diff --git "a/coreml/coreml/g2p.mlmodelc/model.mil" "b/coreml/coreml/g2p.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/coreml/coreml/g2p.mlmodelc/model.mil" @@ -0,0 +1,2100 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3600.16.1"}, {"coremlc-version", "3600.25.2"}})] +{ + func g2p(tensor text) { + int32 var_24 = const()[name = string("op_24"), val = int32(-1)]; + int32 var_26 = const()[name = string("op_26"), val = int32(0)]; + tensor src_pad_mask = equal(x = text, y = var_26)[name = string("src_pad_mask")]; + int32 x_1_batch_dims_0 = const()[name = string("x_1_batch_dims_0"), val = int32(0)]; + bool x_1_validate_indices_0 = const()[name = string("x_1_validate_indices_0"), val = bool(false)]; + tensor module_embedding_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257408))))[name = string("module_embedding_weight_to_fp16_quantized")]; + string text_to_int16_dtype_0 = const()[name = string("text_to_int16_dtype_0"), val = string("int16")]; + string cast_49_dtype_0 = const()[name = string("cast_49_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor text_to_int16 = cast(dtype = text_to_int16_dtype_0, x = text)[name = string("cast_6")]; + tensor cast_49 = cast(dtype = cast_49_dtype_0, x = text_to_int16)[name = string("cast_5")]; + tensor greater_equal_0 = greater_equal(x = cast_49, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(402)]; + tensor add_0 = add(x = cast_49, y = slice_by_index_0)[name = string("add_0")]; + tensor select_0 = select(a = cast_49, b = add_0, cond = greater_equal_0)[name = string("select_0")]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + string cast_0_dtype_0 = const()[name = string("cast_0_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_4")]; + tensor cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = string("cast_3")]; + tensor greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")]; + int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(402)]; + tensor add_0_1 = add(x = cast_0, y = slice_by_index_0_1)[name = string("add_0_1")]; + tensor select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")]; + int32 x_1_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = string("x_1_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = int32(0)]; + tensor x_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = x_1_cast_fp16_cast_uint16_cast_uint16_axis_0, batch_dims = x_1_batch_dims_0, indices = select_0_1, validate_indices = x_1_validate_indices_0, x = module_embedding_weight_to_fp16_quantized)[name = string("x_1_cast_fp16_cast_uint16_cast_uint16")]; + tensor op_286_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27017920))), scale = tensor([[[0x1.724p-7]]]))[name = string("op_286_to_fp16_quantized")]; + tensor input_1_cast_fp16 = add(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = op_286_to_fp16_quantized)[name = string("input_1_cast_fp16")]; + tensor module_layers_0_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1528192))))[name = string("module_layers_0_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_0_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_0_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1532096)))]; + tensor linear_0_cast_fp16 = linear(bias = module_layers_0_self_attn_in_proj_bias_to_fp16, weight = module_layers_0_self_attn_in_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor var_292_split_sizes_0 = const()[name = string("op_292_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_292_axis_0 = const()[name = string("op_292_axis_0"), val = int32(-1)]; + tensor var_292_cast_fp16_0, tensor var_292_cast_fp16_1, tensor var_292_cast_fp16_2 = split(axis = var_292_axis_0, split_sizes = var_292_split_sizes_0, x = linear_0_cast_fp16)[name = string("op_292_cast_fp16")]; + tensor var_295 = const()[name = string("op_295"), val = tensor([1, 512, 4, 160])]; + tensor var_296_cast_fp16 = reshape(shape = var_295, x = var_292_cast_fp16_0)[name = string("op_296_cast_fp16")]; + tensor var_298 = const()[name = string("op_298"), val = tensor([1, 512, 4, 160])]; + tensor var_299_cast_fp16 = reshape(shape = var_298, x = var_292_cast_fp16_1)[name = string("op_299_cast_fp16")]; + tensor var_301 = const()[name = string("op_301"), val = tensor([1, 512, 4, 160])]; + tensor var_302_cast_fp16 = reshape(shape = var_301, x = var_292_cast_fp16_2)[name = string("op_302_cast_fp16")]; + tensor v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_305_transpose_x_0 = const()[name = string("op_305_transpose_x_0"), val = bool(false)]; + bool var_305_transpose_y_0 = const()[name = string("op_305_transpose_y_0"), val = bool(false)]; + tensor transpose_32_perm_0 = const()[name = string("transpose_32_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_33 = transpose(perm = transpose_33_perm_0, x = var_299_cast_fp16)[name = string("transpose_78")]; + tensor transpose_32 = transpose(perm = transpose_32_perm_0, x = var_296_cast_fp16)[name = string("transpose_79")]; + tensor var_305_cast_fp16 = matmul(transpose_x = var_305_transpose_x_0, transpose_y = var_305_transpose_y_0, x = transpose_32, y = transpose_33)[name = string("op_305_cast_fp16")]; + fp16 var_306_to_fp16 = const()[name = string("op_306_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_1_cast_fp16 = mul(x = var_305_cast_fp16, y = var_306_to_fp16)[name = string("scores_1_cast_fp16")]; + tensor weights_1_cast_fp16 = softmax(axis = var_24, x = scores_1_cast_fp16)[name = string("weights_1_cast_fp16")]; + fp16 var_13_to_fp16 = const()[name = string("op_13_to_fp16"), val = fp16(0x1p+0)]; + string var_309_to_fp16_dtype_0 = const()[name = string("op_309_to_fp16_dtype_0"), val = string("fp16")]; + tensor src_pad_mask_to_fp16 = cast(dtype = var_309_to_fp16_dtype_0, x = src_pad_mask)[name = string("cast_2")]; + tensor var_310_cast_fp16 = sub(x = var_13_to_fp16, y = src_pad_mask_to_fp16)[name = string("op_310_cast_fp16")]; + tensor var_312_axes_0 = const()[name = string("op_312_axes_0"), val = tensor([1])]; + tensor var_312_cast_fp16 = expand_dims(axes = var_312_axes_0, x = var_310_cast_fp16)[name = string("op_312_cast_fp16")]; + tensor var_313_axes_0 = const()[name = string("op_313_axes_0"), val = tensor([2])]; + tensor var_313_cast_fp16 = expand_dims(axes = var_313_axes_0, x = var_312_cast_fp16)[name = string("op_313_cast_fp16")]; + tensor weights_3_cast_fp16 = mul(x = weights_1_cast_fp16, y = var_313_cast_fp16)[name = string("weights_3_cast_fp16")]; + tensor var_317_axes_0 = const()[name = string("op_317_axes_0"), val = tensor([-1])]; + bool var_317_keep_dims_0 = const()[name = string("op_317_keep_dims_0"), val = bool(true)]; + tensor var_317_cast_fp16 = reduce_sum(axes = var_317_axes_0, keep_dims = var_317_keep_dims_0, x = weights_3_cast_fp16)[name = string("op_317_cast_fp16")]; + fp16 var_11_to_fp16 = const()[name = string("op_11_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_318_cast_fp16 = maximum(x = var_317_cast_fp16, y = var_11_to_fp16)[name = string("op_318_cast_fp16")]; + tensor weights_5_cast_fp16 = real_div(x = weights_3_cast_fp16, y = var_318_cast_fp16)[name = string("weights_5_cast_fp16")]; + bool context_1_transpose_x_0 = const()[name = string("context_1_transpose_x_0"), val = bool(false)]; + bool context_1_transpose_y_0 = const()[name = string("context_1_transpose_y_0"), val = bool(false)]; + tensor v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_302_cast_fp16)[name = string("transpose_77")]; + tensor context_1_cast_fp16 = matmul(transpose_x = context_1_transpose_x_0, transpose_y = context_1_transpose_y_0, x = weights_5_cast_fp16, y = v_3_cast_fp16)[name = string("context_1_cast_fp16")]; + tensor var_321_perm_0 = const()[name = string("op_321_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_322 = const()[name = string("op_322"), val = tensor([1, 512, 640])]; + tensor var_321_cast_fp16 = transpose(perm = var_321_perm_0, x = context_1_cast_fp16)[name = string("transpose_76")]; + tensor context_3_cast_fp16 = reshape(shape = var_322, x = var_321_cast_fp16)[name = string("context_3_cast_fp16")]; + tensor module_layers_0_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1536000))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1945664))))[name = string("module_layers_0_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_0_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_0_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1947008)))]; + tensor linear_1_cast_fp16 = linear(bias = module_layers_0_self_attn_out_proj_bias_to_fp16, weight = module_layers_0_self_attn_out_proj_weight_to_fp16_quantized, x = context_3_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor input_3_cast_fp16 = add(x = input_1_cast_fp16, y = linear_1_cast_fp16)[name = string("input_3_cast_fp16")]; + tensor input_5_axes_0 = const()[name = string("input_5_axes_0"), val = tensor([-1])]; + tensor module_layers_0_norm1_weight_to_fp16 = const()[name = string("module_layers_0_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1948352)))]; + tensor module_layers_0_norm1_bias_to_fp16 = const()[name = string("module_layers_0_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1949696)))]; + fp16 var_9_to_fp16 = const()[name = string("op_9_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = module_layers_0_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_0_norm1_weight_to_fp16, x = input_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor module_layers_0_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1951040))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2770304))))[name = string("module_layers_0_linear1_weight_to_fp16_quantized")]; + tensor module_layers_0_linear1_bias_to_fp16 = const()[name = string("module_layers_0_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2772928)))]; + tensor linear_2_cast_fp16 = linear(bias = module_layers_0_linear1_bias_to_fp16, weight = module_layers_0_linear1_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_2_cast_fp16")]; + tensor input_9_cast_fp16 = relu(x = linear_2_cast_fp16)[name = string("input_9_cast_fp16")]; + tensor module_layers_0_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2775552))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3594816))))[name = string("module_layers_0_linear2_weight_to_fp16_quantized")]; + tensor module_layers_0_linear2_bias_to_fp16 = const()[name = string("module_layers_0_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3596160)))]; + tensor linear_3_cast_fp16 = linear(bias = module_layers_0_linear2_bias_to_fp16, weight = module_layers_0_linear2_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("linear_3_cast_fp16")]; + tensor input_11_cast_fp16 = add(x = input_5_cast_fp16, y = linear_3_cast_fp16)[name = string("input_11_cast_fp16")]; + tensor x_5_axes_0 = const()[name = string("x_5_axes_0"), val = tensor([-1])]; + tensor module_layers_0_norm2_weight_to_fp16 = const()[name = string("module_layers_0_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3597504)))]; + tensor module_layers_0_norm2_bias_to_fp16 = const()[name = string("module_layers_0_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3598848)))]; + tensor x_5_cast_fp16 = layer_norm(axes = x_5_axes_0, beta = module_layers_0_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_0_norm2_weight_to_fp16, x = input_11_cast_fp16)[name = string("x_5_cast_fp16")]; + tensor module_layers_1_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3600192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4829056))))[name = string("module_layers_1_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_1_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_1_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4832960)))]; + tensor linear_4_cast_fp16 = linear(bias = module_layers_1_self_attn_in_proj_bias_to_fp16, weight = module_layers_1_self_attn_in_proj_weight_to_fp16_quantized, x = x_5_cast_fp16)[name = string("linear_4_cast_fp16")]; + tensor var_345_split_sizes_0 = const()[name = string("op_345_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_345_axis_0 = const()[name = string("op_345_axis_0"), val = int32(-1)]; + tensor var_345_cast_fp16_0, tensor var_345_cast_fp16_1, tensor var_345_cast_fp16_2 = split(axis = var_345_axis_0, split_sizes = var_345_split_sizes_0, x = linear_4_cast_fp16)[name = string("op_345_cast_fp16")]; + tensor var_348 = const()[name = string("op_348"), val = tensor([1, 512, 4, 160])]; + tensor var_349_cast_fp16 = reshape(shape = var_348, x = var_345_cast_fp16_0)[name = string("op_349_cast_fp16")]; + tensor var_351 = const()[name = string("op_351"), val = tensor([1, 512, 4, 160])]; + tensor var_352_cast_fp16 = reshape(shape = var_351, x = var_345_cast_fp16_1)[name = string("op_352_cast_fp16")]; + tensor var_354 = const()[name = string("op_354"), val = tensor([1, 512, 4, 160])]; + tensor var_355_cast_fp16 = reshape(shape = var_354, x = var_345_cast_fp16_2)[name = string("op_355_cast_fp16")]; + tensor v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_358_transpose_x_0 = const()[name = string("op_358_transpose_x_0"), val = bool(false)]; + bool var_358_transpose_y_0 = const()[name = string("op_358_transpose_y_0"), val = bool(false)]; + tensor transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_35_perm_0 = const()[name = string("transpose_35_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_35 = transpose(perm = transpose_35_perm_0, x = var_352_cast_fp16)[name = string("transpose_74")]; + tensor transpose_34 = transpose(perm = transpose_34_perm_0, x = var_349_cast_fp16)[name = string("transpose_75")]; + tensor var_358_cast_fp16 = matmul(transpose_x = var_358_transpose_x_0, transpose_y = var_358_transpose_y_0, x = transpose_34, y = transpose_35)[name = string("op_358_cast_fp16")]; + fp16 var_359_to_fp16 = const()[name = string("op_359_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_3_cast_fp16 = mul(x = var_358_cast_fp16, y = var_359_to_fp16)[name = string("scores_3_cast_fp16")]; + tensor weights_7_cast_fp16 = softmax(axis = var_24, x = scores_3_cast_fp16)[name = string("weights_7_cast_fp16")]; + tensor weights_9_cast_fp16 = mul(x = weights_7_cast_fp16, y = var_313_cast_fp16)[name = string("weights_9_cast_fp16")]; + tensor var_370_axes_0 = const()[name = string("op_370_axes_0"), val = tensor([-1])]; + bool var_370_keep_dims_0 = const()[name = string("op_370_keep_dims_0"), val = bool(true)]; + tensor var_370_cast_fp16 = reduce_sum(axes = var_370_axes_0, keep_dims = var_370_keep_dims_0, x = weights_9_cast_fp16)[name = string("op_370_cast_fp16")]; + tensor var_371_cast_fp16 = maximum(x = var_370_cast_fp16, y = var_11_to_fp16)[name = string("op_371_cast_fp16")]; + tensor weights_11_cast_fp16 = real_div(x = weights_9_cast_fp16, y = var_371_cast_fp16)[name = string("weights_11_cast_fp16")]; + bool context_5_transpose_x_0 = const()[name = string("context_5_transpose_x_0"), val = bool(false)]; + bool context_5_transpose_y_0 = const()[name = string("context_5_transpose_y_0"), val = bool(false)]; + tensor v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = var_355_cast_fp16)[name = string("transpose_73")]; + tensor context_5_cast_fp16 = matmul(transpose_x = context_5_transpose_x_0, transpose_y = context_5_transpose_y_0, x = weights_11_cast_fp16, y = v_7_cast_fp16)[name = string("context_5_cast_fp16")]; + tensor var_374_perm_0 = const()[name = string("op_374_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_375 = const()[name = string("op_375"), val = tensor([1, 512, 640])]; + tensor var_374_cast_fp16 = transpose(perm = var_374_perm_0, x = context_5_cast_fp16)[name = string("transpose_72")]; + tensor context_7_cast_fp16 = reshape(shape = var_375, x = var_374_cast_fp16)[name = string("context_7_cast_fp16")]; + tensor module_layers_1_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4836864))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5246528))))[name = string("module_layers_1_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_1_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_1_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5247872)))]; + tensor linear_5_cast_fp16 = linear(bias = module_layers_1_self_attn_out_proj_bias_to_fp16, weight = module_layers_1_self_attn_out_proj_weight_to_fp16_quantized, x = context_7_cast_fp16)[name = string("linear_5_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = x_5_cast_fp16, y = linear_5_cast_fp16)[name = string("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = string("input_15_axes_0"), val = tensor([-1])]; + tensor module_layers_1_norm1_weight_to_fp16 = const()[name = string("module_layers_1_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5249216)))]; + tensor module_layers_1_norm1_bias_to_fp16 = const()[name = string("module_layers_1_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5250560)))]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = module_layers_1_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_1_norm1_weight_to_fp16, x = input_13_cast_fp16)[name = string("input_15_cast_fp16")]; + tensor module_layers_1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5251904))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6071168))))[name = string("module_layers_1_linear1_weight_to_fp16_quantized")]; + tensor module_layers_1_linear1_bias_to_fp16 = const()[name = string("module_layers_1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6073792)))]; + tensor linear_6_cast_fp16 = linear(bias = module_layers_1_linear1_bias_to_fp16, weight = module_layers_1_linear1_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_6_cast_fp16")]; + tensor input_19_cast_fp16 = relu(x = linear_6_cast_fp16)[name = string("input_19_cast_fp16")]; + tensor module_layers_1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6076416))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6895680))))[name = string("module_layers_1_linear2_weight_to_fp16_quantized")]; + tensor module_layers_1_linear2_bias_to_fp16 = const()[name = string("module_layers_1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6897024)))]; + tensor linear_7_cast_fp16 = linear(bias = module_layers_1_linear2_bias_to_fp16, weight = module_layers_1_linear2_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_7_cast_fp16")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_7_cast_fp16)[name = string("input_21_cast_fp16")]; + tensor x_7_axes_0 = const()[name = string("x_7_axes_0"), val = tensor([-1])]; + tensor module_layers_1_norm2_weight_to_fp16 = const()[name = string("module_layers_1_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6898368)))]; + tensor module_layers_1_norm2_bias_to_fp16 = const()[name = string("module_layers_1_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6899712)))]; + tensor x_7_cast_fp16 = layer_norm(axes = x_7_axes_0, beta = module_layers_1_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_1_norm2_weight_to_fp16, x = input_21_cast_fp16)[name = string("x_7_cast_fp16")]; + tensor module_layers_2_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6901056))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8129920))))[name = string("module_layers_2_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_2_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_2_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8133824)))]; + tensor linear_8_cast_fp16 = linear(bias = module_layers_2_self_attn_in_proj_bias_to_fp16, weight = module_layers_2_self_attn_in_proj_weight_to_fp16_quantized, x = x_7_cast_fp16)[name = string("linear_8_cast_fp16")]; + tensor var_398_split_sizes_0 = const()[name = string("op_398_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_398_axis_0 = const()[name = string("op_398_axis_0"), val = int32(-1)]; + tensor var_398_cast_fp16_0, tensor var_398_cast_fp16_1, tensor var_398_cast_fp16_2 = split(axis = var_398_axis_0, split_sizes = var_398_split_sizes_0, x = linear_8_cast_fp16)[name = string("op_398_cast_fp16")]; + tensor var_401 = const()[name = string("op_401"), val = tensor([1, 512, 4, 160])]; + tensor var_402_cast_fp16 = reshape(shape = var_401, x = var_398_cast_fp16_0)[name = string("op_402_cast_fp16")]; + tensor var_404 = const()[name = string("op_404"), val = tensor([1, 512, 4, 160])]; + tensor var_405_cast_fp16 = reshape(shape = var_404, x = var_398_cast_fp16_1)[name = string("op_405_cast_fp16")]; + tensor var_407 = const()[name = string("op_407"), val = tensor([1, 512, 4, 160])]; + tensor var_408_cast_fp16 = reshape(shape = var_407, x = var_398_cast_fp16_2)[name = string("op_408_cast_fp16")]; + tensor v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_411_transpose_x_0 = const()[name = string("op_411_transpose_x_0"), val = bool(false)]; + bool var_411_transpose_y_0 = const()[name = string("op_411_transpose_y_0"), val = bool(false)]; + tensor transpose_36_perm_0 = const()[name = string("transpose_36_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_37_perm_0 = const()[name = string("transpose_37_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_37 = transpose(perm = transpose_37_perm_0, x = var_405_cast_fp16)[name = string("transpose_70")]; + tensor transpose_36 = transpose(perm = transpose_36_perm_0, x = var_402_cast_fp16)[name = string("transpose_71")]; + tensor var_411_cast_fp16 = matmul(transpose_x = var_411_transpose_x_0, transpose_y = var_411_transpose_y_0, x = transpose_36, y = transpose_37)[name = string("op_411_cast_fp16")]; + fp16 var_412_to_fp16 = const()[name = string("op_412_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_5_cast_fp16 = mul(x = var_411_cast_fp16, y = var_412_to_fp16)[name = string("scores_5_cast_fp16")]; + tensor weights_13_cast_fp16 = softmax(axis = var_24, x = scores_5_cast_fp16)[name = string("weights_13_cast_fp16")]; + tensor weights_15_cast_fp16 = mul(x = weights_13_cast_fp16, y = var_313_cast_fp16)[name = string("weights_15_cast_fp16")]; + tensor var_423_axes_0 = const()[name = string("op_423_axes_0"), val = tensor([-1])]; + bool var_423_keep_dims_0 = const()[name = string("op_423_keep_dims_0"), val = bool(true)]; + tensor var_423_cast_fp16 = reduce_sum(axes = var_423_axes_0, keep_dims = var_423_keep_dims_0, x = weights_15_cast_fp16)[name = string("op_423_cast_fp16")]; + tensor var_424_cast_fp16 = maximum(x = var_423_cast_fp16, y = var_11_to_fp16)[name = string("op_424_cast_fp16")]; + tensor weights_17_cast_fp16 = real_div(x = weights_15_cast_fp16, y = var_424_cast_fp16)[name = string("weights_17_cast_fp16")]; + bool context_9_transpose_x_0 = const()[name = string("context_9_transpose_x_0"), val = bool(false)]; + bool context_9_transpose_y_0 = const()[name = string("context_9_transpose_y_0"), val = bool(false)]; + tensor v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = var_408_cast_fp16)[name = string("transpose_69")]; + tensor context_9_cast_fp16 = matmul(transpose_x = context_9_transpose_x_0, transpose_y = context_9_transpose_y_0, x = weights_17_cast_fp16, y = v_11_cast_fp16)[name = string("context_9_cast_fp16")]; + tensor var_427_perm_0 = const()[name = string("op_427_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_428 = const()[name = string("op_428"), val = tensor([1, 512, 640])]; + tensor var_427_cast_fp16 = transpose(perm = var_427_perm_0, x = context_9_cast_fp16)[name = string("transpose_68")]; + tensor context_11_cast_fp16 = reshape(shape = var_428, x = var_427_cast_fp16)[name = string("context_11_cast_fp16")]; + tensor module_layers_2_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8137728))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8547392))))[name = string("module_layers_2_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_2_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_2_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8548736)))]; + tensor linear_9_cast_fp16 = linear(bias = module_layers_2_self_attn_out_proj_bias_to_fp16, weight = module_layers_2_self_attn_out_proj_weight_to_fp16_quantized, x = context_11_cast_fp16)[name = string("linear_9_cast_fp16")]; + tensor input_23_cast_fp16 = add(x = x_7_cast_fp16, y = linear_9_cast_fp16)[name = string("input_23_cast_fp16")]; + tensor input_25_axes_0 = const()[name = string("input_25_axes_0"), val = tensor([-1])]; + tensor module_layers_2_norm1_weight_to_fp16 = const()[name = string("module_layers_2_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8550080)))]; + tensor module_layers_2_norm1_bias_to_fp16 = const()[name = string("module_layers_2_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8551424)))]; + tensor input_25_cast_fp16 = layer_norm(axes = input_25_axes_0, beta = module_layers_2_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_2_norm1_weight_to_fp16, x = input_23_cast_fp16)[name = string("input_25_cast_fp16")]; + tensor module_layers_2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8552768))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9372032))))[name = string("module_layers_2_linear1_weight_to_fp16_quantized")]; + tensor module_layers_2_linear1_bias_to_fp16 = const()[name = string("module_layers_2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9374656)))]; + tensor linear_10_cast_fp16 = linear(bias = module_layers_2_linear1_bias_to_fp16, weight = module_layers_2_linear1_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_10_cast_fp16")]; + tensor input_29_cast_fp16 = relu(x = linear_10_cast_fp16)[name = string("input_29_cast_fp16")]; + tensor module_layers_2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9377280))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10196544))))[name = string("module_layers_2_linear2_weight_to_fp16_quantized")]; + tensor module_layers_2_linear2_bias_to_fp16 = const()[name = string("module_layers_2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10197888)))]; + tensor linear_11_cast_fp16 = linear(bias = module_layers_2_linear2_bias_to_fp16, weight = module_layers_2_linear2_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_11_cast_fp16")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_11_cast_fp16)[name = string("input_31_cast_fp16")]; + tensor x_9_axes_0 = const()[name = string("x_9_axes_0"), val = tensor([-1])]; + tensor module_layers_2_norm2_weight_to_fp16 = const()[name = string("module_layers_2_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10199232)))]; + tensor module_layers_2_norm2_bias_to_fp16 = const()[name = string("module_layers_2_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10200576)))]; + tensor x_9_cast_fp16 = layer_norm(axes = x_9_axes_0, beta = module_layers_2_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_2_norm2_weight_to_fp16, x = input_31_cast_fp16)[name = string("x_9_cast_fp16")]; + tensor module_layers_3_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10201920))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11430784))))[name = string("module_layers_3_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_3_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_3_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11434688)))]; + tensor linear_12_cast_fp16 = linear(bias = module_layers_3_self_attn_in_proj_bias_to_fp16, weight = module_layers_3_self_attn_in_proj_weight_to_fp16_quantized, x = x_9_cast_fp16)[name = string("linear_12_cast_fp16")]; + tensor var_451_split_sizes_0 = const()[name = string("op_451_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_451_axis_0 = const()[name = string("op_451_axis_0"), val = int32(-1)]; + tensor var_451_cast_fp16_0, tensor var_451_cast_fp16_1, tensor var_451_cast_fp16_2 = split(axis = var_451_axis_0, split_sizes = var_451_split_sizes_0, x = linear_12_cast_fp16)[name = string("op_451_cast_fp16")]; + tensor var_454 = const()[name = string("op_454"), val = tensor([1, 512, 4, 160])]; + tensor var_455_cast_fp16 = reshape(shape = var_454, x = var_451_cast_fp16_0)[name = string("op_455_cast_fp16")]; + tensor var_457 = const()[name = string("op_457"), val = tensor([1, 512, 4, 160])]; + tensor var_458_cast_fp16 = reshape(shape = var_457, x = var_451_cast_fp16_1)[name = string("op_458_cast_fp16")]; + tensor var_460 = const()[name = string("op_460"), val = tensor([1, 512, 4, 160])]; + tensor var_461_cast_fp16 = reshape(shape = var_460, x = var_451_cast_fp16_2)[name = string("op_461_cast_fp16")]; + tensor v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_464_transpose_x_0 = const()[name = string("op_464_transpose_x_0"), val = bool(false)]; + bool var_464_transpose_y_0 = const()[name = string("op_464_transpose_y_0"), val = bool(false)]; + tensor transpose_38_perm_0 = const()[name = string("transpose_38_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_39_perm_0 = const()[name = string("transpose_39_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_39 = transpose(perm = transpose_39_perm_0, x = var_458_cast_fp16)[name = string("transpose_66")]; + tensor transpose_38 = transpose(perm = transpose_38_perm_0, x = var_455_cast_fp16)[name = string("transpose_67")]; + tensor var_464_cast_fp16 = matmul(transpose_x = var_464_transpose_x_0, transpose_y = var_464_transpose_y_0, x = transpose_38, y = transpose_39)[name = string("op_464_cast_fp16")]; + fp16 var_465_to_fp16 = const()[name = string("op_465_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_7_cast_fp16 = mul(x = var_464_cast_fp16, y = var_465_to_fp16)[name = string("scores_7_cast_fp16")]; + tensor weights_19_cast_fp16 = softmax(axis = var_24, x = scores_7_cast_fp16)[name = string("weights_19_cast_fp16")]; + tensor weights_21_cast_fp16 = mul(x = weights_19_cast_fp16, y = var_313_cast_fp16)[name = string("weights_21_cast_fp16")]; + tensor var_476_axes_0 = const()[name = string("op_476_axes_0"), val = tensor([-1])]; + bool var_476_keep_dims_0 = const()[name = string("op_476_keep_dims_0"), val = bool(true)]; + tensor var_476_cast_fp16 = reduce_sum(axes = var_476_axes_0, keep_dims = var_476_keep_dims_0, x = weights_21_cast_fp16)[name = string("op_476_cast_fp16")]; + tensor var_477_cast_fp16 = maximum(x = var_476_cast_fp16, y = var_11_to_fp16)[name = string("op_477_cast_fp16")]; + tensor weights_23_cast_fp16 = real_div(x = weights_21_cast_fp16, y = var_477_cast_fp16)[name = string("weights_23_cast_fp16")]; + bool context_13_transpose_x_0 = const()[name = string("context_13_transpose_x_0"), val = bool(false)]; + bool context_13_transpose_y_0 = const()[name = string("context_13_transpose_y_0"), val = bool(false)]; + tensor v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = var_461_cast_fp16)[name = string("transpose_65")]; + tensor context_13_cast_fp16 = matmul(transpose_x = context_13_transpose_x_0, transpose_y = context_13_transpose_y_0, x = weights_23_cast_fp16, y = v_15_cast_fp16)[name = string("context_13_cast_fp16")]; + tensor var_480_perm_0 = const()[name = string("op_480_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_481 = const()[name = string("op_481"), val = tensor([1, 512, 640])]; + tensor var_480_cast_fp16 = transpose(perm = var_480_perm_0, x = context_13_cast_fp16)[name = string("transpose_64")]; + tensor context_15_cast_fp16 = reshape(shape = var_481, x = var_480_cast_fp16)[name = string("context_15_cast_fp16")]; + tensor module_layers_3_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11438592))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11848256))))[name = string("module_layers_3_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_3_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_3_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11849600)))]; + tensor linear_13_cast_fp16 = linear(bias = module_layers_3_self_attn_out_proj_bias_to_fp16, weight = module_layers_3_self_attn_out_proj_weight_to_fp16_quantized, x = context_15_cast_fp16)[name = string("linear_13_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = x_9_cast_fp16, y = linear_13_cast_fp16)[name = string("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = string("input_35_axes_0"), val = tensor([-1])]; + tensor module_layers_3_norm1_weight_to_fp16 = const()[name = string("module_layers_3_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11850944)))]; + tensor module_layers_3_norm1_bias_to_fp16 = const()[name = string("module_layers_3_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11852288)))]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = module_layers_3_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_3_norm1_weight_to_fp16, x = input_33_cast_fp16)[name = string("input_35_cast_fp16")]; + tensor module_layers_3_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11853632))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12672896))))[name = string("module_layers_3_linear1_weight_to_fp16_quantized")]; + tensor module_layers_3_linear1_bias_to_fp16 = const()[name = string("module_layers_3_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12675520)))]; + tensor linear_14_cast_fp16 = linear(bias = module_layers_3_linear1_bias_to_fp16, weight = module_layers_3_linear1_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_14_cast_fp16")]; + tensor input_39_cast_fp16 = relu(x = linear_14_cast_fp16)[name = string("input_39_cast_fp16")]; + tensor module_layers_3_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12678144))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13497408))))[name = string("module_layers_3_linear2_weight_to_fp16_quantized")]; + tensor module_layers_3_linear2_bias_to_fp16 = const()[name = string("module_layers_3_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13498752)))]; + tensor linear_15_cast_fp16 = linear(bias = module_layers_3_linear2_bias_to_fp16, weight = module_layers_3_linear2_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_15_cast_fp16")]; + tensor input_41_cast_fp16 = add(x = input_35_cast_fp16, y = linear_15_cast_fp16)[name = string("input_41_cast_fp16")]; + tensor x_11_axes_0 = const()[name = string("x_11_axes_0"), val = tensor([-1])]; + tensor module_layers_3_norm2_weight_to_fp16 = const()[name = string("module_layers_3_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13500096)))]; + tensor module_layers_3_norm2_bias_to_fp16 = const()[name = string("module_layers_3_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13501440)))]; + tensor x_11_cast_fp16 = layer_norm(axes = x_11_axes_0, beta = module_layers_3_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_3_norm2_weight_to_fp16, x = input_41_cast_fp16)[name = string("x_11_cast_fp16")]; + tensor module_layers_4_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13502784))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14731648))))[name = string("module_layers_4_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_4_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_4_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14735552)))]; + tensor linear_16_cast_fp16 = linear(bias = module_layers_4_self_attn_in_proj_bias_to_fp16, weight = module_layers_4_self_attn_in_proj_weight_to_fp16_quantized, x = x_11_cast_fp16)[name = string("linear_16_cast_fp16")]; + tensor var_504_split_sizes_0 = const()[name = string("op_504_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_504_axis_0 = const()[name = string("op_504_axis_0"), val = int32(-1)]; + tensor var_504_cast_fp16_0, tensor var_504_cast_fp16_1, tensor var_504_cast_fp16_2 = split(axis = var_504_axis_0, split_sizes = var_504_split_sizes_0, x = linear_16_cast_fp16)[name = string("op_504_cast_fp16")]; + tensor var_507 = const()[name = string("op_507"), val = tensor([1, 512, 4, 160])]; + tensor var_508_cast_fp16 = reshape(shape = var_507, x = var_504_cast_fp16_0)[name = string("op_508_cast_fp16")]; + tensor var_510 = const()[name = string("op_510"), val = tensor([1, 512, 4, 160])]; + tensor var_511_cast_fp16 = reshape(shape = var_510, x = var_504_cast_fp16_1)[name = string("op_511_cast_fp16")]; + tensor var_513 = const()[name = string("op_513"), val = tensor([1, 512, 4, 160])]; + tensor var_514_cast_fp16 = reshape(shape = var_513, x = var_504_cast_fp16_2)[name = string("op_514_cast_fp16")]; + tensor v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_517_transpose_x_0 = const()[name = string("op_517_transpose_x_0"), val = bool(false)]; + bool var_517_transpose_y_0 = const()[name = string("op_517_transpose_y_0"), val = bool(false)]; + tensor transpose_40_perm_0 = const()[name = string("transpose_40_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_41_perm_0 = const()[name = string("transpose_41_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_41 = transpose(perm = transpose_41_perm_0, x = var_511_cast_fp16)[name = string("transpose_62")]; + tensor transpose_40 = transpose(perm = transpose_40_perm_0, x = var_508_cast_fp16)[name = string("transpose_63")]; + tensor var_517_cast_fp16 = matmul(transpose_x = var_517_transpose_x_0, transpose_y = var_517_transpose_y_0, x = transpose_40, y = transpose_41)[name = string("op_517_cast_fp16")]; + fp16 var_518_to_fp16 = const()[name = string("op_518_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_9_cast_fp16 = mul(x = var_517_cast_fp16, y = var_518_to_fp16)[name = string("scores_9_cast_fp16")]; + tensor weights_25_cast_fp16 = softmax(axis = var_24, x = scores_9_cast_fp16)[name = string("weights_25_cast_fp16")]; + tensor weights_27_cast_fp16 = mul(x = weights_25_cast_fp16, y = var_313_cast_fp16)[name = string("weights_27_cast_fp16")]; + tensor var_529_axes_0 = const()[name = string("op_529_axes_0"), val = tensor([-1])]; + bool var_529_keep_dims_0 = const()[name = string("op_529_keep_dims_0"), val = bool(true)]; + tensor var_529_cast_fp16 = reduce_sum(axes = var_529_axes_0, keep_dims = var_529_keep_dims_0, x = weights_27_cast_fp16)[name = string("op_529_cast_fp16")]; + tensor var_530_cast_fp16 = maximum(x = var_529_cast_fp16, y = var_11_to_fp16)[name = string("op_530_cast_fp16")]; + tensor weights_29_cast_fp16 = real_div(x = weights_27_cast_fp16, y = var_530_cast_fp16)[name = string("weights_29_cast_fp16")]; + bool context_17_transpose_x_0 = const()[name = string("context_17_transpose_x_0"), val = bool(false)]; + bool context_17_transpose_y_0 = const()[name = string("context_17_transpose_y_0"), val = bool(false)]; + tensor v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = var_514_cast_fp16)[name = string("transpose_61")]; + tensor context_17_cast_fp16 = matmul(transpose_x = context_17_transpose_x_0, transpose_y = context_17_transpose_y_0, x = weights_29_cast_fp16, y = v_19_cast_fp16)[name = string("context_17_cast_fp16")]; + tensor var_533_perm_0 = const()[name = string("op_533_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_534 = const()[name = string("op_534"), val = tensor([1, 512, 640])]; + tensor var_533_cast_fp16 = transpose(perm = var_533_perm_0, x = context_17_cast_fp16)[name = string("transpose_60")]; + tensor context_19_cast_fp16 = reshape(shape = var_534, x = var_533_cast_fp16)[name = string("context_19_cast_fp16")]; + tensor module_layers_4_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14739456))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15149120))))[name = string("module_layers_4_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_4_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_4_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15150464)))]; + tensor linear_17_cast_fp16 = linear(bias = module_layers_4_self_attn_out_proj_bias_to_fp16, weight = module_layers_4_self_attn_out_proj_weight_to_fp16_quantized, x = context_19_cast_fp16)[name = string("linear_17_cast_fp16")]; + tensor input_43_cast_fp16 = add(x = x_11_cast_fp16, y = linear_17_cast_fp16)[name = string("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = string("input_45_axes_0"), val = tensor([-1])]; + tensor module_layers_4_norm1_weight_to_fp16 = const()[name = string("module_layers_4_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15151808)))]; + tensor module_layers_4_norm1_bias_to_fp16 = const()[name = string("module_layers_4_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15153152)))]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = module_layers_4_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_4_norm1_weight_to_fp16, x = input_43_cast_fp16)[name = string("input_45_cast_fp16")]; + tensor module_layers_4_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15154496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15973760))))[name = string("module_layers_4_linear1_weight_to_fp16_quantized")]; + tensor module_layers_4_linear1_bias_to_fp16 = const()[name = string("module_layers_4_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15976384)))]; + tensor linear_18_cast_fp16 = linear(bias = module_layers_4_linear1_bias_to_fp16, weight = module_layers_4_linear1_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_18_cast_fp16")]; + tensor input_49_cast_fp16 = relu(x = linear_18_cast_fp16)[name = string("input_49_cast_fp16")]; + tensor module_layers_4_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15979008))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16798272))))[name = string("module_layers_4_linear2_weight_to_fp16_quantized")]; + tensor module_layers_4_linear2_bias_to_fp16 = const()[name = string("module_layers_4_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16799616)))]; + tensor linear_19_cast_fp16 = linear(bias = module_layers_4_linear2_bias_to_fp16, weight = module_layers_4_linear2_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_19_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_19_cast_fp16)[name = string("input_51_cast_fp16")]; + tensor x_13_axes_0 = const()[name = string("x_13_axes_0"), val = tensor([-1])]; + tensor module_layers_4_norm2_weight_to_fp16 = const()[name = string("module_layers_4_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16800960)))]; + tensor module_layers_4_norm2_bias_to_fp16 = const()[name = string("module_layers_4_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16802304)))]; + tensor x_13_cast_fp16 = layer_norm(axes = x_13_axes_0, beta = module_layers_4_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_4_norm2_weight_to_fp16, x = input_51_cast_fp16)[name = string("x_13_cast_fp16")]; + tensor module_layers_5_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16803648))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18032512))))[name = string("module_layers_5_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_5_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_5_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18036416)))]; + tensor linear_20_cast_fp16 = linear(bias = module_layers_5_self_attn_in_proj_bias_to_fp16, weight = module_layers_5_self_attn_in_proj_weight_to_fp16_quantized, x = x_13_cast_fp16)[name = string("linear_20_cast_fp16")]; + tensor var_557_split_sizes_0 = const()[name = string("op_557_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_557_axis_0 = const()[name = string("op_557_axis_0"), val = int32(-1)]; + tensor var_557_cast_fp16_0, tensor var_557_cast_fp16_1, tensor var_557_cast_fp16_2 = split(axis = var_557_axis_0, split_sizes = var_557_split_sizes_0, x = linear_20_cast_fp16)[name = string("op_557_cast_fp16")]; + tensor var_560 = const()[name = string("op_560"), val = tensor([1, 512, 4, 160])]; + tensor var_561_cast_fp16 = reshape(shape = var_560, x = var_557_cast_fp16_0)[name = string("op_561_cast_fp16")]; + tensor var_563 = const()[name = string("op_563"), val = tensor([1, 512, 4, 160])]; + tensor var_564_cast_fp16 = reshape(shape = var_563, x = var_557_cast_fp16_1)[name = string("op_564_cast_fp16")]; + tensor var_566 = const()[name = string("op_566"), val = tensor([1, 512, 4, 160])]; + tensor var_567_cast_fp16 = reshape(shape = var_566, x = var_557_cast_fp16_2)[name = string("op_567_cast_fp16")]; + tensor v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_570_transpose_x_0 = const()[name = string("op_570_transpose_x_0"), val = bool(false)]; + bool var_570_transpose_y_0 = const()[name = string("op_570_transpose_y_0"), val = bool(false)]; + tensor transpose_42_perm_0 = const()[name = string("transpose_42_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_43_perm_0 = const()[name = string("transpose_43_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_43 = transpose(perm = transpose_43_perm_0, x = var_564_cast_fp16)[name = string("transpose_58")]; + tensor transpose_42 = transpose(perm = transpose_42_perm_0, x = var_561_cast_fp16)[name = string("transpose_59")]; + tensor var_570_cast_fp16 = matmul(transpose_x = var_570_transpose_x_0, transpose_y = var_570_transpose_y_0, x = transpose_42, y = transpose_43)[name = string("op_570_cast_fp16")]; + fp16 var_571_to_fp16 = const()[name = string("op_571_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_11_cast_fp16 = mul(x = var_570_cast_fp16, y = var_571_to_fp16)[name = string("scores_11_cast_fp16")]; + tensor weights_31_cast_fp16 = softmax(axis = var_24, x = scores_11_cast_fp16)[name = string("weights_31_cast_fp16")]; + tensor weights_33_cast_fp16 = mul(x = weights_31_cast_fp16, y = var_313_cast_fp16)[name = string("weights_33_cast_fp16")]; + tensor var_582_axes_0 = const()[name = string("op_582_axes_0"), val = tensor([-1])]; + bool var_582_keep_dims_0 = const()[name = string("op_582_keep_dims_0"), val = bool(true)]; + tensor var_582_cast_fp16 = reduce_sum(axes = var_582_axes_0, keep_dims = var_582_keep_dims_0, x = weights_33_cast_fp16)[name = string("op_582_cast_fp16")]; + tensor var_583_cast_fp16 = maximum(x = var_582_cast_fp16, y = var_11_to_fp16)[name = string("op_583_cast_fp16")]; + tensor weights_35_cast_fp16 = real_div(x = weights_33_cast_fp16, y = var_583_cast_fp16)[name = string("weights_35_cast_fp16")]; + bool context_21_transpose_x_0 = const()[name = string("context_21_transpose_x_0"), val = bool(false)]; + bool context_21_transpose_y_0 = const()[name = string("context_21_transpose_y_0"), val = bool(false)]; + tensor v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = var_567_cast_fp16)[name = string("transpose_57")]; + tensor context_21_cast_fp16 = matmul(transpose_x = context_21_transpose_x_0, transpose_y = context_21_transpose_y_0, x = weights_35_cast_fp16, y = v_23_cast_fp16)[name = string("context_21_cast_fp16")]; + tensor var_586_perm_0 = const()[name = string("op_586_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_587 = const()[name = string("op_587"), val = tensor([1, 512, 640])]; + tensor var_586_cast_fp16 = transpose(perm = var_586_perm_0, x = context_21_cast_fp16)[name = string("transpose_56")]; + tensor context_23_cast_fp16 = reshape(shape = var_587, x = var_586_cast_fp16)[name = string("context_23_cast_fp16")]; + tensor module_layers_5_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18040320))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18449984))))[name = string("module_layers_5_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_5_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_5_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18451328)))]; + tensor linear_21_cast_fp16 = linear(bias = module_layers_5_self_attn_out_proj_bias_to_fp16, weight = module_layers_5_self_attn_out_proj_weight_to_fp16_quantized, x = context_23_cast_fp16)[name = string("linear_21_cast_fp16")]; + tensor input_53_cast_fp16 = add(x = x_13_cast_fp16, y = linear_21_cast_fp16)[name = string("input_53_cast_fp16")]; + tensor input_55_axes_0 = const()[name = string("input_55_axes_0"), val = tensor([-1])]; + tensor module_layers_5_norm1_weight_to_fp16 = const()[name = string("module_layers_5_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18452672)))]; + tensor module_layers_5_norm1_bias_to_fp16 = const()[name = string("module_layers_5_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18454016)))]; + tensor input_55_cast_fp16 = layer_norm(axes = input_55_axes_0, beta = module_layers_5_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_5_norm1_weight_to_fp16, x = input_53_cast_fp16)[name = string("input_55_cast_fp16")]; + tensor module_layers_5_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18455360))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19274624))))[name = string("module_layers_5_linear1_weight_to_fp16_quantized")]; + tensor module_layers_5_linear1_bias_to_fp16 = const()[name = string("module_layers_5_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19277248)))]; + tensor linear_22_cast_fp16 = linear(bias = module_layers_5_linear1_bias_to_fp16, weight = module_layers_5_linear1_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_22_cast_fp16")]; + tensor input_59_cast_fp16 = relu(x = linear_22_cast_fp16)[name = string("input_59_cast_fp16")]; + tensor module_layers_5_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19279872))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20099136))))[name = string("module_layers_5_linear2_weight_to_fp16_quantized")]; + tensor module_layers_5_linear2_bias_to_fp16 = const()[name = string("module_layers_5_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20100480)))]; + tensor linear_23_cast_fp16 = linear(bias = module_layers_5_linear2_bias_to_fp16, weight = module_layers_5_linear2_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_23_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_23_cast_fp16)[name = string("input_61_cast_fp16")]; + tensor x_15_axes_0 = const()[name = string("x_15_axes_0"), val = tensor([-1])]; + tensor module_layers_5_norm2_weight_to_fp16 = const()[name = string("module_layers_5_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20101824)))]; + tensor module_layers_5_norm2_bias_to_fp16 = const()[name = string("module_layers_5_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20103168)))]; + tensor x_15_cast_fp16 = layer_norm(axes = x_15_axes_0, beta = module_layers_5_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_5_norm2_weight_to_fp16, x = input_61_cast_fp16)[name = string("x_15_cast_fp16")]; + tensor module_layers_6_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20104512))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21333376))))[name = string("module_layers_6_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_6_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_6_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21337280)))]; + tensor linear_24_cast_fp16 = linear(bias = module_layers_6_self_attn_in_proj_bias_to_fp16, weight = module_layers_6_self_attn_in_proj_weight_to_fp16_quantized, x = x_15_cast_fp16)[name = string("linear_24_cast_fp16")]; + tensor var_610_split_sizes_0 = const()[name = string("op_610_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_610_axis_0 = const()[name = string("op_610_axis_0"), val = int32(-1)]; + tensor var_610_cast_fp16_0, tensor var_610_cast_fp16_1, tensor var_610_cast_fp16_2 = split(axis = var_610_axis_0, split_sizes = var_610_split_sizes_0, x = linear_24_cast_fp16)[name = string("op_610_cast_fp16")]; + tensor var_613 = const()[name = string("op_613"), val = tensor([1, 512, 4, 160])]; + tensor var_614_cast_fp16 = reshape(shape = var_613, x = var_610_cast_fp16_0)[name = string("op_614_cast_fp16")]; + tensor var_616 = const()[name = string("op_616"), val = tensor([1, 512, 4, 160])]; + tensor var_617_cast_fp16 = reshape(shape = var_616, x = var_610_cast_fp16_1)[name = string("op_617_cast_fp16")]; + tensor var_619 = const()[name = string("op_619"), val = tensor([1, 512, 4, 160])]; + tensor var_620_cast_fp16 = reshape(shape = var_619, x = var_610_cast_fp16_2)[name = string("op_620_cast_fp16")]; + tensor v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_623_transpose_x_0 = const()[name = string("op_623_transpose_x_0"), val = bool(false)]; + bool var_623_transpose_y_0 = const()[name = string("op_623_transpose_y_0"), val = bool(false)]; + tensor transpose_44_perm_0 = const()[name = string("transpose_44_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_45_perm_0 = const()[name = string("transpose_45_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_45 = transpose(perm = transpose_45_perm_0, x = var_617_cast_fp16)[name = string("transpose_54")]; + tensor transpose_44 = transpose(perm = transpose_44_perm_0, x = var_614_cast_fp16)[name = string("transpose_55")]; + tensor var_623_cast_fp16 = matmul(transpose_x = var_623_transpose_x_0, transpose_y = var_623_transpose_y_0, x = transpose_44, y = transpose_45)[name = string("op_623_cast_fp16")]; + fp16 var_624_to_fp16 = const()[name = string("op_624_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_13_cast_fp16 = mul(x = var_623_cast_fp16, y = var_624_to_fp16)[name = string("scores_13_cast_fp16")]; + tensor weights_37_cast_fp16 = softmax(axis = var_24, x = scores_13_cast_fp16)[name = string("weights_37_cast_fp16")]; + tensor weights_39_cast_fp16 = mul(x = weights_37_cast_fp16, y = var_313_cast_fp16)[name = string("weights_39_cast_fp16")]; + tensor var_635_axes_0 = const()[name = string("op_635_axes_0"), val = tensor([-1])]; + bool var_635_keep_dims_0 = const()[name = string("op_635_keep_dims_0"), val = bool(true)]; + tensor var_635_cast_fp16 = reduce_sum(axes = var_635_axes_0, keep_dims = var_635_keep_dims_0, x = weights_39_cast_fp16)[name = string("op_635_cast_fp16")]; + tensor var_636_cast_fp16 = maximum(x = var_635_cast_fp16, y = var_11_to_fp16)[name = string("op_636_cast_fp16")]; + tensor weights_41_cast_fp16 = real_div(x = weights_39_cast_fp16, y = var_636_cast_fp16)[name = string("weights_41_cast_fp16")]; + bool context_25_transpose_x_0 = const()[name = string("context_25_transpose_x_0"), val = bool(false)]; + bool context_25_transpose_y_0 = const()[name = string("context_25_transpose_y_0"), val = bool(false)]; + tensor v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = var_620_cast_fp16)[name = string("transpose_53")]; + tensor context_25_cast_fp16 = matmul(transpose_x = context_25_transpose_x_0, transpose_y = context_25_transpose_y_0, x = weights_41_cast_fp16, y = v_27_cast_fp16)[name = string("context_25_cast_fp16")]; + tensor var_639_perm_0 = const()[name = string("op_639_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_640 = const()[name = string("op_640"), val = tensor([1, 512, 640])]; + tensor var_639_cast_fp16 = transpose(perm = var_639_perm_0, x = context_25_cast_fp16)[name = string("transpose_52")]; + tensor context_27_cast_fp16 = reshape(shape = var_640, x = var_639_cast_fp16)[name = string("context_27_cast_fp16")]; + tensor module_layers_6_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21341184))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21750848))))[name = string("module_layers_6_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_6_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_6_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21752192)))]; + tensor linear_25_cast_fp16 = linear(bias = module_layers_6_self_attn_out_proj_bias_to_fp16, weight = module_layers_6_self_attn_out_proj_weight_to_fp16_quantized, x = context_27_cast_fp16)[name = string("linear_25_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = x_15_cast_fp16, y = linear_25_cast_fp16)[name = string("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor([-1])]; + tensor module_layers_6_norm1_weight_to_fp16 = const()[name = string("module_layers_6_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21753536)))]; + tensor module_layers_6_norm1_bias_to_fp16 = const()[name = string("module_layers_6_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21754880)))]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = module_layers_6_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_6_norm1_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")]; + tensor module_layers_6_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21756224))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22575488))))[name = string("module_layers_6_linear1_weight_to_fp16_quantized")]; + tensor module_layers_6_linear1_bias_to_fp16 = const()[name = string("module_layers_6_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22578112)))]; + tensor linear_26_cast_fp16 = linear(bias = module_layers_6_linear1_bias_to_fp16, weight = module_layers_6_linear1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_26_cast_fp16")]; + tensor input_69_cast_fp16 = relu(x = linear_26_cast_fp16)[name = string("input_69_cast_fp16")]; + tensor module_layers_6_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22580736))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23400000))))[name = string("module_layers_6_linear2_weight_to_fp16_quantized")]; + tensor module_layers_6_linear2_bias_to_fp16 = const()[name = string("module_layers_6_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23401344)))]; + tensor linear_27_cast_fp16 = linear(bias = module_layers_6_linear2_bias_to_fp16, weight = module_layers_6_linear2_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_27_cast_fp16")]; + tensor input_71_cast_fp16 = add(x = input_65_cast_fp16, y = linear_27_cast_fp16)[name = string("input_71_cast_fp16")]; + tensor x_axes_0 = const()[name = string("x_axes_0"), val = tensor([-1])]; + tensor module_layers_6_norm2_weight_to_fp16 = const()[name = string("module_layers_6_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23402688)))]; + tensor module_layers_6_norm2_bias_to_fp16 = const()[name = string("module_layers_6_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23404032)))]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = module_layers_6_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_6_norm2_weight_to_fp16, x = input_71_cast_fp16)[name = string("x_cast_fp16")]; + tensor module_layers_7_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23405376))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24634240))))[name = string("module_layers_7_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_7_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_7_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24638144)))]; + tensor linear_28_cast_fp16 = linear(bias = module_layers_7_self_attn_in_proj_bias_to_fp16, weight = module_layers_7_self_attn_in_proj_weight_to_fp16_quantized, x = x_cast_fp16)[name = string("linear_28_cast_fp16")]; + tensor var_663_split_sizes_0 = const()[name = string("op_663_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_663_axis_0 = const()[name = string("op_663_axis_0"), val = int32(-1)]; + tensor var_663_cast_fp16_0, tensor var_663_cast_fp16_1, tensor var_663_cast_fp16_2 = split(axis = var_663_axis_0, split_sizes = var_663_split_sizes_0, x = linear_28_cast_fp16)[name = string("op_663_cast_fp16")]; + tensor var_666 = const()[name = string("op_666"), val = tensor([1, 512, 4, 160])]; + tensor var_667_cast_fp16 = reshape(shape = var_666, x = var_663_cast_fp16_0)[name = string("op_667_cast_fp16")]; + tensor var_669 = const()[name = string("op_669"), val = tensor([1, 512, 4, 160])]; + tensor var_670_cast_fp16 = reshape(shape = var_669, x = var_663_cast_fp16_1)[name = string("op_670_cast_fp16")]; + tensor var_672 = const()[name = string("op_672"), val = tensor([1, 512, 4, 160])]; + tensor var_673_cast_fp16 = reshape(shape = var_672, x = var_663_cast_fp16_2)[name = string("op_673_cast_fp16")]; + tensor v_perm_0 = const()[name = string("v_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_676_transpose_x_0 = const()[name = string("op_676_transpose_x_0"), val = bool(false)]; + bool var_676_transpose_y_0 = const()[name = string("op_676_transpose_y_0"), val = bool(false)]; + tensor transpose_46_perm_0 = const()[name = string("transpose_46_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_47_perm_0 = const()[name = string("transpose_47_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_47 = transpose(perm = transpose_47_perm_0, x = var_670_cast_fp16)[name = string("transpose_50")]; + tensor transpose_46 = transpose(perm = transpose_46_perm_0, x = var_667_cast_fp16)[name = string("transpose_51")]; + tensor var_676_cast_fp16 = matmul(transpose_x = var_676_transpose_x_0, transpose_y = var_676_transpose_y_0, x = transpose_46, y = transpose_47)[name = string("op_676_cast_fp16")]; + fp16 var_677_to_fp16 = const()[name = string("op_677_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_cast_fp16 = mul(x = var_676_cast_fp16, y = var_677_to_fp16)[name = string("scores_cast_fp16")]; + tensor weights_43_cast_fp16 = softmax(axis = var_24, x = scores_cast_fp16)[name = string("weights_43_cast_fp16")]; + tensor weights_45_cast_fp16 = mul(x = weights_43_cast_fp16, y = var_313_cast_fp16)[name = string("weights_45_cast_fp16")]; + tensor var_688_axes_0 = const()[name = string("op_688_axes_0"), val = tensor([-1])]; + bool var_688_keep_dims_0 = const()[name = string("op_688_keep_dims_0"), val = bool(true)]; + tensor var_688_cast_fp16 = reduce_sum(axes = var_688_axes_0, keep_dims = var_688_keep_dims_0, x = weights_45_cast_fp16)[name = string("op_688_cast_fp16")]; + tensor var_689_cast_fp16 = maximum(x = var_688_cast_fp16, y = var_11_to_fp16)[name = string("op_689_cast_fp16")]; + tensor weights_cast_fp16 = real_div(x = weights_45_cast_fp16, y = var_689_cast_fp16)[name = string("weights_cast_fp16")]; + bool context_29_transpose_x_0 = const()[name = string("context_29_transpose_x_0"), val = bool(false)]; + bool context_29_transpose_y_0 = const()[name = string("context_29_transpose_y_0"), val = bool(false)]; + tensor v_cast_fp16 = transpose(perm = v_perm_0, x = var_673_cast_fp16)[name = string("transpose_49")]; + tensor context_29_cast_fp16 = matmul(transpose_x = context_29_transpose_x_0, transpose_y = context_29_transpose_y_0, x = weights_cast_fp16, y = v_cast_fp16)[name = string("context_29_cast_fp16")]; + tensor var_692_perm_0 = const()[name = string("op_692_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_693 = const()[name = string("op_693"), val = tensor([1, 512, 640])]; + tensor var_692_cast_fp16 = transpose(perm = var_692_perm_0, x = context_29_cast_fp16)[name = string("transpose_48")]; + tensor context_cast_fp16 = reshape(shape = var_693, x = var_692_cast_fp16)[name = string("context_cast_fp16")]; + tensor module_layers_7_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24642048))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25051712))))[name = string("module_layers_7_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_7_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_7_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25053056)))]; + tensor linear_29_cast_fp16 = linear(bias = module_layers_7_self_attn_out_proj_bias_to_fp16, weight = module_layers_7_self_attn_out_proj_weight_to_fp16_quantized, x = context_cast_fp16)[name = string("linear_29_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = x_cast_fp16, y = linear_29_cast_fp16)[name = string("input_73_cast_fp16")]; + tensor input_75_axes_0 = const()[name = string("input_75_axes_0"), val = tensor([-1])]; + tensor module_layers_7_norm1_weight_to_fp16 = const()[name = string("module_layers_7_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25054400)))]; + tensor module_layers_7_norm1_bias_to_fp16 = const()[name = string("module_layers_7_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25055744)))]; + tensor input_75_cast_fp16 = layer_norm(axes = input_75_axes_0, beta = module_layers_7_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_7_norm1_weight_to_fp16, x = input_73_cast_fp16)[name = string("input_75_cast_fp16")]; + tensor module_layers_7_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25057088))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25876352))))[name = string("module_layers_7_linear1_weight_to_fp16_quantized")]; + tensor module_layers_7_linear1_bias_to_fp16 = const()[name = string("module_layers_7_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25878976)))]; + tensor linear_30_cast_fp16 = linear(bias = module_layers_7_linear1_bias_to_fp16, weight = module_layers_7_linear1_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_30_cast_fp16")]; + tensor input_79_cast_fp16 = relu(x = linear_30_cast_fp16)[name = string("input_79_cast_fp16")]; + tensor module_layers_7_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25881600))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26700864))))[name = string("module_layers_7_linear2_weight_to_fp16_quantized")]; + tensor module_layers_7_linear2_bias_to_fp16 = const()[name = string("module_layers_7_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26702208)))]; + tensor linear_31_cast_fp16 = linear(bias = module_layers_7_linear2_bias_to_fp16, weight = module_layers_7_linear2_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_31_cast_fp16")]; + tensor input_81_cast_fp16 = add(x = input_75_cast_fp16, y = linear_31_cast_fp16)[name = string("input_81_cast_fp16")]; + tensor input_83_axes_0 = const()[name = string("input_83_axes_0"), val = tensor([-1])]; + tensor module_layers_7_norm2_weight_to_fp16 = const()[name = string("module_layers_7_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26703552)))]; + tensor module_layers_7_norm2_bias_to_fp16 = const()[name = string("module_layers_7_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26704896)))]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = module_layers_7_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_7_norm2_weight_to_fp16, x = input_81_cast_fp16)[name = string("input_83_cast_fp16")]; + tensor input_axes_0 = const()[name = string("input_axes_0"), val = tensor([-1])]; + tensor module_encoder_norm_weight_to_fp16 = const()[name = string("module_encoder_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26706240)))]; + tensor module_encoder_norm_bias_to_fp16 = const()[name = string("module_encoder_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26707584)))]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = module_encoder_norm_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_encoder_norm_weight_to_fp16, x = input_83_cast_fp16)[name = string("input_cast_fp16")]; + tensor module_fc_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26708928))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26770432))))[name = string("module_fc_out_weight_to_fp16_quantized")]; + tensor module_fc_out_bias_to_fp16 = const()[name = string("module_fc_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26770688)))]; + tensor linear_32_cast_fp16 = linear(bias = module_fc_out_bias_to_fp16, weight = module_fc_out_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("linear_32_cast_fp16")]; + tensor padded_axes_0 = const()[name = string("padded_axes_0"), val = tensor([-1])]; + tensor padded = expand_dims(axes = padded_axes_0, x = src_pad_mask)[name = string("padded")]; + tensor var_723_begin_0 = const()[name = string("op_723_begin_0"), val = tensor([0, 0, 0])]; + tensor var_723_end_0 = const()[name = string("op_723_end_0"), val = tensor([1, 512, 1])]; + tensor var_723_end_mask_0 = const()[name = string("op_723_end_mask_0"), val = tensor([true, true, false])]; + tensor var_723_cast_fp16 = slice_by_index(begin = var_723_begin_0, end = var_723_end_0, end_mask = var_723_end_mask_0, x = linear_32_cast_fp16)[name = string("op_723_cast_fp16")]; + tensor var_722_to_fp16 = const()[name = string("op_722_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27345664)))]; + tensor blank_logits_cast_fp16 = select(a = var_722_to_fp16, b = var_723_cast_fp16, cond = padded)[name = string("blank_logits_cast_fp16")]; + tensor var_725_begin_0 = const()[name = string("op_725_begin_0"), val = tensor([0, 0, 1])]; + tensor var_725_end_0 = const()[name = string("op_725_end_0"), val = tensor([1, 512, 96])]; + tensor var_725_end_mask_0 = const()[name = string("op_725_end_mask_0"), val = tensor([true, true, true])]; + tensor var_725_cast_fp16 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, x = linear_32_cast_fp16)[name = string("op_725_cast_fp16")]; + fp16 var_726_value_0_to_fp16 = const()[name = string("op_726_value_0_to_fp16"), val = fp16(-0x1.388p+13)]; + tensor var_726_cast_fp16 = fill_like(ref_tensor = var_725_cast_fp16, value = var_726_value_0_to_fp16)[name = string("op_726_cast_fp16")]; + tensor padded_after_broadcast_reps_0 = const()[name = string("padded_after_broadcast_reps_0"), val = tensor([1, 1, 95])]; + tensor padded_after_broadcast = tile(reps = padded_after_broadcast_reps_0, x = padded)[name = string("padded_after_broadcast")]; + tensor nonblank_logits_cast_fp16 = select(a = var_726_cast_fp16, b = var_725_cast_fp16, cond = padded_after_broadcast)[name = string("nonblank_logits_cast_fp16")]; + bool var_730_interleave_0 = const()[name = string("op_730_interleave_0"), val = bool(false)]; + tensor var_730_cast_fp16 = concat(axis = var_24, interleave = var_730_interleave_0, values = (blank_logits_cast_fp16, nonblank_logits_cast_fp16))[name = string("op_730_cast_fp16")]; + string