| program(1.0) |
| [buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})] |
| { |
| func main<ios17>(tensor<fp32, [1, 80000]> audio) { |
| tensor<int32, [1]> input_1_axes_0 = const()[name = tensor<string, []>("input_1_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<string, []> audio_to_fp16_dtype_0 = const()[name = tensor<string, []>("audio_to_fp16_dtype_0"), val = tensor<string, []>("fp16")]; |
| tensor<fp16, [1, 80000]> audio_to_fp16 = cast(dtype = audio_to_fp16_dtype_0, x = audio)[name = tensor<string, []>("cast_0")]; |
| tensor<fp16, [1, 1, 80000]> input_1_cast_fp16 = expand_dims(axes = input_1_axes_0, x = audio_to_fp16)[name = tensor<string, []>("input_1_cast_fp16")]; |
| tensor<string, []> seqs_1_pad_type_0 = const()[name = tensor<string, []>("seqs_1_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> seqs_1_strides_0 = const()[name = tensor<string, []>("seqs_1_strides_0"), val = tensor<int32, [1]>([5])]; |
| tensor<int32, [2]> seqs_1_pad_0 = const()[name = tensor<string, []>("seqs_1_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> seqs_1_dilations_0 = const()[name = tensor<string, []>("seqs_1_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> seqs_1_groups_0 = const()[name = tensor<string, []>("seqs_1_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [512, 1, 10]> model_encoder_frontend_feature_extractor_layers_0_conv_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5248))), name = tensor<string, []>("model_encoder_frontend_feature_extractor_layers_0_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([512, 1, 10])]; |
| tensor<fp16, [512]> model_encoder_frontend_feature_extractor_layers_0_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_frontend_feature_extractor_layers_0_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5824)))]; |
| tensor<fp16, [1, 512, 15999]> seqs_1_cast_fp16 = conv(bias = model_encoder_frontend_feature_extractor_layers_0_conv_bias_to_fp16, dilations = seqs_1_dilations_0, groups = seqs_1_groups_0, pad = seqs_1_pad_0, pad_type = seqs_1_pad_type_0, strides = seqs_1_strides_0, weight = model_encoder_frontend_feature_extractor_layers_0_conv_weight_to_fp16_palettized, x = input_1_cast_fp16)[name = tensor<string, []>("seqs_1_cast_fp16")]; |
| tensor<int32, [3]> x_1_perm_0 = const()[name = tensor<string, []>("x_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [1]> x_3_axes_0 = const()[name = tensor<string, []>("x_3_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [512]> const_0_to_fp16 = const()[name = tensor<string, []>("const_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6912)))]; |
| tensor<fp16, [512]> const_1_to_fp16 = const()[name = tensor<string, []>("const_1_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8000)))]; |
| tensor<fp16, []> var_18_to_fp16 = const()[name = tensor<string, []>("op_18_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| tensor<fp16, [1, 15999, 512]> x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = seqs_1_cast_fp16)[name = tensor<string, []>("transpose_254")]; |
| tensor<fp16, [1, 15999, 512]> x_3_cast_fp16 = layer_norm(axes = x_3_axes_0, beta = const_1_to_fp16, epsilon = var_18_to_fp16, gamma = const_0_to_fp16, x = x_1_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")]; |
| tensor<int32, [3]> input_5_perm_0 = const()[name = tensor<string, []>("input_5_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> input_7_mode_0 = const()[name = tensor<string, []>("input_7_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 512, 15999]> input_5_cast_fp16 = transpose(perm = input_5_perm_0, x = x_3_cast_fp16)[name = tensor<string, []>("transpose_253")]; |
| tensor<fp16, [1, 512, 15999]> input_7_cast_fp16 = gelu(mode = input_7_mode_0, x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")]; |
| tensor<string, []> seqs_5_pad_type_0 = const()[name = tensor<string, []>("seqs_5_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> seqs_5_strides_0 = const()[name = tensor<string, []>("seqs_5_strides_0"), val = tensor<int32, [1]>([2])]; |
| tensor<int32, [2]> seqs_5_pad_0 = const()[name = tensor<string, []>("seqs_5_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> seqs_5_dilations_0 = const()[name = tensor<string, []>("seqs_5_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> seqs_5_groups_0 = const()[name = tensor<string, []>("seqs_5_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [512, 512, 3]> model_encoder_frontend_feature_extractor_layers_1_conv_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9088))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(795584))), name = tensor<string, []>("model_encoder_frontend_feature_extractor_layers_1_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([512, 512, 3])]; |
| tensor<fp16, [512]> model_encoder_frontend_feature_extractor_layers_1_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_frontend_feature_extractor_layers_1_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(796160)))]; |
| tensor<fp16, [1, 512, 7999]> seqs_5_cast_fp16 = conv(bias = model_encoder_frontend_feature_extractor_layers_1_conv_bias_to_fp16, dilations = seqs_5_dilations_0, groups = seqs_5_groups_0, pad = seqs_5_pad_0, pad_type = seqs_5_pad_type_0, strides = seqs_5_strides_0, weight = model_encoder_frontend_feature_extractor_layers_1_conv_weight_to_fp16_palettized, x = input_7_cast_fp16)[name = tensor<string, []>("seqs_5_cast_fp16")]; |
| tensor<int32, [3]> x_5_perm_0 = const()[name = tensor<string, []>("x_5_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [1]> x_7_axes_0 = const()[name = tensor<string, []>("x_7_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [512]> const_2_to_fp16 = const()[name = tensor<string, []>("const_2_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(797248)))]; |
| tensor<fp16, [512]> const_3_to_fp16 = const()[name = tensor<string, []>("const_3_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(798336)))]; |
| tensor<fp16, [1, 7999, 512]> x_5_cast_fp16 = transpose(perm = x_5_perm_0, x = seqs_5_cast_fp16)[name = tensor<string, []>("transpose_252")]; |
| tensor<fp16, [1, 7999, 512]> x_7_cast_fp16 = layer_norm(axes = x_7_axes_0, beta = const_3_to_fp16, epsilon = var_18_to_fp16, gamma = const_2_to_fp16, x = x_5_cast_fp16)[name = tensor<string, []>("x_7_cast_fp16")]; |
| tensor<int32, [3]> input_11_perm_0 = const()[name = tensor<string, []>("input_11_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> input_13_mode_0 = const()[name = tensor<string, []>("input_13_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 512, 7999]> input_11_cast_fp16 = transpose(perm = input_11_perm_0, x = x_7_cast_fp16)[name = tensor<string, []>("transpose_251")]; |
| tensor<fp16, [1, 512, 7999]> input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = input_11_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")]; |
| tensor<string, []> seqs_9_pad_type_0 = const()[name = tensor<string, []>("seqs_9_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> seqs_9_strides_0 = const()[name = tensor<string, []>("seqs_9_strides_0"), val = tensor<int32, [1]>([2])]; |
| tensor<int32, [2]> seqs_9_pad_0 = const()[name = tensor<string, []>("seqs_9_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> seqs_9_dilations_0 = const()[name = tensor<string, []>("seqs_9_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> seqs_9_groups_0 = const()[name = tensor<string, []>("seqs_9_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [512, 512, 3]> model_encoder_frontend_feature_extractor_layers_2_conv_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(799424))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1585920))), name = tensor<string, []>("model_encoder_frontend_feature_extractor_layers_2_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([512, 512, 3])]; |
| tensor<fp16, [512]> model_encoder_frontend_feature_extractor_layers_2_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_frontend_feature_extractor_layers_2_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1586496)))]; |
| tensor<fp16, [1, 512, 3999]> seqs_9_cast_fp16 = conv(bias = model_encoder_frontend_feature_extractor_layers_2_conv_bias_to_fp16, dilations = seqs_9_dilations_0, groups = seqs_9_groups_0, pad = seqs_9_pad_0, pad_type = seqs_9_pad_type_0, strides = seqs_9_strides_0, weight = model_encoder_frontend_feature_extractor_layers_2_conv_weight_to_fp16_palettized, x = input_13_cast_fp16)[name = tensor<string, []>("seqs_9_cast_fp16")]; |
| tensor<int32, [3]> x_9_perm_0 = const()[name = tensor<string, []>("x_9_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [1]> x_11_axes_0 = const()[name = tensor<string, []>("x_11_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [512]> const_4_to_fp16 = const()[name = tensor<string, []>("const_4_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1587584)))]; |
| tensor<fp16, [512]> const_5_to_fp16 = const()[name = tensor<string, []>("const_5_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1588672)))]; |
| tensor<fp16, [1, 3999, 512]> x_9_cast_fp16 = transpose(perm = x_9_perm_0, x = seqs_9_cast_fp16)[name = tensor<string, []>("transpose_250")]; |
| tensor<fp16, [1, 3999, 512]> x_11_cast_fp16 = layer_norm(axes = x_11_axes_0, beta = const_5_to_fp16, epsilon = var_18_to_fp16, gamma = const_4_to_fp16, x = x_9_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")]; |
| tensor<int32, [3]> input_17_perm_0 = const()[name = tensor<string, []>("input_17_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> input_19_mode_0 = const()[name = tensor<string, []>("input_19_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 512, 3999]> input_17_cast_fp16 = transpose(perm = input_17_perm_0, x = x_11_cast_fp16)[name = tensor<string, []>("transpose_249")]; |
| tensor<fp16, [1, 512, 3999]> input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")]; |
| tensor<string, []> seqs_13_pad_type_0 = const()[name = tensor<string, []>("seqs_13_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> seqs_13_strides_0 = const()[name = tensor<string, []>("seqs_13_strides_0"), val = tensor<int32, [1]>([2])]; |
| tensor<int32, [2]> seqs_13_pad_0 = const()[name = tensor<string, []>("seqs_13_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> seqs_13_dilations_0 = const()[name = tensor<string, []>("seqs_13_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> seqs_13_groups_0 = const()[name = tensor<string, []>("seqs_13_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [512, 512, 3]> model_encoder_frontend_feature_extractor_layers_3_conv_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1589760))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2376256))), name = tensor<string, []>("model_encoder_frontend_feature_extractor_layers_3_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([512, 512, 3])]; |
| tensor<fp16, [512]> model_encoder_frontend_feature_extractor_layers_3_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_frontend_feature_extractor_layers_3_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2376832)))]; |
| tensor<fp16, [1, 512, 1999]> seqs_13_cast_fp16 = conv(bias = model_encoder_frontend_feature_extractor_layers_3_conv_bias_to_fp16, dilations = seqs_13_dilations_0, groups = seqs_13_groups_0, pad = seqs_13_pad_0, pad_type = seqs_13_pad_type_0, strides = seqs_13_strides_0, weight = model_encoder_frontend_feature_extractor_layers_3_conv_weight_to_fp16_palettized, x = input_19_cast_fp16)[name = tensor<string, []>("seqs_13_cast_fp16")]; |
| tensor<int32, [3]> x_13_perm_0 = const()[name = tensor<string, []>("x_13_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [1]> x_15_axes_0 = const()[name = tensor<string, []>("x_15_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [512]> const_6_to_fp16 = const()[name = tensor<string, []>("const_6_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2377920)))]; |
| tensor<fp16, [512]> const_7_to_fp16 = const()[name = tensor<string, []>("const_7_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2379008)))]; |
| tensor<fp16, [1, 1999, 512]> x_13_cast_fp16 = transpose(perm = x_13_perm_0, x = seqs_13_cast_fp16)[name = tensor<string, []>("transpose_248")]; |
| tensor<fp16, [1, 1999, 512]> x_15_cast_fp16 = layer_norm(axes = x_15_axes_0, beta = const_7_to_fp16, epsilon = var_18_to_fp16, gamma = const_6_to_fp16, x = x_13_cast_fp16)[name = tensor<string, []>("x_15_cast_fp16")]; |
| tensor<int32, [3]> input_23_perm_0 = const()[name = tensor<string, []>("input_23_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> input_25_mode_0 = const()[name = tensor<string, []>("input_25_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 512, 1999]> input_23_cast_fp16 = transpose(perm = input_23_perm_0, x = x_15_cast_fp16)[name = tensor<string, []>("transpose_247")]; |
| tensor<fp16, [1, 512, 1999]> input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = input_23_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")]; |
| tensor<string, []> seqs_17_pad_type_0 = const()[name = tensor<string, []>("seqs_17_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> seqs_17_strides_0 = const()[name = tensor<string, []>("seqs_17_strides_0"), val = tensor<int32, [1]>([2])]; |
| tensor<int32, [2]> seqs_17_pad_0 = const()[name = tensor<string, []>("seqs_17_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> seqs_17_dilations_0 = const()[name = tensor<string, []>("seqs_17_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> seqs_17_groups_0 = const()[name = tensor<string, []>("seqs_17_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [512, 512, 3]> model_encoder_frontend_feature_extractor_layers_4_conv_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2380096))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3166592))), name = tensor<string, []>("model_encoder_frontend_feature_extractor_layers_4_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([512, 512, 3])]; |
| tensor<fp16, [512]> model_encoder_frontend_feature_extractor_layers_4_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_frontend_feature_extractor_layers_4_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3167168)))]; |
| tensor<fp16, [1, 512, 999]> seqs_17_cast_fp16 = conv(bias = model_encoder_frontend_feature_extractor_layers_4_conv_bias_to_fp16, dilations = seqs_17_dilations_0, groups = seqs_17_groups_0, pad = seqs_17_pad_0, pad_type = seqs_17_pad_type_0, strides = seqs_17_strides_0, weight = model_encoder_frontend_feature_extractor_layers_4_conv_weight_to_fp16_palettized, x = input_25_cast_fp16)[name = tensor<string, []>("seqs_17_cast_fp16")]; |
| tensor<int32, [3]> x_17_perm_0 = const()[name = tensor<string, []>("x_17_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [1]> x_19_axes_0 = const()[name = tensor<string, []>("x_19_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [512]> const_8_to_fp16 = const()[name = tensor<string, []>("const_8_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3168256)))]; |
| tensor<fp16, [512]> const_9_to_fp16 = const()[name = tensor<string, []>("const_9_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3169344)))]; |
| tensor<fp16, [1, 999, 512]> x_17_cast_fp16 = transpose(perm = x_17_perm_0, x = seqs_17_cast_fp16)[name = tensor<string, []>("transpose_246")]; |
| tensor<fp16, [1, 999, 512]> x_19_cast_fp16 = layer_norm(axes = x_19_axes_0, beta = const_9_to_fp16, epsilon = var_18_to_fp16, gamma = const_8_to_fp16, x = x_17_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")]; |
| tensor<int32, [3]> input_29_perm_0 = const()[name = tensor<string, []>("input_29_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> input_31_mode_0 = const()[name = tensor<string, []>("input_31_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 512, 999]> input_29_cast_fp16 = transpose(perm = input_29_perm_0, x = x_19_cast_fp16)[name = tensor<string, []>("transpose_245")]; |
| tensor<fp16, [1, 512, 999]> input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")]; |
| tensor<string, []> seqs_21_pad_type_0 = const()[name = tensor<string, []>("seqs_21_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> seqs_21_strides_0 = const()[name = tensor<string, []>("seqs_21_strides_0"), val = tensor<int32, [1]>([2])]; |
| tensor<int32, [2]> seqs_21_pad_0 = const()[name = tensor<string, []>("seqs_21_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> seqs_21_dilations_0 = const()[name = tensor<string, []>("seqs_21_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> seqs_21_groups_0 = const()[name = tensor<string, []>("seqs_21_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [512, 512, 2]> model_encoder_frontend_feature_extractor_layers_5_conv_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [524288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3170432))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3694784))), name = tensor<string, []>("model_encoder_frontend_feature_extractor_layers_5_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([512, 512, 2])]; |
| tensor<fp16, [512]> model_encoder_frontend_feature_extractor_layers_5_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_frontend_feature_extractor_layers_5_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3695360)))]; |
| tensor<fp16, [1, 512, 499]> seqs_21_cast_fp16 = conv(bias = model_encoder_frontend_feature_extractor_layers_5_conv_bias_to_fp16, dilations = seqs_21_dilations_0, groups = seqs_21_groups_0, pad = seqs_21_pad_0, pad_type = seqs_21_pad_type_0, strides = seqs_21_strides_0, weight = model_encoder_frontend_feature_extractor_layers_5_conv_weight_to_fp16_palettized, x = input_31_cast_fp16)[name = tensor<string, []>("seqs_21_cast_fp16")]; |
| tensor<int32, [3]> x_21_perm_0 = const()[name = tensor<string, []>("x_21_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [1]> x_23_axes_0 = const()[name = tensor<string, []>("x_23_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [512]> const_10_to_fp16 = const()[name = tensor<string, []>("const_10_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3696448)))]; |
| tensor<fp16, [512]> const_11_to_fp16 = const()[name = tensor<string, []>("const_11_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3697536)))]; |
| tensor<fp16, [1, 499, 512]> x_21_cast_fp16 = transpose(perm = x_21_perm_0, x = seqs_21_cast_fp16)[name = tensor<string, []>("transpose_244")]; |
| tensor<fp16, [1, 499, 512]> x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = const_11_to_fp16, epsilon = var_18_to_fp16, gamma = const_10_to_fp16, x = x_21_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")]; |
| tensor<int32, [3]> input_35_perm_0 = const()[name = tensor<string, []>("input_35_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> input_37_mode_0 = const()[name = tensor<string, []>("input_37_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 512, 499]> input_35_cast_fp16 = transpose(perm = input_35_perm_0, x = x_23_cast_fp16)[name = tensor<string, []>("transpose_243")]; |
| tensor<fp16, [1, 512, 499]> input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = input_35_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")]; |
| tensor<string, []> seqs_25_pad_type_0 = const()[name = tensor<string, []>("seqs_25_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> seqs_25_strides_0 = const()[name = tensor<string, []>("seqs_25_strides_0"), val = tensor<int32, [1]>([2])]; |
| tensor<int32, [2]> seqs_25_pad_0 = const()[name = tensor<string, []>("seqs_25_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> seqs_25_dilations_0 = const()[name = tensor<string, []>("seqs_25_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> seqs_25_groups_0 = const()[name = tensor<string, []>("seqs_25_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp16, [512, 512, 2]> model_encoder_frontend_feature_extractor_layers_6_conv_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [524288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3698624))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4222976))), name = tensor<string, []>("model_encoder_frontend_feature_extractor_layers_6_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([512, 512, 2])]; |
| tensor<fp16, [512]> model_encoder_frontend_feature_extractor_layers_6_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_frontend_feature_extractor_layers_6_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4223552)))]; |
| tensor<fp16, [1, 512, 249]> seqs_25_cast_fp16 = conv(bias = model_encoder_frontend_feature_extractor_layers_6_conv_bias_to_fp16, dilations = seqs_25_dilations_0, groups = seqs_25_groups_0, pad = seqs_25_pad_0, pad_type = seqs_25_pad_type_0, strides = seqs_25_strides_0, weight = model_encoder_frontend_feature_extractor_layers_6_conv_weight_to_fp16_palettized, x = input_37_cast_fp16)[name = tensor<string, []>("seqs_25_cast_fp16")]; |
| tensor<int32, [3]> x_25_perm_0 = const()[name = tensor<string, []>("x_25_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [1]> x_27_axes_0 = const()[name = tensor<string, []>("x_27_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [512]> const_12_to_fp16 = const()[name = tensor<string, []>("const_12_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4224640)))]; |
| tensor<fp16, [512]> const_13_to_fp16 = const()[name = tensor<string, []>("const_13_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4225728)))]; |
| tensor<fp16, [1, 249, 512]> x_25_cast_fp16 = transpose(perm = x_25_perm_0, x = seqs_25_cast_fp16)[name = tensor<string, []>("transpose_242")]; |
| tensor<fp16, [1, 249, 512]> x_27_cast_fp16 = layer_norm(axes = x_27_axes_0, beta = const_13_to_fp16, epsilon = var_18_to_fp16, gamma = const_12_to_fp16, x = x_25_cast_fp16)[name = tensor<string, []>("x_27_cast_fp16")]; |
| tensor<string, []> seqs_29_mode_0 = const()[name = tensor<string, []>("seqs_29_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 512]> seqs_29_cast_fp16 = gelu(mode = seqs_29_mode_0, x = x_27_cast_fp16)[name = tensor<string, []>("seqs_29_cast_fp16")]; |
| tensor<int32, [1]> x_29_axes_0 = const()[name = tensor<string, []>("x_29_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [512]> model_encoder_frontend_post_extract_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_frontend_post_extract_layer_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4226816)))]; |
| tensor<fp16, [512]> model_encoder_frontend_post_extract_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_frontend_post_extract_layer_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4227904)))]; |
| tensor<fp16, [1, 249, 512]> x_29_cast_fp16 = layer_norm(axes = x_29_axes_0, beta = model_encoder_frontend_post_extract_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_frontend_post_extract_layer_norm_weight_to_fp16, x = seqs_29_cast_fp16)[name = tensor<string, []>("x_29_cast_fp16")]; |
| tensor<fp16, [1024, 512]> model_encoder_frontend_model_dim_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [524288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4228992))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4753344))), name = tensor<string, []>("model_encoder_frontend_model_dim_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 512])]; |
| tensor<fp16, [1024]> model_encoder_frontend_model_dim_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_frontend_model_dim_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4753920)))]; |
| tensor<fp16, [1, 249, 1024]> linear_0_cast_fp16 = linear(bias = model_encoder_frontend_model_dim_proj_bias_to_fp16, weight = model_encoder_frontend_model_dim_proj_weight_to_fp16_palettized, x = x_29_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")]; |
| tensor<int32, [3]> input_43_perm_0 = const()[name = tensor<string, []>("input_43_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> encodings_1_pad_type_0 = const()[name = tensor<string, []>("encodings_1_pad_type_0"), val = tensor<string, []>("custom")]; |
| tensor<int32, [2]> encodings_1_pad_0 = const()[name = tensor<string, []>("encodings_1_pad_0"), val = tensor<int32, [2]>([64, 64])]; |
| tensor<int32, []> encodings_1_groups_0 = const()[name = tensor<string, []>("encodings_1_groups_0"), val = tensor<int32, []>(16)]; |
| tensor<int32, [1]> encodings_1_strides_0 = const()[name = tensor<string, []>("encodings_1_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [1]> encodings_1_dilations_0 = const()[name = tensor<string, []>("encodings_1_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [1024, 64, 128]> weight_31_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [8388608]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4756032))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13144704))), name = tensor<string, []>("weight_31_to_fp16_palettized"), shape = tensor<uint32, [3]>([1024, 64, 128])]; |
| tensor<fp16, [1024]> model_encoder_frontend_pos_encoder_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_frontend_pos_encoder_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13145280)))]; |
| tensor<fp16, [1, 1024, 249]> input_43_cast_fp16 = transpose(perm = input_43_perm_0, x = linear_0_cast_fp16)[name = tensor<string, []>("transpose_241")]; |
| tensor<fp16, [1, 1024, 250]> encodings_1_cast_fp16 = conv(bias = model_encoder_frontend_pos_encoder_conv_bias_to_fp16, dilations = encodings_1_dilations_0, groups = encodings_1_groups_0, pad = encodings_1_pad_0, pad_type = encodings_1_pad_type_0, strides = encodings_1_strides_0, weight = weight_31_to_fp16_palettized, x = input_43_cast_fp16)[name = tensor<string, []>("encodings_1_cast_fp16")]; |
| tensor<int32, [3]> input_45_begin_0 = const()[name = tensor<string, []>("input_45_begin_0"), val = tensor<int32, [3]>([0, 0, 0])]; |
| tensor<int32, [3]> input_45_end_0 = const()[name = tensor<string, []>("input_45_end_0"), val = tensor<int32, [3]>([1, 1024, 249])]; |
