diff --git "a/streaming_encoder.mlmodelc/model.mil" "b/streaming_encoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/streaming_encoder.mlmodelc/model.mil" @@ -0,0 +1,2964 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.4.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor cache_last_channel, tensor cache_last_channel_len, tensor cache_last_time, tensor mel, tensor mel_length) { + tensor var_40 = const()[name = tensor("op_40"), val = tensor(-1)]; + tensor var_42 = const()[name = tensor("op_42"), val = tensor(1)]; + tensor x_1_perm_0 = const()[name = tensor("x_1_perm_0"), val = tensor([0, 2, 1])]; + tensor mel_to_fp16_dtype_0 = const()[name = tensor("mel_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cast_0_to_fp16_dtype_0 = const()[name = tensor("cast_0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor _inversed_86_y_0_to_fp16 = const()[name = tensor("_inversed_86_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor mel_length_to_fp16 = cast(dtype = cast_0_to_fp16_dtype_0, x = mel_length)[name = tensor("cast_190")]; + tensor _inversed_86_cast_fp16 = mul(x = mel_length_to_fp16, y = _inversed_86_y_0_to_fp16)[name = tensor("_inversed_86_cast_fp16")]; + tensor var_87_to_fp16 = const()[name = tensor("op_87_to_fp16"), val = tensor(0x1p+0)]; + tensor lengths_1_cast_fp16 = add(x = _inversed_86_cast_fp16, y = var_87_to_fp16)[name = tensor("lengths_1_cast_fp16")]; + tensor lengths_3_cast_fp16 = floor(x = lengths_1_cast_fp16)[name = tensor("lengths_3_cast_fp16")]; + tensor _inversed_94_y_0_to_fp16 = const()[name = tensor("_inversed_94_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_94_cast_fp16 = mul(x = lengths_3_cast_fp16, y = _inversed_94_y_0_to_fp16)[name = tensor("_inversed_94_cast_fp16")]; + tensor var_95_to_fp16 = const()[name = tensor("op_95_to_fp16"), val = tensor(0x1p+0)]; + tensor lengths_7_cast_fp16 = add(x = _inversed_94_cast_fp16, y = var_95_to_fp16)[name = tensor("lengths_7_cast_fp16")]; + tensor lengths_9_cast_fp16 = floor(x = lengths_7_cast_fp16)[name = tensor("lengths_9_cast_fp16")]; + tensor _inversed_102_y_0_to_fp16 = const()[name = tensor("_inversed_102_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_102_cast_fp16 = mul(x = lengths_9_cast_fp16, y = _inversed_102_y_0_to_fp16)[name = tensor("_inversed_102_cast_fp16")]; + tensor var_103_to_fp16 = const()[name = tensor("op_103_to_fp16"), val = tensor(0x1p+0)]; + tensor lengths_13_cast_fp16 = add(x = _inversed_102_cast_fp16, y = var_103_to_fp16)[name = tensor("lengths_13_cast_fp16")]; + tensor lengths_cast_fp16 = floor(x = lengths_13_cast_fp16)[name = tensor("lengths_cast_fp16")]; + tensor cast_9_dtype_0 = const()[name = tensor("cast_9_dtype_0"), val = tensor("int32")]; + tensor input_1_axes_0 = const()[name = tensor("input_1_axes_0"), val = tensor([1])]; + tensor mel_to_fp16 = cast(dtype = mel_to_fp16_dtype_0, x = mel)[name = tensor("cast_191")]; + tensor x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = mel_to_fp16)[name = tensor("transpose_241")]; + tensor input_1_cast_fp16 = expand_dims(axes = input_1_axes_0, x = x_1_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_3_pad_0 = const()[name = tensor("input_3_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + tensor input_3_mode_0 = const()[name = tensor("input_3_mode_0"), val = tensor("constant")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(0x0p+0)]; + tensor input_3_cast_fp16 = pad(constant_val = const_0_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor input_5_pad_type_0 = const()[name = tensor("input_5_pad_type_0"), val = tensor("valid")]; + tensor input_5_strides_0 = const()[name = tensor("input_5_strides_0"), val = tensor([2, 2])]; + tensor input_5_pad_0 = const()[name = tensor("input_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_5_dilations_0 = const()[name = tensor("input_5_dilations_0"), val = tensor([1, 1])]; + tensor input_5_groups_0 = const()[name = tensor("input_5_groups_0"), val = tensor(1)]; + tensor encoder_pre_encode_conv_0_weight_to_fp16 = const()[name = tensor("encoder_pre_encode_conv_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = tensor("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4736)))]; + tensor input_5_cast_fp16 = conv(bias = encoder_pre_encode_conv_0_bias_to_fp16, dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = encoder_pre_encode_conv_0_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_pad_0 = const()[name = tensor("input_9_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + tensor input_9_mode_0 = const()[name = tensor("input_9_mode_0"), val = tensor("constant")]; + tensor const_1_to_fp16 = const()[name = tensor("const_1_to_fp16"), val = tensor(0x0p+0)]; + tensor input_9_cast_fp16 = pad(constant_val = const_1_to_fp16, mode = input_9_mode_0, pad = input_9_pad_0, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_pad_type_0 = const()[name = tensor("input_11_pad_type_0"), val = tensor("valid")]; + tensor input_11_strides_0 = const()[name = tensor("input_11_strides_0"), val = tensor([2, 2])]; + tensor input_11_groups_0 = const()[name = tensor("input_11_groups_0"), val = tensor(256)]; + tensor input_11_pad_0 = const()[name = tensor("input_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_11_dilations_0 = const()[name = tensor("input_11_dilations_0"), val = tensor([1, 1])]; + tensor encoder_pre_encode_conv_2_weight_to_fp16 = const()[name = tensor("encoder_pre_encode_conv_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5312)))]; + tensor encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = tensor("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9984)))]; + tensor input_11_cast_fp16 = conv(bias = encoder_pre_encode_conv_2_bias_to_fp16, dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = encoder_pre_encode_conv_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor input_13_pad_type_0 = const()[name = tensor("input_13_pad_type_0"), val = tensor("valid")]; + tensor input_13_strides_0 = const()[name = tensor("input_13_strides_0"), val = tensor([1, 1])]; + tensor input_13_pad_0 = const()[name = tensor("input_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_13_dilations_0 = const()[name = tensor("input_13_dilations_0"), val = tensor([1, 1])]; + tensor input_13_groups_0 = const()[name = tensor("input_13_groups_0"), val = tensor(1)]; + tensor encoder_pre_encode_conv_3_weight_to_fp16 = const()[name = tensor("encoder_pre_encode_conv_3_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10560)))]; + tensor encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = tensor("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141696)))]; + tensor input_13_cast_fp16 = conv(bias = encoder_pre_encode_conv_3_bias_to_fp16, dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = encoder_pre_encode_conv_3_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_cast_fp16 = relu(x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_pad_0 = const()[name = tensor("input_17_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + tensor input_17_mode_0 = const()[name = tensor("input_17_mode_0"), val = tensor("constant")]; + tensor const_2_to_fp16 = const()[name = tensor("const_2_to_fp16"), val = tensor(0x0p+0)]; + tensor input_17_cast_fp16 = pad(constant_val = const_2_to_fp16, mode = input_17_mode_0, pad = input_17_pad_0, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_pad_type_0 = const()[name = tensor("input_19_pad_type_0"), val = tensor("valid")]; + tensor input_19_strides_0 = const()[name = tensor("input_19_strides_0"), val = tensor([2, 2])]; + tensor input_19_groups_0 = const()[name = tensor("input_19_groups_0"), val = tensor(256)]; + tensor input_19_pad_0 = const()[name = tensor("input_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_19_dilations_0 = const()[name = tensor("input_19_dilations_0"), val = tensor([1, 1])]; + tensor encoder_pre_encode_conv_5_weight_to_fp16 = const()[name = tensor("encoder_pre_encode_conv_5_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142272)))]; + tensor encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = tensor("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146944)))]; + tensor input_19_cast_fp16 = conv(bias = encoder_pre_encode_conv_5_bias_to_fp16, dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = encoder_pre_encode_conv_5_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_pad_type_0 = const()[name = tensor("input_21_pad_type_0"), val = tensor("valid")]; + tensor input_21_strides_0 = const()[name = tensor("input_21_strides_0"), val = tensor([1, 1])]; + tensor input_21_pad_0 = const()[name = tensor("input_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_21_dilations_0 = const()[name = tensor("input_21_dilations_0"), val = tensor([1, 1])]; + tensor input_21_groups_0 = const()[name = tensor("input_21_groups_0"), val = tensor(1)]; + tensor encoder_pre_encode_conv_6_weight_to_fp16 = const()[name = tensor("encoder_pre_encode_conv_6_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147520)))]; + tensor encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = tensor("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278656)))]; + tensor input_21_cast_fp16 = conv(bias = encoder_pre_encode_conv_6_bias_to_fp16, dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = encoder_pre_encode_conv_6_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor x_3_cast_fp16 = relu(x = input_21_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor var_159_perm_0 = const()[name = tensor("op_159_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_160 = const()[name = tensor("op_160"), val = tensor([1, 6, -1])]; + tensor var_159_cast_fp16 = transpose(perm = var_159_perm_0, x = x_3_cast_fp16)[name = tensor("transpose_240")]; + tensor input_23_cast_fp16 = reshape(shape = var_160, x = var_159_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor encoder_pre_encode_out_weight_to_fp16 = const()[name = tensor("encoder_pre_encode_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279232)))]; + tensor encoder_pre_encode_out_bias_to_fp16 = const()[name = tensor("encoder_pre_encode_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4735744)))]; + tensor linear_0_cast_fp16 = linear(bias = encoder_pre_encode_out_bias_to_fp16, weight = encoder_pre_encode_out_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_170_begin_0 = const()[name = tensor("op_170_begin_0"), val = tensor([0, 2, 0])]; + tensor var_170_end_0 = const()[name = tensor("op_170_end_0"), val = tensor([1, 6, 512])]; + tensor var_170_end_mask_0 = const()[name = tensor("op_170_end_mask_0"), val = tensor([true, true, true])]; + tensor var_170_cast_fp16 = slice_by_index(begin = var_170_begin_0, end = var_170_end_0, end_mask = var_170_end_mask_0, x = linear_0_cast_fp16)[name = tensor("op_170_cast_fp16")]; + tensor var_172 = const()[name = tensor("op_172"), val = tensor(2)]; + tensor lengths_cast_fp16_to_int32 = cast(dtype = cast_9_dtype_0, x = lengths_cast_fp16)[name = tensor("cast_189")]; + tensor var_173 = sub(x = lengths_cast_fp16_to_int32, y = var_172)[name = tensor("op_173")]; + tensor var_173_promoted_to_fp16_dtype_0 = const()[name = tensor("op_173_promoted_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_38_promoted_to_fp16 = const()[name = tensor("op_38_promoted_to_fp16"), val = tensor(0x0p+0)]; + tensor const_5_to_fp16 = const()[name = tensor("const_5_to_fp16"), val = tensor(inf)]; + tensor var_173_to_fp16 = cast(dtype = var_173_promoted_to_fp16_dtype_0, x = var_173)[name = tensor("cast_188")]; + tensor clip_0_cast_fp16 = clip(alpha = var_38_promoted_to_fp16, beta = const_5_to_fp16, x = var_173_to_fp16)[name = tensor("clip_0_cast_fp16")]; + tensor max_audio_length_1 = const()[name = tensor("max_audio_length_1"), val = tensor([4])]; + tensor var_189_promoted_to_fp16 = const()[name = tensor("op_189_promoted_to_fp16"), val = tensor(0x1.18p+6)]; + tensor padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_189_promoted_to_fp16)[name = tensor("padding_length_cast_fp16")]; + tensor const_7 = const()[name = tensor("const_7"), val = tensor(-1)]; + tensor var_191 = mul(x = cache_last_channel_len, y = const_7)[name = tensor("op_191")]; + tensor var_192 = const()[name = tensor("op_192"), val = tensor(70)]; + tensor offset = add(x = var_191, y = var_192)[name = tensor("offset")]; + tensor var_232_axes_0 = const()[name = tensor("op_232_axes_0"), val = tensor([-1])]; + tensor var_232_cast_fp16 = expand_dims(axes = var_232_axes_0, x = padding_length_cast_fp16)[name = tensor("op_232_cast_fp16")]; + tensor var_231_promoted_to_fp16 = const()[name = tensor("op_231_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4736832)))]; + tensor pad_mask_1_cast_fp16 = less(x = var_231_promoted_to_fp16, y = var_232_cast_fp16)[name = tensor("pad_mask_1_cast_fp16")]; + tensor expand_dims_1 = const()[name = tensor("expand_dims_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73]])]; + tensor var_238_axes_0 = const()[name = tensor("op_238_axes_0"), val = tensor([-1])]; + tensor var_238 = expand_dims(axes = var_238_axes_0, x = offset)[name = tensor("op_238")]; + tensor pad_mask_off = greater_equal(x = expand_dims_1, y = var_238)[name = tensor("pad_mask_off")]; + tensor pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1_cast_fp16)[name = tensor("pad_mask_3")]; + tensor var_241_axes_0 = const()[name = tensor("op_241_axes_0"), val = tensor([1])]; + tensor var_241 = expand_dims(axes = var_241_axes_0, x = pad_mask_3)[name = tensor("op_241")]; + tensor var_242 = const()[name = tensor("op_242"), val = tensor([1, 74, 1])]; + tensor pad_mask_for_att_mask_1 = tile(reps = var_242, x = var_241)[name = tensor("pad_mask_for_att_mask_1")]; + tensor var_244_perm_0 = const()[name = tensor("op_244_perm_0"), val = tensor([0, 2, 1])]; + tensor var_244 = transpose(perm = var_244_perm_0, x = pad_mask_for_att_mask_1)[name = tensor("transpose_239")]; + tensor pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_244)[name = tensor("pad_mask_for_att_mask")]; + tensor const_15 = const()[name = tensor("const_15"), val = tensor([[[true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false], [false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true]]])]; + tensor att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_15)[name = tensor("att_mask_9")]; + tensor att_mask = logical_not(x = att_mask_9)[name = tensor("att_mask")]; + tensor pad_mask_5 = logical_not(x = pad_mask_3)[name = tensor("pad_mask_5")]; + tensor pad_mask_begin_0 = const()[name = tensor("pad_mask_begin_0"), val = tensor([0, 70])]; + tensor pad_mask_end_0 = const()[name = tensor("pad_mask_end_0"), val = tensor([1, 74])]; + tensor pad_mask_end_mask_0 = const()[name = tensor("pad_mask_end_mask_0"), val = tensor([true, true])]; + tensor pad_mask = slice_by_index(begin = pad_mask_begin_0, end = pad_mask_end_0, end_mask = pad_mask_end_mask_0, x = pad_mask_5)[name = tensor("pad_mask")]; + tensor mask_1_begin_0 = const()[name = tensor("mask_1_begin_0"), val = tensor([0, 70, 0])]; + tensor mask_1_end_0 = const()[name = tensor("mask_1_end_0"), val = tensor([1, 74, 74])]; + tensor mask_1_end_mask_0 = const()[name = tensor("mask_1_end_mask_0"), val = tensor([true, true, true])]; + tensor mask_1 = slice_by_index(begin = mask_1_begin_0, end = mask_1_end_0, end_mask = mask_1_end_mask_0, x = att_mask)[name = tensor("mask_1")]; + tensor cache_1_begin_0 = const()[name = tensor("cache_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor cache_1_end_0 = const()[name = tensor("cache_1_end_0"), val = tensor([1, 1, 70, 512])]; + tensor cache_1_end_mask_0 = const()[name = tensor("cache_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_1_squeeze_mask_0 = const()[name = tensor("cache_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_last_channel_to_fp16_dtype_0 = const()[name = tensor("cache_last_channel_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache_last_channel_to_fp16 = cast(dtype = cache_last_channel_to_fp16_dtype_0, x = cache_last_channel)[name = tensor("cast_187")]; + tensor cache_1_cast_fp16 = slice_by_index(begin = cache_1_begin_0, end = cache_1_end_0, end_mask = cache_1_end_mask_0, squeeze_mask = cache_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor("cache_1_cast_fp16")]; + tensor cache_3_begin_0 = const()[name = tensor("cache_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor cache_3_end_0 = const()[name = tensor("cache_3_end_0"), val = tensor([1, 1, 512, 8])]; + tensor cache_3_end_mask_0 = const()[name = tensor("cache_3_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_3_squeeze_mask_0 = const()[name = tensor("cache_3_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_last_time_to_fp16_dtype_0 = const()[name = tensor("cache_last_time_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache_last_time_to_fp16 = cast(dtype = cache_last_time_to_fp16_dtype_0, x = cache_last_time)[name = tensor("cast_186")]; + tensor cache_3_cast_fp16 = slice_by_index(begin = cache_3_begin_0, end = cache_3_end_0, end_mask = cache_3_end_mask_0, squeeze_mask = cache_3_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor("cache_3_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = tensor("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4737088)))]; + tensor encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = tensor("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4738176)))]; + tensor var_16_to_fp16 = const()[name = tensor("op_16_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_170_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor encoder_layers_0_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_0_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4739264)))]; + 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(6836480)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_0_feed_forward1_linear1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor input_31_cast_fp16 = silu(x = linear_1_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor encoder_layers_0_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_0_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6840640)))]; + tensor linear_2_bias_0_to_fp16 = const()[name = tensor("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8937856)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_feed_forward1_linear2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_281_to_fp16 = const()[name = tensor("op_281_to_fp16"), val = tensor(0x1p-1)]; + tensor var_282_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_281_to_fp16)[name = tensor("op_282_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = var_170_cast_fp16, y = var_282_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor key_1_axes_0 = const()[name = tensor("key_1_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = tensor("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8938944)))]; + tensor encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = tensor("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8940032)))]; + tensor key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("key_1_cast_fp16")]; + tensor input_39_interleave_0 = const()[name = tensor("input_39_interleave_0"), val = tensor(false)]; + tensor input_39_cast_fp16 = concat(axis = var_42, interleave = input_39_interleave_0, values = (cache_1_cast_fp16, key_1_cast_fp16))[name = tensor("input_39_cast_fp16")]; + tensor var_304_begin_0 = const()[name = tensor("op_304_begin_0"), val = tensor([0, 4, 0])]; + tensor var_304_end_0 = const()[name = tensor("op_304_end_0"), val = tensor([1, 70, 512])]; + tensor var_304_end_mask_0 = const()[name = tensor("op_304_end_mask_0"), val = tensor([true, true, true])]; + tensor var_304_cast_fp16 = slice_by_index(begin = var_304_begin_0, end = var_304_end_0, end_mask = var_304_end_mask_0, x = cache_1_cast_fp16)[name = tensor("op_304_cast_fp16")]; + tensor var_310_interleave_0 = const()[name = tensor("op_310_interleave_0"), val = tensor(false)]; + tensor var_310_cast_fp16 = concat(axis = var_42, interleave = var_310_interleave_0, values = (var_304_cast_fp16, key_1_cast_fp16))[name = tensor("op_310_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_q_weight_to_fp16 = const()[name = tensor("encoder_layers_0_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8941120)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_q_weight_to_fp16, x = key_1_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor var_314 = const()[name = tensor("op_314"), val = tensor([1, -1, 8, 64])]; + tensor q_1_cast_fp16 = reshape(shape = var_314, x = linear_3_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_k_weight_to_fp16 = const()[name = tensor("encoder_layers_0_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9465472)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_k_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_318 = const()[name = tensor("op_318"), val = tensor([1, -1, 8, 64])]; + tensor k_1_cast_fp16 = reshape(shape = var_318, x = linear_4_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_v_weight_to_fp16 = const()[name = tensor("encoder_layers_0_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9989824)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_v_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_322 = const()[name = tensor("op_322"), val = tensor([1, -1, 8, 64])]; + tensor v_1_cast_fp16 = reshape(shape = var_322, x = linear_5_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor value_1_perm_0 = const()[name = tensor("value_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor encoder_layers_0_self_attn_pos_bias_u_to_fp16 = const()[name = tensor("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10514176)))]; + tensor var_334_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = tensor("op_334_cast_fp16")]; + tensor encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = tensor("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10515264)))]; + tensor var_336_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = tensor("op_336_cast_fp16")]; + tensor q_with_bias_v_1_perm_0 = const()[name = tensor("q_with_bias_v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor x_7_transpose_x_0 = const()[name = tensor("x_7_transpose_x_0"), val = tensor(false)]; + tensor x_7_transpose_y_0 = const()[name = tensor("x_7_transpose_y_0"), val = tensor(false)]; + tensor var_338_to_fp16 = const()[name = tensor("op_338_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10516352)))]; + tensor q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_336_cast_fp16)[name = tensor("transpose_237")]; + tensor x_7_cast_fp16 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = var_338_to_fp16)[name = tensor("x_7_cast_fp16")]; + tensor x_9_pad_0 = const()[name = tensor("x_9_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor x_9_mode_0 = const()[name = tensor("x_9_mode_0"), val = tensor("constant")]; + tensor const_23_to_fp16 = const()[name = tensor("const_23_to_fp16"), val = tensor(0x0p+0)]; + tensor x_9_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = x_9_mode_0, pad = x_9_pad_0, x = x_7_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor var_346 = const()[name = tensor("op_346"), val = tensor([1, 8, -1, 4])]; + tensor x_11_cast_fp16 = reshape(shape = var_346, x = x_9_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor var_350_begin_0 = const()[name = tensor("op_350_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_350_end_0 = const()[name = tensor("op_350_end_0"), val = tensor([1, 8, 148, 4])]; + tensor var_350_end_mask_0 = const()[name = tensor("op_350_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_350_cast_fp16 = slice_by_index(begin = var_350_begin_0, end = var_350_end_0, end_mask = var_350_end_mask_0, x = x_11_cast_fp16)[name = tensor("op_350_cast_fp16")]; + tensor var_351 = const()[name = tensor("op_351"), val = tensor([1, 8, 4, 147])]; + tensor matrix_bd_1_cast_fp16 = reshape(shape = var_351, x = var_350_cast_fp16)[name = tensor("matrix_bd_1_cast_fp16")]; + tensor matrix_ac_1_transpose_x_0 = const()[name = tensor("matrix_ac_1_transpose_x_0"), val = tensor(false)]; + tensor matrix_ac_1_transpose_y_0 = const()[name = tensor("matrix_ac_1_transpose_y_0"), val = tensor(false)]; + tensor transpose_51_perm_0 = const()[name = tensor("transpose_51_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_52_perm_0 = const()[name = tensor("transpose_52_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_52 = transpose(perm = transpose_52_perm_0, x = k_1_cast_fp16)[name = tensor("transpose_235")]; + tensor transpose_51 = transpose(perm = transpose_51_perm_0, x = var_334_cast_fp16)[name = tensor("transpose_236")]; + tensor matrix_ac_1_cast_fp16 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_51, y = transpose_52)[name = tensor("matrix_ac_1_cast_fp16")]; + tensor matrix_bd_3_begin_0 = const()[name = tensor("matrix_bd_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_3_end_0 = const()[name = tensor("matrix_bd_3_end_0"), val = tensor([1, 8, 4, 74])]; + tensor matrix_bd_3_end_mask_0 = const()[name = tensor("matrix_bd_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_3_cast_fp16 = slice_by_index(begin = matrix_bd_3_begin_0, end = matrix_bd_3_end_0, end_mask = matrix_bd_3_end_mask_0, x = matrix_bd_1_cast_fp16)[name = tensor("matrix_bd_3_cast_fp16")]; + tensor var_360_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = tensor("op_360_cast_fp16")]; + tensor _inversed_scores_1_y_0_to_fp16 = const()[name = tensor("_inversed_scores_1_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor _inversed_scores_1_cast_fp16 = mul(x = var_360_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = tensor("_inversed_scores_1_cast_fp16")]; + tensor mask_3_axes_0 = const()[name = tensor("mask_3_axes_0"), val = tensor([1])]; + tensor mask_3 = expand_dims(axes = mask_3_axes_0, x = mask_1)[name = tensor("mask_3")]; + tensor var_19_to_fp16 = const()[name = tensor("op_19_to_fp16"), val = tensor(-0x1.388p+13)]; + tensor scores_3_cast_fp16 = select(a = var_19_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_3)[name = tensor("scores_3_cast_fp16")]; + tensor var_366_cast_fp16 = softmax(axis = var_40, x = scores_3_cast_fp16)[name = tensor("op_366_cast_fp16")]; + tensor var_18_to_fp16 = const()[name = tensor("op_18_to_fp16"), val = tensor(0x0p+0)]; + tensor input_41_cast_fp16 = select(a = var_18_to_fp16, b = var_366_cast_fp16, cond = mask_3)[name = tensor("input_41_cast_fp16")]; + tensor x_13_transpose_x_0 = const()[name = tensor("x_13_transpose_x_0"), val = tensor(false)]; + tensor x_13_transpose_y_0 = const()[name = tensor("x_13_transpose_y_0"), val = tensor(false)]; + tensor value_1_cast_fp16 = transpose(perm = value_1_perm_0, x = v_1_cast_fp16)[name = tensor("transpose_238")]; + tensor x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = input_41_cast_fp16, y = value_1_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor var_370_perm_0 = const()[name = tensor("op_370_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_371 = const()[name = tensor("op_371"), val = tensor([1, -1, 512])]; + tensor var_370_cast_fp16 = transpose(perm = var_370_perm_0, x = x_13_cast_fp16)[name = tensor("transpose_234")]; + tensor input_43_cast_fp16 = reshape(shape = var_371, x = var_370_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_out_weight_to_fp16 = const()[name = tensor("encoder_layers_0_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10666944)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_out_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_47_cast_fp16 = add(x = input_37_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11191296)))]; + tensor encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = tensor("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11192384)))]; + tensor x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor input_49_perm_0 = const()[name = tensor("input_49_perm_0"), val = tensor([0, 2, 1])]; + tensor input_51_pad_type_0 = const()[name = tensor("input_51_pad_type_0"), val = tensor("valid")]; + tensor input_51_strides_0 = const()[name = tensor("input_51_strides_0"), val = tensor([1])]; + tensor input_51_pad_0 = const()[name = tensor("input_51_pad_0"), val = tensor([0, 0])]; + tensor input_51_dilations_0 = const()[name = tensor("input_51_dilations_0"), val = tensor([1])]; + tensor input_51_groups_0 = const()[name = tensor("input_51_groups_0"), val = tensor(1)]; + tensor encoder_layers_0_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11193472)))]; + tensor input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = tensor("transpose_233")]; + tensor input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = encoder_layers_0_conv_pointwise_conv1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor x_19_split_num_splits_0 = const()[name = tensor("x_19_split_num_splits_0"), val = tensor(2)]; + tensor x_19_split_axis_0 = const()[name = tensor("x_19_split_axis_0"), val = tensor(1)]; + tensor x_19_split_cast_fp16_0, tensor x_19_split_cast_fp16_1 = split(axis = x_19_split_axis_0, num_splits = x_19_split_num_splits_0, x = input_51_cast_fp16)[name = tensor("x_19_split_cast_fp16")]; + tensor x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = tensor("x_19_split_1_sigmoid_cast_fp16")]; + tensor x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor var_396_axes_0 = const()[name = tensor("op_396_axes_0"), val = tensor([1])]; + tensor var_396 = expand_dims(axes = var_396_axes_0, x = pad_mask)[name = tensor("op_396")]; + tensor input_53_cast_fp16 = select(a = var_18_to_fp16, b = x_19_cast_fp16, cond = var_396)[name = tensor("input_53_cast_fp16")]; + tensor new_x_3_interleave_0 = const()[name = tensor("new_x_3_interleave_0"), val = tensor(false)]; + tensor new_x_3_cast_fp16 = concat(axis = var_40, interleave = new_x_3_interleave_0, values = (cache_3_cast_fp16, input_53_cast_fp16))[name = tensor("new_x_3_cast_fp16")]; + tensor var_409_begin_0 = const()[name = tensor("op_409_begin_0"), val = tensor([0, 0, 4])]; + tensor var_409_end_0 = const()[name = tensor("op_409_end_0"), val = tensor([1, 512, 12])]; + tensor var_409_end_mask_0 = const()[name = tensor("op_409_end_mask_0"), val = tensor([true, true, true])]; + tensor var_409_cast_fp16 = slice_by_index(begin = var_409_begin_0, end = var_409_end_0, end_mask = var_409_end_mask_0, x = new_x_3_cast_fp16)[name = tensor("op_409_cast_fp16")]; + tensor x_21_pad_type_0 = const()[name = tensor("x_21_pad_type_0"), val = tensor("valid")]; + tensor x_21_groups_0 = const()[name = tensor("x_21_groups_0"), val = tensor(512)]; + tensor x_21_strides_0 = const()[name = tensor("x_21_strides_0"), val = tensor([1])]; + tensor x_21_pad_0 = const()[name = tensor("x_21_pad_0"), val = tensor([0, 0])]; + tensor x_21_dilations_0 = const()[name = tensor("x_21_dilations_0"), val = tensor([1])]; + tensor encoder_layers_0_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_0_conv_depthwise_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12242112)))]; + tensor x_21_cast_fp16 = conv(dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = encoder_layers_0_conv_depthwise_conv_weight_to_fp16, x = new_x_3_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor input_55_perm_0 = const()[name = tensor("input_55_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = tensor("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12251392)))]; + tensor encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = tensor("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12252480)))]; + tensor input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = tensor("transpose_232")]; + tensor x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor input_57_perm_0 = const()[name = tensor("input_57_perm_0"), val = tensor([0, 2, 1])]; + tensor input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = tensor("transpose_231")]; + tensor input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor x_25_pad_type_0 = const()[name = tensor("x_25_pad_type_0"), val = tensor("valid")]; + tensor x_25_strides_0 = const()[name = tensor("x_25_strides_0"), val = tensor([1])]; + tensor x_25_pad_0 = const()[name = tensor("x_25_pad_0"), val = tensor([0, 0])]; + tensor x_25_dilations_0 = const()[name = tensor("x_25_dilations_0"), val = tensor([1])]; + tensor x_25_groups_0 = const()[name = tensor("x_25_groups_0"), val = tensor(1)]; + tensor encoder_layers_0_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12253568)))]; + tensor x_25_cast_fp16 = conv(dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = encoder_layers_0_conv_pointwise_conv2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor input_61_perm_0 = const()[name = tensor("input_61_perm_0"), val = tensor([0, 2, 1])]; + tensor input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = tensor("transpose_230")]; + tensor input_63_cast_fp16 = add(x = input_47_cast_fp16, y = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = tensor("input_65_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = tensor("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12777920)))]; + tensor encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = tensor("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12779008)))]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor encoder_layers_0_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_0_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12780096)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_0_feed_forward2_linear1_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_69_cast_fp16 = silu(x = linear_8_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor encoder_layers_0_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_0_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14877312)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_feed_forward2_linear2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_450_to_fp16 = const()[name = tensor("op_450_to_fp16"), val = tensor(0x1p-1)]; + tensor var_451_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_450_to_fp16)[name = tensor("op_451_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_451_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_77_axes_0 = const()[name = tensor("input_77_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_out_weight_to_fp16 = const()[name = tensor("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16974528)))]; + tensor encoder_layers_0_norm_out_bias_to_fp16 = const()[name = tensor("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16975616)))]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor cache_5_begin_0 = const()[name = tensor("cache_5_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor cache_5_end_0 = const()[name = tensor("cache_5_end_0"), val = tensor([2, 1, 70, 512])]; + tensor cache_5_end_mask_0 = const()[name = tensor("cache_5_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_5_squeeze_mask_0 = const()[name = tensor("cache_5_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_5_cast_fp16 = slice_by_index(begin = cache_5_begin_0, end = cache_5_end_0, end_mask = cache_5_end_mask_0, squeeze_mask = cache_5_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor("cache_5_cast_fp16")]; + tensor cache_7_begin_0 = const()[name = tensor("cache_7_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor cache_7_end_0 = const()[name = tensor("cache_7_end_0"), val = tensor([2, 1, 512, 8])]; + tensor cache_7_end_mask_0 = const()[name = tensor("cache_7_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_7_squeeze_mask_0 = const()[name = tensor("cache_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_7_cast_fp16 = slice_by_index(begin = cache_7_begin_0, end = cache_7_end_0, end_mask = cache_7_end_mask_0, squeeze_mask = cache_7_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor("cache_7_cast_fp16")]; + tensor input_79_axes_0 = const()[name = tensor("input_79_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = tensor("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16976704)))]; + tensor encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = tensor("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16977792)))]; + tensor input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor encoder_layers_1_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_1_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16978880)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_1_feed_forward1_linear1_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_83_cast_fp16 = silu(x = linear_10_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor encoder_layers_1_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_1_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19076096)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_feed_forward1_linear2_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor var_485_to_fp16 = const()[name = tensor("op_485_to_fp16"), val = tensor(0x1p-1)]; + tensor var_486_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_485_to_fp16)[name = tensor("op_486_cast_fp16")]; + tensor input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_486_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor key_3_axes_0 = const()[name = tensor("key_3_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = tensor("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21173312)))]; + tensor encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = tensor("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21174400)))]; + tensor key_3_cast_fp16 = layer_norm(axes = key_3_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("key_3_cast_fp16")]; + tensor input_91_interleave_0 = const()[name = tensor("input_91_interleave_0"), val = tensor(false)]; + tensor input_91_cast_fp16 = concat(axis = var_42, interleave = input_91_interleave_0, values = (cache_5_cast_fp16, key_3_cast_fp16))[name = tensor("input_91_cast_fp16")]; + tensor var_508_begin_0 = const()[name = tensor("op_508_begin_0"), val = tensor([0, 4, 0])]; + tensor var_508_end_0 = const()[name = tensor("op_508_end_0"), val = tensor([1, 70, 512])]; + tensor var_508_end_mask_0 = const()[name = tensor("op_508_end_mask_0"), val = tensor([true, true, true])]; + tensor var_508_cast_fp16 = slice_by_index(begin = var_508_begin_0, end = var_508_end_0, end_mask = var_508_end_mask_0, x = cache_5_cast_fp16)[name = tensor("op_508_cast_fp16")]; + tensor var_514_interleave_0 = const()[name = tensor("op_514_interleave_0"), val = tensor(false)]; + tensor var_514_cast_fp16 = concat(axis = var_42, interleave = var_514_interleave_0, values = (var_508_cast_fp16, key_3_cast_fp16))[name = tensor("op_514_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_q_weight_to_fp16 = const()[name = tensor("encoder_layers_1_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21175488)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_q_weight_to_fp16, x = key_3_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_518 = const()[name = tensor("op_518"), val = tensor([1, -1, 8, 64])]; + tensor q_7_cast_fp16 = reshape(shape = var_518, x = linear_12_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_k_weight_to_fp16 = const()[name = tensor("encoder_layers_1_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21699840)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_k_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_522 = const()[name = tensor("op_522"), val = tensor([1, -1, 8, 64])]; + tensor k_5_cast_fp16 = reshape(shape = var_522, x = linear_13_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_v_weight_to_fp16 = const()[name = tensor("encoder_layers_1_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22224192)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_v_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_526 = const()[name = tensor("op_526"), val = tensor([1, -1, 8, 64])]; + tensor v_3_cast_fp16 = reshape(shape = var_526, x = linear_14_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor value_3_perm_0 = const()[name = tensor("value_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor encoder_layers_1_self_attn_pos_bias_u_to_fp16 = const()[name = tensor("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22748544)))]; + tensor var_538_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = tensor("op_538_cast_fp16")]; + tensor encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = tensor("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22749632)))]; + tensor var_540_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = tensor("op_540_cast_fp16")]; + tensor q_with_bias_v_3_perm_0 = const()[name = tensor("q_with_bias_v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor x_33_transpose_x_0 = const()[name = tensor("x_33_transpose_x_0"), val = tensor(false)]; + tensor x_33_transpose_y_0 = const()[name = tensor("x_33_transpose_y_0"), val = tensor(false)]; + tensor var_542_to_fp16 = const()[name = tensor("op_542_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22750720)))]; + tensor q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_540_cast_fp16)[name = tensor("transpose_228")]; + tensor x_33_cast_fp16 = matmul(transpose_x = x_33_transpose_x_0, transpose_y = x_33_transpose_y_0, x = q_with_bias_v_3_cast_fp16, y = var_542_to_fp16)[name = tensor("x_33_cast_fp16")]; + tensor x_35_pad_0 = const()[name = tensor("x_35_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor x_35_mode_0 = const()[name = tensor("x_35_mode_0"), val = tensor("constant")]; + tensor const_36_to_fp16 = const()[name = tensor("const_36_to_fp16"), val = tensor(0x0p+0)]; + tensor x_35_cast_fp16 = pad(constant_val = const_36_to_fp16, mode = x_35_mode_0, pad = x_35_pad_0, x = x_33_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor var_550 = const()[name = tensor("op_550"), val = tensor([1, 8, -1, 4])]; + tensor x_37_cast_fp16 = reshape(shape = var_550, x = x_35_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor var_554_begin_0 = const()[name = tensor("op_554_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_554_end_0 = const()[name = tensor("op_554_end_0"), val = tensor([1, 8, 148, 4])]; + tensor var_554_end_mask_0 = const()[name = tensor("op_554_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_554_cast_fp16 = slice_by_index(begin = var_554_begin_0, end = var_554_end_0, end_mask = var_554_end_mask_0, x = x_37_cast_fp16)[name = tensor("op_554_cast_fp16")]; + tensor var_555 = const()[name = tensor("op_555"), val = tensor([1, 8, 4, 147])]; + tensor matrix_bd_5_cast_fp16 = reshape(shape = var_555, x = var_554_cast_fp16)[name = tensor("matrix_bd_5_cast_fp16")]; + tensor matrix_ac_3_transpose_x_0 = const()[name = tensor("matrix_ac_3_transpose_x_0"), val = tensor(false)]; + tensor matrix_ac_3_transpose_y_0 = const()[name = tensor("matrix_ac_3_transpose_y_0"), val = tensor(false)]; + tensor transpose_53_perm_0 = const()[name = tensor("transpose_53_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_54_perm_0 = const()[name = tensor("transpose_54_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_54 = transpose(perm = transpose_54_perm_0, x = k_5_cast_fp16)[name = tensor("transpose_226")]; + tensor transpose_53 = transpose(perm = transpose_53_perm_0, x = var_538_cast_fp16)[name = tensor("transpose_227")]; + tensor matrix_ac_3_cast_fp16 = matmul(transpose_x = matrix_ac_3_transpose_x_0, transpose_y = matrix_ac_3_transpose_y_0, x = transpose_53, y = transpose_54)[name = tensor("matrix_ac_3_cast_fp16")]; + tensor matrix_bd_7_begin_0 = const()[name = tensor("matrix_bd_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_7_end_0 = const()[name = tensor("matrix_bd_7_end_0"), val = tensor([1, 8, 4, 74])]; + tensor matrix_bd_7_end_mask_0 = const()[name = tensor("matrix_bd_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_7_cast_fp16 = slice_by_index(begin = matrix_bd_7_begin_0, end = matrix_bd_7_end_0, end_mask = matrix_bd_7_end_mask_0, x = matrix_bd_5_cast_fp16)[name = tensor("matrix_bd_7_cast_fp16")]; + tensor var_564_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = tensor("op_564_cast_fp16")]; + tensor _inversed_scores_5_y_0_to_fp16 = const()[name = tensor("_inversed_scores_5_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor _inversed_scores_5_cast_fp16 = mul(x = var_564_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = tensor("_inversed_scores_5_cast_fp16")]; + tensor scores_7_cast_fp16 = select(a = var_19_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_3)[name = tensor("scores_7_cast_fp16")]; + tensor var_570_cast_fp16 = softmax(axis = var_40, x = scores_7_cast_fp16)[name = tensor("op_570_cast_fp16")]; + tensor input_93_cast_fp16 = select(a = var_18_to_fp16, b = var_570_cast_fp16, cond = mask_3)[name = tensor("input_93_cast_fp16")]; + tensor x_39_transpose_x_0 = const()[name = tensor("x_39_transpose_x_0"), val = tensor(false)]; + tensor x_39_transpose_y_0 = const()[name = tensor("x_39_transpose_y_0"), val = tensor(false)]; + tensor value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = v_3_cast_fp16)[name = tensor("transpose_229")]; + tensor x_39_cast_fp16 = matmul(transpose_x = x_39_transpose_x_0, transpose_y = x_39_transpose_y_0, x = input_93_cast_fp16, y = value_3_cast_fp16)[name = tensor("x_39_cast_fp16")]; + tensor var_574_perm_0 = const()[name = tensor("op_574_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, -1, 512])]; + tensor var_574_cast_fp16 = transpose(perm = var_574_perm_0, x = x_39_cast_fp16)[name = tensor("transpose_225")]; + tensor input_95_cast_fp16 = reshape(shape = var_575, x = var_574_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_out_weight_to_fp16 = const()[name = tensor("encoder_layers_1_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22901312)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_out_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_99_cast_fp16 = add(x = input_89_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor x_43_axes_0 = const()[name = tensor("x_43_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23425664)))]; + tensor encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = tensor("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23426752)))]; + tensor x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor input_101_perm_0 = const()[name = tensor("input_101_perm_0"), val = tensor([0, 2, 1])]; + tensor input_103_pad_type_0 = const()[name = tensor("input_103_pad_type_0"), val = tensor("valid")]; + tensor input_103_strides_0 = const()[name = tensor("input_103_strides_0"), val = tensor([1])]; + tensor input_103_pad_0 = const()[name = tensor("input_103_pad_0"), val = tensor([0, 0])]; + tensor input_103_dilations_0 = const()[name = tensor("input_103_dilations_0"), val = tensor([1])]; + tensor input_103_groups_0 = const()[name = tensor("input_103_groups_0"), val = tensor(1)]; + tensor encoder_layers_1_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23427840)))]; + tensor input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = tensor("transpose_224")]; + tensor input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = encoder_layers_1_conv_pointwise_conv1_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor x_45_split_num_splits_0 = const()[name = tensor("x_45_split_num_splits_0"), val = tensor(2)]; + tensor x_45_split_axis_0 = const()[name = tensor("x_45_split_axis_0"), val = tensor(1)]; + tensor x_45_split_cast_fp16_0, tensor x_45_split_cast_fp16_1 = split(axis = x_45_split_axis_0, num_splits = x_45_split_num_splits_0, x = input_103_cast_fp16)[name = tensor("x_45_split_cast_fp16")]; + tensor x_45_split_1_sigmoid_cast_fp16 = sigmoid(x = x_45_split_cast_fp16_1)[name = tensor("x_45_split_1_sigmoid_cast_fp16")]; + tensor x_45_cast_fp16 = mul(x = x_45_split_cast_fp16_0, y = x_45_split_1_sigmoid_cast_fp16)[name = tensor("x_45_cast_fp16")]; + tensor input_105_cast_fp16 = select(a = var_18_to_fp16, b = x_45_cast_fp16, cond = var_396)[name = tensor("input_105_cast_fp16")]; + tensor new_x_7_interleave_0 = const()[name = tensor("new_x_7_interleave_0"), val = tensor(false)]; + tensor new_x_7_cast_fp16 = concat(axis = var_40, interleave = new_x_7_interleave_0, values = (cache_7_cast_fp16, input_105_cast_fp16))[name = tensor("new_x_7_cast_fp16")]; + tensor var_613_begin_0 = const()[name = tensor("op_613_begin_0"), val = tensor([0, 0, 4])]; + tensor var_613_end_0 = const()[name = tensor("op_613_end_0"), val = tensor([1, 512, 12])]; + tensor var_613_end_mask_0 = const()[name = tensor("op_613_end_mask_0"), val = tensor([true, true, true])]; + tensor var_613_cast_fp16 = slice_by_index(begin = var_613_begin_0, end = var_613_end_0, end_mask = var_613_end_mask_0, x = new_x_7_cast_fp16)[name = tensor("op_613_cast_fp16")]; + tensor x_47_pad_type_0 = const()[name = tensor("x_47_pad_type_0"), val = tensor("valid")]; + tensor x_47_groups_0 = const()[name = tensor("x_47_groups_0"), val = tensor(512)]; + tensor x_47_strides_0 = const()[name = tensor("x_47_strides_0"), val = tensor([1])]; + tensor x_47_pad_0 = const()[name = tensor("x_47_pad_0"), val = tensor([0, 0])]; + tensor x_47_dilations_0 = const()[name = tensor("x_47_dilations_0"), val = tensor([1])]; + tensor encoder_layers_1_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_1_conv_depthwise_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24476480)))]; + tensor x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = encoder_layers_1_conv_depthwise_conv_weight_to_fp16, x = new_x_7_cast_fp16)[name = tensor("x_47_cast_fp16")]; + tensor input_107_perm_0 = const()[name = tensor("input_107_perm_0"), val = tensor([0, 2, 1])]; + tensor x_49_axes_0 = const()[name = tensor("x_49_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = tensor("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24485760)))]; + tensor encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = tensor("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24486848)))]; + tensor input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = tensor("transpose_223")]; + tensor x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("x_49_cast_fp16")]; + tensor input_109_perm_0 = const()[name = tensor("input_109_perm_0"), val = tensor([0, 2, 1])]; + tensor input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = tensor("transpose_222")]; + tensor input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor x_51_pad_type_0 = const()[name = tensor("x_51_pad_type_0"), val = tensor("valid")]; + tensor x_51_strides_0 = const()[name = tensor("x_51_strides_0"), val = tensor([1])]; + tensor x_51_pad_0 = const()[name = tensor("x_51_pad_0"), val = tensor([0, 0])]; + tensor x_51_dilations_0 = const()[name = tensor("x_51_dilations_0"), val = tensor([1])]; + tensor x_51_groups_0 = const()[name = tensor("x_51_groups_0"), val = tensor(1)]; + tensor encoder_layers_1_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24487936)))]; + tensor x_51_cast_fp16 = conv(dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = encoder_layers_1_conv_pointwise_conv2_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor input_113_perm_0 = const()[name = tensor("input_113_perm_0"), val = tensor([0, 2, 1])]; + tensor input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = tensor("transpose_221")]; + tensor input_115_cast_fp16 = add(x = input_99_cast_fp16, y = input_113_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor input_117_axes_0 = const()[name = tensor("input_117_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = tensor("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25012288)))]; + tensor encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = tensor("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25013376)))]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor encoder_layers_1_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_1_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25014464)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_1_feed_forward2_linear1_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor input_121_cast_fp16 = silu(x = linear_17_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor encoder_layers_1_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_1_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27111680)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_feed_forward2_linear2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_654_to_fp16 = const()[name = tensor("op_654_to_fp16"), val = tensor(0x1p-1)]; + tensor var_655_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_654_to_fp16)[name = tensor("op_655_cast_fp16")]; + tensor input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_655_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = tensor("input_129_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_out_weight_to_fp16 = const()[name = tensor("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29208896)))]; + tensor encoder_layers_1_norm_out_bias_to_fp16 = const()[name = tensor("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29209984)))]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor cache_9_begin_0 = const()[name = tensor("cache_9_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor cache_9_end_0 = const()[name = tensor("cache_9_end_0"), val = tensor([3, 1, 70, 512])]; + tensor cache_9_end_mask_0 = const()[name = tensor("cache_9_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_9_squeeze_mask_0 = const()[name = tensor("cache_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_9_cast_fp16 = slice_by_index(begin = cache_9_begin_0, end = cache_9_end_0, end_mask = cache_9_end_mask_0, squeeze_mask = cache_9_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor("cache_9_cast_fp16")]; + tensor cache_11_begin_0 = const()[name = tensor("cache_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor cache_11_end_0 = const()[name = tensor("cache_11_end_0"), val = tensor([3, 1, 512, 8])]; + tensor cache_11_end_mask_0 = const()[name = tensor("cache_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_11_squeeze_mask_0 = const()[name = tensor("cache_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_11_cast_fp16 = slice_by_index(begin = cache_11_begin_0, end = cache_11_end_0, end_mask = cache_11_end_mask_0, squeeze_mask = cache_11_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor("cache_11_cast_fp16")]; + tensor input_131_axes_0 = const()[name = tensor("input_131_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = tensor("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29211072)))]; + tensor encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = tensor("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29212160)))]; + tensor input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor encoder_layers_2_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_2_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29213248)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_2_feed_forward1_linear1_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_135_cast_fp16 = silu(x = linear_19_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor encoder_layers_2_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_2_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31310464)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_feed_forward1_linear2_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_689_to_fp16 = const()[name = tensor("op_689_to_fp16"), val = tensor(0x1p-1)]; + tensor var_690_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_689_to_fp16)[name = tensor("op_690_cast_fp16")]; + tensor input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_690_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor key_5_axes_0 = const()[name = tensor("key_5_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = tensor("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33407680)))]; + tensor encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = tensor("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33408768)))]; + tensor key_5_cast_fp16 = layer_norm(axes = key_5_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("key_5_cast_fp16")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143_cast_fp16 = concat(axis = var_42, interleave = input_143_interleave_0, values = (cache_9_cast_fp16, key_5_cast_fp16))[name = tensor("input_143_cast_fp16")]; + tensor var_712_begin_0 = const()[name = tensor("op_712_begin_0"), val = tensor([0, 4, 0])]; + tensor var_712_end_0 = const()[name = tensor("op_712_end_0"), val = tensor([1, 70, 512])]; + tensor var_712_end_mask_0 = const()[name = tensor("op_712_end_mask_0"), val = tensor([true, true, true])]; + tensor var_712_cast_fp16 = slice_by_index(begin = var_712_begin_0, end = var_712_end_0, end_mask = var_712_end_mask_0, x = cache_9_cast_fp16)[name = tensor("op_712_cast_fp16")]; + tensor var_718_interleave_0 = const()[name = tensor("op_718_interleave_0"), val = tensor(false)]; + tensor var_718_cast_fp16 = concat(axis = var_42, interleave = var_718_interleave_0, values = (var_712_cast_fp16, key_5_cast_fp16))[name = tensor("op_718_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_q_weight_to_fp16 = const()[name = tensor("encoder_layers_2_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33409856)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_q_weight_to_fp16, x = key_5_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_722 = const()[name = tensor("op_722"), val = tensor([1, -1, 8, 64])]; + tensor q_13_cast_fp16 = reshape(shape = var_722, x = linear_21_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_k_weight_to_fp16 = const()[name = tensor("encoder_layers_2_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33934208)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_k_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_726 = const()[name = tensor("op_726"), val = tensor([1, -1, 8, 64])]; + tensor k_9_cast_fp16 = reshape(shape = var_726, x = linear_22_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_v_weight_to_fp16 = const()[name = tensor("encoder_layers_2_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34458560)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_v_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor var_730 = const()[name = tensor("op_730"), val = tensor([1, -1, 8, 64])]; + tensor v_5_cast_fp16 = reshape(shape = var_730, x = linear_23_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor value_5_perm_0 = const()[name = tensor("value_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor encoder_layers_2_self_attn_pos_bias_u_to_fp16 = const()[name = tensor("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34982912)))]; + tensor var_742_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = tensor("op_742_cast_fp16")]; + tensor encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = tensor("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34984000)))]; + tensor var_744_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = tensor("op_744_cast_fp16")]; + tensor q_with_bias_v_5_perm_0 = const()[name = tensor("q_with_bias_v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor x_59_transpose_x_0 = const()[name = tensor("x_59_transpose_x_0"), val = tensor(false)]; + tensor x_59_transpose_y_0 = const()[name = tensor("x_59_transpose_y_0"), val = tensor(false)]; + tensor var_746_to_fp16 = const()[name = tensor("op_746_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34985088)))]; + tensor q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_744_cast_fp16)[name = tensor("transpose_219")]; + tensor x_59_cast_fp16 = matmul(transpose_x = x_59_transpose_x_0, transpose_y = x_59_transpose_y_0, x = q_with_bias_v_5_cast_fp16, y = var_746_to_fp16)[name = tensor("x_59_cast_fp16")]; + tensor x_61_pad_0 = const()[name = tensor("x_61_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor x_61_mode_0 = const()[name = tensor("x_61_mode_0"), val = tensor("constant")]; + tensor const_49_to_fp16 = const()[name = tensor("const_49_to_fp16"), val = tensor(0x0p+0)]; + tensor x_61_cast_fp16 = pad(constant_val = const_49_to_fp16, mode = x_61_mode_0, pad = x_61_pad_0, x = x_59_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor var_754 = const()[name = tensor("op_754"), val = tensor([1, 8, -1, 4])]; + tensor x_63_cast_fp16 = reshape(shape = var_754, x = x_61_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor var_758_begin_0 = const()[name = tensor("op_758_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_758_end_0 = const()[name = tensor("op_758_end_0"), val = tensor([1, 8, 148, 4])]; + tensor var_758_end_mask_0 = const()[name = tensor("op_758_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_758_cast_fp16 = slice_by_index(begin = var_758_begin_0, end = var_758_end_0, end_mask = var_758_end_mask_0, x = x_63_cast_fp16)[name = tensor("op_758_cast_fp16")]; + tensor var_759 = const()[name = tensor("op_759"), val = tensor([1, 8, 4, 147])]; + tensor matrix_bd_9_cast_fp16 = reshape(shape = var_759, x = var_758_cast_fp16)[name = tensor("matrix_bd_9_cast_fp16")]; + tensor matrix_ac_5_transpose_x_0 = const()[name = tensor("matrix_ac_5_transpose_x_0"), val = tensor(false)]; + tensor matrix_ac_5_transpose_y_0 = const()[name = tensor("matrix_ac_5_transpose_y_0"), val = tensor(false)]; + tensor transpose_55_perm_0 = const()[name = tensor("transpose_55_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_56_perm_0 = const()[name = tensor("transpose_56_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_56 = transpose(perm = transpose_56_perm_0, x = k_9_cast_fp16)[name = tensor("transpose_217")]; + tensor transpose_55 = transpose(perm = transpose_55_perm_0, x = var_742_cast_fp16)[name = tensor("transpose_218")]; + tensor matrix_ac_5_cast_fp16 = matmul(transpose_x = matrix_ac_5_transpose_x_0, transpose_y = matrix_ac_5_transpose_y_0, x = transpose_55, y = transpose_56)[name = tensor("matrix_ac_5_cast_fp16")]; + tensor matrix_bd_11_begin_0 = const()[name = tensor("matrix_bd_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_11_end_0 = const()[name = tensor("matrix_bd_11_end_0"), val = tensor([1, 8, 4, 74])]; + tensor matrix_bd_11_end_mask_0 = const()[name = tensor("matrix_bd_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_11_cast_fp16 = slice_by_index(begin = matrix_bd_11_begin_0, end = matrix_bd_11_end_0, end_mask = matrix_bd_11_end_mask_0, x = matrix_bd_9_cast_fp16)[name = tensor("matrix_bd_11_cast_fp16")]; + tensor var_768_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = tensor("op_768_cast_fp16")]; + tensor _inversed_scores_9_y_0_to_fp16 = const()[name = tensor("_inversed_scores_9_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor _inversed_scores_9_cast_fp16 = mul(x = var_768_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = tensor("_inversed_scores_9_cast_fp16")]; + tensor scores_11_cast_fp16 = select(a = var_19_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_3)[name = tensor("scores_11_cast_fp16")]; + tensor var_774_cast_fp16 = softmax(axis = var_40, x = scores_11_cast_fp16)[name = tensor("op_774_cast_fp16")]; + tensor input_145_cast_fp16 = select(a = var_18_to_fp16, b = var_774_cast_fp16, cond = mask_3)[name = tensor("input_145_cast_fp16")]; + tensor x_65_transpose_x_0 = const()[name = tensor("x_65_transpose_x_0"), val = tensor(false)]; + tensor x_65_transpose_y_0 = const()[name = tensor("x_65_transpose_y_0"), val = tensor(false)]; + tensor value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = v_5_cast_fp16)[name = tensor("transpose_220")]; + tensor x_65_cast_fp16 = matmul(transpose_x = x_65_transpose_x_0, transpose_y = x_65_transpose_y_0, x = input_145_cast_fp16, y = value_5_cast_fp16)[name = tensor("x_65_cast_fp16")]; + tensor var_778_perm_0 = const()[name = tensor("op_778_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_779 = const()[name = tensor("op_779"), val = tensor([1, -1, 512])]; + tensor var_778_cast_fp16 = transpose(perm = var_778_perm_0, x = x_65_cast_fp16)[name = tensor("transpose_216")]; + tensor input_147_cast_fp16 = reshape(shape = var_779, x = var_778_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_out_weight_to_fp16 = const()[name = tensor("encoder_layers_2_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35135680)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_out_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor input_151_cast_fp16 = add(x = input_141_cast_fp16, y = linear_25_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor x_69_axes_0 = const()[name = tensor("x_69_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35660032)))]; + tensor encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = tensor("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35661120)))]; + tensor x_69_cast_fp16 = layer_norm(axes = x_69_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor input_153_perm_0 = const()[name = tensor("input_153_perm_0"), val = tensor([0, 2, 1])]; + tensor input_155_pad_type_0 = const()[name = tensor("input_155_pad_type_0"), val = tensor("valid")]; + tensor input_155_strides_0 = const()[name = tensor("input_155_strides_0"), val = tensor([1])]; + tensor input_155_pad_0 = const()[name = tensor("input_155_pad_0"), val = tensor([0, 0])]; + tensor input_155_dilations_0 = const()[name = tensor("input_155_dilations_0"), val = tensor([1])]; + tensor input_155_groups_0 = const()[name = tensor("input_155_groups_0"), val = tensor(1)]; + tensor encoder_layers_2_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35662208)))]; + tensor input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = tensor("transpose_215")]; + tensor input_155_cast_fp16 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = encoder_layers_2_conv_pointwise_conv1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor x_71_split_num_splits_0 = const()[name = tensor("x_71_split_num_splits_0"), val = tensor(2)]; + tensor x_71_split_axis_0 = const()[name = tensor("x_71_split_axis_0"), val = tensor(1)]; + tensor x_71_split_cast_fp16_0, tensor x_71_split_cast_fp16_1 = split(axis = x_71_split_axis_0, num_splits = x_71_split_num_splits_0, x = input_155_cast_fp16)[name = tensor("x_71_split_cast_fp16")]; + tensor x_71_split_1_sigmoid_cast_fp16 = sigmoid(x = x_71_split_cast_fp16_1)[name = tensor("x_71_split_1_sigmoid_cast_fp16")]; + tensor x_71_cast_fp16 = mul(x = x_71_split_cast_fp16_0, y = x_71_split_1_sigmoid_cast_fp16)[name = tensor("x_71_cast_fp16")]; + tensor input_157_cast_fp16 = select(a = var_18_to_fp16, b = x_71_cast_fp16, cond = var_396)[name = tensor("input_157_cast_fp16")]; + tensor new_x_11_interleave_0 = const()[name = tensor("new_x_11_interleave_0"), val = tensor(false)]; + tensor new_x_11_cast_fp16 = concat(axis = var_40, interleave = new_x_11_interleave_0, values = (cache_11_cast_fp16, input_157_cast_fp16))[name = tensor("new_x_11_cast_fp16")]; + tensor var_817_begin_0 = const()[name = tensor("op_817_begin_0"), val = tensor([0, 0, 4])]; + tensor var_817_end_0 = const()[name = tensor("op_817_end_0"), val = tensor([1, 512, 12])]; + tensor var_817_end_mask_0 = const()[name = tensor("op_817_end_mask_0"), val = tensor([true, true, true])]; + tensor var_817_cast_fp16 = slice_by_index(begin = var_817_begin_0, end = var_817_end_0, end_mask = var_817_end_mask_0, x = new_x_11_cast_fp16)[name = tensor("op_817_cast_fp16")]; + tensor x_73_pad_type_0 = const()[name = tensor("x_73_pad_type_0"), val = tensor("valid")]; + tensor x_73_groups_0 = const()[name = tensor("x_73_groups_0"), val = tensor(512)]; + tensor x_73_strides_0 = const()[name = tensor("x_73_strides_0"), val = tensor([1])]; + tensor x_73_pad_0 = const()[name = tensor("x_73_pad_0"), val = tensor([0, 0])]; + tensor x_73_dilations_0 = const()[name = tensor("x_73_dilations_0"), val = tensor([1])]; + tensor encoder_layers_2_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_2_conv_depthwise_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36710848)))]; + tensor x_73_cast_fp16 = conv(dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = encoder_layers_2_conv_depthwise_conv_weight_to_fp16, x = new_x_11_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor input_159_perm_0 = const()[name = tensor("input_159_perm_0"), val = tensor([0, 2, 1])]; + tensor x_75_axes_0 = const()[name = tensor("x_75_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = tensor("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36720128)))]; + tensor encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = tensor("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36721216)))]; + tensor input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = tensor("transpose_214")]; + tensor x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor input_161_perm_0 = const()[name = tensor("input_161_perm_0"), val = tensor([0, 2, 1])]; + tensor input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = tensor("transpose_213")]; + tensor input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor x_77_pad_type_0 = const()[name = tensor("x_77_pad_type_0"), val = tensor("valid")]; + tensor x_77_strides_0 = const()[name = tensor("x_77_strides_0"), val = tensor([1])]; + tensor x_77_pad_0 = const()[name = tensor("x_77_pad_0"), val = tensor([0, 0])]; + tensor x_77_dilations_0 = const()[name = tensor("x_77_dilations_0"), val = tensor([1])]; + tensor x_77_groups_0 = const()[name = tensor("x_77_groups_0"), val = tensor(1)]; + tensor encoder_layers_2_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36722304)))]; + tensor x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = encoder_layers_2_conv_pointwise_conv2_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("x_77_cast_fp16")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = tensor("transpose_212")]; + tensor input_167_cast_fp16 = add(x = input_151_cast_fp16, y = input_165_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor input_169_axes_0 = const()[name = tensor("input_169_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = tensor("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37246656)))]; + tensor encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = tensor("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37247744)))]; + tensor input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor encoder_layers_2_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_2_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37248832)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_2_feed_forward2_linear1_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_173_cast_fp16 = silu(x = linear_26_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor encoder_layers_2_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_2_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39346048)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_feed_forward2_linear2_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor var_858_to_fp16 = const()[name = tensor("op_858_to_fp16"), val = tensor(0x1p-1)]; + tensor var_859_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_858_to_fp16)[name = tensor("op_859_cast_fp16")]; + tensor input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_859_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor input_181_axes_0 = const()[name = tensor("input_181_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_out_weight_to_fp16 = const()[name = tensor("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41443264)))]; + tensor encoder_layers_2_norm_out_bias_to_fp16 = const()[name = tensor("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41444352)))]; + tensor input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor cache_13_begin_0 = const()[name = tensor("cache_13_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor cache_13_end_0 = const()[name = tensor("cache_13_end_0"), val = tensor([4, 1, 70, 512])]; + tensor cache_13_end_mask_0 = const()[name = tensor("cache_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_13_squeeze_mask_0 = const()[name = tensor("cache_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_13_cast_fp16 = slice_by_index(begin = cache_13_begin_0, end = cache_13_end_0, end_mask = cache_13_end_mask_0, squeeze_mask = cache_13_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor("cache_13_cast_fp16")]; + tensor cache_15_begin_0 = const()[name = tensor("cache_15_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor cache_15_end_0 = const()[name = tensor("cache_15_end_0"), val = tensor([4, 1, 512, 8])]; + tensor cache_15_end_mask_0 = const()[name = tensor("cache_15_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_15_squeeze_mask_0 = const()[name = tensor("cache_15_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_15_cast_fp16 = slice_by_index(begin = cache_15_begin_0, end = cache_15_end_0, end_mask = cache_15_end_mask_0, squeeze_mask = cache_15_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor("cache_15_cast_fp16")]; + tensor input_183_axes_0 = const()[name = tensor("input_183_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = tensor("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41445440)))]; + tensor encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = tensor("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41446528)))]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor encoder_layers_3_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_3_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41447616)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_3_feed_forward1_linear1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor input_187_cast_fp16 = silu(x = linear_28_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor encoder_layers_3_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_3_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43544832)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_feed_forward1_linear2_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_893_to_fp16 = const()[name = tensor("op_893_to_fp16"), val = tensor(0x1p-1)]; + tensor var_894_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_893_to_fp16)[name = tensor("op_894_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_894_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor key_7_axes_0 = const()[name = tensor("key_7_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = tensor("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(45642048)))]; + tensor encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = tensor("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(45643136)))]; + tensor key_7_cast_fp16 = layer_norm(axes = key_7_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("key_7_cast_fp16")]; + tensor input_195_interleave_0 = const()[name = tensor("input_195_interleave_0"), val = tensor(false)]; + tensor input_195_cast_fp16 = concat(axis = var_42, interleave = input_195_interleave_0, values = (cache_13_cast_fp16, key_7_cast_fp16))[name = tensor("input_195_cast_fp16")]; + tensor var_916_begin_0 = const()[name = tensor("op_916_begin_0"), val = tensor([0, 4, 0])]; + tensor var_916_end_0 = const()[name = tensor("op_916_end_0"), val = tensor([1, 70, 512])]; + tensor var_916_end_mask_0 = const()[name = tensor("op_916_end_mask_0"), val = tensor([true, true, true])]; + tensor var_916_cast_fp16 = slice_by_index(begin = var_916_begin_0, end = var_916_end_0, end_mask = var_916_end_mask_0, x = cache_13_cast_fp16)[name = tensor("op_916_cast_fp16")]; + tensor var_922_interleave_0 = const()[name = tensor("op_922_interleave_0"), val = tensor(false)]; + tensor var_922_cast_fp16 = concat(axis = var_42, interleave = var_922_interleave_0, values = (var_916_cast_fp16, key_7_cast_fp16))[name = tensor("op_922_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_q_weight_to_fp16 = const()[name = tensor("encoder_layers_3_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(45644224)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_q_weight_to_fp16, x = key_7_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor var_926 = const()[name = tensor("op_926"), val = tensor([1, -1, 8, 64])]; + tensor q_19_cast_fp16 = reshape(shape = var_926, x = linear_30_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_k_weight_to_fp16 = const()[name = tensor("encoder_layers_3_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46168576)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_k_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor var_930 = const()[name = tensor("op_930"), val = tensor([1, -1, 8, 64])]; + tensor k_13_cast_fp16 = reshape(shape = var_930, x = linear_31_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_v_weight_to_fp16 = const()[name = tensor("encoder_layers_3_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46692928)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_v_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_934 = const()[name = tensor("op_934"), val = tensor([1, -1, 8, 64])]; + tensor v_7_cast_fp16 = reshape(shape = var_934, x = linear_32_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor value_7_perm_0 = const()[name = tensor("value_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor encoder_layers_3_self_attn_pos_bias_u_to_fp16 = const()[name = tensor("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47217280)))]; + tensor var_946_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = tensor("op_946_cast_fp16")]; + tensor encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = tensor("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47218368)))]; + tensor var_948_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = tensor("op_948_cast_fp16")]; + tensor q_with_bias_v_7_perm_0 = const()[name = tensor("q_with_bias_v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor x_85_transpose_x_0 = const()[name = tensor("x_85_transpose_x_0"), val = tensor(false)]; + tensor x_85_transpose_y_0 = const()[name = tensor("x_85_transpose_y_0"), val = tensor(false)]; + tensor var_950_to_fp16 = const()[name = tensor("op_950_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47219456)))]; + tensor q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_948_cast_fp16)[name = tensor("transpose_210")]; + tensor x_85_cast_fp16 = matmul(transpose_x = x_85_transpose_x_0, transpose_y = x_85_transpose_y_0, x = q_with_bias_v_7_cast_fp16, y = var_950_to_fp16)[name = tensor("x_85_cast_fp16")]; + tensor x_87_pad_0 = const()[name = tensor("x_87_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor x_87_mode_0 = const()[name = tensor("x_87_mode_0"), val = tensor("constant")]; + tensor const_62_to_fp16 = const()[name = tensor("const_62_to_fp16"), val = tensor(0x0p+0)]; + tensor x_87_cast_fp16 = pad(constant_val = const_62_to_fp16, mode = x_87_mode_0, pad = x_87_pad_0, x = x_85_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor var_958 = const()[name = tensor("op_958"), val = tensor([1, 8, -1, 4])]; + tensor x_89_cast_fp16 = reshape(shape = var_958, x = x_87_cast_fp16)[name = tensor("x_89_cast_fp16")]; + tensor var_962_begin_0 = const()[name = tensor("op_962_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_962_end_0 = const()[name = tensor("op_962_end_0"), val = tensor([1, 8, 148, 4])]; + tensor var_962_end_mask_0 = const()[name = tensor("op_962_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_962_cast_fp16 = slice_by_index(begin = var_962_begin_0, end = var_962_end_0, end_mask = var_962_end_mask_0, x = x_89_cast_fp16)[name = tensor("op_962_cast_fp16")]; + tensor var_963 = const()[name = tensor("op_963"), val = tensor([1, 8, 4, 147])]; + tensor matrix_bd_13_cast_fp16 = reshape(shape = var_963, x = var_962_cast_fp16)[name = tensor("matrix_bd_13_cast_fp16")]; + tensor matrix_ac_7_transpose_x_0 = const()[name = tensor("matrix_ac_7_transpose_x_0"), val = tensor(false)]; + tensor matrix_ac_7_transpose_y_0 = const()[name = tensor("matrix_ac_7_transpose_y_0"), val = tensor(false)]; + tensor transpose_57_perm_0 = const()[name = tensor("transpose_57_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_58_perm_0 = const()[name = tensor("transpose_58_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_58 = transpose(perm = transpose_58_perm_0, x = k_13_cast_fp16)[name = tensor("transpose_208")]; + tensor transpose_57 = transpose(perm = transpose_57_perm_0, x = var_946_cast_fp16)[name = tensor("transpose_209")]; + tensor matrix_ac_7_cast_fp16 = matmul(transpose_x = matrix_ac_7_transpose_x_0, transpose_y = matrix_ac_7_transpose_y_0, x = transpose_57, y = transpose_58)[name = tensor("matrix_ac_7_cast_fp16")]; + tensor matrix_bd_15_begin_0 = const()[name = tensor("matrix_bd_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_15_end_0 = const()[name = tensor("matrix_bd_15_end_0"), val = tensor([1, 8, 4, 74])]; + tensor matrix_bd_15_end_mask_0 = const()[name = tensor("matrix_bd_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_15_cast_fp16 = slice_by_index(begin = matrix_bd_15_begin_0, end = matrix_bd_15_end_0, end_mask = matrix_bd_15_end_mask_0, x = matrix_bd_13_cast_fp16)[name = tensor("matrix_bd_15_cast_fp16")]; + tensor var_972_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = tensor("op_972_cast_fp16")]; + tensor _inversed_scores_13_y_0_to_fp16 = const()[name = tensor("_inversed_scores_13_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor _inversed_scores_13_cast_fp16 = mul(x = var_972_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = tensor("_inversed_scores_13_cast_fp16")]; + tensor scores_15_cast_fp16 = select(a = var_19_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_3)[name = tensor("scores_15_cast_fp16")]; + tensor var_978_cast_fp16 = softmax(axis = var_40, x = scores_15_cast_fp16)[name = tensor("op_978_cast_fp16")]; + tensor input_197_cast_fp16 = select(a = var_18_to_fp16, b = var_978_cast_fp16, cond = mask_3)[name = tensor("input_197_cast_fp16")]; + tensor x_91_transpose_x_0 = const()[name = tensor("x_91_transpose_x_0"), val = tensor(false)]; + tensor x_91_transpose_y_0 = const()[name = tensor("x_91_transpose_y_0"), val = tensor(false)]; + tensor value_7_cast_fp16 = transpose(perm = value_7_perm_0, x = v_7_cast_fp16)[name = tensor("transpose_211")]; + tensor x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = input_197_cast_fp16, y = value_7_cast_fp16)[name = tensor("x_91_cast_fp16")]; + tensor var_982_perm_0 = const()[name = tensor("op_982_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_983 = const()[name = tensor("op_983"), val = tensor([1, -1, 512])]; + tensor var_982_cast_fp16 = transpose(perm = var_982_perm_0, x = x_91_cast_fp16)[name = tensor("transpose_207")]; + tensor input_199_cast_fp16 = reshape(shape = var_983, x = var_982_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_out_weight_to_fp16 = const()[name = tensor("encoder_layers_3_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47370048)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_out_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_193_cast_fp16, y = linear_34_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor x_95_axes_0 = const()[name = tensor("x_95_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47894400)))]; + tensor encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = tensor("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47895488)))]; + tensor x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("x_95_cast_fp16")]; + tensor input_205_perm_0 = const()[name = tensor("input_205_perm_0"), val = tensor([0, 2, 1])]; + tensor input_207_pad_type_0 = const()[name = tensor("input_207_pad_type_0"), val = tensor("valid")]; + tensor input_207_strides_0 = const()[name = tensor("input_207_strides_0"), val = tensor([1])]; + tensor input_207_pad_0 = const()[name = tensor("input_207_pad_0"), val = tensor([0, 0])]; + tensor input_207_dilations_0 = const()[name = tensor("input_207_dilations_0"), val = tensor([1])]; + tensor input_207_groups_0 = const()[name = tensor("input_207_groups_0"), val = tensor(1)]; + tensor encoder_layers_3_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47896576)))]; + tensor input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = tensor("transpose_206")]; + tensor input_207_cast_fp16 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = encoder_layers_3_conv_pointwise_conv1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor x_97_split_num_splits_0 = const()[name = tensor("x_97_split_num_splits_0"), val = tensor(2)]; + tensor x_97_split_axis_0 = const()[name = tensor("x_97_split_axis_0"), val = tensor(1)]; + tensor x_97_split_cast_fp16_0, tensor x_97_split_cast_fp16_1 = split(axis = x_97_split_axis_0, num_splits = x_97_split_num_splits_0, x = input_207_cast_fp16)[name = tensor("x_97_split_cast_fp16")]; + tensor x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = tensor("x_97_split_1_sigmoid_cast_fp16")]; + tensor x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = tensor("x_97_cast_fp16")]; + tensor input_209_cast_fp16 = select(a = var_18_to_fp16, b = x_97_cast_fp16, cond = var_396)[name = tensor("input_209_cast_fp16")]; + tensor new_x_15_interleave_0 = const()[name = tensor("new_x_15_interleave_0"), val = tensor(false)]; + tensor new_x_15_cast_fp16 = concat(axis = var_40, interleave = new_x_15_interleave_0, values = (cache_15_cast_fp16, input_209_cast_fp16))[name = tensor("new_x_15_cast_fp16")]; + tensor var_1021_begin_0 = const()[name = tensor("op_1021_begin_0"), val = tensor([0, 0, 4])]; + tensor var_1021_end_0 = const()[name = tensor("op_1021_end_0"), val = tensor([1, 512, 12])]; + tensor var_1021_end_mask_0 = const()[name = tensor("op_1021_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1021_cast_fp16 = slice_by_index(begin = var_1021_begin_0, end = var_1021_end_0, end_mask = var_1021_end_mask_0, x = new_x_15_cast_fp16)[name = tensor("op_1021_cast_fp16")]; + tensor x_99_pad_type_0 = const()[name = tensor("x_99_pad_type_0"), val = tensor("valid")]; + tensor x_99_groups_0 = const()[name = tensor("x_99_groups_0"), val = tensor(512)]; + tensor x_99_strides_0 = const()[name = tensor("x_99_strides_0"), val = tensor([1])]; + tensor x_99_pad_0 = const()[name = tensor("x_99_pad_0"), val = tensor([0, 0])]; + tensor x_99_dilations_0 = const()[name = tensor("x_99_dilations_0"), val = tensor([1])]; + tensor encoder_layers_3_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_3_conv_depthwise_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48945216)))]; + tensor x_99_cast_fp16 = conv(dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = encoder_layers_3_conv_depthwise_conv_weight_to_fp16, x = new_x_15_cast_fp16)[name = tensor("x_99_cast_fp16")]; + tensor input_211_perm_0 = const()[name = tensor("input_211_perm_0"), val = tensor([0, 2, 1])]; + tensor x_101_axes_0 = const()[name = tensor("x_101_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = tensor("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48954496)))]; + tensor encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = tensor("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48955584)))]; + tensor input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = tensor("transpose_205")]; + tensor x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("x_101_cast_fp16")]; + tensor input_213_perm_0 = const()[name = tensor("input_213_perm_0"), val = tensor([0, 2, 1])]; + tensor input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = tensor("transpose_204")]; + tensor input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor x_103_pad_type_0 = const()[name = tensor("x_103_pad_type_0"), val = tensor("valid")]; + tensor x_103_strides_0 = const()[name = tensor("x_103_strides_0"), val = tensor([1])]; + tensor x_103_pad_0 = const()[name = tensor("x_103_pad_0"), val = tensor([0, 0])]; + tensor x_103_dilations_0 = const()[name = tensor("x_103_dilations_0"), val = tensor([1])]; + tensor x_103_groups_0 = const()[name = tensor("x_103_groups_0"), val = tensor(1)]; + tensor encoder_layers_3_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48956672)))]; + tensor x_103_cast_fp16 = conv(dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = encoder_layers_3_conv_pointwise_conv2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("x_103_cast_fp16")]; + tensor input_217_perm_0 = const()[name = tensor("input_217_perm_0"), val = tensor([0, 2, 1])]; + tensor input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = tensor("transpose_203")]; + tensor input_219_cast_fp16 = add(x = input_203_cast_fp16, y = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor input_221_axes_0 = const()[name = tensor("input_221_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = tensor("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49481024)))]; + tensor encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = tensor("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49482112)))]; + tensor input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor encoder_layers_3_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_3_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49483200)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_3_feed_forward2_linear1_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_225_cast_fp16 = silu(x = linear_35_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor encoder_layers_3_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_3_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51580416)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_feed_forward2_linear2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor var_1062_to_fp16 = const()[name = tensor("op_1062_to_fp16"), val = tensor(0x1p-1)]; + tensor var_1063_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1062_to_fp16)[name = tensor("op_1063_cast_fp16")]; + tensor input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1063_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor input_233_axes_0 = const()[name = tensor("input_233_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_out_weight_to_fp16 = const()[name = tensor("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53677632)))]; + tensor encoder_layers_3_norm_out_bias_to_fp16 = const()[name = tensor("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53678720)))]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor cache_17_begin_0 = const()[name = tensor("cache_17_begin_0"), val = tensor([4, 0, 0, 0])]; + tensor cache_17_end_0 = const()[name = tensor("cache_17_end_0"), val = tensor([5, 1, 70, 512])]; + tensor cache_17_end_mask_0 = const()[name = tensor("cache_17_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_17_squeeze_mask_0 = const()[name = tensor("cache_17_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_17_cast_fp16 = slice_by_index(begin = cache_17_begin_0, end = cache_17_end_0, end_mask = cache_17_end_mask_0, squeeze_mask = cache_17_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor("cache_17_cast_fp16")]; + tensor cache_19_begin_0 = const()[name = tensor("cache_19_begin_0"), val = tensor([4, 0, 0, 0])]; + tensor cache_19_end_0 = const()[name = tensor("cache_19_end_0"), val = tensor([5, 1, 512, 8])]; + tensor cache_19_end_mask_0 = const()[name = tensor("cache_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_19_squeeze_mask_0 = const()[name = tensor("cache_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_19_cast_fp16 = slice_by_index(begin = cache_19_begin_0, end = cache_19_end_0, end_mask = cache_19_end_mask_0, squeeze_mask = cache_19_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor("cache_19_cast_fp16")]; + tensor input_235_axes_0 = const()[name = tensor("input_235_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = tensor("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53679808)))]; + tensor encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = tensor("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53680896)))]; + tensor input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor encoder_layers_4_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_4_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53681984)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_4_feed_forward1_linear1_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor input_239_cast_fp16 = silu(x = linear_37_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor encoder_layers_4_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_4_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(55779200)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_feed_forward1_linear2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_1097_to_fp16 = const()[name = tensor("op_1097_to_fp16"), val = tensor(0x1p-1)]; + tensor var_1098_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1097_to_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1098_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor key_9_axes_0 = const()[name = tensor("key_9_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = tensor("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57876416)))]; + tensor encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = tensor("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57877504)))]; + tensor key_9_cast_fp16 = layer_norm(axes = key_9_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("key_9_cast_fp16")]; + tensor input_247_interleave_0 = const()[name = tensor("input_247_interleave_0"), val = tensor(false)]; + tensor input_247_cast_fp16 = concat(axis = var_42, interleave = input_247_interleave_0, values = (cache_17_cast_fp16, key_9_cast_fp16))[name = tensor("input_247_cast_fp16")]; + tensor var_1120_begin_0 = const()[name = tensor("op_1120_begin_0"), val = tensor([0, 4, 0])]; + tensor var_1120_end_0 = const()[name = tensor("op_1120_end_0"), val = tensor([1, 70, 512])]; + tensor var_1120_end_mask_0 = const()[name = tensor("op_1120_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1120_cast_fp16 = slice_by_index(begin = var_1120_begin_0, end = var_1120_end_0, end_mask = var_1120_end_mask_0, x = cache_17_cast_fp16)[name = tensor("op_1120_cast_fp16")]; + tensor var_1126_interleave_0 = const()[name = tensor("op_1126_interleave_0"), val = tensor(false)]; + tensor var_1126_cast_fp16 = concat(axis = var_42, interleave = var_1126_interleave_0, values = (var_1120_cast_fp16, key_9_cast_fp16))[name = tensor("op_1126_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_q_weight_to_fp16 = const()[name = tensor("encoder_layers_4_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57878592)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_q_weight_to_fp16, x = key_9_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor var_1130 = const()[name = tensor("op_1130"), val = tensor([1, -1, 8, 64])]; + tensor q_25_cast_fp16 = reshape(shape = var_1130, x = linear_39_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_k_weight_to_fp16 = const()[name = tensor("encoder_layers_4_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58402944)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_k_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_1134 = const()[name = tensor("op_1134"), val = tensor([1, -1, 8, 64])]; + tensor k_17_cast_fp16 = reshape(shape = var_1134, x = linear_40_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_v_weight_to_fp16 = const()[name = tensor("encoder_layers_4_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58927296)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_v_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_1138 = const()[name = tensor("op_1138"), val = tensor([1, -1, 8, 64])]; + tensor v_9_cast_fp16 = reshape(shape = var_1138, x = linear_41_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor value_9_perm_0 = const()[name = tensor("value_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor encoder_layers_4_self_attn_pos_bias_u_to_fp16 = const()[name = tensor("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59451648)))]; + tensor var_1150_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = tensor("op_1150_cast_fp16")]; + tensor encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = tensor("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59452736)))]; + tensor var_1152_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = tensor("op_1152_cast_fp16")]; + tensor q_with_bias_v_9_perm_0 = const()[name = tensor("q_with_bias_v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor x_111_transpose_x_0 = const()[name = tensor("x_111_transpose_x_0"), val = tensor(false)]; + tensor x_111_transpose_y_0 = const()[name = tensor("x_111_transpose_y_0"), val = tensor(false)]; + tensor var_1154_to_fp16 = const()[name = tensor("op_1154_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59453824)))]; + tensor q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1152_cast_fp16)[name = tensor("transpose_201")]; + tensor x_111_cast_fp16 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = q_with_bias_v_9_cast_fp16, y = var_1154_to_fp16)[name = tensor("x_111_cast_fp16")]; + tensor x_113_pad_0 = const()[name = tensor("x_113_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor x_113_mode_0 = const()[name = tensor("x_113_mode_0"), val = tensor("constant")]; + tensor const_75_to_fp16 = const()[name = tensor("const_75_to_fp16"), val = tensor(0x0p+0)]; + tensor x_113_cast_fp16 = pad(constant_val = const_75_to_fp16, mode = x_113_mode_0, pad = x_113_pad_0, x = x_111_cast_fp16)[name = tensor("x_113_cast_fp16")]; + tensor var_1162 = const()[name = tensor("op_1162"), val = tensor([1, 8, -1, 4])]; + tensor x_115_cast_fp16 = reshape(shape = var_1162, x = x_113_cast_fp16)[name = tensor("x_115_cast_fp16")]; + tensor var_1166_begin_0 = const()[name = tensor("op_1166_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1166_end_0 = const()[name = tensor("op_1166_end_0"), val = tensor([1, 8, 148, 4])]; + tensor var_1166_end_mask_0 = const()[name = tensor("op_1166_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1166_cast_fp16 = slice_by_index(begin = var_1166_begin_0, end = var_1166_end_0, end_mask = var_1166_end_mask_0, x = x_115_cast_fp16)[name = tensor("op_1166_cast_fp16")]; + tensor var_1167 = const()[name = tensor("op_1167"), val = tensor([1, 8, 4, 147])]; + tensor matrix_bd_17_cast_fp16 = reshape(shape = var_1167, x = var_1166_cast_fp16)[name = tensor("matrix_bd_17_cast_fp16")]; + tensor matrix_ac_9_transpose_x_0 = const()[name = tensor("matrix_ac_9_transpose_x_0"), val = tensor(false)]; + tensor matrix_ac_9_transpose_y_0 = const()[name = tensor("matrix_ac_9_transpose_y_0"), val = tensor(false)]; + tensor transpose_59_perm_0 = const()[name = tensor("transpose_59_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_60_perm_0 = const()[name = tensor("transpose_60_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_60 = transpose(perm = transpose_60_perm_0, x = k_17_cast_fp16)[name = tensor("transpose_199")]; + tensor transpose_59 = transpose(perm = transpose_59_perm_0, x = var_1150_cast_fp16)[name = tensor("transpose_200")]; + tensor matrix_ac_9_cast_fp16 = matmul(transpose_x = matrix_ac_9_transpose_x_0, transpose_y = matrix_ac_9_transpose_y_0, x = transpose_59, y = transpose_60)[name = tensor("matrix_ac_9_cast_fp16")]; + tensor matrix_bd_19_begin_0 = const()[name = tensor("matrix_bd_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_19_end_0 = const()[name = tensor("matrix_bd_19_end_0"), val = tensor([1, 8, 4, 74])]; + tensor matrix_bd_19_end_mask_0 = const()[name = tensor("matrix_bd_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_19_cast_fp16 = slice_by_index(begin = matrix_bd_19_begin_0, end = matrix_bd_19_end_0, end_mask = matrix_bd_19_end_mask_0, x = matrix_bd_17_cast_fp16)[name = tensor("matrix_bd_19_cast_fp16")]; + tensor var_1176_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = tensor("op_1176_cast_fp16")]; + tensor _inversed_scores_17_y_0_to_fp16 = const()[name = tensor("_inversed_scores_17_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor _inversed_scores_17_cast_fp16 = mul(x = var_1176_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = tensor("_inversed_scores_17_cast_fp16")]; + tensor scores_19_cast_fp16 = select(a = var_19_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_3)[name = tensor("scores_19_cast_fp16")]; + tensor var_1182_cast_fp16 = softmax(axis = var_40, x = scores_19_cast_fp16)[name = tensor("op_1182_cast_fp16")]; + tensor input_249_cast_fp16 = select(a = var_18_to_fp16, b = var_1182_cast_fp16, cond = mask_3)[name = tensor("input_249_cast_fp16")]; + tensor x_117_transpose_x_0 = const()[name = tensor("x_117_transpose_x_0"), val = tensor(false)]; + tensor x_117_transpose_y_0 = const()[name = tensor("x_117_transpose_y_0"), val = tensor(false)]; + tensor value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = v_9_cast_fp16)[name = tensor("transpose_202")]; + tensor x_117_cast_fp16 = matmul(transpose_x = x_117_transpose_x_0, transpose_y = x_117_transpose_y_0, x = input_249_cast_fp16, y = value_9_cast_fp16)[name = tensor("x_117_cast_fp16")]; + tensor var_1186_perm_0 = const()[name = tensor("op_1186_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1187 = const()[name = tensor("op_1187"), val = tensor([1, -1, 512])]; + tensor var_1186_cast_fp16 = transpose(perm = var_1186_perm_0, x = x_117_cast_fp16)[name = tensor("transpose_198")]; + tensor input_251_cast_fp16 = reshape(shape = var_1187, x = var_1186_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_out_weight_to_fp16 = const()[name = tensor("encoder_layers_4_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59604416)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_out_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_43_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor x_121_axes_0 = const()[name = tensor("x_121_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60128768)))]; + tensor encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = tensor("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60129856)))]; + tensor x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("x_121_cast_fp16")]; + tensor input_257_perm_0 = const()[name = tensor("input_257_perm_0"), val = tensor([0, 2, 1])]; + tensor input_259_pad_type_0 = const()[name = tensor("input_259_pad_type_0"), val = tensor("valid")]; + tensor input_259_strides_0 = const()[name = tensor("input_259_strides_0"), val = tensor([1])]; + tensor input_259_pad_0 = const()[name = tensor("input_259_pad_0"), val = tensor([0, 0])]; + tensor input_259_dilations_0 = const()[name = tensor("input_259_dilations_0"), val = tensor([1])]; + tensor input_259_groups_0 = const()[name = tensor("input_259_groups_0"), val = tensor(1)]; + tensor encoder_layers_4_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60130944)))]; + tensor input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = tensor("transpose_197")]; + tensor input_259_cast_fp16 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = encoder_layers_4_conv_pointwise_conv1_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor x_123_split_num_splits_0 = const()[name = tensor("x_123_split_num_splits_0"), val = tensor(2)]; + tensor x_123_split_axis_0 = const()[name = tensor("x_123_split_axis_0"), val = tensor(1)]; + tensor x_123_split_cast_fp16_0, tensor x_123_split_cast_fp16_1 = split(axis = x_123_split_axis_0, num_splits = x_123_split_num_splits_0, x = input_259_cast_fp16)[name = tensor("x_123_split_cast_fp16")]; + tensor x_123_split_1_sigmoid_cast_fp16 = sigmoid(x = x_123_split_cast_fp16_1)[name = tensor("x_123_split_1_sigmoid_cast_fp16")]; + tensor x_123_cast_fp16 = mul(x = x_123_split_cast_fp16_0, y = x_123_split_1_sigmoid_cast_fp16)[name = tensor("x_123_cast_fp16")]; + tensor input_261_cast_fp16 = select(a = var_18_to_fp16, b = x_123_cast_fp16, cond = var_396)[name = tensor("input_261_cast_fp16")]; + tensor new_x_19_interleave_0 = const()[name = tensor("new_x_19_interleave_0"), val = tensor(false)]; + tensor new_x_19_cast_fp16 = concat(axis = var_40, interleave = new_x_19_interleave_0, values = (cache_19_cast_fp16, input_261_cast_fp16))[name = tensor("new_x_19_cast_fp16")]; + tensor var_1225_begin_0 = const()[name = tensor("op_1225_begin_0"), val = tensor([0, 0, 4])]; + tensor var_1225_end_0 = const()[name = tensor("op_1225_end_0"), val = tensor([1, 512, 12])]; + tensor var_1225_end_mask_0 = const()[name = tensor("op_1225_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1225_cast_fp16 = slice_by_index(begin = var_1225_begin_0, end = var_1225_end_0, end_mask = var_1225_end_mask_0, x = new_x_19_cast_fp16)[name = tensor("op_1225_cast_fp16")]; + tensor x_125_pad_type_0 = const()[name = tensor("x_125_pad_type_0"), val = tensor("valid")]; + tensor x_125_groups_0 = const()[name = tensor("x_125_groups_0"), val = tensor(512)]; + tensor x_125_strides_0 = const()[name = tensor("x_125_strides_0"), val = tensor([1])]; + tensor x_125_pad_0 = const()[name = tensor("x_125_pad_0"), val = tensor([0, 0])]; + tensor x_125_dilations_0 = const()[name = tensor("x_125_dilations_0"), val = tensor([1])]; + tensor encoder_layers_4_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_4_conv_depthwise_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61179584)))]; + tensor x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = encoder_layers_4_conv_depthwise_conv_weight_to_fp16, x = new_x_19_cast_fp16)[name = tensor("x_125_cast_fp16")]; + tensor input_263_perm_0 = const()[name = tensor("input_263_perm_0"), val = tensor([0, 2, 1])]; + tensor x_127_axes_0 = const()[name = tensor("x_127_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = tensor("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61188864)))]; + tensor encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = tensor("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61189952)))]; + tensor input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = tensor("transpose_196")]; + tensor x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("x_127_cast_fp16")]; + tensor input_265_perm_0 = const()[name = tensor("input_265_perm_0"), val = tensor([0, 2, 1])]; + tensor input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = tensor("transpose_195")]; + tensor input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor x_129_pad_type_0 = const()[name = tensor("x_129_pad_type_0"), val = tensor("valid")]; + tensor x_129_strides_0 = const()[name = tensor("x_129_strides_0"), val = tensor([1])]; + tensor x_129_pad_0 = const()[name = tensor("x_129_pad_0"), val = tensor([0, 0])]; + tensor x_129_dilations_0 = const()[name = tensor("x_129_dilations_0"), val = tensor([1])]; + tensor x_129_groups_0 = const()[name = tensor("x_129_groups_0"), val = tensor(1)]; + tensor encoder_layers_4_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61191040)))]; + tensor x_129_cast_fp16 = conv(dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = encoder_layers_4_conv_pointwise_conv2_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("x_129_cast_fp16")]; + tensor input_269_perm_0 = const()[name = tensor("input_269_perm_0"), val = tensor([0, 2, 1])]; + tensor input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = tensor("transpose_194")]; + tensor input_271_cast_fp16 = add(x = input_255_cast_fp16, y = input_269_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor input_273_axes_0 = const()[name = tensor("input_273_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = tensor("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61715392)))]; + tensor encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = tensor("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61716480)))]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor encoder_layers_4_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_4_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61717568)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_4_feed_forward2_linear1_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor input_277_cast_fp16 = silu(x = linear_44_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor encoder_layers_4_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_4_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63814784)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_feed_forward2_linear2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_1266_to_fp16 = const()[name = tensor("op_1266_to_fp16"), val = tensor(0x1p-1)]; + tensor var_1267_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1266_to_fp16)[name = tensor("op_1267_cast_fp16")]; + tensor input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1267_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor input_285_axes_0 = const()[name = tensor("input_285_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_out_weight_to_fp16 = const()[name = tensor("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65912000)))]; + tensor encoder_layers_4_norm_out_bias_to_fp16 = const()[name = tensor("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65913088)))]; + tensor input_285_cast_fp16 = layer_norm(axes = input_285_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor cache_21_begin_0 = const()[name = tensor("cache_21_begin_0"), val = tensor([5, 0, 0, 0])]; + tensor cache_21_end_0 = const()[name = tensor("cache_21_end_0"), val = tensor([6, 1, 70, 512])]; + tensor cache_21_end_mask_0 = const()[name = tensor("cache_21_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_21_squeeze_mask_0 = const()[name = tensor("cache_21_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_21_cast_fp16 = slice_by_index(begin = cache_21_begin_0, end = cache_21_end_0, end_mask = cache_21_end_mask_0, squeeze_mask = cache_21_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor("cache_21_cast_fp16")]; + tensor cache_23_begin_0 = const()[name = tensor("cache_23_begin_0"), val = tensor([5, 0, 0, 0])]; + tensor cache_23_end_0 = const()[name = tensor("cache_23_end_0"), val = tensor([6, 1, 512, 8])]; + tensor cache_23_end_mask_0 = const()[name = tensor("cache_23_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_23_squeeze_mask_0 = const()[name = tensor("cache_23_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_23_cast_fp16 = slice_by_index(begin = cache_23_begin_0, end = cache_23_end_0, end_mask = cache_23_end_mask_0, squeeze_mask = cache_23_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor("cache_23_cast_fp16")]; + tensor input_287_axes_0 = const()[name = tensor("input_287_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = tensor("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65914176)))]; + tensor encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = tensor("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65915264)))]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor encoder_layers_5_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_5_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65916352)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_5_feed_forward1_linear1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor input_291_cast_fp16 = silu(x = linear_46_cast_fp16)[name = tensor("input_291_cast_fp16")]; + tensor encoder_layers_5_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_5_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68013568)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_feed_forward1_linear2_weight_to_fp16, x = input_291_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor var_1301_to_fp16 = const()[name = tensor("op_1301_to_fp16"), val = tensor(0x1p-1)]; + tensor var_1302_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1301_to_fp16)[name = tensor("op_1302_cast_fp16")]; + tensor input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1302_cast_fp16)[name = tensor("input_297_cast_fp16")]; + tensor key_11_axes_0 = const()[name = tensor("key_11_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = tensor("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70110784)))]; + tensor encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = tensor("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70111872)))]; + tensor key_11_cast_fp16 = layer_norm(axes = key_11_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_297_cast_fp16)[name = tensor("key_11_cast_fp16")]; + tensor input_299_interleave_0 = const()[name = tensor("input_299_interleave_0"), val = tensor(false)]; + tensor input_299_cast_fp16 = concat(axis = var_42, interleave = input_299_interleave_0, values = (cache_21_cast_fp16, key_11_cast_fp16))[name = tensor("input_299_cast_fp16")]; + tensor var_1324_begin_0 = const()[name = tensor("op_1324_begin_0"), val = tensor([0, 4, 0])]; + tensor var_1324_end_0 = const()[name = tensor("op_1324_end_0"), val = tensor([1, 70, 512])]; + tensor var_1324_end_mask_0 = const()[name = tensor("op_1324_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1324_cast_fp16 = slice_by_index(begin = var_1324_begin_0, end = var_1324_end_0, end_mask = var_1324_end_mask_0, x = cache_21_cast_fp16)[name = tensor("op_1324_cast_fp16")]; + tensor var_1330_interleave_0 = const()[name = tensor("op_1330_interleave_0"), val = tensor(false)]; + tensor var_1330_cast_fp16 = concat(axis = var_42, interleave = var_1330_interleave_0, values = (var_1324_cast_fp16, key_11_cast_fp16))[name = tensor("op_1330_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_q_weight_to_fp16 = const()[name = tensor("encoder_layers_5_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70112960)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_q_weight_to_fp16, x = key_11_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, -1, 8, 64])]; + tensor q_31_cast_fp16 = reshape(shape = var_1334, x = linear_48_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_k_weight_to_fp16 = const()[name = tensor("encoder_layers_5_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70637312)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_k_weight_to_fp16, x = input_299_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_1338 = const()[name = tensor("op_1338"), val = tensor([1, -1, 8, 64])]; + tensor k_21_cast_fp16 = reshape(shape = var_1338, x = linear_49_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_v_weight_to_fp16 = const()[name = tensor("encoder_layers_5_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71161664)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_v_weight_to_fp16, x = input_299_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_1342 = const()[name = tensor("op_1342"), val = tensor([1, -1, 8, 64])]; + tensor v_11_cast_fp16 = reshape(shape = var_1342, x = linear_50_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor value_11_perm_0 = const()[name = tensor("value_11_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor encoder_layers_5_self_attn_pos_bias_u_to_fp16 = const()[name = tensor("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71686016)))]; + tensor var_1354_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = tensor("op_1354_cast_fp16")]; + tensor encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = tensor("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71687104)))]; + tensor var_1356_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = tensor("op_1356_cast_fp16")]; + tensor q_with_bias_v_11_perm_0 = const()[name = tensor("q_with_bias_v_11_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor x_137_transpose_x_0 = const()[name = tensor("x_137_transpose_x_0"), val = tensor(false)]; + tensor x_137_transpose_y_0 = const()[name = tensor("x_137_transpose_y_0"), val = tensor(false)]; + tensor var_1358_to_fp16 = const()[name = tensor("op_1358_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71688192)))]; + tensor q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1356_cast_fp16)[name = tensor("transpose_192")]; + tensor x_137_cast_fp16 = matmul(transpose_x = x_137_transpose_x_0, transpose_y = x_137_transpose_y_0, x = q_with_bias_v_11_cast_fp16, y = var_1358_to_fp16)[name = tensor("x_137_cast_fp16")]; + tensor x_139_pad_0 = const()[name = tensor("x_139_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor x_139_mode_0 = const()[name = tensor("x_139_mode_0"), val = tensor("constant")]; + tensor const_88_to_fp16 = const()[name = tensor("const_88_to_fp16"), val = tensor(0x0p+0)]; + tensor x_139_cast_fp16 = pad(constant_val = const_88_to_fp16, mode = x_139_mode_0, pad = x_139_pad_0, x = x_137_cast_fp16)[name = tensor("x_139_cast_fp16")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 8, -1, 4])]; + tensor x_141_cast_fp16 = reshape(shape = var_1366, x = x_139_cast_fp16)[name = tensor("x_141_cast_fp16")]; + tensor var_1370_begin_0 = const()[name = tensor("op_1370_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1370_end_0 = const()[name = tensor("op_1370_end_0"), val = tensor([1, 8, 148, 4])]; + tensor var_1370_end_mask_0 = const()[name = tensor("op_1370_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1370_cast_fp16 = slice_by_index(begin = var_1370_begin_0, end = var_1370_end_0, end_mask = var_1370_end_mask_0, x = x_141_cast_fp16)[name = tensor("op_1370_cast_fp16")]; + tensor var_1371 = const()[name = tensor("op_1371"), val = tensor([1, 8, 4, 147])]; + tensor matrix_bd_21_cast_fp16 = reshape(shape = var_1371, x = var_1370_cast_fp16)[name = tensor("matrix_bd_21_cast_fp16")]; + tensor matrix_ac_11_transpose_x_0 = const()[name = tensor("matrix_ac_11_transpose_x_0"), val = tensor(false)]; + tensor matrix_ac_11_transpose_y_0 = const()[name = tensor("matrix_ac_11_transpose_y_0"), val = tensor(false)]; + tensor transpose_61_perm_0 = const()[name = tensor("transpose_61_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_62_perm_0 = const()[name = tensor("transpose_62_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_62 = transpose(perm = transpose_62_perm_0, x = k_21_cast_fp16)[name = tensor("transpose_190")]; + tensor transpose_61 = transpose(perm = transpose_61_perm_0, x = var_1354_cast_fp16)[name = tensor("transpose_191")]; + tensor matrix_ac_11_cast_fp16 = matmul(transpose_x = matrix_ac_11_transpose_x_0, transpose_y = matrix_ac_11_transpose_y_0, x = transpose_61, y = transpose_62)[name = tensor("matrix_ac_11_cast_fp16")]; + tensor matrix_bd_23_begin_0 = const()[name = tensor("matrix_bd_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_23_end_0 = const()[name = tensor("matrix_bd_23_end_0"), val = tensor([1, 8, 4, 74])]; + tensor matrix_bd_23_end_mask_0 = const()[name = tensor("matrix_bd_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_23_cast_fp16 = slice_by_index(begin = matrix_bd_23_begin_0, end = matrix_bd_23_end_0, end_mask = matrix_bd_23_end_mask_0, x = matrix_bd_21_cast_fp16)[name = tensor("matrix_bd_23_cast_fp16")]; + tensor var_1380_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = tensor("op_1380_cast_fp16")]; + tensor _inversed_scores_21_y_0_to_fp16 = const()[name = tensor("_inversed_scores_21_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor _inversed_scores_21_cast_fp16 = mul(x = var_1380_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = tensor("_inversed_scores_21_cast_fp16")]; + tensor scores_23_cast_fp16 = select(a = var_19_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_3)[name = tensor("scores_23_cast_fp16")]; + tensor var_1386_cast_fp16 = softmax(axis = var_40, x = scores_23_cast_fp16)[name = tensor("op_1386_cast_fp16")]; + tensor input_301_cast_fp16 = select(a = var_18_to_fp16, b = var_1386_cast_fp16, cond = mask_3)[name = tensor("input_301_cast_fp16")]; + tensor x_143_transpose_x_0 = const()[name = tensor("x_143_transpose_x_0"), val = tensor(false)]; + tensor x_143_transpose_y_0 = const()[name = tensor("x_143_transpose_y_0"), val = tensor(false)]; + tensor value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = v_11_cast_fp16)[name = tensor("transpose_193")]; + tensor x_143_cast_fp16 = matmul(transpose_x = x_143_transpose_x_0, transpose_y = x_143_transpose_y_0, x = input_301_cast_fp16, y = value_11_cast_fp16)[name = tensor("x_143_cast_fp16")]; + tensor var_1390_perm_0 = const()[name = tensor("op_1390_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1391 = const()[name = tensor("op_1391"), val = tensor([1, -1, 512])]; + tensor var_1390_cast_fp16 = transpose(perm = var_1390_perm_0, x = x_143_cast_fp16)[name = tensor("transpose_189")]; + tensor input_303_cast_fp16 = reshape(shape = var_1391, x = var_1390_cast_fp16)[name = tensor("input_303_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_out_weight_to_fp16 = const()[name = tensor("encoder_layers_5_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71838784)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_out_weight_to_fp16, x = input_303_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_307_cast_fp16 = add(x = input_297_cast_fp16, y = linear_52_cast_fp16)[name = tensor("input_307_cast_fp16")]; + tensor x_147_axes_0 = const()[name = tensor("x_147_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72363136)))]; + tensor encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = tensor("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72364224)))]; + tensor x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input_307_cast_fp16)[name = tensor("x_147_cast_fp16")]; + tensor input_309_perm_0 = const()[name = tensor("input_309_perm_0"), val = tensor([0, 2, 1])]; + tensor input_311_pad_type_0 = const()[name = tensor("input_311_pad_type_0"), val = tensor("valid")]; + tensor input_311_strides_0 = const()[name = tensor("input_311_strides_0"), val = tensor([1])]; + tensor input_311_pad_0 = const()[name = tensor("input_311_pad_0"), val = tensor([0, 0])]; + tensor input_311_dilations_0 = const()[name = tensor("input_311_dilations_0"), val = tensor([1])]; + tensor input_311_groups_0 = const()[name = tensor("input_311_groups_0"), val = tensor(1)]; + tensor encoder_layers_5_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72365312)))]; + tensor input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = tensor("transpose_188")]; + tensor input_311_cast_fp16 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = encoder_layers_5_conv_pointwise_conv1_weight_to_fp16, x = input_309_cast_fp16)[name = tensor("input_311_cast_fp16")]; + tensor x_149_split_num_splits_0 = const()[name = tensor("x_149_split_num_splits_0"), val = tensor(2)]; + tensor x_149_split_axis_0 = const()[name = tensor("x_149_split_axis_0"), val = tensor(1)]; + tensor x_149_split_cast_fp16_0, tensor x_149_split_cast_fp16_1 = split(axis = x_149_split_axis_0, num_splits = x_149_split_num_splits_0, x = input_311_cast_fp16)[name = tensor("x_149_split_cast_fp16")]; + tensor x_149_split_1_sigmoid_cast_fp16 = sigmoid(x = x_149_split_cast_fp16_1)[name = tensor("x_149_split_1_sigmoid_cast_fp16")]; + tensor x_149_cast_fp16 = mul(x = x_149_split_cast_fp16_0, y = x_149_split_1_sigmoid_cast_fp16)[name = tensor("x_149_cast_fp16")]; + tensor input_313_cast_fp16 = select(a = var_18_to_fp16, b = x_149_cast_fp16, cond = var_396)[name = tensor("input_313_cast_fp16")]; + tensor new_x_23_interleave_0 = const()[name = tensor("new_x_23_interleave_0"), val = tensor(false)]; + tensor new_x_23_cast_fp16 = concat(axis = var_40, interleave = new_x_23_interleave_0, values = (cache_23_cast_fp16, input_313_cast_fp16))[name = tensor("new_x_23_cast_fp16")]; + tensor var_1429_begin_0 = const()[name = tensor("op_1429_begin_0"), val = tensor([0, 0, 4])]; + tensor var_1429_end_0 = const()[name = tensor("op_1429_end_0"), val = tensor([1, 512, 12])]; + tensor var_1429_end_mask_0 = const()[name = tensor("op_1429_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1429_cast_fp16 = slice_by_index(begin = var_1429_begin_0, end = var_1429_end_0, end_mask = var_1429_end_mask_0, x = new_x_23_cast_fp16)[name = tensor("op_1429_cast_fp16")]; + tensor x_151_pad_type_0 = const()[name = tensor("x_151_pad_type_0"), val = tensor("valid")]; + tensor x_151_groups_0 = const()[name = tensor("x_151_groups_0"), val = tensor(512)]; + tensor x_151_strides_0 = const()[name = tensor("x_151_strides_0"), val = tensor([1])]; + tensor x_151_pad_0 = const()[name = tensor("x_151_pad_0"), val = tensor([0, 0])]; + tensor x_151_dilations_0 = const()[name = tensor("x_151_dilations_0"), val = tensor([1])]; + tensor encoder_layers_5_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_5_conv_depthwise_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73413952)))]; + tensor x_151_cast_fp16 = conv(dilations = x_151_dilations_0, groups = x_151_groups_0, pad = x_151_pad_0, pad_type = x_151_pad_type_0, strides = x_151_strides_0, weight = encoder_layers_5_conv_depthwise_conv_weight_to_fp16, x = new_x_23_cast_fp16)[name = tensor("x_151_cast_fp16")]; + tensor input_315_perm_0 = const()[name = tensor("input_315_perm_0"), val = tensor([0, 2, 1])]; + tensor x_153_axes_0 = const()[name = tensor("x_153_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = tensor("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73423232)))]; + tensor encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = tensor("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73424320)))]; + tensor input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = tensor("transpose_187")]; + tensor x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input_315_cast_fp16)[name = tensor("x_153_cast_fp16")]; + tensor input_317_perm_0 = const()[name = tensor("input_317_perm_0"), val = tensor([0, 2, 1])]; + tensor input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = tensor("transpose_186")]; + tensor input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = tensor("input_319_cast_fp16")]; + tensor x_155_pad_type_0 = const()[name = tensor("x_155_pad_type_0"), val = tensor("valid")]; + tensor x_155_strides_0 = const()[name = tensor("x_155_strides_0"), val = tensor([1])]; + tensor x_155_pad_0 = const()[name = tensor("x_155_pad_0"), val = tensor([0, 0])]; + tensor x_155_dilations_0 = const()[name = tensor("x_155_dilations_0"), val = tensor([1])]; + tensor x_155_groups_0 = const()[name = tensor("x_155_groups_0"), val = tensor(1)]; + tensor encoder_layers_5_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73425408)))]; + tensor x_155_cast_fp16 = conv(dilations = x_155_dilations_0, groups = x_155_groups_0, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = x_155_strides_0, weight = encoder_layers_5_conv_pointwise_conv2_weight_to_fp16, x = input_319_cast_fp16)[name = tensor("x_155_cast_fp16")]; + tensor input_321_perm_0 = const()[name = tensor("input_321_perm_0"), val = tensor([0, 2, 1])]; + tensor input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = tensor("transpose_185")]; + tensor input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = tensor("input_323_cast_fp16")]; + tensor input_325_axes_0 = const()[name = tensor("input_325_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = tensor("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73949760)))]; + tensor encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = tensor("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73950848)))]; + tensor input_325_cast_fp16 = layer_norm(axes = input_325_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input_323_cast_fp16)[name = tensor("input_325_cast_fp16")]; + tensor encoder_layers_5_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_5_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73951936)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_5_feed_forward2_linear1_weight_to_fp16, x = input_325_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor input_329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = tensor("input_329_cast_fp16")]; + tensor encoder_layers_5_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_5_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76049152)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_feed_forward2_linear2_weight_to_fp16, x = input_329_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor var_1470_to_fp16 = const()[name = tensor("op_1470_to_fp16"), val = tensor(0x1p-1)]; + tensor var_1471_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1470_to_fp16)[name = tensor("op_1471_cast_fp16")]; + tensor input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1471_cast_fp16)[name = tensor("input_335_cast_fp16")]; + tensor input_337_axes_0 = const()[name = tensor("input_337_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_out_weight_to_fp16 = const()[name = tensor("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78146368)))]; + tensor encoder_layers_5_norm_out_bias_to_fp16 = const()[name = tensor("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78147456)))]; + tensor input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input_335_cast_fp16)[name = tensor("input_337_cast_fp16")]; + tensor cache_25_begin_0 = const()[name = tensor("cache_25_begin_0"), val = tensor([6, 0, 0, 0])]; + tensor cache_25_end_0 = const()[name = tensor("cache_25_end_0"), val = tensor([7, 1, 70, 512])]; + tensor cache_25_end_mask_0 = const()[name = tensor("cache_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_25_squeeze_mask_0 = const()[name = tensor("cache_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_25_cast_fp16 = slice_by_index(begin = cache_25_begin_0, end = cache_25_end_0, end_mask = cache_25_end_mask_0, squeeze_mask = cache_25_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor("cache_25_cast_fp16")]; + tensor cache_27_begin_0 = const()[name = tensor("cache_27_begin_0"), val = tensor([6, 0, 0, 0])]; + tensor cache_27_end_0 = const()[name = tensor("cache_27_end_0"), val = tensor([7, 1, 512, 8])]; + tensor cache_27_end_mask_0 = const()[name = tensor("cache_27_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_27_squeeze_mask_0 = const()[name = tensor("cache_27_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_27_cast_fp16 = slice_by_index(begin = cache_27_begin_0, end = cache_27_end_0, end_mask = cache_27_end_mask_0, squeeze_mask = cache_27_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor("cache_27_cast_fp16")]; + tensor input_339_axes_0 = const()[name = tensor("input_339_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = tensor("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78148544)))]; + tensor encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = tensor("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78149632)))]; + tensor input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input_337_cast_fp16)[name = tensor("input_339_cast_fp16")]; + tensor encoder_layers_6_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_6_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78150720)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_6_feed_forward1_linear1_weight_to_fp16, x = input_339_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_343_cast_fp16 = silu(x = linear_55_cast_fp16)[name = tensor("input_343_cast_fp16")]; + tensor encoder_layers_6_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_6_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80247936)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_feed_forward1_linear2_weight_to_fp16, x = input_343_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor var_1505_to_fp16 = const()[name = tensor("op_1505_to_fp16"), val = tensor(0x1p-1)]; + tensor var_1506_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1505_to_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1506_cast_fp16)[name = tensor("input_349_cast_fp16")]; + tensor key_13_axes_0 = const()[name = tensor("key_13_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = tensor("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82345152)))]; + tensor encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = tensor("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82346240)))]; + tensor key_13_cast_fp16 = layer_norm(axes = key_13_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = tensor("key_13_cast_fp16")]; + tensor input_351_interleave_0 = const()[name = tensor("input_351_interleave_0"), val = tensor(false)]; + tensor input_351_cast_fp16 = concat(axis = var_42, interleave = input_351_interleave_0, values = (cache_25_cast_fp16, key_13_cast_fp16))[name = tensor("input_351_cast_fp16")]; + tensor var_1528_begin_0 = const()[name = tensor("op_1528_begin_0"), val = tensor([0, 4, 0])]; + tensor var_1528_end_0 = const()[name = tensor("op_1528_end_0"), val = tensor([1, 70, 512])]; + tensor var_1528_end_mask_0 = const()[name = tensor("op_1528_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1528_cast_fp16 = slice_by_index(begin = var_1528_begin_0, end = var_1528_end_0, end_mask = var_1528_end_mask_0, x = cache_25_cast_fp16)[name = tensor("op_1528_cast_fp16")]; + tensor var_1534_interleave_0 = const()[name = tensor("op_1534_interleave_0"), val = tensor(false)]; + tensor var_1534_cast_fp16 = concat(axis = var_42, interleave = var_1534_interleave_0, values = (var_1528_cast_fp16, key_13_cast_fp16))[name = tensor("op_1534_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_q_weight_to_fp16 = const()[name = tensor("encoder_layers_6_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82347328)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_q_weight_to_fp16, x = key_13_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_1538 = const()[name = tensor("op_1538"), val = tensor([1, -1, 8, 64])]; + tensor q_37_cast_fp16 = reshape(shape = var_1538, x = linear_57_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_k_weight_to_fp16 = const()[name = tensor("encoder_layers_6_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82871680)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_k_weight_to_fp16, x = input_351_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_1542 = const()[name = tensor("op_1542"), val = tensor([1, -1, 8, 64])]; + tensor k_25_cast_fp16 = reshape(shape = var_1542, x = linear_58_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_v_weight_to_fp16 = const()[name = tensor("encoder_layers_6_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83396032)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_v_weight_to_fp16, x = input_351_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor var_1546 = const()[name = tensor("op_1546"), val = tensor([1, -1, 8, 64])]; + tensor v_13_cast_fp16 = reshape(shape = var_1546, x = linear_59_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor value_13_perm_0 = const()[name = tensor("value_13_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor encoder_layers_6_self_attn_pos_bias_u_to_fp16 = const()[name = tensor("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83920384)))]; + tensor var_1558_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = tensor("op_1558_cast_fp16")]; + tensor encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = tensor("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83921472)))]; + tensor var_1560_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = tensor("op_1560_cast_fp16")]; + tensor q_with_bias_v_13_perm_0 = const()[name = tensor("q_with_bias_v_13_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor x_163_transpose_x_0 = const()[name = tensor("x_163_transpose_x_0"), val = tensor(false)]; + tensor x_163_transpose_y_0 = const()[name = tensor("x_163_transpose_y_0"), val = tensor(false)]; + tensor var_1562_to_fp16 = const()[name = tensor("op_1562_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83922560)))]; + tensor q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1560_cast_fp16)[name = tensor("transpose_183")]; + tensor x_163_cast_fp16 = matmul(transpose_x = x_163_transpose_x_0, transpose_y = x_163_transpose_y_0, x = q_with_bias_v_13_cast_fp16, y = var_1562_to_fp16)[name = tensor("x_163_cast_fp16")]; + tensor x_165_pad_0 = const()[name = tensor("x_165_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor x_165_mode_0 = const()[name = tensor("x_165_mode_0"), val = tensor("constant")]; + tensor const_101_to_fp16 = const()[name = tensor("const_101_to_fp16"), val = tensor(0x0p+0)]; + tensor x_165_cast_fp16 = pad(constant_val = const_101_to_fp16, mode = x_165_mode_0, pad = x_165_pad_0, x = x_163_cast_fp16)[name = tensor("x_165_cast_fp16")]; + tensor var_1570 = const()[name = tensor("op_1570"), val = tensor([1, 8, -1, 4])]; + tensor x_167_cast_fp16 = reshape(shape = var_1570, x = x_165_cast_fp16)[name = tensor("x_167_cast_fp16")]; + tensor var_1574_begin_0 = const()[name = tensor("op_1574_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1574_end_0 = const()[name = tensor("op_1574_end_0"), val = tensor([1, 8, 148, 4])]; + tensor var_1574_end_mask_0 = const()[name = tensor("op_1574_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1574_cast_fp16 = slice_by_index(begin = var_1574_begin_0, end = var_1574_end_0, end_mask = var_1574_end_mask_0, x = x_167_cast_fp16)[name = tensor("op_1574_cast_fp16")]; + tensor var_1575 = const()[name = tensor("op_1575"), val = tensor([1, 8, 4, 147])]; + tensor matrix_bd_25_cast_fp16 = reshape(shape = var_1575, x = var_1574_cast_fp16)[name = tensor("matrix_bd_25_cast_fp16")]; + tensor matrix_ac_13_transpose_x_0 = const()[name = tensor("matrix_ac_13_transpose_x_0"), val = tensor(false)]; + tensor matrix_ac_13_transpose_y_0 = const()[name = tensor("matrix_ac_13_transpose_y_0"), val = tensor(false)]; + tensor transpose_63_perm_0 = const()[name = tensor("transpose_63_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_64_perm_0 = const()[name = tensor("transpose_64_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_64 = transpose(perm = transpose_64_perm_0, x = k_25_cast_fp16)[name = tensor("transpose_181")]; + tensor transpose_63 = transpose(perm = transpose_63_perm_0, x = var_1558_cast_fp16)[name = tensor("transpose_182")]; + tensor matrix_ac_13_cast_fp16 = matmul(transpose_x = matrix_ac_13_transpose_x_0, transpose_y = matrix_ac_13_transpose_y_0, x = transpose_63, y = transpose_64)[name = tensor("matrix_ac_13_cast_fp16")]; + tensor matrix_bd_27_begin_0 = const()[name = tensor("matrix_bd_27_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_27_end_0 = const()[name = tensor("matrix_bd_27_end_0"), val = tensor([1, 8, 4, 74])]; + tensor matrix_bd_27_end_mask_0 = const()[name = tensor("matrix_bd_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_27_cast_fp16 = slice_by_index(begin = matrix_bd_27_begin_0, end = matrix_bd_27_end_0, end_mask = matrix_bd_27_end_mask_0, x = matrix_bd_25_cast_fp16)[name = tensor("matrix_bd_27_cast_fp16")]; + tensor var_1584_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = tensor("op_1584_cast_fp16")]; + tensor _inversed_scores_25_y_0_to_fp16 = const()[name = tensor("_inversed_scores_25_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor _inversed_scores_25_cast_fp16 = mul(x = var_1584_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = tensor("_inversed_scores_25_cast_fp16")]; + tensor scores_27_cast_fp16 = select(a = var_19_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_3)[name = tensor("scores_27_cast_fp16")]; + tensor var_1590_cast_fp16 = softmax(axis = var_40, x = scores_27_cast_fp16)[name = tensor("op_1590_cast_fp16")]; + tensor input_353_cast_fp16 = select(a = var_18_to_fp16, b = var_1590_cast_fp16, cond = mask_3)[name = tensor("input_353_cast_fp16")]; + tensor x_169_transpose_x_0 = const()[name = tensor("x_169_transpose_x_0"), val = tensor(false)]; + tensor x_169_transpose_y_0 = const()[name = tensor("x_169_transpose_y_0"), val = tensor(false)]; + tensor value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = v_13_cast_fp16)[name = tensor("transpose_184")]; + tensor x_169_cast_fp16 = matmul(transpose_x = x_169_transpose_x_0, transpose_y = x_169_transpose_y_0, x = input_353_cast_fp16, y = value_13_cast_fp16)[name = tensor("x_169_cast_fp16")]; + tensor var_1594_perm_0 = const()[name = tensor("op_1594_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1595 = const()[name = tensor("op_1595"), val = tensor([1, -1, 512])]; + tensor var_1594_cast_fp16 = transpose(perm = var_1594_perm_0, x = x_169_cast_fp16)[name = tensor("transpose_180")]; + tensor input_355_cast_fp16 = reshape(shape = var_1595, x = var_1594_cast_fp16)[name = tensor("input_355_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_out_weight_to_fp16 = const()[name = tensor("encoder_layers_6_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84073152)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_out_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor input_359_cast_fp16 = add(x = input_349_cast_fp16, y = linear_61_cast_fp16)[name = tensor("input_359_cast_fp16")]; + tensor x_173_axes_0 = const()[name = tensor("x_173_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84597504)))]; + tensor encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = tensor("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84598592)))]; + tensor x_173_cast_fp16 = layer_norm(axes = x_173_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input_359_cast_fp16)[name = tensor("x_173_cast_fp16")]; + tensor input_361_perm_0 = const()[name = tensor("input_361_perm_0"), val = tensor([0, 2, 1])]; + tensor input_363_pad_type_0 = const()[name = tensor("input_363_pad_type_0"), val = tensor("valid")]; + tensor input_363_strides_0 = const()[name = tensor("input_363_strides_0"), val = tensor([1])]; + tensor input_363_pad_0 = const()[name = tensor("input_363_pad_0"), val = tensor([0, 0])]; + tensor input_363_dilations_0 = const()[name = tensor("input_363_dilations_0"), val = tensor([1])]; + tensor input_363_groups_0 = const()[name = tensor("input_363_groups_0"), val = tensor(1)]; + tensor encoder_layers_6_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84599680)))]; + tensor input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = tensor("transpose_179")]; + tensor input_363_cast_fp16 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = encoder_layers_6_conv_pointwise_conv1_weight_to_fp16, x = input_361_cast_fp16)[name = tensor("input_363_cast_fp16")]; + tensor x_175_split_num_splits_0 = const()[name = tensor("x_175_split_num_splits_0"), val = tensor(2)]; + tensor x_175_split_axis_0 = const()[name = tensor("x_175_split_axis_0"), val = tensor(1)]; + tensor x_175_split_cast_fp16_0, tensor x_175_split_cast_fp16_1 = split(axis = x_175_split_axis_0, num_splits = x_175_split_num_splits_0, x = input_363_cast_fp16)[name = tensor("x_175_split_cast_fp16")]; + tensor x_175_split_1_sigmoid_cast_fp16 = sigmoid(x = x_175_split_cast_fp16_1)[name = tensor("x_175_split_1_sigmoid_cast_fp16")]; + tensor x_175_cast_fp16 = mul(x = x_175_split_cast_fp16_0, y = x_175_split_1_sigmoid_cast_fp16)[name = tensor("x_175_cast_fp16")]; + tensor input_365_cast_fp16 = select(a = var_18_to_fp16, b = x_175_cast_fp16, cond = var_396)[name = tensor("input_365_cast_fp16")]; + tensor new_x_27_interleave_0 = const()[name = tensor("new_x_27_interleave_0"), val = tensor(false)]; + tensor new_x_27_cast_fp16 = concat(axis = var_40, interleave = new_x_27_interleave_0, values = (cache_27_cast_fp16, input_365_cast_fp16))[name = tensor("new_x_27_cast_fp16")]; + tensor var_1633_begin_0 = const()[name = tensor("op_1633_begin_0"), val = tensor([0, 0, 4])]; + tensor var_1633_end_0 = const()[name = tensor("op_1633_end_0"), val = tensor([1, 512, 12])]; + tensor var_1633_end_mask_0 = const()[name = tensor("op_1633_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1633_cast_fp16 = slice_by_index(begin = var_1633_begin_0, end = var_1633_end_0, end_mask = var_1633_end_mask_0, x = new_x_27_cast_fp16)[name = tensor("op_1633_cast_fp16")]; + tensor x_177_pad_type_0 = const()[name = tensor("x_177_pad_type_0"), val = tensor("valid")]; + tensor x_177_groups_0 = const()[name = tensor("x_177_groups_0"), val = tensor(512)]; + tensor x_177_strides_0 = const()[name = tensor("x_177_strides_0"), val = tensor([1])]; + tensor x_177_pad_0 = const()[name = tensor("x_177_pad_0"), val = tensor([0, 0])]; + tensor x_177_dilations_0 = const()[name = tensor("x_177_dilations_0"), val = tensor([1])]; + tensor encoder_layers_6_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_6_conv_depthwise_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85648320)))]; + tensor x_177_cast_fp16 = conv(dilations = x_177_dilations_0, groups = x_177_groups_0, pad = x_177_pad_0, pad_type = x_177_pad_type_0, strides = x_177_strides_0, weight = encoder_layers_6_conv_depthwise_conv_weight_to_fp16, x = new_x_27_cast_fp16)[name = tensor("x_177_cast_fp16")]; + tensor input_367_perm_0 = const()[name = tensor("input_367_perm_0"), val = tensor([0, 2, 1])]; + tensor x_179_axes_0 = const()[name = tensor("x_179_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = tensor("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85657600)))]; + tensor encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = tensor("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85658688)))]; + tensor input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = tensor("transpose_178")]; + tensor x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input_367_cast_fp16)[name = tensor("x_179_cast_fp16")]; + tensor input_369_perm_0 = const()[name = tensor("input_369_perm_0"), val = tensor([0, 2, 1])]; + tensor input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = tensor("transpose_177")]; + tensor input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = tensor("input_371_cast_fp16")]; + tensor x_181_pad_type_0 = const()[name = tensor("x_181_pad_type_0"), val = tensor("valid")]; + tensor x_181_strides_0 = const()[name = tensor("x_181_strides_0"), val = tensor([1])]; + tensor x_181_pad_0 = const()[name = tensor("x_181_pad_0"), val = tensor([0, 0])]; + tensor x_181_dilations_0 = const()[name = tensor("x_181_dilations_0"), val = tensor([1])]; + tensor x_181_groups_0 = const()[name = tensor("x_181_groups_0"), val = tensor(1)]; + tensor encoder_layers_6_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85659776)))]; + tensor x_181_cast_fp16 = conv(dilations = x_181_dilations_0, groups = x_181_groups_0, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = x_181_strides_0, weight = encoder_layers_6_conv_pointwise_conv2_weight_to_fp16, x = input_371_cast_fp16)[name = tensor("x_181_cast_fp16")]; + tensor input_373_perm_0 = const()[name = tensor("input_373_perm_0"), val = tensor([0, 2, 1])]; + tensor input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = tensor("transpose_176")]; + tensor input_375_cast_fp16 = add(x = input_359_cast_fp16, y = input_373_cast_fp16)[name = tensor("input_375_cast_fp16")]; + tensor input_377_axes_0 = const()[name = tensor("input_377_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = tensor("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86184128)))]; + tensor encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = tensor("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86185216)))]; + tensor input_377_cast_fp16 = layer_norm(axes = input_377_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input_375_cast_fp16)[name = tensor("input_377_cast_fp16")]; + tensor encoder_layers_6_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_6_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86186304)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_6_feed_forward2_linear1_weight_to_fp16, x = input_377_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_381_cast_fp16 = silu(x = linear_62_cast_fp16)[name = tensor("input_381_cast_fp16")]; + tensor encoder_layers_6_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_6_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88283520)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_feed_forward2_linear2_weight_to_fp16, x = input_381_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor var_1674_to_fp16 = const()[name = tensor("op_1674_to_fp16"), val = tensor(0x1p-1)]; + tensor var_1675_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1674_to_fp16)[name = tensor("op_1675_cast_fp16")]; + tensor input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1675_cast_fp16)[name = tensor("input_387_cast_fp16")]; + tensor input_389_axes_0 = const()[name = tensor("input_389_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_out_weight_to_fp16 = const()[name = tensor("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90380736)))]; + tensor encoder_layers_6_norm_out_bias_to_fp16 = const()[name = tensor("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90381824)))]; + tensor input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = tensor("input_389_cast_fp16")]; + tensor cache_29_begin_0 = const()[name = tensor("cache_29_begin_0"), val = tensor([7, 0, 0, 0])]; + tensor cache_29_end_0 = const()[name = tensor("cache_29_end_0"), val = tensor([8, 1, 70, 512])]; + tensor cache_29_end_mask_0 = const()[name = tensor("cache_29_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_29_squeeze_mask_0 = const()[name = tensor("cache_29_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_29_cast_fp16 = slice_by_index(begin = cache_29_begin_0, end = cache_29_end_0, end_mask = cache_29_end_mask_0, squeeze_mask = cache_29_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor("cache_29_cast_fp16")]; + tensor cache_31_begin_0 = const()[name = tensor("cache_31_begin_0"), val = tensor([7, 0, 0, 0])]; + tensor cache_31_end_0 = const()[name = tensor("cache_31_end_0"), val = tensor([8, 1, 512, 8])]; + tensor cache_31_end_mask_0 = const()[name = tensor("cache_31_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_31_squeeze_mask_0 = const()[name = tensor("cache_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_31_cast_fp16 = slice_by_index(begin = cache_31_begin_0, end = cache_31_end_0, end_mask = cache_31_end_mask_0, squeeze_mask = cache_31_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor("cache_31_cast_fp16")]; + tensor input_391_axes_0 = const()[name = tensor("input_391_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = tensor("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90382912)))]; + tensor encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = tensor("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90384000)))]; + tensor input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = tensor("input_391_cast_fp16")]; + tensor encoder_layers_7_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_7_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90385088)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_7_feed_forward1_linear1_weight_to_fp16, x = input_391_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor input_395_cast_fp16 = silu(x = linear_64_cast_fp16)[name = tensor("input_395_cast_fp16")]; + tensor encoder_layers_7_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_7_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92482304)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_feed_forward1_linear2_weight_to_fp16, x = input_395_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_1709_to_fp16 = const()[name = tensor("op_1709_to_fp16"), val = tensor(0x1p-1)]; + tensor var_1710_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1709_to_fp16)[name = tensor("op_1710_cast_fp16")]; + tensor input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1710_cast_fp16)[name = tensor("input_401_cast_fp16")]; + tensor key_15_axes_0 = const()[name = tensor("key_15_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = tensor("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94579520)))]; + tensor encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = tensor("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94580608)))]; + tensor key_15_cast_fp16 = layer_norm(axes = key_15_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_401_cast_fp16)[name = tensor("key_15_cast_fp16")]; + tensor input_403_interleave_0 = const()[name = tensor("input_403_interleave_0"), val = tensor(false)]; + tensor input_403_cast_fp16 = concat(axis = var_42, interleave = input_403_interleave_0, values = (cache_29_cast_fp16, key_15_cast_fp16))[name = tensor("input_403_cast_fp16")]; + tensor var_1732_begin_0 = const()[name = tensor("op_1732_begin_0"), val = tensor([0, 4, 0])]; + tensor var_1732_end_0 = const()[name = tensor("op_1732_end_0"), val = tensor([1, 70, 512])]; + tensor var_1732_end_mask_0 = const()[name = tensor("op_1732_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1732_cast_fp16 = slice_by_index(begin = var_1732_begin_0, end = var_1732_end_0, end_mask = var_1732_end_mask_0, x = cache_29_cast_fp16)[name = tensor("op_1732_cast_fp16")]; + tensor var_1738_interleave_0 = const()[name = tensor("op_1738_interleave_0"), val = tensor(false)]; + tensor var_1738_cast_fp16 = concat(axis = var_42, interleave = var_1738_interleave_0, values = (var_1732_cast_fp16, key_15_cast_fp16))[name = tensor("op_1738_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_q_weight_to_fp16 = const()[name = tensor("encoder_layers_7_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94581696)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_q_weight_to_fp16, x = key_15_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_1742 = const()[name = tensor("op_1742"), val = tensor([1, -1, 8, 64])]; + tensor q_43_cast_fp16 = reshape(shape = var_1742, x = linear_66_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_k_weight_to_fp16 = const()[name = tensor("encoder_layers_7_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(95106048)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_k_weight_to_fp16, x = input_403_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor var_1746 = const()[name = tensor("op_1746"), val = tensor([1, -1, 8, 64])]; + tensor k_29_cast_fp16 = reshape(shape = var_1746, x = linear_67_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_v_weight_to_fp16 = const()[name = tensor("encoder_layers_7_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(95630400)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_v_weight_to_fp16, x = input_403_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor var_1750 = const()[name = tensor("op_1750"), val = tensor([1, -1, 8, 64])]; + tensor v_15_cast_fp16 = reshape(shape = var_1750, x = linear_68_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor value_15_perm_0 = const()[name = tensor("value_15_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor encoder_layers_7_self_attn_pos_bias_u_to_fp16 = const()[name = tensor("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96154752)))]; + tensor var_1762_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = tensor("op_1762_cast_fp16")]; + tensor encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = tensor("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96155840)))]; + tensor var_1764_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = tensor("op_1764_cast_fp16")]; + tensor q_with_bias_v_15_perm_0 = const()[name = tensor("q_with_bias_v_15_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor x_189_transpose_x_0 = const()[name = tensor("x_189_transpose_x_0"), val = tensor(false)]; + tensor x_189_transpose_y_0 = const()[name = tensor("x_189_transpose_y_0"), val = tensor(false)]; + tensor var_1766_to_fp16 = const()[name = tensor("op_1766_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96156928)))]; + tensor q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_1764_cast_fp16)[name = tensor("transpose_174")]; + tensor x_189_cast_fp16 = matmul(transpose_x = x_189_transpose_x_0, transpose_y = x_189_transpose_y_0, x = q_with_bias_v_15_cast_fp16, y = var_1766_to_fp16)[name = tensor("x_189_cast_fp16")]; + tensor x_191_pad_0 = const()[name = tensor("x_191_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor x_191_mode_0 = const()[name = tensor("x_191_mode_0"), val = tensor("constant")]; + tensor const_114_to_fp16 = const()[name = tensor("const_114_to_fp16"), val = tensor(0x0p+0)]; + tensor x_191_cast_fp16 = pad(constant_val = const_114_to_fp16, mode = x_191_mode_0, pad = x_191_pad_0, x = x_189_cast_fp16)[name = tensor("x_191_cast_fp16")]; + tensor var_1774 = const()[name = tensor("op_1774"), val = tensor([1, 8, -1, 4])]; + tensor x_193_cast_fp16 = reshape(shape = var_1774, x = x_191_cast_fp16)[name = tensor("x_193_cast_fp16")]; + tensor var_1778_begin_0 = const()[name = tensor("op_1778_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1778_end_0 = const()[name = tensor("op_1778_end_0"), val = tensor([1, 8, 148, 4])]; + tensor var_1778_end_mask_0 = const()[name = tensor("op_1778_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1778_cast_fp16 = slice_by_index(begin = var_1778_begin_0, end = var_1778_end_0, end_mask = var_1778_end_mask_0, x = x_193_cast_fp16)[name = tensor("op_1778_cast_fp16")]; + tensor var_1779 = const()[name = tensor("op_1779"), val = tensor([1, 8, 4, 147])]; + tensor matrix_bd_29_cast_fp16 = reshape(shape = var_1779, x = var_1778_cast_fp16)[name = tensor("matrix_bd_29_cast_fp16")]; + tensor matrix_ac_15_transpose_x_0 = const()[name = tensor("matrix_ac_15_transpose_x_0"), val = tensor(false)]; + tensor matrix_ac_15_transpose_y_0 = const()[name = tensor("matrix_ac_15_transpose_y_0"), val = tensor(false)]; + tensor transpose_65_perm_0 = const()[name = tensor("transpose_65_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_66_perm_0 = const()[name = tensor("transpose_66_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_66 = transpose(perm = transpose_66_perm_0, x = k_29_cast_fp16)[name = tensor("transpose_172")]; + tensor transpose_65 = transpose(perm = transpose_65_perm_0, x = var_1762_cast_fp16)[name = tensor("transpose_173")]; + tensor matrix_ac_15_cast_fp16 = matmul(transpose_x = matrix_ac_15_transpose_x_0, transpose_y = matrix_ac_15_transpose_y_0, x = transpose_65, y = transpose_66)[name = tensor("matrix_ac_15_cast_fp16")]; + tensor matrix_bd_31_begin_0 = const()[name = tensor("matrix_bd_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_31_end_0 = const()[name = tensor("matrix_bd_31_end_0"), val = tensor([1, 8, 4, 74])]; + tensor matrix_bd_31_end_mask_0 = const()[name = tensor("matrix_bd_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_31_cast_fp16 = slice_by_index(begin = matrix_bd_31_begin_0, end = matrix_bd_31_end_0, end_mask = matrix_bd_31_end_mask_0, x = matrix_bd_29_cast_fp16)[name = tensor("matrix_bd_31_cast_fp16")]; + tensor var_1788_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = tensor("op_1788_cast_fp16")]; + tensor _inversed_scores_29_y_0_to_fp16 = const()[name = tensor("_inversed_scores_29_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor _inversed_scores_29_cast_fp16 = mul(x = var_1788_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = tensor("_inversed_scores_29_cast_fp16")]; + tensor scores_31_cast_fp16 = select(a = var_19_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_3)[name = tensor("scores_31_cast_fp16")]; + tensor var_1794_cast_fp16 = softmax(axis = var_40, x = scores_31_cast_fp16)[name = tensor("op_1794_cast_fp16")]; + tensor input_405_cast_fp16 = select(a = var_18_to_fp16, b = var_1794_cast_fp16, cond = mask_3)[name = tensor("input_405_cast_fp16")]; + tensor x_195_transpose_x_0 = const()[name = tensor("x_195_transpose_x_0"), val = tensor(false)]; + tensor x_195_transpose_y_0 = const()[name = tensor("x_195_transpose_y_0"), val = tensor(false)]; + tensor value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = v_15_cast_fp16)[name = tensor("transpose_175")]; + tensor x_195_cast_fp16 = matmul(transpose_x = x_195_transpose_x_0, transpose_y = x_195_transpose_y_0, x = input_405_cast_fp16, y = value_15_cast_fp16)[name = tensor("x_195_cast_fp16")]; + tensor var_1798_perm_0 = const()[name = tensor("op_1798_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1799 = const()[name = tensor("op_1799"), val = tensor([1, -1, 512])]; + tensor var_1798_cast_fp16 = transpose(perm = var_1798_perm_0, x = x_195_cast_fp16)[name = tensor("transpose_171")]; + tensor input_407_cast_fp16 = reshape(shape = var_1799, x = var_1798_cast_fp16)[name = tensor("input_407_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_out_weight_to_fp16 = const()[name = tensor("encoder_layers_7_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96307520)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_out_weight_to_fp16, x = input_407_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor input_411_cast_fp16 = add(x = input_401_cast_fp16, y = linear_70_cast_fp16)[name = tensor("input_411_cast_fp16")]; + tensor x_199_axes_0 = const()[name = tensor("x_199_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96831872)))]; + tensor encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = tensor("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96832960)))]; + tensor x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input_411_cast_fp16)[name = tensor("x_199_cast_fp16")]; + tensor input_413_perm_0 = const()[name = tensor("input_413_perm_0"), val = tensor([0, 2, 1])]; + tensor input_415_pad_type_0 = const()[name = tensor("input_415_pad_type_0"), val = tensor("valid")]; + tensor input_415_strides_0 = const()[name = tensor("input_415_strides_0"), val = tensor([1])]; + tensor input_415_pad_0 = const()[name = tensor("input_415_pad_0"), val = tensor([0, 0])]; + tensor input_415_dilations_0 = const()[name = tensor("input_415_dilations_0"), val = tensor([1])]; + tensor input_415_groups_0 = const()[name = tensor("input_415_groups_0"), val = tensor(1)]; + tensor encoder_layers_7_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96834048)))]; + tensor input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = tensor("transpose_170")]; + tensor input_415_cast_fp16 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = encoder_layers_7_conv_pointwise_conv1_weight_to_fp16, x = input_413_cast_fp16)[name = tensor("input_415_cast_fp16")]; + tensor x_201_split_num_splits_0 = const()[name = tensor("x_201_split_num_splits_0"), val = tensor(2)]; + tensor x_201_split_axis_0 = const()[name = tensor("x_201_split_axis_0"), val = tensor(1)]; + tensor x_201_split_cast_fp16_0, tensor x_201_split_cast_fp16_1 = split(axis = x_201_split_axis_0, num_splits = x_201_split_num_splits_0, x = input_415_cast_fp16)[name = tensor("x_201_split_cast_fp16")]; + tensor x_201_split_1_sigmoid_cast_fp16 = sigmoid(x = x_201_split_cast_fp16_1)[name = tensor("x_201_split_1_sigmoid_cast_fp16")]; + tensor x_201_cast_fp16 = mul(x = x_201_split_cast_fp16_0, y = x_201_split_1_sigmoid_cast_fp16)[name = tensor("x_201_cast_fp16")]; + tensor input_417_cast_fp16 = select(a = var_18_to_fp16, b = x_201_cast_fp16, cond = var_396)[name = tensor("input_417_cast_fp16")]; + tensor new_x_31_interleave_0 = const()[name = tensor("new_x_31_interleave_0"), val = tensor(false)]; + tensor new_x_31_cast_fp16 = concat(axis = var_40, interleave = new_x_31_interleave_0, values = (cache_31_cast_fp16, input_417_cast_fp16))[name = tensor("new_x_31_cast_fp16")]; + tensor var_1837_begin_0 = const()[name = tensor("op_1837_begin_0"), val = tensor([0, 0, 4])]; + tensor var_1837_end_0 = const()[name = tensor("op_1837_end_0"), val = tensor([1, 512, 12])]; + tensor var_1837_end_mask_0 = const()[name = tensor("op_1837_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1837_cast_fp16 = slice_by_index(begin = var_1837_begin_0, end = var_1837_end_0, end_mask = var_1837_end_mask_0, x = new_x_31_cast_fp16)[name = tensor("op_1837_cast_fp16")]; + tensor x_203_pad_type_0 = const()[name = tensor("x_203_pad_type_0"), val = tensor("valid")]; + tensor x_203_groups_0 = const()[name = tensor("x_203_groups_0"), val = tensor(512)]; + tensor x_203_strides_0 = const()[name = tensor("x_203_strides_0"), val = tensor([1])]; + tensor x_203_pad_0 = const()[name = tensor("x_203_pad_0"), val = tensor([0, 0])]; + tensor x_203_dilations_0 = const()[name = tensor("x_203_dilations_0"), val = tensor([1])]; + tensor encoder_layers_7_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_7_conv_depthwise_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(97882688)))]; + tensor x_203_cast_fp16 = conv(dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = encoder_layers_7_conv_depthwise_conv_weight_to_fp16, x = new_x_31_cast_fp16)[name = tensor("x_203_cast_fp16")]; + tensor input_419_perm_0 = const()[name = tensor("input_419_perm_0"), val = tensor([0, 2, 1])]; + tensor x_205_axes_0 = const()[name = tensor("x_205_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = tensor("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(97891968)))]; + tensor encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = tensor("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(97893056)))]; + tensor input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = tensor("transpose_169")]; + tensor x_205_cast_fp16 = layer_norm(axes = x_205_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input_419_cast_fp16)[name = tensor("x_205_cast_fp16")]; + tensor input_421_perm_0 = const()[name = tensor("input_421_perm_0"), val = tensor([0, 2, 1])]; + tensor input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = tensor("transpose_168")]; + tensor input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = tensor("input_423_cast_fp16")]; + tensor x_207_pad_type_0 = const()[name = tensor("x_207_pad_type_0"), val = tensor("valid")]; + tensor x_207_strides_0 = const()[name = tensor("x_207_strides_0"), val = tensor([1])]; + tensor x_207_pad_0 = const()[name = tensor("x_207_pad_0"), val = tensor([0, 0])]; + tensor x_207_dilations_0 = const()[name = tensor("x_207_dilations_0"), val = tensor([1])]; + tensor x_207_groups_0 = const()[name = tensor("x_207_groups_0"), val = tensor(1)]; + tensor encoder_layers_7_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(97894144)))]; + tensor x_207_cast_fp16 = conv(dilations = x_207_dilations_0, groups = x_207_groups_0, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = x_207_strides_0, weight = encoder_layers_7_conv_pointwise_conv2_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("x_207_cast_fp16")]; + tensor input_425_perm_0 = const()[name = tensor("input_425_perm_0"), val = tensor([0, 2, 1])]; + tensor input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = tensor("transpose_167")]; + tensor input_427_cast_fp16 = add(x = input_411_cast_fp16, y = input_425_cast_fp16)[name = tensor("input_427_cast_fp16")]; + tensor input_429_axes_0 = const()[name = tensor("input_429_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = tensor("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(98418496)))]; + tensor encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = tensor("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(98419584)))]; + tensor input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input_427_cast_fp16)[name = tensor("input_429_cast_fp16")]; + tensor encoder_layers_7_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_7_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(98420672)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_7_feed_forward2_linear1_weight_to_fp16, x = input_429_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_433_cast_fp16 = silu(x = linear_71_cast_fp16)[name = tensor("input_433_cast_fp16")]; + tensor encoder_layers_7_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_7_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100517888)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_feed_forward2_linear2_weight_to_fp16, x = input_433_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_1878_to_fp16 = const()[name = tensor("op_1878_to_fp16"), val = tensor(0x1p-1)]; + tensor var_1879_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_1878_to_fp16)[name = tensor("op_1879_cast_fp16")]; + tensor input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_1879_cast_fp16)[name = tensor("input_439_cast_fp16")]; + tensor input_441_axes_0 = const()[name = tensor("input_441_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_out_weight_to_fp16 = const()[name = tensor("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102615104)))]; + tensor encoder_layers_7_norm_out_bias_to_fp16 = const()[name = tensor("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102616192)))]; + tensor input_441_cast_fp16 = layer_norm(axes = input_441_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input_439_cast_fp16)[name = tensor("input_441_cast_fp16")]; + tensor cache_33_begin_0 = const()[name = tensor("cache_33_begin_0"), val = tensor([8, 0, 0, 0])]; + tensor cache_33_end_0 = const()[name = tensor("cache_33_end_0"), val = tensor([9, 1, 70, 512])]; + tensor cache_33_end_mask_0 = const()[name = tensor("cache_33_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_33_squeeze_mask_0 = const()[name = tensor("cache_33_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_33_cast_fp16 = slice_by_index(begin = cache_33_begin_0, end = cache_33_end_0, end_mask = cache_33_end_mask_0, squeeze_mask = cache_33_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor("cache_33_cast_fp16")]; + tensor cache_35_begin_0 = const()[name = tensor("cache_35_begin_0"), val = tensor([8, 0, 0, 0])]; + tensor cache_35_end_0 = const()[name = tensor("cache_35_end_0"), val = tensor([9, 1, 512, 8])]; + tensor cache_35_end_mask_0 = const()[name = tensor("cache_35_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_35_squeeze_mask_0 = const()[name = tensor("cache_35_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_35_cast_fp16 = slice_by_index(begin = cache_35_begin_0, end = cache_35_end_0, end_mask = cache_35_end_mask_0, squeeze_mask = cache_35_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor("cache_35_cast_fp16")]; + tensor input_443_axes_0 = const()[name = tensor("input_443_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = tensor("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102617280)))]; + tensor encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = tensor("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102618368)))]; + tensor input_443_cast_fp16 = layer_norm(axes = input_443_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input_441_cast_fp16)[name = tensor("input_443_cast_fp16")]; + tensor encoder_layers_8_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_8_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102619456)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_8_feed_forward1_linear1_weight_to_fp16, x = input_443_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor input_447_cast_fp16 = silu(x = linear_73_cast_fp16)[name = tensor("input_447_cast_fp16")]; + tensor encoder_layers_8_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_8_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(104716672)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_feed_forward1_linear2_weight_to_fp16, x = input_447_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor var_1913_to_fp16 = const()[name = tensor("op_1913_to_fp16"), val = tensor(0x1p-1)]; + tensor var_1914_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_1913_to_fp16)[name = tensor("op_1914_cast_fp16")]; + tensor input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_1914_cast_fp16)[name = tensor("input_453_cast_fp16")]; + tensor key_17_axes_0 = const()[name = tensor("key_17_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = tensor("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106813888)))]; + tensor encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = tensor("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106814976)))]; + tensor key_17_cast_fp16 = layer_norm(axes = key_17_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_453_cast_fp16)[name = tensor("key_17_cast_fp16")]; + tensor input_455_interleave_0 = const()[name = tensor("input_455_interleave_0"), val = tensor(false)]; + tensor input_455_cast_fp16 = concat(axis = var_42, interleave = input_455_interleave_0, values = (cache_33_cast_fp16, key_17_cast_fp16))[name = tensor("input_455_cast_fp16")]; + tensor var_1936_begin_0 = const()[name = tensor("op_1936_begin_0"), val = tensor([0, 4, 0])]; + tensor var_1936_end_0 = const()[name = tensor("op_1936_end_0"), val = tensor([1, 70, 512])]; + tensor var_1936_end_mask_0 = const()[name = tensor("op_1936_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1936_cast_fp16 = slice_by_index(begin = var_1936_begin_0, end = var_1936_end_0, end_mask = var_1936_end_mask_0, x = cache_33_cast_fp16)[name = tensor("op_1936_cast_fp16")]; + tensor var_1942_interleave_0 = const()[name = tensor("op_1942_interleave_0"), val = tensor(false)]; + tensor var_1942_cast_fp16 = concat(axis = var_42, interleave = var_1942_interleave_0, values = (var_1936_cast_fp16, key_17_cast_fp16))[name = tensor("op_1942_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_q_weight_to_fp16 = const()[name = tensor("encoder_layers_8_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106816064)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_q_weight_to_fp16, x = key_17_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor var_1946 = const()[name = tensor("op_1946"), val = tensor([1, -1, 8, 64])]; + tensor q_49_cast_fp16 = reshape(shape = var_1946, x = linear_75_cast_fp16)[name = tensor("q_49_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_k_weight_to_fp16 = const()[name = tensor("encoder_layers_8_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107340416)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_k_weight_to_fp16, x = input_455_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_1950 = const()[name = tensor("op_1950"), val = tensor([1, -1, 8, 64])]; + tensor k_33_cast_fp16 = reshape(shape = var_1950, x = linear_76_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_v_weight_to_fp16 = const()[name = tensor("encoder_layers_8_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107864768)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_v_weight_to_fp16, x = input_455_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_1954 = const()[name = tensor("op_1954"), val = tensor([1, -1, 8, 64])]; + tensor v_17_cast_fp16 = reshape(shape = var_1954, x = linear_77_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor value_17_perm_0 = const()[name = tensor("value_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor encoder_layers_8_self_attn_pos_bias_u_to_fp16 = const()[name = tensor("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108389120)))]; + tensor var_1966_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = tensor("op_1966_cast_fp16")]; + tensor encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = tensor("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108390208)))]; + tensor var_1968_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = tensor("op_1968_cast_fp16")]; + tensor q_with_bias_v_17_perm_0 = const()[name = tensor("q_with_bias_v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor x_215_transpose_x_0 = const()[name = tensor("x_215_transpose_x_0"), val = tensor(false)]; + tensor x_215_transpose_y_0 = const()[name = tensor("x_215_transpose_y_0"), val = tensor(false)]; + tensor var_1970_to_fp16 = const()[name = tensor("op_1970_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108391296)))]; + tensor q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_1968_cast_fp16)[name = tensor("transpose_165")]; + tensor x_215_cast_fp16 = matmul(transpose_x = x_215_transpose_x_0, transpose_y = x_215_transpose_y_0, x = q_with_bias_v_17_cast_fp16, y = var_1970_to_fp16)[name = tensor("x_215_cast_fp16")]; + tensor x_217_pad_0 = const()[name = tensor("x_217_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor x_217_mode_0 = const()[name = tensor("x_217_mode_0"), val = tensor("constant")]; + tensor const_127_to_fp16 = const()[name = tensor("const_127_to_fp16"), val = tensor(0x0p+0)]; + tensor x_217_cast_fp16 = pad(constant_val = const_127_to_fp16, mode = x_217_mode_0, pad = x_217_pad_0, x = x_215_cast_fp16)[name = tensor("x_217_cast_fp16")]; + tensor var_1978 = const()[name = tensor("op_1978"), val = tensor([1, 8, -1, 4])]; + tensor x_219_cast_fp16 = reshape(shape = var_1978, x = x_217_cast_fp16)[name = tensor("x_219_cast_fp16")]; + tensor var_1982_begin_0 = const()[name = tensor("op_1982_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1982_end_0 = const()[name = tensor("op_1982_end_0"), val = tensor([1, 8, 148, 4])]; + tensor var_1982_end_mask_0 = const()[name = tensor("op_1982_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1982_cast_fp16 = slice_by_index(begin = var_1982_begin_0, end = var_1982_end_0, end_mask = var_1982_end_mask_0, x = x_219_cast_fp16)[name = tensor("op_1982_cast_fp16")]; + tensor var_1983 = const()[name = tensor("op_1983"), val = tensor([1, 8, 4, 147])]; + tensor matrix_bd_33_cast_fp16 = reshape(shape = var_1983, x = var_1982_cast_fp16)[name = tensor("matrix_bd_33_cast_fp16")]; + tensor matrix_ac_17_transpose_x_0 = const()[name = tensor("matrix_ac_17_transpose_x_0"), val = tensor(false)]; + tensor matrix_ac_17_transpose_y_0 = const()[name = tensor("matrix_ac_17_transpose_y_0"), val = tensor(false)]; + tensor transpose_67_perm_0 = const()[name = tensor("transpose_67_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_68_perm_0 = const()[name = tensor("transpose_68_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_68 = transpose(perm = transpose_68_perm_0, x = k_33_cast_fp16)[name = tensor("transpose_163")]; + tensor transpose_67 = transpose(perm = transpose_67_perm_0, x = var_1966_cast_fp16)[name = tensor("transpose_164")]; + tensor matrix_ac_17_cast_fp16 = matmul(transpose_x = matrix_ac_17_transpose_x_0, transpose_y = matrix_ac_17_transpose_y_0, x = transpose_67, y = transpose_68)[name = tensor("matrix_ac_17_cast_fp16")]; + tensor matrix_bd_35_begin_0 = const()[name = tensor("matrix_bd_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_35_end_0 = const()[name = tensor("matrix_bd_35_end_0"), val = tensor([1, 8, 4, 74])]; + tensor matrix_bd_35_end_mask_0 = const()[name = tensor("matrix_bd_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_35_cast_fp16 = slice_by_index(begin = matrix_bd_35_begin_0, end = matrix_bd_35_end_0, end_mask = matrix_bd_35_end_mask_0, x = matrix_bd_33_cast_fp16)[name = tensor("matrix_bd_35_cast_fp16")]; + tensor var_1992_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = tensor("op_1992_cast_fp16")]; + tensor _inversed_scores_33_y_0_to_fp16 = const()[name = tensor("_inversed_scores_33_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor _inversed_scores_33_cast_fp16 = mul(x = var_1992_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = tensor("_inversed_scores_33_cast_fp16")]; + tensor scores_35_cast_fp16 = select(a = var_19_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_3)[name = tensor("scores_35_cast_fp16")]; + tensor var_1998_cast_fp16 = softmax(axis = var_40, x = scores_35_cast_fp16)[name = tensor("op_1998_cast_fp16")]; + tensor input_457_cast_fp16 = select(a = var_18_to_fp16, b = var_1998_cast_fp16, cond = mask_3)[name = tensor("input_457_cast_fp16")]; + tensor x_221_transpose_x_0 = const()[name = tensor("x_221_transpose_x_0"), val = tensor(false)]; + tensor x_221_transpose_y_0 = const()[name = tensor("x_221_transpose_y_0"), val = tensor(false)]; + tensor value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = v_17_cast_fp16)[name = tensor("transpose_166")]; + tensor x_221_cast_fp16 = matmul(transpose_x = x_221_transpose_x_0, transpose_y = x_221_transpose_y_0, x = input_457_cast_fp16, y = value_17_cast_fp16)[name = tensor("x_221_cast_fp16")]; + tensor var_2002_perm_0 = const()[name = tensor("op_2002_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2003 = const()[name = tensor("op_2003"), val = tensor([1, -1, 512])]; + tensor var_2002_cast_fp16 = transpose(perm = var_2002_perm_0, x = x_221_cast_fp16)[name = tensor("transpose_162")]; + tensor input_459_cast_fp16 = reshape(shape = var_2003, x = var_2002_cast_fp16)[name = tensor("input_459_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_out_weight_to_fp16 = const()[name = tensor("encoder_layers_8_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108541888)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_out_weight_to_fp16, x = input_459_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_463_cast_fp16 = add(x = input_453_cast_fp16, y = linear_79_cast_fp16)[name = tensor("input_463_cast_fp16")]; + tensor x_225_axes_0 = const()[name = tensor("x_225_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109066240)))]; + tensor encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = tensor("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109067328)))]; + tensor x_225_cast_fp16 = layer_norm(axes = x_225_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input_463_cast_fp16)[name = tensor("x_225_cast_fp16")]; + tensor input_465_perm_0 = const()[name = tensor("input_465_perm_0"), val = tensor([0, 2, 1])]; + tensor input_467_pad_type_0 = const()[name = tensor("input_467_pad_type_0"), val = tensor("valid")]; + tensor input_467_strides_0 = const()[name = tensor("input_467_strides_0"), val = tensor([1])]; + tensor input_467_pad_0 = const()[name = tensor("input_467_pad_0"), val = tensor([0, 0])]; + tensor input_467_dilations_0 = const()[name = tensor("input_467_dilations_0"), val = tensor([1])]; + tensor input_467_groups_0 = const()[name = tensor("input_467_groups_0"), val = tensor(1)]; + tensor encoder_layers_8_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109068416)))]; + tensor input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = tensor("transpose_161")]; + tensor input_467_cast_fp16 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = encoder_layers_8_conv_pointwise_conv1_weight_to_fp16, x = input_465_cast_fp16)[name = tensor("input_467_cast_fp16")]; + tensor x_227_split_num_splits_0 = const()[name = tensor("x_227_split_num_splits_0"), val = tensor(2)]; + tensor x_227_split_axis_0 = const()[name = tensor("x_227_split_axis_0"), val = tensor(1)]; + tensor x_227_split_cast_fp16_0, tensor x_227_split_cast_fp16_1 = split(axis = x_227_split_axis_0, num_splits = x_227_split_num_splits_0, x = input_467_cast_fp16)[name = tensor("x_227_split_cast_fp16")]; + tensor x_227_split_1_sigmoid_cast_fp16 = sigmoid(x = x_227_split_cast_fp16_1)[name = tensor("x_227_split_1_sigmoid_cast_fp16")]; + tensor x_227_cast_fp16 = mul(x = x_227_split_cast_fp16_0, y = x_227_split_1_sigmoid_cast_fp16)[name = tensor("x_227_cast_fp16")]; + tensor input_469_cast_fp16 = select(a = var_18_to_fp16, b = x_227_cast_fp16, cond = var_396)[name = tensor("input_469_cast_fp16")]; + tensor new_x_35_interleave_0 = const()[name = tensor("new_x_35_interleave_0"), val = tensor(false)]; + tensor new_x_35_cast_fp16 = concat(axis = var_40, interleave = new_x_35_interleave_0, values = (cache_35_cast_fp16, input_469_cast_fp16))[name = tensor("new_x_35_cast_fp16")]; + tensor var_2041_begin_0 = const()[name = tensor("op_2041_begin_0"), val = tensor([0, 0, 4])]; + tensor var_2041_end_0 = const()[name = tensor("op_2041_end_0"), val = tensor([1, 512, 12])]; + tensor var_2041_end_mask_0 = const()[name = tensor("op_2041_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2041_cast_fp16 = slice_by_index(begin = var_2041_begin_0, end = var_2041_end_0, end_mask = var_2041_end_mask_0, x = new_x_35_cast_fp16)[name = tensor("op_2041_cast_fp16")]; + tensor x_229_pad_type_0 = const()[name = tensor("x_229_pad_type_0"), val = tensor("valid")]; + tensor x_229_groups_0 = const()[name = tensor("x_229_groups_0"), val = tensor(512)]; + tensor x_229_strides_0 = const()[name = tensor("x_229_strides_0"), val = tensor([1])]; + tensor x_229_pad_0 = const()[name = tensor("x_229_pad_0"), val = tensor([0, 0])]; + tensor x_229_dilations_0 = const()[name = tensor("x_229_dilations_0"), val = tensor([1])]; + tensor encoder_layers_8_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_8_conv_depthwise_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110117056)))]; + tensor x_229_cast_fp16 = conv(dilations = x_229_dilations_0, groups = x_229_groups_0, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = x_229_strides_0, weight = encoder_layers_8_conv_depthwise_conv_weight_to_fp16, x = new_x_35_cast_fp16)[name = tensor("x_229_cast_fp16")]; + tensor input_471_perm_0 = const()[name = tensor("input_471_perm_0"), val = tensor([0, 2, 1])]; + tensor x_231_axes_0 = const()[name = tensor("x_231_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = tensor("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110126336)))]; + tensor encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = tensor("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110127424)))]; + tensor input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = tensor("transpose_160")]; + tensor x_231_cast_fp16 = layer_norm(axes = x_231_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input_471_cast_fp16)[name = tensor("x_231_cast_fp16")]; + tensor input_473_perm_0 = const()[name = tensor("input_473_perm_0"), val = tensor([0, 2, 1])]; + tensor input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = tensor("transpose_159")]; + tensor input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = tensor("input_475_cast_fp16")]; + tensor x_233_pad_type_0 = const()[name = tensor("x_233_pad_type_0"), val = tensor("valid")]; + tensor x_233_strides_0 = const()[name = tensor("x_233_strides_0"), val = tensor([1])]; + tensor x_233_pad_0 = const()[name = tensor("x_233_pad_0"), val = tensor([0, 0])]; + tensor x_233_dilations_0 = const()[name = tensor("x_233_dilations_0"), val = tensor([1])]; + tensor x_233_groups_0 = const()[name = tensor("x_233_groups_0"), val = tensor(1)]; + tensor encoder_layers_8_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110128512)))]; + tensor x_233_cast_fp16 = conv(dilations = x_233_dilations_0, groups = x_233_groups_0, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = x_233_strides_0, weight = encoder_layers_8_conv_pointwise_conv2_weight_to_fp16, x = input_475_cast_fp16)[name = tensor("x_233_cast_fp16")]; + tensor input_477_perm_0 = const()[name = tensor("input_477_perm_0"), val = tensor([0, 2, 1])]; + tensor input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = tensor("transpose_158")]; + tensor input_479_cast_fp16 = add(x = input_463_cast_fp16, y = input_477_cast_fp16)[name = tensor("input_479_cast_fp16")]; + tensor input_481_axes_0 = const()[name = tensor("input_481_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = tensor("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110652864)))]; + tensor encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = tensor("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110653952)))]; + tensor input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input_479_cast_fp16)[name = tensor("input_481_cast_fp16")]; + tensor encoder_layers_8_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_8_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110655040)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_8_feed_forward2_linear1_weight_to_fp16, x = input_481_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor input_485_cast_fp16 = silu(x = linear_80_cast_fp16)[name = tensor("input_485_cast_fp16")]; + tensor encoder_layers_8_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_8_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(112752256)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_feed_forward2_linear2_weight_to_fp16, x = input_485_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_2082_to_fp16 = const()[name = tensor("op_2082_to_fp16"), val = tensor(0x1p-1)]; + tensor var_2083_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2082_to_fp16)[name = tensor("op_2083_cast_fp16")]; + tensor input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2083_cast_fp16)[name = tensor("input_491_cast_fp16")]; + tensor input_493_axes_0 = const()[name = tensor("input_493_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_out_weight_to_fp16 = const()[name = tensor("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114849472)))]; + tensor encoder_layers_8_norm_out_bias_to_fp16 = const()[name = tensor("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114850560)))]; + tensor input_493_cast_fp16 = layer_norm(axes = input_493_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("input_493_cast_fp16")]; + tensor cache_37_begin_0 = const()[name = tensor("cache_37_begin_0"), val = tensor([9, 0, 0, 0])]; + tensor cache_37_end_0 = const()[name = tensor("cache_37_end_0"), val = tensor([10, 1, 70, 512])]; + tensor cache_37_end_mask_0 = const()[name = tensor("cache_37_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_37_squeeze_mask_0 = const()[name = tensor("cache_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_37_cast_fp16 = slice_by_index(begin = cache_37_begin_0, end = cache_37_end_0, end_mask = cache_37_end_mask_0, squeeze_mask = cache_37_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor("cache_37_cast_fp16")]; + tensor cache_39_begin_0 = const()[name = tensor("cache_39_begin_0"), val = tensor([9, 0, 0, 0])]; + tensor cache_39_end_0 = const()[name = tensor("cache_39_end_0"), val = tensor([10, 1, 512, 8])]; + tensor cache_39_end_mask_0 = const()[name = tensor("cache_39_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_39_squeeze_mask_0 = const()[name = tensor("cache_39_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_39_cast_fp16 = slice_by_index(begin = cache_39_begin_0, end = cache_39_end_0, end_mask = cache_39_end_mask_0, squeeze_mask = cache_39_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor("cache_39_cast_fp16")]; + tensor input_495_axes_0 = const()[name = tensor("input_495_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = tensor("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114851648)))]; + tensor encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = tensor("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114852736)))]; + tensor input_495_cast_fp16 = layer_norm(axes = input_495_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input_493_cast_fp16)[name = tensor("input_495_cast_fp16")]; + tensor encoder_layers_9_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_9_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114853824)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_9_feed_forward1_linear1_weight_to_fp16, x = input_495_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor input_499_cast_fp16 = silu(x = linear_82_cast_fp16)[name = tensor("input_499_cast_fp16")]; + tensor encoder_layers_9_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_9_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116951040)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_feed_forward1_linear2_weight_to_fp16, x = input_499_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor var_2117_to_fp16 = const()[name = tensor("op_2117_to_fp16"), val = tensor(0x1p-1)]; + tensor var_2118_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2117_to_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2118_cast_fp16)[name = tensor("input_505_cast_fp16")]; + tensor key_19_axes_0 = const()[name = tensor("key_19_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = tensor("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119048256)))]; + tensor encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = tensor("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119049344)))]; + tensor key_19_cast_fp16 = layer_norm(axes = key_19_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_505_cast_fp16)[name = tensor("key_19_cast_fp16")]; + tensor input_507_interleave_0 = const()[name = tensor("input_507_interleave_0"), val = tensor(false)]; + tensor input_507_cast_fp16 = concat(axis = var_42, interleave = input_507_interleave_0, values = (cache_37_cast_fp16, key_19_cast_fp16))[name = tensor("input_507_cast_fp16")]; + tensor var_2140_begin_0 = const()[name = tensor("op_2140_begin_0"), val = tensor([0, 4, 0])]; + tensor var_2140_end_0 = const()[name = tensor("op_2140_end_0"), val = tensor([1, 70, 512])]; + tensor var_2140_end_mask_0 = const()[name = tensor("op_2140_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2140_cast_fp16 = slice_by_index(begin = var_2140_begin_0, end = var_2140_end_0, end_mask = var_2140_end_mask_0, x = cache_37_cast_fp16)[name = tensor("op_2140_cast_fp16")]; + tensor var_2146_interleave_0 = const()[name = tensor("op_2146_interleave_0"), val = tensor(false)]; + tensor var_2146_cast_fp16 = concat(axis = var_42, interleave = var_2146_interleave_0, values = (var_2140_cast_fp16, key_19_cast_fp16))[name = tensor("op_2146_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_q_weight_to_fp16 = const()[name = tensor("encoder_layers_9_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119050432)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_q_weight_to_fp16, x = key_19_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_2150 = const()[name = tensor("op_2150"), val = tensor([1, -1, 8, 64])]; + tensor q_55_cast_fp16 = reshape(shape = var_2150, x = linear_84_cast_fp16)[name = tensor("q_55_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_k_weight_to_fp16 = const()[name = tensor("encoder_layers_9_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119574784)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_k_weight_to_fp16, x = input_507_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_2154 = const()[name = tensor("op_2154"), val = tensor([1, -1, 8, 64])]; + tensor k_37_cast_fp16 = reshape(shape = var_2154, x = linear_85_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_v_weight_to_fp16 = const()[name = tensor("encoder_layers_9_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120099136)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_v_weight_to_fp16, x = input_507_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_2158 = const()[name = tensor("op_2158"), val = tensor([1, -1, 8, 64])]; + tensor v_19_cast_fp16 = reshape(shape = var_2158, x = linear_86_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor value_19_perm_0 = const()[name = tensor("value_19_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor encoder_layers_9_self_attn_pos_bias_u_to_fp16 = const()[name = tensor("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120623488)))]; + tensor var_2170_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = tensor("op_2170_cast_fp16")]; + tensor encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = tensor("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120624576)))]; + tensor var_2172_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = tensor("op_2172_cast_fp16")]; + tensor q_with_bias_v_19_perm_0 = const()[name = tensor("q_with_bias_v_19_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor x_241_transpose_x_0 = const()[name = tensor("x_241_transpose_x_0"), val = tensor(false)]; + tensor x_241_transpose_y_0 = const()[name = tensor("x_241_transpose_y_0"), val = tensor(false)]; + tensor var_2174_to_fp16 = const()[name = tensor("op_2174_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120625664)))]; + tensor q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2172_cast_fp16)[name = tensor("transpose_156")]; + tensor x_241_cast_fp16 = matmul(transpose_x = x_241_transpose_x_0, transpose_y = x_241_transpose_y_0, x = q_with_bias_v_19_cast_fp16, y = var_2174_to_fp16)[name = tensor("x_241_cast_fp16")]; + tensor x_243_pad_0 = const()[name = tensor("x_243_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor x_243_mode_0 = const()[name = tensor("x_243_mode_0"), val = tensor("constant")]; + tensor const_140_to_fp16 = const()[name = tensor("const_140_to_fp16"), val = tensor(0x0p+0)]; + tensor x_243_cast_fp16 = pad(constant_val = const_140_to_fp16, mode = x_243_mode_0, pad = x_243_pad_0, x = x_241_cast_fp16)[name = tensor("x_243_cast_fp16")]; + tensor var_2182 = const()[name = tensor("op_2182"), val = tensor([1, 8, -1, 4])]; + tensor x_245_cast_fp16 = reshape(shape = var_2182, x = x_243_cast_fp16)[name = tensor("x_245_cast_fp16")]; + tensor var_2186_begin_0 = const()[name = tensor("op_2186_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2186_end_0 = const()[name = tensor("op_2186_end_0"), val = tensor([1, 8, 148, 4])]; + tensor var_2186_end_mask_0 = const()[name = tensor("op_2186_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2186_cast_fp16 = slice_by_index(begin = var_2186_begin_0, end = var_2186_end_0, end_mask = var_2186_end_mask_0, x = x_245_cast_fp16)[name = tensor("op_2186_cast_fp16")]; + tensor var_2187 = const()[name = tensor("op_2187"), val = tensor([1, 8, 4, 147])]; + tensor matrix_bd_37_cast_fp16 = reshape(shape = var_2187, x = var_2186_cast_fp16)[name = tensor("matrix_bd_37_cast_fp16")]; + tensor matrix_ac_19_transpose_x_0 = const()[name = tensor("matrix_ac_19_transpose_x_0"), val = tensor(false)]; + tensor matrix_ac_19_transpose_y_0 = const()[name = tensor("matrix_ac_19_transpose_y_0"), val = tensor(false)]; + tensor transpose_69_perm_0 = const()[name = tensor("transpose_69_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_70_perm_0 = const()[name = tensor("transpose_70_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_70 = transpose(perm = transpose_70_perm_0, x = k_37_cast_fp16)[name = tensor("transpose_154")]; + tensor transpose_69 = transpose(perm = transpose_69_perm_0, x = var_2170_cast_fp16)[name = tensor("transpose_155")]; + tensor matrix_ac_19_cast_fp16 = matmul(transpose_x = matrix_ac_19_transpose_x_0, transpose_y = matrix_ac_19_transpose_y_0, x = transpose_69, y = transpose_70)[name = tensor("matrix_ac_19_cast_fp16")]; + tensor matrix_bd_39_begin_0 = const()[name = tensor("matrix_bd_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_39_end_0 = const()[name = tensor("matrix_bd_39_end_0"), val = tensor([1, 8, 4, 74])]; + tensor matrix_bd_39_end_mask_0 = const()[name = tensor("matrix_bd_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_39_cast_fp16 = slice_by_index(begin = matrix_bd_39_begin_0, end = matrix_bd_39_end_0, end_mask = matrix_bd_39_end_mask_0, x = matrix_bd_37_cast_fp16)[name = tensor("matrix_bd_39_cast_fp16")]; + tensor var_2196_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = tensor("op_2196_cast_fp16")]; + tensor _inversed_scores_37_y_0_to_fp16 = const()[name = tensor("_inversed_scores_37_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor _inversed_scores_37_cast_fp16 = mul(x = var_2196_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = tensor("_inversed_scores_37_cast_fp16")]; + tensor scores_39_cast_fp16 = select(a = var_19_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_3)[name = tensor("scores_39_cast_fp16")]; + tensor var_2202_cast_fp16 = softmax(axis = var_40, x = scores_39_cast_fp16)[name = tensor("op_2202_cast_fp16")]; + tensor input_509_cast_fp16 = select(a = var_18_to_fp16, b = var_2202_cast_fp16, cond = mask_3)[name = tensor("input_509_cast_fp16")]; + tensor x_247_transpose_x_0 = const()[name = tensor("x_247_transpose_x_0"), val = tensor(false)]; + tensor x_247_transpose_y_0 = const()[name = tensor("x_247_transpose_y_0"), val = tensor(false)]; + tensor value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = v_19_cast_fp16)[name = tensor("transpose_157")]; + tensor x_247_cast_fp16 = matmul(transpose_x = x_247_transpose_x_0, transpose_y = x_247_transpose_y_0, x = input_509_cast_fp16, y = value_19_cast_fp16)[name = tensor("x_247_cast_fp16")]; + tensor var_2206_perm_0 = const()[name = tensor("op_2206_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2207 = const()[name = tensor("op_2207"), val = tensor([1, -1, 512])]; + tensor var_2206_cast_fp16 = transpose(perm = var_2206_perm_0, x = x_247_cast_fp16)[name = tensor("transpose_153")]; + tensor input_511_cast_fp16 = reshape(shape = var_2207, x = var_2206_cast_fp16)[name = tensor("input_511_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_out_weight_to_fp16 = const()[name = tensor("encoder_layers_9_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120776256)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_out_weight_to_fp16, x = input_511_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_515_cast_fp16 = add(x = input_505_cast_fp16, y = linear_88_cast_fp16)[name = tensor("input_515_cast_fp16")]; + tensor x_251_axes_0 = const()[name = tensor("x_251_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121300608)))]; + tensor encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = tensor("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121301696)))]; + tensor x_251_cast_fp16 = layer_norm(axes = x_251_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input_515_cast_fp16)[name = tensor("x_251_cast_fp16")]; + tensor input_517_perm_0 = const()[name = tensor("input_517_perm_0"), val = tensor([0, 2, 1])]; + tensor input_519_pad_type_0 = const()[name = tensor("input_519_pad_type_0"), val = tensor("valid")]; + tensor input_519_strides_0 = const()[name = tensor("input_519_strides_0"), val = tensor([1])]; + tensor input_519_pad_0 = const()[name = tensor("input_519_pad_0"), val = tensor([0, 0])]; + tensor input_519_dilations_0 = const()[name = tensor("input_519_dilations_0"), val = tensor([1])]; + tensor input_519_groups_0 = const()[name = tensor("input_519_groups_0"), val = tensor(1)]; + tensor encoder_layers_9_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121302784)))]; + tensor input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = tensor("transpose_152")]; + tensor input_519_cast_fp16 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = encoder_layers_9_conv_pointwise_conv1_weight_to_fp16, x = input_517_cast_fp16)[name = tensor("input_519_cast_fp16")]; + tensor x_253_split_num_splits_0 = const()[name = tensor("x_253_split_num_splits_0"), val = tensor(2)]; + tensor x_253_split_axis_0 = const()[name = tensor("x_253_split_axis_0"), val = tensor(1)]; + tensor x_253_split_cast_fp16_0, tensor x_253_split_cast_fp16_1 = split(axis = x_253_split_axis_0, num_splits = x_253_split_num_splits_0, x = input_519_cast_fp16)[name = tensor("x_253_split_cast_fp16")]; + tensor x_253_split_1_sigmoid_cast_fp16 = sigmoid(x = x_253_split_cast_fp16_1)[name = tensor("x_253_split_1_sigmoid_cast_fp16")]; + tensor x_253_cast_fp16 = mul(x = x_253_split_cast_fp16_0, y = x_253_split_1_sigmoid_cast_fp16)[name = tensor("x_253_cast_fp16")]; + tensor input_521_cast_fp16 = select(a = var_18_to_fp16, b = x_253_cast_fp16, cond = var_396)[name = tensor("input_521_cast_fp16")]; + tensor new_x_39_interleave_0 = const()[name = tensor("new_x_39_interleave_0"), val = tensor(false)]; + tensor new_x_39_cast_fp16 = concat(axis = var_40, interleave = new_x_39_interleave_0, values = (cache_39_cast_fp16, input_521_cast_fp16))[name = tensor("new_x_39_cast_fp16")]; + tensor var_2245_begin_0 = const()[name = tensor("op_2245_begin_0"), val = tensor([0, 0, 4])]; + tensor var_2245_end_0 = const()[name = tensor("op_2245_end_0"), val = tensor([1, 512, 12])]; + tensor var_2245_end_mask_0 = const()[name = tensor("op_2245_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2245_cast_fp16 = slice_by_index(begin = var_2245_begin_0, end = var_2245_end_0, end_mask = var_2245_end_mask_0, x = new_x_39_cast_fp16)[name = tensor("op_2245_cast_fp16")]; + tensor x_255_pad_type_0 = const()[name = tensor("x_255_pad_type_0"), val = tensor("valid")]; + tensor x_255_groups_0 = const()[name = tensor("x_255_groups_0"), val = tensor(512)]; + tensor x_255_strides_0 = const()[name = tensor("x_255_strides_0"), val = tensor([1])]; + tensor x_255_pad_0 = const()[name = tensor("x_255_pad_0"), val = tensor([0, 0])]; + tensor x_255_dilations_0 = const()[name = tensor("x_255_dilations_0"), val = tensor([1])]; + tensor encoder_layers_9_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_9_conv_depthwise_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122351424)))]; + tensor x_255_cast_fp16 = conv(dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = encoder_layers_9_conv_depthwise_conv_weight_to_fp16, x = new_x_39_cast_fp16)[name = tensor("x_255_cast_fp16")]; + tensor input_523_perm_0 = const()[name = tensor("input_523_perm_0"), val = tensor([0, 2, 1])]; + tensor x_257_axes_0 = const()[name = tensor("x_257_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = tensor("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122360704)))]; + tensor encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = tensor("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122361792)))]; + tensor input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = tensor("transpose_151")]; + tensor x_257_cast_fp16 = layer_norm(axes = x_257_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input_523_cast_fp16)[name = tensor("x_257_cast_fp16")]; + tensor input_525_perm_0 = const()[name = tensor("input_525_perm_0"), val = tensor([0, 2, 1])]; + tensor input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = tensor("transpose_150")]; + tensor input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = tensor("input_527_cast_fp16")]; + tensor x_259_pad_type_0 = const()[name = tensor("x_259_pad_type_0"), val = tensor("valid")]; + tensor x_259_strides_0 = const()[name = tensor("x_259_strides_0"), val = tensor([1])]; + tensor x_259_pad_0 = const()[name = tensor("x_259_pad_0"), val = tensor([0, 0])]; + tensor x_259_dilations_0 = const()[name = tensor("x_259_dilations_0"), val = tensor([1])]; + tensor x_259_groups_0 = const()[name = tensor("x_259_groups_0"), val = tensor(1)]; + tensor encoder_layers_9_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122362880)))]; + tensor x_259_cast_fp16 = conv(dilations = x_259_dilations_0, groups = x_259_groups_0, pad = x_259_pad_0, pad_type = x_259_pad_type_0, strides = x_259_strides_0, weight = encoder_layers_9_conv_pointwise_conv2_weight_to_fp16, x = input_527_cast_fp16)[name = tensor("x_259_cast_fp16")]; + tensor input_529_perm_0 = const()[name = tensor("input_529_perm_0"), val = tensor([0, 2, 1])]; + tensor input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = tensor("transpose_149")]; + tensor input_531_cast_fp16 = add(x = input_515_cast_fp16, y = input_529_cast_fp16)[name = tensor("input_531_cast_fp16")]; + tensor input_533_axes_0 = const()[name = tensor("input_533_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = tensor("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122887232)))]; + tensor encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = tensor("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122888320)))]; + tensor input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input_531_cast_fp16)[name = tensor("input_533_cast_fp16")]; + tensor encoder_layers_9_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_9_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122889408)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_9_feed_forward2_linear1_weight_to_fp16, x = input_533_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor input_537_cast_fp16 = silu(x = linear_89_cast_fp16)[name = tensor("input_537_cast_fp16")]; + tensor encoder_layers_9_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_9_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124986624)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_feed_forward2_linear2_weight_to_fp16, x = input_537_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_2286_to_fp16 = const()[name = tensor("op_2286_to_fp16"), val = tensor(0x1p-1)]; + tensor var_2287_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2286_to_fp16)[name = tensor("op_2287_cast_fp16")]; + tensor input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2287_cast_fp16)[name = tensor("input_543_cast_fp16")]; + tensor input_545_axes_0 = const()[name = tensor("input_545_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_out_weight_to_fp16 = const()[name = tensor("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127083840)))]; + tensor encoder_layers_9_norm_out_bias_to_fp16 = const()[name = tensor("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127084928)))]; + tensor input_545_cast_fp16 = layer_norm(axes = input_545_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input_543_cast_fp16)[name = tensor("input_545_cast_fp16")]; + tensor cache_41_begin_0 = const()[name = tensor("cache_41_begin_0"), val = tensor([10, 0, 0, 0])]; + tensor cache_41_end_0 = const()[name = tensor("cache_41_end_0"), val = tensor([11, 1, 70, 512])]; + tensor cache_41_end_mask_0 = const()[name = tensor("cache_41_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_41_squeeze_mask_0 = const()[name = tensor("cache_41_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_41_cast_fp16 = slice_by_index(begin = cache_41_begin_0, end = cache_41_end_0, end_mask = cache_41_end_mask_0, squeeze_mask = cache_41_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor("cache_41_cast_fp16")]; + tensor cache_43_begin_0 = const()[name = tensor("cache_43_begin_0"), val = tensor([10, 0, 0, 0])]; + tensor cache_43_end_0 = const()[name = tensor("cache_43_end_0"), val = tensor([11, 1, 512, 8])]; + tensor cache_43_end_mask_0 = const()[name = tensor("cache_43_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_43_squeeze_mask_0 = const()[name = tensor("cache_43_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_43_cast_fp16 = slice_by_index(begin = cache_43_begin_0, end = cache_43_end_0, end_mask = cache_43_end_mask_0, squeeze_mask = cache_43_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor("cache_43_cast_fp16")]; + tensor input_547_axes_0 = const()[name = tensor("input_547_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = tensor("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127086016)))]; + tensor encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = tensor("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127087104)))]; + tensor input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input_545_cast_fp16)[name = tensor("input_547_cast_fp16")]; + tensor encoder_layers_10_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_10_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127088192)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_10_feed_forward1_linear1_weight_to_fp16, x = input_547_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_551_cast_fp16 = silu(x = linear_91_cast_fp16)[name = tensor("input_551_cast_fp16")]; + tensor encoder_layers_10_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_10_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129185408)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_feed_forward1_linear2_weight_to_fp16, x = input_551_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_2321_to_fp16 = const()[name = tensor("op_2321_to_fp16"), val = tensor(0x1p-1)]; + tensor var_2322_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2321_to_fp16)[name = tensor("op_2322_cast_fp16")]; + tensor input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2322_cast_fp16)[name = tensor("input_557_cast_fp16")]; + tensor key_21_axes_0 = const()[name = tensor("key_21_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = tensor("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131282624)))]; + tensor encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = tensor("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131283712)))]; + tensor key_21_cast_fp16 = layer_norm(axes = key_21_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = tensor("key_21_cast_fp16")]; + tensor input_559_interleave_0 = const()[name = tensor("input_559_interleave_0"), val = tensor(false)]; + tensor input_559_cast_fp16 = concat(axis = var_42, interleave = input_559_interleave_0, values = (cache_41_cast_fp16, key_21_cast_fp16))[name = tensor("input_559_cast_fp16")]; + tensor var_2344_begin_0 = const()[name = tensor("op_2344_begin_0"), val = tensor([0, 4, 0])]; + tensor var_2344_end_0 = const()[name = tensor("op_2344_end_0"), val = tensor([1, 70, 512])]; + tensor var_2344_end_mask_0 = const()[name = tensor("op_2344_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2344_cast_fp16 = slice_by_index(begin = var_2344_begin_0, end = var_2344_end_0, end_mask = var_2344_end_mask_0, x = cache_41_cast_fp16)[name = tensor("op_2344_cast_fp16")]; + tensor var_2350_interleave_0 = const()[name = tensor("op_2350_interleave_0"), val = tensor(false)]; + tensor var_2350_cast_fp16 = concat(axis = var_42, interleave = var_2350_interleave_0, values = (var_2344_cast_fp16, key_21_cast_fp16))[name = tensor("op_2350_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_q_weight_to_fp16 = const()[name = tensor("encoder_layers_10_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131284800)))]; + tensor linear_93_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_q_weight_to_fp16, x = key_21_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_2354 = const()[name = tensor("op_2354"), val = tensor([1, -1, 8, 64])]; + tensor q_61_cast_fp16 = reshape(shape = var_2354, x = linear_93_cast_fp16)[name = tensor("q_61_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_k_weight_to_fp16 = const()[name = tensor("encoder_layers_10_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131809152)))]; + tensor linear_94_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_k_weight_to_fp16, x = input_559_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor var_2358 = const()[name = tensor("op_2358"), val = tensor([1, -1, 8, 64])]; + tensor k_41_cast_fp16 = reshape(shape = var_2358, x = linear_94_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_v_weight_to_fp16 = const()[name = tensor("encoder_layers_10_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132333504)))]; + tensor linear_95_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_v_weight_to_fp16, x = input_559_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor var_2362 = const()[name = tensor("op_2362"), val = tensor([1, -1, 8, 64])]; + tensor v_21_cast_fp16 = reshape(shape = var_2362, x = linear_95_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor value_21_perm_0 = const()[name = tensor("value_21_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor encoder_layers_10_self_attn_pos_bias_u_to_fp16 = const()[name = tensor("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132857856)))]; + tensor var_2374_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = tensor("op_2374_cast_fp16")]; + tensor encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = tensor("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132858944)))]; + tensor var_2376_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = tensor("op_2376_cast_fp16")]; + tensor q_with_bias_v_21_perm_0 = const()[name = tensor("q_with_bias_v_21_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor x_267_transpose_x_0 = const()[name = tensor("x_267_transpose_x_0"), val = tensor(false)]; + tensor x_267_transpose_y_0 = const()[name = tensor("x_267_transpose_y_0"), val = tensor(false)]; + tensor var_2378_to_fp16 = const()[name = tensor("op_2378_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132860032)))]; + tensor q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2376_cast_fp16)[name = tensor("transpose_147")]; + tensor x_267_cast_fp16 = matmul(transpose_x = x_267_transpose_x_0, transpose_y = x_267_transpose_y_0, x = q_with_bias_v_21_cast_fp16, y = var_2378_to_fp16)[name = tensor("x_267_cast_fp16")]; + tensor x_269_pad_0 = const()[name = tensor("x_269_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor x_269_mode_0 = const()[name = tensor("x_269_mode_0"), val = tensor("constant")]; + tensor const_153_to_fp16 = const()[name = tensor("const_153_to_fp16"), val = tensor(0x0p+0)]; + tensor x_269_cast_fp16 = pad(constant_val = const_153_to_fp16, mode = x_269_mode_0, pad = x_269_pad_0, x = x_267_cast_fp16)[name = tensor("x_269_cast_fp16")]; + tensor var_2386 = const()[name = tensor("op_2386"), val = tensor([1, 8, -1, 4])]; + tensor x_271_cast_fp16 = reshape(shape = var_2386, x = x_269_cast_fp16)[name = tensor("x_271_cast_fp16")]; + tensor var_2390_begin_0 = const()[name = tensor("op_2390_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2390_end_0 = const()[name = tensor("op_2390_end_0"), val = tensor([1, 8, 148, 4])]; + tensor var_2390_end_mask_0 = const()[name = tensor("op_2390_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2390_cast_fp16 = slice_by_index(begin = var_2390_begin_0, end = var_2390_end_0, end_mask = var_2390_end_mask_0, x = x_271_cast_fp16)[name = tensor("op_2390_cast_fp16")]; + tensor var_2391 = const()[name = tensor("op_2391"), val = tensor([1, 8, 4, 147])]; + tensor matrix_bd_41_cast_fp16 = reshape(shape = var_2391, x = var_2390_cast_fp16)[name = tensor("matrix_bd_41_cast_fp16")]; + tensor matrix_ac_21_transpose_x_0 = const()[name = tensor("matrix_ac_21_transpose_x_0"), val = tensor(false)]; + tensor matrix_ac_21_transpose_y_0 = const()[name = tensor("matrix_ac_21_transpose_y_0"), val = tensor(false)]; + tensor transpose_71_perm_0 = const()[name = tensor("transpose_71_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = k_41_cast_fp16)[name = tensor("transpose_145")]; + tensor transpose_71 = transpose(perm = transpose_71_perm_0, x = var_2374_cast_fp16)[name = tensor("transpose_146")]; + tensor matrix_ac_21_cast_fp16 = matmul(transpose_x = matrix_ac_21_transpose_x_0, transpose_y = matrix_ac_21_transpose_y_0, x = transpose_71, y = transpose_72)[name = tensor("matrix_ac_21_cast_fp16")]; + tensor matrix_bd_43_begin_0 = const()[name = tensor("matrix_bd_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_43_end_0 = const()[name = tensor("matrix_bd_43_end_0"), val = tensor([1, 8, 4, 74])]; + tensor matrix_bd_43_end_mask_0 = const()[name = tensor("matrix_bd_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_43_cast_fp16 = slice_by_index(begin = matrix_bd_43_begin_0, end = matrix_bd_43_end_0, end_mask = matrix_bd_43_end_mask_0, x = matrix_bd_41_cast_fp16)[name = tensor("matrix_bd_43_cast_fp16")]; + tensor var_2400_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = tensor("op_2400_cast_fp16")]; + tensor _inversed_scores_41_y_0_to_fp16 = const()[name = tensor("_inversed_scores_41_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor _inversed_scores_41_cast_fp16 = mul(x = var_2400_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = tensor("_inversed_scores_41_cast_fp16")]; + tensor scores_43_cast_fp16 = select(a = var_19_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_3)[name = tensor("scores_43_cast_fp16")]; + tensor var_2406_cast_fp16 = softmax(axis = var_40, x = scores_43_cast_fp16)[name = tensor("op_2406_cast_fp16")]; + tensor input_561_cast_fp16 = select(a = var_18_to_fp16, b = var_2406_cast_fp16, cond = mask_3)[name = tensor("input_561_cast_fp16")]; + tensor x_273_transpose_x_0 = const()[name = tensor("x_273_transpose_x_0"), val = tensor(false)]; + tensor x_273_transpose_y_0 = const()[name = tensor("x_273_transpose_y_0"), val = tensor(false)]; + tensor value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = v_21_cast_fp16)[name = tensor("transpose_148")]; + tensor x_273_cast_fp16 = matmul(transpose_x = x_273_transpose_x_0, transpose_y = x_273_transpose_y_0, x = input_561_cast_fp16, y = value_21_cast_fp16)[name = tensor("x_273_cast_fp16")]; + tensor var_2410_perm_0 = const()[name = tensor("op_2410_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2411 = const()[name = tensor("op_2411"), val = tensor([1, -1, 512])]; + tensor var_2410_cast_fp16 = transpose(perm = var_2410_perm_0, x = x_273_cast_fp16)[name = tensor("transpose_144")]; + tensor input_563_cast_fp16 = reshape(shape = var_2411, x = var_2410_cast_fp16)[name = tensor("input_563_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_out_weight_to_fp16 = const()[name = tensor("encoder_layers_10_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133010624)))]; + tensor