var_730_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_730_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = var_730_cast_fp16_to_fp32_dtype_0, x = var_730_cast_fp16)[name = string("cast_1")]; + } -> (logits); + func g2p_128(tensor text) { + int32 var_24 = const()[name = string("op_24"), val = int32(-1)]; + int32 var_26 = const()[name = string("op_26"), val = int32(0)]; + tensor src_pad_mask = equal(x = text, y = var_26)[name = string("src_pad_mask")]; + int32 x_1_batch_dims_0 = const()[name = string("x_1_batch_dims_0"), val = int32(0)]; + bool x_1_validate_indices_0 = const()[name = string("x_1_validate_indices_0"), val = bool(false)]; + tensor module_embedding_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257408))))[name = string("module_embedding_weight_to_fp16_quantized")]; + string text_to_int16_dtype_0 = const()[name = string("text_to_int16_dtype_0"), val = string("int16")]; + string cast_49_dtype_0 = const()[name = string("cast_49_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor text_to_int16 = cast(dtype = text_to_int16_dtype_0, x = text)[name = string("cast_6")]; + tensor cast_49 = cast(dtype = cast_49_dtype_0, x = text_to_int16)[name = string("cast_5")]; + tensor greater_equal_0 = greater_equal(x = cast_49, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(402)]; + tensor add_0 = add(x = cast_49, y = slice_by_index_0)[name = string("add_0")]; + tensor select_0 = select(a = cast_49, b = add_0, cond = greater_equal_0)[name = string("select_0")]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + string cast_0_dtype_0 = const()[name = string("cast_0_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_4")]; + tensor cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = string("cast_3")]; + tensor greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")]; + int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(402)]; + tensor add_0_1 = add(x = cast_0, y = slice_by_index_0_1)[name = string("add_0_1")]; + tensor select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")]; + int32 x_1_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = string("x_1_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = int32(0)]; + tensor x_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = x_1_cast_fp16_cast_uint16_cast_uint16_axis_0, batch_dims = x_1_batch_dims_0, indices = select_0_1, validate_indices = x_1_validate_indices_0, x = module_embedding_weight_to_fp16_quantized)[name = string("x_1_cast_fp16_cast_uint16_cast_uint16")]; + tensor op_286_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26771136))), scale = tensor([[[0x1.724p-7]]]))[name = string("op_286_to_fp16_quantized")]; + tensor input_1_cast_fp16 = add(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = op_286_to_fp16_quantized)[name = string("input_1_cast_fp16")]; + tensor module_layers_0_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1528192))))[name = string("module_layers_0_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_0_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_0_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1532096)))]; + tensor linear_0_cast_fp16 = linear(bias = module_layers_0_self_attn_in_proj_bias_to_fp16, weight = module_layers_0_self_attn_in_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor var_292_split_sizes_0 = const()[name = string("op_292_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_292_axis_0 = const()[name = string("op_292_axis_0"), val = int32(-1)]; + tensor var_292_cast_fp16_0, tensor var_292_cast_fp16_1, tensor var_292_cast_fp16_2 = split(axis = var_292_axis_0, split_sizes = var_292_split_sizes_0, x = linear_0_cast_fp16)[name = string("op_292_cast_fp16")]; + tensor var_295 = const()[name = string("op_295"), val = tensor([1, 128, 4, 160])]; + tensor var_296_cast_fp16 = reshape(shape = var_295, x = var_292_cast_fp16_0)[name = string("op_296_cast_fp16")]; + tensor var_298 = const()[name = string("op_298"), val = tensor([1, 128, 4, 160])]; + tensor var_299_cast_fp16 = reshape(shape = var_298, x = var_292_cast_fp16_1)[name = string("op_299_cast_fp16")]; + tensor var_301 = const()[name = string("op_301"), val = tensor([1, 128, 4, 160])]; + tensor var_302_cast_fp16 = reshape(shape = var_301, x = var_292_cast_fp16_2)[name = string("op_302_cast_fp16")]; + tensor v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_305_transpose_x_0 = const()[name = string("op_305_transpose_x_0"), val = bool(false)]; + bool var_305_transpose_y_0 = const()[name = string("op_305_transpose_y_0"), val = bool(false)]; + tensor transpose_32_perm_0 = const()[name = string("transpose_32_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_33 = transpose(perm = transpose_33_perm_0, x = var_299_cast_fp16)[name = string("transpose_78")]; + tensor transpose_32 = transpose(perm = transpose_32_perm_0, x = var_296_cast_fp16)[name = string("transpose_79")]; + tensor var_305_cast_fp16 = matmul(transpose_x = var_305_transpose_x_0, transpose_y = var_305_transpose_y_0, x = transpose_32, y = transpose_33)[name = string("op_305_cast_fp16")]; + fp16 var_306_to_fp16 = const()[name = string("op_306_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_1_cast_fp16 = mul(x = var_305_cast_fp16, y = var_306_to_fp16)[name = string("scores_1_cast_fp16")]; + tensor weights_1_cast_fp16 = softmax(axis = var_24, x = scores_1_cast_fp16)[name = string("weights_1_cast_fp16")]; + fp16 var_13_to_fp16 = const()[name = string("op_13_to_fp16"), val = fp16(0x1p+0)]; + string var_309_to_fp16_dtype_0 = const()[name = string("op_309_to_fp16_dtype_0"), val = string("fp16")]; + tensor src_pad_mask_to_fp16 = cast(dtype = var_309_to_fp16_dtype_0, x = src_pad_mask)[name = string("cast_2")]; + tensor var_310_cast_fp16 = sub(x = var_13_to_fp16, y = src_pad_mask_to_fp16)[name = string("op_310_cast_fp16")]; + tensor var_312_axes_0 = const()[name = string("op_312_axes_0"), val = tensor([1])]; + tensor var_312_cast_fp16 = expand_dims(axes = var_312_axes_0, x = var_310_cast_fp16)[name = string("op_312_cast_fp16")]; + tensor var_313_axes_0 = const()[name = string("op_313_axes_0"), val = tensor([2])]; + tensor var_313_cast_fp16 = expand_dims(axes = var_313_axes_0, x = var_312_cast_fp16)[name = string("op_313_cast_fp16")]; + tensor weights_3_cast_fp16 = mul(x = weights_1_cast_fp16, y = var_313_cast_fp16)[name = string("weights_3_cast_fp16")]; + tensor var_317_axes_0 = const()[name = string("op_317_axes_0"), val = tensor([-1])]; + bool var_317_keep_dims_0 = const()[name = string("op_317_keep_dims_0"), val = bool(true)]; + tensor var_317_cast_fp16 = reduce_sum(axes = var_317_axes_0, keep_dims = var_317_keep_dims_0, x = weights_3_cast_fp16)[name = string("op_317_cast_fp16")]; + fp16 var_11_to_fp16 = const()[name = string("op_11_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_318_cast_fp16 = maximum(x = var_317_cast_fp16, y = var_11_to_fp16)[name = string("op_318_cast_fp16")]; + tensor weights_5_cast_fp16 = real_div(x = weights_3_cast_fp16, y = var_318_cast_fp16)[name = string("weights_5_cast_fp16")]; + bool context_1_transpose_x_0 = const()[name = string("context_1_transpose_x_0"), val = bool(false)]; + bool context_1_transpose_y_0 = const()[name = string("context_1_transpose_y_0"), val = bool(false)]; + tensor v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_302_cast_fp16)[name = string("transpose_77")]; + tensor context_1_cast_fp16 = matmul(transpose_x = context_1_transpose_x_0, transpose_y = context_1_transpose_y_0, x = weights_5_cast_fp16, y = v_3_cast_fp16)[name = string("context_1_cast_fp16")]; + tensor var_321_perm_0 = const()[name = string("op_321_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_322 = const()[name = string("op_322"), val = tensor([1, 128, 640])]; + tensor var_321_cast_fp16 = transpose(perm = var_321_perm_0, x = context_1_cast_fp16)[name = string("transpose_76")]; + tensor context_3_cast_fp16 = reshape(shape = var_322, x = var_321_cast_fp16)[name = string("context_3_cast_fp16")]; + tensor module_layers_0_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1536000))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1945664))))[name = string("module_layers_0_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_0_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_0_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1947008)))]; + tensor linear_1_cast_fp16 = linear(bias = module_layers_0_self_attn_out_proj_bias_to_fp16, weight = module_layers_0_self_attn_out_proj_weight_to_fp16_quantized, x = context_3_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor input_3_cast_fp16 = add(x = input_1_cast_fp16, y = linear_1_cast_fp16)[name = string("input_3_cast_fp16")]; + tensor input_5_axes_0 = const()[name = string("input_5_axes_0"), val = tensor([-1])]; + tensor module_layers_0_norm1_weight_to_fp16 = const()[name = string("module_layers_0_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1948352)))]; + tensor module_layers_0_norm1_bias_to_fp16 = const()[name = string("module_layers_0_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1949696)))]; + fp16 var_9_to_fp16 = const()[name = string("op_9_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = module_layers_0_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_0_norm1_weight_to_fp16, x = input_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor module_layers_0_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1951040))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2770304))))[name = string("module_layers_0_linear1_weight_to_fp16_quantized")]; + tensor module_layers_0_linear1_bias_to_fp16 = const()[name = string("module_layers_0_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2772928)))]; + tensor linear_2_cast_fp16 = linear(bias = module_layers_0_linear1_bias_to_fp16, weight = module_layers_0_linear1_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_2_cast_fp16")]; + tensor input_9_cast_fp16 = relu(x = linear_2_cast_fp16)[name = string("input_9_cast_fp16")]; + tensor module_layers_0_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2775552))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3594816))))[name = string("module_layers_0_linear2_weight_to_fp16_quantized")]; + tensor module_layers_0_linear2_bias_to_fp16 = const()[name = string("module_layers_0_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3596160)))]; + tensor linear_3_cast_fp16 = linear(bias = module_layers_0_linear2_bias_to_fp16, weight = module_layers_0_linear2_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("linear_3_cast_fp16")]; + tensor input_11_cast_fp16 = add(x = input_5_cast_fp16, y = linear_3_cast_fp16)[name = string("input_11_cast_fp16")]; + tensor x_5_axes_0 = const()[name = string("x_5_axes_0"), val = tensor([-1])]; + tensor module_layers_0_norm2_weight_to_fp16 = const()[name = string("module_layers_0_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3597504)))]; + tensor module_layers_0_norm2_bias_to_fp16 = const()[name = string("module_layers_0_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3598848)))]; + tensor x_5_cast_fp16 = layer_norm(axes = x_5_axes_0, beta = module_layers_0_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_0_norm2_weight_to_fp16, x = input_11_cast_fp16)[name = string("x_5_cast_fp16")]; + tensor module_layers_1_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3600192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4829056))))[name = string("module_layers_1_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_1_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_1_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4832960)))]; + tensor linear_4_cast_fp16 = linear(bias = module_layers_1_self_attn_in_proj_bias_to_fp16, weight = module_layers_1_self_attn_in_proj_weight_to_fp16_quantized, x = x_5_cast_fp16)[name = string("linear_4_cast_fp16")]; + tensor var_345_split_sizes_0 = const()[name = string("op_345_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_345_axis_0 = const()[name = string("op_345_axis_0"), val = int32(-1)]; + tensor var_345_cast_fp16_0, tensor var_345_cast_fp16_1, tensor var_345_cast_fp16_2 = split(axis = var_345_axis_0, split_sizes = var_345_split_sizes_0, x = linear_4_cast_fp16)[name = string("op_345_cast_fp16")]; + tensor var_348 = const()[name = string("op_348"), val = tensor([1, 128, 4, 160])]; + tensor var_349_cast_fp16 = reshape(shape = var_348, x = var_345_cast_fp16_0)[name = string("op_349_cast_fp16")]; + tensor var_351 = const()[name = string("op_351"), val = tensor([1, 128, 4, 160])]; + tensor var_352_cast_fp16 = reshape(shape = var_351, x = var_345_cast_fp16_1)[name = string("op_352_cast_fp16")]; + tensor var_354 = const()[name = string("op_354"), val = tensor([1, 128, 4, 160])]; + tensor var_355_cast_fp16 = reshape(shape = var_354, x = var_345_cast_fp16_2)[name = string("op_355_cast_fp16")]; + tensor v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_358_transpose_x_0 = const()[name = string("op_358_transpose_x_0"), val = bool(false)]; + bool var_358_transpose_y_0 = const()[name = string("op_358_transpose_y_0"), val = bool(false)]; + tensor transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_35_perm_0 = const()[name = string("transpose_35_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_35 = transpose(perm = transpose_35_perm_0, x = var_352_cast_fp16)[name = string("transpose_74")]; + tensor transpose_34 = transpose(perm = transpose_34_perm_0, x = var_349_cast_fp16)[name = string("transpose_75")]; + tensor var_358_cast_fp16 = matmul(transpose_x = var_358_transpose_x_0, transpose_y = var_358_transpose_y_0, x = transpose_34, y = transpose_35)[name = string("op_358_cast_fp16")]; + fp16 var_359_to_fp16 = const()[name = string("op_359_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_3_cast_fp16 = mul(x = var_358_cast_fp16, y = var_359_to_fp16)[name = string("scores_3_cast_fp16")]; + tensor weights_7_cast_fp16 = softmax(axis = var_24, x = scores_3_cast_fp16)[name = string("weights_7_cast_fp16")]; + tensor weights_9_cast_fp16 = mul(x = weights_7_cast_fp16, y = var_313_cast_fp16)[name = string("weights_9_cast_fp16")]; + tensor var_370_axes_0 = const()[name = string("op_370_axes_0"), val = tensor([-1])]; + bool var_370_keep_dims_0 = const()[name = string("op_370_keep_dims_0"), val = bool(true)]; + tensor var_370_cast_fp16 = reduce_sum(axes = var_370_axes_0, keep_dims = var_370_keep_dims_0, x = weights_9_cast_fp16)[name = string("op_370_cast_fp16")]; + tensor var_371_cast_fp16 = maximum(x = var_370_cast_fp16, y = var_11_to_fp16)[name = string("op_371_cast_fp16")]; + tensor weights_11_cast_fp16 = real_div(x = weights_9_cast_fp16, y = var_371_cast_fp16)[name = string("weights_11_cast_fp16")]; + bool context_5_transpose_x_0 = const()[name = string("context_5_transpose_x_0"), val = bool(false)]; + bool context_5_transpose_y_0 = const()[name = string("context_5_transpose_y_0"), val = bool(false)]; + tensor v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = var_355_cast_fp16)[name = string("transpose_73")]; + tensor context_5_cast_fp16 = matmul(transpose_x = context_5_transpose_x_0, transpose_y = context_5_transpose_y_0, x = weights_11_cast_fp16, y = v_7_cast_fp16)[name = string("context_5_cast_fp16")]; + tensor var_374_perm_0 = const()[name = string("op_374_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_375 = const()[name = string("op_375"), val = tensor([1, 128, 640])]; + tensor var_374_cast_fp16 = transpose(perm = var_374_perm_0, x = context_5_cast_fp16)[name = string("transpose_72")]; + tensor context_7_cast_fp16 = reshape(shape = var_375, x = var_374_cast_fp16)[name = string("context_7_cast_fp16")]; + tensor module_layers_1_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4836864))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5246528))))[name = string("module_layers_1_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_1_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_1_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5247872)))]; + tensor linear_5_cast_fp16 = linear(bias = module_layers_1_self_attn_out_proj_bias_to_fp16, weight = module_layers_1_self_attn_out_proj_weight_to_fp16_quantized, x = context_7_cast_fp16)[name = string("linear_5_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = x_5_cast_fp16, y = linear_5_cast_fp16)[name = string("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = string("input_15_axes_0"), val = tensor([-1])]; + tensor module_layers_1_norm1_weight_to_fp16 = const()[name = string("module_layers_1_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5249216)))]; + tensor module_layers_1_norm1_bias_to_fp16 = const()[name = string("module_layers_1_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5250560)))]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = module_layers_1_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_1_norm1_weight_to_fp16, x = input_13_cast_fp16)[name = string("input_15_cast_fp16")]; + tensor module_layers_1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5251904))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6071168))))[name = string("module_layers_1_linear1_weight_to_fp16_quantized")]; + tensor module_layers_1_linear1_bias_to_fp16 = const()[name = string("module_layers_1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6073792)))]; + tensor linear_6_cast_fp16 = linear(bias = module_layers_1_linear1_bias_to_fp16, weight = module_layers_1_linear1_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_6_cast_fp16")]; + tensor input_19_cast_fp16 = relu(x = linear_6_cast_fp16)[name = string("input_19_cast_fp16")]; + tensor module_layers_1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6076416))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6895680))))[name = string("module_layers_1_linear2_weight_to_fp16_quantized")]; + tensor module_layers_1_linear2_bias_to_fp16 = const()[name = string("module_layers_1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6897024)))]; + tensor linear_7_cast_fp16 = linear(bias = module_layers_1_linear2_bias_to_fp16, weight = module_layers_1_linear2_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_7_cast_fp16")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_7_cast_fp16)[name = string("input_21_cast_fp16")]; + tensor x_7_axes_0 = const()[name = string("x_7_axes_0"), val = tensor([-1])]; + tensor module_layers_1_norm2_weight_to_fp16 = const()[name = string("module_layers_1_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6898368)))]; + tensor module_layers_1_norm2_bias_to_fp16 = const()[name = string("module_layers_1_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6899712)))]; + tensor x_7_cast_fp16 = layer_norm(axes = x_7_axes_0, beta = module_layers_1_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_1_norm2_weight_to_fp16, x = input_21_cast_fp16)[name = string("x_7_cast_fp16")]; + tensor module_layers_2_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6901056))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8129920))))[name = string("module_layers_2_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_2_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_2_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8133824)))]; + tensor linear_8_cast_fp16 = linear(bias = module_layers_2_self_attn_in_proj_bias_to_fp16, weight = module_layers_2_self_attn_in_proj_weight_to_fp16_quantized, x = x_7_cast_fp16)[name = string("linear_8_cast_fp16")]; + tensor var_398_split_sizes_0 = const()[name = string("op_398_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_398_axis_0 = const()[name = string("op_398_axis_0"), val = int32(-1)]; + tensor var_398_cast_fp16_0, tensor var_398_cast_fp16_1, tensor var_398_cast_fp16_2 = split(axis = var_398_axis_0, split_sizes = var_398_split_sizes_0, x = linear_8_cast_fp16)[name = string("op_398_cast_fp16")]; + tensor var_401 = const()[name = string("op_401"), val = tensor([1, 128, 4, 160])]; + tensor var_402_cast_fp16 = reshape(shape = var_401, x = var_398_cast_fp16_0)[name = string("op_402_cast_fp16")]; + tensor var_404 = const()[name = string("op_404"), val = tensor([1, 128, 4, 160])]; + tensor var_405_cast_fp16 = reshape(shape = var_404, x = var_398_cast_fp16_1)[name = string("op_405_cast_fp16")]; + tensor var_407 = const()[name = string("op_407"), val = tensor([1, 128, 4, 160])]; + tensor var_408_cast_fp16 = reshape(shape = var_407, x = var_398_cast_fp16_2)[name = string("op_408_cast_fp16")]; + tensor v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_411_transpose_x_0 = const()[name = string("op_411_transpose_x_0"), val = bool(false)]; + bool var_411_transpose_y_0 = const()[name = string("op_411_transpose_y_0"), val = bool(false)]; + tensor transpose_36_perm_0 = const()[name = string("transpose_36_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_37_perm_0 = const()[name = string("transpose_37_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_37 = transpose(perm = transpose_37_perm_0, x = var_405_cast_fp16)[name = string("transpose_70")]; + tensor transpose_36 = transpose(perm = transpose_36_perm_0, x = var_402_cast_fp16)[name = string("transpose_71")]; + tensor var_411_cast_fp16 = matmul(transpose_x = var_411_transpose_x_0, transpose_y = var_411_transpose_y_0, x = transpose_36, y = transpose_37)[name = string("op_411_cast_fp16")]; + fp16 var_412_to_fp16 = const()[name = string("op_412_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_5_cast_fp16 = mul(x = var_411_cast_fp16, y = var_412_to_fp16)[name = string("scores_5_cast_fp16")]; + tensor weights_13_cast_fp16 = softmax(axis = var_24, x = scores_5_cast_fp16)[name = string("weights_13_cast_fp16")]; + tensor weights_15_cast_fp16 = mul(x = weights_13_cast_fp16, y = var_313_cast_fp16)[name = string("weights_15_cast_fp16")]; + tensor var_423_axes_0 = const()[name = string("op_423_axes_0"), val = tensor([-1])]; + bool var_423_keep_dims_0 = const()[name = string("op_423_keep_dims_0"), val = bool(true)]; + tensor var_423_cast_fp16 = reduce_sum(axes = var_423_axes_0, keep_dims = var_423_keep_dims_0, x = weights_15_cast_fp16)[name = string("op_423_cast_fp16")]; + tensor var_424_cast_fp16 = maximum(x = var_423_cast_fp16, y = var_11_to_fp16)[name = string("op_424_cast_fp16")]; + tensor weights_17_cast_fp16 = real_div(x = weights_15_cast_fp16, y = var_424_cast_fp16)[name = string("weights_17_cast_fp16")]; + bool context_9_transpose_x_0 = const()[name = string("context_9_transpose_x_0"), val = bool(false)]; + bool context_9_transpose_y_0 = const()[name = string("context_9_transpose_y_0"), val = bool(false)]; + tensor v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = var_408_cast_fp16)[name = string("transpose_69")]; + tensor context_9_cast_fp16 = matmul(transpose_x = context_9_transpose_x_0, transpose_y = context_9_transpose_y_0, x = weights_17_cast_fp16, y = v_11_cast_fp16)[name = string("context_9_cast_fp16")]; + tensor var_427_perm_0 = const()[name = string("op_427_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_428 = const()[name = string("op_428"), val = tensor([1, 128, 640])]; + tensor var_427_cast_fp16 = transpose(perm = var_427_perm_0, x = context_9_cast_fp16)[name = string("transpose_68")]; + tensor context_11_cast_fp16 = reshape(shape = var_428, x = var_427_cast_fp16)[name = string("context_11_cast_fp16")]; + tensor module_layers_2_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8137728))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8547392))))[name = string("module_layers_2_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_2_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_2_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8548736)))]; + tensor linear_9_cast_fp16 = linear(bias = module_layers_2_self_attn_out_proj_bias_to_fp16, weight = module_layers_2_self_attn_out_proj_weight_to_fp16_quantized, x = context_11_cast_fp16)[name = string("linear_9_cast_fp16")]; + tensor input_23_cast_fp16 = add(x = x_7_cast_fp16, y = linear_9_cast_fp16)[name = string("input_23_cast_fp16")]; + tensor input_25_axes_0 = const()[name = string("input_25_axes_0"), val = tensor([-1])]; + tensor module_layers_2_norm1_weight_to_fp16 = const()[name = string("module_layers_2_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8550080)))]; + tensor module_layers_2_norm1_bias_to_fp16 = const()[name = string("module_layers_2_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8551424)))]; + tensor input_25_cast_fp16 = layer_norm(axes = input_25_axes_0, beta = module_layers_2_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_2_norm1_weight_to_fp16, x = input_23_cast_fp16)[name = string("input_25_cast_fp16")]; + tensor module_layers_2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8552768))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9372032))))[name = string("module_layers_2_linear1_weight_to_fp16_quantized")]; + tensor module_layers_2_linear1_bias_to_fp16 = const()[name = string("module_layers_2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9374656)))]; + tensor linear_10_cast_fp16 = linear(bias = module_layers_2_linear1_bias_to_fp16, weight = module_layers_2_linear1_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_10_cast_fp16")]; + tensor input_29_cast_fp16 = relu(x = linear_10_cast_fp16)[name = string("input_29_cast_fp16")]; + tensor module_layers_2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9377280))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10196544))))[name = string("module_layers_2_linear2_weight_to_fp16_quantized")]; + tensor module_layers_2_linear2_bias_to_fp16 = const()[name = string("module_layers_2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10197888)))]; + tensor linear_11_cast_fp16 = linear(bias = module_layers_2_linear2_bias_to_fp16, weight = module_layers_2_linear2_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_11_cast_fp16")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_11_cast_fp16)[name = string("input_31_cast_fp16")]; + tensor x_9_axes_0 = const()[name = string("x_9_axes_0"), val = tensor([-1])]; + tensor module_layers_2_norm2_weight_to_fp16 = const()[name = string("module_layers_2_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10199232)))]; + tensor module_layers_2_norm2_bias_to_fp16 = const()[name = string("module_layers_2_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10200576)))]; + tensor x_9_cast_fp16 = layer_norm(axes = x_9_axes_0, beta = module_layers_2_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_2_norm2_weight_to_fp16, x = input_31_cast_fp16)[name = string("x_9_cast_fp16")]; + tensor module_layers_3_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10201920))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11430784))))[name = string("module_layers_3_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_3_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_3_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11434688)))]; + tensor linear_12_cast_fp16 = linear(bias = module_layers_3_self_attn_in_proj_bias_to_fp16, weight = module_layers_3_self_attn_in_proj_weight_to_fp16_quantized, x = x_9_cast_fp16)[name = string("linear_12_cast_fp16")]; + tensor var_451_split_sizes_0 = const()[name = string("op_451_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_451_axis_0 = const()[name = string("op_451_axis_0"), val = int32(-1)]; + tensor var_451_cast_fp16_0, tensor var_451_cast_fp16_1, tensor var_451_cast_fp16_2 = split(axis = var_451_axis_0, split_sizes = var_451_split_sizes_0, x = linear_12_cast_fp16)[name = string("op_451_cast_fp16")]; + tensor var_454 = const()[name = string("op_454"), val = tensor([1, 128, 4, 160])]; + tensor var_455_cast_fp16 = reshape(shape = var_454, x = var_451_cast_fp16_0)[name = string("op_455_cast_fp16")]; + tensor var_457 = const()[name = string("op_457"), val = tensor([1, 128, 4, 160])]; + tensor var_458_cast_fp16 = reshape(shape = var_457, x = var_451_cast_fp16_1)[name = string("op_458_cast_fp16")]; + tensor var_460 = const()[name = string("op_460"), val = tensor([1, 128, 4, 160])]; + tensor var_461_cast_fp16 = reshape(shape = var_460, x = var_451_cast_fp16_2)[name = string("op_461_cast_fp16")]; + tensor v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_464_transpose_x_0 = const()[name = string("op_464_transpose_x_0"), val = bool(false)]; + bool var_464_transpose_y_0 = const()[name = string("op_464_transpose_y_0"), val = bool(false)]; + tensor transpose_38_perm_0 = const()[name = string("transpose_38_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_39_perm_0 = const()[name = string("transpose_39_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_39 = transpose(perm = transpose_39_perm_0, x = var_458_cast_fp16)[name = string("transpose_66")]; + tensor transpose_38 = transpose(perm = transpose_38_perm_0, x = var_455_cast_fp16)[name = string("transpose_67")]; + tensor var_464_cast_fp16 = matmul(transpose_x = var_464_transpose_x_0, transpose_y = var_464_transpose_y_0, x = transpose_38, y = transpose_39)[name = string("op_464_cast_fp16")]; + fp16 var_465_to_fp16 = const()[name = string("op_465_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_7_cast_fp16 = mul(x = var_464_cast_fp16, y = var_465_to_fp16)[name = string("scores_7_cast_fp16")]; + tensor weights_19_cast_fp16 = softmax(axis = var_24, x = scores_7_cast_fp16)[name = string("weights_19_cast_fp16")]; + tensor weights_21_cast_fp16 = mul(x = weights_19_cast_fp16, y = var_313_cast_fp16)[name = string("weights_21_cast_fp16")]; + tensor var_476_axes_0 = const()[name = string("op_476_axes_0"), val = tensor([-1])]; + bool var_476_keep_dims_0 = const()[name = string("op_476_keep_dims_0"), val = bool(true)]; + tensor var_476_cast_fp16 = reduce_sum(axes = var_476_axes_0, keep_dims = var_476_keep_dims_0, x = weights_21_cast_fp16)[name = string("op_476_cast_fp16")]; + tensor var_477_cast_fp16 = maximum(x = var_476_cast_fp16, y = var_11_to_fp16)[name = string("op_477_cast_fp16")]; + tensor weights_23_cast_fp16 = real_div(x = weights_21_cast_fp16, y = var_477_cast_fp16)[name = string("weights_23_cast_fp16")]; + bool context_13_transpose_x_0 = const()[name = string("context_13_transpose_x_0"), val = bool(false)]; + bool context_13_transpose_y_0 = const()[name = string("context_13_transpose_y_0"), val = bool(false)]; + tensor v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = var_461_cast_fp16)[name = string("transpose_65")]; + tensor context_13_cast_fp16 = matmul(transpose_x = context_13_transpose_x_0, transpose_y = context_13_transpose_y_0, x = weights_23_cast_fp16, y = v_15_cast_fp16)[name = string("context_13_cast_fp16")]; + tensor var_480_perm_0 = const()[name = string("op_480_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_481 = const()[name = string("op_481"), val = tensor([1, 128, 640])]; + tensor var_480_cast_fp16 = transpose(perm = var_480_perm_0, x = context_13_cast_fp16)[name = string("transpose_64")]; + tensor context_15_cast_fp16 = reshape(shape = var_481, x = var_480_cast_fp16)[name = string("context_15_cast_fp16")]; + tensor module_layers_3_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11438592))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11848256))))[name = string("module_layers_3_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_3_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_3_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11849600)))]; + tensor linear_13_cast_fp16 = linear(bias = module_layers_3_self_attn_out_proj_bias_to_fp16, weight = module_layers_3_self_attn_out_proj_weight_to_fp16_quantized, x = context_15_cast_fp16)[name = string("linear_13_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = x_9_cast_fp16, y = linear_13_cast_fp16)[name = string("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = string("input_35_axes_0"), val = tensor([-1])]; + tensor module_layers_3_norm1_weight_to_fp16 = const()[name = string("module_layers_3_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11850944)))]; + tensor module_layers_3_norm1_bias_to_fp16 = const()[name = string("module_layers_3_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11852288)))]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = module_layers_3_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_3_norm1_weight_to_fp16, x = input_33_cast_fp16)[name = string("input_35_cast_fp16")]; + tensor module_layers_3_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11853632))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12672896))))[name = string("module_layers_3_linear1_weight_to_fp16_quantized")]; + tensor module_layers_3_linear1_bias_to_fp16 = const()[name = string("module_layers_3_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12675520)))]; + tensor linear_14_cast_fp16 = linear(bias = module_layers_3_linear1_bias_to_fp16, weight = module_layers_3_linear1_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_14_cast_fp16")]; + tensor input_39_cast_fp16 = relu(x = linear_14_cast_fp16)[name = string("input_39_cast_fp16")]; + tensor module_layers_3_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12678144))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13497408))))[name = string("module_layers_3_linear2_weight_to_fp16_quantized")]; + tensor module_layers_3_linear2_bias_to_fp16 = const()[name = string("module_layers_3_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13498752)))]; + tensor linear_15_cast_fp16 = linear(bias = module_layers_3_linear2_bias_to_fp16, weight = module_layers_3_linear2_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_15_cast_fp16")]; + tensor input_41_cast_fp16 = add(x = input_35_cast_fp16, y = linear_15_cast_fp16)[name = string("input_41_cast_fp16")]; + tensor x_11_axes_0 = const()[name = string("x_11_axes_0"), val = tensor([-1])]; + tensor module_layers_3_norm2_weight_to_fp16 = const()[name = string("module_layers_3_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13500096)))]; + tensor module_layers_3_norm2_bias_to_fp16 = const()[name = string("module_layers_3_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13501440)))]; + tensor x_11_cast_fp16 = layer_norm(axes = x_11_axes_0, beta = module_layers_3_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_3_norm2_weight_to_fp16, x = input_41_cast_fp16)[name = string("x_11_cast_fp16")]; + tensor