| tensor<bool, [3]> input_45_end_mask_0 = const()[name = tensor<string, []>("input_45_end_mask_0"), val = tensor<bool, [3]>([true, true, false])]; |
| tensor<fp16, [1, 1024, 249]> input_45_cast_fp16 = slice_by_index(begin = input_45_begin_0, end = input_45_end_0, end_mask = input_45_end_mask_0, x = encodings_1_cast_fp16)[name = tensor<string, []>("input_45_cast_fp16")]; |
| tensor<string, []> encodings_3_mode_0 = const()[name = tensor<string, []>("encodings_3_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 1024, 249]> encodings_3_cast_fp16 = gelu(mode = encodings_3_mode_0, x = input_45_cast_fp16)[name = tensor<string, []>("encodings_3_cast_fp16")]; |
| tensor<int32, [3]> encodings_perm_0 = const()[name = tensor<string, []>("encodings_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp16, [1, 249, 1024]> encodings_cast_fp16 = transpose(perm = encodings_perm_0, x = encodings_3_cast_fp16)[name = tensor<string, []>("transpose_240")]; |
| tensor<fp16, [1, 249, 1024]> input_47_cast_fp16 = add(x = linear_0_cast_fp16, y = encodings_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")]; |
| tensor<int32, [1]> x_31_axes_0 = const()[name = tensor<string, []>("x_31_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_0_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13147392)))]; |
| tensor<fp16, [1024]> model_encoder_layers_0_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13149504)))]; |
| tensor<fp16, [1, 249, 1024]> x_31_cast_fp16 = layer_norm(axes = x_31_axes_0, beta = model_encoder_layers_0_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_0_self_attn_layer_norm_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("x_31_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_0_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13151616))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14200256))), name = tensor<string, []>("model_encoder_layers_0_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14200832)))]; |
| tensor<fp16, [1, 249, 1024]> linear_1_cast_fp16 = linear(bias = model_encoder_layers_0_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_0_self_attn_q_proj_weight_to_fp16_palettized, x = x_31_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")]; |
| tensor<int32, [4]> concat_0 = const()[name = tensor<string, []>("concat_0"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_1_cast_fp16 = reshape(shape = concat_0, x = linear_1_cast_fp16)[name = tensor<string, []>("q_1_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_0_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14202944))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15251584))), name = tensor<string, []>("model_encoder_layers_0_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_0_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15252160)))]; |
| tensor<fp16, [1, 249, 1024]> linear_2_cast_fp16 = linear(bias = model_encoder_layers_0_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_0_self_attn_k_proj_weight_to_fp16_palettized, x = x_31_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_0_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15254272))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16302912))), name = tensor<string, []>("model_encoder_layers_0_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16303488)))]; |
| tensor<fp16, [1, 249, 1024]> linear_3_cast_fp16 = linear(bias = model_encoder_layers_0_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_0_self_attn_v_proj_weight_to_fp16_palettized, x = x_31_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")]; |
| tensor<int32, [4]> concat_1 = const()[name = tensor<string, []>("concat_1"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_1_cast_fp16 = reshape(shape = concat_1, x = linear_2_cast_fp16)[name = tensor<string, []>("k_1_cast_fp16")]; |
| tensor<int32, [4]> concat_2 = const()[name = tensor<string, []>("concat_2"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_3_cast_fp16 = reshape(shape = concat_2, x = linear_3_cast_fp16)[name = tensor<string, []>("v_3_cast_fp16")]; |
| tensor<int32, [4]> v_5_perm_0 = const()[name = tensor<string, []>("v_5_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_0_y_0_to_fp16 = const()[name = tensor<string, []>("mul_0_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_0_cast_fp16 = mul(x = q_1_cast_fp16, y = mul_0_y_0_to_fp16)[name = tensor<string, []>("mul_0_cast_fp16")]; |
| tensor<bool, []> matmul_0_transpose_y_0 = const()[name = tensor<string, []>("matmul_0_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_0_transpose_x_0 = const()[name = tensor<string, []>("matmul_0_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_96_perm_0 = const()[name = tensor<string, []>("transpose_96_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_97_perm_0 = const()[name = tensor<string, []>("transpose_97_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_97 = transpose(perm = transpose_97_perm_0, x = k_1_cast_fp16)[name = tensor<string, []>("transpose_238")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_96 = transpose(perm = transpose_96_perm_0, x = mul_0_cast_fp16)[name = tensor<string, []>("transpose_239")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = transpose_96, y = transpose_97)[name = tensor<string, []>("matmul_0_cast_fp16")]; |
| tensor<int32, []> softmax_0_axis_0 = const()[name = tensor<string, []>("softmax_0_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = matmul_0_cast_fp16)[name = tensor<string, []>("softmax_0_cast_fp16")]; |
| tensor<bool, []> attns_1_transpose_x_0 = const()[name = tensor<string, []>("attns_1_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_1_transpose_y_0 = const()[name = tensor<string, []>("attns_1_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_5_cast_fp16 = transpose(perm = v_5_perm_0, x = v_3_cast_fp16)[name = tensor<string, []>("transpose_237")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_1_cast_fp16 = matmul(transpose_x = attns_1_transpose_x_0, transpose_y = attns_1_transpose_y_0, x = softmax_0_cast_fp16, y = v_5_cast_fp16)[name = tensor<string, []>("attns_1_cast_fp16")]; |
| tensor<int32, [4]> attns_3_perm_0 = const()[name = tensor<string, []>("attns_3_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_3 = const()[name = tensor<string, []>("concat_3"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_3_cast_fp16 = transpose(perm = attns_3_perm_0, x = attns_1_cast_fp16)[name = tensor<string, []>("transpose_236")]; |
| tensor<fp16, [1, 249, 1024]> x_33_cast_fp16 = reshape(shape = concat_3, x = attns_3_cast_fp16)[name = tensor<string, []>("x_33_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_0_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16305600))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17354240))), name = tensor<string, []>("model_encoder_layers_0_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_0_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17354816)))]; |
| tensor<fp16, [1, 249, 1024]> linear_4_cast_fp16 = linear(bias = model_encoder_layers_0_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_0_self_attn_output_proj_weight_to_fp16_palettized, x = x_33_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_49_cast_fp16 = add(x = linear_4_cast_fp16, y = input_47_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")]; |
| tensor<int32, [1]> x_35_axes_0 = const()[name = tensor<string, []>("x_35_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_0_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17356928)))]; |
| tensor<fp16, [1024]> model_encoder_layers_0_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17359040)))]; |
| tensor<fp16, [1, 249, 1024]> x_35_cast_fp16 = layer_norm(axes = x_35_axes_0, beta = model_encoder_layers_0_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_0_ffn_layer_norm_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("x_35_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_0_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17361152))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21555520))), name = tensor<string, []>("model_encoder_layers_0_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_0_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21556096)))]; |
| tensor<fp16, [1, 249, 4096]> linear_5_cast_fp16 = linear(bias = model_encoder_layers_0_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_0_ffn_inner_proj_weight_to_fp16_palettized, x = x_35_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")]; |
| tensor<string, []> input_53_mode_0 = const()[name = tensor<string, []>("input_53_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_53_cast_fp16 = gelu(mode = input_53_mode_0, x = linear_5_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_0_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21564352))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25758720))), name = tensor<string, []>("model_encoder_layers_0_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_0_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25759296)))]; |
| tensor<fp16, [1, 249, 1024]> linear_6_cast_fp16 = linear(bias = model_encoder_layers_0_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_0_ffn_output_proj_weight_to_fp16_palettized, x = input_53_cast_fp16)[name = tensor<string, []>("linear_6_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_55_cast_fp16 = add(x = linear_6_cast_fp16, y = input_49_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")]; |
| tensor<int32, [1]> x_39_axes_0 = const()[name = tensor<string, []>("x_39_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_1_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25761408)))]; |
| tensor<fp16, [1024]> model_encoder_layers_1_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25763520)))]; |
| tensor<fp16, [1, 249, 1024]> x_39_cast_fp16 = layer_norm(axes = x_39_axes_0, beta = model_encoder_layers_1_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_1_self_attn_layer_norm_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("x_39_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_1_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25765632))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26814272))), name = tensor<string, []>("model_encoder_layers_1_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26814848)))]; |
| tensor<fp16, [1, 249, 1024]> linear_7_cast_fp16 = linear(bias = model_encoder_layers_1_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_1_self_attn_q_proj_weight_to_fp16_palettized, x = x_39_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")]; |
| tensor<int32, [4]> concat_4 = const()[name = tensor<string, []>("concat_4"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_5_cast_fp16 = reshape(shape = concat_4, x = linear_7_cast_fp16)[name = tensor<string, []>("q_5_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_1_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26816960))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27865600))), name = tensor<string, []>("model_encoder_layers_1_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_1_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27866176)))]; |
| tensor<fp16, [1, 249, 1024]> linear_8_cast_fp16 = linear(bias = model_encoder_layers_1_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_1_self_attn_k_proj_weight_to_fp16_palettized, x = x_39_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_1_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27868288))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28916928))), name = tensor<string, []>("model_encoder_layers_1_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28917504)))]; |
| tensor<fp16, [1, 249, 1024]> linear_9_cast_fp16 = linear(bias = model_encoder_layers_1_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_1_self_attn_v_proj_weight_to_fp16_palettized, x = x_39_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")]; |
| tensor<int32, [4]> concat_5 = const()[name = tensor<string, []>("concat_5"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_5_cast_fp16 = reshape(shape = concat_5, x = linear_8_cast_fp16)[name = tensor<string, []>("k_5_cast_fp16")]; |
| tensor<int32, [4]> concat_6 = const()[name = tensor<string, []>("concat_6"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_7_cast_fp16 = reshape(shape = concat_6, x = linear_9_cast_fp16)[name = tensor<string, []>("v_7_cast_fp16")]; |
| tensor<int32, [4]> v_9_perm_0 = const()[name = tensor<string, []>("v_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_1_y_0_to_fp16 = const()[name = tensor<string, []>("mul_1_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_1_cast_fp16 = mul(x = q_5_cast_fp16, y = mul_1_y_0_to_fp16)[name = tensor<string, []>("mul_1_cast_fp16")]; |
| tensor<bool, []> matmul_1_transpose_y_0 = const()[name = tensor<string, []>("matmul_1_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_1_transpose_x_0 = const()[name = tensor<string, []>("matmul_1_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_98_perm_0 = const()[name = tensor<string, []>("transpose_98_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_99_perm_0 = const()[name = tensor<string, []>("transpose_99_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_99 = transpose(perm = transpose_99_perm_0, x = k_5_cast_fp16)[name = tensor<string, []>("transpose_234")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_98 = transpose(perm = transpose_98_perm_0, x = mul_1_cast_fp16)[name = tensor<string, []>("transpose_235")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = transpose_98, y = transpose_99)[name = tensor<string, []>("matmul_1_cast_fp16")]; |
| tensor<int32, []> softmax_1_axis_0 = const()[name = tensor<string, []>("softmax_1_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = matmul_1_cast_fp16)[name = tensor<string, []>("softmax_1_cast_fp16")]; |
| tensor<bool, []> attns_5_transpose_x_0 = const()[name = tensor<string, []>("attns_5_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_5_transpose_y_0 = const()[name = tensor<string, []>("attns_5_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_9_cast_fp16 = transpose(perm = v_9_perm_0, x = v_7_cast_fp16)[name = tensor<string, []>("transpose_233")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_5_cast_fp16 = matmul(transpose_x = attns_5_transpose_x_0, transpose_y = attns_5_transpose_y_0, x = softmax_1_cast_fp16, y = v_9_cast_fp16)[name = tensor<string, []>("attns_5_cast_fp16")]; |
| tensor<int32, [4]> attns_7_perm_0 = const()[name = tensor<string, []>("attns_7_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_7 = const()[name = tensor<string, []>("concat_7"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_7_cast_fp16 = transpose(perm = attns_7_perm_0, x = attns_5_cast_fp16)[name = tensor<string, []>("transpose_232")]; |
| tensor<fp16, [1, 249, 1024]> x_41_cast_fp16 = reshape(shape = concat_7, x = attns_7_cast_fp16)[name = tensor<string, []>("x_41_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_1_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28919616))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29968256))), name = tensor<string, []>("model_encoder_layers_1_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_1_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29968832)))]; |
| tensor<fp16, [1, 249, 1024]> linear_10_cast_fp16 = linear(bias = model_encoder_layers_1_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_1_self_attn_output_proj_weight_to_fp16_palettized, x = x_41_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_57_cast_fp16 = add(x = linear_10_cast_fp16, y = input_55_cast_fp16)[name = tensor<string, []>("input_57_cast_fp16")]; |
| tensor<int32, [1]> x_43_axes_0 = const()[name = tensor<string, []>("x_43_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_1_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29970944)))]; |
| tensor<fp16, [1024]> model_encoder_layers_1_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29973056)))]; |
| tensor<fp16, [1, 249, 1024]> x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = model_encoder_layers_1_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_1_ffn_layer_norm_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("x_43_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_1_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29975168))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34169536))), name = tensor<string, []>("model_encoder_layers_1_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_1_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34170112)))]; |
| tensor<fp16, [1, 249, 4096]> linear_11_cast_fp16 = linear(bias = model_encoder_layers_1_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_1_ffn_inner_proj_weight_to_fp16_palettized, x = x_43_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")]; |
| tensor<string, []> input_61_mode_0 = const()[name = tensor<string, []>("input_61_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_11_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_1_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34178368))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38372736))), name = tensor<string, []>("model_encoder_layers_1_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_1_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38373312)))]; |
| tensor<fp16, [1, 249, 1024]> linear_12_cast_fp16 = linear(bias = model_encoder_layers_1_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_1_ffn_output_proj_weight_to_fp16_palettized, x = input_61_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_63_cast_fp16 = add(x = linear_12_cast_fp16, y = input_57_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")]; |
| tensor<int32, [1]> x_47_axes_0 = const()[name = tensor<string, []>("x_47_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_2_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_2_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38375424)))]; |
| tensor<fp16, [1024]> model_encoder_layers_2_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_2_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38377536)))]; |
| tensor<fp16, [1, 249, 1024]> x_47_cast_fp16 = layer_norm(axes = x_47_axes_0, beta = model_encoder_layers_2_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_2_self_attn_layer_norm_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("x_47_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_2_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38379648))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39428288))), name = tensor<string, []>("model_encoder_layers_2_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39428864)))]; |
| tensor<fp16, [1, 249, 1024]> linear_13_cast_fp16 = linear(bias = model_encoder_layers_2_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_2_self_attn_q_proj_weight_to_fp16_palettized, x = x_47_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")]; |
| tensor<int32, [4]> concat_8 = const()[name = tensor<string, []>("concat_8"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_9_cast_fp16 = reshape(shape = concat_8, x = linear_13_cast_fp16)[name = tensor<string, []>("q_9_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_2_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39430976))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40479616))), name = tensor<string, []>("model_encoder_layers_2_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_2_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_2_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40480192)))]; |
| tensor<fp16, [1, 249, 1024]> linear_14_cast_fp16 = linear(bias = model_encoder_layers_2_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_2_self_attn_k_proj_weight_to_fp16_palettized, x = x_47_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_2_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40482304))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41530944))), name = tensor<string, []>("model_encoder_layers_2_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41531520)))]; |
| tensor<fp16, [1, 249, 1024]> linear_15_cast_fp16 = linear(bias = model_encoder_layers_2_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_2_self_attn_v_proj_weight_to_fp16_palettized, x = x_47_cast_fp16)[name = tensor<string, []>("linear_15_cast_fp16")]; |
| tensor<int32, [4]> concat_9 = const()[name = tensor<string, []>("concat_9"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_9_cast_fp16 = reshape(shape = concat_9, x = linear_14_cast_fp16)[name = tensor<string, []>("k_9_cast_fp16")]; |
| tensor<int32, [4]> concat_10 = const()[name = tensor<string, []>("concat_10"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_11_cast_fp16 = reshape(shape = concat_10, x = linear_15_cast_fp16)[name = tensor<string, []>("v_11_cast_fp16")]; |
| tensor<int32, [4]> v_13_perm_0 = const()[name = tensor<string, []>("v_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_2_y_0_to_fp16 = const()[name = tensor<string, []>("mul_2_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_2_cast_fp16 = mul(x = q_9_cast_fp16, y = mul_2_y_0_to_fp16)[name = tensor<string, []>("mul_2_cast_fp16")]; |
| tensor<bool, []> matmul_2_transpose_y_0 = const()[name = tensor<string, []>("matmul_2_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_2_transpose_x_0 = const()[name = tensor<string, []>("matmul_2_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_100_perm_0 = const()[name = tensor<string, []>("transpose_100_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_101_perm_0 = const()[name = tensor<string, []>("transpose_101_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_101 = transpose(perm = transpose_101_perm_0, x = k_9_cast_fp16)[name = tensor<string, []>("transpose_230")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_100 = transpose(perm = transpose_100_perm_0, x = mul_2_cast_fp16)[name = tensor<string, []>("transpose_231")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = transpose_100, y = transpose_101)[name = tensor<string, []>("matmul_2_cast_fp16")]; |
| tensor<int32, []> softmax_2_axis_0 = const()[name = tensor<string, []>("softmax_2_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = matmul_2_cast_fp16)[name = tensor<string, []>("softmax_2_cast_fp16")]; |
| tensor<bool, []> attns_9_transpose_x_0 = const()[name = tensor<string, []>("attns_9_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_9_transpose_y_0 = const()[name = tensor<string, []>("attns_9_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_13_cast_fp16 = transpose(perm = v_13_perm_0, x = v_11_cast_fp16)[name = tensor<string, []>("transpose_229")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_9_cast_fp16 = matmul(transpose_x = attns_9_transpose_x_0, transpose_y = attns_9_transpose_y_0, x = softmax_2_cast_fp16, y = v_13_cast_fp16)[name = tensor<string, []>("attns_9_cast_fp16")]; |
| tensor<int32, [4]> attns_11_perm_0 = const()[name = tensor<string, []>("attns_11_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_11 = const()[name = tensor<string, []>("concat_11"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_11_cast_fp16 = transpose(perm = attns_11_perm_0, x = attns_9_cast_fp16)[name = tensor<string, []>("transpose_228")]; |
| tensor<fp16, [1, 249, 1024]> x_49_cast_fp16 = reshape(shape = concat_11, x = attns_11_cast_fp16)[name = tensor<string, []>("x_49_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_2_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41533632))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42582272))), name = tensor<string, []>("model_encoder_layers_2_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_2_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_2_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42582848)))]; |
| tensor<fp16, [1, 249, 1024]> linear_16_cast_fp16 = linear(bias = model_encoder_layers_2_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_2_self_attn_output_proj_weight_to_fp16_palettized, x = x_49_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_65_cast_fp16 = add(x = linear_16_cast_fp16, y = input_63_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")]; |
| tensor<int32, [1]> x_51_axes_0 = const()[name = tensor<string, []>("x_51_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_2_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_2_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42584960)))]; |
| tensor<fp16, [1024]> model_encoder_layers_2_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_2_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42587072)))]; |
| tensor<fp16, [1, 249, 1024]> x_51_cast_fp16 = layer_norm(axes = x_51_axes_0, beta = model_encoder_layers_2_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_2_ffn_layer_norm_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("x_51_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_2_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42589184))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46783552))), name = tensor<string, []>("model_encoder_layers_2_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_2_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_2_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46784128)))]; |
| tensor<fp16, [1, 249, 4096]> linear_17_cast_fp16 = linear(bias = model_encoder_layers_2_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_2_ffn_inner_proj_weight_to_fp16_palettized, x = x_51_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")]; |
