p+0])]; tensor x_9_cast_fp16 = layer_norm(axes = x_9_axes_0, beta = model_transformer_h_0_ln_2_bias_to_fp16, epsilon = var_19_to_fp16, gamma = model_transformer_h_0_ln_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("x_9_cast_fp16")]; tensor var_149 = const()[name = tensor("op_149"), val = tensor([-1, 2])]; tensor var_150_cast_fp16 = reshape(shape = var_149, x = x_9_cast_fp16)[name = tensor("op_150_cast_fp16")]; tensor matmul_3_transpose_x_0 = const()[name = tensor("matmul_3_transpose_x_0"), val = tensor(false)]; tensor matmul_3_transpose_y_0 = const()[name = tensor("matmul_3_transpose_y_0"), val = tensor(false)]; tensor model_transformer_h_0_mlp_c_fc_weight_to_fp16 = const()[name = tensor("model_transformer_h_0_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201600)))]; tensor matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = var_150_cast_fp16, y = model_transformer_h_0_mlp_c_fc_weight_to_fp16)[name = tensor("matmul_3_cast_fp16")]; tensor var_152 = const()[name = tensor("op_152"), val = tensor([1, 5, 8])]; tensor input_9_cast_fp16 = reshape(shape = var_152, x = matmul_3_cast_fp16)[name = tensor("input_9_cast_fp16")]; tensor x_13_mode_0 = const()[name = tensor("x_13_mode_0"), val = tensor("TANH_APPROXIMATION")]; tensor x_13_cast_fp16 = gelu(mode = x_13_mode_0, x = input_9_cast_fp16)[name = tensor("x_13_cast_fp16")]; tensor var_171 = const()[name = tensor("op_171"), val = tensor([-1, 8])]; tensor var_172_cast_fp16 = reshape(shape = var_171, x = x_13_cast_fp16)[name = tensor("op_172_cast_fp16")]; tensor matmul_4_transpose_x_0 = const()[name = tensor("matmul_4_transpose_x_0"), val = tensor(false)]; tensor matmul_4_transpose_y_0 = const()[name = tensor("matmul_4_transpose_y_0"), val = tensor(false)]; tensor model_transformer_h_0_mlp_c_proj_weight_to_fp16 = const()[name = tensor("model_transformer_h_0_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201728)))]; tensor matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = var_172_cast_fp16, y = model_transformer_h_0_mlp_c_proj_weight_to_fp16)[name = tensor("matmul_4_cast_fp16")]; tensor var_174 = const()[name = tensor("op_174"), val = tensor([1, 5, 2])]; tensor input_11_cast_fp16 = reshape(shape = var_174, x = matmul_4_cast_fp16)[name = tensor("input_11_cast_fp16")]; tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = input_11_cast_fp16)[name = tensor("input_13_cast_fp16")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; tensor model_transformer_h_1_ln_1_weight_to_fp16 = const()[name = tensor("model_transformer_h_1_ln_1_weight_to_fp16"), val = tensor([0x1p+0, 0x1p+0])]; tensor model_transformer_h_1_ln_1_bias_to_fp16 = const()[name = tensor("model_transformer_h_1_ln_1_bias_to_fp16"), val = tensor([0x0p+0, 0x0p+0])]; tensor x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = model_transformer_h_1_ln_1_bias_to_fp16, epsilon = var_19_to_fp16, gamma = model_transformer_h_1_ln_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("x_17_cast_fp16")]; tensor var_193 = const()[name = tensor("op_193"), val = tensor([-1, 2])]; tensor var_194_cast_fp16 = reshape(shape = var_193, x = x_17_cast_fp16)[name = tensor("op_194_cast_fp16")]; tensor matmul_5_transpose_x_0 = const()[name = tensor("matmul_5_transpose_x_0"), val = tensor(false)]; tensor matmul_5_transpose_y_0 = const()[name = tensor("matmul_5_transpose_y_0"), val = tensor(false)]; tensor model_transformer_h_1_attn_c_attn_weight_to_fp16 = const()[name = tensor("model_transformer_h_1_attn_c_attn_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201856)))]; tensor matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = var_194_cast_fp16, y = model_transformer_h_1_attn_c_attn_weight_to_fp16)[name = tensor("matmul_5_cast_fp16")]; tensor var_196 = const()[name = tensor("op_196"), val = tensor([1, 5, 6])]; tensor var_197_cast_fp16 = reshape(shape = var_196, x = matmul_5_cast_fp16)[name = tensor("op_197_cast_fp16")]; tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([2, 2, 2])]; tensor var_198_axis_0 = const()[name = tensor("op_198_axis_0"), val = tensor(2)]; tensor var_198_cast_fp16_0, tensor var_198_cast_fp16_1, tensor var_198_cast_fp16_2 = split(axis = var_198_axis_0, split_sizes = tile_1, x = var_197_cast_fp16)[name = tensor("op_198_cast_fp16")]; tensor var_206 = const()[name = tensor("op_206"), val = tensor([1, 5, -1, 1])]; tensor var_207_cast_fp16 = reshape(shape = var_206, x = var_198_cast_fp16_0)[name = tensor("op_207_cast_fp16")]; tensor var_209 = const()[name = tensor("op_209"), val = tensor([1, 5, -1, 1])]; tensor var_210_cast_fp16 = reshape(shape = var_209, x = var_198_cast_fp16_1)[name = tensor("op_210_cast_fp16")]; tensor var_212 = const()[name = tensor("op_212"), val = tensor([1, 5, -1, 1])]; tensor var_213_cast_fp16 = reshape(shape = var_212, x = var_198_cast_fp16_2)[name = tensor("op_213_cast_fp16")]; tensor value_5_perm_0 = const()[name = tensor("value_5_perm_0"), val = tensor([0, 2, 1, 3])]; tensor matmul_6_transpose_y_0 = const()[name = tensor("matmul_6_transpose_y_0"), val = tensor(true)]; tensor matmul_6_transpose_x_0 = const()[name = tensor("matmul_6_transpose_x_0"), val = tensor(false)]; tensor transpose_10_perm_0 = const()[name = tensor("transpose_10_perm_0"), val = tensor([0, 2, -3, -1])]; tensor transpose_11_perm_0 = const()[name = tensor("transpose_11_perm_0"), val = tensor([0, 2, -3, -1])]; tensor transpose_11 = transpose(perm = transpose_11_perm_0, x = var_210_cast_fp16)[name = tensor("transpose_13")]; tensor transpose_10 = transpose(perm = transpose_10_perm_0, x = var_207_cast_fp16)[name = tensor("transpose_14")]; tensor matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = transpose_10, y = transpose_11)[name = tensor("matmul_6_cast_fp16")]; tensor add_1_cast_fp16 = add(x = matmul_6_cast_fp16, y = var_64_to_fp16)[name = tensor("add_1_cast_fp16")]; tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; tensor softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = add_1_cast_fp16)[name = tensor("softmax_1_cast_fp16")]; tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; tensor value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = var_213_cast_fp16)[name = tensor("transpose_15")]; tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = softmax_1_cast_fp16, y = value_5_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; tensor var_224_perm_0 = const()[name = tensor("op_224_perm_0"), val = tensor([0, 2, 1, 3])]; tensor var_236 = const()[name = tensor("op_236"), val = tensor([-1, 2])]; tensor var_224_cast_fp16 = transpose(perm = var_224_perm_0, x = attn_output_7_cast_fp16)[name = tensor("transpose_12")]; tensor var_237_cast_fp16 = reshape(shape = var_236, x = var_224_cast_fp16)[name = tensor("op_237_cast_fp16")]; tensor matmul_7_transpose_x_0 = const()[name = tensor("matmul_7_transpose_x_0"), val = tensor(false)]; tensor matmul_7_transpose_y_0 = const()[name = tensor("matmul_7_transpose_y_0"), val = tensor(false)]; tensor model_transformer_h_1_attn_c_proj_weight_to_fp16 = const()[name = tensor("model_transformer_h_1_attn_c_proj_weight_to_fp16"), val = tensor([[0x1.e5cp-8, 0x1.54p-7], [0x1.2d8p-6, -0x1.594p-6]])]; tensor matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = var_237_cast_fp16, y = model_transformer_h_1_attn_c_proj_weight_to_fp16)[name = tensor("matmul_7_cast_fp16")]; tensor var_239 = const()[name = tensor("op_239"), val = tensor([1, 5, 2])]; tensor input_15_cast_fp16 = reshape(shape = var_239, x = matmul_7_cast_fp16)[name = tensor("input_15_cast_fp16")]; tensor input_17_cast_fp16 = add(x = input_15_cast_fp16, y = input_13_cast_fp16)[name = tensor("input_17_cast_fp16")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; tensor model_transformer_h_1_ln_2_weight_to_fp16 = const()[name = tensor("model_transformer_h_1_ln_2_weight_to_fp16"), val = tensor([0x1p+0, 0x1p+0])]; tensor model_transformer_h_1_ln_2_bias_to_fp16 = const()[name = tensor("model_transformer_h_1_ln_2_bias_to_fp16"), val = tensor([0x0p+0, 0x0p+0])]; tensor x_25_cast_fp16 = layer_norm(axes = x_25_axes_0, beta = model_transformer_h_1_ln_2_bias_to_fp16, epsilon = var_19_to_fp16, gamma = model_transformer_h_1_ln_2_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("x_25_cast_fp16")]; tensor var_254 = const()[name = tensor("op_254"), val = tensor([-1, 2])]; tensor var_255_cast_fp16 = reshape(shape = var_254, x = x_25_cast_fp16)[name = tensor("op_255_cast_fp16")]; tensor matmul_8_transpose_x_0 = const()[name = tensor("matmul_8_transpose_x_0"), val = tensor(false)]; tensor matmul_8_transpose_y_0 = const()[name = tensor("matmul_8_transpose_y_0"), val = tensor(false)]; tensor model_transformer_h_1_mlp_c_fc_weight_to_fp16 = const()[name = tensor("model_transformer_h_1_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201984)))]; tensor matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = var_255_cast_fp16, y = model_transformer_h_1_mlp_c_fc_weight_to_fp16)[name = tensor("matmul_8_cast_fp16")]; tensor var_257 = const()[name = tensor("op_257"), val = tensor([1, 5, 8])]; tensor input_19_cast_fp16 = reshape(shape = var_257, x = matmul_8_cast_fp16)[name = tensor("input_19_cast_fp16")]; tensor x_29_mode_0 = const()[name = tensor("x_29_mode_0"), val = tensor("TANH_APPROXIMATION")]; tensor x_29_cast_fp16 = gelu(mode = x_29_mode_0, x = input_19_cast_fp16)[name = tensor("x_29_cast_fp16")]; tensor var_276 = const()[name = tensor("op_276"), val = tensor([-1, 8])]; tensor var_277_cast_fp16 = reshape(shape = var_276, x = x_29_cast_fp16)[name = tensor("op_277_cast_fp16")]; tensor matmul_9_transpose_x_0 = const()[name = tensor("matmul_9_transpose_x_0"), val = tensor