module_layers_4_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13502784))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14731648))))[name = string("module_layers_4_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_4_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_4_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14735552)))]; + tensor linear_16_cast_fp16 = linear(bias = module_layers_4_self_attn_in_proj_bias_to_fp16, weight = module_layers_4_self_attn_in_proj_weight_to_fp16_quantized, x = x_11_cast_fp16)[name = string("linear_16_cast_fp16")]; + tensor var_504_split_sizes_0 = const()[name = string("op_504_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_504_axis_0 = const()[name = string("op_504_axis_0"), val = int32(-1)]; + tensor var_504_cast_fp16_0, tensor var_504_cast_fp16_1, tensor var_504_cast_fp16_2 = split(axis = var_504_axis_0, split_sizes = var_504_split_sizes_0, x = linear_16_cast_fp16)[name = string("op_504_cast_fp16")]; + tensor var_507 = const()[name = string("op_507"), val = tensor([1, 128, 4, 160])]; + tensor var_508_cast_fp16 = reshape(shape = var_507, x = var_504_cast_fp16_0)[name = string("op_508_cast_fp16")]; + tensor var_510 = const()[name = string("op_510"), val = tensor([1, 128, 4, 160])]; + tensor var_511_cast_fp16 = reshape(shape = var_510, x = var_504_cast_fp16_1)[name = string("op_511_cast_fp16")]; + tensor var_513 = const()[name = string("op_513"), val = tensor([1, 128, 4, 160])]; + tensor var_514_cast_fp16 = reshape(shape = var_513, x = var_504_cast_fp16_2)[name = string("op_514_cast_fp16")]; + tensor v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_517_transpose_x_0 = const()[name = string("op_517_transpose_x_0"), val = bool(false)]; + bool var_517_transpose_y_0 = const()[name = string("op_517_transpose_y_0"), val = bool(false)]; + tensor transpose_40_perm_0 = const()[name = string("transpose_40_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_41_perm_0 = const()[name = string("transpose_41_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_41 = transpose(perm = transpose_41_perm_0, x = var_511_cast_fp16)[name = string("transpose_62")]; + tensor transpose_40 = transpose(perm = transpose_40_perm_0, x = var_508_cast_fp16)[name = string("transpose_63")]; + tensor var_517_cast_fp16 = matmul(transpose_x = var_517_transpose_x_0, transpose_y = var_517_transpose_y_0, x = transpose_40, y = transpose_41)[name = string("op_517_cast_fp16")]; + fp16 var_518_to_fp16 = const()[name = string("op_518_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_9_cast_fp16 = mul(x = var_517_cast_fp16, y = var_518_to_fp16)[name = string("scores_9_cast_fp16")]; + tensor weights_25_cast_fp16 = softmax(axis = var_24, x = scores_9_cast_fp16)[name = string("weights_25_cast_fp16")]; + tensor weights_27_cast_fp16 = mul(x = weights_25_cast_fp16, y = var_313_cast_fp16)[name = string("weights_27_cast_fp16")]; + tensor var_529_axes_0 = const()[name = string("op_529_axes_0"), val = tensor([-1])]; + bool var_529_keep_dims_0 = const()[name = string("op_529_keep_dims_0"), val = bool(true)]; + tensor var_529_cast_fp16 = reduce_sum(axes = var_529_axes_0, keep_dims = var_529_keep_dims_0, x = weights_27_cast_fp16)[name = string("op_529_cast_fp16")]; + tensor var_530_cast_fp16 = maximum(x = var_529_cast_fp16, y = var_11_to_fp16)[name = string("op_530_cast_fp16")]; + tensor weights_29_cast_fp16 = real_div(x = weights_27_cast_fp16, y = var_530_cast_fp16)[name = string("weights_29_cast_fp16")]; + bool context_17_transpose_x_0 = const()[name = string("context_17_transpose_x_0"), val = bool(false)]; + bool context_17_transpose_y_0 = const()[name = string("context_17_transpose_y_0"), val = bool(false)]; + tensor v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = var_514_cast_fp16)[name = string("transpose_61")]; + tensor context_17_cast_fp16 = matmul(transpose_x = context_17_transpose_x_0, transpose_y = context_17_transpose_y_0, x = weights_29_cast_fp16, y = v_19_cast_fp16)[name = string("context_17_cast_fp16")]; + tensor var_533_perm_0 = const()[name = string("op_533_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_534 = const()[name = string("op_534"), val = tensor([1, 128, 640])]; + tensor var_533_cast_fp16 = transpose(perm = var_533_perm_0, x = context_17_cast_fp16)[name = string("transpose_60")]; + tensor context_19_cast_fp16 = reshape(shape = var_534, x = var_533_cast_fp16)[name = string("context_19_cast_fp16")]; + tensor module_layers_4_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14739456))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15149120))))[name = string("module_layers_4_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_4_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_4_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15150464)))]; + tensor linear_17_cast_fp16 = linear(bias = module_layers_4_self_attn_out_proj_bias_to_fp16, weight = module_layers_4_self_attn_out_proj_weight_to_fp16_quantized, x = context_19_cast_fp16)[name = string("linear_17_cast_fp16")]; + tensor input_43_cast_fp16 = add(x = x_11_cast_fp16, y = linear_17_cast_fp16)[name = string("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = string("input_45_axes_0"), val = tensor([-1])]; + tensor module_layers_4_norm1_weight_to_fp16 = const()[name = string("module_layers_4_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15151808)))]; + tensor module_layers_4_norm1_bias_to_fp16 = const()[name = string("module_layers_4_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15153152)))]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = module_layers_4_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_4_norm1_weight_to_fp16, x = input_43_cast_fp16)[name = string("input_45_cast_fp16")]; + tensor module_layers_4_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15154496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15973760))))[name = string("module_layers_4_linear1_weight_to_fp16_quantized")]; + tensor module_layers_4_linear1_bias_to_fp16 = const()[name = string("module_layers_4_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15976384)))]; + tensor linear_18_cast_fp16 = linear(bias = module_layers_4_linear1_bias_to_fp16, weight = module_layers_4_linear1_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_18_cast_fp16")]; + tensor input_49_cast_fp16 = relu(x = linear_18_cast_fp16)[name = string("input_49_cast_fp16")]; + tensor module_layers_4_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15979008))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16798272))))[name = string("module_layers_4_linear2_weight_to_fp16_quantized")]; + tensor module_layers_4_linear2_bias_to_fp16 = const()[name = string("module_layers_4_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16799616)))]; + tensor linear_19_cast_fp16 = linear(bias = module_layers_4_linear2_bias_to_fp16, weight = module_layers_4_linear2_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_19_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_19_cast_fp16)[name = string("input_51_cast_fp16")]; + tensor x_13_axes_0 = const()[name = string("x_13_axes_0"), val = tensor([-1])]; + tensor module_layers_4_norm2_weight_to_fp16 = const()[name = string("module_layers_4_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16800960)))]; + tensor module_layers_4_norm2_bias_to_fp16 = const()[name = string("module_layers_4_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16802304)))]; + tensor x_13_cast_fp16 = layer_norm(axes = x_13_axes_0, beta = module_layers_4_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_4_norm2_weight_to_fp16, x = input_51_cast_fp16)[name = string("x_13_cast_fp16")]; + tensor module_layers_5_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16803648))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18032512))))[name = string("module_layers_5_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_5_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_5_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18036416)))]; + tensor linear_20_cast_fp16 = linear(bias = module_layers_5_self_attn_in_proj_bias_to_fp16, weight = module_layers_5_self_attn_in_proj_weight_to_fp16_quantized, x = x_13_cast_fp16)[name = string("linear_20_cast_fp16")]; + tensor var_557_split_sizes_0 = const()[name = string("op_557_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_557_axis_0 = const()[name = string("op_557_axis_0"), val = int32(-1)]; + tensor var_557_cast_fp16_0, tensor var_557_cast_fp16_1, tensor var_557_cast_fp16_2 = split(axis = var_557_axis_0, split_sizes = var_557_split_sizes_0, x = linear_20_cast_fp16)[name = string("op_557_cast_fp16")]; + tensor var_560 = const()[name = string("op_560"), val = tensor([1, 128, 4, 160])]; + tensor var_561_cast_fp16 = reshape(shape = var_560, x = var_557_cast_fp16_0)[name = string("op_561_cast_fp16")]; + tensor var_563 = const()[name = string("op_563"), val = tensor([1, 128, 4, 160])]; + tensor var_564_cast_fp16 = reshape(shape = var_563, x = var_557_cast_fp16_1)[name = string("op_564_cast_fp16")]; + tensor var_566 = const()[name = string("op_566"), val = tensor([1, 128, 4, 160])]; + tensor var_567_cast_fp16 = reshape(shape = var_566, x = var_557_cast_fp16_2)[name = string("op_567_cast_fp16")]; + tensor v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_570_transpose_x_0 = const()[name = string("op_570_transpose_x_0"), val = bool(false)]; + bool var_570_transpose_y_0 = const()[name = string("op_570_transpose_y_0"), val = bool(false)]; + tensor transpose_42_perm_0 = const()[name = string("transpose_42_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_43_perm_0 = const()[name = string("transpose_43_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_43 = transpose(perm = transpose_43_perm_0, x = var_564_cast_fp16)[name = string("transpose_58")]; + tensor transpose_42 = transpose(perm = transpose_42_perm_0, x = var_561_cast_fp16)[name = string("transpose_59")]; + tensor var_570_cast_fp16 = matmul(transpose_x = var_570_transpose_x_0, transpose_y = var_570_transpose_y_0, x = transpose_42, y = transpose_43)[name = string("op_570_cast_fp16")]; + fp16 var_571_to_fp16 = const()[name = string("op_571_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_11_cast_fp16 = mul(x = var_570_cast_fp16, y = var_571_to_fp16)[name = string("scores_11_cast_fp16")]; + tensor weights_31_cast_fp16 = softmax(axis = var_24, x = scores_11_cast_fp16)[name = string("weights_31_cast_fp16")]; + tensor weights_33_cast_fp16 = mul(x = weights_31_cast_fp16, y = var_313_cast_fp16)[name = string("weights_33_cast_fp16")]; + tensor var_582_axes_0 = const()[name = string("op_582_axes_0"), val = tensor([-1])]; + bool var_582_keep_dims_0 = const()[name = string("op_582_keep_dims_0"), val = bool(true)]; + tensor var_582_cast_fp16 = reduce_sum(axes = var_582_axes_0, keep_dims = var_582_keep_dims_0, x = weights_33_cast_fp16)[name = string("op_582_cast_fp16")]; + tensor var_583_cast_fp16 = maximum(x = var_582_cast_fp16, y = var_11_to_fp16)[name = string("op_583_cast_fp16")]; + tensor weights_35_cast_fp16 = real_div(x = weights_33_cast_fp16, y = var_583_cast_fp16)[name = string("weights_35_cast_fp16")]; + bool context_21_transpose_x_0 = const()[name = string("context_21_transpose_x_0"), val = bool(false)]; + bool context_21_transpose_y_0 = const()[name = string("context_21_transpose_y_0"), val = bool(false)]; + tensor v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = var_567_cast_fp16)[name = string("transpose_57")]; + tensor context_21_cast_fp16 = matmul(transpose_x = context_21_transpose_x_0, transpose_y = context_21_transpose_y_0, x = weights_35_cast_fp16, y = v_23_cast_fp16)[name = string("context_21_cast_fp16")]; + tensor var_586_perm_0 = const()[name = string("op_586_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_587 = const()[name = string("op_587"), val = tensor([1, 128, 640])]; + tensor var_586_cast_fp16 = transpose(perm = var_586_perm_0, x = context_21_cast_fp16)[name = string("transpose_56")]; + tensor context_23_cast_fp16 = reshape(shape = var_587, x = var_586_cast_fp16)[name = string("context_23_cast_fp16")]; + tensor module_layers_5_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18040320))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18449984))))[name = string("module_layers_5_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_5_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_5_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18451328)))]; + tensor linear_21_cast_fp16 = linear(bias = module_layers_5_self_attn_out_proj_bias_to_fp16, weight = module_layers_5_self_attn_out_proj_weight_to_fp16_quantized, x = context_23_cast_fp16)[name = string("linear_21_cast_fp16")]; + tensor input_53_cast_fp16 = add(x = x_13_cast_fp16, y = linear_21_cast_fp16)[name = string("input_53_cast_fp16")]; + tensor input_55_axes_0 = const()[name = string("input_55_axes_0"), val = tensor([-1])]; + tensor module_layers_5_norm1_weight_to_fp16 = const()[name = string("module_layers_5_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18452672)))]; + tensor module_layers_5_norm1_bias_to_fp16 = const()[name = string("module_layers_5_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18454016)))]; + tensor input_55_cast_fp16 = layer_norm(axes = input_55_axes_0, beta = module_layers_5_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_5_norm1_weight_to_fp16, x = input_53_cast_fp16)[name = string("input_55_cast_fp16")]; + tensor module_layers_5_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18455360))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19274624))))[name = string("module_layers_5_linear1_weight_to_fp16_quantized")]; + tensor module_layers_5_linear1_bias_to_fp16 = const()[name = string("module_layers_5_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19277248)))]; + tensor linear_22_cast_fp16 = linear(bias = module_layers_5_linear1_bias_to_fp16, weight = module_layers_5_linear1_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_22_cast_fp16")]; + tensor input_59_cast_fp16 = relu(x = linear_22_cast_fp16)[name = string("input_59_cast_fp16")]; + tensor module_layers_5_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19279872))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20099136))))[name = string("module_layers_5_linear2_weight_to_fp16_quantized")]; + tensor module_layers_5_linear2_bias_to_fp16 = const()[name = string("module_layers_5_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20100480)))]; + tensor linear_23_cast_fp16 = linear(bias = module_layers_5_linear2_bias_to_fp16, weight = module_layers_5_linear2_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_23_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_23_cast_fp16)[name = string("input_61_cast_fp16")]; + tensor x_15_axes_0 = const()[name = string("x_15_axes_0"), val = tensor([-1])]; + tensor module_layers_5_norm2_weight_to_fp16 = const()[name = string("module_layers_5_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20101824)))]; + tensor module_layers_5_norm2_bias_to_fp16 = const()[name = string("module_layers_5_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20103168)))]; + tensor x_15_cast_fp16 = layer_norm(axes = x_15_axes_0, beta = module_layers_5_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_5_norm2_weight_to_fp16, x = input_61_cast_fp16)[name = string("x_15_cast_fp16")]; + tensor module_layers_6_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20104512))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21333376))))[name = string("module_layers_6_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_6_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_6_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21337280)))]; + tensor linear_24_cast_fp16 = linear(bias = module_layers_6_self_attn_in_proj_bias_to_fp16, weight = module_layers_6_self_attn_in_proj_weight_to_fp16_quantized, x = x_15_cast_fp16)[name = string("linear_24_cast_fp16")]; + tensor var_610_split_sizes_0 = const()[name = string("op_610_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_610_axis_0 = const()[name = string("op_610_axis_0"), val = int32(-1)]; + tensor var_610_cast_fp16_0, tensor var_610_cast_fp16_1, tensor var_610_cast_fp16_2 = split(axis = var_610_axis_0, split_sizes = var_610_split_sizes_0, x = linear_24_cast_fp16)[name = string("op_610_cast_fp16")]; + tensor var_613 = const()[name = string("op_613"), val = tensor([1, 128, 4, 160])]; + tensor var_614_cast_fp16 = reshape(shape = var_613, x = var_610_cast_fp16_0)[name = string("op_614_cast_fp16")]; + tensor var_616 = const()[name = string("op_616"), val = tensor([1, 128, 4, 160])]; + tensor var_617_cast_fp16 = reshape(shape = var_616, x = var_610_cast_fp16_1)[name = string("op_617_cast_fp16")]; + tensor var_619 = const()[name = string("op_619"), val = tensor([1, 128, 4, 160])]; + tensor var_620_cast_fp16 = reshape(shape = var_619, x = var_610_cast_fp16_2)[name = string("op_620_cast_fp16")]; + tensor v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_623_transpose_x_0 = const()[name = string("op_623_transpose_x_0"), val = bool(false)]; + bool var_623_transpose_y_0 = const()[name = string("op_623_transpose_y_0"), val = bool(false)]; + tensor transpose_44_perm_0 = const()[name = string("transpose_44_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_45_perm_0 = const()[name = string("transpose_45_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_45 = transpose(perm = transpose_45_perm_0, x = var_617_cast_fp16)[name = string("transpose_54")]; + tensor transpose_44 = transpose(perm = transpose_44_perm_0, x = var_614_cast_fp16)[name = string("transpose_55")]; + tensor var_623_cast_fp16 = matmul(transpose_x = var_623_transpose_x_0, transpose_y = var_623_transpose_y_0, x = transpose_44, y = transpose_45)[name = string("op_623_cast_fp16")]; + fp16 var_624_to_fp16 = const()[name = string("op_624_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_13_cast_fp16 = mul(x = var_623_cast_fp16, y = var_624_to_fp16)[name = string("scores_13_cast_fp16")]; + tensor weights_37_cast_fp16 = softmax(axis = var_24, x = scores_13_cast_fp16)[name = string("weights_37_cast_fp16")]; + tensor weights_39_cast_fp16 = mul(x = weights_37_cast_fp16, y = var_313_cast_fp16)[name = string("weights_39_cast_fp16")]; + tensor var_635_axes_0 = const()[name = string("op_635_axes_0"), val = tensor([-1])]; + bool var_635_keep_dims_0 = const()[name = string("op_635_keep_dims_0"), val = bool(true)]; + tensor var_635_cast_fp16 = reduce_sum(axes = var_635_axes_0, keep_dims = var_635_keep_dims_0, x = weights_39_cast_fp16)[name = string("op_635_cast_fp16")]; + tensor var_636_cast_fp16 = maximum(x = var_635_cast_fp16, y = var_11_to_fp16)[name = string("op_636_cast_fp16")]; + tensor weights_41_cast_fp16 = real_div(x = weights_39_cast_fp16, y = var_636_cast_fp16)[name = string("weights_41_cast_fp16")]; + bool context_25_transpose_x_0 = const()[name = string("context_25_transpose_x_0"), val = bool(false)]; + bool context_25_transpose_y_0 = const()[name = string("context_25_transpose_y_0"), val = bool(false)]; + tensor v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = var_620_cast_fp16)[name = string("transpose_53")]; + tensor context_25_cast_fp16 = matmul(transpose_x = context_25_transpose_x_0, transpose_y = context_25_transpose_y_0, x = weights_41_cast_fp16, y = v_27_cast_fp16)[name = string("context_25_cast_fp16")]; + tensor var_639_perm_0 = const()[name = string("op_639_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_640 = const()[name = string("op_640"), val = tensor([1, 128, 640])]; + tensor var_639_cast_fp16 = transpose(perm = var_639_perm_0, x = context_25_cast_fp16)[name = string("transpose_52")]; + tensor context_27_cast_fp16 = reshape(shape = var_640, x = var_639_cast_fp16)[name = string("context_27_cast_fp16")]; + tensor module_layers_6_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21341184))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21750848))))[name = string("module_layers_6_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_6_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_6_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21752192)))]; + tensor linear_25_cast_fp16 = linear(bias = module_layers_6_self_attn_out_proj_bias_to_fp16, weight = module_layers_6_self_attn_out_proj_weight_to_fp16_quantized, x = context_27_cast_fp16)[name = string("linear_25_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = x_15_cast_fp16, y = linear_25_cast_fp16)[name = string("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor([-1])]; + tensor module_layers_6_norm1_weight_to_fp16 = const()[name = string("module_layers_6_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21753536)))]; + tensor module_layers_6_norm1_bias_to_fp16 = const()[name = string("module_layers_6_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21754880)))]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = module_layers_6_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_6_norm1_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")]; + tensor module_layers_6_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21756224))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22575488))))[name = string("module_layers_6_linear1_weight_to_fp16_quantized")]; + tensor module_layers_6_linear1_bias_to_fp16 = const()[name = string("module_layers_6_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22578112)))]; + tensor linear_26_cast_fp16 = linear(bias = module_layers_6_linear1_bias_to_fp16, weight = module_layers_6_linear1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_26_cast_fp16")]; + tensor input_69_cast_fp16 = relu(x = linear_26_cast_fp16)[name = string("input_69_cast_fp16")]; + tensor module_layers_6_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22580736))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23400000))))[name = string("module_layers_6_linear2_weight_to_fp16_quantized")]; + tensor module_layers_6_linear2_bias_to_fp16 = const()[name = string("module_layers_6_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23401344)))]; + tensor linear_27_cast_fp16 = linear(bias = module_layers_6_linear2_bias_to_fp16, weight = module_layers_6_linear2_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_27_cast_fp16")]; + tensor input_71_cast_fp16 = add(x = input_65_cast_fp16, y = linear_27_cast_fp16)[name = string("input_71_cast_fp16")]; + tensor x_axes_0 = const()[name = string("x_axes_0"), val = tensor([-1])]; + tensor module_layers_6_norm2_weight_to_fp16 = const()[name = string("module_layers_6_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23402688)))]; + tensor module_layers_6_norm2_bias_to_fp16 = const()[name = string("module_layers_6_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23404032)))]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = module_layers_6_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_6_norm2_weight_to_fp16, x = input_71_cast_fp16)[name = string("x_cast_fp16")]; + tensor module_layers_7_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23405376))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24634240))))[name = string("module_layers_7_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_7_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_7_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24638144)))]; + tensor linear_28_cast_fp16 = linear(bias = module_layers_7_self_attn_in_proj_bias_to_fp16, weight = module_layers_7_self_attn_in_proj_weight_to_fp16_quantized, x = x_cast_fp16)[name = string("linear_28_cast_fp16")]; + tensor var_663_split_sizes_0 = const()[name = string("op_663_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_663_axis_0 = const()[name = string("op_663_axis_0"), val = int32(-1)]; + tensor var_663_cast_fp16_0, tensor var_663_cast_fp16_1, tensor var_663_cast_fp16_2 = split(axis = var_663_axis_0, split_sizes = var_663_split_sizes_0, x = linear_28_cast_fp16)[name = string("op_663_cast_fp16")]; + tensor var_666 = const()[name = string("op_666"), val = tensor([1, 128, 4, 160])]; + tensor var_667_cast_fp16 = reshape(shape = var_666, x = var_663_cast_fp16_0)[name = string("op_667_cast_fp16")]; + tensor var_669 = const()[name = string("op_669"), val = tensor([1, 128, 4, 160])]; + tensor var_670_cast_fp16 = reshape(shape = var_669, x = var_663_cast_fp16_1)[name = string("op_670_cast_fp16")]; + tensor var_672 = const()[name = string("op_672"), val = tensor([1, 128, 4, 160])]; + tensor var_673_cast_fp16 = reshape(shape = var_672, x = var_663_cast_fp16_2)[name = string("op_673_cast_fp16")]; + tensor v_perm_0 = const()[name = string("v_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_676_transpose_x_0 = const()[name = string("op_676_transpose_x_0"), val = bool(false)]; + bool var_676_transpose_y_0 = const()[name = string("op_676_transpose_y_0"), val = bool(false)]; + tensor transpose_46_perm_0 = const()[name = string("transpose_46_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_47_perm_0 = const()[name = string("transpose_47_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_47 = transpose(perm = transpose_47_perm_0, x = var_670_cast_fp16)[name = string("transpose_50")]; + tensor transpose_46 = transpose(perm = transpose_46_perm_0, x = var_667_cast_fp16)[name = string("transpose_51")]; + tensor var_676_cast_fp16 = matmul(transpose_x = var_676_transpose_x_0, transpose_y = var_676_transpose_y_0, x = transpose_46, y = transpose_47)[name = string("op_676_cast_fp16")]; + fp16 var_677_to_fp16 = const()[name = string("op_677_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_cast_fp16 = mul(x = var_676_cast_fp16, y = var_677_to_fp16)[name = string("scores_cast_fp16")]; + tensor weights_43_cast_fp16 = softmax(axis = var_24, x = scores_cast_fp16)[name = string("weights_43_cast_fp16")]; + tensor weights_45_cast_fp16 = mul(x = weights_43_cast_fp16, y = var_313_cast_fp16)[name = string("weights_45_cast_fp16")]; + tensor var_688_axes_0 = const()[name = string("op_688_axes_0"), val = tensor([-1])]; + bool var_688_keep_dims_0 = const()[name = string("op_688_keep_dims_0"), val = bool(true)]; + tensor var_688_cast_fp16 = reduce_sum(axes = var_688_axes_0, keep_dims = var_688_keep_dims_0, x = weights_45_cast_fp16)[name = string("op_688_cast_fp16")]; + tensor var_689_cast_fp16 = maximum(x = var_688_cast_fp16, y = var_11_to_fp16)[name = string("op_689_cast_fp16")]; + tensor weights_cast_fp16 = real_div(x = weights_45_cast_fp16, y = var_689_cast_fp16)[name = string("weights_cast_fp16")]; + bool context_29_transpose_x_0 = const()[name = string("context_29_transpose_x_0"), val = bool(false)]; + bool context_29_transpose_y_0 = const()[name = string("context_29_transpose_y_0"), val = bool(false)]; + tensor v_cast_fp16 = transpose(perm = v_perm_0, x = var_673_cast_fp16)[name = string("transpose_49")]; + tensor context_29_cast_fp16 = matmul(transpose_x = context_29_transpose_x_0, transpose_y = context_29_transpose_y_0, x = weights_cast_fp16, y = v_cast_fp16)[name = string("context_29_cast_fp16")]; + tensor var_692_perm_0 = const()[name = string("op_692_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_693 = const()[name = string("op_693"), val = tensor([1, 128, 640])]; + tensor var_692_cast_fp16 = transpose(perm = var_692_perm_0, x = context_29_cast_fp16)[name = string("transpose_48")]; + tensor context_cast_fp16 = reshape(shape = var_693, x = var_692_cast_fp16)[name = string("context_cast_fp16")]; + tensor module_layers_7_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24642048))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25051712))))[name = string("module_layers_7_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_7_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_7_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25053056)))]; + tensor linear_29_cast_fp16 = linear(bias = module_layers_7_self_attn_out_proj_bias_to_fp16, weight = module_layers_7_self_attn_out_proj_weight_to_fp16_quantized, x = context_cast_fp16)[name = string("linear_29_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = x_cast_fp16, y = linear_29_cast_fp16)[name = string("input_73_cast_fp16")]; + tensor input_75_axes_0 = const()[name = string("input_75_axes_0"), val = tensor([-1])]; + tensor module_layers_7_norm1_weight_to_fp16 = const()[name = string("module_layers_7_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25054400)))]; + tensor module_layers_7_norm1_bias_to_fp16 = const()[name = string("module_layers_7_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25055744)))]; + tensor input_75_cast_fp16 = layer_norm(axes = input_75_axes_0, beta = module_layers_7_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_7_norm1_weight_to_fp16, x = input_73_cast_fp16)[name = string("input_75_cast_fp16")]; + tensor module_layers_7_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25057088))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25876352))))[name = string("module_layers_7_linear1_weight_to_fp16_quantized")]; + tensor module_layers_7_linear1_bias_to_fp16 = const()[name = string("module_layers_7_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25878976)))]; + tensor linear_30_cast_fp16 = linear(bias = module_layers_7_linear1_bias_to_fp16, weight = module_layers_7_linear1_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_30_cast_fp16")]; + tensor input_79_cast_fp16 = relu(x = linear_30_cast_fp16)[name = string("input_79_cast_fp16")]; + tensor module_layers_7_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25881600))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26700864))))[name = string("module_layers_7_linear2_weight_to_fp16_quantized")]; + tensor module_layers_7_linear2_bias_to_fp16 = const()[name = string("module_layers_7_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26702208)))]; + tensor linear_31_cast_fp16 = linear(bias = module_layers_7_linear2_bias_to_fp16, weight = module_layers_7_linear2_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_31_cast_fp16")]; + tensor input_81_cast_fp16 = add(x = input_75_cast_fp16, y = linear_31_cast_fp16)[name = string("input_81_cast_fp16")]; + tensor input_83_axes_0 = const()[name = string("input_83_axes_0"), val = tensor([-1])]; + tensor module_layers_7_norm2_weight_to_fp16 = const()[name = string("module_layers_7_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26703552)))]; + tensor module_layers_7_norm2_bias_to_fp16 = const()[name = string("module_layers_7_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26704896)))]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = module_layers_7_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_7_norm2_weight_to_fp16, x = input_81_cast_fp16)[name = string("input_83_cast_fp16")]; + tensor input_axes_0 = const()[name = string("input_axes_0"), val = tensor([-1])]; + tensor module_encoder_norm_weight_to_fp16 = const()[name = string("module_encoder_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26706240)))]; + tensor module_encoder_norm_bias_to_fp16 = const()[name = string("module_encoder_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26707584)))]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = module_encoder_norm_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_encoder_norm_weight_to_fp16, x = input_83_cast_fp16)[name = string("input_cast_fp16")]; + tensor module_fc_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26708928))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26770432))))[name = string("module_fc_out_weight_to_fp16_quantized")]; + tensor module_fc_out_bias_to_fp16 = const()[name = string("module_fc_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26770688)))]; + tensor linear_32_cast_fp16 = linear(bias = module_fc_out_bias_to_fp16, weight = module_fc_out_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("linear_32_cast_fp16")]; + tensor padded_axes_0 = const()[name = string("padded_axes_0"), val = tensor([-1])]; + tensor padded = expand_dims(axes = padded_axes_0, x = src_pad_mask)[name = string("padded")]; + tensor var_723_begin_0 = const()[name = string("op_723_begin_0"), val = tensor([0, 0, 0])]; + tensor var_723_end_0 = const()[name = string("op_723_end_0"), val = tensor([1, 128, 1])]; + tensor var_723_end_mask_0 = const()[name = string("op_723_end_mask_0"), val = tensor([true, true, false])]; + tensor var_723_cast_fp16 = slice_by_index(begin = var_723_begin_0, end = var_723_end_0, end_mask = var_723_end_mask_0, x = linear_32_cast_fp16)[name = string("op_723_cast_fp16")]; + tensor var_722_to_fp16 = const()[name = string("op_722_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26853120)))]; + tensor blank_logits_cast_fp16 = select(a = var_722_to_fp16, b = var_723_cast_fp16, cond = padded)[name = string("blank_logits_cast_fp16")]; + tensor var_725_begin_0 = const()[name = string("op_725_begin_0"), val = tensor([0, 0, 1])]; + tensor var_725_end_0 = const()[name = string("op_725_end_0"), val = tensor([1, 128, 96])]; + tensor var_725_end_mask_0 = const()[name = string("op_725_end_mask_0"), val = tensor([true, true, true])]; + tensor var_725_cast_fp16 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, x = linear_32_cast_fp16)[name = string("op_725_cast_fp16")]; + fp16 var_726_value_0_to_fp16 = const()[name = string("op_726_value_0_to_fp16"), val = fp16(-0x1.388p+13)]; + tensor var_726_cast_fp16 = fill_like(ref_tensor = var_725_cast_fp16, value = var_726_value_0_to_fp16)[name = string("op_726_cast_fp16")]; + tensor padded_after_broadcast_reps_0 = const()[name = string("padded_after_broadcast_reps_0"), val = tensor([1, 1, 95])]; + tensor padded_after_broadcast = tile(reps = padded_after_broadcast_reps_0, x = padded)[name = string("padded_after_broadcast")]; + tensor nonblank_logits_cast_fp16 = select(a = var_726_cast_fp16, b = var_725_cast_fp16, cond = padded_after_broadcast)[name = string("nonblank_logits_cast_fp16")]; + bool var_730_interleave_0 = const()[name = string("op_730_interleave_0"), val = bool(false)]; + tensor var_730_cast_fp16 = concat(axis = var_24, interleave = var_730_interleave_0, values = (blank_logits_cast_fp16, nonblank_logits_cast_fp16))[name = string("op_730_cast_fp16")]; + string var_730_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_730_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = var_730_cast_fp16_to_fp32_dtype_0, x = var_730_cast_fp16)[name = string("cast_1")]; + } -> (logits); + func g2p_256(tensor text) { + int32 var_24 = const()[name = string("op_24"), val = int32(-1)]; + int32 var_26 = const()[name = string("op_26"), val = int32(0)]; + tensor src_pad_mask = equal(x = text, y = var_26)[name = string("src_pad_mask")]; + int32 x_1_batch_dims_0 = const()[name = string("x_1_batch_dims_0"), val = int32(0)]; + bool x_1_validate_indices_0 = const()[name = string("x_1_validate_indices_0"), val = bool(false)]; + tensor module_embedding_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257408))))[name = string("module_embedding_weight_to_fp16_quantized")]; + string text_to_int16_dtype_0 = const()[name = string("text_to_int16_dtype_0"), val = string("int16")]; + string cast_49_dtype_0 = const()[name = string("cast_49_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor text_to_int16 = cast(dtype = text_to_int16_dtype_0, x = text)[name = string("cast_6")]; + tensor cast_49 = cast(dtype = cast_49_dtype_0, x = text_to_int16)[name = string("cast_5")]; + tensor greater_equal_0 = greater_equal(x = cast_49, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(402)]; + tensor add_0 = add(x = cast_49, y = slice_by_index_0)[name = string("add_0")]; + tensor select_0 = select(a = cast_49, b = add_0, cond = greater_equal_0)[name = string("select_0")]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + string cast_0_dtype_0 = const()[name = string("cast_0_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_4")]; + tensor cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = string("cast_3")]; + tensor greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")]; + int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(402)]; + tensor add_0_1 = add(x = cast_0, y = slice_by_index_0_1)[name = string("add_0_1")]; + tensor select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")]; + int32 x_1_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = string("x_1_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = int32(0)]; + tensor x_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = x_1_cast_fp16_cast_uint16_cast_uint16_axis_0, batch_dims = x_1_batch_dims_0, indices = select_0_1, validate_indices = x_1_validate_indices_0, x = module_embedding_weight_to_fp16_quantized)[name = string("x_1_cast_fp16_cast_uint16_cast_uint16")]; + tensor op_286_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26853440))), scale = tensor([[[0x1.724p-7]]]))[name = string("op_286_to_fp16_quantized")]; + tensor input_1_cast_fp16 = add(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = op_286_to_fp16_quantized)[name = string("input_1_cast_fp16")]; + tensor module_layers_0_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1528192))))[name = string("module_layers_0_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_0_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_0_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1532096)))]; + tensor linear_0_cast_fp16 = linear(bias = module_layers_0_self_attn_in_proj_bias_to_fp16, weight = module_layers_0_self_attn_in_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor var_292_split_sizes_0 = const()[name = string("op_292_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_292_axis_0 = const()[name = string("op_292_axis_0"), val = int32(-1)]; + tensor var_292_cast_fp16_0, tensor var_292_cast_fp16_1, tensor var_292_cast_fp16_2 = split(axis = var_292_axis_0, split_sizes = var_292_split_sizes_0, x = linear_0_cast_fp16)[name = string("op_292_cast_fp16")]; + tensor var_295 = const()[name = string("op_295"), val = tensor([1, 256, 4, 160])]; + tensor var_296_cast_fp16 = reshape(shape = var_295, x = var_292_cast_fp16_0)[name = string("op_296_cast_fp16")]; + tensor var_298 = const()[name = string("op_298"), val = tensor([1, 256, 4, 160])]; + tensor var_299_cast_fp16 = reshape(shape = var_298, x = var_292_cast_fp16_1)[name = string("op_299_cast_fp16")]; + tensor var_301 = const()[name = string("op_301"), val = tensor([1, 256, 4, 160])]; + tensor var_302_cast_fp16 = reshape(shape = var_301, x = var_292_cast_fp16_2)[name = string("op_302_cast_fp16")]; + tensor v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_305_transpose_x_0 = const()[name = string("op_305_transpose_x_0"), val = bool(false)]; + bool var_305_transpose_y_0 = const()[name = string("op_305_transpose_y_0"), val = bool(false)]; + tensor transpose_32_perm_0 = const()[name = string("transpose_32_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_33 = transpose(perm = transpose_33_perm_0, x = var_299_cast_fp16)[name = string("transpose_78")]; + tensor transpose_32 = transpose(perm = transpose_32_perm_0, x = var_296_cast_fp16)[name = string("transpose_79")]; + tensor var_305_cast_fp16 = matmul(transpose_x = var_305_transpose_x_0, transpose_y = var_305_transpose_y_0, x = transpose_32, y = transpose_33)[name = string("op_305_cast_fp16")]; + fp16 var_306_to_fp16 = const()[name = string("op_306_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_1_cast_fp16 = mul(x = var_305_cast_fp16, y = var_306_to_fp16)[name = string("scores_1_cast_fp16")]; + tensor weights_1_cast_fp16 = softmax(axis = var_24, x = scores_1_cast_fp16)[name = string("weights_1_cast_fp16")]; + fp16 var_13_to_fp16 = const()[name = string("op_13_to_fp16"), val = fp16(0x1p+0)]; + string var_309_to_fp16_dtype_0 = const()[name = string("op_309_to_fp16_dtype_0"), val = string("fp16")]; + tensor src_pad_mask_to_fp16 = cast(dtype = var_309_to_fp16_dtype_0, x = src_pad_mask)[name = string("cast_2")]; + tensor var_310_cast_fp16 = sub(x = var_13_to_fp16, y = src_pad_mask_to_fp16)[name = string("op_310_cast_fp16")]; + tensor var_312_axes_0 = const()[name = string("op_312_axes_0"), val = tensor([1])]; + tensor var_312_cast_fp16 = expand_dims(axes = var_312_axes_0, x = var_310_cast_fp16)[name = string("op_312_cast_fp16")]; + tensor var_313_axes_0 = const()[name = string("op_313_axes_0"), val = tensor([2])]; + tensor var_313_cast_fp16 = expand_dims(axes = var_313_axes_0, x = var_312_cast_fp16)[name = string("op_313_cast_fp16")]; + tensor weights_3_cast_fp16 = mul(x = weights_1_cast_fp16, y = var_313_cast_fp16)[name = string("weights_3_cast_fp16")]; + tensor var_317_axes_0 = const()[name = string("op_317_axes_0"), val = tensor([-1])]; + bool var_317_keep_dims_0 = const()[name = string("op_317_keep_dims_0"), val = bool(true)]; + tensor var_317_cast_fp16 = reduce_sum(axes = var_317_axes_0, keep_dims = var_317_keep_dims_0, x = weights_3_cast_fp16)[name = string("op_317_cast_fp16")]; + fp16 var_11_to_fp16 = const()[name = string("op_11_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_318_cast_fp16 = maximum(x = var_317_cast_fp16, y = var_11_to_fp16)[name = string("op_318_cast_fp16")]; + tensor weights_5_cast_fp16 = real_div(x = weights_3_cast_fp16, y = var_318_cast_fp16)[name = string("weights_5_cast_fp16")]; + bool context_1_transpose_x_0 = const()[name = string("context_1_transpose_x_0"), val = bool(false)]; + bool context_1_transpose_y_0 = const()[name = string("context_1_transpose_y_0"), val = bool(false)]; + tensor v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_302_cast_fp16)[name = string("transpose_77")]; + tensor context_1_cast_fp16 = matmul(transpose_x = context_1_transpose_x_0, transpose_y = context_1_transpose_y_0, x = weights_5_cast_fp16, y = v_3_cast_fp16)[name = string("context_1_cast_fp16")]; + tensor var_321_perm_0 = const()[name = string("op_321_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_322 = const()[name = string("op_322"), val = tensor([1, 256, 640])]; + tensor var_321_cast_fp16 = transpose(perm = var_321_perm_0, x = context_1_cast_fp16)[name = string("transpose_76")]; + tensor context_3_cast_fp16 = reshape(shape = var_322, x = var_321_cast_fp16)[name = string("context_3_cast_fp16")]; + tensor module_layers_0_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1536000))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1945664))))[name = string("module_layers_0_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_0_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_0_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1947008)))]; + tensor linear_1_cast_fp16 = linear(bias = module_layers_0_self_attn_out_proj_bias_to_fp16, weight = module_layers_0_self_attn_out_proj_weight_to_fp16_quantized, x = context_3_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor input_3_cast_fp16 = add(x = input_1_cast_fp16, y = linear_1_cast_fp16)[name = string("input_3_cast_fp16")]; + tensor input_5_axes_0 = const()[name = string("input_5_axes_0"), val = tensor([-1])]; + tensor module_layers_0_norm1_weight_to_fp16 = const()[name = string("module_layers_0_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1948352)))]; + tensor module_layers_0_norm1_bias_to_fp16 = const()[name = string("module_layers_0_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1949696)))]; + fp16 var_9_to_fp16 = const()[name = string("op_9_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = module_layers_0_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_0_norm1_weight_to_fp16, x = input_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor module_layers_0_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1951040))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2770304))))[name = string("module_layers_0_linear1_weight_to_fp16_quantized")]; + tensor module_layers_0_linear1_bias_to_fp16 = const()[name = string("module_layers_0_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2772928)))]; + tensor linear_2_cast_fp16 = linear(bias = module_layers_0_linear1_bias_to_fp16, weight = module_layers_0_linear1_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_2_cast_fp16")]; + tensor input_9_cast_fp16 = relu(x = linear_2_cast_fp16)[name = string("input_9_cast_fp16")]; + tensor module_layers_0_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2775552))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3594816))))[name = string("module_layers_0_linear2_weight_to_fp16_quantized")]; + tensor module_layers_0_linear2_bias_to_fp16 = const()[name = string("module_layers_0_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3596160)))]; + tensor linear_3_cast_fp16 = linear(bias = module_layers_0_linear2_bias_to_fp16, weight = module_layers_0_linear2_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("linear_3_cast_fp16")]; + tensor input_11_cast_fp16 = add(x = input_5_cast_fp16, y = linear_3_cast_fp16)[name = string("input_11_cast_fp16")]; + tensor x_5_axes_0 = const()[name = string("x_5_axes_0"), val = tensor([-1])]; + tensor module_layers_0_norm2_weight_to_fp16 = const()[name = string("module_layers_0_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3597504)))]; + tensor module_layers_0_norm2_bias_to_fp16 = const()[name = string("module_layers_0_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3598848)))]; + tensor x_5_cast_fp16 = layer_norm(axes = x_5_axes_0, beta = module_layers_0_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_0_norm2_weight_to_fp16, x = input_11_cast_fp16)[name = string("x_5_cast_fp16")]; + tensor module_layers_1_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3600192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4829056))))[name = string("module_layers_1_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_1_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_1_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4832960)))]; + tensor linear_4_cast_fp16 = linear(bias = module_layers_1_self_attn_in_proj_bias_to_fp16, weight = module_layers_1_self_attn_in_proj_weight_to_fp16_quantized, x = x_5_cast_fp16)[name = string("linear_4_cast_fp16")]; + tensor var_345_split_sizes_0 = const()[name = string("op_345_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_345_axis_0 = const()[name = string("op_345_axis_0"), val = int32(-1)]; + tensor var_345_cast_fp16_0, tensor var_345_cast_fp16_1, tensor var_345_cast_fp16_2 = split(axis = var_345_axis_0, split_sizes = var_345_split_sizes_0, x = linear_4_cast_fp16)[name = string("op_345_cast_fp16")]; + tensor var_348 = const()[name = string("op_348"), val = tensor([1, 256, 4, 160])]; + tensor var_349_cast_fp16 = reshape(shape = var_348, x = var_345_cast_fp16_0)[name = string("op_349_cast_fp16")]; + tensor var_351 = const()[name = string("op_351"), val = tensor([1, 256, 4, 160])]; + tensor var_352_cast_fp16 = reshape(shape = var_351, x = var_345_cast_fp16_1)[name = string("op_352_cast_fp16")]; + tensor var_354 = const()[name = string("op_354"), val = tensor([1, 256, 4, 160])]; + tensor var_355_cast_fp16 = reshape(shape = var_354, x = var_345_cast_fp16_2)[name = string("op_355_cast_fp16")]; + tensor v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_358_transpose_x_0 = const()[name = string("op_358_transpose_x_0"), val = bool(false)]; + bool var_358_transpose_y_0 = const()[name = string("op_358_transpose_y_0"), val = bool(false)]; + tensor transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_35_perm_0 = const()[name = string("transpose_35_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_35 = transpose(perm = transpose_35_perm_0, x = var_352_cast_fp16)[name = string("transpose_74")]; + tensor transpose_34 = transpose(perm = transpose_34_perm_0, x = var_349_cast_fp16)[name = string("transpose_75")]; + tensor var_358_cast_fp16 = matmul(transpose_x = var_358_transpose_x_0, transpose_y = var_358_transpose_y_0, x = transpose_34, y = transpose_35)[name = string("op_358_cast_fp16")]; + fp16 var_359_to_fp16 = const()[name = string("op_359_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_3_cast_fp16 = mul(x = var_358_cast_fp16, y = var_359_to_fp16)[name = string("scores_3_cast_fp16")]; + tensor weights_7_cast_fp16 = softmax(axis = var_24, x = scores_3_cast_fp16)[name = string("weights_7_cast_fp16")]; + tensor weights_9_cast_fp16 = mul(x = weights_7_cast_fp16, y = var_313_cast_fp16)[name = string("weights_9_cast_fp16")]; + tensor var_370_axes_0 = const()[name = string("op_370_axes_0"), val = tensor([-1])]; + bool var_370_keep_dims_0 = const()[name = string("op_370_keep_dims_0"), val = bool(true)]; + tensor var_370_cast_fp16 = reduce_sum(axes = var_370_axes_0, keep_dims = var_370_keep_dims_0, x = weights_9_cast_fp16)[name = string("op_370_cast_fp16")]; + tensor var_371_cast_fp16 = maximum(x = var_370_cast_fp16, y = var_11_to_fp16)[name = string("op_371_cast_fp16")]; + tensor weights_11_cast_fp16 = real_div(x = weights_9_cast_fp16, y = var_371_cast_fp16)[name = string("weights_11_cast_fp16")]; + bool context_5_transpose_x_0 = const()[name = string("context_5_transpose_x_0"), val = bool(false)]; + bool context_5_transpose_y_0 = const()[name = string("context_5_transpose_y_0"), val = bool(false)]; + tensor v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = var_355_cast_fp16)[name = string("transpose_73")]; + tensor context_5_cast_fp16 = matmul(transpose_x = context_5_transpose_x_0, transpose_y = context_5_transpose_y_0, x = weights_11_cast_fp16, y = v_7_cast_fp16)[name = string("context_5_cast_fp16")]; + tensor var_374_perm_0 = const()[name = string("op_374_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_375 = const()[name = string("op_375"), val = tensor([1, 256, 640])]; + tensor var_374_cast_fp16 = transpose(perm = var_374_perm_0, x = context_5_cast_fp16)[name = string("transpose_72")]; + tensor context_7_cast_fp16 = reshape(shape = var_375, x = var_374_cast_fp16)[name = string("context_7_cast_fp16")]; + tensor module_layers_1_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4836864))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5246528))))[name = string("module_layers_1_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_1_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_1_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5247872)))]; + tensor linear_5_cast_fp16 = linear(bias = module_layers_1_self_attn_out_proj_bias_to_fp16, weight = module_layers_1_self_attn_out_proj_weight_to_fp16_quantized, x = context_7_cast_fp16)[name = string("linear_5_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = x_5_cast_fp16, y = linear_5_cast_fp16)[name = string("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = string("input_15_axes_0"), val = tensor([-1])]; + tensor module_layers_1_norm1_weight_to_fp16 = const()[name = string("module_layers_1_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5249216)))]; + tensor module_layers_1_norm1_bias_to_fp16 = const()[name = string("module_layers_1_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5250560)))]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = module_layers_1_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_1_norm1_weight_to_fp16, x = input_13_cast_fp16)[name = string("input_15_cast_fp16")]; + tensor module_layers_1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5251904))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6071168))))[name = string("module_layers_1_linear1_weight_to_fp16_quantized")]; + tensor module_layers_1_linear1_bias_to_fp16 = const()[name = string("module_layers_1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6073792)))]; + tensor linear_6_cast_fp16 = linear(bias = module_layers_1_linear1_bias_to_fp16, weight = module_layers_1_linear1_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_6_cast_fp16")]; + tensor input_19_cast_fp16 = relu(x = linear_6_cast_fp16)[name = string("input_19_cast_fp16")]; + tensor module_layers_1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6076416))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6895680))))[name = string("module_layers_1_linear2_weight_to_fp16_quantized")]; + tensor module_layers_1_linear2_bias_to_fp16 = const()[name = string("module_layers_1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6897024)))]; + tensor linear_7_cast_fp16 = linear(bias = module_layers_1_linear2_bias_to_fp16, weight = module_layers_1_linear2_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_7_cast_fp16")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_7_cast_fp16)[name = string("input_21_cast_fp16")]; + tensor x_7_axes_0 = const()[name = string("x_7_axes_0"), val = tensor([-1])]; + tensor module_layers_1_norm2_weight_to_fp16 = const()[name = string("module_layers_1_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6898368)))]; + tensor module_layers_1_norm2_bias_to_fp16 = const()[name = string("module_layers_1_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6899712)))]; + tensor x_7_cast_fp16 = layer_norm(axes = x_7_axes_0, beta = module_layers_1_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_1_norm2_weight_to_fp16, x = input_21_cast_fp16)[name = string("x_7_cast_fp16")]; + tensor module_layers_2_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6901056))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8129920))))[name = string("module_layers_2_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_2_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_2_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8133824)))]; + tensor linear_8_cast_fp16 = linear(bias = module_layers_2_self_attn_in_proj_bias_to_fp16, weight = module_layers_2_self_attn_in_proj_weight_to_fp16_quantized, x = x_7_cast_fp16)[name = string("linear_8_cast_fp16")]; + tensor var_398_split_sizes_0 = const()[name = string("op_398_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_398_axis_0 = const()[name = string("op_398_axis_0"), val = int32(-1)]; + tensor var_398_cast_fp16_0, tensor var_398_cast_fp16_1, tensor var_398_cast_fp16_2 = split(axis = var_398_axis_0, split_sizes = var_398_split_sizes_0, x = linear_8_cast_fp16)[name = string("op_398_cast_fp16")]; + tensor var_401 = const()[name = string("op_401"), val = tensor([1, 256, 4, 160])]; + tensor var_402_cast_fp16 = reshape(shape = var_401, x = var_398_cast_fp16_0)[name = string("op_402_cast_fp16")]; + tensor var_404 = const()[name = string("op_404"), val = tensor([1, 256, 4, 160])]; + tensor var_405_cast_fp16 = reshape(shape = var_404, x = var_398_cast_fp16_1)[name = string("op_405_cast_fp16")]; + tensor var_407 = const()[name = string("op_407"), val = tensor([1, 256, 4, 160])]; + tensor var_408_cast_fp16 = reshape(shape = var_407, x = var_398_cast_fp16_2)[name = string("op_408_cast_fp16")]; + tensor v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_411_transpose_x_0 = const()[name = string("op_411_transpose_x_0"), val = bool(false)]; + bool var_411_transpose_y_0 = const()[name = string("op_411_transpose_y_0"), val = bool(false)]; + tensor transpose_36_perm_0 = const()[name = string("transpose_36_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_37_perm_0 = const()[name = string("transpose_37_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_37 = transpose(perm = transpose_37_perm_0, x = var_405_cast_fp16)[name = string("transpose_70")]; + tensor transpose_36 = transpose(perm = transpose_36_perm_0, x = var_402_cast_fp16)[name = string("transpose_71")]; + tensor var_411_cast_fp16 = matmul(transpose_x = var_411_transpose_x_0, transpose_y = var_411_transpose_y_0, x = transpose_36, y = transpose_37)[name = string("op_411_cast_fp16")]; + fp16 var_412_to_fp16 = const()[name = string("op_412_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_5_cast_fp16 = mul(x = var_411_cast_fp16, y = var_412_to_fp16)[name = string("scores_5_cast_fp16")]; + tensor weights_13_cast_fp16 = softmax(axis = var_24, x = scores_5_cast_fp16)[name = string("weights_13_cast_fp16")]; + tensor weights_15_cast_fp16 = mul(x = weights_13_cast_fp16, y = var_313_cast_fp16)[name = string("weights_15_cast_fp16")]; + tensor var_423_axes_0 = const()[name = string("op_423_axes_0"), val = tensor([-1])]; + bool var_423_keep_dims_0 = const()[name = string("op_423_keep_dims_0"), val = bool(true)]; + tensor var_423_cast_fp16 = reduce_sum(axes = var_423_axes_0, keep_dims = var_423_keep_dims_0, x = weights_15_cast_fp16)[name = string("op_423_cast_fp16")]; + tensor var_424_cast_fp16 = maximum(x = var_423_cast_fp16, y = var_11_to_fp16)[name = string("op_424_cast_fp16")]; + tensor weights_17_cast_fp16 = real_div(x = weights_15_cast_fp16, y = var_424_cast_fp16)[name = string("weights_17_cast_fp16")]; + bool context_9_transpose_x_0 = const()[name = string("context_9_transpose_x_0"), val = bool(false)]; + bool context_9_transpose_y_0 = const()[name = string("context_9_transpose_y_0"), val = bool(false)]; + tensor v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = var_408_cast_fp16)[name = string("transpose_69")]; + tensor context_9_cast_fp16 = matmul(transpose_x = context_9_transpose_x_0, transpose_y = context_9_transpose_y_0, x = weights_17_cast_fp16, y = v_11_cast_fp16)[name = string("context_9_cast_fp16")]; + tensor var_427_perm_0 = const()[name = string("op_427_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_428 = const()[name = string("op_428"), val = tensor([1, 256, 640])]; + tensor var_427_cast_fp16 = transpose(perm = var_427_perm_0, x = context_9_cast_fp16)[name = string("transpose_68")]; + tensor context_11_cast_fp16 = reshape(shape = var_428, x = var_427_cast_fp16)[name = string("context_11_cast_fp16")]; + tensor module_layers_2_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8137728))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8547392))))[name = string("module_layers_2_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_2_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_2_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8548736)))]; + tensor linear_9_cast_fp16 = linear(bias = module_layers_2_self_attn_out_proj_bias_to_fp16, weight = module_layers_2_self_attn_out_proj_weight_to_fp16_quantized, x = context_11_cast_fp16)[name = string("linear_9_cast_fp16")]; + tensor input_23_cast_fp16 = add(x = x_7_cast_fp16, y = linear_9_cast_fp16)[name = string("input_23_cast_fp16")]; + tensor input_25_axes_0 = const()[name = string("input_25_axes_0"), val = tensor([-1])]; + tensor module_layers_2_norm1_weight_to_fp16 = const()[name = string("module_layers_2_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8550080)))]; + tensor module_layers_2_norm1_bias_to_fp16 = const()[name = string("module_layers_2_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8551424)))]; + tensor input_25_cast_fp16 = layer_norm(axes = input_25_axes_0, beta = module_layers_2_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_2_norm1_weight_to_fp16, x = input_23_cast_fp16)[name = string("input_25_cast_fp16")]; + tensor module_layers_2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8552768))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9372032))))[name = string("module_layers_2_linear1_weight_to_fp16_quantized")]; + tensor module_layers_2_linear1_bias_to_fp16 = const()[name = string("module_layers_2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9374656)))]; + tensor linear_10_cast_fp16 = linear(bias = module_layers_2_linear1_bias_to_fp16, weight = module_layers_2_linear1_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_10_cast_fp16")]; + tensor input_29_cast_fp16 = relu(x = linear_10_cast_fp16)[name = string("input_29_cast_fp16")]; + tensor module_layers_2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9377280))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10196544))))[name = string("module_layers_2_linear2_weight_to_fp16_quantized")]; + tensor module_layers_2_linear2_bias_to_fp16 = const()[name = string("module_layers_2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10197888)))]; + tensor linear_11_cast_fp16 = linear(bias = module_layers_2_linear2_bias_to_fp16, weight = module_layers_2_linear2_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_11_cast_fp16")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_11_cast_fp16)[name = string("input_31_cast_fp16")]; + tensor x_9_axes_0 = const()[name = string("x_9_axes_0"), val = tensor([-1])]; + tensor module_layers_2_norm2_weight_to_fp16 = const()[name = string("module_layers_2_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10199232)))]; + tensor module_layers_2_norm2_bias_to_fp16 = const()[name = string("module_layers_2_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10200576)))]; + tensor x_9_cast_fp16 = layer_norm(axes = x_9_axes_0, beta = module_layers_2_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_2_norm2_weight_to_fp16, x = input_31_cast_fp16)[name = string("x_9_cast_fp16")]; + tensor module_layers_3_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10201920))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11430784))))[name = string("module_layers_3_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_3_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_3_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11434688)))]; + tensor linear_12_cast_fp16 = linear(bias = module_layers_3_self_attn_in_proj_bias_to_fp16, weight = module_layers_3_self_attn_in_proj_weight_to_fp16_quantized, x = x_9_cast_fp16)[name = string("linear_12_cast_fp16")]; + tensor var_451_split_sizes_0 = const()[name = string("op_451_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_451_axis_0 = const()[name = string("op_451_axis_0"), val = int32(-1)]; + tensor var_451_cast_fp16_0, tensor var_451_cast_fp16_1, tensor var_451_cast_fp16_2 = split(axis = var_451_axis_0, split_sizes = var_451_split_sizes_0, x = linear_12_cast_fp16)[name = string("op_451_cast_fp16")]; + tensor var_454 = const()[name = string("op_454"), val = tensor([1, 256, 4, 160])]; + tensor var_455_cast_fp16 = reshape(shape = var_454, x = var_451_cast_fp16_0)[name = string("op_455_cast_fp16")]; + tensor var_457 = const()[name = string("op_457"), val = tensor([1, 256, 4, 160])]; + tensor var_458_cast_fp16 = reshape(shape = var_457, x = var_451_cast_fp16_1)[name = string("op_458_cast_fp16")]; + tensor var_460 = const()[name = string("op_460"), val = tensor([1, 256, 4, 160])]; + tensor var_461_cast_fp16 = reshape(shape = var_460, x = var_451_cast_fp16_2)[name = string("op_461_cast_fp16")]; + tensor v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_464_transpose_x_0 = const()[name = string("op_464_transpose_x_0"), val = bool(false)]; + bool var_464_transpose_y_0 = const()[name = string("op_464_transpose_y_0"), val = bool(false)]; + tensor transpose_38_perm_0 = const()[name = string("transpose_38_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_39_perm_0 = const()[name = string("transpose_39_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_39 = transpose(perm = transpose_39_perm_0, x = var_458_cast_fp16)[name = string("transpose_66")]; + tensor transpose_38 = transpose(perm = transpose_38_perm_0, x = var_455_cast_fp16)[name = string("transpose_67")]; + tensor var_464_cast_fp16 = matmul(transpose_x = var_464_transpose_x_0, transpose_y = var_464_transpose_y_0, x = transpose_38, y = transpose_39)[name = string("op_464_cast_fp16")]; + fp16 var_465_to_fp16 = const()[name = string("op_465_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_7_cast_fp16 = mul(x = var_464_cast_fp16, y = var_465_to_fp16)[name = string("scores_7_cast_fp16")]; + tensor weights_19_cast_fp16 = softmax(axis = var_24, x = scores_7_cast_fp16)[name = string("weights_19_cast_fp16")]; + tensor weights_21_cast_fp16 = mul(x = weights_19_cast_fp16, y = var_313_cast_fp16)[name = string("weights_21_cast_fp16")]; + tensor var_476_axes_0 = const()[name = string("op_476_axes_0"), val = tensor([-1])]; + bool var_476_keep_dims_0 = const()[name = string("op_476_keep_dims_0"), val = bool(true)]; + tensor var_476_cast_fp16 = reduce_sum(axes = var_476_axes_0, keep_dims = var_476_keep_dims_0, x = weights_21_cast_fp16)[name = string("op_476_cast_fp16")]; + tensor var_477_cast_fp16 = maximum(x = var_476_cast_fp16, y = var_11_to_fp16)[name = string("op_477_cast_fp16")]; + tensor weights_23_cast_fp16 = real_div(x = weights_21_cast_fp16, y = var_477_cast_fp16)[name = string("weights_23_cast_fp16")]; + bool context_13_transpose_x_0 = const()[name = string("context_13_transpose_x_0"), val = bool(false)]; + bool context_13_transpose_y_0 = const()[name = string("context_13_transpose_y_0"), val = bool(false)]; + tensor v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = var_461_cast_fp16)[name = string("transpose_65")]; + tensor context_13_cast_fp16 = matmul(transpose_x = context_13_transpose_x_0, transpose_y = context_13_transpose_y_0, x = weights_23_cast_fp16, y = v_15_cast_fp16)[name = string("context_13_cast_fp16")]; + tensor var_480_perm_0 = const()[name = string("op_480_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_481 = const()[name = string("op_481"), val = tensor([1, 256, 640])]; + tensor var_480_cast_fp16 = transpose(perm = var_480_perm_0, x = context_13_cast_fp16)[name = string("transpose_64")]; + tensor context_15_cast_fp16 = reshape(shape = var_481, x = var_480_cast_fp16)[name = string("context_15_cast_fp16")]; + tensor module_layers_3_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11438592))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11848256))))[name = string("module_layers_3_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_3_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_3_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11849600)))]; + tensor linear_13_cast_fp16 = linear(bias = module_layers_3_self_attn_out_proj_bias_to_fp16, weight = module_layers_3_self_attn_out_proj_weight_to_fp16_quantized, x = context_15_cast_fp16)[name = string("linear_13_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = x_9_cast_fp16, y = linear_13_cast_fp16)[name = string("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = string("input_35_axes_0"), val = tensor([-1])]; + tensor module_layers_3_norm1_weight_to_fp16 = const()[name = string("module_layers_3_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11850944)))]; + tensor module_layers_3_norm1_bias_to_fp16 = const()[name = string("module_layers_3_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11852288)))]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = module_layers_3_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_3_norm1_weight_to_fp16, x = input_33_cast_fp16)[name = string("input_35_cast_fp16")]; + tensor module_layers_3_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11853632))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12672896))))[name = string("module_layers_3_linear1_weight_to_fp16_quantized")]; + tensor module_layers_3_linear1_bias_to_fp16 = const()[name = string("module_layers_3_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12675520)))]; + tensor linear_14_cast_fp16 = linear(bias = module_layers_3_linear1_bias_to_fp16, weight = module_layers_3_linear1_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_14_cast_fp16")]; + tensor input_39_cast_fp16 = relu(x = linear_14_cast_fp16)[name = string("input_39_cast_fp16")]; + tensor module_layers_3_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12678144))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13497408))))[name = string("module_layers_3_linear2_weight_to_fp16_quantized")]; + tensor module_layers_3_linear2_bias_to_fp16 = const()[name = string("module_layers_3_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13498752)))]; + tensor linear_15_cast_fp16 = linear(bias = module_layers_3_linear2_bias_to_fp16, weight = module_layers_3_linear2_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_15_cast_fp16")]; + tensor input_41_cast_fp16 = add(x = input_35_cast_fp16, y = linear_15_cast_fp16)[name = string("input_41_cast_fp16")]; + tensor x_11_axes_0 = const()[name = string("x_11_axes_0"), val = tensor([-1])]; + tensor module_layers_3_norm2_weight_to_fp16 = const()[name = string("module_layers_3_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13500096)))]; + tensor module_layers_3_norm2_bias_to_fp16 = const()[name = string("module_layers_3_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13501440)))]; + tensor x_11_cast_fp16 = layer_norm(axes = x_11_axes_0, beta = module_layers_3_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_3_norm2_weight_to_fp16, x = input_41_cast_fp16)[name = string("x_11_cast_fp16")]; + tensor module_layers_4_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13502784))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14731648))))[name = string("module_layers_4_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_4_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_4_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14735552)))]; + tensor linear_16_cast_fp16 = linear(bias = module_layers_4_self_attn_in_proj_bias_to_fp16, weight = module_layers_4_self_attn_in_proj_weight_to_fp16_quantized, x = x_11_cast_fp16)[name = string("linear_16_cast_fp16")]; + tensor var_504_split_sizes_0 = const()[name = string("op_504_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_504_axis_0 = const()[name = string("op_504_axis_0"), val = int32(-1)]; + tensor var_504_cast_fp16_0, tensor var_504_cast_fp16_1, tensor var_504_cast_fp16_2 = split(axis = var_504_axis_0, split_sizes = var_504_split_sizes_0, x = linear_16_cast_fp16)[name = string("op_504_cast_fp16")]; + tensor var_507 = const()[name = string("op_507"), val = tensor([1, 256, 4, 160])]; + tensor var_508_cast_fp16 = reshape(shape = var_507, x = var_504_cast_fp16_0)[name = string("op_508_cast_fp16")]; + tensor var_510 = const()[name = string("op_510"), val = tensor([1, 256, 4, 160])]; + tensor var_511_cast_fp16 = reshape(shape = var_510, x = var_504_cast_fp16_1)[name = string("op_511_cast_fp16")]; + tensor var_513 = const()[name = string("op_513"), val = tensor([1, 256, 4, 160])]; + tensor var_514_cast_fp16 = reshape(shape = var_513, x = var_504_cast_fp16_2)[name = string("op_514_cast_fp16")]; + tensor v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_517_transpose_x_0 = const()[name = string("op_517_transpose_x_0"), val = bool(false)]; + bool var_517_transpose_y_0 = const()[name = string("op_517_transpose_y_0"), val = bool(false)]; + tensor transpose_40_perm_0 = const()[name = string("transpose_40_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_41_perm_0 = const()[name = string("transpose_41_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_41 = transpose(perm = transpose_41_perm_0, x = var_511_cast_fp16)[name = string("transpose_62")]; + tensor transpose_40 = transpose(perm = transpose_40_perm_0, x = var_508_cast_fp16)[name = string("transpose_63")]; + tensor var_517_cast_fp16 = matmul(transpose_x = var_517_transpose_x_0, transpose_y = var_517_transpose_y_0, x = transpose_40, y = transpose_41)[name = string("op_517_cast_fp16")]; + fp16 var_518_to_fp16 = const()[name = string("op_518_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_9_cast_fp16 = mul(x = var_517_cast_fp16, y = var_518_to_fp16)[name = string("scores_9_cast_fp16")]; + tensor weights_25_cast_fp16 = softmax(axis = var_24, x = scores_9_cast_fp16)[name = string("weights_25_cast_fp16")]; + tensor weights_27_cast_fp16 = mul(x = weights_25_cast_fp16, y = var_313_cast_fp16)[name = string("weights_27_cast_fp16")]; + tensor var_529_axes_0 = const()[name = string("op_529_axes_0"), val = tensor([-1])]; + bool var_529_keep_dims_0 = const()[name = string("op_529_keep_dims_0"), val = bool(true)]; + tensor var_529_cast_fp16 = reduce_sum(axes = var_529_axes_0, keep_dims = var_529_keep_dims_0, x = weights_27_cast_fp16)[name = string("op_529_cast_fp16")]; + tensor var_530_cast_fp16 = maximum(x = var_529_cast_fp16, y = var_11_to_fp16)[name = string("op_530_cast_fp16")]; + tensor weights_29_cast_fp16 = real_div(x = weights_27_cast_fp16, y = var_530_cast_fp16)[name = string("weights_29_cast_fp16")]; + bool context_17_transpose_x_0 = const()[name = string("context_17_transpose_x_0"), val = bool(false)]; + bool context_17_transpose_y_0 = const()[name = string("context_17_transpose_y_0"), val = bool(false)]; + tensor v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = var_514_cast_fp16)[name = string("transpose_61")]; + tensor context_17_cast_fp16 = matmul(transpose_x = context_17_transpose_x_0, transpose_y = context_17_transpose_y_0, x = weights_29_cast_fp16, y = v_19_cast_fp16)[name = string("context_17_cast_fp16")]; + tensor var_533_perm_0 = const()[name = string("op_533_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_534 = const()[name = string("op_534"), val = tensor([1, 256, 640])]; + tensor var_533_cast_fp16 = transpose(perm = var_533_perm_0, x = context_17_cast_fp16)[name = string("transpose_60")]; + tensor context_19_cast_fp16 = reshape(shape = var_534, x = var_533_cast_fp16)[name = string("context_19_cast_fp16")]; + tensor module_layers_4_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14739456))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15149120))))[name = string("module_layers_4_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_4_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_4_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15150464)))]; + tensor linear_17_cast_fp16 = linear(bias = module_layers_4_self_attn_out_proj_bias_to_fp16, weight = module_layers_4_self_attn_out_proj_weight_to_fp16_quantized, x = context_19_cast_fp16)[name = string("linear_17_cast_fp16")]; + tensor input_43_cast_fp16 = add(x = x_11_cast_fp16, y = linear_17_cast_fp16)[name = string("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = string("input_45_axes_0"), val = tensor([-1])]; + tensor module_layers_4_norm1_weight_to_fp16 = const()[name = string("module_layers_4_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15151808)))]; + tensor module_layers_4_norm1_bias_to_fp16 = const()[name = string("module_layers_4_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15153152)))]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = module_layers_4_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_4_norm1_weight_to_fp16, x = input_43_cast_fp16)[name = string("input_45_cast_fp16")]; + tensor module_layers_4_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15154496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15973760))))[name = string("module_layers_4_linear1_weight_to_fp16_quantized")]; + tensor module_layers_4_linear1_bias_to_fp16 = const()[name = string("module_layers_4_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15976384)))]; + tensor linear_18_cast_fp16 = linear(bias = module_layers_4_linear1_bias_to_fp16, weight = module_layers_4_linear1_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_18_cast_fp16")]; + tensor input_49_cast_fp16 = relu(x = linear_18_cast_fp16)[name = string("input_49_cast_fp16")]; + tensor module_layers_4_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15979008))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16798272))))[name = string("module_layers_4_linear2_weight_to_fp16_quantized")]; + tensor module_layers_4_linear2_bias_to_fp16 = const()[name = string("module_layers_4_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16799616)))]; + tensor linear_19_cast_fp16 = linear(bias = module_layers_4_linear2_bias_to_fp16, weight = module_layers_4_linear2_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_19_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_19_cast_fp16)[name = string("input_51_cast_fp16")]; + tensor x_13_axes_0 = const()[name = string("x_13_axes_0"), val = tensor([-1])]; + tensor module_layers_4_norm2_weight_to_fp16 = const()[name = string("module_layers_4_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16800960)))]; + tensor module_layers_4_norm2_bias_to_fp16 = const()[name = string("module_layers_4_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16802304)))]; + tensor x_13_cast_fp16 = layer_norm(axes = x_13_axes_0, beta = module_layers_4_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_4_norm2_weight_to_fp16, x = input_51_cast_fp16)[name = string("x_13_cast_fp16")]; + tensor module_layers_5_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16803648))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18032512))))[name = string("module_layers_5_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_5_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_5_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18036416)))]; + tensor linear_20_cast_fp16 = linear(bias = module_layers_5_self_attn_in_proj_bias_to_fp16, weight = module_layers_5_self_attn_in_proj_weight_to_fp16_quantized, x = x_13_cast_fp16)[name = string("linear_20_cast_fp16")]; + tensor var_557_split_sizes_0 = const()[name = string("op_557_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_557_axis_0 = const()[name = string("op_557_axis_0"), val = int32(-1)]; + tensor var_557_cast_fp16_0, tensor var_557_cast_fp16_1, tensor var_557_cast_fp16_2 = split(axis = var_557_axis_0, split_sizes = var_557_split_sizes_0, x = linear_20_cast_fp16)[name = string("op_557_cast_fp16")]; + tensor var_560 = const()[name = string("op_560"), val = tensor([1, 256, 4, 160])]; + tensor var_561_cast_fp16 = reshape(shape = var_560, x = var_557_cast_fp16_0)[name = string("op_561_cast_fp16")]; + tensor var_563 = const()[name = string("op_563"), val = tensor([1, 256, 4, 160])]; + tensor var_564_cast_fp16 = reshape(shape = var_563, x = var_557_cast_fp16_1)[name = string("op_564_cast_fp16")]; + tensor var_566 = const()[name = string("op_566"), val = tensor([1, 256, 4, 160])]; + tensor var_567_cast_fp16 = reshape(shape = var_566, x = var_557_cast_fp16_2)[name = string("op_567_cast_fp16")]; + tensor v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_570_transpose_x_0 = const()[name = string("op_570_transpose_x_0"), val = bool(false)]; + bool var_570_transpose_y_0 = const()[name = string("op_570_transpose_y_0"), val = bool(false)]; + tensor transpose_42_perm_0 = const()[name = string("transpose_42_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_43_perm_0 = const()[name = string("transpose_43_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_43 = transpose(perm = transpose_43_perm_0, x = var_564_cast_fp16)[name = string("transpose_58")]; + tensor transpose_42 = transpose(perm = transpose_42_perm_0, x = var_561_cast_fp16)[name = string("transpose_59")]; + tensor var_570_cast_fp16 = matmul(transpose_x = var_570_transpose_x_0, transpose_y = var_570_transpose_y_0, x = transpose_42, y = transpose_43)[name = string("op_570_cast_fp16")]; + fp16 var_571_to_fp16 = const()[name = string("op_571_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_11_cast_fp16 = mul(x = var_570_cast_fp16, y = var_571_to_fp16)[name = string("scores_11_cast_fp16")]; + tensor weights_31_cast_fp16 = softmax(axis = var_24, x = scores_11_cast_fp16)[name = string("weights_31_cast_fp16")]; + tensor weights_33_cast_fp16 = mul(x = weights_31_cast_fp16, y = var_313_cast_fp16)[name = string("weights_33_cast_fp16")]; + tensor var_582_axes_0 = const()[name = string("op_582_axes_0"), val = tensor([-1])]; + bool var_582_keep_dims_0 = const()[name = string("op_582_keep_dims_0"), val = bool(true)]; + tensor var_582_cast_fp16 = reduce_sum(axes = var_582_axes_0, keep_dims = var_582_keep_dims_0, x = weights_33_cast_fp16)[name = string("op_582_cast_fp16")]; + tensor var_583_cast_fp16 = maximum(x = var_582_cast_fp16, y = var_11_to_fp16)[name = string("op_583_cast_fp16")]; + tensor weights_35_cast_fp16 = real_div(x = weights_33_cast_fp16, y = var_583_cast_fp16)[name = string("weights_35_cast_fp16")]; + bool context_21_transpose_x_0 = const()[name = string("context_21_transpose_x_0"), val = bool(false)]; + bool context_21_transpose_y_0 = const()[name = string("context_21_transpose_y_0"), val = bool(false)]; + tensor v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = var_567_cast_fp16)[name = string("transpose_57")]; + tensor context_21_cast_fp16 = matmul(transpose_x = context_21_transpose_x_0, transpose_y = context_21_transpose_y_0, x = weights_35_cast_fp16, y = v_23_cast_fp16)[name = string("context_21_cast_fp16")]; + tensor var_586_perm_0 = const()[name = string("op_586_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_587 = const()[name = string("op_587"), val = tensor([1, 256, 640])]; + tensor var_586_cast_fp16 = transpose(perm = var_586_perm_0, x = context_21_cast_fp16)[name = string("transpose_56")]; + tensor context_23_cast_fp16 = reshape(shape = var_587, x = var_586_cast_fp16)[name = string("context_23_cast_fp16")]; + tensor module_layers_5_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18040320))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18449984))))[name = string("module_layers_5_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_5_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_5_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18451328)))]; + tensor linear_21_cast_fp16 = linear(bias = module_layers_5_self_attn_out_proj_bias_to_fp16, weight = module_layers_5_self_attn_out_proj_weight_to_fp16_quantized, x = context_23_cast_fp16)[name = string("linear_21_cast_fp16")]; + tensor input_53_cast_fp16 = add(x = x_13_cast_fp16, y = linear_21_cast_fp16)[name = string("input_53_cast_fp16")]; + tensor input_55_axes_0 = const()[name = string("input_55_axes_0"), val = tensor([-1])]; + tensor module_layers_5_norm1_weight_to_fp16 = const()[name = string("module_layers_5_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18452672)))]; + tensor module_layers_5_norm1_bias_to_fp16 = const()[name = string("module_layers_5_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18454016)))]; + tensor input_55_cast_fp16 = layer_norm(axes = input_55_axes_0, beta = module_layers_5_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_5_norm1_weight_to_fp16, x = input_53_cast_fp16)[name = string("input_55_cast_fp16")]; + tensor module_layers_5_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18455360))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19274624))))[name = string("module_layers_5_linear1_weight_to_fp16_quantized")]; + tensor module_layers_5_linear1_bias_to_fp16 = const()[name = string("module_layers_5_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19277248)))]; + tensor linear_22_cast_fp16 = linear(bias = module_layers_5_linear1_bias_to_fp16, weight = module_layers_5_linear1_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_22_cast_fp16")]; + tensor input_59_cast_fp16 = relu(x = linear_22_cast_fp16)[name = string("input_59_cast_fp16")]; + tensor module_layers_5_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19279872))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20099136))))[name = string("module_layers_5_linear2_weight_to_fp16_quantized")]; + tensor module_layers_5_linear2_bias_to_fp16 = const()[name = string("module_layers_5_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20100480)))]; + tensor linear_23_cast_fp16 = linear(bias = module_layers_5_linear2_bias_to_fp16, weight = module_layers_5_linear2_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_23_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_23_cast_fp16)[name = string("input_61_cast_fp16")]; + tensor x_15_axes_0 = const()[name = string("x_15_axes_0"), val = tensor([-1])]; + tensor module_layers_5_norm2_weight_to_fp16 = const()[name = string("module_layers_5_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20101824)))]; + tensor module_layers_5_norm2_bias_to_fp16 = const()[name = string("module_layers_5_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20103168)))]; + tensor x_15_cast_fp16 = layer_norm(axes = x_15_axes_0, beta = module_layers_5_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_5_norm2_weight_to_fp16, x = input_61_cast_fp16)[name = string("x_15_cast_fp16")]; + tensor module_layers_6_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20104512))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21333376))))[name = string("module_layers_6_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_6_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_6_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21337280)))]; + tensor linear_24_cast_fp16 = linear(bias = module_layers_6_self_attn_in_proj_bias_to_fp16, weight = module_layers_6_self_attn_in_proj_weight_to_fp16_quantized, x = x_15_cast_fp16)[name = string("linear_24_cast_fp16")]; + tensor var_610_split_sizes_0 = const()[name = string("op_610_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_610_axis_0 = const()[name = string("op_610_axis_0"), val = int32(-1)]; + tensor var_610_cast_fp16_0, tensor var_610_cast_fp16_1, tensor var_610_cast_fp16_2 = split(axis = var_610_axis_0, split_sizes = var_610_split_sizes_0, x = linear_24_cast_fp16)[name = string("op_610_cast_fp16")]; + tensor var_613 = const()[name = string("op_613"), val = tensor([1, 256, 4, 160])]; + tensor var_614_cast_fp16 = reshape(shape = var_613, x = var_610_cast_fp16_0)[name = string("op_614_cast_fp16")]; + tensor var_616 = const()[name = string("op_616"), val = tensor([1, 256, 4, 160])]; + tensor var_617_cast_fp16 = reshape(shape = var_616, x = var_610_cast_fp16_1)[name = string("op_617_cast_fp16")]; + tensor var_619 = const()[name = string("op_619"), val = tensor([1, 256, 4, 160])]; + tensor var_620_cast_fp16 = reshape(shape = var_619, x = var_610_cast_fp16_2)[name = string("op_620_cast_fp16")]; + tensor v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_623_transpose_x_0 = const()[name = string("op_623_transpose_x_0"), val = bool(false)]; + bool var_623_transpose_y_0 = const()[name = string("op_623_transpose_y_0"), val = bool(false)]; + tensor transpose_44_perm_0 = const()[name = string("transpose_44_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_45_perm_0 = const()[name = string("transpose_45_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_45 = transpose(perm = transpose_45_perm_0, x = var_617_cast_fp16)[name = string("transpose_54")]; + tensor transpose_44 = transpose(perm = transpose_44_perm_0, x = var_614_cast_fp16)[name = string("transpose_55")]; + tensor var_623_cast_fp16 = matmul(transpose_x = var_623_transpose_x_0, transpose_y = var_623_transpose_y_0, x = transpose_44, y = transpose_45)[name = string("op_623_cast_fp16")]; + fp16 var_624_to_fp16 = const()[name = string("op_624_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_13_cast_fp16 = mul(x = var_623_cast_fp16, y = var_624_to_fp16)[name = string("scores_13_cast_fp16")]; + tensor weights_37_cast_fp16 = softmax(axis = var_24, x = scores_13_cast_fp16)[name = string("weights_37_cast_fp16")]; + tensor weights_39_cast_fp16 = mul(x = weights_37_cast_fp16, y = var_313_cast_fp16)[name = string("weights_39_cast_fp16")]; + tensor var_635_axes_0 = const()[name = string("op_635_axes_0"), val = tensor([-1])]; + bool var_635_keep_dims_0 = const()[name = string("op_635_keep_dims_0"), val = bool(true)]; + tensor var_635_cast_fp16 = reduce_sum(axes = var_635_axes_0, keep_dims = var_635_keep_dims_0, x = weights_39_cast_fp16)[name = string("op_635_cast_fp16")]; + tensor var_636_cast_fp16 = maximum(x = var_635_cast_fp16, y = var_11_to_fp16)[name = string("op_636_cast_fp16")]; + tensor weights_41_cast_fp16 = real_div(x = weights_39_cast_fp16, y = var_636_cast_fp16)[name = string("weights_41_cast_fp16")]; + bool context_25_transpose_x_0 = const()[name = string("context_25_transpose_x_0"), val = bool(false)]; + bool context_25_transpose_y_0 = const()[name = string("context_25_transpose_y_0"), val = bool(false)]; + tensor v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = var_620_cast_fp16)[name = string("transpose_53")]; + tensor context_25_cast_fp16 = matmul(transpose_x = context_25_transpose_x_0, transpose_y = context_25_transpose_y_0, x = weights_41_cast_fp16, y = v_27_cast_fp16)[name = string("context_25_cast_fp16")]; + tensor var_639_perm_0 = const()[name = string("op_639_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_640 = const()[name = string("op_640"), val = tensor([1, 256, 640])]; + tensor var_639_cast_fp16 = transpose(perm = var_639_perm_0, x = context_25_cast_fp16)[name = string("transpose_52")]; + tensor context_27_cast_fp16 = reshape(shape = var_640, x = var_639_cast_fp16)[name = string("context_27_cast_fp16")]; + tensor module_layers_6_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21341184))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21750848))))[name = string("module_layers_6_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_6_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_6_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21752192)))]; + tensor linear_25_cast_fp16 = linear(bias = module_layers_6_self_attn_out_proj_bias_to_fp16, weight = module_layers_6_self_attn_out_proj_weight_to_fp16_quantized, x = context_27_cast_fp16)[name = string("linear_25_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = x_15_cast_fp16, y = linear_25_cast_fp16)[name = string("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor([-1])]; + tensor module_layers_6_norm1_weight_to_fp16 = const()[name = string("module_layers_6_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21753536)))]; + tensor module_layers_6_norm1_bias_to_fp16 = const()[name = string("module_layers_6_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21754880)))]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = module_layers_6_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_6_norm1_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")]; + tensor module_layers_6_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21756224))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22575488))))[name = string("module_layers_6_linear1_weight_to_fp16_quantized")]; + tensor module_layers_6_linear1_bias_to_fp16 = const()[name = string("module_layers_6_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22578112)))]; + tensor linear_26_cast_fp16 = linear(bias = module_layers_6_linear1_bias_to_fp16, weight = module_layers_6_linear1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_26_cast_fp16")]; + tensor input_69_cast_fp16 = relu(x = linear_26_cast_fp16)[name = string("input_69_cast_fp16")]; + tensor module_layers_6_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22580736))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23400000))))[name = string("module_layers_6_linear2_weight_to_fp16_quantized")]; + tensor module_layers_6_linear2_bias_to_fp16 = const()[name = string("module_layers_6_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23401344)))]; + tensor linear_27_cast_fp16 = linear(bias = module_layers_6_linear2_bias_to_fp16, weight = module_layers_6_linear2_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_27_cast_fp16")]; + tensor input_71_cast_fp16 = add(x = input_65_cast_fp16, y = linear_27_cast_fp16)[name = string("input_71_cast_fp16")]; + tensor x_axes_0 = const()[name = string("x_axes_0"), val = tensor([-1])]; + tensor module_layers_6_norm2_weight_to_fp16 = const()[name = string("module_layers_6_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23402688)))]; + tensor module_layers_6_norm2_bias_to_fp16 = const()[name = string("module_layers_6_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23404032)))]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = module_layers_6_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_6_norm2_weight_to_fp16, x = input_71_cast_fp16)[name = string("x_cast_fp16")]; + tensor module_layers_7_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23405376))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24634240))))[name = string("module_layers_7_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_7_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_7_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24638144)))]; + tensor linear_28_cast_fp16 = linear(bias = module_layers_7_self_attn_in_proj_bias_to_fp16, weight = module_layers_7_self_attn_in_proj_weight_to_fp16_quantized, x = x_cast_fp16)[name = string("linear_28_cast_fp16")]; + tensor var_663_split_sizes_0 = const()[name = string("op_663_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_663_axis_0 = const()[name = string("op_663_axis_0"), val = int32(-1)]; + tensor var_663_cast_fp16_0, tensor var_663_cast_fp16_1, tensor var_663_cast_fp16_2 = split(axis = var_663_axis_0, split_sizes = var_663_split_sizes_0, x = linear_28_cast_fp16)[name = string("op_663_cast_fp16")]; + tensor var_666 = const()[name = string("op_666"), val = tensor([1, 256, 4, 160])]; + tensor var_667_cast_fp16 = reshape(shape = var_666, x = var_663_cast_fp16_0)[name = string("op_667_cast_fp16")]; + tensor var_669 = const()[name = string("op_669"), val = tensor([1, 256, 4, 160])]; + tensor var_670_cast_fp16 = reshape(shape = var_669, x = var_663_cast_fp16_1)[name = string("op_670_cast_fp16")]; + tensor var_672 = const()[name = string("op_672"), val = tensor([1, 256, 4, 160])]; + tensor var_673_cast_fp16 = reshape(shape = var_672, x = var_663_cast_fp16_2)[name = string("op_673_cast_fp16")]; + tensor v_perm_0 = const()[name = string("v_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_676_transpose_x_0 = const()[name = string("op_676_transpose_x_0"), val = bool(false)]; + bool var_676_transpose_y_0 = const()[name = string("op_676_transpose_y_0"), val = bool(false)]; + tensor transpose_46_perm_0 = const()[name = string("transpose_46_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_47_perm_0 = const()[name = string("transpose_47_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_47 = transpose(perm = transpose_47_perm_0, x = var_670_cast_fp16)[name = string("transpose_50")]; + tensor transpose_46 = transpose(perm = transpose_46_perm_0, x = var_667_cast_fp16)[name = string("transpose_51")]; + tensor var_676_cast_fp16 = matmul(transpose_x = var_676_transpose_x_0, transpose_y = var_676_transpose_y_0, x = transpose_46, y = transpose_47)[name = string("op_676_cast_fp16")]; + fp16 var_677_to_fp16 = const()[name = string("op_677_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_cast_fp16 = mul(x = var_676_cast_fp16, y = var_677_to_fp16)[name = string("scores_cast_fp16")]; + tensor weights_43_cast_fp16 = softmax(axis = var_24, x = scores_cast_fp16)[name = string("weights_43_cast_fp16")]; + tensor weights_45_cast_fp16 = mul(x = weights_43_cast_fp16, y = var_313_cast_fp16)[name = string("weights_45_cast_fp16")]; + tensor var_688_axes_0 = const()[name = string("op_688_axes_0"), val = tensor([-1])]; + bool var_688_keep_dims_0 = const()[name = string("op_688_keep_dims_0"), val = bool(true)]; + tensor var_688_cast_fp16 = reduce_sum(axes = var_688_axes_0, keep_dims = var_688_keep_dims_0, x = weights_45_cast_fp16)[name = string("op_688_cast_fp16")]; + tensor var_689_cast_fp16 = maximum(x = var_688_cast_fp16, y = var_11_to_fp16)[name = string("op_689_cast_fp16")]; + tensor weights_cast_fp16 = real_div(x = weights_45_cast_fp16, y = var_689_cast_fp16)[name = string("weights_cast_fp16")]; + bool context_29_transpose_x_0 = const()[name = string("context_29_transpose_x_0"), val = bool(false)]; + bool context_29_transpose_y_0 = const()[name = string("context_29_transpose_y_0"), val = bool(false)]; + tensor v_cast_fp16 = transpose(perm = v_perm_0, x = var_673_cast_fp16)[name = string("transpose_49")]; + tensor context_29_cast_fp16 = matmul(transpose_x = context_29_transpose_x_0, transpose_y = context_29_transpose_y_0, x = weights_cast_fp16, y = v_cast_fp16)[name = string("context_29_cast_fp16")]; + tensor var_692_perm_0 = const()[name = string("op_692_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_693 = const()[name = string("op_693"), val = tensor([1, 256, 640])]; + tensor var_692_cast_fp16 = transpose(perm = var_692_perm_0, x = context_29_cast_fp16)[name = string("transpose_48")]; + tensor context_cast_fp16 = reshape(shape = var_693, x = var_692_cast_fp16)[name = string("context_cast_fp16")]; + tensor module_layers_7_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24642048))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25051712))))[name = string("module_layers_7_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_7_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_7_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25053056)))]; + tensor linear_29_cast_fp16 = linear(bias = module_layers_7_self_attn_out_proj_bias_to_fp16, weight = module_layers_7_self_attn_out_proj_weight_to_fp16_quantized, x = context_cast_fp16)[name = string("linear_29_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = x_cast_fp16, y = linear_29_cast_fp16)[name = string("input_73_cast_fp16")]; + tensor input_75_axes_0 = const()[name = string("input_75_axes_0"), val = tensor([-1])]; + tensor module_layers_7_norm1_weight_to_fp16 = const()[name = string("module_layers_7_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25054400)))]; + tensor module_layers_7_norm1_bias_to_fp16 = const()[name = string("module_layers_7_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25055744)))]; + tensor input_75_cast_fp16 = layer_norm(axes = input_75_axes_0, beta = module_layers_7_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_7_norm1_weight_to_fp16, x = input_73_cast_fp16)[name = string("input_75_cast_fp16")]; + tensor module_layers_7_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25057088))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25876352))))[name = string("module_layers_7_linear1_weight_to_fp16_quantized")]; + tensor module_layers_7_linear1_bias_to_fp16 = const()[name = string("module_layers_7_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25878976)))]; + tensor linear_30_cast_fp16 = linear(bias = module_layers_7_linear1_bias_to_fp16, weight = module_layers_7_linear1_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_30_cast_fp16")]; + tensor input_79_cast_fp16 = relu(x = linear_30_cast_fp16)[name = string("input_79_cast_fp16")]; + tensor module_layers_7_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25881600))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26700864))))[name = string("module_layers_7_linear2_weight_to_fp16_quantized")]; + tensor module_layers_7_linear2_bias_to_fp16 = const()[name = string("module_layers_7_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26702208)))]; + tensor linear_31_cast_fp16 = linear(bias = module_layers_7_linear2_bias_to_fp16, weight = module_layers_7_linear2_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_31_cast_fp16")]; + tensor input_81_cast_fp16 = add(x = input_75_cast_fp16, y = linear_31_cast_fp16)[name = string("input_81_cast_fp16")]; + tensor input_83_axes_0 = const()[name = string("input_83_axes_0"), val = tensor([-1])]; + tensor module_layers_7_norm2_weight_to_fp16 = const()[name = string("module_layers_7_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26703552)))]; + tensor module_layers_7_norm2_bias_to_fp16 = const()[name = string("module_layers_7_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26704896)))]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = module_layers_7_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_7_norm2_weight_to_fp16, x = input_81_cast_fp16)[name = string("input_83_cast_fp16")]; + tensor input_axes_0 = const()[name = string("input_axes_0"), val = tensor([-1])]; + tensor module_encoder_norm_weight_to_fp16 = const()[name = string("module_encoder_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26706240)))]; + tensor module_encoder_norm_bias_to_fp16 = const()[name = string("module_encoder_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26707584)))]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = module_encoder_norm_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_encoder_norm_weight_to_fp16, x = input_83_cast_fp16)[name = string("input_cast_fp16")]; + tensor module_fc_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26708928))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26770432))))[name = string("module_fc_out_weight_to_fp16_quantized")]; + tensor module_fc_out_bias_to_fp16 = const()[name = string("module_fc_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26770688)))]; + tensor linear_32_cast_fp16 = linear(bias = module_fc_out_bias_to_fp16, weight = module_fc_out_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("linear_32_cast_fp16")]; + tensor padded_axes_0 = const()[name = string("padded_axes_0"), val = tensor([-1])]; + tensor padded = expand_dims(axes = padded_axes_0, x = src_pad_mask)[name = string("padded")]; + tensor var_723_begin_0 = const()[name = string("op_723_begin_0"), val = tensor([0, 0, 0])]; + tensor var_723_end_0 = const()[name = string("op_723_end_0"), val = tensor([1, 256, 1])]; + tensor var_723_end_mask_0 = const()[name = string("op_723_end_mask_0"), val = tensor([true, true, false])]; + tensor var_723_cast_fp16 = slice_by_index(begin = var_723_begin_0, end = var_723_end_0, end_mask = var_723_end_mask_0, x = linear_32_cast_fp16)[name = string("op_723_cast_fp16")]; + tensor var_722_to_fp16 = const()[name = string("op_722_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27017344)))]; + tensor blank_logits_cast_fp16 = select(a = var_722_to_fp16, b = var_723_cast_fp16, cond = padded)[name = string("blank_logits_cast_fp16")]; + tensor var_725_begin_0 = const()[name = string("op_725_begin_0"), val = tensor([0, 0, 1])]; + tensor var_725_end_0 = const()[name = string("op_725_end_0"), val = tensor([1, 256, 96])]; + tensor var_725_end_mask_0 = const()[name = string("op_725_end_mask_0"), val = tensor([true, true, true])]; + tensor var_725_cast_fp16 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, x = linear_32_cast_fp16)[name = string("op_725_cast_fp16")]; + fp16 var_726_value_0_to_fp16 = const()[name = string("op_726_value_0_to_fp16"), val = fp16(-0x1.388p+13)]; + tensor var_726_cast_fp16 = fill_like(ref_tensor = var_725_cast_fp16, value = var_726_value_0_to_fp16)[name = string("op_726_cast_fp16")]; + tensor padded_after_broadcast_reps_0 = const()[name = string("padded_after_broadcast_reps_0"), val = tensor([1, 1, 95])]; + tensor padded_after_broadcast = tile(reps = padded_after_broadcast_reps_0, x = padded)[name = string("padded_after_broadcast")]; + tensor nonblank_logits_cast_fp16 = select(a = var_726_cast_fp16, b = var_725_cast_fp16, cond = padded_after_broadcast)[name = string("nonblank_logits_cast_fp16")]; + bool var_730_interleave_0 = const()[name = string("op_730_interleave_0"), val = bool(false)]; + tensor var_730_cast_fp16 = concat(axis = var_24, interleave = var_730_interleave_0, values = (blank_logits_cast_fp16, nonblank_logits_cast_fp16))[name = string("op_730_cast_fp16")]; + string var_730_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_730_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = var_730_cast_fp16_to_fp32_dtype_0, x = var_730_cast_fp16)[name = string("cast_1")]; + } -> (logits); + func g2p_64(tensor