| tensor<string, []> input_69_mode_0 = const()[name = tensor<string, []>("input_69_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_69_cast_fp16 = gelu(mode = input_69_mode_0, x = linear_17_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_2_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46792384))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50986752))), name = tensor<string, []>("model_encoder_layers_2_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_2_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_2_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50987328)))]; |
| tensor<fp16, [1, 249, 1024]> linear_18_cast_fp16 = linear(bias = model_encoder_layers_2_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_2_ffn_output_proj_weight_to_fp16_palettized, x = input_69_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_71_cast_fp16 = add(x = linear_18_cast_fp16, y = input_65_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")]; |
| tensor<int32, [1]> x_55_axes_0 = const()[name = tensor<string, []>("x_55_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_3_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_3_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50989440)))]; |
| tensor<fp16, [1024]> model_encoder_layers_3_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_3_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50991552)))]; |
| tensor<fp16, [1, 249, 1024]> x_55_cast_fp16 = layer_norm(axes = x_55_axes_0, beta = model_encoder_layers_3_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_3_self_attn_layer_norm_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("x_55_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_3_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50993664))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52042304))), name = tensor<string, []>("model_encoder_layers_3_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52042880)))]; |
| tensor<fp16, [1, 249, 1024]> linear_19_cast_fp16 = linear(bias = model_encoder_layers_3_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_3_self_attn_q_proj_weight_to_fp16_palettized, x = x_55_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")]; |
| tensor<int32, [4]> concat_12 = const()[name = tensor<string, []>("concat_12"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_13_cast_fp16 = reshape(shape = concat_12, x = linear_19_cast_fp16)[name = tensor<string, []>("q_13_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_3_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52044992))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53093632))), name = tensor<string, []>("model_encoder_layers_3_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_3_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_3_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53094208)))]; |
| tensor<fp16, [1, 249, 1024]> linear_20_cast_fp16 = linear(bias = model_encoder_layers_3_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_3_self_attn_k_proj_weight_to_fp16_palettized, x = x_55_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_3_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53096320))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54144960))), name = tensor<string, []>("model_encoder_layers_3_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54145536)))]; |
| tensor<fp16, [1, 249, 1024]> linear_21_cast_fp16 = linear(bias = model_encoder_layers_3_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_3_self_attn_v_proj_weight_to_fp16_palettized, x = x_55_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")]; |
| tensor<int32, [4]> concat_13 = const()[name = tensor<string, []>("concat_13"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_13_cast_fp16 = reshape(shape = concat_13, x = linear_20_cast_fp16)[name = tensor<string, []>("k_13_cast_fp16")]; |
| tensor<int32, [4]> concat_14 = const()[name = tensor<string, []>("concat_14"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_15_cast_fp16 = reshape(shape = concat_14, x = linear_21_cast_fp16)[name = tensor<string, []>("v_15_cast_fp16")]; |
| tensor<int32, [4]> v_17_perm_0 = const()[name = tensor<string, []>("v_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_3_y_0_to_fp16 = const()[name = tensor<string, []>("mul_3_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_3_cast_fp16 = mul(x = q_13_cast_fp16, y = mul_3_y_0_to_fp16)[name = tensor<string, []>("mul_3_cast_fp16")]; |
| tensor<bool, []> matmul_3_transpose_y_0 = const()[name = tensor<string, []>("matmul_3_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_3_transpose_x_0 = const()[name = tensor<string, []>("matmul_3_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_102_perm_0 = const()[name = tensor<string, []>("transpose_102_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_103_perm_0 = const()[name = tensor<string, []>("transpose_103_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_103 = transpose(perm = transpose_103_perm_0, x = k_13_cast_fp16)[name = tensor<string, []>("transpose_226")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_102 = transpose(perm = transpose_102_perm_0, x = mul_3_cast_fp16)[name = tensor<string, []>("transpose_227")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = transpose_102, y = transpose_103)[name = tensor<string, []>("matmul_3_cast_fp16")]; |
| tensor<int32, []> softmax_3_axis_0 = const()[name = tensor<string, []>("softmax_3_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = matmul_3_cast_fp16)[name = tensor<string, []>("softmax_3_cast_fp16")]; |
| tensor<bool, []> attns_13_transpose_x_0 = const()[name = tensor<string, []>("attns_13_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_13_transpose_y_0 = const()[name = tensor<string, []>("attns_13_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_17_cast_fp16 = transpose(perm = v_17_perm_0, x = v_15_cast_fp16)[name = tensor<string, []>("transpose_225")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_13_cast_fp16 = matmul(transpose_x = attns_13_transpose_x_0, transpose_y = attns_13_transpose_y_0, x = softmax_3_cast_fp16, y = v_17_cast_fp16)[name = tensor<string, []>("attns_13_cast_fp16")]; |
| tensor<int32, [4]> attns_15_perm_0 = const()[name = tensor<string, []>("attns_15_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_15 = const()[name = tensor<string, []>("concat_15"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_15_cast_fp16 = transpose(perm = attns_15_perm_0, x = attns_13_cast_fp16)[name = tensor<string, []>("transpose_224")]; |
| tensor<fp16, [1, 249, 1024]> x_57_cast_fp16 = reshape(shape = concat_15, x = attns_15_cast_fp16)[name = tensor<string, []>("x_57_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_3_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54147648))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55196288))), name = tensor<string, []>("model_encoder_layers_3_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_3_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_3_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55196864)))]; |
| tensor<fp16, [1, 249, 1024]> linear_22_cast_fp16 = linear(bias = model_encoder_layers_3_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_3_self_attn_output_proj_weight_to_fp16_palettized, x = x_57_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_73_cast_fp16 = add(x = linear_22_cast_fp16, y = input_71_cast_fp16)[name = tensor<string, []>("input_73_cast_fp16")]; |
| tensor<int32, [1]> x_59_axes_0 = const()[name = tensor<string, []>("x_59_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_3_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_3_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55198976)))]; |
| tensor<fp16, [1024]> model_encoder_layers_3_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_3_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55201088)))]; |
| tensor<fp16, [1, 249, 1024]> x_59_cast_fp16 = layer_norm(axes = x_59_axes_0, beta = model_encoder_layers_3_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_3_ffn_layer_norm_weight_to_fp16, x = input_73_cast_fp16)[name = tensor<string, []>("x_59_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_3_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55203200))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59397568))), name = tensor<string, []>("model_encoder_layers_3_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_3_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_3_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59398144)))]; |
| tensor<fp16, [1, 249, 4096]> linear_23_cast_fp16 = linear(bias = model_encoder_layers_3_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_3_ffn_inner_proj_weight_to_fp16_palettized, x = x_59_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")]; |
| tensor<string, []> input_77_mode_0 = const()[name = tensor<string, []>("input_77_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_77_cast_fp16 = gelu(mode = input_77_mode_0, x = linear_23_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_3_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59406400))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63600768))), name = tensor<string, []>("model_encoder_layers_3_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_3_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_3_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63601344)))]; |
| tensor<fp16, [1, 249, 1024]> linear_24_cast_fp16 = linear(bias = model_encoder_layers_3_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_3_ffn_output_proj_weight_to_fp16_palettized, x = input_77_cast_fp16)[name = tensor<string, []>("linear_24_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_79_cast_fp16 = add(x = linear_24_cast_fp16, y = input_73_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")]; |
| tensor<int32, [1]> x_63_axes_0 = const()[name = tensor<string, []>("x_63_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_4_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_4_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63603456)))]; |
| tensor<fp16, [1024]> model_encoder_layers_4_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_4_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63605568)))]; |
| tensor<fp16, [1, 249, 1024]> x_63_cast_fp16 = layer_norm(axes = x_63_axes_0, beta = model_encoder_layers_4_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_4_self_attn_layer_norm_weight_to_fp16, x = input_79_cast_fp16)[name = tensor<string, []>("x_63_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_4_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63607680))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64656320))), name = tensor<string, []>("model_encoder_layers_4_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64656896)))]; |
| tensor<fp16, [1, 249, 1024]> linear_25_cast_fp16 = linear(bias = model_encoder_layers_4_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_4_self_attn_q_proj_weight_to_fp16_palettized, x = x_63_cast_fp16)[name = tensor<string, []>("linear_25_cast_fp16")]; |
| tensor<int32, [4]> concat_16 = const()[name = tensor<string, []>("concat_16"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_17_cast_fp16 = reshape(shape = concat_16, x = linear_25_cast_fp16)[name = tensor<string, []>("q_17_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_4_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64659008))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65707648))), name = tensor<string, []>("model_encoder_layers_4_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_4_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_4_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65708224)))]; |
| tensor<fp16, [1, 249, 1024]> linear_26_cast_fp16 = linear(bias = model_encoder_layers_4_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_4_self_attn_k_proj_weight_to_fp16_palettized, x = x_63_cast_fp16)[name = tensor<string, []>("linear_26_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_4_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65710336))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66758976))), name = tensor<string, []>("model_encoder_layers_4_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66759552)))]; |
| tensor<fp16, [1, 249, 1024]> linear_27_cast_fp16 = linear(bias = model_encoder_layers_4_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_4_self_attn_v_proj_weight_to_fp16_palettized, x = x_63_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")]; |
| tensor<int32, [4]> concat_17 = const()[name = tensor<string, []>("concat_17"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_17_cast_fp16 = reshape(shape = concat_17, x = linear_26_cast_fp16)[name = tensor<string, []>("k_17_cast_fp16")]; |
| tensor<int32, [4]> concat_18 = const()[name = tensor<string, []>("concat_18"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_19_cast_fp16 = reshape(shape = concat_18, x = linear_27_cast_fp16)[name = tensor<string, []>("v_19_cast_fp16")]; |
| tensor<int32, [4]> v_21_perm_0 = const()[name = tensor<string, []>("v_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_4_y_0_to_fp16 = const()[name = tensor<string, []>("mul_4_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_4_cast_fp16 = mul(x = q_17_cast_fp16, y = mul_4_y_0_to_fp16)[name = tensor<string, []>("mul_4_cast_fp16")]; |
| tensor<bool, []> matmul_4_transpose_y_0 = const()[name = tensor<string, []>("matmul_4_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_4_transpose_x_0 = const()[name = tensor<string, []>("matmul_4_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_104_perm_0 = const()[name = tensor<string, []>("transpose_104_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_105_perm_0 = const()[name = tensor<string, []>("transpose_105_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_105 = transpose(perm = transpose_105_perm_0, x = k_17_cast_fp16)[name = tensor<string, []>("transpose_222")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_104 = transpose(perm = transpose_104_perm_0, x = mul_4_cast_fp16)[name = tensor<string, []>("transpose_223")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = transpose_104, y = transpose_105)[name = tensor<string, []>("matmul_4_cast_fp16")]; |
| tensor<int32, []> softmax_4_axis_0 = const()[name = tensor<string, []>("softmax_4_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = matmul_4_cast_fp16)[name = tensor<string, []>("softmax_4_cast_fp16")]; |
| tensor<bool, []> attns_17_transpose_x_0 = const()[name = tensor<string, []>("attns_17_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_17_transpose_y_0 = const()[name = tensor<string, []>("attns_17_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_21_cast_fp16 = transpose(perm = v_21_perm_0, x = v_19_cast_fp16)[name = tensor<string, []>("transpose_221")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_17_cast_fp16 = matmul(transpose_x = attns_17_transpose_x_0, transpose_y = attns_17_transpose_y_0, x = softmax_4_cast_fp16, y = v_21_cast_fp16)[name = tensor<string, []>("attns_17_cast_fp16")]; |
| tensor<int32, [4]> attns_19_perm_0 = const()[name = tensor<string, []>("attns_19_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_19 = const()[name = tensor<string, []>("concat_19"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_19_cast_fp16 = transpose(perm = attns_19_perm_0, x = attns_17_cast_fp16)[name = tensor<string, []>("transpose_220")]; |
| tensor<fp16, [1, 249, 1024]> x_65_cast_fp16 = reshape(shape = concat_19, x = attns_19_cast_fp16)[name = tensor<string, []>("x_65_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_4_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66761664))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67810304))), name = tensor<string, []>("model_encoder_layers_4_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_4_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_4_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67810880)))]; |
| tensor<fp16, [1, 249, 1024]> linear_28_cast_fp16 = linear(bias = model_encoder_layers_4_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_4_self_attn_output_proj_weight_to_fp16_palettized, x = x_65_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_81_cast_fp16 = add(x = linear_28_cast_fp16, y = input_79_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")]; |
| tensor<int32, [1]> x_67_axes_0 = const()[name = tensor<string, []>("x_67_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_4_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_4_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67812992)))]; |
| tensor<fp16, [1024]> model_encoder_layers_4_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_4_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67815104)))]; |
| tensor<fp16, [1, 249, 1024]> x_67_cast_fp16 = layer_norm(axes = x_67_axes_0, beta = model_encoder_layers_4_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_4_ffn_layer_norm_weight_to_fp16, x = input_81_cast_fp16)[name = tensor<string, []>("x_67_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_4_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67817216))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72011584))), name = tensor<string, []>("model_encoder_layers_4_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_4_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_4_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72012160)))]; |
| tensor<fp16, [1, 249, 4096]> linear_29_cast_fp16 = linear(bias = model_encoder_layers_4_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_4_ffn_inner_proj_weight_to_fp16_palettized, x = x_67_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")]; |
| tensor<string, []> input_85_mode_0 = const()[name = tensor<string, []>("input_85_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_85_cast_fp16 = gelu(mode = input_85_mode_0, x = linear_29_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_4_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72020416))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76214784))), name = tensor<string, []>("model_encoder_layers_4_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_4_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_4_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76215360)))]; |
| tensor<fp16, [1, 249, 1024]> linear_30_cast_fp16 = linear(bias = model_encoder_layers_4_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_4_ffn_output_proj_weight_to_fp16_palettized, x = input_85_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_87_cast_fp16 = add(x = linear_30_cast_fp16, y = input_81_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")]; |
| tensor<int32, [1]> x_71_axes_0 = const()[name = tensor<string, []>("x_71_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_5_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_5_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76217472)))]; |
| tensor<fp16, [1024]> model_encoder_layers_5_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_5_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76219584)))]; |
| tensor<fp16, [1, 249, 1024]> x_71_cast_fp16 = layer_norm(axes = x_71_axes_0, beta = model_encoder_layers_5_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_5_self_attn_layer_norm_weight_to_fp16, x = input_87_cast_fp16)[name = tensor<string, []>("x_71_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_5_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76221696))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77270336))), name = tensor<string, []>("model_encoder_layers_5_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77270912)))]; |
| tensor<fp16, [1, 249, 1024]> linear_31_cast_fp16 = linear(bias = model_encoder_layers_5_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_5_self_attn_q_proj_weight_to_fp16_palettized, x = x_71_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")]; |
| tensor<int32, [4]> concat_20 = const()[name = tensor<string, []>("concat_20"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_21_cast_fp16 = reshape(shape = concat_20, x = linear_31_cast_fp16)[name = tensor<string, []>("q_21_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_5_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77273024))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78321664))), name = tensor<string, []>("model_encoder_layers_5_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_5_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_5_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78322240)))]; |
| tensor<fp16, [1, 249, 1024]> linear_32_cast_fp16 = linear(bias = model_encoder_layers_5_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_5_self_attn_k_proj_weight_to_fp16_palettized, x = x_71_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_5_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78324352))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(79372992))), name = tensor<string, []>("model_encoder_layers_5_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(79373568)))]; |
| tensor<fp16, [1, 249, 1024]> linear_33_cast_fp16 = linear(bias = model_encoder_layers_5_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_5_self_attn_v_proj_weight_to_fp16_palettized, x = x_71_cast_fp16)[name = tensor<string, []>("linear_33_cast_fp16")]; |
| tensor<int32, [4]> concat_21 = const()[name = tensor<string, []>("concat_21"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_21_cast_fp16 = reshape(shape = concat_21, x = linear_32_cast_fp16)[name = tensor<string, []>("k_21_cast_fp16")]; |
| tensor<int32, [4]> concat_22 = const()[name = tensor<string, []>("concat_22"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_23_cast_fp16 = reshape(shape = concat_22, x = linear_33_cast_fp16)[name = tensor<string, []>("v_23_cast_fp16")]; |
| tensor<int32, [4]> v_25_perm_0 = const()[name = tensor<string, []>("v_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_5_y_0_to_fp16 = const()[name = tensor<string, []>("mul_5_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_5_cast_fp16 = mul(x = q_21_cast_fp16, y = mul_5_y_0_to_fp16)[name = tensor<string, []>("mul_5_cast_fp16")]; |
| tensor<bool, []> matmul_5_transpose_y_0 = const()[name = tensor<string, []>("matmul_5_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_5_transpose_x_0 = const()[name = tensor<string, []>("matmul_5_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_106_perm_0 = const()[name = tensor<string, []>("transpose_106_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_107_perm_0 = const()[name = tensor<string, []>("transpose_107_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_107 = transpose(perm = transpose_107_perm_0, x = k_21_cast_fp16)[name = tensor<string, []>("transpose_218")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_106 = transpose(perm = transpose_106_perm_0, x = mul_5_cast_fp16)[name = tensor<string, []>("transpose_219")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = transpose_106, y = transpose_107)[name = tensor<string, []>("matmul_5_cast_fp16")]; |
| tensor<int32, []> softmax_5_axis_0 = const()[name = tensor<string, []>("softmax_5_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = matmul_5_cast_fp16)[name = tensor<string, []>("softmax_5_cast_fp16")]; |
| tensor<bool, []> attns_21_transpose_x_0 = const()[name = tensor<string, []>("attns_21_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_21_transpose_y_0 = const()[name = tensor<string, []>("attns_21_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_25_cast_fp16 = transpose(perm = v_25_perm_0, x = v_23_cast_fp16)[name = tensor<string, []>("transpose_217")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_21_cast_fp16 = matmul(transpose_x = attns_21_transpose_x_0, transpose_y = attns_21_transpose_y_0, x = softmax_5_cast_fp16, y = v_25_cast_fp16)[name = tensor<string, []>("attns_21_cast_fp16")]; |
| tensor<int32, [4]> attns_23_perm_0 = const()[name = tensor<string, []>("attns_23_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_23 = const()[name = tensor<string, []>("concat_23"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_23_cast_fp16 = transpose(perm = attns_23_perm_0, x = attns_21_cast_fp16)[name = tensor<string, []>("transpose_216")]; |
| tensor<fp16, [1, 249, 1024]> x_73_cast_fp16 = reshape(shape = concat_23, x = attns_23_cast_fp16)[name = tensor<string, []>("x_73_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_5_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(79375680))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80424320))), name = tensor<string, []>("model_encoder_layers_5_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_5_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_5_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80424896)))]; |
| tensor<fp16, [1, 249, 1024]> linear_34_cast_fp16 = linear(bias = model_encoder_layers_5_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_5_self_attn_output_proj_weight_to_fp16_palettized, x = x_73_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_89_cast_fp16 = add(x = linear_34_cast_fp16, y = input_87_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")]; |
| tensor<int32, [1]> x_75_axes_0 = const()[name = tensor<string, []>("x_75_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_5_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_5_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80427008)))]; |
| tensor<fp16, [1024]> model_encoder_layers_5_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_5_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80429120)))]; |
