program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})] { func main(tensor mask, tensor tokens) { tensor x_1_batch_dims_0 = const()[name = tensor("x_1_batch_dims_0"), val = tensor(0)]; tensor x_1_validate_indices_0 = const()[name = tensor("x_1_validate_indices_0"), val = tensor(false)]; tensor text_embedding_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("text_embedding_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1816576))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1814144)))]; tensor tokens_to_int16_dtype_0 = const()[name = tensor("tokens_to_int16_dtype_0"), val = tensor("int16")]; tensor cast_14_dtype_0 = const()[name = tensor("cast_14_dtype_0"), val = tensor("int32")]; tensor greater_equal_0_y_0 = const()[name = tensor("greater_equal_0_y_0"), val = tensor(0)]; tensor tokens_to_int16 = cast(dtype = tokens_to_int16_dtype_0, x = tokens)[name = tensor("cast_6")]; tensor cast_14 = cast(dtype = cast_14_dtype_0, x = tokens_to_int16)[name = tensor("cast_5")]; tensor greater_equal_0 = greater_equal(x = cast_14, y = greater_equal_0_y_0)[name = tensor("greater_equal_0")]; tensor slice_by_index_0 = const()[name = tensor("slice_by_index_0"), val = tensor(2362)]; tensor add_0 = add(x = cast_14, y = slice_by_index_0)[name = tensor("add_0")]; tensor select_0 = select(a = cast_14, b = add_0, cond = greater_equal_0)[name = tensor("select_0")]; tensor select_0_to_int16_dtype_0 = const()[name = tensor("select_0_to_int16_dtype_0"), val = tensor("int16")]; tensor cast_0_dtype_0 = const()[name = tensor("cast_0_dtype_0"), val = tensor("int32")]; tensor greater_equal_0_y_0_1 = const()[name = tensor("greater_equal_0_y_0_1"), val = tensor(0)]; tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = tensor("cast_4")]; tensor cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = tensor("cast_3")]; tensor greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = tensor("greater_equal_0_1")]; tensor slice_by_index_0_1 = const()[name = tensor("slice_by_index_0_1"), val = tensor(2362)]; tensor add_0_1 = add(x = cast_0, y = slice_by_index_0_1)[name = tensor("add_0_1")]; tensor select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = tensor("select_0_1")]; tensor x_1_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = tensor("x_1_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = tensor(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 = text_embedding_weight_to_fp16_quantized)[name = tensor("x_1_cast_fp16_cast_uint16_cast_uint16")]; tensor var_26 = const()[name = tensor("op_26"), val = tensor(-1)]; tensor op_50_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("op_50_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1821376))), scale = tensor(0x1.544p-6), zero_point = tensor(0)]; tensor input_5_cast_fp16 = add(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = op_50_to_fp16_quantized)[name = tensor("input_5_cast_fp16")]; tensor var_57_axes_0 = const()[name = tensor("op_57_axes_0"), val = tensor([-1])]; tensor mask_to_fp16_dtype_0 = const()[name = tensor("mask_to_fp16_dtype_0"), val = tensor("fp16")]; tensor mask_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = mask)[name = tensor("cast_2")]; tensor var_57_cast_fp16 = expand_dims(axes = var_57_axes_0, x = mask_to_fp16)[name = tensor("op_57_cast_fp16")]; tensor input_7_cast_fp16 = mul(x = input_5_cast_fp16, y = var_57_cast_fp16)[name = tensor("input_7_cast_fp16")]; tensor query_1_axes_0 = const()[name = tensor("query_1_axes_0"), val = tensor([-1])]; tensor encoder_layers_0_norm_self_weight_to_fp16 = const()[name = tensor("encoder_layers_0_norm_self_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2018048)))]; tensor var_15_to_fp16 = const()[name = tensor("op_15_to_fp16"), val = tensor(0x1.5p-17)]; tensor query_1_cast_fp16 = layer_norm(axes = query_1_axes_0, epsilon = var_15_to_fp16, gamma = encoder_layers_0_norm_self_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("query_1_cast_fp16")]; tensor encoder_layers_0_self_attention_qkv_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_0_self_attention_qkv_net_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2019648))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3791552))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3789184)))]; tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3796224)))]; tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = encoder_layers_0_self_attention_qkv_net_weight_to_fp16_quantized, x = query_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; tensor var_69 = const()[name = tensor("op_69"), val = tensor([1, 256, 3, 12, 64])]; tensor qkv_1_cast_fp16 = reshape(shape = var_69, x = linear_0_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; tensor var_71_split_sizes_0 = const()[name = tensor("op_71_split_sizes_0"), val = tensor([1, 1, 1])]; tensor var_71_axis_0 = const()[name = tensor("op_71_axis_0"), val = tensor(2)]; tensor var_71_cast_fp16_0, tensor var_71_cast_fp16_1, tensor var_71_cast_fp16_2 = split(axis = var_71_axis_0, split_sizes = var_71_split_sizes_0, x = qkv_1_cast_fp16)[name = tensor("op_71_cast_fp16")]; tensor q_3_axes_0 = const()[name = tensor("q_3_axes_0"), val = tensor([2])]; tensor q_3_cast_fp16 = squeeze(axes = q_3_axes_0, x = var_71_cast_fp16_0)[name = tensor("q_3_cast_fp16")]; tensor k_3_axes_0 = const()[name = tensor("k_3_axes_0"), val = tensor([2])]; tensor k_3_cast_fp16 = squeeze(axes = k_3_axes_0, x = var_71_cast_fp16_1)[name = tensor("k_3_cast_fp16")]; tensor v_3_axes_0 = const()[name = tensor("v_3_axes_0"), val = tensor([2])]; tensor v_3_cast_fp16 = squeeze(axes = v_3_axes_0, x = var_71_cast_fp16_2)[name = tensor("v_3_cast_fp16")]; tensor var_77_axes_0 = const()[name = tensor("op_77_axes_0"), val = tensor([1])]; tensor var_77_cast_fp16 = expand_dims(axes = var_77_axes_0, x = mask_to_fp16)[name = tensor("op_77_cast_fp16")]; tensor var_78_axes_0 = const()[name = tensor("op_78_axes_0"), val = tensor([2])]; tensor var_78_cast_fp16 = expand_dims(axes = var_78_axes_0, x = mask_to_fp16)[name = tensor("op_78_cast_fp16")]; tensor mask_3_cast_fp16 = mul(x = var_77_cast_fp16, y = var_78_cast_fp16)[name = tensor("mask_3_cast_fp16")]; tensor mask_5_axes_0 = const()[name = tensor("mask_5_axes_0"), val = tensor([1])]; tensor mask_5_cast_fp16 = expand_dims(axes = mask_5_axes_0, x = mask_3_cast_fp16)[name = tensor("mask_5_cast_fp16")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, -3, -1])]; tensor var_87_transpose_x_0 = const()[name = tensor("op_87_transpose_x_0"), val = tensor(false)]; tensor var_87_transpose_y_0 = const()[name = tensor("op_87_transpose_y_0"), val = tensor(false)]; tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, -3, -1])]; tensor transpose_25_perm_0 = const()[name = tensor("transpose_25_perm_0"), val = tensor([0, 2, -1, -3])]; tensor transpose_25 = transpose(perm = transpose_25_perm_0, x = k_3_cast_fp16)[name = tensor("transpose_70")]; tensor transpose_24 = transpose(perm = transpose_24_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_71")]; tensor var_87_cast_fp16 = matmul(transpose_x = var_87_transpose_x_0, transpose_y = var_87_transpose_y_0, x = transpose_24, y = transpose_25)[name = tensor("op_87_cast_fp16")]; tensor var_88_to_fp16 = const()[name = tensor("op_88_to_fp16"), val = tensor(0x1p-3)]; tensor attn_score_1_cast_fp16 = mul(x = var_87_cast_fp16, y = var_88_to_fp16)[name = tensor("attn_score_1_cast_fp16")]; tensor var_27_promoted_to_fp16 = const()[name = tensor("op_27_promoted_to_fp16"), val = tensor(0x0p+0)]; tensor var_90_cast_fp16 = equal(x = mask_5_cast_fp16, y = var_27_promoted_to_fp16)[name = tensor("op_90_cast_fp16")]; tensor var_18_to_fp16 = const()[name = tensor("op_18_to_fp16"), val = tensor(-inf)]; tensor attn_score_3_cast_fp16 = select(a = var_18_to_fp16, b = attn_score_1_cast_fp16, cond = var_90_cast_fp16)[name = tensor("attn_score_3_cast_fp16")]; tensor input_9_cast_fp16_x_0 = const()[name = tensor("input_9_cast_fp16_x_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3800896)))]; tensor input_9_cast_fp16 = add(x = input_9_cast_fp16_x_0, y = attn_score_3_cast_fp16)[name = tensor("input_9_cast_fp16")]; tensor attn_prob_1_cast_fp16 = softmax(axis = var_26, x = input_9_cast_fp16)[name = tensor("attn_prob_1_cast_fp16")]; tensor var_17_to_fp16 = const()[name = tensor("op_17_to_fp16"), val = tensor(0x0p+0)]; tensor input_11_cast_fp16 = select(a = var_17_to_fp16, b = attn_prob_1_cast_fp16, cond = var_90_cast_fp16)[name = tensor("input_11_cast_fp16")]; tensor y_1_transpose_x_0 = const()[name = tensor("y_1_transpose_x_0"), val = tensor(false)]; tensor y_1_transpose_y_0 = const()[name = tensor("y_1_transpose_y_0"), val = tensor(false)]; tensor v_5_cast_fp16 = transpose(perm = v_5_perm_0, x = v_3_cast_fp16)[name = tensor("transpose_69")]; tensor y_1_cast_fp16 = matmul(transpose_x = y_1_transpose_x_0, transpose_y = y_1_transpose_y_0, x = input_11_cast_fp16, y = v_5_cast_fp16)[name = tensor("y_1_cast_fp16")]; tensor var_101_perm_0 = const()[name = tensor("op_101_perm_0"), val = tensor([0, 2, 1, 3])]; tensor var_103 = const()[name = tensor("op_103"), val = tensor([1, 256, -1])]; tensor var_101_cast_fp16 = transpose(perm = var_101_perm_0, x = y_1_cast_fp16)[name = tensor("transpose_68")]; tensor input_13_cast_fp16 = reshape(shape = var_103, x = var_101_cast_fp16)[name = tensor("input_13_cast_fp16")]; tensor encoder_layers_0_self_attention_o_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_0_self_attention_o_net_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3932032))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4522752))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4521920)))]; tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4524352)))]; tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_0_self_attention_o_net_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = tensor("linear_1_cast_fp16")]; tensor input_17_cast_fp16 = add(x = input_7_cast_fp16, y = linear_1_cast_fp16)[name = tensor("input_17_cast_fp16")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; tensor encoder_layers_0_norm_pos_ff_weight_to_fp16 = const()[name = tensor("encoder_layers_0_norm_pos_ff_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4525952)))]; tensor x_5_cast_fp16 = layer_norm(axes = x_5_axes_0, epsilon = var_15_to_fp16, gamma = encoder_layers_0_norm_pos_ff_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("x_5_cast_fp16")]; tensor signal_1_perm_0 = const()[name = tensor("signal_1_perm_0"), val = tensor([0, 2, 1])]; tensor signal_1_cast_fp16 = transpose(perm = signal_1_perm_0, x = x_5_cast_fp16)[name = tensor("transpose_67")]; tensor input_19_cast_fp16 = mul(x = signal_1_cast_fp16, y = var_77_cast_fp16)[name = tensor("input_19_cast_fp16")]; tensor input_21_pad_0 = const()[name = tensor("input_21_pad_0"), val = tensor([0, 0, 0, 0, 2, 0])]; tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("constant")]; tensor const_7_to_fp16 = const()[name = tensor("const_7_to_fp16"), val = tensor(0x0p+0)]; tensor input_21_cast_fp16 = pad(constant_val = const_7_to_fp16, mode = input_21_mode_0, pad = input_21_pad_0, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; tensor conv_signal_1_pad_type_0 = const()[name = tensor("conv_signal_1_pad_type_0"), val = tensor("valid")]; tensor conv_signal_1_strides_0 = const()[name = tensor("conv_signal_1_strides_0"), val = tensor([1])]; tensor conv_signal_1_pad_0 = const()[name = tensor("conv_signal_1_pad_0"), val = tensor([0, 0])]; tensor conv_signal_1_dilations_0 = const()[name = tensor("conv_signal_1_dilations_0"), val = tensor([1])]; tensor conv_signal_1_groups_0 = const()[name = tensor("conv_signal_1_groups_0"), val = tensor(1)]; tensor encoder_layers_0_pos_ff_proj_conv_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_0_pos_ff_proj_conv_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4527552))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11608640))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11605504)))]; tensor conv_signal_1_cast_fp16 = conv(dilations = conv_signal_1_dilations_0, groups = conv_signal_1_groups_0, pad = conv_signal_1_pad_0, pad_type = conv_signal_1_pad_type_0, strides = conv_signal_1_strides_0, weight = encoder_layers_0_pos_ff_proj_conv_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = tensor("conv_signal_1_cast_fp16")]; tensor input_23_cast_fp16 = mul(x = conv_signal_1_cast_fp16, y = var_77_cast_fp16)[name = tensor("input_23_cast_fp16")]; tensor signal_3_mode_0 = const()[name = tensor("signal_3_mode_0"), val = tensor("TANH_APPROXIMATION")]; tensor signal_3_cast_fp16 = gelu(mode = signal_3_mode_0, x = input_23_cast_fp16)[name = tensor("signal_3_cast_fp16")]; tensor input_25_cast_fp16 = mul(x = signal_3_cast_fp16, y = var_77_cast_fp16)[name = tensor("input_25_cast_fp16")]; tensor input_27_pad_0 = const()[name = tensor("input_27_pad_0"), val = tensor([0, 0, 0, 0, 2, 0])]; tensor input_27_mode_0 = const()[name = tensor("input_27_mode_0"), val = tensor("constant")]; tensor const_8_to_fp16 = const()[name = tensor("const_8_to_fp16"), val = tensor(0x0p+0)]; tensor input_27_cast_fp16 = pad(constant_val = const_8_to_fp16, mode = input_27_mode_0, pad = input_27_pad_0, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; tensor conv_signal_3_pad_type_0 = const()[name = tensor("conv_signal_3_pad_type_0"), val = tensor("valid")]; tensor conv_signal_3_strides_0 = const()[name = tensor("conv_signal_3_strides_0"), val = tensor([1])]; tensor conv_signal_3_pad_0 = const()[name = tensor("conv_signal_3_pad_0"), val = tensor([0, 0])]; tensor conv_signal_3_dilations_0 = const()[name = tensor("conv_signal_3_dilations_0"), val = tensor([1])]; tensor conv_signal_3_groups_0 = const()[name = tensor("conv_signal_3_groups_0"), val = tensor(1)]; tensor encoder_layers_0_pos_ff_o_net_conv_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_0_pos_ff_o_net_conv_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11614848))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18692800))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4521920)))]; tensor conv_signal_3_cast_fp16 = conv(dilations = conv_signal_3_dilations_0, groups = conv_signal_3_groups_0, pad = conv_signal_3_pad_0, pad_type = conv_signal_3_pad_type_0, strides = conv_signal_3_strides_0, weight = encoder_layers_0_pos_ff_o_net_conv_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = tensor("conv_signal_3_cast_fp16")]; tensor var_141_cast_fp16 = mul(x = conv_signal_3_cast_fp16, y = var_77_cast_fp16)[name = tensor("op_141_cast_fp16")]; tensor input_29_perm_0 = const()[name = tensor("input_29_perm_0"), val = tensor([0, 2, 1])]; tensor input_29_cast_fp16 = transpose(perm = input_29_perm_0, x = var_141_cast_fp16)[name = tensor("transpose_66")]; tensor x_7_cast_fp16 = add(x = input_17_cast_fp16, y = input_29_cast_fp16)[name = tensor("x_7_cast_fp16")]; tensor x_9_cast_fp16 = mul(x = x_7_cast_fp16, y = var_57_cast_fp16)[name = tensor("x_9_cast_fp16")]; tensor input_31_cast_fp16 = mul(x = x_9_cast_fp16, y = var_57_cast_fp16)[name = tensor("input_31_cast_fp16")]; tensor query_3_axes_0 = const()[name = tensor("query_3_axes_0"), val = tensor([-1])]; tensor encoder_layers_1_norm_self_weight_to_fp16 = const()[name = tensor("encoder_layers_1_norm_self_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18694400)))]; tensor query_3_cast_fp16 = layer_norm(axes = query_3_axes_0, epsilon = var_15_to_fp16, gamma = encoder_layers_1_norm_self_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("query_3_cast_fp16")]; tensor encoder_layers_1_self_attention_qkv_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_1_self_attention_qkv_net_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18696000))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20465536))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3789184)))]; tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = encoder_layers_1_self_attention_qkv_net_weight_to_fp16_quantized, x = query_3_cast_fp16)[name = tensor("linear_2_cast_fp16")]; tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 256, 3, 12, 64])]; tensor qkv_3_cast_fp16 = reshape(shape = var_163, x = linear_2_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; tensor var_165_split_sizes_0 = const()[name = tensor("op_165_split_sizes_0"), val = tensor([1, 1, 1])]; tensor var_165_axis_0 = const()[name = tensor("op_165_axis_0"), val = tensor(2)]; tensor var_165_cast_fp16_0, tensor var_165_cast_fp16_1, tensor var_165_cast_fp16_2 = split(axis = var_165_axis_0, split_sizes = var_165_split_sizes_0, x = qkv_3_cast_fp16)[name = tensor("op_165_cast_fp16")]; tensor q_9_axes_0 = const()[name = tensor("q_9_axes_0"), val = tensor([2])]; tensor q_9_cast_fp16 = squeeze(axes = q_9_axes_0, x = var_165_cast_fp16_0)[name = tensor("q_9_cast_fp16")]; tensor k_9_axes_0 = const()[name = tensor("k_9_axes_0"), val = tensor([2])]; tensor k_9_cast_fp16 = squeeze(axes = k_9_axes_0, x = var_165_cast_fp16_1)[name = tensor("k_9_cast_fp16")]; tensor v_9_axes_0 = const()[name = tensor("v_9_axes_0"), val = tensor([2])]; tensor v_9_cast_fp16 = squeeze(axes = v_9_axes_0, x = var_165_cast_fp16_2)[name = tensor("v_9_cast_fp16")]; tensor v_11_perm_0 = const()[name = tensor("v_11_perm_0"), val = tensor([0, 2, -3, -1])]; tensor var_181_transpose_x_0 = const()[name = tensor("op_181_transpose_x_0"), val = tensor(false)]; tensor var_181_transpose_y_0 = const()[name = tensor("op_181_transpose_y_0"), val = tensor(false)]; tensor transpose_26_perm_0 = const()[name = tensor("transpose_26_perm_0"), val = tensor([0, 2, -3, -1])]; tensor transpose_27_perm_0 = const()[name = tensor("transpose_27_perm_0"), val = tensor([0, 2, -1, -3])]; tensor transpose_27 = transpose(perm = transpose_27_perm_0, x = k_9_cast_fp16)[name = tensor("transpose_64")]; tensor transpose_26 = transpose(perm = transpose_26_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_65")]; tensor var_181_cast_fp16 = matmul(transpose_x = var_181_transpose_x_0, transpose_y = var_181_transpose_y_0, x = transpose_26, y = transpose_27)[name = tensor("op_181_cast_fp16")]; tensor var_182_to_fp16 = const()[name = tensor("op_182_to_fp16"), val = tensor(0x1p-3)]; tensor attn_score_5_cast_fp16 = mul(x = var_181_cast_fp16, y = var_182_to_fp16)[name = tensor("attn_score_5_cast_fp16")]; tensor attn_score_7_cast_fp16 = select(a = var_18_to_fp16, b = attn_score_5_cast_fp16, cond = var_90_cast_fp16)[name = tensor("attn_score_7_cast_fp16")]; tensor input_33_cast_fp16 = add(x = input_9_cast_fp16_x_0, y = attn_score_7_cast_fp16)[name = tensor("input_33_cast_fp16")]; tensor attn_prob_5_cast_fp16 = softmax(axis = var_26, x = input_33_cast_fp16)[name = tensor("attn_prob_5_cast_fp16")]; tensor input_35_cast_fp16 = select(a = var_17_to_fp16, b = attn_prob_5_cast_fp16, cond = var_90_cast_fp16)[name = tensor("input_35_cast_fp16")]; tensor y_3_transpose_x_0 = const()[name = tensor("y_3_transpose_x_0"), val = tensor(false)]; tensor y_3_transpose_y_0 = const()[name = tensor("y_3_transpose_y_0"), val = tensor(false)]; tensor v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = v_9_cast_fp16)[name = tensor("transpose_63")]; tensor y_3_cast_fp16 = matmul(transpose_x = y_3_transpose_x_0, transpose_y = y_3_transpose_y_0, x = input_35_cast_fp16, y = v_11_cast_fp16)[name = tensor("y_3_cast_fp16")]; tensor var_195_perm_0 = const()[name = tensor("op_195_perm_0"), val = tensor([0, 2, 1, 3])]; tensor var_197 = const()[name = tensor("op_197"), val = tensor([1, 256, -1])]; tensor var_195_cast_fp16 = transpose(perm = var_195_perm_0, x = y_3_cast_fp16)[name = tensor("transpose_62")]; tensor input_37_cast_fp16 = reshape(shape = var_197, x = var_195_cast_fp16)[name = tensor("input_37_cast_fp16")]; tensor encoder_layers_1_self_attention_o_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_1_self_attention_o_net_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20470208))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21060096))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4521920)))]; tensor linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_1_self_attention_o_net_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = tensor("linear_3_cast_fp16")]; tensor input_41_cast_fp16 = add(x = input_31_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_41_cast_fp16")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; tensor encoder_layers_1_norm_pos_ff_weight_to_fp16 = const()[name = tensor("encoder_layers_1_norm_pos_ff_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21061696)))]; tensor x_11_cast_fp16 = layer_norm(axes = x_11_axes_0, epsilon = var_15_to_fp16, gamma = encoder_layers_1_norm_pos_ff_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("x_11_cast_fp16")]; tensor signal_5_perm_0 = const()[name = tensor("signal_5_perm_0"), val = tensor([0, 2, 1])]; tensor signal_5_cast_fp16 = transpose(perm = signal_5_perm_0, x = x_11_cast_fp16)[name = tensor("transpose_61")]; tensor input_43_cast_fp16 = mul(x = signal_5_cast_fp16, y = var_77_cast_fp16)[name = tensor("input_43_cast_fp16")]; tensor input_45_pad_0 = const()[name = tensor("input_45_pad_0"), val = tensor([0, 0, 0, 0, 2, 0])]; tensor input_45_mode_0 = const()[name = tensor("input_45_mode_0"), val = tensor("constant")]; tensor const_13_to_fp16 = const()[name = tensor("const_13_to_fp16"), val = tensor(0x0p+0)]; tensor input_45_cast_fp16 = pad(constant_val = const_13_to_fp16, mode = input_45_mode_0, pad = input_45_pad_0, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; tensor conv_signal_5_pad_type_0 = const()[name = tensor("conv_signal_5_pad_type_0"), val = tensor("valid")]; tensor conv_signal_5_strides_0 = const()[name = tensor("conv_signal_5_strides_0"), val = tensor([1])]; tensor conv_signal_5_pad_0 = const()[name = tensor("conv_signal_5_pad_0"), val = tensor([0, 0])]; tensor conv_signal_5_dilations_0 = const()[name = tensor("conv_signal_5_dilations_0"), val = tensor([1])]; tensor conv_signal_5_groups_0 = const()[name = tensor("conv_signal_5_groups_0"), val = tensor(1)]; tensor encoder_layers_1_pos_ff_proj_conv_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_1_pos_ff_proj_conv_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21063296))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28141248))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11605504)))]; tensor conv_signal_5_cast_fp16 = conv(dilations = conv_signal_5_dilations_0, groups = conv_signal_5_groups_0, pad = conv_signal_5_pad_0, pad_type = conv_signal_5_pad_type_0, strides = conv_signal_5_strides_0, weight = encoder_layers_1_pos_ff_proj_conv_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = tensor("conv_signal_5_cast_fp16")]; tensor input_47_cast_fp16 = mul(x = conv_signal_5_cast_fp16, y = var_77_cast_fp16)[name = tensor("input_47_cast_fp16")]; tensor signal_7_mode_0 = const()[name = tensor("signal_7_mode_0"), val = tensor("TANH_APPROXIMATION")]; tensor signal_7_cast_fp16 = gelu(mode = signal_7_mode_0, x = input_47_cast_fp16)[name = tensor("signal_7_cast_fp16")]; tensor input_49_cast_fp16 = mul(x = signal_7_cast_fp16, y = var_77_cast_fp16)[name = tensor("input_49_cast_fp16")]; tensor input_51_pad_0 = const()[name = tensor("input_51_pad_0"), val = tensor([0, 0, 0, 0, 2, 0])]; tensor input_51_mode_0 = const()[name = tensor("input_51_mode_0"), val = tensor("constant")]; tensor const_14_to_fp16 = const()[name = tensor("const_14_to_fp16"), val = tensor(0x0p+0)]; tensor input_51_cast_fp16 = pad(constant_val = const_14_to_fp16, mode = input_51_mode_0, pad = input_51_pad_0, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; tensor conv_signal_7_pad_type_0 = const()[name = tensor("conv_signal_7_pad_type_0"), val = tensor("valid")]; tensor conv_signal_7_strides_0 = const()[name = tensor("conv_signal_7_strides_0"), val = tensor([1])]; tensor conv_signal_7_pad_0 = const()[name = tensor("conv_signal_7_pad_0"), val = tensor([0, 0])]; tensor conv_signal_7_dilations_0 = const()[name = tensor("conv_signal_7_dilations_0"), val = tensor([1])]; tensor conv_signal_7_groups_0 = const()[name = tensor("conv_signal_7_groups_0"), val = tensor(1)]; tensor encoder_layers_1_pos_ff_o_net_conv_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_1_pos_ff_o_net_conv_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28147456))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35225408))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4521920)))]; tensor conv_signal_7_cast_fp16 = conv(dilations = conv_signal_7_dilations_0, groups = conv_signal_7_groups_0, pad = conv_signal_7_pad_0, pad_type = conv_signal_7_pad_type_0, strides = conv_signal_7_strides_0, weight = encoder_layers_1_pos_ff_o_net_conv_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = tensor("conv_signal_7_cast_fp16")]; tensor var_235_cast_fp16 = mul(x = conv_signal_7_cast_fp16, y = var_77_cast_fp16)[name = tensor("op_235_cast_fp16")]; tensor input_53_perm_0 = const()[name = tensor("input_53_perm_0"), val = tensor([0, 2, 1])]; tensor input_53_cast_fp16 = transpose(perm = input_53_perm_0, x = var_235_cast_fp16)[name = tensor("transpose_60")]; tensor x_13_cast_fp16 = add(x = input_41_cast_fp16, y = input_53_cast_fp16)[name = tensor("x_13_cast_fp16")]; tensor x_15_cast_fp16 = mul(x = x_13_cast_fp16, y = var_57_cast_fp16)[name = tensor("x_15_cast_fp16")]; tensor input_55_cast_fp16 = mul(x = x_15_cast_fp16, y = var_57_cast_fp16)[name = tensor("input_55_cast_fp16")]; tensor query_5_axes_0 = const()[name = tensor("query_5_axes_0"), val = tensor([-1])]; tensor encoder_layers_2_norm_self_weight_to_fp16 = const()[name = tensor("encoder_layers_2_norm_self_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35227008)))]; tensor query_5_cast_fp16 = layer_norm(axes = query_5_axes_0, epsilon = var_15_to_fp16, gamma = encoder_layers_2_norm_self_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("query_5_cast_fp16")]; tensor encoder_layers_2_self_attention_qkv_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_2_self_attention_qkv_net_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35228608))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36998144))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3789184)))]; tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = encoder_layers_2_self_attention_qkv_net_weight_to_fp16_quantized, x = query_5_cast_fp16)[name = tensor("linear_4_cast_fp16")]; tensor var_257 = const()[name = tensor("op_257"), val = tensor([1, 256, 3, 12, 64])]; tensor qkv_5_cast_fp16 = reshape(shape = var_257, x = linear_4_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; tensor var_259_split_sizes_0 = const()[name = tensor("op_259_split_sizes_0"), val = tensor([1, 1, 1])]; tensor var_259_axis_0 = const()[name = tensor("op_259_axis_0"), val = tensor(2)]; tensor var_259_cast_fp16_0, tensor var_259_cast_fp16_1, tensor var_259_cast_fp16_2 = split(axis = var_259_axis_0, split_sizes = var_259_split_sizes_0, x = qkv_5_cast_fp16)[name = tensor("op_259_cast_fp16")]; tensor q_15_axes_0 = const()[name = tensor("q_15_axes_0"), val = tensor([2])]; tensor q_15_cast_fp16 = squeeze(axes = q_15_axes_0, x = var_259_cast_fp16_0)[name = tensor("q_15_cast_fp16")]; tensor k_15_axes_0 = const()[name = tensor("k_15_axes_0"), val = tensor([2])]; tensor k_15_cast_fp16 = squeeze(axes = k_15_axes_0, x = var_259_cast_fp16_1)[name = tensor("k_15_cast_fp16")]; tensor v_15_axes_0 = const()[name = tensor("v_15_axes_0"), val = tensor([2])]; tensor v_15_cast_fp16 = squeeze(axes = v_15_axes_0, x = var_259_cast_fp16_2)[name = tensor("v_15_cast_fp16")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, -3, -1])]; tensor var_275_transpose_x_0 = const()[name = tensor("op_275_transpose_x_0"), val = tensor(false)]; tensor var_275_transpose_y_0 = const()[name = tensor("op_275_transpose_y_0"), val = tensor(false)]; tensor transpose_28_perm_0 = const()[name = tensor("transpose_28_perm_0"), val = tensor([0, 2, -3, -1])]; tensor transpose_29_perm_0 = const()[name = tensor("transpose_29_perm_0"), val = tensor([0, 2, -1, -3])]; tensor transpose_29 = transpose(perm = transpose_29_perm_0, x = k_15_cast_fp16)[name = tensor("transpose_58")]; tensor transpose_28 = transpose(perm = transpose_28_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_59")]; tensor var_275_cast_fp16 = matmul(transpose_x = var_275_transpose_x_0, transpose_y = var_275_transpose_y_0, x = transpose_28, y = transpose_29)[name = tensor("op_275_cast_fp16")]; tensor var_276_to_fp16 = const()[name = tensor("op_276_to_fp16"), val = tensor(0x1p-3)]; tensor attn_score_9_cast_fp16 = mul(x = var_275_cast_fp16, y = var_276_to_fp16)[name = tensor("attn_score_9_cast_fp16")]; tensor attn_score_11_cast_fp16 = select(a = var_18_to_fp16, b = attn_score_9_cast_fp16, cond = var_90_cast_fp16)[name = tensor("attn_score_11_cast_fp16")]; tensor input_57_cast_fp16 = add(x = input_9_cast_fp16_x_0, y = attn_score_11_cast_fp16)[name = tensor("input_57_cast_fp16")]; tensor attn_prob_9_cast_fp16 = softmax(axis = var_26, x = input_57_cast_fp16)[name = tensor("attn_prob_9_cast_fp16")]; tensor input_59_cast_fp16 = select(a = var_17_to_fp16, b = attn_prob_9_cast_fp16, cond = var_90_cast_fp16)[name = tensor("input_59_cast_fp16")]; tensor y_5_transpose_x_0 = const()[name = tensor("y_5_transpose_x_0"), val = tensor(false)]; tensor y_5_transpose_y_0 = const()[name = tensor("y_5_transpose_y_0"), val = tensor(false)]; tensor v_17_cast_fp16 = transpose(perm = v_17_perm_0, x = v_15_cast_fp16)[name = tensor("transpose_57")]; tensor y_5_cast_fp16 = matmul(transpose_x = y_5_transpose_x_0, transpose_y = y_5_transpose_y_0, x = input_59_cast_fp16, y = v_17_cast_fp16)[name = tensor("y_5_cast_fp16")]; tensor var_289_perm_0 = const()[name = tensor("op_289_perm_0"), val = tensor([0, 2, 1, 3])]; tensor var_291 = const()[name = tensor("op_291"), val = tensor([1, 256, -1])]; tensor var_289_cast_fp16 = transpose(perm = var_289_perm_0, x = y_5_cast_fp16)[name = tensor("transpose_56")]; tensor input_61_cast_fp16 = reshape(shape = var_291, x = var_289_cast_fp16)[name = tensor("input_61_cast_fp16")]; tensor encoder_layers_2_self_attention_o_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_2_self_attention_o_net_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37002816))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37592704))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4521920)))]; tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_2_self_attention_o_net_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = tensor("linear_5_cast_fp16")]; tensor input_65_cast_fp16 = add(x = input_55_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_65_cast_fp16")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; tensor encoder_layers_2_norm_pos_ff_weight_to_fp16 = const()[name = tensor("encoder_layers_2_norm_pos_ff_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37594304)))]; tensor x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, epsilon = var_15_to_fp16, gamma = encoder_layers_2_norm_pos_ff_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("x_17_cast_fp16")]; tensor signal_9_perm_0 = const()[name = tensor("signal_9_perm_0"), val = tensor([0, 2, 1])]; tensor signal_9_cast_fp16 = transpose(perm = signal_9_perm_0, x = x_17_cast_fp16)[name = tensor("transpose_55")]; tensor input_67_cast_fp16 = mul(x = signal_9_cast_fp16, y = var_77_cast_fp16)[name = tensor("input_67_cast_fp16")]; tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0, 0, 0, 2, 0])]; tensor input_69_mode_0 = const()[name = tensor("input_69_mode_0"), val = tensor("constant")]; tensor const_19_to_fp16 = const()[name = tensor("const_19_to_fp16"), val = tensor(0x0p+0)]; tensor input_69_cast_fp16 = pad(constant_val = const_19_to_fp16, mode = input_69_mode_0, pad = input_69_pad_0, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; tensor conv_signal_9_pad_type_0 = const()[name = tensor("conv_signal_9_pad_type_0"), val = tensor("valid")]; tensor conv_signal_9_strides_0 = const()[name = tensor("conv_signal_9_strides_0"), val = tensor([1])]; tensor conv_signal_9_pad_0 = const()[name = tensor("conv_signal_9_pad_0"), val = tensor([0, 0])]; tensor conv_signal_9_dilations_0 = const()[name = tensor("conv_signal_9_dilations_0"), val = tensor([1])]; tensor conv_signal_9_groups_0 = const()[name = tensor("conv_signal_9_groups_0"), val = tensor(1)]; tensor encoder_layers_2_pos_ff_proj_conv_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_2_pos_ff_proj_conv_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37595904))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44673856))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11605504)))]; tensor conv_signal_9_cast_fp16 = conv(dilations = conv_signal_9_dilations_0, groups = conv_signal_9_groups_0, pad = conv_signal_9_pad_0, pad_type = conv_signal_9_pad_type_0, strides = conv_signal_9_strides_0, weight = encoder_layers_2_pos_ff_proj_conv_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = tensor("conv_signal_9_cast_fp16")]; tensor input_71_cast_fp16 = mul(x = conv_signal_9_cast_fp16, y = var_77_cast_fp16)[name = tensor("input_71_cast_fp16")]; tensor signal_11_mode_0 = const()[name = tensor("signal_11_mode_0"), val = tensor("TANH_APPROXIMATION")]; tensor signal_11_cast_fp16 = gelu(mode = signal_11_mode_0, x = input_71_cast_fp16)[name = tensor("signal_11_cast_fp16")]; tensor input_73_cast_fp16 = mul(x = signal_11_cast_fp16, y = var_77_cast_fp16)[name = tensor("input_73_cast_fp16")]; tensor input_75_pad_0 = const()[name = tensor("input_75_pad_0"), val = tensor([0, 0, 0, 0, 2, 0])]; tensor input_75_mode_0 = const()[name = tensor("input_75_mode_0"), val = tensor("constant")]; tensor const_20_to_fp16 = const()[name = tensor("const_20_to_fp16"), val = tensor(0x0p+0)]; tensor input_75_cast_fp16 = pad(constant_val = const_20_to_fp16, mode = input_75_mode_0, pad = input_75_pad_0, x = input_73_cast_fp16)[name = tensor("input_75_cast_fp16")]; tensor conv_signal_11_pad_type_0 = const()[name = tensor("conv_signal_11_pad_type_0"), val = tensor("valid")]; tensor conv_signal_11_strides_0 = const()[name = tensor("conv_signal_11_strides_0"), val = tensor([1])]; tensor conv_signal_11_pad_0 = const()[name = tensor("conv_signal_11_pad_0"), val = tensor([0, 0])]; tensor conv_signal_11_dilations_0 = const()[name = tensor("conv_signal_11_dilations_0"), val = tensor([1])]; tensor conv_signal_11_groups_0 = const()[name = tensor("conv_signal_11_groups_0"), val = tensor(1)]; tensor encoder_layers_2_pos_ff_o_net_conv_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_2_pos_ff_o_net_conv_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44680064))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51758016))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4521920)))]; tensor conv_signal_11_cast_fp16 = conv(dilations = conv_signal_11_dilations_0, groups = conv_signal_11_groups_0, pad = conv_signal_11_pad_0, pad_type = conv_signal_11_pad_type_0, strides = conv_signal_11_strides_0, weight = encoder_layers_2_pos_ff_o_net_conv_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = tensor("conv_signal_11_cast_fp16")]; tensor var_329_cast_fp16 = mul(x = conv_signal_11_cast_fp16, y = var_77_cast_fp16)[name = tensor("op_329_cast_fp16")]; tensor input_77_perm_0 = const()[name = tensor("input_77_perm_0"), val = tensor([0, 2, 1])]; tensor input_77_cast_fp16 = transpose(perm = input_77_perm_0, x = var_329_cast_fp16)[name = tensor("transpose_54")]; tensor x_19_cast_fp16 = add(x = input_65_cast_fp16, y = input_77_cast_fp16)[name = tensor("x_19_cast_fp16")]; tensor x_21_cast_fp16 = mul(x = x_19_cast_fp16, y = var_57_cast_fp16)[name = tensor("x_21_cast_fp16")]; tensor input_79_cast_fp16 = mul(x = x_21_cast_fp16, y = var_57_cast_fp16)[name = tensor("input_79_cast_fp16")]; tensor query_7_axes_0 = const()[name = tensor("query_7_axes_0"), val = tensor([-1])]; tensor encoder_layers_3_norm_self_weight_to_fp16 = const()[name = tensor("encoder_layers_3_norm_self_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51759616)))]; tensor query_7_cast_fp16 = layer_norm(axes = query_7_axes_0, epsilon = var_15_to_fp16, gamma = encoder_layers_3_norm_self_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("query_7_cast_fp16")]; tensor encoder_layers_3_self_attention_qkv_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_3_self_attention_qkv_net_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51761216))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53530752))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3789184)))]; tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = encoder_layers_3_self_attention_qkv_net_weight_to_fp16_quantized, x = query_7_cast_fp16)[name = tensor("linear_6_cast_fp16")]; tensor var_351 = const()[name = tensor("op_351"), val = tensor([1, 256, 3, 12, 64])]; tensor qkv_7_cast_fp16 = reshape(shape = var_351, x = linear_6_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; tensor var_353_split_sizes_0 = const()[name = tensor("op_353_split_sizes_0"), val = tensor([1, 1, 1])]; tensor var_353_axis_0 = const()[name = tensor("op_353_axis_0"), val = tensor(2)]; tensor var_353_cast_fp16_0, tensor var_353_cast_fp16_1, tensor var_353_cast_fp16_2 = split(axis = var_353_axis_0, split_sizes = var_353_split_sizes_0, x = qkv_7_cast_fp16)[name = tensor("op_353_cast_fp16")]; tensor q_21_axes_0 = const()[name = tensor("q_21_axes_0"), val = tensor([2])]; tensor q_21_cast_fp16 = squeeze(axes = q_21_axes_0, x = var_353_cast_fp16_0)[name = tensor("q_21_cast_fp16")]; tensor k_21_axes_0 = const()[name = tensor("k_21_axes_0"), val = tensor([2])]; tensor k_21_cast_fp16 = squeeze(axes = k_21_axes_0, x = var_353_cast_fp16_1)[name = tensor("k_21_cast_fp16")]; tensor v_21_axes_0 = const()[name = tensor("v_21_axes_0"), val = tensor([2])]; tensor v_21_cast_fp16 = squeeze(axes = v_21_axes_0, x = var_353_cast_fp16_2)[name = tensor("v_21_cast_fp16")]; tensor v_23_perm_0 = const()[name = tensor("v_23_perm_0"), val = tensor([0, 2, -3, -1])]; tensor var_369_transpose_x_0 = const()[name = tensor("op_369_transpose_x_0"), val = tensor(false)]; tensor var_369_transpose_y_0 = const()[name = tensor("op_369_transpose_y_0"), val = tensor(false)]; tensor transpose_30_perm_0 = const()[name = tensor("transpose_30_perm_0"), val = tensor([0, 2, -3, -1])]; tensor transpose_31_perm_0 = const()[name = tensor("transpose_31_perm_0"), val = tensor([0, 2, -1, -3])]; tensor transpose_31 = transpose(perm = transpose_31_perm_0, x = k_21_cast_fp16)[name = tensor("transpose_52")]; tensor transpose_30 = transpose(perm = transpose_30_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_53")]; tensor var_369_cast_fp16 = matmul(transpose_x = var_369_transpose_x_0, transpose_y = var_369_transpose_y_0, x = transpose_30, y = transpose_31)[name = tensor("op_369_cast_fp16")]; tensor var_370_to_fp16 = const()[name = tensor("op_370_to_fp16"), val = tensor(0x1p-3)]; tensor attn_score_13_cast_fp16 = mul(x = var_369_cast_fp16, y = var_370_to_fp16)[name = tensor("attn_score_13_cast_fp16")]; tensor attn_score_15_cast_fp16 = select(a = var_18_to_fp16, b = attn_score_13_cast_fp16, cond = var_90_cast_fp16)[name = tensor("attn_score_15_cast_fp16")]; tensor input_81_cast_fp16 = add(x = input_9_cast_fp16_x_0, y = attn_score_15_cast_fp16)[name = tensor("input_81_cast_fp16")]; tensor attn_prob_13_cast_fp16 = softmax(axis = var_26, x = input_81_cast_fp16)[name = tensor("attn_prob_13_cast_fp16")]; tensor input_83_cast_fp16 = select(a = var_17_to_fp16, b = attn_prob_13_cast_fp16, cond = var_90_cast_fp16)[name = tensor("input_83_cast_fp16")]; tensor y_7_transpose_x_0 = const()[name = tensor("y_7_transpose_x_0"), val = tensor(false)]; tensor y_7_transpose_y_0 = const()[name = tensor("y_7_transpose_y_0"), val = tensor(false)]; tensor v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = v_21_cast_fp16)[name = tensor("transpose_51")]; tensor y_7_cast_fp16 = matmul(transpose_x = y_7_transpose_x_0, transpose_y = y_7_transpose_y_0, x = input_83_cast_fp16, y = v_23_cast_fp16)[name = tensor("y_7_cast_fp16")]; tensor var_383_perm_0 = const()[name = tensor("op_383_perm_0"), val = tensor([0, 2, 1, 3])]; tensor var_385 = const()[name = tensor("op_385"), val = tensor([1, 256, -1])]; tensor var_383_cast_fp16 = transpose(perm = var_383_perm_0, x = y_7_cast_fp16)[name = tensor("transpose_50")]; tensor input_85_cast_fp16 = reshape(shape = var_385, x = var_383_cast_fp16)[name = tensor("input_85_cast_fp16")]; tensor encoder_layers_3_self_attention_o_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_3_self_attention_o_net_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53535424))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54125312))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4521920)))]; tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_3_self_attention_o_net_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = tensor("linear_7_cast_fp16")]; tensor input_89_cast_fp16 = add(x = input_79_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_89_cast_fp16")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; tensor encoder_layers_3_norm_pos_ff_weight_to_fp16 = const()[name = tensor("encoder_layers_3_norm_pos_ff_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54126912)))]; tensor x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, epsilon = var_15_to_fp16, gamma = encoder_layers_3_norm_pos_ff_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("x_23_cast_fp16")]; tensor signal_13_perm_0 = const()[name = tensor("signal_13_perm_0"), val = tensor([0, 2, 1])]; tensor signal_13_cast_fp16 = transpose(perm = signal_13_perm_0, x = x_23_cast_fp16)[name = tensor("transpose_49")]; tensor input_91_cast_fp16 = mul(x = signal_13_cast_fp16, y = var_77_cast_fp16)[name = tensor("input_91_cast_fp16")]; tensor input_93_pad_0 = const()[name = tensor("input_93_pad_0"), val = tensor([0, 0, 0, 0, 2, 0])]; tensor input_93_mode_0 = const()[name = tensor("input_93_mode_0"), val = tensor("constant")]; tensor const_25_to_fp16 = const()[name = tensor("const_25_to_fp16"), val = tensor(0x0p+0)]; tensor input_93_cast_fp16 = pad(constant_val = const_25_to_fp16, mode = input_93_mode_0, pad = input_93_pad_0, x = input_91_cast_fp16)[name = tensor("input_93_cast_fp16")]; tensor conv_signal_13_pad_type_0 = const()[name = tensor("conv_signal_13_pad_type_0"), val = tensor("valid")]; tensor conv_signal_13_strides_0 = const()[name = tensor("conv_signal_13_strides_0"), val = tensor([1])]; tensor conv_signal_13_pad_0 = const()[name = tensor("conv_signal_13_pad_0"), val = tensor([0, 0])]; tensor conv_signal_13_dilations_0 = const()[name = tensor("conv_signal_13_dilations_0"), val = tensor([1])]; tensor conv_signal_13_groups_0 = const()[name = tensor("conv_signal_13_groups_0"), val = tensor(1)]; tensor encoder_layers_3_pos_ff_proj_conv_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_3_pos_ff_proj_conv_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54128512))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61206464))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11605504)))]; tensor conv_signal_13_cast_fp16 = conv(dilations = conv_signal_13_dilations_0, groups = conv_signal_13_groups_0, pad = conv_signal_13_pad_0, pad_type = conv_signal_13_pad_type_0, strides = conv_signal_13_strides_0, weight = encoder_layers_3_pos_ff_proj_conv_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = tensor("conv_signal_13_cast_fp16")]; tensor input_95_cast_fp16 = mul(x = conv_signal_13_cast_fp16, y = var_77_cast_fp16)[name = tensor("input_95_cast_fp16")]; tensor signal_15_mode_0 = const()[name = tensor("signal_15_mode_0"), val = tensor("TANH_APPROXIMATION")]; tensor signal_15_cast_fp16 = gelu(mode = signal_15_mode_0, x = input_95_cast_fp16)[name = tensor("signal_15_cast_fp16")]; tensor input_97_cast_fp16 = mul(x = signal_15_cast_fp16, y = var_77_cast_fp16)[name = tensor("input_97_cast_fp16")]; tensor input_99_pad_0 = const()[name = tensor("input_99_pad_0"), val = tensor([0, 0, 0, 0, 2, 0])]; tensor input_99_mode_0 = const()[name = tensor("input_99_mode_0"), val = tensor("constant")]; tensor const_26_to_fp16 = const()[name = tensor("const_26_to_fp16"), val = tensor(0x0p+0)]; tensor input_99_cast_fp16 = pad(constant_val = const_26_to_fp16, mode = input_99_mode_0, pad = input_99_pad_0, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; tensor conv_signal_15_pad_type_0 = const()[name = tensor("conv_signal_15_pad_type_0"), val = tensor("valid")]; tensor conv_signal_15_strides_0 = const()[name = tensor("conv_signal_15_strides_0"), val = tensor([1])]; tensor conv_signal_15_pad_0 = const()[name = tensor("conv_signal_15_pad_0"), val = tensor([0, 0])]; tensor conv_signal_15_dilations_0 = const()[name = tensor("conv_signal_15_dilations_0"), val = tensor([1])]; tensor conv_signal_15_groups_0 = const()[name = tensor("conv_signal_15_groups_0"), val = tensor(1)]; tensor encoder_layers_3_pos_ff_o_net_conv_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_3_pos_ff_o_net_conv_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61212672))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68290624))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4521920)))]; tensor conv_signal_15_cast_fp16 = conv(dilations = conv_signal_15_dilations_0, groups = conv_signal_15_groups_0, pad = conv_signal_15_pad_0, pad_type = conv_signal_15_pad_type_0, strides = conv_signal_15_strides_0, weight = encoder_layers_3_pos_ff_o_net_conv_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = tensor("conv_signal_15_cast_fp16")]; tensor var_423_cast_fp16 = mul(x = conv_signal_15_cast_fp16, y = var_77_cast_fp16)[name = tensor("op_423_cast_fp16")]; tensor input_101_perm_0 = const()[name = tensor("input_101_perm_0"), val = tensor([0, 2, 1])]; tensor input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = var_423_cast_fp16)[name = tensor("transpose_48")]; tensor x_25_cast_fp16 = add(x = input_89_cast_fp16, y = input_101_cast_fp16)[name = tensor("x_25_cast_fp16")]; tensor x_27_cast_fp16 = mul(x = x_25_cast_fp16, y = var_57_cast_fp16)[name = tensor("x_27_cast_fp16")]; tensor input_103_cast_fp16 = mul(x = x_27_cast_fp16, y = var_57_cast_fp16)[name = tensor("input_103_cast_fp16")]; tensor query_9_axes_0 = const()[name = tensor("query_9_axes_0"), val = tensor([-1])]; tensor encoder_layers_4_norm_self_weight_to_fp16 = const()[name = tensor("encoder_layers_4_norm_self_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68292224)))]; tensor query_9_cast_fp16 = layer_norm(axes = query_9_axes_0, epsilon = var_15_to_fp16, gamma = encoder_layers_4_norm_self_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("query_9_cast_fp16")]; tensor encoder_layers_4_self_attention_qkv_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_4_self_attention_qkv_net_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68293824))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70063360))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3789184)))]; tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = encoder_layers_4_self_attention_qkv_net_weight_to_fp16_quantized, x = query_9_cast_fp16)[name = tensor("linear_8_cast_fp16")]; tensor var_445 = const()[name = tensor("op_445"), val = tensor([1, 256, 3, 12, 64])]; tensor qkv_9_cast_fp16 = reshape(shape = var_445, x = linear_8_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; tensor var_447_split_sizes_0 = const()[name = tensor("op_447_split_sizes_0"), val = tensor([1, 1, 1])]; tensor var_447_axis_0 = const()[name = tensor("op_447_axis_0"), val = tensor(2)]; tensor var_447_cast_fp16_0, tensor var_447_cast_fp16_1, tensor var_447_cast_fp16_2 = split(axis = var_447_axis_0, split_sizes = var_447_split_sizes_0, x = qkv_9_cast_fp16)[name = tensor("op_447_cast_fp16")]; tensor q_27_axes_0 = const()[name = tensor("q_27_axes_0"), val = tensor([2])]; tensor q_27_cast_fp16 = squeeze(axes = q_27_axes_0, x = var_447_cast_fp16_0)[name = tensor("q_27_cast_fp16")]; tensor k_27_axes_0 = const()[name = tensor("k_27_axes_0"), val = tensor([2])]; tensor k_27_cast_fp16 = squeeze(axes = k_27_axes_0, x = var_447_cast_fp16_1)[name = tensor("k_27_cast_fp16")]; tensor v_27_axes_0 = const()[name = tensor("v_27_axes_0"), val = tensor([2])]; tensor v_27_cast_fp16 = squeeze(axes = v_27_axes_0, x = var_447_cast_fp16_2)[name = tensor("v_27_cast_fp16")]; tensor v_29_perm_0 = const()[name = tensor("v_29_perm_0"), val = tensor([0, 2, -3, -1])]; tensor var_463_transpose_x_0 = const()[name = tensor("op_463_transpose_x_0"), val = tensor(false)]; tensor var_463_transpose_y_0 = const()[name = tensor("op_463_transpose_y_0"), val = tensor(false)]; tensor transpose_32_perm_0 = const()[name = tensor("transpose_32_perm_0"), val = tensor([0, 2, -3, -1])]; tensor transpose_33_perm_0 = const()[name = tensor("transpose_33_perm_0"), val = tensor([0, 2, -1, -3])]; tensor transpose_33 = transpose(perm = transpose_33_perm_0, x = k_27_cast_fp16)[name = tensor("transpose_46")]; tensor transpose_32 = transpose(perm = transpose_32_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_47")]; tensor var_463_cast_fp16 = matmul(transpose_x = var_463_transpose_x_0, transpose_y = var_463_transpose_y_0, x = transpose_32, y = transpose_33)[name = tensor("op_463_cast_fp16")]; tensor var_464_to_fp16 = const()[name = tensor("op_464_to_fp16"), val = tensor(0x1p-3)]; tensor attn_score_17_cast_fp16 = mul(x = var_463_cast_fp16, y = var_464_to_fp16)[name = tensor("attn_score_17_cast_fp16")]; tensor attn_score_19_cast_fp16 = select(a = var_18_to_fp16, b = attn_score_17_cast_fp16, cond = var_90_cast_fp16)[name = tensor("attn_score_19_cast_fp16")]; tensor input_105_cast_fp16 = add(x = input_9_cast_fp16_x_0, y = attn_score_19_cast_fp16)[name = tensor("input_105_cast_fp16")]; tensor attn_prob_17_cast_fp16 = softmax(axis = var_26, x = input_105_cast_fp16)[name = tensor("attn_prob_17_cast_fp16")]; tensor input_107_cast_fp16 = select(a = var_17_to_fp16, b = attn_prob_17_cast_fp16, cond = var_90_cast_fp16)[name = tensor("input_107_cast_fp16")]; tensor y_9_transpose_x_0 = const()[name = tensor("y_9_transpose_x_0"), val = tensor(false)]; tensor y_9_transpose_y_0 = const()[name = tensor("y_9_transpose_y_0"), val = tensor(false)]; tensor v_29_cast_fp16 = transpose(perm = v_29_perm_0, x = v_27_cast_fp16)[name = tensor("transpose_45")]; tensor y_9_cast_fp16 = matmul(transpose_x = y_9_transpose_x_0, transpose_y = y_9_transpose_y_0, x = input_107_cast_fp16, y = v_29_cast_fp16)[name = tensor("y_9_cast_fp16")]; tensor var_477_perm_0 = const()[name = tensor("op_477_perm_0"), val = tensor([0, 2, 1, 3])]; tensor var_479 = const()[name = tensor("op_479"), val = tensor([1, 256, -1])]; tensor var_477_cast_fp16 = transpose(perm = var_477_perm_0, x = y_9_cast_fp16)[name = tensor("transpose_44")]; tensor