text) { + int32 var_24 = const()[name = string("op_24"), val = int32(-1)]; + int32 var_26 = const()[name = string("op_26"), val = int32(0)]; + tensor src_pad_mask = equal(x = text, y = var_26)[name = string("src_pad_mask")]; + int32 x_1_batch_dims_0 = const()[name = string("x_1_batch_dims_0"), val = int32(0)]; + bool x_1_validate_indices_0 = const()[name = string("x_1_validate_indices_0"), val = bool(false)]; + tensor module_embedding_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257408))))[name = string("module_embedding_weight_to_fp16_quantized")]; + string text_to_int16_dtype_0 = const()[name = string("text_to_int16_dtype_0"), val = string("int16")]; + string cast_49_dtype_0 = const()[name = string("cast_49_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor text_to_int16 = cast(dtype = text_to_int16_dtype_0, x = text)[name = string("cast_6")]; + tensor cast_49 = cast(dtype = cast_49_dtype_0, x = text_to_int16)[name = string("cast_5")]; + tensor greater_equal_0 = greater_equal(x = cast_49, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(402)]; + tensor add_0 = add(x = cast_49, y = slice_by_index_0)[name = string("add_0")]; + tensor select_0 = select(a = cast_49, b = add_0, cond = greater_equal_0)[name = string("select_0")]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + string cast_0_dtype_0 = const()[name = string("cast_0_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_4")]; + tensor cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = string("cast_3")]; + tensor greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")]; + int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(402)]; + tensor add_0_1 = add(x = cast_0, y = slice_by_index_0_1)[name = string("add_0_1")]; + tensor select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")]; + int32 x_1_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = string("x_1_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = int32(0)]; + tensor x_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = x_1_cast_fp16_cast_uint16_cast_uint16_axis_0, batch_dims = x_1_batch_dims_0, indices = select_0_1, validate_indices = x_1_validate_indices_0, x = module_embedding_weight_to_fp16_quantized)[name = string("x_1_cast_fp16_cast_uint16_cast_uint16")]; + tensor op_286_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258304))), scale = tensor([[[0x1.724p-7]]]))[name = string("op_286_to_fp16_quantized")]; + tensor input_1_cast_fp16 = add(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = op_286_to_fp16_quantized)[name = string("input_1_cast_fp16")]; + tensor module_layers_0_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1528192))))[name = string("module_layers_0_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_0_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_0_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1532096)))]; + tensor linear_0_cast_fp16 = linear(bias = module_layers_0_self_attn_in_proj_bias_to_fp16, weight = module_layers_0_self_attn_in_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor var_292_split_sizes_0 = const()[name = string("op_292_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_292_axis_0 = const()[name = string("op_292_axis_0"), val = int32(-1)]; + tensor var_292_cast_fp16_0, tensor var_292_cast_fp16_1, tensor var_292_cast_fp16_2 = split(axis = var_292_axis_0, split_sizes = var_292_split_sizes_0, x = linear_0_cast_fp16)[name = string("op_292_cast_fp16")]; + tensor var_295 = const()[name = string("op_295"), val = tensor([1, 64, 4, 160])]; + tensor var_296_cast_fp16 = reshape(shape = var_295, x = var_292_cast_fp16_0)[name = string("op_296_cast_fp16")]; + tensor var_298 = const()[name = string("op_298"), val = tensor([1, 64, 4, 160])]; + tensor var_299_cast_fp16 = reshape(shape = var_298, x = var_292_cast_fp16_1)[name = string("op_299_cast_fp16")]; + tensor var_301 = const()[name = string("op_301"), val = tensor([1, 64, 4, 160])]; + tensor var_302_cast_fp16 = reshape(shape = var_301, x = var_292_cast_fp16_2)[name = string("op_302_cast_fp16")]; + tensor v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_305_transpose_x_0 = const()[name = string("op_305_transpose_x_0"), val = bool(false)]; + bool var_305_transpose_y_0 = const()[name = string("op_305_transpose_y_0"), val = bool(false)]; + tensor transpose_32_perm_0 = const()[name = string("transpose_32_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_33 = transpose(perm = transpose_33_perm_0, x = var_299_cast_fp16)[name = string("transpose_78")]; + tensor transpose_32 = transpose(perm = transpose_32_perm_0, x = var_296_cast_fp16)[name = string("transpose_79")]; + tensor var_305_cast_fp16 = matmul(transpose_x = var_305_transpose_x_0, transpose_y = var_305_transpose_y_0, x = transpose_32, y = transpose_33)[name = string("op_305_cast_fp16")]; + fp16 var_306_to_fp16 = const()[name = string("op_306_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_1_cast_fp16 = mul(x = var_305_cast_fp16, y = var_306_to_fp16)[name = string("scores_1_cast_fp16")]; + tensor weights_1_cast_fp16 = softmax(axis = var_24, x = scores_1_cast_fp16)[name = string("weights_1_cast_fp16")]; + fp16 var_13_to_fp16 = const()[name = string("op_13_to_fp16"), val = fp16(0x1p+0)]; + string var_309_to_fp16_dtype_0 = const()[name = string("op_309_to_fp16_dtype_0"), val = string("fp16")]; + tensor src_pad_mask_to_fp16 = cast(dtype = var_309_to_fp16_dtype_0, x = src_pad_mask)[name = string("cast_2")]; + tensor var_310_cast_fp16 = sub(x = var_13_to_fp16, y = src_pad_mask_to_fp16)[name = string("op_310_cast_fp16")]; + tensor var_312_axes_0 = const()[name = string("op_312_axes_0"), val = tensor([1])]; + tensor var_312_cast_fp16 = expand_dims(axes = var_312_axes_0, x = var_310_cast_fp16)[name = string("op_312_cast_fp16")]; + tensor var_313_axes_0 = const()[name = string("op_313_axes_0"), val = tensor([2])]; + tensor var_313_cast_fp16 = expand_dims(axes = var_313_axes_0, x = var_312_cast_fp16)[name = string("op_313_cast_fp16")]; + tensor weights_3_cast_fp16 = mul(x = weights_1_cast_fp16, y = var_313_cast_fp16)[name = string("weights_3_cast_fp16")]; + tensor var_317_axes_0 = const()[name = string("op_317_axes_0"), val = tensor([-1])]; + bool var_317_keep_dims_0 = const()[name = string("op_317_keep_dims_0"), val = bool(true)]; + tensor var_317_cast_fp16 = reduce_sum(axes = var_317_axes_0, keep_dims = var_317_keep_dims_0, x = weights_3_cast_fp16)[name = string("op_317_cast_fp16")]; + fp16 var_11_to_fp16 = const()[name = string("op_11_to_fp16"), val = fp16(0x1.1p-20)]; + tensor var_318_cast_fp16 = maximum(x = var_317_cast_fp16, y = var_11_to_fp16)[name = string("op_318_cast_fp16")]; + tensor weights_5_cast_fp16 = real_div(x = weights_3_cast_fp16, y = var_318_cast_fp16)[name = string("weights_5_cast_fp16")]; + bool context_1_transpose_x_0 = const()[name = string("context_1_transpose_x_0"), val = bool(false)]; + bool context_1_transpose_y_0 = const()[name = string("context_1_transpose_y_0"), val = bool(false)]; + tensor v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_302_cast_fp16)[name = string("transpose_77")]; + tensor context_1_cast_fp16 = matmul(transpose_x = context_1_transpose_x_0, transpose_y = context_1_transpose_y_0, x = weights_5_cast_fp16, y = v_3_cast_fp16)[name = string("context_1_cast_fp16")]; + tensor var_321_perm_0 = const()[name = string("op_321_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_322 = const()[name = string("op_322"), val = tensor([1, 64, 640])]; + tensor var_321_cast_fp16 = transpose(perm = var_321_perm_0, x = context_1_cast_fp16)[name = string("transpose_76")]; + tensor context_3_cast_fp16 = reshape(shape = var_322, x = var_321_cast_fp16)[name = string("context_3_cast_fp16")]; + tensor module_layers_0_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1536000))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1945664))))[name = string("module_layers_0_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_0_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_0_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1947008)))]; + tensor linear_1_cast_fp16 = linear(bias = module_layers_0_self_attn_out_proj_bias_to_fp16, weight = module_layers_0_self_attn_out_proj_weight_to_fp16_quantized, x = context_3_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor input_3_cast_fp16 = add(x = input_1_cast_fp16, y = linear_1_cast_fp16)[name = string("input_3_cast_fp16")]; + tensor input_5_axes_0 = const()[name = string("input_5_axes_0"), val = tensor([-1])]; + tensor module_layers_0_norm1_weight_to_fp16 = const()[name = string("module_layers_0_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1948352)))]; + tensor module_layers_0_norm1_bias_to_fp16 = const()[name = string("module_layers_0_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1949696)))]; + fp16 var_9_to_fp16 = const()[name = string("op_9_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = module_layers_0_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_0_norm1_weight_to_fp16, x = input_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor module_layers_0_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1951040))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2770304))))[name = string("module_layers_0_linear1_weight_to_fp16_quantized")]; + tensor module_layers_0_linear1_bias_to_fp16 = const()[name = string("module_layers_0_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2772928)))]; + tensor linear_2_cast_fp16 = linear(bias = module_layers_0_linear1_bias_to_fp16, weight = module_layers_0_linear1_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_2_cast_fp16")]; + tensor input_9_cast_fp16 = relu(x = linear_2_cast_fp16)[name = string("input_9_cast_fp16")]; + tensor module_layers_0_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2775552))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3594816))))[name = string("module_layers_0_linear2_weight_to_fp16_quantized")]; + tensor module_layers_0_linear2_bias_to_fp16 = const()[name = string("module_layers_0_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3596160)))]; + tensor linear_3_cast_fp16 = linear(bias = module_layers_0_linear2_bias_to_fp16, weight = module_layers_0_linear2_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("linear_3_cast_fp16")]; + tensor input_11_cast_fp16 = add(x = input_5_cast_fp16, y = linear_3_cast_fp16)[name = string("input_11_cast_fp16")]; + tensor x_5_axes_0 = const()[name = string("x_5_axes_0"), val = tensor([-1])]; + tensor module_layers_0_norm2_weight_to_fp16 = const()[name = string("module_layers_0_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3597504)))]; + tensor module_layers_0_norm2_bias_to_fp16 = const()[name = string("module_layers_0_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3598848)))]; + tensor x_5_cast_fp16 = layer_norm(axes = x_5_axes_0, beta = module_layers_0_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_0_norm2_weight_to_fp16, x = input_11_cast_fp16)[name = string("x_5_cast_fp16")]; + tensor module_layers_1_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3600192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4829056))))[name = string("module_layers_1_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_1_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_1_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4832960)))]; + tensor linear_4_cast_fp16 = linear(bias = module_layers_1_self_attn_in_proj_bias_to_fp16, weight = module_layers_1_self_attn_in_proj_weight_to_fp16_quantized, x = x_5_cast_fp16)[name = string("linear_4_cast_fp16")]; + tensor var_345_split_sizes_0 = const()[name = string("op_345_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_345_axis_0 = const()[name = string("op_345_axis_0"), val = int32(-1)]; + tensor var_345_cast_fp16_0, tensor var_345_cast_fp16_1, tensor var_345_cast_fp16_2 = split(axis = var_345_axis_0, split_sizes = var_345_split_sizes_0, x = linear_4_cast_fp16)[name = string("op_345_cast_fp16")]; + tensor var_348 = const()[name = string("op_348"), val = tensor([1, 64, 4, 160])]; + tensor var_349_cast_fp16 = reshape(shape = var_348, x = var_345_cast_fp16_0)[name = string("op_349_cast_fp16")]; + tensor var_351 = const()[name = string("op_351"), val = tensor([1, 64, 4, 160])]; + tensor var_352_cast_fp16 = reshape(shape = var_351, x = var_345_cast_fp16_1)[name = string("op_352_cast_fp16")]; + tensor var_354 = const()[name = string("op_354"), val = tensor([1, 64, 4, 160])]; + tensor var_355_cast_fp16 = reshape(shape = var_354, x = var_345_cast_fp16_2)[name = string("op_355_cast_fp16")]; + tensor v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_358_transpose_x_0 = const()[name = string("op_358_transpose_x_0"), val = bool(false)]; + bool var_358_transpose_y_0 = const()[name = string("op_358_transpose_y_0"), val = bool(false)]; + tensor transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_35_perm_0 = const()[name = string("transpose_35_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_35 = transpose(perm = transpose_35_perm_0, x = var_352_cast_fp16)[name = string("transpose_74")]; + tensor transpose_34 = transpose(perm = transpose_34_perm_0, x = var_349_cast_fp16)[name = string("transpose_75")]; + tensor var_358_cast_fp16 = matmul(transpose_x = var_358_transpose_x_0, transpose_y = var_358_transpose_y_0, x = transpose_34, y = transpose_35)[name = string("op_358_cast_fp16")]; + fp16 var_359_to_fp16 = const()[name = string("op_359_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_3_cast_fp16 = mul(x = var_358_cast_fp16, y = var_359_to_fp16)[name = string("scores_3_cast_fp16")]; + tensor weights_7_cast_fp16 = softmax(axis = var_24, x = scores_3_cast_fp16)[name = string("weights_7_cast_fp16")]; + tensor weights_9_cast_fp16 = mul(x = weights_7_cast_fp16, y = var_313_cast_fp16)[name = string("weights_9_cast_fp16")]; + tensor var_370_axes_0 = const()[name = string("op_370_axes_0"), val = tensor([-1])]; + bool var_370_keep_dims_0 = const()[name = string("op_370_keep_dims_0"), val = bool(true)]; + tensor var_370_cast_fp16 = reduce_sum(axes = var_370_axes_0, keep_dims = var_370_keep_dims_0, x = weights_9_cast_fp16)[name = string("op_370_cast_fp16")]; + tensor var_371_cast_fp16 = maximum(x = var_370_cast_fp16, y = var_11_to_fp16)[name = string("op_371_cast_fp16")]; + tensor weights_11_cast_fp16 = real_div(x = weights_9_cast_fp16, y = var_371_cast_fp16)[name = string("weights_11_cast_fp16")]; + bool context_5_transpose_x_0 = const()[name = string("context_5_transpose_x_0"), val = bool(false)]; + bool context_5_transpose_y_0 = const()[name = string("context_5_transpose_y_0"), val = bool(false)]; + tensor v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = var_355_cast_fp16)[name = string("transpose_73")]; + tensor context_5_cast_fp16 = matmul(transpose_x = context_5_transpose_x_0, transpose_y = context_5_transpose_y_0, x = weights_11_cast_fp16, y = v_7_cast_fp16)[name = string("context_5_cast_fp16")]; + tensor var_374_perm_0 = const()[name = string("op_374_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_375 = const()[name = string("op_375"), val = tensor([1, 64, 640])]; + tensor var_374_cast_fp16 = transpose(perm = var_374_perm_0, x = context_5_cast_fp16)[name = string("transpose_72")]; + tensor context_7_cast_fp16 = reshape(shape = var_375, x = var_374_cast_fp16)[name = string("context_7_cast_fp16")]; + tensor module_layers_1_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4836864))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5246528))))[name = string("module_layers_1_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_1_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_1_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5247872)))]; + tensor linear_5_cast_fp16 = linear(bias = module_layers_1_self_attn_out_proj_bias_to_fp16, weight = module_layers_1_self_attn_out_proj_weight_to_fp16_quantized, x = context_7_cast_fp16)[name = string("linear_5_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = x_5_cast_fp16, y = linear_5_cast_fp16)[name = string("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = string("input_15_axes_0"), val = tensor([-1])]; + tensor module_layers_1_norm1_weight_to_fp16 = const()[name = string("module_layers_1_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5249216)))]; + tensor module_layers_1_norm1_bias_to_fp16 = const()[name = string("module_layers_1_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5250560)))]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = module_layers_1_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_1_norm1_weight_to_fp16, x = input_13_cast_fp16)[name = string("input_15_cast_fp16")]; + tensor module_layers_1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5251904))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6071168))))[name = string("module_layers_1_linear1_weight_to_fp16_quantized")]; + tensor module_layers_1_linear1_bias_to_fp16 = const()[name = string("module_layers_1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6073792)))]; + tensor linear_6_cast_fp16 = linear(bias = module_layers_1_linear1_bias_to_fp16, weight = module_layers_1_linear1_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_6_cast_fp16")]; + tensor input_19_cast_fp16 = relu(x = linear_6_cast_fp16)[name = string("input_19_cast_fp16")]; + tensor module_layers_1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6076416))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6895680))))[name = string("module_layers_1_linear2_weight_to_fp16_quantized")]; + tensor module_layers_1_linear2_bias_to_fp16 = const()[name = string("module_layers_1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6897024)))]; + tensor linear_7_cast_fp16 = linear(bias = module_layers_1_linear2_bias_to_fp16, weight = module_layers_1_linear2_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_7_cast_fp16")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_7_cast_fp16)[name = string("input_21_cast_fp16")]; + tensor x_7_axes_0 = const()[name = string("x_7_axes_0"), val = tensor([-1])]; + tensor module_layers_1_norm2_weight_to_fp16 = const()[name = string("module_layers_1_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6898368)))]; + tensor module_layers_1_norm2_bias_to_fp16 = const()[name = string("module_layers_1_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6899712)))]; + tensor x_7_cast_fp16 = layer_norm(axes = x_7_axes_0, beta = module_layers_1_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_1_norm2_weight_to_fp16, x = input_21_cast_fp16)[name = string("x_7_cast_fp16")]; + tensor module_layers_2_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6901056))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8129920))))[name = string("module_layers_2_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_2_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_2_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8133824)))]; + tensor linear_8_cast_fp16 = linear(bias = module_layers_2_self_attn_in_proj_bias_to_fp16, weight = module_layers_2_self_attn_in_proj_weight_to_fp16_quantized, x = x_7_cast_fp16)[name = string("linear_8_cast_fp16")]; + tensor var_398_split_sizes_0 = const()[name = string("op_398_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_398_axis_0 = const()[name = string("op_398_axis_0"), val = int32(-1)]; + tensor var_398_cast_fp16_0, tensor var_398_cast_fp16_1, tensor var_398_cast_fp16_2 = split(axis = var_398_axis_0, split_sizes = var_398_split_sizes_0, x = linear_8_cast_fp16)[name = string("op_398_cast_fp16")]; + tensor var_401 = const()[name = string("op_401"), val = tensor([1, 64, 4, 160])]; + tensor var_402_cast_fp16 = reshape(shape = var_401, x = var_398_cast_fp16_0)[name = string("op_402_cast_fp16")]; + tensor var_404 = const()[name = string("op_404"), val = tensor([1, 64, 4, 160])]; + tensor var_405_cast_fp16 = reshape(shape = var_404, x = var_398_cast_fp16_1)[name = string("op_405_cast_fp16")]; + tensor var_407 = const()[name = string("op_407"), val = tensor([1, 64, 4, 160])]; + tensor var_408_cast_fp16 = reshape(shape = var_407, x = var_398_cast_fp16_2)[name = string("op_408_cast_fp16")]; + tensor v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_411_transpose_x_0 = const()[name = string("op_411_transpose_x_0"), val = bool(false)]; + bool var_411_transpose_y_0 = const()[name = string("op_411_transpose_y_0"), val = bool(false)]; + tensor transpose_36_perm_0 = const()[name = string("transpose_36_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_37_perm_0 = const()[name = string("transpose_37_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_37 = transpose(perm = transpose_37_perm_0, x = var_405_cast_fp16)[name = string("transpose_70")]; + tensor transpose_36 = transpose(perm = transpose_36_perm_0, x = var_402_cast_fp16)[name = string("transpose_71")]; + tensor var_411_cast_fp16 = matmul(transpose_x = var_411_transpose_x_0, transpose_y = var_411_transpose_y_0, x = transpose_36, y = transpose_37)[name = string("op_411_cast_fp16")]; + fp16 var_412_to_fp16 = const()[name = string("op_412_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_5_cast_fp16 = mul(x = var_411_cast_fp16, y = var_412_to_fp16)[name = string("scores_5_cast_fp16")]; + tensor weights_13_cast_fp16 = softmax(axis = var_24, x = scores_5_cast_fp16)[name = string("weights_13_cast_fp16")]; + tensor weights_15_cast_fp16 = mul(x = weights_13_cast_fp16, y = var_313_cast_fp16)[name = string("weights_15_cast_fp16")]; + tensor var_423_axes_0 = const()[name = string("op_423_axes_0"), val = tensor([-1])]; + bool var_423_keep_dims_0 = const()[name = string("op_423_keep_dims_0"), val = bool(true)]; + tensor var_423_cast_fp16 = reduce_sum(axes = var_423_axes_0, keep_dims = var_423_keep_dims_0, x = weights_15_cast_fp16)[name = string("op_423_cast_fp16")]; + tensor var_424_cast_fp16 = maximum(x = var_423_cast_fp16, y = var_11_to_fp16)[name = string("op_424_cast_fp16")]; + tensor weights_17_cast_fp16 = real_div(x = weights_15_cast_fp16, y = var_424_cast_fp16)[name = string("weights_17_cast_fp16")]; + bool context_9_transpose_x_0 = const()[name = string("context_9_transpose_x_0"), val = bool(false)]; + bool context_9_transpose_y_0 = const()[name = string("context_9_transpose_y_0"), val = bool(false)]; + tensor v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = var_408_cast_fp16)[name = string("transpose_69")]; + tensor context_9_cast_fp16 = matmul(transpose_x = context_9_transpose_x_0, transpose_y = context_9_transpose_y_0, x = weights_17_cast_fp16, y = v_11_cast_fp16)[name = string("context_9_cast_fp16")]; + tensor var_427_perm_0 = const()[name = string("op_427_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_428 = const()[name = string("op_428"), val = tensor([1, 64, 640])]; + tensor var_427_cast_fp16 = transpose(perm = var_427_perm_0, x = context_9_cast_fp16)[name = string("transpose_68")]; + tensor context_11_cast_fp16 = reshape(shape = var_428, x = var_427_cast_fp16)[name = string("context_11_cast_fp16")]; + tensor module_layers_2_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8137728))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8547392))))[name = string("module_layers_2_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_2_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_2_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8548736)))]; + tensor linear_9_cast_fp16 = linear(bias = module_layers_2_self_attn_out_proj_bias_to_fp16, weight = module_layers_2_self_attn_out_proj_weight_to_fp16_quantized, x = context_11_cast_fp16)[name = string("linear_9_cast_fp16")]; + tensor input_23_cast_fp16 = add(x = x_7_cast_fp16, y = linear_9_cast_fp16)[name = string("input_23_cast_fp16")]; + tensor input_25_axes_0 = const()[name = string("input_25_axes_0"), val = tensor([-1])]; + tensor module_layers_2_norm1_weight_to_fp16 = const()[name = string("module_layers_2_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8550080)))]; + tensor module_layers_2_norm1_bias_to_fp16 = const()[name = string("module_layers_2_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8551424)))]; + tensor input_25_cast_fp16 = layer_norm(axes = input_25_axes_0, beta = module_layers_2_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_2_norm1_weight_to_fp16, x = input_23_cast_fp16)[name = string("input_25_cast_fp16")]; + tensor module_layers_2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8552768))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9372032))))[name = string("module_layers_2_linear1_weight_to_fp16_quantized")]; + tensor module_layers_2_linear1_bias_to_fp16 = const()[name = string("module_layers_2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9374656)))]; + tensor linear_10_cast_fp16 = linear(bias = module_layers_2_linear1_bias_to_fp16, weight = module_layers_2_linear1_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_10_cast_fp16")]; + tensor input_29_cast_fp16 = relu(x = linear_10_cast_fp16)[name = string("input_29_cast_fp16")]; + tensor module_layers_2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9377280))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10196544))))[name = string("module_layers_2_linear2_weight_to_fp16_quantized")]; + tensor module_layers_2_linear2_bias_to_fp16 = const()[name = string("module_layers_2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10197888)))]; + tensor linear_11_cast_fp16 = linear(bias = module_layers_2_linear2_bias_to_fp16, weight = module_layers_2_linear2_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_11_cast_fp16")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_11_cast_fp16)[name = string("input_31_cast_fp16")]; + tensor x_9_axes_0 = const()[name = string("x_9_axes_0"), val = tensor([-1])]; + tensor module_layers_2_norm2_weight_to_fp16 = const()[name = string("module_layers_2_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10199232)))]; + tensor module_layers_2_norm2_bias_to_fp16 = const()[name = string("module_layers_2_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10200576)))]; + tensor x_9_cast_fp16 = layer_norm(axes = x_9_axes_0, beta = module_layers_2_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_2_norm2_weight_to_fp16, x = input_31_cast_fp16)[name = string("x_9_cast_fp16")]; + tensor module_layers_3_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10201920))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11430784))))[name = string("module_layers_3_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_3_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_3_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11434688)))]; + tensor linear_12_cast_fp16 = linear(bias = module_layers_3_self_attn_in_proj_bias_to_fp16, weight = module_layers_3_self_attn_in_proj_weight_to_fp16_quantized, x = x_9_cast_fp16)[name = string("linear_12_cast_fp16")]; + tensor var_451_split_sizes_0 = const()[name = string("op_451_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_451_axis_0 = const()[name = string("op_451_axis_0"), val = int32(-1)]; + tensor var_451_cast_fp16_0, tensor var_451_cast_fp16_1, tensor var_451_cast_fp16_2 = split(axis = var_451_axis_0, split_sizes = var_451_split_sizes_0, x = linear_12_cast_fp16)[name = string("op_451_cast_fp16")]; + tensor var_454 = const()[name = string("op_454"), val = tensor([1, 64, 4, 160])]; + tensor var_455_cast_fp16 = reshape(shape = var_454, x = var_451_cast_fp16_0)[name = string("op_455_cast_fp16")]; + tensor var_457 = const()[name = string("op_457"), val = tensor([1, 64, 4, 160])]; + tensor var_458_cast_fp16 = reshape(shape = var_457, x = var_451_cast_fp16_1)[name = string("op_458_cast_fp16")]; + tensor var_460 = const()[name = string("op_460"), val = tensor([1, 64, 4, 160])]; + tensor var_461_cast_fp16 = reshape(shape = var_460, x = var_451_cast_fp16_2)[name = string("op_461_cast_fp16")]; + tensor v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_464_transpose_x_0 = const()[name = string("op_464_transpose_x_0"), val = bool(false)]; + bool var_464_transpose_y_0 = const()[name = string("op_464_transpose_y_0"), val = bool(false)]; + tensor transpose_38_perm_0 = const()[name = string("transpose_38_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_39_perm_0 = const()[name = string("transpose_39_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_39 = transpose(perm = transpose_39_perm_0, x = var_458_cast_fp16)[name = string("transpose_66")]; + tensor transpose_38 = transpose(perm = transpose_38_perm_0, x = var_455_cast_fp16)[name = string("transpose_67")]; + tensor var_464_cast_fp16 = matmul(transpose_x = var_464_transpose_x_0, transpose_y = var_464_transpose_y_0, x = transpose_38, y = transpose_39)[name = string("op_464_cast_fp16")]; + fp16 var_465_to_fp16 = const()[name = string("op_465_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_7_cast_fp16 = mul(x = var_464_cast_fp16, y = var_465_to_fp16)[name = string("scores_7_cast_fp16")]; + tensor weights_19_cast_fp16 = softmax(axis = var_24, x = scores_7_cast_fp16)[name = string("weights_19_cast_fp16")]; + tensor weights_21_cast_fp16 = mul(x = weights_19_cast_fp16, y = var_313_cast_fp16)[name = string("weights_21_cast_fp16")]; + tensor var_476_axes_0 = const()[name = string("op_476_axes_0"), val = tensor([-1])]; + bool var_476_keep_dims_0 = const()[name = string("op_476_keep_dims_0"), val = bool(true)]; + tensor var_476_cast_fp16 = reduce_sum(axes = var_476_axes_0, keep_dims = var_476_keep_dims_0, x = weights_21_cast_fp16)[name = string("op_476_cast_fp16")]; + tensor var_477_cast_fp16 = maximum(x = var_476_cast_fp16, y = var_11_to_fp16)[name = string("op_477_cast_fp16")]; + tensor weights_23_cast_fp16 = real_div(x = weights_21_cast_fp16, y = var_477_cast_fp16)[name = string("weights_23_cast_fp16")]; + bool context_13_transpose_x_0 = const()[name = string("context_13_transpose_x_0"), val = bool(false)]; + bool context_13_transpose_y_0 = const()[name = string("context_13_transpose_y_0"), val = bool(false)]; + tensor v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = var_461_cast_fp16)[name = string("transpose_65")]; + tensor context_13_cast_fp16 = matmul(transpose_x = context_13_transpose_x_0, transpose_y = context_13_transpose_y_0, x = weights_23_cast_fp16, y = v_15_cast_fp16)[name = string("context_13_cast_fp16")]; + tensor var_480_perm_0 = const()[name = string("op_480_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_481 = const()[name = string("op_481"), val = tensor([1, 64, 640])]; + tensor var_480_cast_fp16 = transpose(perm = var_480_perm_0, x = context_13_cast_fp16)[name = string("transpose_64")]; + tensor context_15_cast_fp16 = reshape(shape = var_481, x = var_480_cast_fp16)[name = string("context_15_cast_fp16")]; + tensor module_layers_3_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11438592))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11848256))))[name = string("module_layers_3_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_3_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_3_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11849600)))]; + tensor linear_13_cast_fp16 = linear(bias = module_layers_3_self_attn_out_proj_bias_to_fp16, weight = module_layers_3_self_attn_out_proj_weight_to_fp16_quantized, x = context_15_cast_fp16)[name = string("linear_13_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = x_9_cast_fp16, y = linear_13_cast_fp16)[name = string("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = string("input_35_axes_0"), val = tensor([-1])]; + tensor module_layers_3_norm1_weight_to_fp16 = const()[name = string("module_layers_3_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11850944)))]; + tensor module_layers_3_norm1_bias_to_fp16 = const()[name = string("module_layers_3_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11852288)))]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = module_layers_3_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_3_norm1_weight_to_fp16, x = input_33_cast_fp16)[name = string("input_35_cast_fp16")]; + tensor module_layers_3_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11853632))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12672896))))[name = string("module_layers_3_linear1_weight_to_fp16_quantized")]; + tensor module_layers_3_linear1_bias_to_fp16 = const()[name = string("module_layers_3_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12675520)))]; + tensor linear_14_cast_fp16 = linear(bias = module_layers_3_linear1_bias_to_fp16, weight = module_layers_3_linear1_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_14_cast_fp16")]; + tensor input_39_cast_fp16 = relu(x = linear_14_cast_fp16)[name = string("input_39_cast_fp16")]; + tensor module_layers_3_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12678144))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13497408))))[name = string("module_layers_3_linear2_weight_to_fp16_quantized")]; + tensor module_layers_3_linear2_bias_to_fp16 = const()[name = string("module_layers_3_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13498752)))]; + tensor linear_15_cast_fp16 = linear(bias = module_layers_3_linear2_bias_to_fp16, weight = module_layers_3_linear2_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_15_cast_fp16")]; + tensor input_41_cast_fp16 = add(x = input_35_cast_fp16, y = linear_15_cast_fp16)[name = string("input_41_cast_fp16")]; + tensor x_11_axes_0 = const()[name = string("x_11_axes_0"), val = tensor([-1])]; + tensor module_layers_3_norm2_weight_to_fp16 = const()[name = string("module_layers_3_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13500096)))]; + tensor module_layers_3_norm2_bias_to_fp16 = const()[name = string("module_layers_3_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13501440)))]; + tensor x_11_cast_fp16 = layer_norm(axes = x_11_axes_0, beta = module_layers_3_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_3_norm2_weight_to_fp16, x = input_41_cast_fp16)[name = string("x_11_cast_fp16")]; + tensor module_layers_4_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13502784))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14731648))))[name = string("module_layers_4_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_4_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_4_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14735552)))]; + tensor