| tensor<fp16, [1, 249, 1024]> x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = model_encoder_layers_5_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_5_ffn_layer_norm_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("x_75_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_5_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80431232))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84625600))), name = tensor<string, []>("model_encoder_layers_5_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_5_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_5_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84626176)))]; |
| tensor<fp16, [1, 249, 4096]> linear_35_cast_fp16 = linear(bias = model_encoder_layers_5_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_5_ffn_inner_proj_weight_to_fp16_palettized, x = x_75_cast_fp16)[name = tensor<string, []>("linear_35_cast_fp16")]; |
| tensor<string, []> input_93_mode_0 = const()[name = tensor<string, []>("input_93_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_93_cast_fp16 = gelu(mode = input_93_mode_0, x = linear_35_cast_fp16)[name = tensor<string, []>("input_93_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_5_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84634432))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88828800))), name = tensor<string, []>("model_encoder_layers_5_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_5_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_5_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88829376)))]; |
| tensor<fp16, [1, 249, 1024]> linear_36_cast_fp16 = linear(bias = model_encoder_layers_5_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_5_ffn_output_proj_weight_to_fp16_palettized, x = input_93_cast_fp16)[name = tensor<string, []>("linear_36_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_95_cast_fp16 = add(x = linear_36_cast_fp16, y = input_89_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")]; |
| tensor<int32, [1]> x_79_axes_0 = const()[name = tensor<string, []>("x_79_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_6_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_6_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88831488)))]; |
| tensor<fp16, [1024]> model_encoder_layers_6_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_6_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88833600)))]; |
| tensor<fp16, [1, 249, 1024]> x_79_cast_fp16 = layer_norm(axes = x_79_axes_0, beta = model_encoder_layers_6_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_6_self_attn_layer_norm_weight_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("x_79_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_6_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88835712))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89884352))), name = tensor<string, []>("model_encoder_layers_6_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89884928)))]; |
| tensor<fp16, [1, 249, 1024]> linear_37_cast_fp16 = linear(bias = model_encoder_layers_6_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_6_self_attn_q_proj_weight_to_fp16_palettized, x = x_79_cast_fp16)[name = tensor<string, []>("linear_37_cast_fp16")]; |
| tensor<int32, [4]> concat_24 = const()[name = tensor<string, []>("concat_24"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_25_cast_fp16 = reshape(shape = concat_24, x = linear_37_cast_fp16)[name = tensor<string, []>("q_25_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_6_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89887040))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90935680))), name = tensor<string, []>("model_encoder_layers_6_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_6_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_6_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90936256)))]; |
| tensor<fp16, [1, 249, 1024]> linear_38_cast_fp16 = linear(bias = model_encoder_layers_6_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_6_self_attn_k_proj_weight_to_fp16_palettized, x = x_79_cast_fp16)[name = tensor<string, []>("linear_38_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_6_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90938368))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91987008))), name = tensor<string, []>("model_encoder_layers_6_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91987584)))]; |
| tensor<fp16, [1, 249, 1024]> linear_39_cast_fp16 = linear(bias = model_encoder_layers_6_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_6_self_attn_v_proj_weight_to_fp16_palettized, x = x_79_cast_fp16)[name = tensor<string, []>("linear_39_cast_fp16")]; |
| tensor<int32, [4]> concat_25 = const()[name = tensor<string, []>("concat_25"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_25_cast_fp16 = reshape(shape = concat_25, x = linear_38_cast_fp16)[name = tensor<string, []>("k_25_cast_fp16")]; |
| tensor<int32, [4]> concat_26 = const()[name = tensor<string, []>("concat_26"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_27_cast_fp16 = reshape(shape = concat_26, x = linear_39_cast_fp16)[name = tensor<string, []>("v_27_cast_fp16")]; |
| tensor<int32, [4]> v_29_perm_0 = const()[name = tensor<string, []>("v_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_6_y_0_to_fp16 = const()[name = tensor<string, []>("mul_6_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_6_cast_fp16 = mul(x = q_25_cast_fp16, y = mul_6_y_0_to_fp16)[name = tensor<string, []>("mul_6_cast_fp16")]; |
| tensor<bool, []> matmul_6_transpose_y_0 = const()[name = tensor<string, []>("matmul_6_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_6_transpose_x_0 = const()[name = tensor<string, []>("matmul_6_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_108_perm_0 = const()[name = tensor<string, []>("transpose_108_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_109_perm_0 = const()[name = tensor<string, []>("transpose_109_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_109 = transpose(perm = transpose_109_perm_0, x = k_25_cast_fp16)[name = tensor<string, []>("transpose_214")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_108 = transpose(perm = transpose_108_perm_0, x = mul_6_cast_fp16)[name = tensor<string, []>("transpose_215")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = transpose_108, y = transpose_109)[name = tensor<string, []>("matmul_6_cast_fp16")]; |
| tensor<int32, []> softmax_6_axis_0 = const()[name = tensor<string, []>("softmax_6_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = matmul_6_cast_fp16)[name = tensor<string, []>("softmax_6_cast_fp16")]; |
| tensor<bool, []> attns_25_transpose_x_0 = const()[name = tensor<string, []>("attns_25_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_25_transpose_y_0 = const()[name = tensor<string, []>("attns_25_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_29_cast_fp16 = transpose(perm = v_29_perm_0, x = v_27_cast_fp16)[name = tensor<string, []>("transpose_213")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_25_cast_fp16 = matmul(transpose_x = attns_25_transpose_x_0, transpose_y = attns_25_transpose_y_0, x = softmax_6_cast_fp16, y = v_29_cast_fp16)[name = tensor<string, []>("attns_25_cast_fp16")]; |
| tensor<int32, [4]> attns_27_perm_0 = const()[name = tensor<string, []>("attns_27_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_27 = const()[name = tensor<string, []>("concat_27"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_27_cast_fp16 = transpose(perm = attns_27_perm_0, x = attns_25_cast_fp16)[name = tensor<string, []>("transpose_212")]; |
| tensor<fp16, [1, 249, 1024]> x_81_cast_fp16 = reshape(shape = concat_27, x = attns_27_cast_fp16)[name = tensor<string, []>("x_81_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_6_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91989696))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93038336))), name = tensor<string, []>("model_encoder_layers_6_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_6_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_6_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93038912)))]; |
| tensor<fp16, [1, 249, 1024]> linear_40_cast_fp16 = linear(bias = model_encoder_layers_6_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_6_self_attn_output_proj_weight_to_fp16_palettized, x = x_81_cast_fp16)[name = tensor<string, []>("linear_40_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_97_cast_fp16 = add(x = linear_40_cast_fp16, y = input_95_cast_fp16)[name = tensor<string, []>("input_97_cast_fp16")]; |
| tensor<int32, [1]> x_83_axes_0 = const()[name = tensor<string, []>("x_83_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_6_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_6_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93041024)))]; |
| tensor<fp16, [1024]> model_encoder_layers_6_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_6_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93043136)))]; |
| tensor<fp16, [1, 249, 1024]> x_83_cast_fp16 = layer_norm(axes = x_83_axes_0, beta = model_encoder_layers_6_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_6_ffn_layer_norm_weight_to_fp16, x = input_97_cast_fp16)[name = tensor<string, []>("x_83_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_6_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93045248))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97239616))), name = tensor<string, []>("model_encoder_layers_6_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_6_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_6_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97240192)))]; |
| tensor<fp16, [1, 249, 4096]> linear_41_cast_fp16 = linear(bias = model_encoder_layers_6_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_6_ffn_inner_proj_weight_to_fp16_palettized, x = x_83_cast_fp16)[name = tensor<string, []>("linear_41_cast_fp16")]; |
| tensor<string, []> input_101_mode_0 = const()[name = tensor<string, []>("input_101_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_101_cast_fp16 = gelu(mode = input_101_mode_0, x = linear_41_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_6_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97248448))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101442816))), name = tensor<string, []>("model_encoder_layers_6_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_6_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_6_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101443392)))]; |
| tensor<fp16, [1, 249, 1024]> linear_42_cast_fp16 = linear(bias = model_encoder_layers_6_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_6_ffn_output_proj_weight_to_fp16_palettized, x = input_101_cast_fp16)[name = tensor<string, []>("linear_42_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_103_cast_fp16 = add(x = linear_42_cast_fp16, y = input_97_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")]; |
| tensor<int32, [1]> x_87_axes_0 = const()[name = tensor<string, []>("x_87_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_7_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_7_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101445504)))]; |
| tensor<fp16, [1024]> model_encoder_layers_7_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_7_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101447616)))]; |
| tensor<fp16, [1, 249, 1024]> x_87_cast_fp16 = layer_norm(axes = x_87_axes_0, beta = model_encoder_layers_7_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_7_self_attn_layer_norm_weight_to_fp16, x = input_103_cast_fp16)[name = tensor<string, []>("x_87_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_7_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101449728))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102498368))), name = tensor<string, []>("model_encoder_layers_7_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102498944)))]; |
| tensor<fp16, [1, 249, 1024]> linear_43_cast_fp16 = linear(bias = model_encoder_layers_7_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_7_self_attn_q_proj_weight_to_fp16_palettized, x = x_87_cast_fp16)[name = tensor<string, []>("linear_43_cast_fp16")]; |
| tensor<int32, [4]> concat_28 = const()[name = tensor<string, []>("concat_28"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_29_cast_fp16 = reshape(shape = concat_28, x = linear_43_cast_fp16)[name = tensor<string, []>("q_29_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_7_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102501056))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103549696))), name = tensor<string, []>("model_encoder_layers_7_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_7_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_7_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103550272)))]; |
| tensor<fp16, [1, 249, 1024]> linear_44_cast_fp16 = linear(bias = model_encoder_layers_7_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_7_self_attn_k_proj_weight_to_fp16_palettized, x = x_87_cast_fp16)[name = tensor<string, []>("linear_44_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_7_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103552384))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104601024))), name = tensor<string, []>("model_encoder_layers_7_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104601600)))]; |
| tensor<fp16, [1, 249, 1024]> linear_45_cast_fp16 = linear(bias = model_encoder_layers_7_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_7_self_attn_v_proj_weight_to_fp16_palettized, x = x_87_cast_fp16)[name = tensor<string, []>("linear_45_cast_fp16")]; |
| tensor<int32, [4]> concat_29 = const()[name = tensor<string, []>("concat_29"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_29_cast_fp16 = reshape(shape = concat_29, x = linear_44_cast_fp16)[name = tensor<string, []>("k_29_cast_fp16")]; |
| tensor<int32, [4]> concat_30 = const()[name = tensor<string, []>("concat_30"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_31_cast_fp16 = reshape(shape = concat_30, x = linear_45_cast_fp16)[name = tensor<string, []>("v_31_cast_fp16")]; |
| tensor<int32, [4]> v_33_perm_0 = const()[name = tensor<string, []>("v_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_7_y_0_to_fp16 = const()[name = tensor<string, []>("mul_7_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_7_cast_fp16 = mul(x = q_29_cast_fp16, y = mul_7_y_0_to_fp16)[name = tensor<string, []>("mul_7_cast_fp16")]; |
| tensor<bool, []> matmul_7_transpose_y_0 = const()[name = tensor<string, []>("matmul_7_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_7_transpose_x_0 = const()[name = tensor<string, []>("matmul_7_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_110_perm_0 = const()[name = tensor<string, []>("transpose_110_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_111_perm_0 = const()[name = tensor<string, []>("transpose_111_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_111 = transpose(perm = transpose_111_perm_0, x = k_29_cast_fp16)[name = tensor<string, []>("transpose_210")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_110 = transpose(perm = transpose_110_perm_0, x = mul_7_cast_fp16)[name = tensor<string, []>("transpose_211")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = transpose_110, y = transpose_111)[name = tensor<string, []>("matmul_7_cast_fp16")]; |
| tensor<int32, []> softmax_7_axis_0 = const()[name = tensor<string, []>("softmax_7_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = matmul_7_cast_fp16)[name = tensor<string, []>("softmax_7_cast_fp16")]; |
| tensor<bool, []> attns_29_transpose_x_0 = const()[name = tensor<string, []>("attns_29_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_29_transpose_y_0 = const()[name = tensor<string, []>("attns_29_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_33_cast_fp16 = transpose(perm = v_33_perm_0, x = v_31_cast_fp16)[name = tensor<string, []>("transpose_209")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_29_cast_fp16 = matmul(transpose_x = attns_29_transpose_x_0, transpose_y = attns_29_transpose_y_0, x = softmax_7_cast_fp16, y = v_33_cast_fp16)[name = tensor<string, []>("attns_29_cast_fp16")]; |
| tensor<int32, [4]> attns_31_perm_0 = const()[name = tensor<string, []>("attns_31_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_31 = const()[name = tensor<string, []>("concat_31"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_31_cast_fp16 = transpose(perm = attns_31_perm_0, x = attns_29_cast_fp16)[name = tensor<string, []>("transpose_208")]; |
| tensor<fp16, [1, 249, 1024]> x_89_cast_fp16 = reshape(shape = concat_31, x = attns_31_cast_fp16)[name = tensor<string, []>("x_89_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_7_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104603712))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105652352))), name = tensor<string, []>("model_encoder_layers_7_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_7_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_7_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105652928)))]; |
| tensor<fp16, [1, 249, 1024]> linear_46_cast_fp16 = linear(bias = model_encoder_layers_7_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_7_self_attn_output_proj_weight_to_fp16_palettized, x = x_89_cast_fp16)[name = tensor<string, []>("linear_46_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_105_cast_fp16 = add(x = linear_46_cast_fp16, y = input_103_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")]; |
| tensor<int32, [1]> x_91_axes_0 = const()[name = tensor<string, []>("x_91_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_7_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_7_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105655040)))]; |
| tensor<fp16, [1024]> model_encoder_layers_7_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_7_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105657152)))]; |
| tensor<fp16, [1, 249, 1024]> x_91_cast_fp16 = layer_norm(axes = x_91_axes_0, beta = model_encoder_layers_7_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_7_ffn_layer_norm_weight_to_fp16, x = input_105_cast_fp16)[name = tensor<string, []>("x_91_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_7_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105659264))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109853632))), name = tensor<string, []>("model_encoder_layers_7_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_7_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_7_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109854208)))]; |
| tensor<fp16, [1, 249, 4096]> linear_47_cast_fp16 = linear(bias = model_encoder_layers_7_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_7_ffn_inner_proj_weight_to_fp16_palettized, x = x_91_cast_fp16)[name = tensor<string, []>("linear_47_cast_fp16")]; |
| tensor<string, []> input_109_mode_0 = const()[name = tensor<string, []>("input_109_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = linear_47_cast_fp16)[name = tensor<string, []>("input_109_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_7_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109862464))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114056832))), name = tensor<string, []>("model_encoder_layers_7_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_7_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_7_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114057408)))]; |
| tensor<fp16, [1, 249, 1024]> linear_48_cast_fp16 = linear(bias = model_encoder_layers_7_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_7_ffn_output_proj_weight_to_fp16_palettized, x = input_109_cast_fp16)[name = tensor<string, []>("linear_48_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_111_cast_fp16 = add(x = linear_48_cast_fp16, y = input_105_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")]; |
| tensor<int32, [1]> x_95_axes_0 = const()[name = tensor<string, []>("x_95_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_8_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_8_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114059520)))]; |
| tensor<fp16, [1024]> model_encoder_layers_8_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_8_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114061632)))]; |
| tensor<fp16, [1, 249, 1024]> x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = model_encoder_layers_8_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_8_self_attn_layer_norm_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("x_95_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_8_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114063744))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115112384))), name = tensor<string, []>("model_encoder_layers_8_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115112960)))]; |
| tensor<fp16, [1, 249, 1024]> linear_49_cast_fp16 = linear(bias = model_encoder_layers_8_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_8_self_attn_q_proj_weight_to_fp16_palettized, x = x_95_cast_fp16)[name = tensor<string, []>("linear_49_cast_fp16")]; |
| tensor<int32, [4]> concat_32 = const()[name = tensor<string, []>("concat_32"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_33_cast_fp16 = reshape(shape = concat_32, x = linear_49_cast_fp16)[name = tensor<string, []>("q_33_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_8_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115115072))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116163712))), name = tensor<string, []>("model_encoder_layers_8_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_8_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_8_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116164288)))]; |
| tensor<fp16, [1, 249, 1024]> linear_50_cast_fp16 = linear(bias = model_encoder_layers_8_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_8_self_attn_k_proj_weight_to_fp16_palettized, x = x_95_cast_fp16)[name = tensor<string, []>("linear_50_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_8_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116166400))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117215040))), name = tensor<string, []>("model_encoder_layers_8_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117215616)))]; |
| tensor<fp16, [1, 249, 1024]> linear_51_cast_fp16 = linear(bias = model_encoder_layers_8_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_8_self_attn_v_proj_weight_to_fp16_palettized, x = x_95_cast_fp16)[name = tensor<string, []>("linear_51_cast_fp16")]; |
| tensor<int32, [4]> concat_33 = const()[name = tensor<string, []>("concat_33"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_33_cast_fp16 = reshape(shape = concat_33, x = linear_50_cast_fp16)[name = tensor<string, []>("k_33_cast_fp16")]; |
| tensor<int32, [4]> concat_34 = const()[name = tensor<string, []>("concat_34"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_35_cast_fp16 = reshape(shape = concat_34, x = linear_51_cast_fp16)[name = tensor<string, []>("v_35_cast_fp16")]; |
| tensor<int32, [4]> v_37_perm_0 = const()[name = tensor<string, []>("v_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_8_y_0_to_fp16 = const()[name = tensor<string, []>("mul_8_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_8_cast_fp16 = mul(x = q_33_cast_fp16, y = mul_8_y_0_to_fp16)[name = tensor<string, []>("mul_8_cast_fp16")]; |
| tensor<bool, []> matmul_8_transpose_y_0 = const()[name = tensor<string, []>("matmul_8_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_8_transpose_x_0 = const()[name = tensor<string, []>("matmul_8_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_112_perm_0 = const()[name = tensor<string, []>("transpose_112_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_113_perm_0 = const()[name = tensor<string, []>("transpose_113_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_113 = transpose(perm = transpose_113_perm_0, x = k_33_cast_fp16)[name = tensor<string, []>("transpose_206")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_112 = transpose(perm = transpose_112_perm_0, x = mul_8_cast_fp16)[name = tensor<string, []>("transpose_207")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = transpose_112, y = transpose_113)[name = tensor<string, []>("matmul_8_cast_fp16")]; |
| tensor<int32, []> softmax_8_axis_0 = const()[name = tensor<string, []>("softmax_8_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = matmul_8_cast_fp16)[name = tensor<string, []>("softmax_8_cast_fp16")]; |
| tensor<bool, []> attns_33_transpose_x_0 = const()[name = tensor<string, []>("attns_33_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_33_transpose_y_0 = const()[name = tensor<string, []>("attns_33_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_37_cast_fp16 = transpose(perm = v_37_perm_0, x = v_35_cast_fp16)[name = tensor<string, []>("transpose_205")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_33_cast_fp16 = matmul(transpose_x = attns_33_transpose_x_0, transpose_y = attns_33_transpose_y_0, x = softmax_8_cast_fp16, y = v_37_cast_fp16)[name = tensor<string, []>("attns_33_cast_fp16")]; |
| tensor<int32, [4]> attns_35_perm_0 = const()[name = tensor<string, []>("attns_35_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_35 = const()[name = tensor<string, []>("concat_35"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_35_cast_fp16 = transpose(perm = attns_35_perm_0, x = attns_33_cast_fp16)[name = tensor<string, []>("transpose_204")]; |
| tensor<fp16, [1, 249, 1024]> x_97_cast_fp16 = reshape(shape = concat_35, x = attns_35_cast_fp16)[name = tensor<string, []>("x_97_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_8_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117217728))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118266368))), name = tensor<string, []>("model_encoder_layers_8_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_8_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_8_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118266944)))]; |