linear_97_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_out_weight_to_fp16, x = input_563_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor input_567_cast_fp16 = add(x = input_557_cast_fp16, y = linear_97_cast_fp16)[name = tensor("input_567_cast_fp16")]; + tensor x_277_axes_0 = const()[name = tensor("x_277_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133534976)))]; + tensor encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = tensor("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133536064)))]; + tensor x_277_cast_fp16 = layer_norm(axes = x_277_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input_567_cast_fp16)[name = tensor("x_277_cast_fp16")]; + tensor input_569_perm_0 = const()[name = tensor("input_569_perm_0"), val = tensor([0, 2, 1])]; + tensor input_571_pad_type_0 = const()[name = tensor("input_571_pad_type_0"), val = tensor("valid")]; + tensor input_571_strides_0 = const()[name = tensor("input_571_strides_0"), val = tensor([1])]; + tensor input_571_pad_0 = const()[name = tensor("input_571_pad_0"), val = tensor([0, 0])]; + tensor input_571_dilations_0 = const()[name = tensor("input_571_dilations_0"), val = tensor([1])]; + tensor input_571_groups_0 = const()[name = tensor("input_571_groups_0"), val = tensor(1)]; + tensor encoder_layers_10_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133537152)))]; + tensor input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = tensor("transpose_143")]; + tensor input_571_cast_fp16 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = encoder_layers_10_conv_pointwise_conv1_weight_to_fp16, x = input_569_cast_fp16)[name = tensor("input_571_cast_fp16")]; + tensor x_279_split_num_splits_0 = const()[name = tensor("x_279_split_num_splits_0"), val = tensor(2)]; + tensor x_279_split_axis_0 = const()[name = tensor("x_279_split_axis_0"), val = tensor(1)]; + tensor x_279_split_cast_fp16_0, tensor x_279_split_cast_fp16_1 = split(axis = x_279_split_axis_0, num_splits = x_279_split_num_splits_0, x = input_571_cast_fp16)[name = tensor("x_279_split_cast_fp16")]; + tensor x_279_split_1_sigmoid_cast_fp16 = sigmoid(x = x_279_split_cast_fp16_1)[name = tensor("x_279_split_1_sigmoid_cast_fp16")]; + tensor x_279_cast_fp16 = mul(x = x_279_split_cast_fp16_0, y = x_279_split_1_sigmoid_cast_fp16)[name = tensor("x_279_cast_fp16")]; + tensor input_573_cast_fp16 = select(a = var_18_to_fp16, b = x_279_cast_fp16, cond = var_396)[name = tensor("input_573_cast_fp16")]; + tensor new_x_43_interleave_0 = const()[name = tensor("new_x_43_interleave_0"), val = tensor(false)]; + tensor new_x_43_cast_fp16 = concat(axis = var_40, interleave = new_x_43_interleave_0, values = (cache_43_cast_fp16, input_573_cast_fp16))[name = tensor("new_x_43_cast_fp16")]; + tensor var_2449_begin_0 = const()[name = tensor("op_2449_begin_0"), val = tensor([0, 0, 4])]; + tensor var_2449_end_0 = const()[name = tensor("op_2449_end_0"), val = tensor([1, 512, 12])]; + tensor var_2449_end_mask_0 = const()[name = tensor("op_2449_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2449_cast_fp16 = slice_by_index(begin = var_2449_begin_0, end = var_2449_end_0, end_mask = var_2449_end_mask_0, x = new_x_43_cast_fp16)[name = tensor("op_2449_cast_fp16")]; + tensor x_281_pad_type_0 = const()[name = tensor("x_281_pad_type_0"), val = tensor("valid")]; + tensor x_281_groups_0 = const()[name = tensor("x_281_groups_0"), val = tensor(512)]; + tensor x_281_strides_0 = const()[name = tensor("x_281_strides_0"), val = tensor([1])]; + tensor x_281_pad_0 = const()[name = tensor("x_281_pad_0"), val = tensor([0, 0])]; + tensor x_281_dilations_0 = const()[name = tensor("x_281_dilations_0"), val = tensor([1])]; + tensor encoder_layers_10_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_10_conv_depthwise_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134585792)))]; + tensor x_281_cast_fp16 = conv(dilations = x_281_dilations_0, groups = x_281_groups_0, pad = x_281_pad_0, pad_type = x_281_pad_type_0, strides = x_281_strides_0, weight = encoder_layers_10_conv_depthwise_conv_weight_to_fp16, x = new_x_43_cast_fp16)[name = tensor("x_281_cast_fp16")]; + tensor input_575_perm_0 = const()[name = tensor("input_575_perm_0"), val = tensor([0, 2, 1])]; + tensor x_283_axes_0 = const()[name = tensor("x_283_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = tensor("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134595072)))]; + tensor encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = tensor("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134596160)))]; + tensor input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = tensor("transpose_142")]; + tensor x_283_cast_fp16 = layer_norm(axes = x_283_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input_575_cast_fp16)[name = tensor("x_283_cast_fp16")]; + tensor input_577_perm_0 = const()[name = tensor("input_577_perm_0"), val = tensor([0, 2, 1])]; + tensor input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = tensor("transpose_141")]; + tensor input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = tensor("input_579_cast_fp16")]; + tensor x_285_pad_type_0 = const()[name = tensor("x_285_pad_type_0"), val = tensor("valid")]; + tensor x_285_strides_0 = const()[name = tensor("x_285_strides_0"), val = tensor([1])]; + tensor x_285_pad_0 = const()[name = tensor("x_285_pad_0"), val = tensor([0, 0])]; + tensor x_285_dilations_0 = const()[name = tensor("x_285_dilations_0"), val = tensor([1])]; + tensor x_285_groups_0 = const()[name = tensor("x_285_groups_0"), val = tensor(1)]; + tensor encoder_layers_10_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134597248)))]; + tensor x_285_cast_fp16 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = encoder_layers_10_conv_pointwise_conv2_weight_to_fp16, x = input_579_cast_fp16)[name = tensor("x_285_cast_fp16")]; + tensor input_581_perm_0 = const()[name = tensor("input_581_perm_0"), val = tensor([0, 2, 1])]; + tensor input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = tensor("transpose_140")]; + tensor input_583_cast_fp16 = add(x = input_567_cast_fp16, y = input_581_cast_fp16)[name = tensor("input_583_cast_fp16")]; + tensor input_585_axes_0 = const()[name = tensor("input_585_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = tensor("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135121600)))]; + tensor encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = tensor("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135122688)))]; + tensor input_585_cast_fp16 = layer_norm(axes = input_585_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input_583_cast_fp16)[name = tensor("input_585_cast_fp16")]; + tensor encoder_layers_10_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_10_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135123776)))]; + tensor linear_98_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_10_feed_forward2_linear1_weight_to_fp16, x = input_585_cast_fp16)[name = tensor("linear_98_cast_fp16")]; + tensor input_589_cast_fp16 = silu(x = linear_98_cast_fp16)[name = tensor("input_589_cast_fp16")]; + tensor encoder_layers_10_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_10_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137220992)))]; + tensor linear_99_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_feed_forward2_linear2_weight_to_fp16, x = input_589_cast_fp16)[name = tensor("linear_99_cast_fp16")]; + tensor var_2490_to_fp16 = const()[name = tensor("op_2490_to_fp16"), val = tensor(0x1p-1)]; + tensor var_2491_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2490_to_fp16)[name = tensor("op_2491_cast_fp16")]; + tensor input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2491_cast_fp16)[name = tensor("input_595_cast_fp16")]; + tensor input_597_axes_0 = const()[name = tensor("input_597_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_out_weight_to_fp16 = const()[name = tensor("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139318208)))]; + tensor encoder_layers_10_norm_out_bias_to_fp16 = const()[name = tensor("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139319296)))]; + tensor input_597_cast_fp16 = layer_norm(axes = input_597_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input_595_cast_fp16)[name = tensor("input_597_cast_fp16")]; + tensor cache_45_begin_0 = const()[name = tensor("cache_45_begin_0"), val = tensor([11, 0, 0, 0])]; + tensor cache_45_end_0 = const()[name = tensor("cache_45_end_0"), val = tensor([12, 1, 70, 512])]; + tensor cache_45_end_mask_0 = const()[name = tensor("cache_45_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_45_squeeze_mask_0 = const()[name = tensor("cache_45_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_45_cast_fp16 = slice_by_index(begin = cache_45_begin_0, end = cache_45_end_0, end_mask = cache_45_end_mask_0, squeeze_mask = cache_45_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor("cache_45_cast_fp16")]; + tensor cache_47_begin_0 = const()[name = tensor("cache_47_begin_0"), val = tensor([11, 0, 0, 0])]; + tensor cache_47_end_0 = const()[name = tensor("cache_47_end_0"), val = tensor([12, 1, 512, 8])]; + tensor cache_47_end_mask_0 = const()[name = tensor("cache_47_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_47_squeeze_mask_0 = const()[name = tensor("cache_47_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_47_cast_fp16 = slice_by_index(begin = cache_47_begin_0, end = cache_47_end_0, end_mask = cache_47_end_mask_0, squeeze_mask = cache_47_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor("cache_47_cast_fp16")]; + tensor input_599_axes_0 = const()[name = tensor("input_599_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = tensor("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139320384)))]; + tensor encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = tensor("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139321472)))]; + tensor input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input_597_cast_fp16)[name = tensor("input_599_cast_fp16")]; + tensor encoder_layers_11_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_11_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139322560)))]; + tensor linear_100_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_11_feed_forward1_linear1_weight_to_fp16, x = input_599_cast_fp16)[name = tensor("linear_100_cast_fp16")]; + tensor input_603_cast_fp16 = silu(x = linear_100_cast_fp16)[name = tensor("input_603_cast_fp16")]; + tensor encoder_layers_11_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_11_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141419776)))]; + tensor linear_101_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_feed_forward1_linear2_weight_to_fp16, x = input_603_cast_fp16)[name = tensor("linear_101_cast_fp16")]; + tensor var_2525_to_fp16 = const()[name = tensor("op_2525_to_fp16"), val = tensor(0x1p-1)]; + tensor var_2526_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2525_to_fp16)[name = tensor("op_2526_cast_fp16")]; + tensor input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2526_cast_fp16)[name = tensor("input_609_cast_fp16")]; + tensor key_23_axes_0 = const()[name = tensor("key_23_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = tensor("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143516992)))]; + tensor encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = tensor("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143518080)))]; + tensor key_23_cast_fp16 = layer_norm(axes = key_23_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_609_cast_fp16)[name = tensor("key_23_cast_fp16")]; + tensor input_611_interleave_0 = const()[name = tensor("input_611_interleave_0"), val = tensor(false)]; + tensor input_611_cast_fp16 = concat(axis = var_42, interleave = input_611_interleave_0, values = (cache_45_cast_fp16, key_23_cast_fp16))[name = tensor("input_611_cast_fp16")]; + tensor var_2548_begin_0 = const()[name = tensor("op_2548_begin_0"), val = tensor([0, 4, 0])]; + tensor var_2548_end_0 = const()[name = tensor("op_2548_end_0"), val = tensor([1, 70, 512])]; + tensor var_2548_end_mask_0 = const()[name = tensor("op_2548_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2548_cast_fp16 = slice_by_index(begin = var_2548_begin_0, end = var_2548_end_0, end_mask = var_2548_end_mask_0, x = cache_45_cast_fp16)[name = tensor("op_2548_cast_fp16")]; + tensor var_2554_interleave_0 = const()[name = tensor("op_2554_interleave_0"), val = tensor(false)]; + tensor var_2554_cast_fp16 = concat(axis = var_42, interleave = var_2554_interleave_0, values = (var_2548_cast_fp16, key_23_cast_fp16))[name = tensor("op_2554_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_q_weight_to_fp16 = const()[name = tensor("encoder_layers_11_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143519168)))]; + tensor linear_102_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_q_weight_to_fp16, x = key_23_cast_fp16)[name = tensor("linear_102_cast_fp16")]; + tensor var_2558 = const()[name = tensor("op_2558"), val = tensor([1, -1, 8, 64])]; + tensor q_67_cast_fp16 = reshape(shape = var_2558, x = linear_102_cast_fp16)[name = tensor("q_67_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_k_weight_to_fp16 = const()[name = tensor("encoder_layers_11_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144043520)))]; + tensor linear_103_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_k_weight_to_fp16, x = input_611_cast_fp16)[name = tensor("linear_103_cast_fp16")]; + tensor var_2562 = const()[name = tensor("op_2562"), val = tensor([1, -1, 8, 64])]; + tensor k_45_cast_fp16 = reshape(shape = var_2562, x = linear_103_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_v_weight_to_fp16 = const()[name = tensor("encoder_layers_11_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144567872)))]; + tensor linear_104_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_v_weight_to_fp16, x = input_611_cast_fp16)[name = tensor("linear_104_cast_fp16")]; + tensor var_2566 = const()[name = tensor("op_2566"), val = tensor([1, -1, 8, 64])]; + tensor v_23_cast_fp16 = reshape(shape = var_2566, x = linear_104_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor value_23_perm_0 = const()[name = tensor("value_23_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor encoder_layers_11_self_attn_pos_bias_u_to_fp16 = const()[name = tensor("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145092224)))]; + tensor var_2578_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = tensor("op_2578_cast_fp16")]; + tensor encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = tensor("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145093312)))]; + tensor var_2580_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = tensor("op_2580_cast_fp16")]; + tensor q_with_bias_v_23_perm_0 = const()[name = tensor("q_with_bias_v_23_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor x_293_transpose_x_0 = const()[name = tensor("x_293_transpose_x_0"), val = tensor(false)]; + tensor x_293_transpose_y_0 = const()[name = tensor("x_293_transpose_y_0"), val = tensor(false)]; + tensor var_2582_to_fp16 = const()[name = tensor("op_2582_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145094400)))]; + tensor q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2580_cast_fp16)[name = tensor("transpose_138")]; + tensor x_293_cast_fp16 = matmul(transpose_x = x_293_transpose_x_0, transpose_y = x_293_transpose_y_0, x = q_with_bias_v_23_cast_fp16, y = var_2582_to_fp16)[name = tensor("x_293_cast_fp16")]; + tensor x_295_pad_0 = const()[name = tensor("x_295_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor x_295_mode_0 = const()[name = tensor("x_295_mode_0"), val = tensor("constant")]; + tensor const_166_to_fp16 = const()[name = tensor("const_166_to_fp16"), val = tensor(0x0p+0)]; + tensor x_295_cast_fp16 = pad(constant_val = const_166_to_fp16, mode = x_295_mode_0, pad = x_295_pad_0, x = x_293_cast_fp16)[name = tensor("x_295_cast_fp16")]; + tensor var_2590 = const()[name = tensor("op_2590"), val = tensor([1, 8, -1, 4])]; + tensor x_297_cast_fp16 = reshape(shape = var_2590, x = x_295_cast_fp16)[name = tensor("x_297_cast_fp16")]; + tensor var_2594_begin_0 = const()[name = tensor("op_2594_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2594_end_0 = const()[name = tensor("op_2594_end_0"), val = tensor([1, 8, 148, 4])]; + tensor var_2594_end_mask_0 = const()[name = tensor("op_2594_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2594_cast_fp16 = slice_by_index(begin = var_2594_begin_0, end = var_2594_end_0, end_mask = var_2594_end_mask_0, x = x_297_cast_fp16)[name = tensor("op_2594_cast_fp16")]; + tensor var_2595 = const()[name = tensor("op_2595"), val = tensor([1, 8, 4, 147])]; + tensor matrix_bd_45_cast_fp16 = reshape(shape = var_2595, x = var_2594_cast_fp16)[name = tensor("matrix_bd_45_cast_fp16")]; + tensor matrix_ac_23_transpose_x_0 = const()[name = tensor("matrix_ac_23_transpose_x_0"), val = tensor(false)]; + tensor matrix_ac_23_transpose_y_0 = const()[name = tensor("matrix_ac_23_transpose_y_0"), val = tensor(false)]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = k_45_cast_fp16)[name = tensor("transpose_136")]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = var_2578_cast_fp16)[name = tensor("transpose_137")]; + tensor matrix_ac_23_cast_fp16 = matmul(transpose_x = matrix_ac_23_transpose_x_0, transpose_y = matrix_ac_23_transpose_y_0, x = transpose_73, y = transpose_74)[name = tensor("matrix_ac_23_cast_fp16")]; + tensor matrix_bd_47_begin_0 = const()[name = tensor("matrix_bd_47_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_47_end_0 = const()[name = tensor("matrix_bd_47_end_0"), val = tensor([1, 8, 4, 74])]; + tensor matrix_bd_47_end_mask_0 = const()[name = tensor("matrix_bd_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_47_cast_fp16 = slice_by_index(begin = matrix_bd_47_begin_0, end = matrix_bd_47_end_0, end_mask = matrix_bd_47_end_mask_0, x = matrix_bd_45_cast_fp16)[name = tensor("matrix_bd_47_cast_fp16")]; + tensor var_2604_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = tensor("op_2604_cast_fp16")]; + tensor _inversed_scores_45_y_0_to_fp16 = const()[name = tensor("_inversed_scores_45_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor _inversed_scores_45_cast_fp16 = mul(x = var_2604_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = tensor("_inversed_scores_45_cast_fp16")]; + tensor scores_47_cast_fp16 = select(a = var_19_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_3)[name = tensor("scores_47_cast_fp16")]; + tensor var_2610_cast_fp16 = softmax(axis = var_40, x = scores_47_cast_fp16)[name = tensor("op_2610_cast_fp16")]; + tensor input_613_cast_fp16 = select(a = var_18_to_fp16, b = var_2610_cast_fp16, cond = mask_3)[name = tensor("input_613_cast_fp16")]; + tensor x_299_transpose_x_0 = const()[name = tensor("x_299_transpose_x_0"), val = tensor(false)]; + tensor x_299_transpose_y_0 = const()[name = tensor("x_299_transpose_y_0"), val = tensor(false)]; + tensor value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = v_23_cast_fp16)[name = tensor("transpose_139")]; + tensor x_299_cast_fp16 = matmul(transpose_x = x_299_transpose_x_0, transpose_y = x_299_transpose_y_0, x = input_613_cast_fp16, y = value_23_cast_fp16)[name = tensor("x_299_cast_fp16")]; + tensor var_2614_perm_0 = const()[name = tensor("op_2614_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2615 = const()[name = tensor("op_2615"), val = tensor([1, -1, 512])]; + tensor var_2614_cast_fp16 = transpose(perm = var_2614_perm_0, x = x_299_cast_fp16)[name = tensor("transpose_135")]; + tensor input_615_cast_fp16 = reshape(shape = var_2615, x = var_2614_cast_fp16)[name = tensor("input_615_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_out_weight_to_fp16 = const()[name = tensor("encoder_layers_11_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145244992)))]; + tensor linear_106_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_out_weight_to_fp16, x = input_615_cast_fp16)[name = tensor("linear_106_cast_fp16")]; + tensor input_619_cast_fp16 = add(x = input_609_cast_fp16, y = linear_106_cast_fp16)[name = tensor("input_619_cast_fp16")]; + tensor x_303_axes_0 = const()[name = tensor("x_303_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145769344)))]; + tensor encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = tensor("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145770432)))]; + tensor x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input_619_cast_fp16)[name = tensor("x_303_cast_fp16")]; + tensor input_621_perm_0 = const()[name = tensor("input_621_perm_0"), val = tensor([0, 2, 1])]; + tensor input_623_pad_type_0 = const()[name = tensor("input_623_pad_type_0"), val = tensor("valid")]; + tensor input_623_strides_0 = const()[name = tensor("input_623_strides_0"), val = tensor([1])]; + tensor input_623_pad_0 = const()[name = tensor("input_623_pad_0"), val = tensor([0, 0])]; + tensor input_623_dilations_0 = const()[name = tensor("input_623_dilations_0"), val = tensor([1])]; + tensor input_623_groups_0 = const()[name = tensor("input_623_groups_0"), val = tensor(1)]; + tensor encoder_layers_11_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145771520)))]; + tensor input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = tensor("transpose_134")]; + tensor input_623_cast_fp16 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = encoder_layers_11_conv_pointwise_conv1_weight_to_fp16, x = input_621_cast_fp16)[name = tensor("input_623_cast_fp16")]; + tensor x_305_split_num_splits_0 = const()[name = tensor("x_305_split_num_splits_0"), val = tensor(2)]; + tensor x_305_split_axis_0 = const()[name = tensor("x_305_split_axis_0"), val = tensor(1)]; + tensor x_305_split_cast_fp16_0, tensor x_305_split_cast_fp16_1 = split(axis = x_305_split_axis_0, num_splits = x_305_split_num_splits_0, x = input_623_cast_fp16)[name = tensor("x_305_split_cast_fp16")]; + tensor x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = tensor("x_305_split_1_sigmoid_cast_fp16")]; + tensor x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = tensor("x_305_cast_fp16")]; + tensor input_625_cast_fp16 = select(a = var_18_to_fp16, b = x_305_cast_fp16, cond = var_396)[name = tensor("input_625_cast_fp16")]; + tensor new_x_47_interleave_0 = const()[name = tensor("new_x_47_interleave_0"), val = tensor(false)]; + tensor new_x_47_cast_fp16 = concat(axis = var_40, interleave = new_x_47_interleave_0, values = (cache_47_cast_fp16, input_625_cast_fp16))[name = tensor("new_x_47_cast_fp16")]; + tensor var_2653_begin_0 = const()[name = tensor("op_2653_begin_0"), val = tensor([0, 0, 4])]; + tensor var_2653_end_0 = const()[name = tensor("op_2653_end_0"), val = tensor([1, 512, 12])]; + tensor var_2653_end_mask_0 = const()[name = tensor("op_2653_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2653_cast_fp16 = slice_by_index(begin = var_2653_begin_0, end = var_2653_end_0, end_mask = var_2653_end_mask_0, x = new_x_47_cast_fp16)[name = tensor("op_2653_cast_fp16")]; + tensor x_307_pad_type_0 = const()[name = tensor("x_307_pad_type_0"), val = tensor("valid")]; + tensor x_307_groups_0 = const()[name = tensor("x_307_groups_0"), val = tensor(512)]; + tensor x_307_strides_0 = const()[name = tensor("x_307_strides_0"), val = tensor([1])]; + tensor x_307_pad_0 = const()[name = tensor("x_307_pad_0"), val = tensor([0, 0])]; + tensor x_307_dilations_0 = const()[name = tensor("x_307_dilations_0"), val = tensor([1])]; + tensor encoder_layers_11_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_11_conv_depthwise_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146820160)))]; + tensor x_307_cast_fp16 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = encoder_layers_11_conv_depthwise_conv_weight_to_fp16, x = new_x_47_cast_fp16)[name = tensor("x_307_cast_fp16")]; + tensor input_627_perm_0 = const()[name = tensor("input_627_perm_0"), val = tensor([0, 2, 1])]; + tensor x_309_axes_0 = const()[name = tensor("x_309_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = tensor("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146829440)))]; + tensor encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = tensor("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146830528)))]; + tensor input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = tensor("transpose_133")]; + tensor x_309_cast_fp16 = layer_norm(axes = x_309_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input_627_cast_fp16)[name = tensor("x_309_cast_fp16")]; + tensor input_629_perm_0 = const()[name = tensor("input_629_perm_0"), val = tensor([0, 2, 1])]; + tensor input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = tensor("transpose_132")]; + tensor input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = tensor("input_631_cast_fp16")]; + tensor x_311_pad_type_0 = const()[name = tensor("x_311_pad_type_0"), val = tensor("valid")]; + tensor x_311_strides_0 = const()[name = tensor("x_311_strides_0"), val = tensor([1])]; + tensor x_311_pad_0 = const()[name = tensor("x_311_pad_0"), val = tensor([0, 0])]; + tensor x_311_dilations_0 = const()[name = tensor("x_311_dilations_0"), val = tensor([1])]; + tensor x_311_groups_0 = const()[name = tensor("x_311_groups_0"), val = tensor(1)]; + tensor encoder_layers_11_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146831616)))]; + tensor x_311_cast_fp16 = conv(dilations = x_311_dilations_0, groups = x_311_groups_0, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = x_311_strides_0, weight = encoder_layers_11_conv_pointwise_conv2_weight_to_fp16, x = input_631_cast_fp16)[name = tensor("x_311_cast_fp16")]; + tensor input_633_perm_0 = const()[name = tensor("input_633_perm_0"), val = tensor([0, 2, 1])]; + tensor input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = tensor("transpose_131")]; + tensor input_635_cast_fp16 = add(x = input_619_cast_fp16, y = input_633_cast_fp16)[name = tensor("input_635_cast_fp16")]; + tensor input_637_axes_0 = const()[name = tensor("input_637_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = tensor("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147355968)))]; + tensor encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = tensor("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147357056)))]; + tensor input_637_cast_fp16 = layer_norm(axes = input_637_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input_635_cast_fp16)[name = tensor("input_637_cast_fp16")]; + tensor encoder_layers_11_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_11_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147358144)))]; + tensor linear_107_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_11_feed_forward2_linear1_weight_to_fp16, x = input_637_cast_fp16)[name = tensor("linear_107_cast_fp16")]; + tensor input_641_cast_fp16 = silu(x = linear_107_cast_fp16)[name = tensor("input_641_cast_fp16")]; + tensor encoder_layers_11_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_11_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149455360)))]; + tensor linear_108_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_feed_forward2_linear2_weight_to_fp16, x = input_641_cast_fp16)[name = tensor("linear_108_cast_fp16")]; + tensor var_2694_to_fp16 = const()[name = tensor("op_2694_to_fp16"), val = tensor(0x1p-1)]; + tensor var_2695_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2694_to_fp16)[name = tensor("op_2695_cast_fp16")]; + tensor input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2695_cast_fp16)[name = tensor("input_647_cast_fp16")]; + tensor input_649_axes_0 = const()[name = tensor("input_649_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_out_weight_to_fp16 = const()[name = tensor("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151552576)))]; + tensor encoder_layers_11_norm_out_bias_to_fp16 = const()[name = tensor("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151553664)))]; + tensor input_649_cast_fp16 = layer_norm(axes = input_649_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input_647_cast_fp16)[name = tensor("input_649_cast_fp16")]; + tensor cache_49_begin_0 = const()[name = tensor("cache_49_begin_0"), val = tensor([12, 0, 0, 0])]; + tensor cache_49_end_0 = const()[name = tensor("cache_49_end_0"), val = tensor([13, 1, 70, 512])]; + tensor cache_49_end_mask_0 = const()[name = tensor("cache_49_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_49_squeeze_mask_0 = const()[name = tensor("cache_49_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_49_cast_fp16 = slice_by_index(begin = cache_49_begin_0, end = cache_49_end_0, end_mask = cache_49_end_mask_0, squeeze_mask = cache_49_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor("cache_49_cast_fp16")]; + tensor cache_51_begin_0 = const()[name = tensor("cache_51_begin_0"), val = tensor([12, 0, 0, 0])]; + tensor cache_51_end_0 = const()[name = tensor("cache_51_end_0"), val = tensor([13, 1, 512, 8])]; + tensor cache_51_end_mask_0 = const()[name = tensor("cache_51_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_51_squeeze_mask_0 = const()[name = tensor("cache_51_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_51_cast_fp16 = slice_by_index(begin = cache_51_begin_0, end = cache_51_end_0, end_mask = cache_51_end_mask_0, squeeze_mask = cache_51_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor("cache_51_cast_fp16")]; + tensor input_651_axes_0 = const()[name = tensor("input_651_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = tensor("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151554752)))]; + tensor encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = tensor("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151555840)))]; + tensor input_651_cast_fp16 = layer_norm(axes = input_651_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input_649_cast_fp16)[name = tensor("input_651_cast_fp16")]; + tensor encoder_layers_12_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_12_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151556928)))]; + tensor linear_109_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_12_feed_forward1_linear1_weight_to_fp16, x = input_651_cast_fp16)[name = tensor("linear_109_cast_fp16")]; + tensor input_655_cast_fp16 = silu(x = linear_109_cast_fp16)[name = tensor("input_655_cast_fp16")]; + tensor encoder_layers_12_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_12_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153654144)))]; + tensor linear_110_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_feed_forward1_linear2_weight_to_fp16, x = input_655_cast_fp16)[name = tensor("linear_110_cast_fp16")]; + tensor var_2729_to_fp16 = const()[name = tensor("op_2729_to_fp16"), val = tensor(0x1p-1)]; + tensor var_2730_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_2729_to_fp16)[name = tensor("op_2730_cast_fp16")]; + tensor input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_2730_cast_fp16)[name = tensor("input_661_cast_fp16")]; + tensor key_25_axes_0 = const()[name = tensor("key_25_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = tensor("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155751360)))]; + tensor encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = tensor("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155752448)))]; + tensor key_25_cast_fp16 = layer_norm(axes = key_25_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_661_cast_fp16)[name = tensor("key_25_cast_fp16")]; + tensor input_663_interleave_0 = const()[name = tensor("input_663_interleave_0"), val = tensor(false)]; + tensor input_663_cast_fp16 = concat(axis = var_42, interleave = input_663_interleave_0, values = (cache_49_cast_fp16, key_25_cast_fp16))[name = tensor("input_663_cast_fp16")]; + tensor var_2752_begin_0 = const()[name = tensor("op_2752_begin_0"), val = tensor([0, 4, 0])]; + tensor var_2752_end_0 = const()[name = tensor("op_2752_end_0"), val = tensor([1, 70, 512])]; + tensor var_2752_end_mask_0 = const()[name = tensor("op_2752_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2752_cast_fp16 = slice_by_index(begin = var_2752_begin_0, end = var_2752_end_0, end_mask = var_2752_end_mask_0, x = cache_49_cast_fp16)[name = tensor("op_2752_cast_fp16")]; + tensor var_2758_interleave_0 = const()[name = tensor("op_2758_interleave_0"), val = tensor(false)]; + tensor var_2758_cast_fp16 = concat(axis = var_42, interleave = var_2758_interleave_0, values = (var_2752_cast_fp16, key_25_cast_fp16))[name = tensor("op_2758_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_q_weight_to_fp16 = const()[name = tensor("encoder_layers_12_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155753536)))]; + tensor linear_111_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_q_weight_to_fp16, x = key_25_cast_fp16)[name = tensor("linear_111_cast_fp16")]; + tensor var_2762 = const()[name = tensor("op_2762"), val = tensor([1, -1, 8, 64])]; + tensor q_73_cast_fp16 = reshape(shape = var_2762, x = linear_111_cast_fp16)[name = tensor("q_73_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_k_weight_to_fp16 = const()[name = tensor("encoder_layers_12_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156277888)))]; + tensor linear_112_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_k_weight_to_fp16, x = input_663_cast_fp16)[name = tensor("linear_112_cast_fp16")]; + tensor var_2766 = const()[name = tensor("op_2766"), val = tensor([1, -1, 8, 64])]; + tensor k_49_cast_fp16 = reshape(shape = var_2766, x = linear_112_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_v_weight_to_fp16 = const()[name = tensor("encoder_layers_12_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156802240)))]; + tensor linear_113_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_v_weight_to_fp16, x = input_663_cast_fp16)[name = tensor("linear_113_cast_fp16")]; + tensor var_2770 = const()[name = tensor("op_2770"), val = tensor([1, -1, 8, 64])]; + tensor v_25_cast_fp16 = reshape(shape = var_2770, x = linear_113_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor value_25_perm_0 = const()[name = tensor("value_25_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor encoder_layers_12_self_attn_pos_bias_u_to_fp16 = const()[name = tensor("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157326592)))]; + tensor var_2782_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = tensor("op_2782_cast_fp16")]; + tensor encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = tensor("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157327680)))]; + tensor var_2784_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = tensor("op_2784_cast_fp16")]; + tensor q_with_bias_v_25_perm_0 = const()[name = tensor("q_with_bias_v_25_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor x_319_transpose_x_0 = const()[name = tensor("x_319_transpose_x_0"), val = tensor(false)]; + tensor x_319_transpose_y_0 = const()[name = tensor("x_319_transpose_y_0"), val = tensor(false)]; + tensor var_2786_to_fp16 = const()[name = tensor("op_2786_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157328768)))]; + tensor q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_2784_cast_fp16)[name = tensor("transpose_129")]; + tensor x_319_cast_fp16 = matmul(transpose_x = x_319_transpose_x_0, transpose_y = x_319_transpose_y_0, x = q_with_bias_v_25_cast_fp16, y = var_2786_to_fp16)[name = tensor("x_319_cast_fp16")]; + tensor x_321_pad_0 = const()[name = tensor("x_321_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor x_321_mode_0 = const()[name = tensor("x_321_mode_0"), val = tensor("constant")]; + tensor const_179_to_fp16 = const()[name = tensor("const_179_to_fp16"), val = tensor(0x0p+0)]; + tensor x_321_cast_fp16 = pad(constant_val = const_179_to_fp16, mode = x_321_mode_0, pad = x_321_pad_0, x = x_319_cast_fp16)[name = tensor("x_321_cast_fp16")]; + tensor var_2794 = const()[name = tensor("op_2794"), val = tensor([1, 8, -1, 4])]; + tensor x_323_cast_fp16 = reshape(shape = var_2794, x = x_321_cast_fp16)[name = tensor("x_323_cast_fp16")]; + tensor var_2798_begin_0 = const()[name = tensor("op_2798_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2798_end_0 = const()[name = tensor("op_2798_end_0"), val = tensor([1, 8, 148, 4])]; + tensor var_2798_end_mask_0 = const()[name = tensor("op_2798_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2798_cast_fp16 = slice_by_index(begin = var_2798_begin_0, end = var_2798_end_0, end_mask = var_2798_end_mask_0, x = x_323_cast_fp16)[name = tensor("op_2798_cast_fp16")]; + tensor var_2799 = const()[name = tensor("op_2799"), val = tensor([1, 8, 4, 147])]; + tensor matrix_bd_49_cast_fp16 = reshape(shape = var_2799, x = var_2798_cast_fp16)[name = tensor("matrix_bd_49_cast_fp16")]; + tensor matrix_ac_25_transpose_x_0 = const()[name = tensor("matrix_ac_25_transpose_x_0"), val = tensor(false)]; + tensor matrix_ac_25_transpose_y_0 = const()[name = tensor("matrix_ac_25_transpose_y_0"), val = tensor(false)]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = k_49_cast_fp16)[name = tensor("transpose_127")]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = var_2782_cast_fp16)[name = tensor("transpose_128")]; + tensor matrix_ac_25_cast_fp16 = matmul(transpose_x = matrix_ac_25_transpose_x_0, transpose_y = matrix_ac_25_transpose_y_0, x = transpose_75, y = transpose_76)[name = tensor("matrix_ac_25_cast_fp16")]; + tensor matrix_bd_51_begin_0 = const()[name = tensor("matrix_bd_51_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_51_end_0 = const()[name = tensor("matrix_bd_51_end_0"), val = tensor([1, 8, 4, 74])]; + tensor matrix_bd_51_end_mask_0 = const()[name = tensor("matrix_bd_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_51_cast_fp16 = slice_by_index(begin = matrix_bd_51_begin_0, end = matrix_bd_51_end_0, end_mask = matrix_bd_51_end_mask_0, x = matrix_bd_49_cast_fp16)[name = tensor("matrix_bd_51_cast_fp16")]; + tensor var_2808_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = tensor("op_2808_cast_fp16")]; + tensor _inversed_scores_49_y_0_to_fp16 = const()[name = tensor("_inversed_scores_49_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor _inversed_scores_49_cast_fp16 = mul(x = var_2808_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = tensor("_inversed_scores_49_cast_fp16")]; + tensor scores_51_cast_fp16 = select(a = var_19_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_3)[name = tensor("scores_51_cast_fp16")]; + tensor var_2814_cast_fp16 = softmax(axis = var_40, x = scores_51_cast_fp16)[name = tensor("op_2814_cast_fp16")]; + tensor input_665_cast_fp16 = select(a = var_18_to_fp16, b = var_2814_cast_fp16, cond = mask_3)[name = tensor("input_665_cast_fp16")]; + tensor x_325_transpose_x_0 = const()[name = tensor("x_325_transpose_x_0"), val = tensor(false)]; + tensor x_325_transpose_y_0 = const()[name = tensor("x_325_transpose_y_0"), val = tensor(false)]; + tensor value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = v_25_cast_fp16)[name = tensor("transpose_130")]; + tensor x_325_cast_fp16 = matmul(transpose_x = x_325_transpose_x_0, transpose_y = x_325_transpose_y_0, x = input_665_cast_fp16, y = value_25_cast_fp16)[name = tensor("x_325_cast_fp16")]; + tensor var_2818_perm_0 = const()[name = tensor("op_2818_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2819 = const()[name = tensor("op_2819"), val = tensor([1, -1, 512])]; + tensor var_2818_cast_fp16 = transpose(perm = var_2818_perm_0, x = x_325_cast_fp16)[name = tensor("transpose_126")]; + tensor input_667_cast_fp16 = reshape(shape = var_2819, x = var_2818_cast_fp16)[name = tensor("input_667_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_out_weight_to_fp16 = const()[name = tensor("encoder_layers_12_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157479360)))]; + tensor linear_115_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_out_weight_to_fp16, x = input_667_cast_fp16)[name = tensor("linear_115_cast_fp16")]; + tensor input_671_cast_fp16 = add(x = input_661_cast_fp16, y = linear_115_cast_fp16)[name = tensor("input_671_cast_fp16")]; + tensor x_329_axes_0 = const()[name = tensor("x_329_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158003712)))]; + tensor encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = tensor("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158004800)))]; + tensor x_329_cast_fp16 = layer_norm(axes = x_329_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input_671_cast_fp16)[name = tensor("x_329_cast_fp16")]; + tensor input_673_perm_0 = const()[name = tensor("input_673_perm_0"), val = tensor([0, 2, 1])]; + tensor input_675_pad_type_0 = const()[name = tensor("input_675_pad_type_0"), val = tensor("valid")]; + tensor input_675_strides_0 = const()[name = tensor("input_675_strides_0"), val = tensor([1])]; + tensor input_675_pad_0 = const()[name = tensor("input_675_pad_0"), val = tensor([0, 0])]; + tensor input_675_dilations_0 = const()[name = tensor("input_675_dilations_0"), val = tensor([1])]; + tensor input_675_groups_0 = const()[name = tensor("input_675_groups_0"), val = tensor(1)]; + tensor encoder_layers_12_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158005888)))]; + tensor input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = tensor("transpose_125")]; + tensor input_675_cast_fp16 = conv(dilations = input_675_dilations_0, groups = input_675_groups_0, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = input_675_strides_0, weight = encoder_layers_12_conv_pointwise_conv1_weight_to_fp16, x = input_673_cast_fp16)[name = tensor("input_675_cast_fp16")]; + tensor x_331_split_num_splits_0 = const()[name = tensor("x_331_split_num_splits_0"), val = tensor(2)]; + tensor x_331_split_axis_0 = const()[name = tensor("x_331_split_axis_0"), val = tensor(1)]; + tensor x_331_split_cast_fp16_0, tensor x_331_split_cast_fp16_1 = split(axis = x_331_split_axis_0, num_splits = x_331_split_num_splits_0, x = input_675_cast_fp16)[name = tensor("x_331_split_cast_fp16")]; + tensor x_331_split_1_sigmoid_cast_fp16 = sigmoid(x = x_331_split_cast_fp16_1)[name = tensor("x_331_split_1_sigmoid_cast_fp16")]; + tensor x_331_cast_fp16 = mul(x = x_331_split_cast_fp16_0, y = x_331_split_1_sigmoid_cast_fp16)[name = tensor("x_331_cast_fp16")]; + tensor input_677_cast_fp16 = select(a = var_18_to_fp16, b = x_331_cast_fp16, cond = var_396)[name = tensor("input_677_cast_fp16")]; + tensor new_x_51_interleave_0 = const()[name = tensor("new_x_51_interleave_0"), val = tensor(false)]; + tensor new_x_51_cast_fp16 = concat(axis = var_40, interleave = new_x_51_interleave_0, values = (cache_51_cast_fp16, input_677_cast_fp16))[name = tensor("new_x_51_cast_fp16")]; + tensor var_2857_begin_0 = const()[name = tensor("op_2857_begin_0"), val = tensor([0, 0, 4])]; + tensor var_2857_end_0 = const()[name = tensor("op_2857_end_0"), val = tensor([1, 512, 12])]; + tensor var_2857_end_mask_0 = const()[name = tensor("op_2857_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2857_cast_fp16 = slice_by_index(begin = var_2857_begin_0, end = var_2857_end_0, end_mask = var_2857_end_mask_0, x = new_x_51_cast_fp16)[name = tensor("op_2857_cast_fp16")]; + tensor x_333_pad_type_0 = const()[name = tensor("x_333_pad_type_0"), val = tensor("valid")]; + tensor x_333_groups_0 = const()[name = tensor("x_333_groups_0"), val = tensor(512)]; + tensor x_333_strides_0 = const()[name = tensor("x_333_strides_0"), val = tensor([1])]; + tensor x_333_pad_0 = const()[name = tensor("x_333_pad_0"), val = tensor([0, 0])]; + tensor x_333_dilations_0 = const()[name = tensor("x_333_dilations_0"), val = tensor([1])]; + tensor encoder_layers_12_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_12_conv_depthwise_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159054528)))]; + tensor x_333_cast_fp16 = conv(dilations = x_333_dilations_0, groups = x_333_groups_0, pad = x_333_pad_0, pad_type = x_333_pad_type_0, strides = x_333_strides_0, weight = encoder_layers_12_conv_depthwise_conv_weight_to_fp16, x = new_x_51_cast_fp16)[name = tensor("x_333_cast_fp16")]; + tensor input_679_perm_0 = const()[name = tensor("input_679_perm_0"), val = tensor([0, 2, 1])]; + tensor x_335_axes_0 = const()[name = tensor("x_335_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = tensor("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159063808)))]; + tensor encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = tensor("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159064896)))]; + tensor input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = tensor("transpose_124")]; + tensor x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input_679_cast_fp16)[name = tensor("x_335_cast_fp16")]; + tensor input_681_perm_0 = const()[name = tensor("input_681_perm_0"), val = tensor([0, 2, 1])]; + tensor input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = tensor("transpose_123")]; + tensor input_683_cast_fp16 = silu(x = input_681_cast_fp16)[name = tensor("input_683_cast_fp16")]; + tensor x_337_pad_type_0 = const()[name = tensor("x_337_pad_type_0"), val = tensor("valid")]; + tensor x_337_strides_0 = const()[name = tensor("x_337_strides_0"), val = tensor([1])]; + tensor x_337_pad_0 = const()[name = tensor("x_337_pad_0"), val = tensor([0, 0])]; + tensor x_337_dilations_0 = const()[name = tensor("x_337_dilations_0"), val = tensor([1])]; + tensor x_337_groups_0 = const()[name = tensor("x_337_groups_0"), val = tensor(1)]; + tensor encoder_layers_12_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159065984)))]; + tensor x_337_cast_fp16 = conv(dilations = x_337_dilations_0, groups = x_337_groups_0, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = x_337_strides_0, weight = encoder_layers_12_conv_pointwise_conv2_weight_to_fp16, x = input_683_cast_fp16)[name = tensor("x_337_cast_fp16")]; + tensor input_685_perm_0 = const()[name = tensor("input_685_perm_0"), val = tensor([0, 2, 1])]; + tensor input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = tensor("transpose_122")]; + tensor input_687_cast_fp16 = add(x = input_671_cast_fp16, y = input_685_cast_fp16)[name = tensor("input_687_cast_fp16")]; + tensor input_689_axes_0 = const()[name = tensor("input_689_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = tensor("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159590336)))]; + tensor encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = tensor("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159591424)))]; + tensor input_689_cast_fp16 = layer_norm(axes = input_689_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input_687_cast_fp16)[name = tensor("input_689_cast_fp16")]; + tensor encoder_layers_12_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_12_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159592512)))]; + tensor linear_116_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_12_feed_forward2_linear1_weight_to_fp16, x = input_689_cast_fp16)[name = tensor("linear_116_cast_fp16")]; + tensor input_693_cast_fp16 = silu(x = linear_116_cast_fp16)[name = tensor("input_693_cast_fp16")]; + tensor encoder_layers_12_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_12_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161689728)))]; + tensor linear_117_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_feed_forward2_linear2_weight_to_fp16, x = input_693_cast_fp16)[name = tensor("linear_117_cast_fp16")]; + tensor var_2898_to_fp16 = const()[name = tensor("op_2898_to_fp16"), val = tensor(0x1p-1)]; + tensor var_2899_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_2898_to_fp16)[name = tensor("op_2899_cast_fp16")]; + tensor input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_2899_cast_fp16)[name = tensor("input_699_cast_fp16")]; + tensor input_701_axes_0 = const()[name = tensor("input_701_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_out_weight_to_fp16 = const()[name = tensor("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163786944)))]; + tensor encoder_layers_12_norm_out_bias_to_fp16 = const()[name = tensor("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163788032)))]; + tensor input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input_699_cast_fp16)[name = tensor("input_701_cast_fp16")]; + tensor cache_53_begin_0 = const()[name = tensor("cache_53_begin_0"), val = tensor([13, 0, 0, 0])]; + tensor cache_53_end_0 = const()[name = tensor("cache_53_end_0"), val = tensor([14, 1, 70, 512])]; + tensor cache_53_end_mask_0 = const()[name = tensor("cache_53_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_53_squeeze_mask_0 = const()[name = tensor("cache_53_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_53_cast_fp16 = slice_by_index(begin = cache_53_begin_0, end = cache_53_end_0, end_mask = cache_53_end_mask_0, squeeze_mask = cache_53_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor("cache_53_cast_fp16")]; + tensor cache_55_begin_0 = const()[name = tensor("cache_55_begin_0"), val = tensor([13, 0, 0, 0])]; + tensor cache_55_end_0 = const()[name = tensor("cache_55_end_0"), val = tensor([14, 1, 512, 8])]; + tensor cache_55_end_mask_0 = const()[name = tensor("cache_55_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_55_squeeze_mask_0 = const()[name = tensor("cache_55_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_55_cast_fp16 = slice_by_index(begin = cache_55_begin_0, end = cache_55_end_0, end_mask = cache_55_end_mask_0, squeeze_mask = cache_55_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor("cache_55_cast_fp16")]; + tensor input_703_axes_0 = const()[name = tensor("input_703_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = tensor("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163789120)))]; + tensor encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = tensor("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163790208)))]; + tensor input_703_cast_fp16 = layer_norm(axes = input_703_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input_701_cast_fp16)[name = tensor("input_703_cast_fp16")]; + tensor encoder_layers_13_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_13_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163791296)))]; + tensor linear_118_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_13_feed_forward1_linear1_weight_to_fp16, x = input_703_cast_fp16)[name = tensor("linear_118_cast_fp16")]; + tensor input_707_cast_fp16 = silu(x = linear_118_cast_fp16)[name = tensor("input_707_cast_fp16")]; + tensor encoder_layers_13_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_13_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165888512)))]; + tensor linear_119_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_feed_forward1_linear2_weight_to_fp16, x = input_707_cast_fp16)[name = tensor("linear_119_cast_fp16")]; + tensor var_2933_to_fp16 = const()[name = tensor("op_2933_to_fp16"), val = tensor(0x1p-1)]; + tensor var_2934_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_2933_to_fp16)[name = tensor("op_2934_cast_fp16")]; + tensor input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_2934_cast_fp16)[name = tensor("input_713_cast_fp16")]; + tensor key_27_axes_0 = const()[name = tensor("key_27_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = tensor("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167985728)))]; + tensor encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = tensor("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167986816)))]; + tensor key_27_cast_fp16 = layer_norm(axes = key_27_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_713_cast_fp16)[name = tensor("key_27_cast_fp16")]; + tensor input_715_interleave_0 = const()[name = tensor("input_715_interleave_0"), val = tensor(false)]; + tensor input_715_cast_fp16 = concat(axis = var_42, interleave = input_715_interleave_0, values = (cache_53_cast_fp16, key_27_cast_fp16))[name = tensor("input_715_cast_fp16")]; + tensor var_2956_begin_0 = const()[name = tensor("op_2956_begin_0"), val = tensor([0, 4, 0])]; + tensor var_2956_end_0 = const()[name = tensor("op_2956_end_0"), val = tensor([1, 70, 512])]; + tensor var_2956_end_mask_0 = const()[name = tensor("op_2956_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2956_cast_fp16 = slice_by_index(begin = var_2956_begin_0, end = var_2956_end_0, end_mask = var_2956_end_mask_0, x = cache_53_cast_fp16)[name = tensor("op_2956_cast_fp16")]; + tensor var_2962_interleave_0 = const()[name = tensor("op_2962_interleave_0"), val = tensor(false)]; + tensor var_2962_cast_fp16 = concat(axis = var_42, interleave = var_2962_interleave_0, values = (var_2956_cast_fp16, key_27_cast_fp16))[name = tensor("op_2962_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_q_weight_to_fp16 = const()[name = tensor("encoder_layers_13_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167987904)))]; + tensor linear_120_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_q_weight_to_fp16, x = key_27_cast_fp16)[name = tensor("linear_120_cast_fp16")]; + tensor var_2966 = const()[name = tensor("op_2966"), val = tensor([1, -1, 8, 64])]; + tensor q_79_cast_fp16 = reshape(shape = var_2966, x = linear_120_cast_fp16)[name = tensor("q_79_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_k_weight_to_fp16 = const()[name = tensor("encoder_layers_13_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168512256)))]; + tensor linear_121_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_k_weight_to_fp16, x = input_715_cast_fp16)[name = tensor("linear_121_cast_fp16")]; + tensor var_2970 = const()[name = tensor("op_2970"), val = tensor([1, -1, 8, 64])]; + tensor k_53_cast_fp16 = reshape(shape = var_2970, x = linear_121_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_v_weight_to_fp16 = const()[name = tensor("encoder_layers_13_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169036608)))]; + tensor linear_122_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_v_weight_to_fp16, x = input_715_cast_fp16)[name = tensor("linear_122_cast_fp16")]; + tensor var_2974 = const()[name = tensor("op_2974"), val = tensor([1, -1, 8, 64])]; + tensor v_27_cast_fp16 = reshape(shape = var_2974, x = linear_122_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor value_27_perm_0 = const()[name = tensor("value_27_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor encoder_layers_13_self_attn_pos_bias_u_to_fp16 = const()[name = tensor("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169560960)))]; + tensor var_2986_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = tensor("op_2986_cast_fp16")]; + tensor encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = tensor("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169562048)))]; + tensor var_2988_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = tensor("op_2988_cast_fp16")]; + tensor q_with_bias_v_27_perm_0 = const()[name = tensor("q_with_bias_v_27_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor x_345_transpose_x_0 = const()[name = tensor("x_345_transpose_x_0"), val = tensor(false)]; + tensor x_345_transpose_y_0 = const()[name = tensor("x_345_transpose_y_0"), val = tensor(false)]; + tensor var_2990_to_fp16 = const()[name = tensor("op_2990_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169563136)))]; + tensor q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_2988_cast_fp16)[name = tensor("transpose_120")]; + tensor x_345_cast_fp16 = matmul(transpose_x = x_345_transpose_x_0, transpose_y = x_345_transpose_y_0, x = q_with_bias_v_27_cast_fp16, y = var_2990_to_fp16)[name = tensor("x_345_cast_fp16")]; + tensor x_347_pad_0 = const()[name = tensor("x_347_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor x_347_mode_0 = const()[name = tensor("x_347_mode_0"), val = tensor("constant")]; + tensor const_192_to_fp16 = const()[name = tensor("const_192_to_fp16"), val = tensor(0x0p+0)]; + tensor x_347_cast_fp16 = pad(constant_val = const_192_to_fp16, mode = x_347_mode_0, pad = x_347_pad_0, x = x_345_cast_fp16)[name = tensor("x_347_cast_fp16")]; + tensor var_2998 = const()[name = tensor("op_2998"), val = tensor([1, 8, -1, 4])]; + tensor x_349_cast_fp16 = reshape(shape = var_2998, x = x_347_cast_fp16)[name = tensor("x_349_cast_fp16")]; + tensor var_3002_begin_0 = const()[name = tensor("op_3002_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3002_end_0 = const()[name = tensor("op_3002_end_0"), val = tensor([1, 8, 148, 4])]; + tensor var_3002_end_mask_0 = const()[name = tensor("op_3002_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3002_cast_fp16 = slice_by_index(begin = var_3002_begin_0, end = var_3002_end_0, end_mask = var_3002_end_mask_0, x = x_349_cast_fp16)[name = tensor("op_3002_cast_fp16")]; + tensor var_3003 = const()[name = tensor("op_3003"), val = tensor([1, 8, 4, 147])]; + tensor matrix_bd_53_cast_fp16 = reshape(shape = var_3003, x = var_3002_cast_fp16)[name = tensor("matrix_bd_53_cast_fp16")]; + tensor matrix_ac_27_transpose_x_0 = const()[name = tensor("matrix_ac_27_transpose_x_0"), val = tensor(false)]; + tensor matrix_ac_27_transpose_y_0 = const()[name = tensor("matrix_ac_27_transpose_y_0"), val = tensor(false)]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = k_53_cast_fp16)[name = tensor("transpose_118")]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = var_2986_cast_fp16)[name = tensor("transpose_119")]; + tensor matrix_ac_27_cast_fp16 = matmul(transpose_x = matrix_ac_27_transpose_x_0, transpose_y = matrix_ac_27_transpose_y_0, x = transpose_77, y = transpose_78)[name = tensor("matrix_ac_27_cast_fp16")]; + tensor matrix_bd_55_begin_0 = const()[name = tensor("matrix_bd_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_55_end_0 = const()[name = tensor("matrix_bd_55_end_0"), val = tensor([1, 8, 4, 74])]; + tensor matrix_bd_55_end_mask_0 = const()[name = tensor("matrix_bd_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_55_cast_fp16 = slice_by_index(begin = matrix_bd_55_begin_0, end = matrix_bd_55_end_0, end_mask = matrix_bd_55_end_mask_0, x = matrix_bd_53_cast_fp16)[name = tensor("matrix_bd_55_cast_fp16")]; + tensor var_3012_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = tensor("op_3012_cast_fp16")]; + tensor _inversed_scores_53_y_0_to_fp16 = const()[name = tensor("_inversed_scores_53_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor _inversed_scores_53_cast_fp16 = mul(x = var_3012_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = tensor("_inversed_scores_53_cast_fp16")]; + tensor scores_55_cast_fp16 = select(a = var_19_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_3)[name = tensor("scores_55_cast_fp16")]; + tensor var_3018_cast_fp16 = softmax(axis = var_40, x = scores_55_cast_fp16)[name = tensor("op_3018_cast_fp16")]; + tensor input_717_cast_fp16 = select(a = var_18_to_fp16, b = var_3018_cast_fp16, cond = mask_3)[name = tensor("input_717_cast_fp16")]; + tensor x_351_transpose_x_0 = const()[name = tensor("x_351_transpose_x_0"), val = tensor(false)]; + tensor x_351_transpose_y_0 = const()[name = tensor("x_351_transpose_y_0"), val = tensor(false)]; + tensor value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = v_27_cast_fp16)[name = tensor("transpose_121")]; + tensor x_351_cast_fp16 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = input_717_cast_fp16, y = value_27_cast_fp16)[name = tensor("x_351_cast_fp16")]; + tensor var_3022_perm_0 = const()[name = tensor("op_3022_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3023 = const()[name = tensor("op_3023"), val = tensor([1, -1, 512])]; + tensor var_3022_cast_fp16 = transpose(perm = var_3022_perm_0, x = x_351_cast_fp16)[name = tensor("transpose_117")]; + tensor input_719_cast_fp16 = reshape(shape = var_3023, x = var_3022_cast_fp16)[name = tensor("input_719_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_out_weight_to_fp16 = const()[name = tensor("encoder_layers_13_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169713728)))]; + tensor linear_124_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_out_weight_to_fp16, x = input_719_cast_fp16)[name = tensor("linear_124_cast_fp16")]; + tensor input_723_cast_fp16 = add(x = input_713_cast_fp16, y = linear_124_cast_fp16)[name = tensor("input_723_cast_fp16")]; + tensor x_355_axes_0 = const()[name = tensor("x_355_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(170238080)))]; + tensor encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = tensor("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(170239168)))]; + tensor x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input_723_cast_fp16)[name = tensor("x_355_cast_fp16")]; + tensor input_725_perm_0 = const()[name = tensor("input_725_perm_0"), val = tensor([0, 2, 1])]; + tensor input_727_pad_type_0 = const()[name = tensor("input_727_pad_type_0"), val = tensor("valid")]; + tensor input_727_strides_0 = const()[name = tensor("input_727_strides_0"), val = tensor([1])]; + tensor input_727_pad_0 = const()[name = tensor("input_727_pad_0"), val = tensor([0, 0])]; + tensor input_727_dilations_0 = const()[name = tensor("input_727_dilations_0"), val = tensor([1])]; + tensor input_727_groups_0 = const()[name = tensor("input_727_groups_0"), val = tensor(1)]; + tensor encoder_layers_13_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(170240256)))]; + tensor input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = tensor("transpose_116")]; + tensor input_727_cast_fp16 = conv(dilations = input_727_dilations_0, groups = input_727_groups_0, pad = input_727_pad_0, pad_type = input_727_pad_type_0, strides = input_727_strides_0, weight = encoder_layers_13_conv_pointwise_conv1_weight_to_fp16, x = input_725_cast_fp16)[name = tensor("input_727_cast_fp16")]; + tensor x_357_split_num_splits_0 = const()[name = tensor("x_357_split_num_splits_0"), val = tensor(2)]; + tensor x_357_split_axis_0 = const()[name = tensor("x_357_split_axis_0"), val = tensor(1)]; + tensor x_357_split_cast_fp16_0, tensor x_357_split_cast_fp16_1 = split(axis = x_357_split_axis_0, num_splits = x_357_split_num_splits_0, x = input_727_cast_fp16)[name = tensor("x_357_split_cast_fp16")]; + tensor x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = tensor("x_357_split_1_sigmoid_cast_fp16")]; + tensor x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = tensor("x_357_cast_fp16")]; + tensor input_729_cast_fp16 = select(a = var_18_to_fp16, b = x_357_cast_fp16, cond = var_396)[name = tensor("input_729_cast_fp16")]; + tensor new_x_55_interleave_0 = const()[name = tensor("new_x_55_interleave_0"), val = tensor(false)]; + tensor new_x_55_cast_fp16 = concat(axis = var_40, interleave = new_x_55_interleave_0, values = (cache_55_cast_fp16, input_729_cast_fp16))[name = tensor("new_x_55_cast_fp16")]; + tensor var_3061_begin_0 = const()[name = tensor("op_3061_begin_0"), val = tensor([0, 0, 4])]; + tensor var_3061_end_0 = const()[name = tensor("op_3061_end_0"), val = tensor([1, 512, 12])]; + tensor var_3061_end_mask_0 = const()[name = tensor("op_3061_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3061_cast_fp16 = slice_by_index(begin = var_3061_begin_0, end = var_3061_end_0, end_mask = var_3061_end_mask_0, x = new_x_55_cast_fp16)[name = tensor("op_3061_cast_fp16")]; + tensor x_359_pad_type_0 = const()[name = tensor("x_359_pad_type_0"), val = tensor("valid")]; + tensor x_359_groups_0 = const()[name = tensor("x_359_groups_0"), val = tensor(512)]; + tensor x_359_strides_0 = const()[name = tensor("x_359_strides_0"), val = tensor([1])]; + tensor x_359_pad_0 = const()[name = tensor("x_359_pad_0"), val = tensor([0, 0])]; + tensor x_359_dilations_0 = const()[name = tensor("x_359_dilations_0"), val = tensor([1])]; + tensor encoder_layers_13_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_13_conv_depthwise_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171288896)))]; + tensor x_359_cast_fp16 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = encoder_layers_13_conv_depthwise_conv_weight_to_fp16, x = new_x_55_cast_fp16)[name = tensor("x_359_cast_fp16")]; + tensor input_731_perm_0 = const()[name = tensor("input_731_perm_0"), val = tensor([0, 2, 1])]; + tensor x_361_axes_0 = const()[name = tensor("x_361_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = tensor("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171298176)))]; + tensor encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = tensor("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171299264)))]; + tensor input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = tensor("transpose_115")]; + tensor x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input_731_cast_fp16)[name = tensor("x_361_cast_fp16")]; + tensor input_733_perm_0 = const()[name = tensor("input_733_perm_0"), val = tensor([0, 2, 1])]; + tensor input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = tensor("transpose_114")]; + tensor input_735_cast_fp16 = silu(x = input_733_cast_fp16)[name = tensor("input_735_cast_fp16")]; + tensor x_363_pad_type_0 = const()[name = tensor("x_363_pad_type_0"), val = tensor("valid")]; + tensor x_363_strides_0 = const()[name = tensor("x_363_strides_0"), val = tensor([1])]; + tensor x_363_pad_0 = const()[name = tensor("x_363_pad_0"), val = tensor([0, 0])]; + tensor x_363_dilations_0 = const()[name = tensor("x_363_dilations_0"), val = tensor([1])]; + tensor x_363_groups_0 = const()[name = tensor("x_363_groups_0"), val = tensor(1)]; + tensor encoder_layers_13_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171300352)))]; + tensor x_363_cast_fp16 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = encoder_layers_13_conv_pointwise_conv2_weight_to_fp16, x = input_735_cast_fp16)[name = tensor("x_363_cast_fp16")]; + tensor input_737_perm_0 = const()[name = tensor("input_737_perm_0"), val = tensor([0, 2, 1])]; + tensor input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = tensor("transpose_113")]; + tensor input_739_cast_fp16 = add(x = input_723_cast_fp16, y = input_737_cast_fp16)[name = tensor("input_739_cast_fp16")]; + tensor input_741_axes_0 = const()[name = tensor("input_741_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = tensor("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171824704)))]; + tensor encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = tensor("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171825792)))]; + tensor input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = tensor("input_741_cast_fp16")]; + tensor encoder_layers_13_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_13_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171826880)))]; + tensor linear_125_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_13_feed_forward2_linear1_weight_to_fp16, x = input_741_cast_fp16)[name = tensor("linear_125_cast_fp16")]; + tensor input_745_cast_fp16 = silu(x = linear_125_cast_fp16)[name = tensor("input_745_cast_fp16")]; + tensor encoder_layers_13_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_13_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173924096)))]; + tensor linear_126_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_feed_forward2_linear2_weight_to_fp16, x = input_745_cast_fp16)[name = tensor("linear_126_cast_fp16")]; + tensor var_3102_to_fp16 = const()[name = tensor("op_3102_to_fp16"), val = tensor(0x1p-1)]; + tensor var_3103_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3102_to_fp16)[name = tensor("op_3103_cast_fp16")]; + tensor input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3103_cast_fp16)[name = tensor("input_751_cast_fp16")]; + tensor input_753_axes_0 = const()[name = tensor("input_753_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_out_weight_to_fp16 = const()[name = tensor("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176021312)))]; + tensor encoder_layers_13_norm_out_bias_to_fp16 = const()[name = tensor("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176022400)))]; + tensor input_753_cast_fp16 = layer_norm(axes = input_753_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input_751_cast_fp16)[name = tensor("input_753_cast_fp16")]; + tensor cache_57_begin_0 = const()[name = tensor("cache_57_begin_0"), val = tensor([14, 0, 0, 0])]; + tensor cache_57_end_0 = const()[name = tensor("cache_57_end_0"), val = tensor([15, 1, 70, 512])]; + tensor cache_57_end_mask_0 = const()[name = tensor("cache_57_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_57_squeeze_mask_0 = const()[name = tensor("cache_57_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_57_cast_fp16 = slice_by_index(begin = cache_57_begin_0, end = cache_57_end_0, end_mask = cache_57_end_mask_0, squeeze_mask = cache_57_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor("cache_57_cast_fp16")]; + tensor cache_59_begin_0 = const()[name = tensor("cache_59_begin_0"), val = tensor([14, 0, 0, 0])]; + tensor cache_59_end_0 = const()[name = tensor("cache_59_end_0"), val = tensor([15, 1, 512, 8])]; + tensor cache_59_end_mask_0 = const()[name = tensor("cache_59_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_59_squeeze_mask_0 = const()[name = tensor("cache_59_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_59_cast_fp16 = slice_by_index(begin = cache_59_begin_0, end = cache_59_end_0, end_mask = cache_59_end_mask_0, squeeze_mask = cache_59_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor("cache_59_cast_fp16")]; + tensor input_755_axes_0 = const()[name = tensor("input_755_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = tensor("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176023488)))]; + tensor encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = tensor("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176024576)))]; + tensor input_755_cast_fp16 = layer_norm(axes = input_755_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input_753_cast_fp16)[name = tensor("input_755_cast_fp16")]; + tensor encoder_layers_14_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_14_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176025664)))]; + tensor linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_14_feed_forward1_linear1_weight_to_fp16, x = input_755_cast_fp16)[name = tensor("linear_127_cast_fp16")]; + tensor input_759_cast_fp16 = silu(x = linear_127_cast_fp16)[name = tensor("input_759_cast_fp16")]; + tensor encoder_layers_14_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_14_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178122880)))]; + tensor linear_128_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_feed_forward1_linear2_weight_to_fp16, x = input_759_cast_fp16)[name = tensor("linear_128_cast_fp16")]; + tensor var_3137_to_fp16 = const()[name = tensor("op_3137_to_fp16"), val = tensor(0x1p-1)]; + tensor var_3138_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3137_to_fp16)[name = tensor("op_3138_cast_fp16")]; + tensor input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3138_cast_fp16)[name = tensor("input_765_cast_fp16")]; + tensor key_29_axes_0 = const()[name = tensor("key_29_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = tensor("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180220096)))]; + tensor encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = tensor("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180221184)))]; + tensor key_29_cast_fp16 = layer_norm(axes = key_29_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_765_cast_fp16)[name = tensor("key_29_cast_fp16")]; + tensor input_767_interleave_0 = const()[name = tensor("input_767_interleave_0"), val = tensor(false)]; + tensor input_767_cast_fp16 = concat(axis = var_42, interleave = input_767_interleave_0, values = (cache_57_cast_fp16, key_29_cast_fp16))[name = tensor("input_767_cast_fp16")]; + tensor var_3160_begin_0 = const()[name = tensor("op_3160_begin_0"), val = tensor([0, 4, 0])]; + tensor var_3160_end_0 = const()[name = tensor("op_3160_end_0"), val = tensor([1, 70, 512])]; + tensor var_3160_end_mask_0 = const()[name = tensor("op_3160_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3160_cast_fp16 = slice_by_index(begin = var_3160_begin_0, end = var_3160_end_0, end_mask = var_3160_end_mask_0, x = cache_57_cast_fp16)[name = tensor("op_3160_cast_fp16")]; + tensor var_3166_interleave_0 = const()[name = tensor("op_3166_interleave_0"), val = tensor(false)]; + tensor var_3166_cast_fp16 = concat(axis = var_42, interleave = var_3166_interleave_0, values = (var_3160_cast_fp16, key_29_cast_fp16))[name = tensor("op_3166_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_q_weight_to_fp16 = const()[name = tensor("encoder_layers_14_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180222272)))]; + tensor linear_129_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_q_weight_to_fp16, x = key_29_cast_fp16)[name = tensor("linear_129_cast_fp16")]; + tensor var_3170 = const()[name = tensor("op_3170"), val = tensor([1, -1, 8, 64])]; + tensor q_85_cast_fp16 = reshape(shape = var_3170, x = linear_129_cast_fp16)[name = tensor("q_85_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_k_weight_to_fp16 = const()[name = tensor("encoder_layers_14_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180746624)))]; + tensor linear_130_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_k_weight_to_fp16, x = input_767_cast_fp16)[name = tensor("linear_130_cast_fp16")]; + tensor var_3174 = const()[name = tensor("op_3174"), val = tensor([1, -1, 8, 64])]; + tensor k_57_cast_fp16 = reshape(shape = var_3174, x = linear_130_cast_fp16)[name = tensor("k_57_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_v_weight_to_fp16 = const()[name = tensor("encoder_layers_14_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181270976)))]; + tensor linear_131_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_v_weight_to_fp16, x = input_767_cast_fp16)[name = tensor("linear_131_cast_fp16")]; + tensor var_3178 = const()[name = tensor("op_3178"), val = tensor([1, -1, 8, 64])]; + tensor v_29_cast_fp16 = reshape(shape = var_3178, x = linear_131_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor value_29_perm_0 = const()[name = tensor("value_29_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor encoder_layers_14_self_attn_pos_bias_u_to_fp16 = const()[name = tensor("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181795328)))]; + tensor var_3190_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = tensor("op_3190_cast_fp16")]; + tensor encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = tensor("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181796416)))]; + tensor var_3192_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = tensor("op_3192_cast_fp16")]; + tensor q_with_bias_v_29_perm_0 = const()[name = tensor("q_with_bias_v_29_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor x_371_transpose_x_0 = const()[name = tensor("x_371_transpose_x_0"), val = tensor(false)]; + tensor x_371_transpose_y_0 = const()[name = tensor("x_371_transpose_y_0"), val = tensor(false)]; + tensor var_3194_to_fp16 = const()[name = tensor("op_3194_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181797504)))]; + tensor q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3192_cast_fp16)[name = tensor("transpose_111")]; + tensor x_371_cast_fp16 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = q_with_bias_v_29_cast_fp16, y = var_3194_to_fp16)[name = tensor("x_371_cast_fp16")]; + tensor x_373_pad_0 = const()[name = tensor("x_373_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor x_373_mode_0 = const()[name = tensor("x_373_mode_0"), val = tensor("constant")]; + tensor const_205_to_fp16 = const()[name = tensor("const_205_to_fp16"), val = tensor(0x0p+0)]; + tensor x_373_cast_fp16 = pad(constant_val = const_205_to_fp16, mode = x_373_mode_0, pad = x_373_pad_0, x = x_371_cast_fp16)[name = tensor("x_373_cast_fp16")]; + tensor var_3202 = const()[name = tensor("op_3202"), val = tensor([1, 8, -1, 4])]; + tensor x_375_cast_fp16 = reshape(shape = var_3202, x = x_373_cast_fp16)[name = tensor("x_375_cast_fp16")]; + tensor var_3206_begin_0 = const()[name = tensor("op_3206_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3206_end_0 = const()[name = tensor("op_3206_end_0"), val = tensor([1, 8, 148, 4])]; + tensor var_3206_end_mask_0 = const()[name = tensor("op_3206_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3206_cast_fp16 = slice_by_index(begin = var_3206_begin_0, end = var_3206_end_0, end_mask = var_3206_end_mask_0, x = x_375_cast_fp16)[name = tensor("op_3206_cast_fp16")]; + tensor var_3207 = const()[name = tensor("op_3207"), val = tensor([1, 8, 4, 147])]; + tensor matrix_bd_57_cast_fp16 = reshape(shape = var_3207, x = var_3206_cast_fp16)[name = tensor("matrix_bd_57_cast_fp16")]; + tensor matrix_ac_29_transpose_x_0 = const()[name = tensor("matrix_ac_29_transpose_x_0"), val = tensor(false)]; + tensor matrix_ac_29_transpose_y_0 = const()[name = tensor("matrix_ac_29_transpose_y_0"), val = tensor(false)]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = k_57_cast_fp16)[name = tensor("transpose_109")]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = var_3190_cast_fp16)[name = tensor("transpose_110")]; + tensor matrix_ac_29_cast_fp16 = matmul(transpose_x = matrix_ac_29_transpose_x_0, transpose_y = matrix_ac_29_transpose_y_0, x = transpose_79, y = transpose_80)[name = tensor("matrix_ac_29_cast_fp16")]; + tensor matrix_bd_59_begin_0 = const()[name = tensor("matrix_bd_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_59_end_0 = const()[name = tensor("matrix_bd_59_end_0"), val = tensor([1, 8, 4, 74])]; + tensor matrix_bd_59_end_mask_0 = const()[name = tensor("matrix_bd_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_59_cast_fp16 = slice_by_index(begin = matrix_bd_59_begin_0, end = matrix_bd_59_end_0, end_mask = matrix_bd_59_end_mask_0, x = matrix_bd_57_cast_fp16)[name = tensor("matrix_bd_59_cast_fp16")]; + tensor var_3216_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = tensor("op_3216_cast_fp16")]; + tensor _inversed_scores_57_y_0_to_fp16 = const()[name = tensor("_inversed_scores_57_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor _inversed_scores_57_cast_fp16 = mul(x = var_3216_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = tensor("_inversed_scores_57_cast_fp16")]; + tensor scores_59_cast_fp16 = select(a = var_19_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_3)[name = tensor("scores_59_cast_fp16")]; + tensor var_3222_cast_fp16 = softmax(axis = var_40, x = scores_59_cast_fp16)[name = tensor("op_3222_cast_fp16")]; + tensor input_769_cast_fp16 = select(a = var_18_to_fp16, b = var_3222_cast_fp16, cond = mask_3)[name = tensor("input_769_cast_fp16")]; + tensor x_377_transpose_x_0 = const()[name = tensor("x_377_transpose_x_0"), val = tensor(false)]; + tensor x_377_transpose_y_0 = const()[name = tensor("x_377_transpose_y_0"), val = tensor(false)]; + tensor value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = v_29_cast_fp16)[name = tensor("transpose_112")]; + tensor x_377_cast_fp16 = matmul(transpose_x = x_377_transpose_x_0, transpose_y = x_377_transpose_y_0, x = input_769_cast_fp16, y = value_29_cast_fp16)[name = tensor("x_377_cast_fp16")]; + tensor var_3226_perm_0 = const()[name = tensor("op_3226_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3227 = const()[name = tensor("op_3227"), val = tensor([1, -1, 512])]; + tensor var_3226_cast_fp16 = transpose(perm = var_3226_perm_0, x = x_377_cast_fp16)[name = tensor("transpose_108")]; + tensor input_771_cast_fp16 = reshape(shape = var_3227, x = var_3226_cast_fp16)[name = tensor("input_771_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_out_weight_to_fp16 = const()[name = tensor("encoder_layers_14_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181948096)))]; + tensor linear_133_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_out_weight_to_fp16, x = input_771_cast_fp16)[name = tensor("linear_133_cast_fp16")]; + tensor input_775_cast_fp16 = add(x = input_765_cast_fp16, y = linear_133_cast_fp16)[name = tensor("input_775_cast_fp16")]; + tensor x_381_axes_0 = const()[name = tensor("x_381_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182472448)))]; + tensor encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = tensor("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182473536)))]; + tensor x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = tensor("x_381_cast_fp16")]; + tensor input_777_perm_0 = const()[name = tensor("input_777_perm_0"), val = tensor([0, 2, 1])]; + tensor input_779_pad_type_0 = const()[name = tensor("input_779_pad_type_0"), val = tensor("valid")]; + tensor input_779_strides_0 = const()[name = tensor("input_779_strides_0"), val = tensor([1])]; + tensor input_779_pad_0 = const()[name = tensor("input_779_pad_0"), val = tensor([0, 0])]; + tensor input_779_dilations_0 = const()[name = tensor("input_779_dilations_0"), val = tensor([1])]; + tensor input_779_groups_0 = const()[name = tensor("input_779_groups_0"), val = tensor(1)]; + tensor encoder_layers_14_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182474624)))]; + tensor input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = tensor("transpose_107")]; + tensor input_779_cast_fp16 = conv(dilations = input_779_dilations_0, groups = input_779_groups_0, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = input_779_strides_0, weight = encoder_layers_14_conv_pointwise_conv1_weight_to_fp16, x = input_777_cast_fp16)[name = tensor("input_779_cast_fp16")]; + tensor x_383_split_num_splits_0 = const()[name = tensor("x_383_split_num_splits_0"), val = tensor(2)]; + tensor x_383_split_axis_0 = const()[name = tensor("x_383_split_axis_0"), val = tensor(1)]; + tensor x_383_split_cast_fp16_0, tensor x_383_split_cast_fp16_1 = split(axis = x_383_split_axis_0, num_splits = x_383_split_num_splits_0, x = input_779_cast_fp16)[name = tensor("x_383_split_cast_fp16")]; + tensor x_383_split_1_sigmoid_cast_fp16 = sigmoid(x = x_383_split_cast_fp16_1)[name = tensor("x_383_split_1_sigmoid_cast_fp16")]; + tensor x_383_cast_fp16 = mul(x = x_383_split_cast_fp16_0, y = x_383_split_1_sigmoid_cast_fp16)[name = tensor("x_383_cast_fp16")]; + tensor input_781_cast_fp16 = select(a = var_18_to_fp16, b = x_383_cast_fp16, cond = var_396)[name = tensor("input_781_cast_fp16")]; + tensor new_x_59_interleave_0 = const()[name = tensor("new_x_59_interleave_0"), val = tensor(false)]; + tensor new_x_59_cast_fp16 = concat(axis = var_40, interleave = new_x_59_interleave_0, values = (cache_59_cast_fp16, input_781_cast_fp16))[name = tensor("new_x_59_cast_fp16")]; + tensor var_3265_begin_0 = const()[name = tensor("op_3265_begin_0"), val = tensor([0, 0, 4])]; + tensor var_3265_end_0 = const()[name = tensor("op_3265_end_0"), val = tensor([1, 512, 12])]; + tensor var_3265_end_mask_0 = const()[name = tensor("op_3265_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3265_cast_fp16 = slice_by_index(begin = var_3265_begin_0, end = var_3265_end_0, end_mask = var_3265_end_mask_0, x = new_x_59_cast_fp16)[name = tensor("op_3265_cast_fp16")]; + tensor x_385_pad_type_0 = const()[name = tensor("x_385_pad_type_0"), val = tensor("valid")]; + tensor x_385_groups_0 = const()[name = tensor("x_385_groups_0"), val = tensor(512)]; + tensor x_385_strides_0 = const()[name = tensor("x_385_strides_0"), val = tensor([1])]; + tensor x_385_pad_0 = const()[name = tensor("x_385_pad_0"), val = tensor([0, 0])]; + tensor x_385_dilations_0 = const()[name = tensor("x_385_dilations_0"), val = tensor([1])]; + tensor encoder_layers_14_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_14_conv_depthwise_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183523264)))]; + tensor x_385_cast_fp16 = conv(dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = encoder_layers_14_conv_depthwise_conv_weight_to_fp16, x = new_x_59_cast_fp16)[name = tensor("x_385_cast_fp16")]; + tensor input_783_perm_0 = const()[name = tensor("input_783_perm_0"), val = tensor([0, 2, 1])]; + tensor x_387_axes_0 = const()[name = tensor("x_387_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = tensor("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183532544)))]; + tensor encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = tensor("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183533632)))]; + tensor input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = tensor("transpose_106")]; + tensor x_387_cast_fp16 = layer_norm(axes = x_387_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input_783_cast_fp16)[name = tensor("x_387_cast_fp16")]; + tensor input_785_perm_0 = const()[name = tensor("input_785_perm_0"), val = tensor([0, 2, 1])]; + tensor input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = tensor("transpose_105")]; + tensor input_787_cast_fp16 = silu(x = input_785_cast_fp16)[name = tensor("input_787_cast_fp16")]; + tensor x_389_pad_type_0 = const()[name = tensor("x_389_pad_type_0"), val = tensor("valid")]; + tensor x_389_strides_0 = const()[name = tensor("x_389_strides_0"), val = tensor([1])]; + tensor x_389_pad_0 = const()[name = tensor("x_389_pad_0"), val = tensor([0, 0])]; + tensor x_389_dilations_0 = const()[name = tensor("x_389_dilations_0"), val = tensor([1])]; + tensor x_389_groups_0 = const()[name = tensor("x_389_groups_0"), val = tensor(1)]; + tensor encoder_layers_14_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183534720)))]; + tensor x_389_cast_fp16 = conv(dilations = x_389_dilations_0, groups = x_389_groups_0, pad = x_389_pad_0, pad_type = x_389_pad_type_0, strides = x_389_strides_0, weight = encoder_layers_14_conv_pointwise_conv2_weight_to_fp16, x = input_787_cast_fp16)[name = tensor("x_389_cast_fp16")]; + tensor input_789_perm_0 = const()[name = tensor("input_789_perm_0"), val = tensor([0, 2, 1])]; + tensor input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = tensor("transpose_104")]; + tensor input_791_cast_fp16 = add(x = input_775_cast_fp16, y = input_789_cast_fp16)[name = tensor("input_791_cast_fp16")]; + tensor input_793_axes_0 = const()[name = tensor("input_793_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = tensor("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184059072)))]; + tensor encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = tensor("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184060160)))]; + tensor input_793_cast_fp16 = layer_norm(axes = input_793_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = tensor("input_793_cast_fp16")]; + tensor encoder_layers_14_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_14_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184061248)))]; + tensor linear_134_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_14_feed_forward2_linear1_weight_to_fp16, x = input_793_cast_fp16)[name = tensor("linear_134_cast_fp16")]; + tensor input_797_cast_fp16 = silu(x = linear_134_cast_fp16)[name = tensor("input_797_cast_fp16")]; + tensor encoder_layers_14_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_14_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186158464)))]; + tensor linear_135_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_feed_forward2_linear2_weight_to_fp16, x = input_797_cast_fp16)[name = tensor("linear_135_cast_fp16")]; + tensor var_3306_to_fp16 = const()[name = tensor("op_3306_to_fp16"), val = tensor(0x1p-1)]; + tensor var_3307_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3306_to_fp16)[name = tensor("op_3307_cast_fp16")]; + tensor input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3307_cast_fp16)[name = tensor("input_803_cast_fp16")]; + tensor input_805_axes_0 = const()[name = tensor("input_805_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_out_weight_to_fp16 = const()[name = tensor("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188255680)))]; + tensor