linear_16_cast_fp16 = linear(bias = module_layers_4_self_attn_in_proj_bias_to_fp16, weight = module_layers_4_self_attn_in_proj_weight_to_fp16_quantized, x = x_11_cast_fp16)[name = string("linear_16_cast_fp16")]; + tensor var_504_split_sizes_0 = const()[name = string("op_504_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_504_axis_0 = const()[name = string("op_504_axis_0"), val = int32(-1)]; + tensor var_504_cast_fp16_0, tensor var_504_cast_fp16_1, tensor var_504_cast_fp16_2 = split(axis = var_504_axis_0, split_sizes = var_504_split_sizes_0, x = linear_16_cast_fp16)[name = string("op_504_cast_fp16")]; + tensor var_507 = const()[name = string("op_507"), val = tensor([1, 64, 4, 160])]; + tensor var_508_cast_fp16 = reshape(shape = var_507, x = var_504_cast_fp16_0)[name = string("op_508_cast_fp16")]; + tensor var_510 = const()[name = string("op_510"), val = tensor([1, 64, 4, 160])]; + tensor var_511_cast_fp16 = reshape(shape = var_510, x = var_504_cast_fp16_1)[name = string("op_511_cast_fp16")]; + tensor var_513 = const()[name = string("op_513"), val = tensor([1, 64, 4, 160])]; + tensor var_514_cast_fp16 = reshape(shape = var_513, x = var_504_cast_fp16_2)[name = string("op_514_cast_fp16")]; + tensor v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_517_transpose_x_0 = const()[name = string("op_517_transpose_x_0"), val = bool(false)]; + bool var_517_transpose_y_0 = const()[name = string("op_517_transpose_y_0"), val = bool(false)]; + tensor transpose_40_perm_0 = const()[name = string("transpose_40_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_41_perm_0 = const()[name = string("transpose_41_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_41 = transpose(perm = transpose_41_perm_0, x = var_511_cast_fp16)[name = string("transpose_62")]; + tensor transpose_40 = transpose(perm = transpose_40_perm_0, x = var_508_cast_fp16)[name = string("transpose_63")]; + tensor var_517_cast_fp16 = matmul(transpose_x = var_517_transpose_x_0, transpose_y = var_517_transpose_y_0, x = transpose_40, y = transpose_41)[name = string("op_517_cast_fp16")]; + fp16 var_518_to_fp16 = const()[name = string("op_518_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_9_cast_fp16 = mul(x = var_517_cast_fp16, y = var_518_to_fp16)[name = string("scores_9_cast_fp16")]; + tensor weights_25_cast_fp16 = softmax(axis = var_24, x = scores_9_cast_fp16)[name = string("weights_25_cast_fp16")]; + tensor weights_27_cast_fp16 = mul(x = weights_25_cast_fp16, y = var_313_cast_fp16)[name = string("weights_27_cast_fp16")]; + tensor var_529_axes_0 = const()[name = string("op_529_axes_0"), val = tensor([-1])]; + bool var_529_keep_dims_0 = const()[name = string("op_529_keep_dims_0"), val = bool(true)]; + tensor var_529_cast_fp16 = reduce_sum(axes = var_529_axes_0, keep_dims = var_529_keep_dims_0, x = weights_27_cast_fp16)[name = string("op_529_cast_fp16")]; + tensor var_530_cast_fp16 = maximum(x = var_529_cast_fp16, y = var_11_to_fp16)[name = string("op_530_cast_fp16")]; + tensor weights_29_cast_fp16 = real_div(x = weights_27_cast_fp16, y = var_530_cast_fp16)[name = string("weights_29_cast_fp16")]; + bool context_17_transpose_x_0 = const()[name = string("context_17_transpose_x_0"), val = bool(false)]; + bool context_17_transpose_y_0 = const()[name = string("context_17_transpose_y_0"), val = bool(false)]; + tensor v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = var_514_cast_fp16)[name = string("transpose_61")]; + tensor context_17_cast_fp16 = matmul(transpose_x = context_17_transpose_x_0, transpose_y = context_17_transpose_y_0, x = weights_29_cast_fp16, y = v_19_cast_fp16)[name = string("context_17_cast_fp16")]; + tensor var_533_perm_0 = const()[name = string("op_533_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_534 = const()[name = string("op_534"), val = tensor([1, 64, 640])]; + tensor var_533_cast_fp16 = transpose(perm = var_533_perm_0, x = context_17_cast_fp16)[name = string("transpose_60")]; + tensor context_19_cast_fp16 = reshape(shape = var_534, x = var_533_cast_fp16)[name = string("context_19_cast_fp16")]; + tensor module_layers_4_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14739456))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15149120))))[name = string("module_layers_4_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_4_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_4_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15150464)))]; + tensor linear_17_cast_fp16 = linear(bias = module_layers_4_self_attn_out_proj_bias_to_fp16, weight = module_layers_4_self_attn_out_proj_weight_to_fp16_quantized, x = context_19_cast_fp16)[name = string("linear_17_cast_fp16")]; + tensor input_43_cast_fp16 = add(x = x_11_cast_fp16, y = linear_17_cast_fp16)[name = string("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = string("input_45_axes_0"), val = tensor([-1])]; + tensor module_layers_4_norm1_weight_to_fp16 = const()[name = string("module_layers_4_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15151808)))]; + tensor module_layers_4_norm1_bias_to_fp16 = const()[name = string("module_layers_4_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15153152)))]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = module_layers_4_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_4_norm1_weight_to_fp16, x = input_43_cast_fp16)[name = string("input_45_cast_fp16")]; + tensor module_layers_4_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15154496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15973760))))[name = string("module_layers_4_linear1_weight_to_fp16_quantized")]; + tensor module_layers_4_linear1_bias_to_fp16 = const()[name = string("module_layers_4_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15976384)))]; + tensor linear_18_cast_fp16 = linear(bias = module_layers_4_linear1_bias_to_fp16, weight = module_layers_4_linear1_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_18_cast_fp16")]; + tensor input_49_cast_fp16 = relu(x = linear_18_cast_fp16)[name = string("input_49_cast_fp16")]; + tensor module_layers_4_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15979008))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16798272))))[name = string("module_layers_4_linear2_weight_to_fp16_quantized")]; + tensor module_layers_4_linear2_bias_to_fp16 = const()[name = string("module_layers_4_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16799616)))]; + tensor linear_19_cast_fp16 = linear(bias = module_layers_4_linear2_bias_to_fp16, weight = module_layers_4_linear2_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_19_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_19_cast_fp16)[name = string("input_51_cast_fp16")]; + tensor x_13_axes_0 = const()[name = string("x_13_axes_0"), val = tensor([-1])]; + tensor module_layers_4_norm2_weight_to_fp16 = const()[name = string("module_layers_4_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16800960)))]; + tensor module_layers_4_norm2_bias_to_fp16 = const()[name = string("module_layers_4_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16802304)))]; + tensor x_13_cast_fp16 = layer_norm(axes = x_13_axes_0, beta = module_layers_4_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_4_norm2_weight_to_fp16, x = input_51_cast_fp16)[name = string("x_13_cast_fp16")]; + tensor module_layers_5_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16803648))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18032512))))[name = string("module_layers_5_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_5_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_5_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18036416)))]; + tensor linear_20_cast_fp16 = linear(bias = module_layers_5_self_attn_in_proj_bias_to_fp16, weight = module_layers_5_self_attn_in_proj_weight_to_fp16_quantized, x = x_13_cast_fp16)[name = string("linear_20_cast_fp16")]; + tensor var_557_split_sizes_0 = const()[name = string("op_557_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_557_axis_0 = const()[name = string("op_557_axis_0"), val = int32(-1)]; + tensor var_557_cast_fp16_0, tensor var_557_cast_fp16_1, tensor var_557_cast_fp16_2 = split(axis = var_557_axis_0, split_sizes = var_557_split_sizes_0, x = linear_20_cast_fp16)[name = string("op_557_cast_fp16")]; + tensor var_560 = const()[name = string("op_560"), val = tensor([1, 64, 4, 160])]; + tensor var_561_cast_fp16 = reshape(shape = var_560, x = var_557_cast_fp16_0)[name = string("op_561_cast_fp16")]; + tensor var_563 = const()[name = string("op_563"), val = tensor([1, 64, 4, 160])]; + tensor var_564_cast_fp16 = reshape(shape = var_563, x = var_557_cast_fp16_1)[name = string("op_564_cast_fp16")]; + tensor var_566 = const()[name = string("op_566"), val = tensor([1, 64, 4, 160])]; + tensor var_567_cast_fp16 = reshape(shape = var_566, x = var_557_cast_fp16_2)[name = string("op_567_cast_fp16")]; + tensor v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_570_transpose_x_0 = const()[name = string("op_570_transpose_x_0"), val = bool(false)]; + bool var_570_transpose_y_0 = const()[name = string("op_570_transpose_y_0"), val = bool(false)]; + tensor transpose_42_perm_0 = const()[name = string("transpose_42_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_43_perm_0 = const()[name = string("transpose_43_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_43 = transpose(perm = transpose_43_perm_0, x = var_564_cast_fp16)[name = string("transpose_58")]; + tensor transpose_42 = transpose(perm = transpose_42_perm_0, x = var_561_cast_fp16)[name = string("transpose_59")]; + tensor var_570_cast_fp16 = matmul(transpose_x = var_570_transpose_x_0, transpose_y = var_570_transpose_y_0, x = transpose_42, y = transpose_43)[name = string("op_570_cast_fp16")]; + fp16 var_571_to_fp16 = const()[name = string("op_571_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_11_cast_fp16 = mul(x = var_570_cast_fp16, y = var_571_to_fp16)[name = string("scores_11_cast_fp16")]; + tensor weights_31_cast_fp16 = softmax(axis = var_24, x = scores_11_cast_fp16)[name = string("weights_31_cast_fp16")]; + tensor weights_33_cast_fp16 = mul(x = weights_31_cast_fp16, y = var_313_cast_fp16)[name = string("weights_33_cast_fp16")]; + tensor var_582_axes_0 = const()[name = string("op_582_axes_0"), val = tensor([-1])]; + bool var_582_keep_dims_0 = const()[name = string("op_582_keep_dims_0"), val = bool(true)]; + tensor var_582_cast_fp16 = reduce_sum(axes = var_582_axes_0, keep_dims = var_582_keep_dims_0, x = weights_33_cast_fp16)[name = string("op_582_cast_fp16")]; + tensor var_583_cast_fp16 = maximum(x = var_582_cast_fp16, y = var_11_to_fp16)[name = string("op_583_cast_fp16")]; + tensor weights_35_cast_fp16 = real_div(x = weights_33_cast_fp16, y = var_583_cast_fp16)[name = string("weights_35_cast_fp16")]; + bool context_21_transpose_x_0 = const()[name = string("context_21_transpose_x_0"), val = bool(false)]; + bool context_21_transpose_y_0 = const()[name = string("context_21_transpose_y_0"), val = bool(false)]; + tensor v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = var_567_cast_fp16)[name = string("transpose_57")]; + tensor context_21_cast_fp16 = matmul(transpose_x = context_21_transpose_x_0, transpose_y = context_21_transpose_y_0, x = weights_35_cast_fp16, y = v_23_cast_fp16)[name = string("context_21_cast_fp16")]; + tensor var_586_perm_0 = const()[name = string("op_586_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_587 = const()[name = string("op_587"), val = tensor([1, 64, 640])]; + tensor var_586_cast_fp16 = transpose(perm = var_586_perm_0, x = context_21_cast_fp16)[name = string("transpose_56")]; + tensor context_23_cast_fp16 = reshape(shape = var_587, x = var_586_cast_fp16)[name = string("context_23_cast_fp16")]; + tensor module_layers_5_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18040320))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18449984))))[name = string("module_layers_5_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_5_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_5_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18451328)))]; + tensor linear_21_cast_fp16 = linear(bias = module_layers_5_self_attn_out_proj_bias_to_fp16, weight = module_layers_5_self_attn_out_proj_weight_to_fp16_quantized, x = context_23_cast_fp16)[name = string("linear_21_cast_fp16")]; + tensor input_53_cast_fp16 = add(x = x_13_cast_fp16, y = linear_21_cast_fp16)[name = string("input_53_cast_fp16")]; + tensor input_55_axes_0 = const()[name = string("input_55_axes_0"), val = tensor([-1])]; + tensor module_layers_5_norm1_weight_to_fp16 = const()[name = string("module_layers_5_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18452672)))]; + tensor module_layers_5_norm1_bias_to_fp16 = const()[name = string("module_layers_5_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18454016)))]; + tensor input_55_cast_fp16 = layer_norm(axes = input_55_axes_0, beta = module_layers_5_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_5_norm1_weight_to_fp16, x = input_53_cast_fp16)[name = string("input_55_cast_fp16")]; + tensor module_layers_5_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18455360))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19274624))))[name = string("module_layers_5_linear1_weight_to_fp16_quantized")]; + tensor module_layers_5_linear1_bias_to_fp16 = const()[name = string("module_layers_5_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19277248)))]; + tensor linear_22_cast_fp16 = linear(bias = module_layers_5_linear1_bias_to_fp16, weight = module_layers_5_linear1_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_22_cast_fp16")]; + tensor input_59_cast_fp16 = relu(x = linear_22_cast_fp16)[name = string("input_59_cast_fp16")]; + tensor module_layers_5_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19279872))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20099136))))[name = string("module_layers_5_linear2_weight_to_fp16_quantized")]; + tensor module_layers_5_linear2_bias_to_fp16 = const()[name = string("module_layers_5_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20100480)))]; + tensor linear_23_cast_fp16 = linear(bias = module_layers_5_linear2_bias_to_fp16, weight = module_layers_5_linear2_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_23_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_23_cast_fp16)[name = string("input_61_cast_fp16")]; + tensor x_15_axes_0 = const()[name = string("x_15_axes_0"), val = tensor([-1])]; + tensor module_layers_5_norm2_weight_to_fp16 = const()[name = string("module_layers_5_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20101824)))]; + tensor module_layers_5_norm2_bias_to_fp16 = const()[name = string("module_layers_5_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20103168)))]; + tensor x_15_cast_fp16 = layer_norm(axes = x_15_axes_0, beta = module_layers_5_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_5_norm2_weight_to_fp16, x = input_61_cast_fp16)[name = string("x_15_cast_fp16")]; + tensor module_layers_6_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20104512))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21333376))))[name = string("module_layers_6_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_6_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_6_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21337280)))]; + tensor linear_24_cast_fp16 = linear(bias = module_layers_6_self_attn_in_proj_bias_to_fp16, weight = module_layers_6_self_attn_in_proj_weight_to_fp16_quantized, x = x_15_cast_fp16)[name = string("linear_24_cast_fp16")]; + tensor var_610_split_sizes_0 = const()[name = string("op_610_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_610_axis_0 = const()[name = string("op_610_axis_0"), val = int32(-1)]; + tensor var_610_cast_fp16_0, tensor var_610_cast_fp16_1, tensor var_610_cast_fp16_2 = split(axis = var_610_axis_0, split_sizes = var_610_split_sizes_0, x = linear_24_cast_fp16)[name = string("op_610_cast_fp16")]; + tensor var_613 = const()[name = string("op_613"), val = tensor([1, 64, 4, 160])]; + tensor var_614_cast_fp16 = reshape(shape = var_613, x = var_610_cast_fp16_0)[name = string("op_614_cast_fp16")]; + tensor var_616 = const()[name = string("op_616"), val = tensor([1, 64, 4, 160])]; + tensor var_617_cast_fp16 = reshape(shape = var_616, x = var_610_cast_fp16_1)[name = string("op_617_cast_fp16")]; + tensor var_619 = const()[name = string("op_619"), val = tensor([1, 64, 4, 160])]; + tensor var_620_cast_fp16 = reshape(shape = var_619, x = var_610_cast_fp16_2)[name = string("op_620_cast_fp16")]; + tensor v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_623_transpose_x_0 = const()[name = string("op_623_transpose_x_0"), val = bool(false)]; + bool var_623_transpose_y_0 = const()[name = string("op_623_transpose_y_0"), val = bool(false)]; + tensor transpose_44_perm_0 = const()[name = string("transpose_44_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_45_perm_0 = const()[name = string("transpose_45_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_45 = transpose(perm = transpose_45_perm_0, x = var_617_cast_fp16)[name = string("transpose_54")]; + tensor transpose_44 = transpose(perm = transpose_44_perm_0, x = var_614_cast_fp16)[name = string("transpose_55")]; + tensor var_623_cast_fp16 = matmul(transpose_x = var_623_transpose_x_0, transpose_y = var_623_transpose_y_0, x = transpose_44, y = transpose_45)[name = string("op_623_cast_fp16")]; + fp16 var_624_to_fp16 = const()[name = string("op_624_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_13_cast_fp16 = mul(x = var_623_cast_fp16, y = var_624_to_fp16)[name = string("scores_13_cast_fp16")]; + tensor weights_37_cast_fp16 = softmax(axis = var_24, x = scores_13_cast_fp16)[name = string("weights_37_cast_fp16")]; + tensor weights_39_cast_fp16 = mul(x = weights_37_cast_fp16, y = var_313_cast_fp16)[name = string("weights_39_cast_fp16")]; + tensor var_635_axes_0 = const()[name = string("op_635_axes_0"), val = tensor([-1])]; + bool var_635_keep_dims_0 = const()[name = string("op_635_keep_dims_0"), val = bool(true)]; + tensor var_635_cast_fp16 = reduce_sum(axes = var_635_axes_0, keep_dims = var_635_keep_dims_0, x = weights_39_cast_fp16)[name = string("op_635_cast_fp16")]; + tensor var_636_cast_fp16 = maximum(x = var_635_cast_fp16, y = var_11_to_fp16)[name = string("op_636_cast_fp16")]; + tensor weights_41_cast_fp16 = real_div(x = weights_39_cast_fp16, y = var_636_cast_fp16)[name = string("weights_41_cast_fp16")]; + bool context_25_transpose_x_0 = const()[name = string("context_25_transpose_x_0"), val = bool(false)]; + bool context_25_transpose_y_0 = const()[name = string("context_25_transpose_y_0"), val = bool(false)]; + tensor v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = var_620_cast_fp16)[name = string("transpose_53")]; + tensor context_25_cast_fp16 = matmul(transpose_x = context_25_transpose_x_0, transpose_y = context_25_transpose_y_0, x = weights_41_cast_fp16, y = v_27_cast_fp16)[name = string("context_25_cast_fp16")]; + tensor var_639_perm_0 = const()[name = string("op_639_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_640 = const()[name = string("op_640"), val = tensor([1, 64, 640])]; + tensor var_639_cast_fp16 = transpose(perm = var_639_perm_0, x = context_25_cast_fp16)[name = string("transpose_52")]; + tensor context_27_cast_fp16 = reshape(shape = var_640, x = var_639_cast_fp16)[name = string("context_27_cast_fp16")]; + tensor module_layers_6_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21341184))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21750848))))[name = string("module_layers_6_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_6_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_6_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21752192)))]; + tensor linear_25_cast_fp16 = linear(bias = module_layers_6_self_attn_out_proj_bias_to_fp16, weight = module_layers_6_self_attn_out_proj_weight_to_fp16_quantized, x = context_27_cast_fp16)[name = string("linear_25_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = x_15_cast_fp16, y = linear_25_cast_fp16)[name = string("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor([-1])]; + tensor module_layers_6_norm1_weight_to_fp16 = const()[name = string("module_layers_6_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21753536)))]; + tensor module_layers_6_norm1_bias_to_fp16 = const()[name = string("module_layers_6_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21754880)))]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = module_layers_6_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_6_norm1_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")]; + tensor module_layers_6_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21756224))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22575488))))[name = string("module_layers_6_linear1_weight_to_fp16_quantized")]; + tensor module_layers_6_linear1_bias_to_fp16 = const()[name = string("module_layers_6_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22578112)))]; + tensor linear_26_cast_fp16 = linear(bias = module_layers_6_linear1_bias_to_fp16, weight = module_layers_6_linear1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_26_cast_fp16")]; + tensor input_69_cast_fp16 = relu(x = linear_26_cast_fp16)[name = string("input_69_cast_fp16")]; + tensor module_layers_6_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22580736))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23400000))))[name = string("module_layers_6_linear2_weight_to_fp16_quantized")]; + tensor module_layers_6_linear2_bias_to_fp16 = const()[name = string("module_layers_6_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23401344)))]; + tensor linear_27_cast_fp16 = linear(bias = module_layers_6_linear2_bias_to_fp16, weight = module_layers_6_linear2_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_27_cast_fp16")]; + tensor input_71_cast_fp16 = add(x = input_65_cast_fp16, y = linear_27_cast_fp16)[name = string("input_71_cast_fp16")]; + tensor x_axes_0 = const()[name = string("x_axes_0"), val = tensor([-1])]; + tensor module_layers_6_norm2_weight_to_fp16 = const()[name = string("module_layers_6_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23402688)))]; + tensor module_layers_6_norm2_bias_to_fp16 = const()[name = string("module_layers_6_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23404032)))]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = module_layers_6_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_6_norm2_weight_to_fp16, x = input_71_cast_fp16)[name = string("x_cast_fp16")]; + tensor module_layers_7_self_attn_in_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23405376))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24634240))))[name = string("module_layers_7_self_attn_in_proj_weight_to_fp16_quantized")]; + tensor module_layers_7_self_attn_in_proj_bias_to_fp16 = const()[name = string("module_layers_7_self_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24638144)))]; + tensor linear_28_cast_fp16 = linear(bias = module_layers_7_self_attn_in_proj_bias_to_fp16, weight = module_layers_7_self_attn_in_proj_weight_to_fp16_quantized, x = x_cast_fp16)[name = string("linear_28_cast_fp16")]; + tensor var_663_split_sizes_0 = const()[name = string("op_663_split_sizes_0"), val = tensor([640, 640, 640])]; + int32 var_663_axis_0 = const()[name = string("op_663_axis_0"), val = int32(-1)]; + tensor var_663_cast_fp16_0, tensor var_663_cast_fp16_1, tensor var_663_cast_fp16_2 = split(axis = var_663_axis_0, split_sizes = var_663_split_sizes_0, x = linear_28_cast_fp16)[name = string("op_663_cast_fp16")]; + tensor var_666 = const()[name = string("op_666"), val = tensor([1, 64, 4, 160])]; + tensor var_667_cast_fp16 = reshape(shape = var_666, x = var_663_cast_fp16_0)[name = string("op_667_cast_fp16")]; + tensor var_669 = const()[name = string("op_669"), val = tensor([1, 64, 4, 160])]; + tensor var_670_cast_fp16 = reshape(shape = var_669, x = var_663_cast_fp16_1)[name = string("op_670_cast_fp16")]; + tensor var_672 = const()[name = string("op_672"), val = tensor([1, 64, 4, 160])]; + tensor var_673_cast_fp16 = reshape(shape = var_672, x = var_663_cast_fp16_2)[name = string("op_673_cast_fp16")]; + tensor v_perm_0 = const()[name = string("v_perm_0"), val = tensor([0, 2, -3, -1])]; + bool var_676_transpose_x_0 = const()[name = string("op_676_transpose_x_0"), val = bool(false)]; + bool var_676_transpose_y_0 = const()[name = string("op_676_transpose_y_0"), val = bool(false)]; + tensor transpose_46_perm_0 = const()[name = string("transpose_46_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_47_perm_0 = const()[name = string("transpose_47_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_47 = transpose(perm = transpose_47_perm_0, x = var_670_cast_fp16)[name = string("transpose_50")]; + tensor transpose_46 = transpose(perm = transpose_46_perm_0, x = var_667_cast_fp16)[name = string("transpose_51")]; + tensor var_676_cast_fp16 = matmul(transpose_x = var_676_transpose_x_0, transpose_y = var_676_transpose_y_0, x = transpose_46, y = transpose_47)[name = string("op_676_cast_fp16")]; + fp16 var_677_to_fp16 = const()[name = string("op_677_to_fp16"), val = fp16(0x1.43cp-4)]; + tensor scores_cast_fp16 = mul(x = var_676_cast_fp16, y = var_677_to_fp16)[name = string("scores_cast_fp16")]; + tensor weights_43_cast_fp16 = softmax(axis = var_24, x = scores_cast_fp16)[name = string("weights_43_cast_fp16")]; + tensor weights_45_cast_fp16 = mul(x = weights_43_cast_fp16, y = var_313_cast_fp16)[name = string("weights_45_cast_fp16")]; + tensor var_688_axes_0 = const()[name = string("op_688_axes_0"), val = tensor([-1])]; + bool var_688_keep_dims_0 = const()[name = string("op_688_keep_dims_0"), val = bool(true)]; + tensor var_688_cast_fp16 = reduce_sum(axes = var_688_axes_0, keep_dims = var_688_keep_dims_0, x = weights_45_cast_fp16)[name = string("op_688_cast_fp16")]; + tensor var_689_cast_fp16 = maximum(x = var_688_cast_fp16, y = var_11_to_fp16)[name = string("op_689_cast_fp16")]; + tensor weights_cast_fp16 = real_div(x = weights_45_cast_fp16, y = var_689_cast_fp16)[name = string("weights_cast_fp16")]; + bool context_29_transpose_x_0 = const()[name = string("context_29_transpose_x_0"), val = bool(false)]; + bool context_29_transpose_y_0 = const()[name = string("context_29_transpose_y_0"), val = bool(false)]; + tensor v_cast_fp16 = transpose(perm = v_perm_0, x = var_673_cast_fp16)[name = string("transpose_49")]; + tensor context_29_cast_fp16 = matmul(transpose_x = context_29_transpose_x_0, transpose_y = context_29_transpose_y_0, x = weights_cast_fp16, y = v_cast_fp16)[name = string("context_29_cast_fp16")]; + tensor var_692_perm_0 = const()[name = string("op_692_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_693 = const()[name = string("op_693"), val = tensor([1, 64, 640])]; + tensor var_692_cast_fp16 = transpose(perm = var_692_perm_0, x = context_29_cast_fp16)[name = string("transpose_48")]; + tensor context_cast_fp16 = reshape(shape = var_693, x = var_692_cast_fp16)[name = string("context_cast_fp16")]; + tensor module_layers_7_self_attn_out_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24642048))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25051712))))[name = string("module_layers_7_self_attn_out_proj_weight_to_fp16_quantized")]; + tensor module_layers_7_self_attn_out_proj_bias_to_fp16 = const()[name = string("module_layers_7_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25053056)))]; + tensor linear_29_cast_fp16 = linear(bias = module_layers_7_self_attn_out_proj_bias_to_fp16, weight = module_layers_7_self_attn_out_proj_weight_to_fp16_quantized, x = context_cast_fp16)[name = string("linear_29_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = x_cast_fp16, y = linear_29_cast_fp16)[name = string("input_73_cast_fp16")]; + tensor input_75_axes_0 = const()[name = string("input_75_axes_0"), val = tensor([-1])]; + tensor module_layers_7_norm1_weight_to_fp16 = const()[name = string("module_layers_7_norm1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25054400)))]; + tensor module_layers_7_norm1_bias_to_fp16 = const()[name = string("module_layers_7_norm1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25055744)))]; + tensor input_75_cast_fp16 = layer_norm(axes = input_75_axes_0, beta = module_layers_7_norm1_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_7_norm1_weight_to_fp16, x = input_73_cast_fp16)[name = string("input_75_cast_fp16")]; + tensor module_layers_7_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25057088))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25876352))))[name = string("module_layers_7_linear1_weight_to_fp16_quantized")]; + tensor module_layers_7_linear1_bias_to_fp16 = const()[name = string("module_layers_7_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25878976)))]; + tensor linear_30_cast_fp16 = linear(bias = module_layers_7_linear1_bias_to_fp16, weight = module_layers_7_linear1_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_30_cast_fp16")]; + tensor input_79_cast_fp16 = relu(x = linear_30_cast_fp16)[name = string("input_79_cast_fp16")]; + tensor module_layers_7_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25881600))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26700864))))[name = string("module_layers_7_linear2_weight_to_fp16_quantized")]; + tensor module_layers_7_linear2_bias_to_fp16 = const()[name = string("module_layers_7_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26702208)))]; + tensor linear_31_cast_fp16 = linear(bias = module_layers_7_linear2_bias_to_fp16, weight = module_layers_7_linear2_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_31_cast_fp16")]; + tensor input_81_cast_fp16 = add(x = input_75_cast_fp16, y = linear_31_cast_fp16)[name = string("input_81_cast_fp16")]; + tensor input_83_axes_0 = const()[name = string("input_83_axes_0"), val = tensor([-1])]; + tensor module_layers_7_norm2_weight_to_fp16 = const()[name = string("module_layers_7_norm2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26703552)))]; + tensor module_layers_7_norm2_bias_to_fp16 = const()[name = string("module_layers_7_norm2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26704896)))]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = module_layers_7_norm2_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_layers_7_norm2_weight_to_fp16, x = input_81_cast_fp16)[name = string("input_83_cast_fp16")]; + tensor input_axes_0 = const()[name = string("input_axes_0"), val = tensor([-1])]; + tensor module_encoder_norm_weight_to_fp16 = const()[name = string("module_encoder_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26706240)))]; + tensor module_encoder_norm_bias_to_fp16 = const()[name = string("module_encoder_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26707584)))]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = module_encoder_norm_bias_to_fp16, epsilon = var_9_to_fp16, gamma = module_encoder_norm_weight_to_fp16, x = input_83_cast_fp16)[name = string("input_cast_fp16")]; + tensor module_fc_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26708928))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26770432))))[name = string("module_fc_out_weight_to_fp16_quantized")]; + tensor module_fc_out_bias_to_fp16 = const()[name = string("module_fc_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26770688)))]; + tensor linear_32_cast_fp16 = linear(bias = module_fc_out_bias_to_fp16, weight = module_fc_out_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("linear_32_cast_fp16")]; + tensor padded_axes_0 = const()[name = string("padded_axes_0"), val = tensor([-1])]; + tensor padded = expand_dims(axes = padded_axes_0, x = src_pad_mask)[name = string("padded")]; + tensor var_723_begin_0 = const()[name = string("op_723_begin_0"), val = tensor([0, 0, 0])]; + tensor var_723_end_0 = const()[name = string("op_723_end_0"), val = tensor([1, 64, 1])]; + tensor var_723_end_mask_0 = const()[name = string("op_723_end_mask_0"), val = tensor([true, true, false])]; + tensor var_723_cast_fp16 = slice_by_index(begin = var_723_begin_0, end = var_723_end_0, end_mask = var_723_end_mask_0, x = linear_32_cast_fp16)[name = string("op_723_cast_fp16")]; + tensor var_722_to_fp16 = const()[name = string("op_722_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26770944)))]; + tensor blank_logits_cast_fp16 = select(a = var_722_to_fp16, b = var_723_cast_fp16, cond = padded)[name = string("blank_logits_cast_fp16")]; + tensor var_725_begin_0 = const()[name = string("op_725_begin_0"), val = tensor([0, 0, 1])]; + tensor var_725_end_0 = const()[name = string("op_725_end_0"), val = tensor([1, 64, 96])]; + tensor var_725_end_mask_0 = const()[name = string("op_725_end_mask_0"), val = tensor([true, true, true])]; + tensor var_725_cast_fp16 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, x = linear_32_cast_fp16)[name = string("op_725_cast_fp16")]; + fp16 var_726_value_0_to_fp16 = const()[name = string("op_726_value_0_to_fp16"), val = fp16(-0x1.388p+13)]; + tensor var_726_cast_fp16 = fill_like(ref_tensor = var_725_cast_fp16, value = var_726_value_0_to_fp16)[name = string("op_726_cast_fp16")]; + tensor padded_after_broadcast_reps_0 = const()[name = string("padded_after_broadcast_reps_0"), val = tensor([1, 1, 95])]; + tensor padded_after_broadcast = tile(reps = padded_after_broadcast_reps_0, x = padded)[name = string("padded_after_broadcast")]; + tensor nonblank_logits_cast_fp16 = select(a = var_726_cast_fp16, b = var_725_cast_fp16, cond = padded_after_broadcast)[name = string("nonblank_logits_cast_fp16")]; + bool var_730_interleave_0 = const()[name = string("op_730_interleave_0"), val = bool(false)]; + tensor var_730_cast_fp16 = concat(axis = var_24, interleave = var_730_interleave_0, values = (blank_logits_cast_fp16, nonblank_logits_cast_fp16))[name = string("op_730_cast_fp16")]; + string var_730_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_730_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = var_730_cast_fp16_to_fp32_dtype_0, x = var_730_cast_fp16)[name = string("cast_1")]; + } -> (logits); +} \ No newline at end of file