input_109_cast_fp16 = reshape(shape = var_479, x = var_477_cast_fp16)[name = tensor("input_109_cast_fp16")]; tensor encoder_layers_4_self_attention_o_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_4_self_attention_o_net_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70068032))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70657920))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4521920)))]; tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_4_self_attention_o_net_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = tensor("linear_9_cast_fp16")]; tensor input_113_cast_fp16 = add(x = input_103_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_113_cast_fp16")]; tensor x_29_axes_0 = const()[name = tensor("x_29_axes_0"), val = tensor([-1])]; tensor encoder_layers_4_norm_pos_ff_weight_to_fp16 = const()[name = tensor("encoder_layers_4_norm_pos_ff_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70659520)))]; tensor x_29_cast_fp16 = layer_norm(axes = x_29_axes_0, epsilon = var_15_to_fp16, gamma = encoder_layers_4_norm_pos_ff_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("x_29_cast_fp16")]; tensor signal_17_perm_0 = const()[name = tensor("signal_17_perm_0"), val = tensor([0, 2, 1])]; tensor signal_17_cast_fp16 = transpose(perm = signal_17_perm_0, x = x_29_cast_fp16)[name = tensor("transpose_43")]; tensor input_115_cast_fp16 = mul(x = signal_17_cast_fp16, y = var_77_cast_fp16)[name = tensor("input_115_cast_fp16")]; tensor input_117_pad_0 = const()[name = tensor("input_117_pad_0"), val = tensor([0, 0, 0, 0, 2, 0])]; tensor input_117_mode_0 = const()[name = tensor("input_117_mode_0"), val = tensor("constant")]; tensor const_31_to_fp16 = const()[name = tensor("const_31_to_fp16"), val = tensor(0x0p+0)]; tensor input_117_cast_fp16 = pad(constant_val = const_31_to_fp16, mode = input_117_mode_0, pad = input_117_pad_0, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; tensor conv_signal_17_pad_type_0 = const()[name = tensor("conv_signal_17_pad_type_0"), val = tensor("valid")]; tensor conv_signal_17_strides_0 = const()[name = tensor("conv_signal_17_strides_0"), val = tensor([1])]; tensor conv_signal_17_pad_0 = const()[name = tensor("conv_signal_17_pad_0"), val = tensor([0, 0])]; tensor conv_signal_17_dilations_0 = const()[name = tensor("conv_signal_17_dilations_0"), val = tensor([1])]; tensor conv_signal_17_groups_0 = const()[name = tensor("conv_signal_17_groups_0"), val = tensor(1)]; tensor encoder_layers_4_pos_ff_proj_conv_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_4_pos_ff_proj_conv_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70661120))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77739072))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11605504)))]; tensor conv_signal_17_cast_fp16 = conv(dilations = conv_signal_17_dilations_0, groups = conv_signal_17_groups_0, pad = conv_signal_17_pad_0, pad_type = conv_signal_17_pad_type_0, strides = conv_signal_17_strides_0, weight = encoder_layers_4_pos_ff_proj_conv_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = tensor("conv_signal_17_cast_fp16")]; tensor input_119_cast_fp16 = mul(x = conv_signal_17_cast_fp16, y = var_77_cast_fp16)[name = tensor("input_119_cast_fp16")]; tensor signal_19_mode_0 = const()[name = tensor("signal_19_mode_0"), val = tensor("TANH_APPROXIMATION")]; tensor signal_19_cast_fp16 = gelu(mode = signal_19_mode_0, x = input_119_cast_fp16)[name = tensor("signal_19_cast_fp16")]; tensor input_121_cast_fp16 = mul(x = signal_19_cast_fp16, y = var_77_cast_fp16)[name = tensor("input_121_cast_fp16")]; tensor input_123_pad_0 = const()[name = tensor("input_123_pad_0"), val = tensor([0, 0, 0, 0, 2, 0])]; tensor input_123_mode_0 = const()[name = tensor("input_123_mode_0"), val = tensor("constant")]; tensor const_32_to_fp16 = const()[name = tensor("const_32_to_fp16"), val = tensor(0x0p+0)]; tensor input_123_cast_fp16 = pad(constant_val = const_32_to_fp16, mode = input_123_mode_0, pad = input_123_pad_0, x = input_121_cast_fp16)[name = tensor("input_123_cast_fp16")]; tensor conv_signal_19_pad_type_0 = const()[name = tensor("conv_signal_19_pad_type_0"), val = tensor("valid")]; tensor conv_signal_19_strides_0 = const()[name = tensor("conv_signal_19_strides_0"), val = tensor([1])]; tensor conv_signal_19_pad_0 = const()[name = tensor("conv_signal_19_pad_0"), val = tensor([0, 0])]; tensor conv_signal_19_dilations_0 = const()[name = tensor("conv_signal_19_dilations_0"), val = tensor([1])]; tensor conv_signal_19_groups_0 = const()[name = tensor("conv_signal_19_groups_0"), val = tensor(1)]; tensor encoder_layers_4_pos_ff_o_net_conv_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_4_pos_ff_o_net_conv_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77745280))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84823232))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4521920)))]; tensor conv_signal_19_cast_fp16 = conv(dilations = conv_signal_19_dilations_0, groups = conv_signal_19_groups_0, pad = conv_signal_19_pad_0, pad_type = conv_signal_19_pad_type_0, strides = conv_signal_19_strides_0, weight = encoder_layers_4_pos_ff_o_net_conv_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = tensor("conv_signal_19_cast_fp16")]; tensor var_517_cast_fp16 = mul(x = conv_signal_19_cast_fp16, y = var_77_cast_fp16)[name = tensor("op_517_cast_fp16")]; tensor input_125_perm_0 = const()[name = tensor("input_125_perm_0"), val = tensor([0, 2, 1])]; tensor input_125_cast_fp16 = transpose(perm = input_125_perm_0, x = var_517_cast_fp16)[name = tensor("transpose_42")]; tensor x_31_cast_fp16 = add(x = input_113_cast_fp16, y = input_125_cast_fp16)[name = tensor("x_31_cast_fp16")]; tensor x_33_cast_fp16 = mul(x = x_31_cast_fp16, y = var_57_cast_fp16)[name = tensor("x_33_cast_fp16")]; tensor input_127_cast_fp16 = mul(x = x_33_cast_fp16, y = var_57_cast_fp16)[name = tensor("input_127_cast_fp16")]; tensor query_axes_0 = const()[name = tensor("query_axes_0"), val = tensor([-1])]; tensor encoder_layers_5_norm_self_weight_to_fp16 = const()[name = tensor("encoder_layers_5_norm_self_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84824832)))]; tensor query_cast_fp16 = layer_norm(axes = query_axes_0, epsilon = var_15_to_fp16, gamma = encoder_layers_5_norm_self_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("query_cast_fp16")]; tensor encoder_layers_5_self_attention_qkv_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_5_self_attention_qkv_net_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84826432))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86595968))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3789184)))]; tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = encoder_layers_5_self_attention_qkv_net_weight_to_fp16_quantized, x = query_cast_fp16)[name = tensor("linear_10_cast_fp16")]; tensor var_539 = const()[name = tensor("op_539"), val = tensor([1, 256, 3, 12, 64])]; tensor qkv_cast_fp16 = reshape(shape = var_539, x = linear_10_cast_fp16)[name = tensor("qkv_cast_fp16")]; tensor var_541_split_sizes_0 = const()[name = tensor("op_541_split_sizes_0"), val = tensor([1, 1, 1])]; tensor var_541_axis_0 = const()[name = tensor("op_541_axis_0"), val = tensor(2)]; tensor var_541_cast_fp16_0, tensor var_541_cast_fp16_1, tensor var_541_cast_fp16_2 = split(axis = var_541_axis_0, split_sizes = var_541_split_sizes_0, x = qkv_cast_fp16)[name = tensor("op_541_cast_fp16")]; tensor q_33_axes_0 = const()[name = tensor("q_33_axes_0"), val = tensor([2])]; tensor q_33_cast_fp16 = squeeze(axes = q_33_axes_0, x = var_541_cast_fp16_0)[name = tensor("q_33_cast_fp16")]; tensor k_33_axes_0 = const()[name = tensor("k_33_axes_0"), val = tensor([2])]; tensor k_33_cast_fp16 = squeeze(axes = k_33_axes_0, x = var_541_cast_fp16_1)[name = tensor("k_33_cast_fp16")]; tensor v_33_axes_0 = const()[name = tensor("v_33_axes_0"), val = tensor([2])]; tensor v_33_cast_fp16 = squeeze(axes = v_33_axes_0, x = var_541_cast_fp16_2)[name = tensor("v_33_cast_fp16")]; tensor v_perm_0 = const()[name = tensor("v_perm_0"), val = tensor([0, 2, -3, -1])]; tensor var_557_transpose_x_0 = const()[name = tensor("op_557_transpose_x_0"), val = tensor(false)]; tensor var_557_transpose_y_0 = const()[name = tensor("op_557_transpose_y_0"), val = tensor(false)]; tensor transpose_34_perm_0 = const()[name = tensor("transpose_34_perm_0"), val = tensor([0, 2, -3, -1])]; tensor transpose_35_perm_0 = const()[name = tensor("transpose_35_perm_0"), val = tensor([0, 2, -1, -3])]; tensor transpose_35 = transpose(perm = transpose_35_perm_0, x = k_33_cast_fp16)[name = tensor("transpose_40")]; tensor transpose_34 = transpose(perm = transpose_34_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_41")]; tensor var_557_cast_fp16 = matmul(transpose_x = var_557_transpose_x_0, transpose_y = var_557_transpose_y_0, x = transpose_34, y = transpose_35)[name = tensor("op_557_cast_fp16")]; tensor var_558_to_fp16 = const()[name = tensor("op_558_to_fp16"), val = tensor(0x1p-3)]; tensor attn_score_21_cast_fp16 = mul(x = var_557_cast_fp16, y = var_558_to_fp16)[name = tensor("attn_score_21_cast_fp16")]; tensor attn_score_cast_fp16 = select(a = var_18_to_fp16, b = attn_score_21_cast_fp16, cond = var_90_cast_fp16)[name = tensor("attn_score_cast_fp16")]; tensor input_129_cast_fp16 = add(x = input_9_cast_fp16_x_0, y = attn_score_cast_fp16)[name = tensor("input_129_cast_fp16")]; tensor attn_prob_21_cast_fp16 = softmax(axis = var_26, x = input_129_cast_fp16)[name = tensor("attn_prob_21_cast_fp16")]; tensor input_131_cast_fp16 = select(a = var_17_to_fp16, b = attn_prob_21_cast_fp16, cond = var_90_cast_fp16)[name = tensor("input_131_cast_fp16")]; tensor y_transpose_x_0 = const()[name = tensor("y_transpose_x_0"), val = tensor(false)]; tensor y_transpose_y_0 = const()[name = tensor("y_transpose_y_0"), val = tensor(false)]; tensor v_cast_fp16 = transpose(perm = v_perm_0, x = v_33_cast_fp16)[name = tensor("transpose_39")]; tensor y_cast_fp16 = matmul(transpose_x = y_transpose_x_0, transpose_y = y_transpose_y_0, x = input_131_cast_fp16, y = v_cast_fp16)[name = tensor("y_cast_fp16")]; tensor var_571_perm_0 = const()[name = tensor("op_571_perm_0"), val = tensor([0, 2, 1, 3])]; tensor var_573 = const()[name = tensor("op_573"), val = tensor([1, 256, -1])]; tensor var_571_cast_fp16 = transpose(perm = var_571_perm_0, x = y_cast_fp16)[name = tensor("transpose_38")]; tensor input_133_cast_fp16 = reshape(shape = var_573, x = var_571_cast_fp16)[name = tensor("input_133_cast_fp16")]; tensor encoder_layers_5_self_attention_o_net_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_5_self_attention_o_net_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86600640))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87190528))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4521920)))]; tensor linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_5_self_attention_o_net_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = tensor("linear_11_cast_fp16")]; tensor input_137_cast_fp16 = add(x = input_127_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_137_cast_fp16")]; tensor x_35_axes_0 = const()[name = tensor("x_35_axes_0"), val = tensor([-1])]; tensor encoder_layers_5_norm_pos_ff_weight_to_fp16 = const()[name = tensor("encoder_layers_5_norm_pos_ff_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87192128)))]; tensor x_35_cast_fp16 = layer_norm(axes = x_35_axes_0, epsilon = var_15_to_fp16, gamma = encoder_layers_5_norm_pos_ff_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("x_35_cast_fp16")]; tensor signal_21_perm_0 = const()[name = tensor("signal_21_perm_0"), val = tensor([0, 2, 1])]; tensor signal_21_cast_fp16 = transpose(perm = signal_21_perm_0, x = x_35_cast_fp16)[name = tensor("transpose_37")]; tensor input_139_cast_fp16 = mul(x = signal_21_cast_fp16, y = var_77_cast_fp16)[name = tensor("input_139_cast_fp16")]; tensor input_141_pad_0 = const()[name = tensor("input_141_pad_0"), val = tensor([0, 0, 0, 0, 2, 0])]; tensor input_141_mode_0 = const()[name = tensor("input_141_mode_0"), val = tensor("constant")]; tensor const_37_to_fp16 = const()[name = tensor("const_37_to_fp16"), val = tensor(0x0p+0)]; tensor input_141_cast_fp16 = pad(constant_val = const_37_to_fp16, mode = input_141_mode_0, pad = input_141_pad_0, x = input_139_cast_fp16)[name = tensor("input_141_cast_fp16")]; tensor conv_signal_21_pad_type_0 = const()[name = tensor("conv_signal_21_pad_type_0"), val = tensor("valid")]; tensor conv_signal_21_strides_0 = const()[name = tensor("conv_signal_21_strides_0"), val = tensor([1])]; tensor conv_signal_21_pad_0 = const()[name = tensor("conv_signal_21_pad_0"), val = tensor([0, 0])]; tensor conv_signal_21_dilations_0 = const()[name = tensor("conv_signal_21_dilations_0"), val = tensor([1])]; tensor conv_signal_21_groups_0 = const()[name = tensor("conv_signal_21_groups_0"), val = tensor(1)]; tensor encoder_layers_5_pos_ff_proj_conv_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_5_pos_ff_proj_conv_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87193728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94271680))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11605504)))]; tensor conv_signal_21_cast_fp16 = conv(dilations = conv_signal_21_dilations_0, groups = conv_signal_21_groups_0, pad = conv_signal_21_pad_0, pad_type = conv_signal_21_pad_type_0, strides = conv_signal_21_strides_0, weight = encoder_layers_5_pos_ff_proj_conv_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = tensor("conv_signal_21_cast_fp16")]; tensor input_143_cast_fp16 = mul(x = conv_signal_21_cast_fp16, y = var_77_cast_fp16)[name = tensor("input_143_cast_fp16")]; tensor signal_mode_0 = const()[name = tensor("signal_mode_0"), val = tensor("TANH_APPROXIMATION")]; tensor signal_cast_fp16 = gelu(mode = signal_mode_0, x = input_143_cast_fp16)[name = tensor("signal_cast_fp16")]; tensor input_145_cast_fp16 = mul(x = signal_cast_fp16, y = var_77_cast_fp16)[name = tensor("input_145_cast_fp16")]; tensor input_147_pad_0 = const()[name = tensor("input_147_pad_0"), val = tensor([0, 0, 0, 0, 2, 0])]; tensor input_147_mode_0 = const()[name = tensor("input_147_mode_0"), val = tensor("constant")]; tensor const_38_to_fp16 = const()[name = tensor("const_38_to_fp16"), val = tensor(0x0p+0)]; tensor input_147_cast_fp16 = pad(constant_val = const_38_to_fp16, mode = input_147_mode_0, pad = input_147_pad_0, x = input_145_cast_fp16)[name = tensor("input_147_cast_fp16")]; tensor conv_signal_pad_type_0 = const()[name = tensor("conv_signal_pad_type_0"), val = tensor("valid")]; tensor conv_signal_strides_0 = const()[name = tensor("conv_signal_strides_0"), val = tensor([1])]; tensor conv_signal_pad_0 = const()[name = tensor("conv_signal_pad_0"), val = tensor([0, 0])]; tensor conv_signal_dilations_0 = const()[name = tensor("conv_signal_dilations_0"), val = tensor([1])]; tensor conv_signal_groups_0 = const()[name = tensor("conv_signal_groups_0"), val = tensor(1)]; tensor encoder_layers_5_pos_ff_o_net_conv_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("encoder_layers_5_pos_ff_o_net_conv_weight_to_fp16_quantized"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94277888))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101355840))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4521920)))]; tensor conv_signal_cast_fp16 = conv(dilations = conv_signal_dilations_0, groups = conv_signal_groups_0, pad = conv_signal_pad_0, pad_type = conv_signal_pad_type_0, strides = conv_signal_strides_0, weight = encoder_layers_5_pos_ff_o_net_conv_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = tensor("conv_signal_cast_fp16")]; tensor var_611_cast_fp16 = mul(x = conv_signal_cast_fp16, y = var_77_cast_fp16)[name = tensor("op_611_cast_fp16")]; tensor input_149_perm_0 = const()[name = tensor("input_149_perm_0"), val = tensor([0, 2, 1])]; tensor input_149_cast_fp16 = transpose(perm = input_149_perm_0, x = var_611_cast_fp16)[name = tensor("transpose_36")]; tensor x_cast_fp16 = add(x = input_137_cast_fp16, y = input_149_cast_fp16)[name = tensor("x_cast_fp16")]; tensor input_cast_fp16 = mul(x = x_cast_fp16, y = var_57_cast_fp16)[name = tensor("input_cast_fp16")]; tensor var_619_axes_0 = const()[name = tensor("op_619_axes_0"), val = tensor([-1])]; tensor encoder_norm_out_weight_to_fp16 = const()[name = tensor("encoder_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101357440)))]; tensor var_619_cast_fp16 = layer_norm(axes = var_619_axes_0, epsilon = var_15_to_fp16, gamma = encoder_norm_out_weight_to_fp16, x = input_cast_fp16)[name = tensor("op_619_cast_fp16")]; tensor var_619_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_619_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; tensor encoder_output = cast(dtype = var_619_cast_fp16_to_fp32_dtype_0, x = var_619_cast_fp16)[name = tensor("cast_1")]; } -> (encoder_output); }