| tensor<fp16, [1, 249, 1024]> linear_52_cast_fp16 = linear(bias = model_encoder_layers_8_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_8_self_attn_output_proj_weight_to_fp16_palettized, x = x_97_cast_fp16)[name = tensor<string, []>("linear_52_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_113_cast_fp16 = add(x = linear_52_cast_fp16, y = input_111_cast_fp16)[name = tensor<string, []>("input_113_cast_fp16")]; |
| tensor<int32, [1]> x_99_axes_0 = const()[name = tensor<string, []>("x_99_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_8_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_8_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118269056)))]; |
| tensor<fp16, [1024]> model_encoder_layers_8_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_8_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118271168)))]; |
| tensor<fp16, [1, 249, 1024]> x_99_cast_fp16 = layer_norm(axes = x_99_axes_0, beta = model_encoder_layers_8_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_8_ffn_layer_norm_weight_to_fp16, x = input_113_cast_fp16)[name = tensor<string, []>("x_99_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_8_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118273280))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122467648))), name = tensor<string, []>("model_encoder_layers_8_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_8_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_8_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122468224)))]; |
| tensor<fp16, [1, 249, 4096]> linear_53_cast_fp16 = linear(bias = model_encoder_layers_8_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_8_ffn_inner_proj_weight_to_fp16_palettized, x = x_99_cast_fp16)[name = tensor<string, []>("linear_53_cast_fp16")]; |
| tensor<string, []> input_117_mode_0 = const()[name = tensor<string, []>("input_117_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_117_cast_fp16 = gelu(mode = input_117_mode_0, x = linear_53_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_8_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122476480))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126670848))), name = tensor<string, []>("model_encoder_layers_8_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_8_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_8_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126671424)))]; |
| tensor<fp16, [1, 249, 1024]> linear_54_cast_fp16 = linear(bias = model_encoder_layers_8_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_8_ffn_output_proj_weight_to_fp16_palettized, x = input_117_cast_fp16)[name = tensor<string, []>("linear_54_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_119_cast_fp16 = add(x = linear_54_cast_fp16, y = input_113_cast_fp16)[name = tensor<string, []>("input_119_cast_fp16")]; |
| tensor<int32, [1]> x_103_axes_0 = const()[name = tensor<string, []>("x_103_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_9_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_9_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126673536)))]; |
| tensor<fp16, [1024]> model_encoder_layers_9_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_9_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126675648)))]; |
| tensor<fp16, [1, 249, 1024]> x_103_cast_fp16 = layer_norm(axes = x_103_axes_0, beta = model_encoder_layers_9_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_9_self_attn_layer_norm_weight_to_fp16, x = input_119_cast_fp16)[name = tensor<string, []>("x_103_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_9_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126677760))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127726400))), name = tensor<string, []>("model_encoder_layers_9_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127726976)))]; |
| tensor<fp16, [1, 249, 1024]> linear_55_cast_fp16 = linear(bias = model_encoder_layers_9_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_9_self_attn_q_proj_weight_to_fp16_palettized, x = x_103_cast_fp16)[name = tensor<string, []>("linear_55_cast_fp16")]; |
| tensor<int32, [4]> concat_36 = const()[name = tensor<string, []>("concat_36"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_37_cast_fp16 = reshape(shape = concat_36, x = linear_55_cast_fp16)[name = tensor<string, []>("q_37_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_9_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127729088))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(128777728))), name = tensor<string, []>("model_encoder_layers_9_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_9_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_9_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(128778304)))]; |
| tensor<fp16, [1, 249, 1024]> linear_56_cast_fp16 = linear(bias = model_encoder_layers_9_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_9_self_attn_k_proj_weight_to_fp16_palettized, x = x_103_cast_fp16)[name = tensor<string, []>("linear_56_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_9_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(128780416))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129829056))), name = tensor<string, []>("model_encoder_layers_9_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129829632)))]; |
| tensor<fp16, [1, 249, 1024]> linear_57_cast_fp16 = linear(bias = model_encoder_layers_9_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_9_self_attn_v_proj_weight_to_fp16_palettized, x = x_103_cast_fp16)[name = tensor<string, []>("linear_57_cast_fp16")]; |
| tensor<int32, [4]> concat_37 = const()[name = tensor<string, []>("concat_37"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_37_cast_fp16 = reshape(shape = concat_37, x = linear_56_cast_fp16)[name = tensor<string, []>("k_37_cast_fp16")]; |
| tensor<int32, [4]> concat_38 = const()[name = tensor<string, []>("concat_38"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_39_cast_fp16 = reshape(shape = concat_38, x = linear_57_cast_fp16)[name = tensor<string, []>("v_39_cast_fp16")]; |
| tensor<int32, [4]> v_41_perm_0 = const()[name = tensor<string, []>("v_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_9_y_0_to_fp16 = const()[name = tensor<string, []>("mul_9_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_9_cast_fp16 = mul(x = q_37_cast_fp16, y = mul_9_y_0_to_fp16)[name = tensor<string, []>("mul_9_cast_fp16")]; |
| tensor<bool, []> matmul_9_transpose_y_0 = const()[name = tensor<string, []>("matmul_9_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_9_transpose_x_0 = const()[name = tensor<string, []>("matmul_9_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_114_perm_0 = const()[name = tensor<string, []>("transpose_114_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_115_perm_0 = const()[name = tensor<string, []>("transpose_115_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_115 = transpose(perm = transpose_115_perm_0, x = k_37_cast_fp16)[name = tensor<string, []>("transpose_202")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_114 = transpose(perm = transpose_114_perm_0, x = mul_9_cast_fp16)[name = tensor<string, []>("transpose_203")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = transpose_114, y = transpose_115)[name = tensor<string, []>("matmul_9_cast_fp16")]; |
| tensor<int32, []> softmax_9_axis_0 = const()[name = tensor<string, []>("softmax_9_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = matmul_9_cast_fp16)[name = tensor<string, []>("softmax_9_cast_fp16")]; |
| tensor<bool, []> attns_37_transpose_x_0 = const()[name = tensor<string, []>("attns_37_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_37_transpose_y_0 = const()[name = tensor<string, []>("attns_37_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_41_cast_fp16 = transpose(perm = v_41_perm_0, x = v_39_cast_fp16)[name = tensor<string, []>("transpose_201")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_37_cast_fp16 = matmul(transpose_x = attns_37_transpose_x_0, transpose_y = attns_37_transpose_y_0, x = softmax_9_cast_fp16, y = v_41_cast_fp16)[name = tensor<string, []>("attns_37_cast_fp16")]; |
| tensor<int32, [4]> attns_39_perm_0 = const()[name = tensor<string, []>("attns_39_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_39 = const()[name = tensor<string, []>("concat_39"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_39_cast_fp16 = transpose(perm = attns_39_perm_0, x = attns_37_cast_fp16)[name = tensor<string, []>("transpose_200")]; |
| tensor<fp16, [1, 249, 1024]> x_105_cast_fp16 = reshape(shape = concat_39, x = attns_39_cast_fp16)[name = tensor<string, []>("x_105_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_9_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129831744))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130880384))), name = tensor<string, []>("model_encoder_layers_9_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_9_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_9_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130880960)))]; |
| tensor<fp16, [1, 249, 1024]> linear_58_cast_fp16 = linear(bias = model_encoder_layers_9_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_9_self_attn_output_proj_weight_to_fp16_palettized, x = x_105_cast_fp16)[name = tensor<string, []>("linear_58_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_121_cast_fp16 = add(x = linear_58_cast_fp16, y = input_119_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")]; |
| tensor<int32, [1]> x_107_axes_0 = const()[name = tensor<string, []>("x_107_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_9_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_9_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130883072)))]; |
| tensor<fp16, [1024]> model_encoder_layers_9_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_9_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130885184)))]; |
| tensor<fp16, [1, 249, 1024]> x_107_cast_fp16 = layer_norm(axes = x_107_axes_0, beta = model_encoder_layers_9_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_9_ffn_layer_norm_weight_to_fp16, x = input_121_cast_fp16)[name = tensor<string, []>("x_107_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_9_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130887296))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135081664))), name = tensor<string, []>("model_encoder_layers_9_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_9_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_9_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135082240)))]; |
| tensor<fp16, [1, 249, 4096]> linear_59_cast_fp16 = linear(bias = model_encoder_layers_9_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_9_ffn_inner_proj_weight_to_fp16_palettized, x = x_107_cast_fp16)[name = tensor<string, []>("linear_59_cast_fp16")]; |
| tensor<string, []> input_125_mode_0 = const()[name = tensor<string, []>("input_125_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_125_cast_fp16 = gelu(mode = input_125_mode_0, x = linear_59_cast_fp16)[name = tensor<string, []>("input_125_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_9_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135090496))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139284864))), name = tensor<string, []>("model_encoder_layers_9_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_9_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_9_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139285440)))]; |
| tensor<fp16, [1, 249, 1024]> linear_60_cast_fp16 = linear(bias = model_encoder_layers_9_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_9_ffn_output_proj_weight_to_fp16_palettized, x = input_125_cast_fp16)[name = tensor<string, []>("linear_60_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_127_cast_fp16 = add(x = linear_60_cast_fp16, y = input_121_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")]; |
| tensor<int32, [1]> x_111_axes_0 = const()[name = tensor<string, []>("x_111_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_10_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_10_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139287552)))]; |
| tensor<fp16, [1024]> model_encoder_layers_10_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_10_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139289664)))]; |
| tensor<fp16, [1, 249, 1024]> x_111_cast_fp16 = layer_norm(axes = x_111_axes_0, beta = model_encoder_layers_10_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_10_self_attn_layer_norm_weight_to_fp16, x = input_127_cast_fp16)[name = tensor<string, []>("x_111_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_10_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139291776))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140340416))), name = tensor<string, []>("model_encoder_layers_10_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140340992)))]; |
| tensor<fp16, [1, 249, 1024]> linear_61_cast_fp16 = linear(bias = model_encoder_layers_10_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_10_self_attn_q_proj_weight_to_fp16_palettized, x = x_111_cast_fp16)[name = tensor<string, []>("linear_61_cast_fp16")]; |
| tensor<int32, [4]> concat_40 = const()[name = tensor<string, []>("concat_40"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_41_cast_fp16 = reshape(shape = concat_40, x = linear_61_cast_fp16)[name = tensor<string, []>("q_41_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_10_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140343104))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141391744))), name = tensor<string, []>("model_encoder_layers_10_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_10_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_10_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141392320)))]; |
| tensor<fp16, [1, 249, 1024]> linear_62_cast_fp16 = linear(bias = model_encoder_layers_10_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_10_self_attn_k_proj_weight_to_fp16_palettized, x = x_111_cast_fp16)[name = tensor<string, []>("linear_62_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_10_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141394432))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142443072))), name = tensor<string, []>("model_encoder_layers_10_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142443648)))]; |
| tensor<fp16, [1, 249, 1024]> linear_63_cast_fp16 = linear(bias = model_encoder_layers_10_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_10_self_attn_v_proj_weight_to_fp16_palettized, x = x_111_cast_fp16)[name = tensor<string, []>("linear_63_cast_fp16")]; |
| tensor<int32, [4]> concat_41 = const()[name = tensor<string, []>("concat_41"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_41_cast_fp16 = reshape(shape = concat_41, x = linear_62_cast_fp16)[name = tensor<string, []>("k_41_cast_fp16")]; |
| tensor<int32, [4]> concat_42 = const()[name = tensor<string, []>("concat_42"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_43_cast_fp16 = reshape(shape = concat_42, x = linear_63_cast_fp16)[name = tensor<string, []>("v_43_cast_fp16")]; |
| tensor<int32, [4]> v_45_perm_0 = const()[name = tensor<string, []>("v_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_10_y_0_to_fp16 = const()[name = tensor<string, []>("mul_10_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_10_cast_fp16 = mul(x = q_41_cast_fp16, y = mul_10_y_0_to_fp16)[name = tensor<string, []>("mul_10_cast_fp16")]; |
| tensor<bool, []> matmul_10_transpose_y_0 = const()[name = tensor<string, []>("matmul_10_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_10_transpose_x_0 = const()[name = tensor<string, []>("matmul_10_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_116_perm_0 = const()[name = tensor<string, []>("transpose_116_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_117_perm_0 = const()[name = tensor<string, []>("transpose_117_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_117 = transpose(perm = transpose_117_perm_0, x = k_41_cast_fp16)[name = tensor<string, []>("transpose_198")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_116 = transpose(perm = transpose_116_perm_0, x = mul_10_cast_fp16)[name = tensor<string, []>("transpose_199")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = transpose_116, y = transpose_117)[name = tensor<string, []>("matmul_10_cast_fp16")]; |
| tensor<int32, []> softmax_10_axis_0 = const()[name = tensor<string, []>("softmax_10_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = matmul_10_cast_fp16)[name = tensor<string, []>("softmax_10_cast_fp16")]; |
| tensor<bool, []> attns_41_transpose_x_0 = const()[name = tensor<string, []>("attns_41_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_41_transpose_y_0 = const()[name = tensor<string, []>("attns_41_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_45_cast_fp16 = transpose(perm = v_45_perm_0, x = v_43_cast_fp16)[name = tensor<string, []>("transpose_197")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_41_cast_fp16 = matmul(transpose_x = attns_41_transpose_x_0, transpose_y = attns_41_transpose_y_0, x = softmax_10_cast_fp16, y = v_45_cast_fp16)[name = tensor<string, []>("attns_41_cast_fp16")]; |
| tensor<int32, [4]> attns_43_perm_0 = const()[name = tensor<string, []>("attns_43_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_43 = const()[name = tensor<string, []>("concat_43"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_43_cast_fp16 = transpose(perm = attns_43_perm_0, x = attns_41_cast_fp16)[name = tensor<string, []>("transpose_196")]; |
| tensor<fp16, [1, 249, 1024]> x_113_cast_fp16 = reshape(shape = concat_43, x = attns_43_cast_fp16)[name = tensor<string, []>("x_113_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_10_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142445760))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143494400))), name = tensor<string, []>("model_encoder_layers_10_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_10_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_10_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143494976)))]; |
| tensor<fp16, [1, 249, 1024]> linear_64_cast_fp16 = linear(bias = model_encoder_layers_10_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_10_self_attn_output_proj_weight_to_fp16_palettized, x = x_113_cast_fp16)[name = tensor<string, []>("linear_64_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_129_cast_fp16 = add(x = linear_64_cast_fp16, y = input_127_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")]; |
| tensor<int32, [1]> x_115_axes_0 = const()[name = tensor<string, []>("x_115_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_10_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_10_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143497088)))]; |
| tensor<fp16, [1024]> model_encoder_layers_10_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_10_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143499200)))]; |
| tensor<fp16, [1, 249, 1024]> x_115_cast_fp16 = layer_norm(axes = x_115_axes_0, beta = model_encoder_layers_10_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_10_ffn_layer_norm_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("x_115_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_10_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143501312))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147695680))), name = tensor<string, []>("model_encoder_layers_10_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_10_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_10_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147696256)))]; |
| tensor<fp16, [1, 249, 4096]> linear_65_cast_fp16 = linear(bias = model_encoder_layers_10_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_10_ffn_inner_proj_weight_to_fp16_palettized, x = x_115_cast_fp16)[name = tensor<string, []>("linear_65_cast_fp16")]; |
| tensor<string, []> input_133_mode_0 = const()[name = tensor<string, []>("input_133_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_133_cast_fp16 = gelu(mode = input_133_mode_0, x = linear_65_cast_fp16)[name = tensor<string, []>("input_133_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_10_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147704512))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151898880))), name = tensor<string, []>("model_encoder_layers_10_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_10_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_10_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151899456)))]; |
| tensor<fp16, [1, 249, 1024]> linear_66_cast_fp16 = linear(bias = model_encoder_layers_10_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_10_ffn_output_proj_weight_to_fp16_palettized, x = input_133_cast_fp16)[name = tensor<string, []>("linear_66_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_135_cast_fp16 = add(x = linear_66_cast_fp16, y = input_129_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")]; |
| tensor<int32, [1]> x_119_axes_0 = const()[name = tensor<string, []>("x_119_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_11_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_11_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151901568)))]; |
| tensor<fp16, [1024]> model_encoder_layers_11_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_11_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151903680)))]; |
| tensor<fp16, [1, 249, 1024]> x_119_cast_fp16 = layer_norm(axes = x_119_axes_0, beta = model_encoder_layers_11_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_11_self_attn_layer_norm_weight_to_fp16, x = input_135_cast_fp16)[name = tensor<string, []>("x_119_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_11_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151905792))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152954432))), name = tensor<string, []>("model_encoder_layers_11_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152955008)))]; |
| tensor<fp16, [1, 249, 1024]> linear_67_cast_fp16 = linear(bias = model_encoder_layers_11_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_11_self_attn_q_proj_weight_to_fp16_palettized, x = x_119_cast_fp16)[name = tensor<string, []>("linear_67_cast_fp16")]; |
| tensor<int32, [4]> concat_44 = const()[name = tensor<string, []>("concat_44"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_45_cast_fp16 = reshape(shape = concat_44, x = linear_67_cast_fp16)[name = tensor<string, []>("q_45_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_11_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152957120))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154005760))), name = tensor<string, []>("model_encoder_layers_11_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_11_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_11_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154006336)))]; |
| tensor<fp16, [1, 249, 1024]> linear_68_cast_fp16 = linear(bias = model_encoder_layers_11_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_11_self_attn_k_proj_weight_to_fp16_palettized, x = x_119_cast_fp16)[name = tensor<string, []>("linear_68_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_11_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154008448))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155057088))), name = tensor<string, []>("model_encoder_layers_11_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155057664)))]; |
| tensor<fp16, [1, 249, 1024]> linear_69_cast_fp16 = linear(bias = model_encoder_layers_11_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_11_self_attn_v_proj_weight_to_fp16_palettized, x = x_119_cast_fp16)[name = tensor<string, []>("linear_69_cast_fp16")]; |
| tensor<int32, [4]> concat_45 = const()[name = tensor<string, []>("concat_45"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_45_cast_fp16 = reshape(shape = concat_45, x = linear_68_cast_fp16)[name = tensor<string, []>("k_45_cast_fp16")]; |
| tensor<int32, [4]> concat_46 = const()[name = tensor<string, []>("concat_46"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_47_cast_fp16 = reshape(shape = concat_46, x = linear_69_cast_fp16)[name = tensor<string, []>("v_47_cast_fp16")]; |
| tensor<int32, [4]> v_49_perm_0 = const()[name = tensor<string, []>("v_49_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_11_y_0_to_fp16 = const()[name = tensor<string, []>("mul_11_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_11_cast_fp16 = mul(x = q_45_cast_fp16, y = mul_11_y_0_to_fp16)[name = tensor<string, []>("mul_11_cast_fp16")]; |
| tensor<bool, []> matmul_11_transpose_y_0 = const()[name = tensor<string, []>("matmul_11_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_11_transpose_x_0 = const()[name = tensor<string, []>("matmul_11_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_118_perm_0 = const()[name = tensor<string, []>("transpose_118_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_119_perm_0 = const()[name = tensor<string, []>("transpose_119_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_119 = transpose(perm = transpose_119_perm_0, x = k_45_cast_fp16)[name = tensor<string, []>("transpose_194")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_118 = transpose(perm = transpose_118_perm_0, x = mul_11_cast_fp16)[name = tensor<string, []>("transpose_195")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = transpose_118, y = transpose_119)[name = tensor<string, []>("matmul_11_cast_fp16")]; |