encoder_layers_14_norm_out_bias_to_fp16 = const()[name = tensor("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188256768)))]; + tensor input_805_cast_fp16 = layer_norm(axes = input_805_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input_803_cast_fp16)[name = tensor("input_805_cast_fp16")]; + tensor cache_61_begin_0 = const()[name = tensor("cache_61_begin_0"), val = tensor([15, 0, 0, 0])]; + tensor cache_61_end_0 = const()[name = tensor("cache_61_end_0"), val = tensor([16, 1, 70, 512])]; + tensor cache_61_end_mask_0 = const()[name = tensor("cache_61_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_61_squeeze_mask_0 = const()[name = tensor("cache_61_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_61_cast_fp16 = slice_by_index(begin = cache_61_begin_0, end = cache_61_end_0, end_mask = cache_61_end_mask_0, squeeze_mask = cache_61_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor("cache_61_cast_fp16")]; + tensor cache_63_begin_0 = const()[name = tensor("cache_63_begin_0"), val = tensor([15, 0, 0, 0])]; + tensor cache_63_end_0 = const()[name = tensor("cache_63_end_0"), val = tensor([16, 1, 512, 8])]; + tensor cache_63_end_mask_0 = const()[name = tensor("cache_63_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_63_squeeze_mask_0 = const()[name = tensor("cache_63_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_63_cast_fp16 = slice_by_index(begin = cache_63_begin_0, end = cache_63_end_0, end_mask = cache_63_end_mask_0, squeeze_mask = cache_63_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor("cache_63_cast_fp16")]; + tensor input_807_axes_0 = const()[name = tensor("input_807_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = tensor("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188257856)))]; + tensor encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = tensor("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188258944)))]; + tensor input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input_805_cast_fp16)[name = tensor("input_807_cast_fp16")]; + tensor encoder_layers_15_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_15_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188260032)))]; + tensor linear_136_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_15_feed_forward1_linear1_weight_to_fp16, x = input_807_cast_fp16)[name = tensor("linear_136_cast_fp16")]; + tensor input_811_cast_fp16 = silu(x = linear_136_cast_fp16)[name = tensor("input_811_cast_fp16")]; + tensor encoder_layers_15_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_15_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190357248)))]; + tensor linear_137_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_feed_forward1_linear2_weight_to_fp16, x = input_811_cast_fp16)[name = tensor("linear_137_cast_fp16")]; + tensor var_3341_to_fp16 = const()[name = tensor("op_3341_to_fp16"), val = tensor(0x1p-1)]; + tensor var_3342_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3341_to_fp16)[name = tensor("op_3342_cast_fp16")]; + tensor input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3342_cast_fp16)[name = tensor("input_817_cast_fp16")]; + tensor key_31_axes_0 = const()[name = tensor("key_31_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = tensor("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192454464)))]; + tensor encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = tensor("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192455552)))]; + tensor key_31_cast_fp16 = layer_norm(axes = key_31_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_817_cast_fp16)[name = tensor("key_31_cast_fp16")]; + tensor input_819_interleave_0 = const()[name = tensor("input_819_interleave_0"), val = tensor(false)]; + tensor input_819_cast_fp16 = concat(axis = var_42, interleave = input_819_interleave_0, values = (cache_61_cast_fp16, key_31_cast_fp16))[name = tensor("input_819_cast_fp16")]; + tensor var_3364_begin_0 = const()[name = tensor("op_3364_begin_0"), val = tensor([0, 4, 0])]; + tensor var_3364_end_0 = const()[name = tensor("op_3364_end_0"), val = tensor([1, 70, 512])]; + tensor var_3364_end_mask_0 = const()[name = tensor("op_3364_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3364_cast_fp16 = slice_by_index(begin = var_3364_begin_0, end = var_3364_end_0, end_mask = var_3364_end_mask_0, x = cache_61_cast_fp16)[name = tensor("op_3364_cast_fp16")]; + tensor var_3370_interleave_0 = const()[name = tensor("op_3370_interleave_0"), val = tensor(false)]; + tensor var_3370_cast_fp16 = concat(axis = var_42, interleave = var_3370_interleave_0, values = (var_3364_cast_fp16, key_31_cast_fp16))[name = tensor("op_3370_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_q_weight_to_fp16 = const()[name = tensor("encoder_layers_15_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192456640)))]; + tensor linear_138_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_q_weight_to_fp16, x = key_31_cast_fp16)[name = tensor("linear_138_cast_fp16")]; + tensor var_3374 = const()[name = tensor("op_3374"), val = tensor([1, -1, 8, 64])]; + tensor q_91_cast_fp16 = reshape(shape = var_3374, x = linear_138_cast_fp16)[name = tensor("q_91_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_k_weight_to_fp16 = const()[name = tensor("encoder_layers_15_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192980992)))]; + tensor linear_139_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_k_weight_to_fp16, x = input_819_cast_fp16)[name = tensor("linear_139_cast_fp16")]; + tensor var_3378 = const()[name = tensor("op_3378"), val = tensor([1, -1, 8, 64])]; + tensor k_61_cast_fp16 = reshape(shape = var_3378, x = linear_139_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_v_weight_to_fp16 = const()[name = tensor("encoder_layers_15_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193505344)))]; + tensor linear_140_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_v_weight_to_fp16, x = input_819_cast_fp16)[name = tensor("linear_140_cast_fp16")]; + tensor var_3382 = const()[name = tensor("op_3382"), val = tensor([1, -1, 8, 64])]; + tensor v_31_cast_fp16 = reshape(shape = var_3382, x = linear_140_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor value_31_perm_0 = const()[name = tensor("value_31_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor encoder_layers_15_self_attn_pos_bias_u_to_fp16 = const()[name = tensor("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194029696)))]; + tensor var_3394_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = tensor("op_3394_cast_fp16")]; + tensor encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = tensor("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194030784)))]; + tensor var_3396_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = tensor("op_3396_cast_fp16")]; + tensor q_with_bias_v_31_perm_0 = const()[name = tensor("q_with_bias_v_31_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor x_397_transpose_x_0 = const()[name = tensor("x_397_transpose_x_0"), val = tensor(false)]; + tensor x_397_transpose_y_0 = const()[name = tensor("x_397_transpose_y_0"), val = tensor(false)]; + tensor var_3398_to_fp16 = const()[name = tensor("op_3398_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194031872)))]; + tensor q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3396_cast_fp16)[name = tensor("transpose_102")]; + tensor x_397_cast_fp16 = matmul(transpose_x = x_397_transpose_x_0, transpose_y = x_397_transpose_y_0, x = q_with_bias_v_31_cast_fp16, y = var_3398_to_fp16)[name = tensor("x_397_cast_fp16")]; + tensor x_399_pad_0 = const()[name = tensor("x_399_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor x_399_mode_0 = const()[name = tensor("x_399_mode_0"), val = tensor("constant")]; + tensor const_218_to_fp16 = const()[name = tensor("const_218_to_fp16"), val = tensor(0x0p+0)]; + tensor x_399_cast_fp16 = pad(constant_val = const_218_to_fp16, mode = x_399_mode_0, pad = x_399_pad_0, x = x_397_cast_fp16)[name = tensor("x_399_cast_fp16")]; + tensor var_3406 = const()[name = tensor("op_3406"), val = tensor([1, 8, -1, 4])]; + tensor x_401_cast_fp16 = reshape(shape = var_3406, x = x_399_cast_fp16)[name = tensor("x_401_cast_fp16")]; + tensor var_3410_begin_0 = const()[name = tensor("op_3410_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3410_end_0 = const()[name = tensor("op_3410_end_0"), val = tensor([1, 8, 148, 4])]; + tensor var_3410_end_mask_0 = const()[name = tensor("op_3410_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3410_cast_fp16 = slice_by_index(begin = var_3410_begin_0, end = var_3410_end_0, end_mask = var_3410_end_mask_0, x = x_401_cast_fp16)[name = tensor("op_3410_cast_fp16")]; + tensor var_3411 = const()[name = tensor("op_3411"), val = tensor([1, 8, 4, 147])]; + tensor matrix_bd_61_cast_fp16 = reshape(shape = var_3411, x = var_3410_cast_fp16)[name = tensor("matrix_bd_61_cast_fp16")]; + tensor matrix_ac_31_transpose_x_0 = const()[name = tensor("matrix_ac_31_transpose_x_0"), val = tensor(false)]; + tensor matrix_ac_31_transpose_y_0 = const()[name = tensor("matrix_ac_31_transpose_y_0"), val = tensor(false)]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = k_61_cast_fp16)[name = tensor("transpose_100")]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = var_3394_cast_fp16)[name = tensor("transpose_101")]; + tensor matrix_ac_31_cast_fp16 = matmul(transpose_x = matrix_ac_31_transpose_x_0, transpose_y = matrix_ac_31_transpose_y_0, x = transpose_81, y = transpose_82)[name = tensor("matrix_ac_31_cast_fp16")]; + tensor matrix_bd_63_begin_0 = const()[name = tensor("matrix_bd_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_63_end_0 = const()[name = tensor("matrix_bd_63_end_0"), val = tensor([1, 8, 4, 74])]; + tensor matrix_bd_63_end_mask_0 = const()[name = tensor("matrix_bd_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_63_cast_fp16 = slice_by_index(begin = matrix_bd_63_begin_0, end = matrix_bd_63_end_0, end_mask = matrix_bd_63_end_mask_0, x = matrix_bd_61_cast_fp16)[name = tensor("matrix_bd_63_cast_fp16")]; + tensor var_3420_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = tensor("op_3420_cast_fp16")]; + tensor _inversed_scores_61_y_0_to_fp16 = const()[name = tensor("_inversed_scores_61_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor _inversed_scores_61_cast_fp16 = mul(x = var_3420_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = tensor("_inversed_scores_61_cast_fp16")]; + tensor scores_63_cast_fp16 = select(a = var_19_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_3)[name = tensor("scores_63_cast_fp16")]; + tensor var_3426_cast_fp16 = softmax(axis = var_40, x = scores_63_cast_fp16)[name = tensor("op_3426_cast_fp16")]; + tensor input_821_cast_fp16 = select(a = var_18_to_fp16, b = var_3426_cast_fp16, cond = mask_3)[name = tensor("input_821_cast_fp16")]; + tensor x_403_transpose_x_0 = const()[name = tensor("x_403_transpose_x_0"), val = tensor(false)]; + tensor x_403_transpose_y_0 = const()[name = tensor("x_403_transpose_y_0"), val = tensor(false)]; + tensor value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = v_31_cast_fp16)[name = tensor("transpose_103")]; + tensor x_403_cast_fp16 = matmul(transpose_x = x_403_transpose_x_0, transpose_y = x_403_transpose_y_0, x = input_821_cast_fp16, y = value_31_cast_fp16)[name = tensor("x_403_cast_fp16")]; + tensor var_3430_perm_0 = const()[name = tensor("op_3430_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3431 = const()[name = tensor("op_3431"), val = tensor([1, -1, 512])]; + tensor var_3430_cast_fp16 = transpose(perm = var_3430_perm_0, x = x_403_cast_fp16)[name = tensor("transpose_99")]; + tensor input_823_cast_fp16 = reshape(shape = var_3431, x = var_3430_cast_fp16)[name = tensor("input_823_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_out_weight_to_fp16 = const()[name = tensor("encoder_layers_15_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194182464)))]; + tensor linear_142_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_out_weight_to_fp16, x = input_823_cast_fp16)[name = tensor("linear_142_cast_fp16")]; + tensor input_827_cast_fp16 = add(x = input_817_cast_fp16, y = linear_142_cast_fp16)[name = tensor("input_827_cast_fp16")]; + tensor x_407_axes_0 = const()[name = tensor("x_407_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194706816)))]; + tensor encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = tensor("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194707904)))]; + tensor x_407_cast_fp16 = layer_norm(axes = x_407_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input_827_cast_fp16)[name = tensor("x_407_cast_fp16")]; + tensor input_829_perm_0 = const()[name = tensor("input_829_perm_0"), val = tensor([0, 2, 1])]; + tensor input_831_pad_type_0 = const()[name = tensor("input_831_pad_type_0"), val = tensor("valid")]; + tensor input_831_strides_0 = const()[name = tensor("input_831_strides_0"), val = tensor([1])]; + tensor input_831_pad_0 = const()[name = tensor("input_831_pad_0"), val = tensor([0, 0])]; + tensor input_831_dilations_0 = const()[name = tensor("input_831_dilations_0"), val = tensor([1])]; + tensor input_831_groups_0 = const()[name = tensor("input_831_groups_0"), val = tensor(1)]; + tensor encoder_layers_15_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194708992)))]; + tensor input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = tensor("transpose_98")]; + tensor input_831_cast_fp16 = conv(dilations = input_831_dilations_0, groups = input_831_groups_0, pad = input_831_pad_0, pad_type = input_831_pad_type_0, strides = input_831_strides_0, weight = encoder_layers_15_conv_pointwise_conv1_weight_to_fp16, x = input_829_cast_fp16)[name = tensor("input_831_cast_fp16")]; + tensor x_409_split_num_splits_0 = const()[name = tensor("x_409_split_num_splits_0"), val = tensor(2)]; + tensor x_409_split_axis_0 = const()[name = tensor("x_409_split_axis_0"), val = tensor(1)]; + tensor x_409_split_cast_fp16_0, tensor x_409_split_cast_fp16_1 = split(axis = x_409_split_axis_0, num_splits = x_409_split_num_splits_0, x = input_831_cast_fp16)[name = tensor("x_409_split_cast_fp16")]; + tensor x_409_split_1_sigmoid_cast_fp16 = sigmoid(x = x_409_split_cast_fp16_1)[name = tensor("x_409_split_1_sigmoid_cast_fp16")]; + tensor x_409_cast_fp16 = mul(x = x_409_split_cast_fp16_0, y = x_409_split_1_sigmoid_cast_fp16)[name = tensor("x_409_cast_fp16")]; + tensor input_833_cast_fp16 = select(a = var_18_to_fp16, b = x_409_cast_fp16, cond = var_396)[name = tensor("input_833_cast_fp16")]; + tensor new_x_63_interleave_0 = const()[name = tensor("new_x_63_interleave_0"), val = tensor(false)]; + tensor new_x_63_cast_fp16 = concat(axis = var_40, interleave = new_x_63_interleave_0, values = (cache_63_cast_fp16, input_833_cast_fp16))[name = tensor("new_x_63_cast_fp16")]; + tensor var_3469_begin_0 = const()[name = tensor("op_3469_begin_0"), val = tensor([0, 0, 4])]; + tensor var_3469_end_0 = const()[name = tensor("op_3469_end_0"), val = tensor([1, 512, 12])]; + tensor var_3469_end_mask_0 = const()[name = tensor("op_3469_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3469_cast_fp16 = slice_by_index(begin = var_3469_begin_0, end = var_3469_end_0, end_mask = var_3469_end_mask_0, x = new_x_63_cast_fp16)[name = tensor("op_3469_cast_fp16")]; + tensor x_411_pad_type_0 = const()[name = tensor("x_411_pad_type_0"), val = tensor("valid")]; + tensor x_411_groups_0 = const()[name = tensor("x_411_groups_0"), val = tensor(512)]; + tensor x_411_strides_0 = const()[name = tensor("x_411_strides_0"), val = tensor([1])]; + tensor x_411_pad_0 = const()[name = tensor("x_411_pad_0"), val = tensor([0, 0])]; + tensor x_411_dilations_0 = const()[name = tensor("x_411_dilations_0"), val = tensor([1])]; + tensor encoder_layers_15_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_15_conv_depthwise_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195757632)))]; + tensor x_411_cast_fp16 = conv(dilations = x_411_dilations_0, groups = x_411_groups_0, pad = x_411_pad_0, pad_type = x_411_pad_type_0, strides = x_411_strides_0, weight = encoder_layers_15_conv_depthwise_conv_weight_to_fp16, x = new_x_63_cast_fp16)[name = tensor("x_411_cast_fp16")]; + tensor input_835_perm_0 = const()[name = tensor("input_835_perm_0"), val = tensor([0, 2, 1])]; + tensor x_413_axes_0 = const()[name = tensor("x_413_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = tensor("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195766912)))]; + tensor encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = tensor("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195768000)))]; + tensor input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = tensor("transpose_97")]; + tensor x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input_835_cast_fp16)[name = tensor("x_413_cast_fp16")]; + tensor input_837_perm_0 = const()[name = tensor("input_837_perm_0"), val = tensor([0, 2, 1])]; + tensor input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = tensor("transpose_96")]; + tensor input_839_cast_fp16 = silu(x = input_837_cast_fp16)[name = tensor("input_839_cast_fp16")]; + tensor x_415_pad_type_0 = const()[name = tensor("x_415_pad_type_0"), val = tensor("valid")]; + tensor x_415_strides_0 = const()[name = tensor("x_415_strides_0"), val = tensor([1])]; + tensor x_415_pad_0 = const()[name = tensor("x_415_pad_0"), val = tensor([0, 0])]; + tensor x_415_dilations_0 = const()[name = tensor("x_415_dilations_0"), val = tensor([1])]; + tensor x_415_groups_0 = const()[name = tensor("x_415_groups_0"), val = tensor(1)]; + tensor encoder_layers_15_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195769088)))]; + tensor x_415_cast_fp16 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = encoder_layers_15_conv_pointwise_conv2_weight_to_fp16, x = input_839_cast_fp16)[name = tensor("x_415_cast_fp16")]; + tensor input_841_perm_0 = const()[name = tensor("input_841_perm_0"), val = tensor([0, 2, 1])]; + tensor input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = tensor("transpose_95")]; + tensor input_843_cast_fp16 = add(x = input_827_cast_fp16, y = input_841_cast_fp16)[name = tensor("input_843_cast_fp16")]; + tensor input_845_axes_0 = const()[name = tensor("input_845_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = tensor("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196293440)))]; + tensor encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = tensor("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196294528)))]; + tensor input_845_cast_fp16 = layer_norm(axes = input_845_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input_843_cast_fp16)[name = tensor("input_845_cast_fp16")]; + tensor encoder_layers_15_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_15_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196295616)))]; + tensor linear_143_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_15_feed_forward2_linear1_weight_to_fp16, x = input_845_cast_fp16)[name = tensor("linear_143_cast_fp16")]; + tensor input_849_cast_fp16 = silu(x = linear_143_cast_fp16)[name = tensor("input_849_cast_fp16")]; + tensor encoder_layers_15_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_15_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198392832)))]; + tensor linear_144_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_feed_forward2_linear2_weight_to_fp16, x = input_849_cast_fp16)[name = tensor("linear_144_cast_fp16")]; + tensor var_3510_to_fp16 = const()[name = tensor("op_3510_to_fp16"), val = tensor(0x1p-1)]; + tensor var_3511_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3510_to_fp16)[name = tensor("op_3511_cast_fp16")]; + tensor input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3511_cast_fp16)[name = tensor("input_855_cast_fp16")]; + tensor input_857_axes_0 = const()[name = tensor("input_857_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_out_weight_to_fp16 = const()[name = tensor("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200490048)))]; + tensor encoder_layers_15_norm_out_bias_to_fp16 = const()[name = tensor("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200491136)))]; + tensor input_857_cast_fp16 = layer_norm(axes = input_857_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input_855_cast_fp16)[name = tensor("input_857_cast_fp16")]; + tensor cache_65_begin_0 = const()[name = tensor("cache_65_begin_0"), val = tensor([16, 0, 0, 0])]; + tensor cache_65_end_0 = const()[name = tensor("cache_65_end_0"), val = tensor([17, 1, 70, 512])]; + tensor cache_65_end_mask_0 = const()[name = tensor("cache_65_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_65_squeeze_mask_0 = const()[name = tensor("cache_65_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_65_cast_fp16 = slice_by_index(begin = cache_65_begin_0, end = cache_65_end_0, end_mask = cache_65_end_mask_0, squeeze_mask = cache_65_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor("cache_65_cast_fp16")]; + tensor cache_begin_0 = const()[name = tensor("cache_begin_0"), val = tensor([16, 0, 0, 0])]; + tensor cache_end_0 = const()[name = tensor("cache_end_0"), val = tensor([17, 1, 512, 8])]; + tensor cache_end_mask_0 = const()[name = tensor("cache_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_squeeze_mask_0 = const()[name = tensor("cache_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_cast_fp16 = slice_by_index(begin = cache_begin_0, end = cache_end_0, end_mask = cache_end_mask_0, squeeze_mask = cache_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor("cache_cast_fp16")]; + tensor input_859_axes_0 = const()[name = tensor("input_859_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = tensor("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200492224)))]; + tensor encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = tensor("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200493312)))]; + tensor input_859_cast_fp16 = layer_norm(axes = input_859_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input_857_cast_fp16)[name = tensor("input_859_cast_fp16")]; + tensor encoder_layers_16_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_16_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200494400)))]; + tensor linear_145_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_16_feed_forward1_linear1_weight_to_fp16, x = input_859_cast_fp16)[name = tensor("linear_145_cast_fp16")]; + tensor input_863_cast_fp16 = silu(x = linear_145_cast_fp16)[name = tensor("input_863_cast_fp16")]; + tensor encoder_layers_16_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_16_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202591616)))]; + tensor linear_146_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_feed_forward1_linear2_weight_to_fp16, x = input_863_cast_fp16)[name = tensor("linear_146_cast_fp16")]; + tensor var_3545_to_fp16 = const()[name = tensor("op_3545_to_fp16"), val = tensor(0x1p-1)]; + tensor var_3546_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3545_to_fp16)[name = tensor("op_3546_cast_fp16")]; + tensor input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3546_cast_fp16)[name = tensor("input_869_cast_fp16")]; + tensor key_axes_0 = const()[name = tensor("key_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = tensor("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204688832)))]; + tensor encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = tensor("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204689920)))]; + tensor key_cast_fp16 = layer_norm(axes = key_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_869_cast_fp16)[name = tensor("key_cast_fp16")]; + tensor input_871_interleave_0 = const()[name = tensor("input_871_interleave_0"), val = tensor(false)]; + tensor input_871_cast_fp16 = concat(axis = var_42, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_cast_fp16))[name = tensor("input_871_cast_fp16")]; + tensor var_3568_begin_0 = const()[name = tensor("op_3568_begin_0"), val = tensor([0, 4, 0])]; + tensor var_3568_end_0 = const()[name = tensor("op_3568_end_0"), val = tensor([1, 70, 512])]; + tensor var_3568_end_mask_0 = const()[name = tensor("op_3568_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3568_cast_fp16 = slice_by_index(begin = var_3568_begin_0, end = var_3568_end_0, end_mask = var_3568_end_mask_0, x = cache_65_cast_fp16)[name = tensor("op_3568_cast_fp16")]; + tensor cache_last_channel_cur_interleave_0 = const()[name = tensor("cache_last_channel_cur_interleave_0"), val = tensor(false)]; + tensor cache_last_channel_cur_cast_fp16 = concat(axis = var_42, interleave = cache_last_channel_cur_interleave_0, values = (var_3568_cast_fp16, key_cast_fp16))[name = tensor("cache_last_channel_cur_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_q_weight_to_fp16 = const()[name = tensor("encoder_layers_16_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204691008)))]; + tensor linear_147_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_q_weight_to_fp16, x = key_cast_fp16)[name = tensor("linear_147_cast_fp16")]; + tensor var_3578 = const()[name = tensor("op_3578"), val = tensor([1, -1, 8, 64])]; + tensor q_97_cast_fp16 = reshape(shape = var_3578, x = linear_147_cast_fp16)[name = tensor("q_97_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_k_weight_to_fp16 = const()[name = tensor("encoder_layers_16_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205215360)))]; + tensor linear_148_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_k_weight_to_fp16, x = input_871_cast_fp16)[name = tensor("linear_148_cast_fp16")]; + tensor var_3582 = const()[name = tensor("op_3582"), val = tensor([1, -1, 8, 64])]; + tensor k_65_cast_fp16 = reshape(shape = var_3582, x = linear_148_cast_fp16)[name = tensor("k_65_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_v_weight_to_fp16 = const()[name = tensor("encoder_layers_16_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205739712)))]; + tensor linear_149_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_v_weight_to_fp16, x = input_871_cast_fp16)[name = tensor("linear_149_cast_fp16")]; + tensor var_3586 = const()[name = tensor("op_3586"), val = tensor([1, -1, 8, 64])]; + tensor v_cast_fp16 = reshape(shape = var_3586, x = linear_149_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor value_perm_0 = const()[name = tensor("value_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor encoder_layers_16_self_attn_pos_bias_u_to_fp16 = const()[name = tensor("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206264064)))]; + tensor var_3598_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = tensor("op_3598_cast_fp16")]; + tensor encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = tensor("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206265152)))]; + tensor var_3600_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = tensor("op_3600_cast_fp16")]; + tensor q_with_bias_v_perm_0 = const()[name = tensor("q_with_bias_v_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor x_423_transpose_x_0 = const()[name = tensor("x_423_transpose_x_0"), val = tensor(false)]; + tensor x_423_transpose_y_0 = const()[name = tensor("x_423_transpose_y_0"), val = tensor(false)]; + tensor var_3602_to_fp16 = const()[name = tensor("op_3602_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206266240)))]; + tensor q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_3600_cast_fp16)[name = tensor("transpose_93")]; + tensor x_423_cast_fp16 = matmul(transpose_x = x_423_transpose_x_0, transpose_y = x_423_transpose_y_0, x = q_with_bias_v_cast_fp16, y = var_3602_to_fp16)[name = tensor("x_423_cast_fp16")]; + tensor x_425_pad_0 = const()[name = tensor("x_425_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor x_425_mode_0 = const()[name = tensor("x_425_mode_0"), val = tensor("constant")]; + tensor const_231_to_fp16 = const()[name = tensor("const_231_to_fp16"), val = tensor(0x0p+0)]; + tensor x_425_cast_fp16 = pad(constant_val = const_231_to_fp16, mode = x_425_mode_0, pad = x_425_pad_0, x = x_423_cast_fp16)[name = tensor("x_425_cast_fp16")]; + tensor var_3610 = const()[name = tensor("op_3610"), val = tensor([1, 8, -1, 4])]; + tensor x_427_cast_fp16 = reshape(shape = var_3610, x = x_425_cast_fp16)[name = tensor("x_427_cast_fp16")]; + tensor var_3614_begin_0 = const()[name = tensor("op_3614_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3614_end_0 = const()[name = tensor("op_3614_end_0"), val = tensor([1, 8, 148, 4])]; + tensor var_3614_end_mask_0 = const()[name = tensor("op_3614_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3614_cast_fp16 = slice_by_index(begin = var_3614_begin_0, end = var_3614_end_0, end_mask = var_3614_end_mask_0, x = x_427_cast_fp16)[name = tensor("op_3614_cast_fp16")]; + tensor var_3615 = const()[name = tensor("op_3615"), val = tensor([1, 8, 4, 147])]; + tensor matrix_bd_65_cast_fp16 = reshape(shape = var_3615, x = var_3614_cast_fp16)[name = tensor("matrix_bd_65_cast_fp16")]; + tensor matrix_ac_transpose_x_0 = const()[name = tensor("matrix_ac_transpose_x_0"), val = tensor(false)]; + tensor matrix_ac_transpose_y_0 = const()[name = tensor("matrix_ac_transpose_y_0"), val = tensor(false)]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = k_65_cast_fp16)[name = tensor("transpose_91")]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = var_3598_cast_fp16)[name = tensor("transpose_92")]; + tensor matrix_ac_cast_fp16 = matmul(transpose_x = matrix_ac_transpose_x_0, transpose_y = matrix_ac_transpose_y_0, x = transpose_83, y = transpose_84)[name = tensor("matrix_ac_cast_fp16")]; + tensor matrix_bd_begin_0 = const()[name = tensor("matrix_bd_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_end_0 = const()[name = tensor("matrix_bd_end_0"), val = tensor([1, 8, 4, 74])]; + tensor matrix_bd_end_mask_0 = const()[name = tensor("matrix_bd_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_cast_fp16 = slice_by_index(begin = matrix_bd_begin_0, end = matrix_bd_end_0, end_mask = matrix_bd_end_mask_0, x = matrix_bd_65_cast_fp16)[name = tensor("matrix_bd_cast_fp16")]; + tensor var_3624_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = tensor("op_3624_cast_fp16")]; + tensor _inversed_scores_65_y_0_to_fp16 = const()[name = tensor("_inversed_scores_65_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor _inversed_scores_65_cast_fp16 = mul(x = var_3624_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = tensor("_inversed_scores_65_cast_fp16")]; + tensor scores_cast_fp16 = select(a = var_19_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_3)[name = tensor("scores_cast_fp16")]; + tensor var_3630_cast_fp16 = softmax(axis = var_40, x = scores_cast_fp16)[name = tensor("op_3630_cast_fp16")]; + tensor input_873_cast_fp16 = select(a = var_18_to_fp16, b = var_3630_cast_fp16, cond = mask_3)[name = tensor("input_873_cast_fp16")]; + tensor x_429_transpose_x_0 = const()[name = tensor("x_429_transpose_x_0"), val = tensor(false)]; + tensor x_429_transpose_y_0 = const()[name = tensor("x_429_transpose_y_0"), val = tensor(false)]; + tensor value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = tensor("transpose_94")]; + tensor x_429_cast_fp16 = matmul(transpose_x = x_429_transpose_x_0, transpose_y = x_429_transpose_y_0, x = input_873_cast_fp16, y = value_cast_fp16)[name = tensor("x_429_cast_fp16")]; + tensor var_3634_perm_0 = const()[name = tensor("op_3634_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3635 = const()[name = tensor("op_3635"), val = tensor([1, -1, 512])]; + tensor var_3634_cast_fp16 = transpose(perm = var_3634_perm_0, x = x_429_cast_fp16)[name = tensor("transpose_90")]; + tensor input_875_cast_fp16 = reshape(shape = var_3635, x = var_3634_cast_fp16)[name = tensor("input_875_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_out_weight_to_fp16 = const()[name = tensor("encoder_layers_16_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206416832)))]; + tensor linear_151_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_out_weight_to_fp16, x = input_875_cast_fp16)[name = tensor("linear_151_cast_fp16")]; + tensor input_879_cast_fp16 = add(x = input_869_cast_fp16, y = linear_151_cast_fp16)[name = tensor("input_879_cast_fp16")]; + tensor x_433_axes_0 = const()[name = tensor("x_433_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206941184)))]; + tensor encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = tensor("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206942272)))]; + tensor x_433_cast_fp16 = layer_norm(axes = x_433_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input_879_cast_fp16)[name = tensor("x_433_cast_fp16")]; + tensor input_881_perm_0 = const()[name = tensor("input_881_perm_0"), val = tensor([0, 2, 1])]; + tensor input_883_pad_type_0 = const()[name = tensor("input_883_pad_type_0"), val = tensor("valid")]; + tensor input_883_strides_0 = const()[name = tensor("input_883_strides_0"), val = tensor([1])]; + tensor input_883_pad_0 = const()[name = tensor("input_883_pad_0"), val = tensor([0, 0])]; + tensor input_883_dilations_0 = const()[name = tensor("input_883_dilations_0"), val = tensor([1])]; + tensor input_883_groups_0 = const()[name = tensor("input_883_groups_0"), val = tensor(1)]; + tensor encoder_layers_16_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206943360)))]; + tensor input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = tensor("transpose_89")]; + tensor input_883_cast_fp16 = conv(dilations = input_883_dilations_0, groups = input_883_groups_0, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = input_883_strides_0, weight = encoder_layers_16_conv_pointwise_conv1_weight_to_fp16, x = input_881_cast_fp16)[name = tensor("input_883_cast_fp16")]; + tensor x_435_split_num_splits_0 = const()[name = tensor("x_435_split_num_splits_0"), val = tensor(2)]; + tensor x_435_split_axis_0 = const()[name = tensor("x_435_split_axis_0"), val = tensor(1)]; + tensor x_435_split_cast_fp16_0, tensor x_435_split_cast_fp16_1 = split(axis = x_435_split_axis_0, num_splits = x_435_split_num_splits_0, x = input_883_cast_fp16)[name = tensor("x_435_split_cast_fp16")]; + tensor x_435_split_1_sigmoid_cast_fp16 = sigmoid(x = x_435_split_cast_fp16_1)[name = tensor("x_435_split_1_sigmoid_cast_fp16")]; + tensor x_435_cast_fp16 = mul(x = x_435_split_cast_fp16_0, y = x_435_split_1_sigmoid_cast_fp16)[name = tensor("x_435_cast_fp16")]; + tensor input_885_cast_fp16 = select(a = var_18_to_fp16, b = x_435_cast_fp16, cond = var_396)[name = tensor("input_885_cast_fp16")]; + tensor new_x_interleave_0 = const()[name = tensor("new_x_interleave_0"), val = tensor(false)]; + tensor new_x_cast_fp16 = concat(axis = var_40, interleave = new_x_interleave_0, values = (cache_cast_fp16, input_885_cast_fp16))[name = tensor("new_x_cast_fp16")]; + tensor cache_last_time_cur_begin_0 = const()[name = tensor("cache_last_time_cur_begin_0"), val = tensor([0, 0, 4])]; + tensor cache_last_time_cur_end_0 = const()[name = tensor("cache_last_time_cur_end_0"), val = tensor([1, 512, 12])]; + tensor cache_last_time_cur_end_mask_0 = const()[name = tensor("cache_last_time_cur_end_mask_0"), val = tensor([true, true, true])]; + tensor cache_last_time_cur_cast_fp16 = slice_by_index(begin = cache_last_time_cur_begin_0, end = cache_last_time_cur_end_0, end_mask = cache_last_time_cur_end_mask_0, x = new_x_cast_fp16)[name = tensor("cache_last_time_cur_cast_fp16")]; + tensor x_437_pad_type_0 = const()[name = tensor("x_437_pad_type_0"), val = tensor("valid")]; + tensor x_437_groups_0 = const()[name = tensor("x_437_groups_0"), val = tensor(512)]; + tensor x_437_strides_0 = const()[name = tensor("x_437_strides_0"), val = tensor([1])]; + tensor x_437_pad_0 = const()[name = tensor("x_437_pad_0"), val = tensor([0, 0])]; + tensor x_437_dilations_0 = const()[name = tensor("x_437_dilations_0"), val = tensor([1])]; + tensor encoder_layers_16_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor("encoder_layers_16_conv_depthwise_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207992000)))]; + tensor x_437_cast_fp16 = conv(dilations = x_437_dilations_0, groups = x_437_groups_0, pad = x_437_pad_0, pad_type = x_437_pad_type_0, strides = x_437_strides_0, weight = encoder_layers_16_conv_depthwise_conv_weight_to_fp16, x = new_x_cast_fp16)[name = tensor("x_437_cast_fp16")]; + tensor input_887_perm_0 = const()[name = tensor("input_887_perm_0"), val = tensor([0, 2, 1])]; + tensor x_439_axes_0 = const()[name = tensor("x_439_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = tensor("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208001280)))]; + tensor encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = tensor("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208002368)))]; + tensor input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = tensor("transpose_88")]; + tensor x_439_cast_fp16 = layer_norm(axes = x_439_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input_887_cast_fp16)[name = tensor("x_439_cast_fp16")]; + tensor input_889_perm_0 = const()[name = tensor("input_889_perm_0"), val = tensor([0, 2, 1])]; + tensor input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = tensor("transpose_87")]; + tensor input_891_cast_fp16 = silu(x = input_889_cast_fp16)[name = tensor("input_891_cast_fp16")]; + tensor x_441_pad_type_0 = const()[name = tensor("x_441_pad_type_0"), val = tensor("valid")]; + tensor x_441_strides_0 = const()[name = tensor("x_441_strides_0"), val = tensor([1])]; + tensor x_441_pad_0 = const()[name = tensor("x_441_pad_0"), val = tensor([0, 0])]; + tensor x_441_dilations_0 = const()[name = tensor("x_441_dilations_0"), val = tensor([1])]; + tensor x_441_groups_0 = const()[name = tensor("x_441_groups_0"), val = tensor(1)]; + tensor encoder_layers_16_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208003456)))]; + tensor x_441_cast_fp16 = conv(dilations = x_441_dilations_0, groups = x_441_groups_0, pad = x_441_pad_0, pad_type = x_441_pad_type_0, strides = x_441_strides_0, weight = encoder_layers_16_conv_pointwise_conv2_weight_to_fp16, x = input_891_cast_fp16)[name = tensor("x_441_cast_fp16")]; + tensor input_893_perm_0 = const()[name = tensor("input_893_perm_0"), val = tensor([0, 2, 1])]; + tensor input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = tensor("transpose_86")]; + tensor input_895_cast_fp16 = add(x = input_879_cast_fp16, y = input_893_cast_fp16)[name = tensor("input_895_cast_fp16")]; + tensor input_897_axes_0 = const()[name = tensor("input_897_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = tensor("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208527808)))]; + tensor encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = tensor("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208528896)))]; + tensor input_897_cast_fp16 = layer_norm(axes = input_897_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input_895_cast_fp16)[name = tensor("input_897_cast_fp16")]; + tensor encoder_layers_16_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor("encoder_layers_16_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208529984)))]; + tensor linear_152_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_16_feed_forward2_linear1_weight_to_fp16, x = input_897_cast_fp16)[name = tensor("linear_152_cast_fp16")]; + tensor input_901_cast_fp16 = silu(x = linear_152_cast_fp16)[name = tensor("input_901_cast_fp16")]; + tensor encoder_layers_16_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor("encoder_layers_16_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210627200)))]; + tensor linear_153_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_feed_forward2_linear2_weight_to_fp16, x = input_901_cast_fp16)[name = tensor("linear_153_cast_fp16")]; + tensor var_3714_to_fp16 = const()[name = tensor("op_3714_to_fp16"), val = tensor(0x1p-1)]; + tensor var_3715_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_3714_to_fp16)[name = tensor("op_3715_cast_fp16")]; + tensor input_cast_fp16 = add(x = input_895_cast_fp16, y = var_3715_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor audio_signal_axes_0 = const()[name = tensor("audio_signal_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_out_weight_to_fp16 = const()[name = tensor("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212724416)))]; + tensor encoder_layers_16_norm_out_bias_to_fp16 = const()[name = tensor("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212725504)))]; + tensor audio_signal_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_16_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input_cast_fp16)[name = tensor("audio_signal_cast_fp16")]; + tensor obj_1_perm_0 = const()[name = tensor("obj_1_perm_0"), val = tensor([0, 2, 1])]; + tensor obj_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("obj_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor cast_178_dtype_0 = const()[name = tensor("cast_178_dtype_0"), val = tensor("int32")]; + tensor obj_5_axis_0 = const()[name = tensor("obj_5_axis_0"), val = tensor(0)]; + tensor obj_5_cast_fp16 = stack(axis = obj_5_axis_0, values = (var_310_cast_fp16, var_514_cast_fp16, var_718_cast_fp16, var_922_cast_fp16, var_1126_cast_fp16, var_1330_cast_fp16, var_1534_cast_fp16, var_1738_cast_fp16, var_1942_cast_fp16, var_2146_cast_fp16, var_2350_cast_fp16, var_2554_cast_fp16, var_2758_cast_fp16, var_2962_cast_fp16, var_3166_cast_fp16, var_3370_cast_fp16, cache_last_channel_cur_cast_fp16))[name = tensor("obj_5_cast_fp16")]; + tensor obj_7_axis_0 = const()[name = tensor("obj_7_axis_0"), val = tensor(0)]; + tensor obj_7_cast_fp16 = stack(axis = obj_7_axis_0, values = (var_409_cast_fp16, var_613_cast_fp16, var_817_cast_fp16, var_1021_cast_fp16, var_1225_cast_fp16, var_1429_cast_fp16, var_1633_cast_fp16, var_1837_cast_fp16, var_2041_cast_fp16, var_2245_cast_fp16, var_2449_cast_fp16, var_2653_cast_fp16, var_2857_cast_fp16, var_3061_cast_fp16, var_3265_cast_fp16, var_3469_cast_fp16, cache_last_time_cur_cast_fp16))[name = tensor("obj_7_cast_fp16")]; + tensor obj_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("obj_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3731 = add(x = cache_last_channel_len, y = max_audio_length_1)[name = tensor("op_3731")]; + tensor var_3731_promoted_to_fp16_dtype_0 = const()[name = tensor("op_3731_promoted_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor const_237_to_fp16 = const()[name = tensor("const_237_to_fp16"), val = tensor(-inf)]; + tensor var_23_promoted_to_fp16 = const()[name = tensor("op_23_promoted_to_fp16"), val = tensor(0x1.18p+6)]; + tensor var_3731_to_fp16 = cast(dtype = var_3731_promoted_to_fp16_dtype_0, x = var_3731)[name = tensor("cast_182")]; + tensor clip_1_cast_fp16 = clip(alpha = const_237_to_fp16, beta = var_23_promoted_to_fp16, x = var_3731_to_fp16)[name = tensor("clip_1_cast_fp16")]; + tensor new_channel_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_channel_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor cast_179_dtype_0 = const()[name = tensor("cast_179_dtype_0"), val = tensor("int32")]; + tensor cache_last_channel_len_out = cast(dtype = cast_179_dtype_0, x = clip_1_cast_fp16)[name = tensor("cast_180")]; + tensor cache_last_channel_out = cast(dtype = new_channel_cache_cast_fp16_to_fp32_dtype_0, x = obj_5_cast_fp16)[name = tensor("cast_181")]; + tensor cache_last_time_out = cast(dtype = obj_7_cast_fp16_to_fp32_dtype_0, x = obj_7_cast_fp16)[name = tensor("cast_183")]; + tensor encoder_length = cast(dtype = cast_178_dtype_0, x = clip_0_cast_fp16)[name = tensor("cast_184")]; + tensor obj_1_cast_fp16 = transpose(perm = obj_1_perm_0, x = audio_signal_cast_fp16)[name = tensor("transpose_85")]; + tensor encoder = cast(dtype = obj_1_cast_fp16_to_fp32_dtype_0, x = obj_1_cast_fp16)[name = tensor("cast_185")]; + } -> (encoder, encoder_length, cache_last_channel_out, cache_last_time_out, cache_last_channel_len_out); +} \ No newline at end of file