| tensor<int32, []> softmax_11_axis_0 = const()[name = tensor<string, []>("softmax_11_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = matmul_11_cast_fp16)[name = tensor<string, []>("softmax_11_cast_fp16")]; |
| tensor<bool, []> attns_45_transpose_x_0 = const()[name = tensor<string, []>("attns_45_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_45_transpose_y_0 = const()[name = tensor<string, []>("attns_45_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_49_cast_fp16 = transpose(perm = v_49_perm_0, x = v_47_cast_fp16)[name = tensor<string, []>("transpose_193")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_45_cast_fp16 = matmul(transpose_x = attns_45_transpose_x_0, transpose_y = attns_45_transpose_y_0, x = softmax_11_cast_fp16, y = v_49_cast_fp16)[name = tensor<string, []>("attns_45_cast_fp16")]; |
| tensor<int32, [4]> attns_47_perm_0 = const()[name = tensor<string, []>("attns_47_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_47 = const()[name = tensor<string, []>("concat_47"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_47_cast_fp16 = transpose(perm = attns_47_perm_0, x = attns_45_cast_fp16)[name = tensor<string, []>("transpose_192")]; |
| tensor<fp16, [1, 249, 1024]> x_121_cast_fp16 = reshape(shape = concat_47, x = attns_47_cast_fp16)[name = tensor<string, []>("x_121_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_11_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155059776))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156108416))), name = tensor<string, []>("model_encoder_layers_11_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_11_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_11_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156108992)))]; |
| tensor<fp16, [1, 249, 1024]> linear_70_cast_fp16 = linear(bias = model_encoder_layers_11_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_11_self_attn_output_proj_weight_to_fp16_palettized, x = x_121_cast_fp16)[name = tensor<string, []>("linear_70_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_137_cast_fp16 = add(x = linear_70_cast_fp16, y = input_135_cast_fp16)[name = tensor<string, []>("input_137_cast_fp16")]; |
| tensor<int32, [1]> x_123_axes_0 = const()[name = tensor<string, []>("x_123_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_11_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_11_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156111104)))]; |
| tensor<fp16, [1024]> model_encoder_layers_11_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_11_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156113216)))]; |
| tensor<fp16, [1, 249, 1024]> x_123_cast_fp16 = layer_norm(axes = x_123_axes_0, beta = model_encoder_layers_11_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_11_ffn_layer_norm_weight_to_fp16, x = input_137_cast_fp16)[name = tensor<string, []>("x_123_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_11_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156115328))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160309696))), name = tensor<string, []>("model_encoder_layers_11_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_11_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_11_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160310272)))]; |
| tensor<fp16, [1, 249, 4096]> linear_71_cast_fp16 = linear(bias = model_encoder_layers_11_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_11_ffn_inner_proj_weight_to_fp16_palettized, x = x_123_cast_fp16)[name = tensor<string, []>("linear_71_cast_fp16")]; |
| tensor<string, []> input_141_mode_0 = const()[name = tensor<string, []>("input_141_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_141_cast_fp16 = gelu(mode = input_141_mode_0, x = linear_71_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_11_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160318528))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164512896))), name = tensor<string, []>("model_encoder_layers_11_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_11_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_11_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164513472)))]; |
| tensor<fp16, [1, 249, 1024]> linear_72_cast_fp16 = linear(bias = model_encoder_layers_11_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_11_ffn_output_proj_weight_to_fp16_palettized, x = input_141_cast_fp16)[name = tensor<string, []>("linear_72_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_143_cast_fp16 = add(x = linear_72_cast_fp16, y = input_137_cast_fp16)[name = tensor<string, []>("input_143_cast_fp16")]; |
| tensor<int32, [1]> x_127_axes_0 = const()[name = tensor<string, []>("x_127_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_12_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_12_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164515584)))]; |
| tensor<fp16, [1024]> model_encoder_layers_12_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_12_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164517696)))]; |
| tensor<fp16, [1, 249, 1024]> x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = model_encoder_layers_12_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_12_self_attn_layer_norm_weight_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("x_127_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_12_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164519808))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165568448))), name = tensor<string, []>("model_encoder_layers_12_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165569024)))]; |
| tensor<fp16, [1, 249, 1024]> linear_73_cast_fp16 = linear(bias = model_encoder_layers_12_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_12_self_attn_q_proj_weight_to_fp16_palettized, x = x_127_cast_fp16)[name = tensor<string, []>("linear_73_cast_fp16")]; |
| tensor<int32, [4]> concat_48 = const()[name = tensor<string, []>("concat_48"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_49_cast_fp16 = reshape(shape = concat_48, x = linear_73_cast_fp16)[name = tensor<string, []>("q_49_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_12_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165571136))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166619776))), name = tensor<string, []>("model_encoder_layers_12_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_12_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_12_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166620352)))]; |
| tensor<fp16, [1, 249, 1024]> linear_74_cast_fp16 = linear(bias = model_encoder_layers_12_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_12_self_attn_k_proj_weight_to_fp16_palettized, x = x_127_cast_fp16)[name = tensor<string, []>("linear_74_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_12_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166622464))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167671104))), name = tensor<string, []>("model_encoder_layers_12_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167671680)))]; |
| tensor<fp16, [1, 249, 1024]> linear_75_cast_fp16 = linear(bias = model_encoder_layers_12_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_12_self_attn_v_proj_weight_to_fp16_palettized, x = x_127_cast_fp16)[name = tensor<string, []>("linear_75_cast_fp16")]; |
| tensor<int32, [4]> concat_49 = const()[name = tensor<string, []>("concat_49"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_49_cast_fp16 = reshape(shape = concat_49, x = linear_74_cast_fp16)[name = tensor<string, []>("k_49_cast_fp16")]; |
| tensor<int32, [4]> concat_50 = const()[name = tensor<string, []>("concat_50"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_51_cast_fp16 = reshape(shape = concat_50, x = linear_75_cast_fp16)[name = tensor<string, []>("v_51_cast_fp16")]; |
| tensor<int32, [4]> v_53_perm_0 = const()[name = tensor<string, []>("v_53_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_12_y_0_to_fp16 = const()[name = tensor<string, []>("mul_12_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_12_cast_fp16 = mul(x = q_49_cast_fp16, y = mul_12_y_0_to_fp16)[name = tensor<string, []>("mul_12_cast_fp16")]; |
| tensor<bool, []> matmul_12_transpose_y_0 = const()[name = tensor<string, []>("matmul_12_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_12_transpose_x_0 = const()[name = tensor<string, []>("matmul_12_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_120_perm_0 = const()[name = tensor<string, []>("transpose_120_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_121_perm_0 = const()[name = tensor<string, []>("transpose_121_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_121 = transpose(perm = transpose_121_perm_0, x = k_49_cast_fp16)[name = tensor<string, []>("transpose_190")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_120 = transpose(perm = transpose_120_perm_0, x = mul_12_cast_fp16)[name = tensor<string, []>("transpose_191")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_0, transpose_y = matmul_12_transpose_y_0, x = transpose_120, y = transpose_121)[name = tensor<string, []>("matmul_12_cast_fp16")]; |
| tensor<int32, []> softmax_12_axis_0 = const()[name = tensor<string, []>("softmax_12_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_12_cast_fp16 = softmax(axis = softmax_12_axis_0, x = matmul_12_cast_fp16)[name = tensor<string, []>("softmax_12_cast_fp16")]; |
| tensor<bool, []> attns_49_transpose_x_0 = const()[name = tensor<string, []>("attns_49_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_49_transpose_y_0 = const()[name = tensor<string, []>("attns_49_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_53_cast_fp16 = transpose(perm = v_53_perm_0, x = v_51_cast_fp16)[name = tensor<string, []>("transpose_189")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_49_cast_fp16 = matmul(transpose_x = attns_49_transpose_x_0, transpose_y = attns_49_transpose_y_0, x = softmax_12_cast_fp16, y = v_53_cast_fp16)[name = tensor<string, []>("attns_49_cast_fp16")]; |
| tensor<int32, [4]> attns_51_perm_0 = const()[name = tensor<string, []>("attns_51_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_51 = const()[name = tensor<string, []>("concat_51"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_51_cast_fp16 = transpose(perm = attns_51_perm_0, x = attns_49_cast_fp16)[name = tensor<string, []>("transpose_188")]; |
| tensor<fp16, [1, 249, 1024]> x_129_cast_fp16 = reshape(shape = concat_51, x = attns_51_cast_fp16)[name = tensor<string, []>("x_129_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_12_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167673792))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168722432))), name = tensor<string, []>("model_encoder_layers_12_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_12_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_12_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168723008)))]; |
| tensor<fp16, [1, 249, 1024]> linear_76_cast_fp16 = linear(bias = model_encoder_layers_12_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_12_self_attn_output_proj_weight_to_fp16_palettized, x = x_129_cast_fp16)[name = tensor<string, []>("linear_76_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_145_cast_fp16 = add(x = linear_76_cast_fp16, y = input_143_cast_fp16)[name = tensor<string, []>("input_145_cast_fp16")]; |
| tensor<int32, [1]> x_131_axes_0 = const()[name = tensor<string, []>("x_131_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_12_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_12_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168725120)))]; |
| tensor<fp16, [1024]> model_encoder_layers_12_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_12_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168727232)))]; |
| tensor<fp16, [1, 249, 1024]> x_131_cast_fp16 = layer_norm(axes = x_131_axes_0, beta = model_encoder_layers_12_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_12_ffn_layer_norm_weight_to_fp16, x = input_145_cast_fp16)[name = tensor<string, []>("x_131_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_12_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168729344))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172923712))), name = tensor<string, []>("model_encoder_layers_12_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_12_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_12_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172924288)))]; |
| tensor<fp16, [1, 249, 4096]> linear_77_cast_fp16 = linear(bias = model_encoder_layers_12_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_12_ffn_inner_proj_weight_to_fp16_palettized, x = x_131_cast_fp16)[name = tensor<string, []>("linear_77_cast_fp16")]; |
| tensor<string, []> input_149_mode_0 = const()[name = tensor<string, []>("input_149_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_149_cast_fp16 = gelu(mode = input_149_mode_0, x = linear_77_cast_fp16)[name = tensor<string, []>("input_149_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_12_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172932544))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177126912))), name = tensor<string, []>("model_encoder_layers_12_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_12_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_12_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177127488)))]; |
| tensor<fp16, [1, 249, 1024]> linear_78_cast_fp16 = linear(bias = model_encoder_layers_12_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_12_ffn_output_proj_weight_to_fp16_palettized, x = input_149_cast_fp16)[name = tensor<string, []>("linear_78_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_151_cast_fp16 = add(x = linear_78_cast_fp16, y = input_145_cast_fp16)[name = tensor<string, []>("input_151_cast_fp16")]; |
| tensor<int32, [1]> x_135_axes_0 = const()[name = tensor<string, []>("x_135_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_13_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_13_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177129600)))]; |
| tensor<fp16, [1024]> model_encoder_layers_13_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_13_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177131712)))]; |
| tensor<fp16, [1, 249, 1024]> x_135_cast_fp16 = layer_norm(axes = x_135_axes_0, beta = model_encoder_layers_13_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_13_self_attn_layer_norm_weight_to_fp16, x = input_151_cast_fp16)[name = tensor<string, []>("x_135_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_13_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177133824))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178182464))), name = tensor<string, []>("model_encoder_layers_13_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178183040)))]; |
| tensor<fp16, [1, 249, 1024]> linear_79_cast_fp16 = linear(bias = model_encoder_layers_13_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_13_self_attn_q_proj_weight_to_fp16_palettized, x = x_135_cast_fp16)[name = tensor<string, []>("linear_79_cast_fp16")]; |
| tensor<int32, [4]> concat_52 = const()[name = tensor<string, []>("concat_52"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_53_cast_fp16 = reshape(shape = concat_52, x = linear_79_cast_fp16)[name = tensor<string, []>("q_53_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_13_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178185152))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179233792))), name = tensor<string, []>("model_encoder_layers_13_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_13_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_13_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179234368)))]; |
| tensor<fp16, [1, 249, 1024]> linear_80_cast_fp16 = linear(bias = model_encoder_layers_13_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_13_self_attn_k_proj_weight_to_fp16_palettized, x = x_135_cast_fp16)[name = tensor<string, []>("linear_80_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_13_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179236480))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180285120))), name = tensor<string, []>("model_encoder_layers_13_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180285696)))]; |
| tensor<fp16, [1, 249, 1024]> linear_81_cast_fp16 = linear(bias = model_encoder_layers_13_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_13_self_attn_v_proj_weight_to_fp16_palettized, x = x_135_cast_fp16)[name = tensor<string, []>("linear_81_cast_fp16")]; |
| tensor<int32, [4]> concat_53 = const()[name = tensor<string, []>("concat_53"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_53_cast_fp16 = reshape(shape = concat_53, x = linear_80_cast_fp16)[name = tensor<string, []>("k_53_cast_fp16")]; |
| tensor<int32, [4]> concat_54 = const()[name = tensor<string, []>("concat_54"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_55_cast_fp16 = reshape(shape = concat_54, x = linear_81_cast_fp16)[name = tensor<string, []>("v_55_cast_fp16")]; |
| tensor<int32, [4]> v_57_perm_0 = const()[name = tensor<string, []>("v_57_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_13_y_0_to_fp16 = const()[name = tensor<string, []>("mul_13_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_13_cast_fp16 = mul(x = q_53_cast_fp16, y = mul_13_y_0_to_fp16)[name = tensor<string, []>("mul_13_cast_fp16")]; |
| tensor<bool, []> matmul_13_transpose_y_0 = const()[name = tensor<string, []>("matmul_13_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_13_transpose_x_0 = const()[name = tensor<string, []>("matmul_13_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_122_perm_0 = const()[name = tensor<string, []>("transpose_122_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_123_perm_0 = const()[name = tensor<string, []>("transpose_123_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_123 = transpose(perm = transpose_123_perm_0, x = k_53_cast_fp16)[name = tensor<string, []>("transpose_186")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_122 = transpose(perm = transpose_122_perm_0, x = mul_13_cast_fp16)[name = tensor<string, []>("transpose_187")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_0, transpose_y = matmul_13_transpose_y_0, x = transpose_122, y = transpose_123)[name = tensor<string, []>("matmul_13_cast_fp16")]; |
| tensor<int32, []> softmax_13_axis_0 = const()[name = tensor<string, []>("softmax_13_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_13_cast_fp16 = softmax(axis = softmax_13_axis_0, x = matmul_13_cast_fp16)[name = tensor<string, []>("softmax_13_cast_fp16")]; |
| tensor<bool, []> attns_53_transpose_x_0 = const()[name = tensor<string, []>("attns_53_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_53_transpose_y_0 = const()[name = tensor<string, []>("attns_53_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_57_cast_fp16 = transpose(perm = v_57_perm_0, x = v_55_cast_fp16)[name = tensor<string, []>("transpose_185")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_53_cast_fp16 = matmul(transpose_x = attns_53_transpose_x_0, transpose_y = attns_53_transpose_y_0, x = softmax_13_cast_fp16, y = v_57_cast_fp16)[name = tensor<string, []>("attns_53_cast_fp16")]; |
| tensor<int32, [4]> attns_55_perm_0 = const()[name = tensor<string, []>("attns_55_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_55 = const()[name = tensor<string, []>("concat_55"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_55_cast_fp16 = transpose(perm = attns_55_perm_0, x = attns_53_cast_fp16)[name = tensor<string, []>("transpose_184")]; |
| tensor<fp16, [1, 249, 1024]> x_137_cast_fp16 = reshape(shape = concat_55, x = attns_55_cast_fp16)[name = tensor<string, []>("x_137_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_13_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180287808))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181336448))), name = tensor<string, []>("model_encoder_layers_13_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_13_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_13_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181337024)))]; |
| tensor<fp16, [1, 249, 1024]> linear_82_cast_fp16 = linear(bias = model_encoder_layers_13_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_13_self_attn_output_proj_weight_to_fp16_palettized, x = x_137_cast_fp16)[name = tensor<string, []>("linear_82_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_153_cast_fp16 = add(x = linear_82_cast_fp16, y = input_151_cast_fp16)[name = tensor<string, []>("input_153_cast_fp16")]; |
| tensor<int32, [1]> x_139_axes_0 = const()[name = tensor<string, []>("x_139_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_13_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_13_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181339136)))]; |
| tensor<fp16, [1024]> model_encoder_layers_13_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_13_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181341248)))]; |
| tensor<fp16, [1, 249, 1024]> x_139_cast_fp16 = layer_norm(axes = x_139_axes_0, beta = model_encoder_layers_13_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_13_ffn_layer_norm_weight_to_fp16, x = input_153_cast_fp16)[name = tensor<string, []>("x_139_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_13_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181343360))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185537728))), name = tensor<string, []>("model_encoder_layers_13_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_13_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_13_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185538304)))]; |
| tensor<fp16, [1, 249, 4096]> linear_83_cast_fp16 = linear(bias = model_encoder_layers_13_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_13_ffn_inner_proj_weight_to_fp16_palettized, x = x_139_cast_fp16)[name = tensor<string, []>("linear_83_cast_fp16")]; |
| tensor<string, []> input_157_mode_0 = const()[name = tensor<string, []>("input_157_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_157_cast_fp16 = gelu(mode = input_157_mode_0, x = linear_83_cast_fp16)[name = tensor<string, []>("input_157_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_13_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185546560))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189740928))), name = tensor<string, []>("model_encoder_layers_13_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_13_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_13_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189741504)))]; |
| tensor<fp16, [1, 249, 1024]> linear_84_cast_fp16 = linear(bias = model_encoder_layers_13_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_13_ffn_output_proj_weight_to_fp16_palettized, x = input_157_cast_fp16)[name = tensor<string, []>("linear_84_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_159_cast_fp16 = add(x = linear_84_cast_fp16, y = input_153_cast_fp16)[name = tensor<string, []>("input_159_cast_fp16")]; |
| tensor<int32, [1]> x_143_axes_0 = const()[name = tensor<string, []>("x_143_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_14_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_14_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189743616)))]; |
| tensor<fp16, [1024]> model_encoder_layers_14_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_14_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189745728)))]; |
| tensor<fp16, [1, 249, 1024]> x_143_cast_fp16 = layer_norm(axes = x_143_axes_0, beta = model_encoder_layers_14_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_14_self_attn_layer_norm_weight_to_fp16, x = input_159_cast_fp16)[name = tensor<string, []>("x_143_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_14_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189747840))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190796480))), name = tensor<string, []>("model_encoder_layers_14_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190797056)))]; |
| tensor<fp16, [1, 249, 1024]> linear_85_cast_fp16 = linear(bias = model_encoder_layers_14_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_14_self_attn_q_proj_weight_to_fp16_palettized, x = x_143_cast_fp16)[name = tensor<string, []>("linear_85_cast_fp16")]; |
| tensor<int32, [4]> concat_56 = const()[name = tensor<string, []>("concat_56"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_57_cast_fp16 = reshape(shape = concat_56, x = linear_85_cast_fp16)[name = tensor<string, []>("q_57_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_14_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190799168))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191847808))), name = tensor<string, []>("model_encoder_layers_14_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_14_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_14_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191848384)))]; |
| tensor<fp16, [1, 249, 1024]> linear_86_cast_fp16 = linear(bias = model_encoder_layers_14_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_14_self_attn_k_proj_weight_to_fp16_palettized, x = x_143_cast_fp16)[name = tensor<string, []>("linear_86_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_14_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191850496))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192899136))), name = tensor<string, []>("model_encoder_layers_14_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192899712)))]; |
| tensor<fp16, [1, 249, 1024]> linear_87_cast_fp16 = linear(bias = model_encoder_layers_14_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_14_self_attn_v_proj_weight_to_fp16_palettized, x = x_143_cast_fp16)[name = tensor<string, []>("linear_87_cast_fp16")]; |
| tensor<int32, [4]> concat_57 = const()[name = tensor<string, []>("concat_57"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_57_cast_fp16 = reshape(shape = concat_57, x = linear_86_cast_fp16)[name = tensor<string, []>("k_57_cast_fp16")]; |
| tensor<int32, [4]> concat_58 = const()[name = tensor<string, []>("concat_58"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_59_cast_fp16 = reshape(shape = concat_58, x = linear_87_cast_fp16)[name = tensor<string, []>("v_59_cast_fp16")]; |
| tensor<int32, [4]> v_61_perm_0 = const()[name = tensor<string, []>("v_61_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_14_y_0_to_fp16 = const()[name = tensor<string, []>("mul_14_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_14_cast_fp16 = mul(x = q_57_cast_fp16, y = mul_14_y_0_to_fp16)[name = tensor<string, []>("mul_14_cast_fp16")]; |
| tensor<bool, []> matmul_14_transpose_y_0 = const()[name = tensor<string, []>("matmul_14_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_14_transpose_x_0 = const()[name = tensor<string, []>("matmul_14_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_124_perm_0 = const()[name = tensor<string, []>("transpose_124_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_125_perm_0 = const()[name = tensor<string, []>("transpose_125_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_125 = transpose(perm = transpose_125_perm_0, x = k_57_cast_fp16)[name = tensor<string, []>("transpose_182")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_124 = transpose(perm = transpose_124_perm_0, x = mul_14_cast_fp16)[name = tensor<string, []>("transpose_183")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_14_cast_fp16 = matmul(transpose_x = matmul_14_transpose_x_0, transpose_y = matmul_14_transpose_y_0, x = transpose_124, y = transpose_125)[name = tensor<string, []>("matmul_14_cast_fp16")]; |
| tensor<int32, []> softmax_14_axis_0 = const()[name = tensor<string, []>("softmax_14_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_14_cast_fp16 = softmax(axis = softmax_14_axis_0, x = matmul_14_cast_fp16)[name = tensor<string, []>("softmax_14_cast_fp16")]; |
| tensor<bool, []> attns_57_transpose_x_0 = const()[name = tensor<string, []>("attns_57_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_57_transpose_y_0 = const()[name = tensor<string, []>("attns_57_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_61_cast_fp16 = transpose(perm = v_61_perm_0, x = v_59_cast_fp16)[name = tensor<string, []>("transpose_181")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_57_cast_fp16 = matmul(transpose_x = attns_57_transpose_x_0, transpose_y = attns_57_transpose_y_0, x = softmax_14_cast_fp16, y = v_61_cast_fp16)[name = tensor<string, []>("attns_57_cast_fp16")]; |
| tensor<int32, [4]> attns_59_perm_0 = const()[name = tensor<string, []>("attns_59_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_59 = const()[name = tensor<string, []>("concat_59"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_59_cast_fp16 = transpose(perm = attns_59_perm_0, x = attns_57_cast_fp16)[name = tensor<string, []>("transpose_180")]; |
| tensor<fp16, [1, 249, 1024]> x_145_cast_fp16 = reshape(shape = concat_59, x = attns_59_cast_fp16)[name = tensor<string, []>("x_145_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_14_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192901824))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193950464))), name = tensor<string, []>("model_encoder_layers_14_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_14_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_14_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193951040)))]; |
| tensor<fp16, [1, 249, 1024]> linear_88_cast_fp16 = linear(bias = model_encoder_layers_14_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_14_self_attn_output_proj_weight_to_fp16_palettized, x = x_145_cast_fp16)[name = tensor<string, []>("linear_88_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_161_cast_fp16 = add(x = linear_88_cast_fp16, y = input_159_cast_fp16)[name = tensor<string, []>("input_161_cast_fp16")]; |
| tensor<int32, [1]> x_147_axes_0 = const()[name = tensor<string, []>("x_147_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_14_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_14_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193953152)))]; |
| tensor<fp16, [1024]> model_encoder_layers_14_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_14_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193955264)))]; |
| tensor<fp16, [1, 249, 1024]> x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = model_encoder_layers_14_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_14_ffn_layer_norm_weight_to_fp16, x = input_161_cast_fp16)[name = tensor<string, []>("x_147_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_14_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193957376))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198151744))), name = tensor<string, []>("model_encoder_layers_14_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_14_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_14_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198152320)))]; |
| tensor<fp16, [1, 249, 4096]> linear_89_cast_fp16 = linear(bias = model_encoder_layers_14_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_14_ffn_inner_proj_weight_to_fp16_palettized, x = x_147_cast_fp16)[name = tensor<string, []>("linear_89_cast_fp16")]; |
| tensor<string, []> input_165_mode_0 = const()[name = tensor<string, []>("input_165_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_165_cast_fp16 = gelu(mode = input_165_mode_0, x = linear_89_cast_fp16)[name = tensor<string, []>("input_165_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_14_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198160576))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202354944))), name = tensor<string, []>("model_encoder_layers_14_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_14_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_14_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202355520)))]; |
| tensor<fp16, [1, 249, 1024]> linear_90_cast_fp16 = linear(bias = model_encoder_layers_14_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_14_ffn_output_proj_weight_to_fp16_palettized, x = input_165_cast_fp16)[name = tensor<string, []>("linear_90_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_167_cast_fp16 = add(x = linear_90_cast_fp16, y = input_161_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")]; |
| tensor<int32, [1]> x_151_axes_0 = const()[name = tensor<string, []>("x_151_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_15_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_15_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202357632)))]; |
| tensor<fp16, [1024]> model_encoder_layers_15_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_15_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202359744)))]; |
| tensor<fp16, [1, 249, 1024]> x_151_cast_fp16 = layer_norm(axes = x_151_axes_0, beta = model_encoder_layers_15_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_15_self_attn_layer_norm_weight_to_fp16, x = input_167_cast_fp16)[name = tensor<string, []>("x_151_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_15_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202361856))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203410496))), name = tensor<string, []>("model_encoder_layers_15_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203411072)))]; |
| tensor<fp16, [1, 249, 1024]> linear_91_cast_fp16 = linear(bias = model_encoder_layers_15_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_15_self_attn_q_proj_weight_to_fp16_palettized, x = x_151_cast_fp16)[name = tensor<string, []>("linear_91_cast_fp16")]; |
| tensor<int32, [4]> concat_60 = const()[name = tensor<string, []>("concat_60"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_61_cast_fp16 = reshape(shape = concat_60, x = linear_91_cast_fp16)[name = tensor<string, []>("q_61_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_15_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203413184))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(204461824))), name = tensor<string, []>("model_encoder_layers_15_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_15_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_15_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(204462400)))]; |
| tensor<fp16, [1, 249, 1024]> linear_92_cast_fp16 = linear(bias = model_encoder_layers_15_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_15_self_attn_k_proj_weight_to_fp16_palettized, x = x_151_cast_fp16)[name = tensor<string, []>("linear_92_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_15_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(204464512))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205513152))), name = tensor<string, []>("model_encoder_layers_15_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205513728)))]; |
| tensor<fp16, [1, 249, 1024]> linear_93_cast_fp16 = linear(bias = model_encoder_layers_15_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_15_self_attn_v_proj_weight_to_fp16_palettized, x = x_151_cast_fp16)[name = tensor<string, []>("linear_93_cast_fp16")]; |
| tensor<int32, [4]> concat_61 = const()[name = tensor<string, []>("concat_61"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_61_cast_fp16 = reshape(shape = concat_61, x = linear_92_cast_fp16)[name = tensor<string, []>("k_61_cast_fp16")]; |
| tensor<int32, [4]> concat_62 = const()[name = tensor<string, []>("concat_62"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_63_cast_fp16 = reshape(shape = concat_62, x = linear_93_cast_fp16)[name = tensor<string, []>("v_63_cast_fp16")]; |
| tensor<int32, [4]> v_65_perm_0 = const()[name = tensor<string, []>("v_65_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_15_y_0_to_fp16 = const()[name = tensor<string, []>("mul_15_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_15_cast_fp16 = mul(x = q_61_cast_fp16, y = mul_15_y_0_to_fp16)[name = tensor<string, []>("mul_15_cast_fp16")]; |
| tensor<bool, []> matmul_15_transpose_y_0 = const()[name = tensor<string, []>("matmul_15_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_15_transpose_x_0 = const()[name = tensor<string, []>("matmul_15_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_126_perm_0 = const()[name = tensor<string, []>("transpose_126_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_127_perm_0 = const()[name = tensor<string, []>("transpose_127_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_127 = transpose(perm = transpose_127_perm_0, x = k_61_cast_fp16)[name = tensor<string, []>("transpose_178")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_126 = transpose(perm = transpose_126_perm_0, x = mul_15_cast_fp16)[name = tensor<string, []>("transpose_179")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_15_cast_fp16 = matmul(transpose_x = matmul_15_transpose_x_0, transpose_y = matmul_15_transpose_y_0, x = transpose_126, y = transpose_127)[name = tensor<string, []>("matmul_15_cast_fp16")]; |
| tensor<int32, []> softmax_15_axis_0 = const()[name = tensor<string, []>("softmax_15_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_15_cast_fp16 = softmax(axis = softmax_15_axis_0, x = matmul_15_cast_fp16)[name = tensor<string, []>("softmax_15_cast_fp16")]; |
| tensor<bool, []> attns_61_transpose_x_0 = const()[name = tensor<string, []>("attns_61_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_61_transpose_y_0 = const()[name = tensor<string, []>("attns_61_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_65_cast_fp16 = transpose(perm = v_65_perm_0, x = v_63_cast_fp16)[name = tensor<string, []>("transpose_177")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_61_cast_fp16 = matmul(transpose_x = attns_61_transpose_x_0, transpose_y = attns_61_transpose_y_0, x = softmax_15_cast_fp16, y = v_65_cast_fp16)[name = tensor<string, []>("attns_61_cast_fp16")]; |
| tensor<int32, [4]> attns_63_perm_0 = const()[name = tensor<string, []>("attns_63_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_63 = const()[name = tensor<string, []>("concat_63"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_63_cast_fp16 = transpose(perm = attns_63_perm_0, x = attns_61_cast_fp16)[name = tensor<string, []>("transpose_176")]; |
| tensor<fp16, [1, 249, 1024]> x_153_cast_fp16 = reshape(shape = concat_63, x = attns_63_cast_fp16)[name = tensor<string, []>("x_153_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_15_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205515840))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206564480))), name = tensor<string, []>("model_encoder_layers_15_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_15_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_15_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206565056)))]; |
| tensor<fp16, [1, 249, 1024]> linear_94_cast_fp16 = linear(bias = model_encoder_layers_15_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_15_self_attn_output_proj_weight_to_fp16_palettized, x = x_153_cast_fp16)[name = tensor<string, []>("linear_94_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_169_cast_fp16 = add(x = linear_94_cast_fp16, y = input_167_cast_fp16)[name = tensor<string, []>("input_169_cast_fp16")]; |
| tensor<int32, [1]> x_155_axes_0 = const()[name = tensor<string, []>("x_155_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_15_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_15_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206567168)))]; |
| tensor<fp16, [1024]> model_encoder_layers_15_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_15_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206569280)))]; |
| tensor<fp16, [1, 249, 1024]> x_155_cast_fp16 = layer_norm(axes = x_155_axes_0, beta = model_encoder_layers_15_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_15_ffn_layer_norm_weight_to_fp16, x = input_169_cast_fp16)[name = tensor<string, []>("x_155_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_15_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206571392))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210765760))), name = tensor<string, []>("model_encoder_layers_15_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_15_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_15_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210766336)))]; |
| tensor<fp16, [1, 249, 4096]> linear_95_cast_fp16 = linear(bias = model_encoder_layers_15_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_15_ffn_inner_proj_weight_to_fp16_palettized, x = x_155_cast_fp16)[name = tensor<string, []>("linear_95_cast_fp16")]; |
| tensor<string, []> input_173_mode_0 = const()[name = tensor<string, []>("input_173_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_173_cast_fp16 = gelu(mode = input_173_mode_0, x = linear_95_cast_fp16)[name = tensor<string, []>("input_173_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_15_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210774592))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214968960))), name = tensor<string, []>("model_encoder_layers_15_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_15_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_15_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214969536)))]; |
| tensor<fp16, [1, 249, 1024]> linear_96_cast_fp16 = linear(bias = model_encoder_layers_15_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_15_ffn_output_proj_weight_to_fp16_palettized, x = input_173_cast_fp16)[name = tensor<string, []>("linear_96_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_175_cast_fp16 = add(x = linear_96_cast_fp16, y = input_169_cast_fp16)[name = tensor<string, []>("input_175_cast_fp16")]; |
| tensor<int32, [1]> x_159_axes_0 = const()[name = tensor<string, []>("x_159_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_16_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_16_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214971648)))]; |
| tensor<fp16, [1024]> model_encoder_layers_16_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_16_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214973760)))]; |
| tensor<fp16, [1, 249, 1024]> x_159_cast_fp16 = layer_norm(axes = x_159_axes_0, beta = model_encoder_layers_16_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_16_self_attn_layer_norm_weight_to_fp16, x = input_175_cast_fp16)[name = tensor<string, []>("x_159_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_16_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214975872))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216024512))), name = tensor<string, []>("model_encoder_layers_16_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216025088)))]; |
| tensor<fp16, [1, 249, 1024]> linear_97_cast_fp16 = linear(bias = model_encoder_layers_16_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_16_self_attn_q_proj_weight_to_fp16_palettized, x = x_159_cast_fp16)[name = tensor<string, []>("linear_97_cast_fp16")]; |
| tensor<int32, [4]> concat_64 = const()[name = tensor<string, []>("concat_64"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_65_cast_fp16 = reshape(shape = concat_64, x = linear_97_cast_fp16)[name = tensor<string, []>("q_65_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_16_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216027200))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217075840))), name = tensor<string, []>("model_encoder_layers_16_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_16_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_16_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217076416)))]; |
| tensor<fp16, [1, 249, 1024]> linear_98_cast_fp16 = linear(bias = model_encoder_layers_16_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_16_self_attn_k_proj_weight_to_fp16_palettized, x = x_159_cast_fp16)[name = tensor<string, []>("linear_98_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_16_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217078528))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218127168))), name = tensor<string, []>("model_encoder_layers_16_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218127744)))]; |
| tensor<fp16, [1, 249, 1024]> linear_99_cast_fp16 = linear(bias = model_encoder_layers_16_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_16_self_attn_v_proj_weight_to_fp16_palettized, x = x_159_cast_fp16)[name = tensor<string, []>("linear_99_cast_fp16")]; |
| tensor<int32, [4]> concat_65 = const()[name = tensor<string, []>("concat_65"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_65_cast_fp16 = reshape(shape = concat_65, x = linear_98_cast_fp16)[name = tensor<string, []>("k_65_cast_fp16")]; |
| tensor<int32, [4]> concat_66 = const()[name = tensor<string, []>("concat_66"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_67_cast_fp16 = reshape(shape = concat_66, x = linear_99_cast_fp16)[name = tensor<string, []>("v_67_cast_fp16")]; |
| tensor<int32, [4]> v_69_perm_0 = const()[name = tensor<string, []>("v_69_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_16_y_0_to_fp16 = const()[name = tensor<string, []>("mul_16_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_16_cast_fp16 = mul(x = q_65_cast_fp16, y = mul_16_y_0_to_fp16)[name = tensor<string, []>("mul_16_cast_fp16")]; |
| tensor<bool, []> matmul_16_transpose_y_0 = const()[name = tensor<string, []>("matmul_16_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_16_transpose_x_0 = const()[name = tensor<string, []>("matmul_16_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_128_perm_0 = const()[name = tensor<string, []>("transpose_128_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_129_perm_0 = const()[name = tensor<string, []>("transpose_129_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_129 = transpose(perm = transpose_129_perm_0, x = k_65_cast_fp16)[name = tensor<string, []>("transpose_174")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_128 = transpose(perm = transpose_128_perm_0, x = mul_16_cast_fp16)[name = tensor<string, []>("transpose_175")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_16_cast_fp16 = matmul(transpose_x = matmul_16_transpose_x_0, transpose_y = matmul_16_transpose_y_0, x = transpose_128, y = transpose_129)[name = tensor<string, []>("matmul_16_cast_fp16")]; |
| tensor<int32, []> softmax_16_axis_0 = const()[name = tensor<string, []>("softmax_16_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_16_cast_fp16 = softmax(axis = softmax_16_axis_0, x = matmul_16_cast_fp16)[name = tensor<string, []>("softmax_16_cast_fp16")]; |
| tensor<bool, []> attns_65_transpose_x_0 = const()[name = tensor<string, []>("attns_65_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_65_transpose_y_0 = const()[name = tensor<string, []>("attns_65_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_69_cast_fp16 = transpose(perm = v_69_perm_0, x = v_67_cast_fp16)[name = tensor<string, []>("transpose_173")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_65_cast_fp16 = matmul(transpose_x = attns_65_transpose_x_0, transpose_y = attns_65_transpose_y_0, x = softmax_16_cast_fp16, y = v_69_cast_fp16)[name = tensor<string, []>("attns_65_cast_fp16")]; |
| tensor<int32, [4]> attns_67_perm_0 = const()[name = tensor<string, []>("attns_67_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_67 = const()[name = tensor<string, []>("concat_67"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_67_cast_fp16 = transpose(perm = attns_67_perm_0, x = attns_65_cast_fp16)[name = tensor<string, []>("transpose_172")]; |
| tensor<fp16, [1, 249, 1024]> x_161_cast_fp16 = reshape(shape = concat_67, x = attns_67_cast_fp16)[name = tensor<string, []>("x_161_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_16_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218129856))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(219178496))), name = tensor<string, []>("model_encoder_layers_16_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_16_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_16_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(219179072)))]; |
| tensor<fp16, [1, 249, 1024]> linear_100_cast_fp16 = linear(bias = model_encoder_layers_16_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_16_self_attn_output_proj_weight_to_fp16_palettized, x = x_161_cast_fp16)[name = tensor<string, []>("linear_100_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_177_cast_fp16 = add(x = linear_100_cast_fp16, y = input_175_cast_fp16)[name = tensor<string, []>("input_177_cast_fp16")]; |
| tensor<int32, [1]> x_163_axes_0 = const()[name = tensor<string, []>("x_163_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_16_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_16_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(219181184)))]; |
| tensor<fp16, [1024]> model_encoder_layers_16_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_16_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(219183296)))]; |
| tensor<fp16, [1, 249, 1024]> x_163_cast_fp16 = layer_norm(axes = x_163_axes_0, beta = model_encoder_layers_16_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_16_ffn_layer_norm_weight_to_fp16, x = input_177_cast_fp16)[name = tensor<string, []>("x_163_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_16_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(219185408))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(223379776))), name = tensor<string, []>("model_encoder_layers_16_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_16_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_16_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(223380352)))]; |
| tensor<fp16, [1, 249, 4096]> linear_101_cast_fp16 = linear(bias = model_encoder_layers_16_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_16_ffn_inner_proj_weight_to_fp16_palettized, x = x_163_cast_fp16)[name = tensor<string, []>("linear_101_cast_fp16")]; |
| tensor<string, []> input_181_mode_0 = const()[name = tensor<string, []>("input_181_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = linear_101_cast_fp16)[name = tensor<string, []>("input_181_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_16_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(223388608))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227582976))), name = tensor<string, []>("model_encoder_layers_16_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_16_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_16_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227583552)))]; |
| tensor<fp16, [1, 249, 1024]> linear_102_cast_fp16 = linear(bias = model_encoder_layers_16_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_16_ffn_output_proj_weight_to_fp16_palettized, x = input_181_cast_fp16)[name = tensor<string, []>("linear_102_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_183_cast_fp16 = add(x = linear_102_cast_fp16, y = input_177_cast_fp16)[name = tensor<string, []>("input_183_cast_fp16")]; |
| tensor<int32, [1]> x_167_axes_0 = const()[name = tensor<string, []>("x_167_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_17_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_17_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227585664)))]; |
| tensor<fp16, [1024]> model_encoder_layers_17_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_17_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227587776)))]; |
| tensor<fp16, [1, 249, 1024]> x_167_cast_fp16 = layer_norm(axes = x_167_axes_0, beta = model_encoder_layers_17_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_17_self_attn_layer_norm_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("x_167_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_17_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227589888))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228638528))), name = tensor<string, []>("model_encoder_layers_17_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228639104)))]; |
| tensor<fp16, [1, 249, 1024]> linear_103_cast_fp16 = linear(bias = model_encoder_layers_17_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_17_self_attn_q_proj_weight_to_fp16_palettized, x = x_167_cast_fp16)[name = tensor<string, []>("linear_103_cast_fp16")]; |
| tensor<int32, [4]> concat_68 = const()[name = tensor<string, []>("concat_68"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_69_cast_fp16 = reshape(shape = concat_68, x = linear_103_cast_fp16)[name = tensor<string, []>("q_69_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_17_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228641216))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(229689856))), name = tensor<string, []>("model_encoder_layers_17_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_17_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_17_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(229690432)))]; |
| tensor<fp16, [1, 249, 1024]> linear_104_cast_fp16 = linear(bias = model_encoder_layers_17_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_17_self_attn_k_proj_weight_to_fp16_palettized, x = x_167_cast_fp16)[name = tensor<string, []>("linear_104_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_17_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(229692544))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(230741184))), name = tensor<string, []>("model_encoder_layers_17_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(230741760)))]; |
| tensor<fp16, [1, 249, 1024]> linear_105_cast_fp16 = linear(bias = model_encoder_layers_17_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_17_self_attn_v_proj_weight_to_fp16_palettized, x = x_167_cast_fp16)[name = tensor<string, []>("linear_105_cast_fp16")]; |
| tensor<int32, [4]> concat_69 = const()[name = tensor<string, []>("concat_69"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_69_cast_fp16 = reshape(shape = concat_69, x = linear_104_cast_fp16)[name = tensor<string, []>("k_69_cast_fp16")]; |
| tensor<int32, [4]> concat_70 = const()[name = tensor<string, []>("concat_70"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_71_cast_fp16 = reshape(shape = concat_70, x = linear_105_cast_fp16)[name = tensor<string, []>("v_71_cast_fp16")]; |
| tensor<int32, [4]> v_73_perm_0 = const()[name = tensor<string, []>("v_73_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_17_y_0_to_fp16 = const()[name = tensor<string, []>("mul_17_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_17_cast_fp16 = mul(x = q_69_cast_fp16, y = mul_17_y_0_to_fp16)[name = tensor<string, []>("mul_17_cast_fp16")]; |
| tensor<bool, []> matmul_17_transpose_y_0 = const()[name = tensor<string, []>("matmul_17_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_17_transpose_x_0 = const()[name = tensor<string, []>("matmul_17_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_130_perm_0 = const()[name = tensor<string, []>("transpose_130_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_131_perm_0 = const()[name = tensor<string, []>("transpose_131_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_131 = transpose(perm = transpose_131_perm_0, x = k_69_cast_fp16)[name = tensor<string, []>("transpose_170")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_130 = transpose(perm = transpose_130_perm_0, x = mul_17_cast_fp16)[name = tensor<string, []>("transpose_171")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_17_cast_fp16 = matmul(transpose_x = matmul_17_transpose_x_0, transpose_y = matmul_17_transpose_y_0, x = transpose_130, y = transpose_131)[name = tensor<string, []>("matmul_17_cast_fp16")]; |
| tensor<int32, []> softmax_17_axis_0 = const()[name = tensor<string, []>("softmax_17_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_17_cast_fp16 = softmax(axis = softmax_17_axis_0, x = matmul_17_cast_fp16)[name = tensor<string, []>("softmax_17_cast_fp16")]; |
| tensor<bool, []> attns_69_transpose_x_0 = const()[name = tensor<string, []>("attns_69_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_69_transpose_y_0 = const()[name = tensor<string, []>("attns_69_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_73_cast_fp16 = transpose(perm = v_73_perm_0, x = v_71_cast_fp16)[name = tensor<string, []>("transpose_169")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_69_cast_fp16 = matmul(transpose_x = attns_69_transpose_x_0, transpose_y = attns_69_transpose_y_0, x = softmax_17_cast_fp16, y = v_73_cast_fp16)[name = tensor<string, []>("attns_69_cast_fp16")]; |
| tensor<int32, [4]> attns_71_perm_0 = const()[name = tensor<string, []>("attns_71_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_71 = const()[name = tensor<string, []>("concat_71"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_71_cast_fp16 = transpose(perm = attns_71_perm_0, x = attns_69_cast_fp16)[name = tensor<string, []>("transpose_168")]; |
| tensor<fp16, [1, 249, 1024]> x_169_cast_fp16 = reshape(shape = concat_71, x = attns_71_cast_fp16)[name = tensor<string, []>("x_169_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_17_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(230743872))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231792512))), name = tensor<string, []>("model_encoder_layers_17_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_17_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_17_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231793088)))]; |
| tensor<fp16, [1, 249, 1024]> linear_106_cast_fp16 = linear(bias = model_encoder_layers_17_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_17_self_attn_output_proj_weight_to_fp16_palettized, x = x_169_cast_fp16)[name = tensor<string, []>("linear_106_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_185_cast_fp16 = add(x = linear_106_cast_fp16, y = input_183_cast_fp16)[name = tensor<string, []>("input_185_cast_fp16")]; |
| tensor<int32, [1]> x_171_axes_0 = const()[name = tensor<string, []>("x_171_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_17_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_17_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231795200)))]; |
| tensor<fp16, [1024]> model_encoder_layers_17_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_17_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231797312)))]; |
| tensor<fp16, [1, 249, 1024]> x_171_cast_fp16 = layer_norm(axes = x_171_axes_0, beta = model_encoder_layers_17_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_17_ffn_layer_norm_weight_to_fp16, x = input_185_cast_fp16)[name = tensor<string, []>("x_171_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_17_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231799424))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235993792))), name = tensor<string, []>("model_encoder_layers_17_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_17_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_17_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235994368)))]; |
| tensor<fp16, [1, 249, 4096]> linear_107_cast_fp16 = linear(bias = model_encoder_layers_17_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_17_ffn_inner_proj_weight_to_fp16_palettized, x = x_171_cast_fp16)[name = tensor<string, []>("linear_107_cast_fp16")]; |
| tensor<string, []> input_189_mode_0 = const()[name = tensor<string, []>("input_189_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_189_cast_fp16 = gelu(mode = input_189_mode_0, x = linear_107_cast_fp16)[name = tensor<string, []>("input_189_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_17_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236002624))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240196992))), name = tensor<string, []>("model_encoder_layers_17_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_17_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_17_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240197568)))]; |
| tensor<fp16, [1, 249, 1024]> linear_108_cast_fp16 = linear(bias = model_encoder_layers_17_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_17_ffn_output_proj_weight_to_fp16_palettized, x = input_189_cast_fp16)[name = tensor<string, []>("linear_108_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_191_cast_fp16 = add(x = linear_108_cast_fp16, y = input_185_cast_fp16)[name = tensor<string, []>("input_191_cast_fp16")]; |
| tensor<int32, [1]> x_175_axes_0 = const()[name = tensor<string, []>("x_175_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_18_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_18_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240199680)))]; |
| tensor<fp16, [1024]> model_encoder_layers_18_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_18_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240201792)))]; |
| tensor<fp16, [1, 249, 1024]> x_175_cast_fp16 = layer_norm(axes = x_175_axes_0, beta = model_encoder_layers_18_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_18_self_attn_layer_norm_weight_to_fp16, x = input_191_cast_fp16)[name = tensor<string, []>("x_175_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_18_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240203904))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241252544))), name = tensor<string, []>("model_encoder_layers_18_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241253120)))]; |
| tensor<fp16, [1, 249, 1024]> linear_109_cast_fp16 = linear(bias = model_encoder_layers_18_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_18_self_attn_q_proj_weight_to_fp16_palettized, x = x_175_cast_fp16)[name = tensor<string, []>("linear_109_cast_fp16")]; |
| tensor<int32, [4]> concat_72 = const()[name = tensor<string, []>("concat_72"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_73_cast_fp16 = reshape(shape = concat_72, x = linear_109_cast_fp16)[name = tensor<string, []>("q_73_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_18_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241255232))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(242303872))), name = tensor<string, []>("model_encoder_layers_18_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_18_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_18_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(242304448)))]; |
| tensor<fp16, [1, 249, 1024]> linear_110_cast_fp16 = linear(bias = model_encoder_layers_18_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_18_self_attn_k_proj_weight_to_fp16_palettized, x = x_175_cast_fp16)[name = tensor<string, []>("linear_110_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_18_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(242306560))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243355200))), name = tensor<string, []>("model_encoder_layers_18_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243355776)))]; |
| tensor<fp16, [1, 249, 1024]> linear_111_cast_fp16 = linear(bias = model_encoder_layers_18_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_18_self_attn_v_proj_weight_to_fp16_palettized, x = x_175_cast_fp16)[name = tensor<string, []>("linear_111_cast_fp16")]; |
| tensor<int32, [4]> concat_73 = const()[name = tensor<string, []>("concat_73"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_73_cast_fp16 = reshape(shape = concat_73, x = linear_110_cast_fp16)[name = tensor<string, []>("k_73_cast_fp16")]; |
| tensor<int32, [4]> concat_74 = const()[name = tensor<string, []>("concat_74"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_75_cast_fp16 = reshape(shape = concat_74, x = linear_111_cast_fp16)[name = tensor<string, []>("v_75_cast_fp16")]; |
| tensor<int32, [4]> v_77_perm_0 = const()[name = tensor<string, []>("v_77_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_18_y_0_to_fp16 = const()[name = tensor<string, []>("mul_18_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_18_cast_fp16 = mul(x = q_73_cast_fp16, y = mul_18_y_0_to_fp16)[name = tensor<string, []>("mul_18_cast_fp16")]; |
| tensor<bool, []> matmul_18_transpose_y_0 = const()[name = tensor<string, []>("matmul_18_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_18_transpose_x_0 = const()[name = tensor<string, []>("matmul_18_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_132_perm_0 = const()[name = tensor<string, []>("transpose_132_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_133_perm_0 = const()[name = tensor<string, []>("transpose_133_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_133 = transpose(perm = transpose_133_perm_0, x = k_73_cast_fp16)[name = tensor<string, []>("transpose_166")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_132 = transpose(perm = transpose_132_perm_0, x = mul_18_cast_fp16)[name = tensor<string, []>("transpose_167")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_18_cast_fp16 = matmul(transpose_x = matmul_18_transpose_x_0, transpose_y = matmul_18_transpose_y_0, x = transpose_132, y = transpose_133)[name = tensor<string, []>("matmul_18_cast_fp16")]; |
| tensor<int32, []> softmax_18_axis_0 = const()[name = tensor<string, []>("softmax_18_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_18_cast_fp16 = softmax(axis = softmax_18_axis_0, x = matmul_18_cast_fp16)[name = tensor<string, []>("softmax_18_cast_fp16")]; |
| tensor<bool, []> attns_73_transpose_x_0 = const()[name = tensor<string, []>("attns_73_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_73_transpose_y_0 = const()[name = tensor<string, []>("attns_73_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_77_cast_fp16 = transpose(perm = v_77_perm_0, x = v_75_cast_fp16)[name = tensor<string, []>("transpose_165")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_73_cast_fp16 = matmul(transpose_x = attns_73_transpose_x_0, transpose_y = attns_73_transpose_y_0, x = softmax_18_cast_fp16, y = v_77_cast_fp16)[name = tensor<string, []>("attns_73_cast_fp16")]; |
| tensor<int32, [4]> attns_75_perm_0 = const()[name = tensor<string, []>("attns_75_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_75 = const()[name = tensor<string, []>("concat_75"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_75_cast_fp16 = transpose(perm = attns_75_perm_0, x = attns_73_cast_fp16)[name = tensor<string, []>("transpose_164")]; |
| tensor<fp16, [1, 249, 1024]> x_177_cast_fp16 = reshape(shape = concat_75, x = attns_75_cast_fp16)[name = tensor<string, []>("x_177_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_18_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243357888))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(244406528))), name = tensor<string, []>("model_encoder_layers_18_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_18_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_18_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(244407104)))]; |
| tensor<fp16, [1, 249, 1024]> linear_112_cast_fp16 = linear(bias = model_encoder_layers_18_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_18_self_attn_output_proj_weight_to_fp16_palettized, x = x_177_cast_fp16)[name = tensor<string, []>("linear_112_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_193_cast_fp16 = add(x = linear_112_cast_fp16, y = input_191_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")]; |
| tensor<int32, [1]> x_179_axes_0 = const()[name = tensor<string, []>("x_179_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_18_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_18_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(244409216)))]; |
| tensor<fp16, [1024]> model_encoder_layers_18_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_18_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(244411328)))]; |
| tensor<fp16, [1, 249, 1024]> x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = model_encoder_layers_18_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_18_ffn_layer_norm_weight_to_fp16, x = input_193_cast_fp16)[name = tensor<string, []>("x_179_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_18_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(244413440))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248607808))), name = tensor<string, []>("model_encoder_layers_18_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_18_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_18_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248608384)))]; |
| tensor<fp16, [1, 249, 4096]> linear_113_cast_fp16 = linear(bias = model_encoder_layers_18_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_18_ffn_inner_proj_weight_to_fp16_palettized, x = x_179_cast_fp16)[name = tensor<string, []>("linear_113_cast_fp16")]; |
| tensor<string, []> input_197_mode_0 = const()[name = tensor<string, []>("input_197_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_197_cast_fp16 = gelu(mode = input_197_mode_0, x = linear_113_cast_fp16)[name = tensor<string, []>("input_197_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_18_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248616640))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252811008))), name = tensor<string, []>("model_encoder_layers_18_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_18_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_18_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252811584)))]; |
| tensor<fp16, [1, 249, 1024]> linear_114_cast_fp16 = linear(bias = model_encoder_layers_18_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_18_ffn_output_proj_weight_to_fp16_palettized, x = input_197_cast_fp16)[name = tensor<string, []>("linear_114_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_199_cast_fp16 = add(x = linear_114_cast_fp16, y = input_193_cast_fp16)[name = tensor<string, []>("input_199_cast_fp16")]; |
| tensor<int32, [1]> x_183_axes_0 = const()[name = tensor<string, []>("x_183_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_19_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_19_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252813696)))]; |
| tensor<fp16, [1024]> model_encoder_layers_19_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_19_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252815808)))]; |
| tensor<fp16, [1, 249, 1024]> x_183_cast_fp16 = layer_norm(axes = x_183_axes_0, beta = model_encoder_layers_19_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_19_self_attn_layer_norm_weight_to_fp16, x = input_199_cast_fp16)[name = tensor<string, []>("x_183_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_19_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252817920))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(253866560))), name = tensor<string, []>("model_encoder_layers_19_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(253867136)))]; |
| tensor<fp16, [1, 249, 1024]> linear_115_cast_fp16 = linear(bias = model_encoder_layers_19_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_19_self_attn_q_proj_weight_to_fp16_palettized, x = x_183_cast_fp16)[name = tensor<string, []>("linear_115_cast_fp16")]; |
| tensor<int32, [4]> concat_76 = const()[name = tensor<string, []>("concat_76"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_77_cast_fp16 = reshape(shape = concat_76, x = linear_115_cast_fp16)[name = tensor<string, []>("q_77_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_19_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(253869248))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254917888))), name = tensor<string, []>("model_encoder_layers_19_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_19_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_19_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254918464)))]; |
| tensor<fp16, [1, 249, 1024]> linear_116_cast_fp16 = linear(bias = model_encoder_layers_19_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_19_self_attn_k_proj_weight_to_fp16_palettized, x = x_183_cast_fp16)[name = tensor<string, []>("linear_116_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_19_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254920576))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255969216))), name = tensor<string, []>("model_encoder_layers_19_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255969792)))]; |
| tensor<fp16, [1, 249, 1024]> linear_117_cast_fp16 = linear(bias = model_encoder_layers_19_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_19_self_attn_v_proj_weight_to_fp16_palettized, x = x_183_cast_fp16)[name = tensor<string, []>("linear_117_cast_fp16")]; |
| tensor<int32, [4]> concat_77 = const()[name = tensor<string, []>("concat_77"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_77_cast_fp16 = reshape(shape = concat_77, x = linear_116_cast_fp16)[name = tensor<string, []>("k_77_cast_fp16")]; |
| tensor<int32, [4]> concat_78 = const()[name = tensor<string, []>("concat_78"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_79_cast_fp16 = reshape(shape = concat_78, x = linear_117_cast_fp16)[name = tensor<string, []>("v_79_cast_fp16")]; |
| tensor<int32, [4]> v_81_perm_0 = const()[name = tensor<string, []>("v_81_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_19_y_0_to_fp16 = const()[name = tensor<string, []>("mul_19_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_19_cast_fp16 = mul(x = q_77_cast_fp16, y = mul_19_y_0_to_fp16)[name = tensor<string, []>("mul_19_cast_fp16")]; |
| tensor<bool, []> matmul_19_transpose_y_0 = const()[name = tensor<string, []>("matmul_19_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_19_transpose_x_0 = const()[name = tensor<string, []>("matmul_19_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_134_perm_0 = const()[name = tensor<string, []>("transpose_134_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_135_perm_0 = const()[name = tensor<string, []>("transpose_135_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_135 = transpose(perm = transpose_135_perm_0, x = k_77_cast_fp16)[name = tensor<string, []>("transpose_162")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_134 = transpose(perm = transpose_134_perm_0, x = mul_19_cast_fp16)[name = tensor<string, []>("transpose_163")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_19_cast_fp16 = matmul(transpose_x = matmul_19_transpose_x_0, transpose_y = matmul_19_transpose_y_0, x = transpose_134, y = transpose_135)[name = tensor<string, []>("matmul_19_cast_fp16")]; |
| tensor<int32, []> softmax_19_axis_0 = const()[name = tensor<string, []>("softmax_19_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_19_cast_fp16 = softmax(axis = softmax_19_axis_0, x = matmul_19_cast_fp16)[name = tensor<string, []>("softmax_19_cast_fp16")]; |
| tensor<bool, []> attns_77_transpose_x_0 = const()[name = tensor<string, []>("attns_77_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_77_transpose_y_0 = const()[name = tensor<string, []>("attns_77_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_81_cast_fp16 = transpose(perm = v_81_perm_0, x = v_79_cast_fp16)[name = tensor<string, []>("transpose_161")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_77_cast_fp16 = matmul(transpose_x = attns_77_transpose_x_0, transpose_y = attns_77_transpose_y_0, x = softmax_19_cast_fp16, y = v_81_cast_fp16)[name = tensor<string, []>("attns_77_cast_fp16")]; |
| tensor<int32, [4]> attns_79_perm_0 = const()[name = tensor<string, []>("attns_79_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_79 = const()[name = tensor<string, []>("concat_79"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_79_cast_fp16 = transpose(perm = attns_79_perm_0, x = attns_77_cast_fp16)[name = tensor<string, []>("transpose_160")]; |
| tensor<fp16, [1, 249, 1024]> x_185_cast_fp16 = reshape(shape = concat_79, x = attns_79_cast_fp16)[name = tensor<string, []>("x_185_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_19_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255971904))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257020544))), name = tensor<string, []>("model_encoder_layers_19_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_19_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_19_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257021120)))]; |
| tensor<fp16, [1, 249, 1024]> linear_118_cast_fp16 = linear(bias = model_encoder_layers_19_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_19_self_attn_output_proj_weight_to_fp16_palettized, x = x_185_cast_fp16)[name = tensor<string, []>("linear_118_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_201_cast_fp16 = add(x = linear_118_cast_fp16, y = input_199_cast_fp16)[name = tensor<string, []>("input_201_cast_fp16")]; |
| tensor<int32, [1]> x_187_axes_0 = const()[name = tensor<string, []>("x_187_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_19_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_19_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257023232)))]; |
| tensor<fp16, [1024]> model_encoder_layers_19_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_19_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257025344)))]; |
| tensor<fp16, [1, 249, 1024]> x_187_cast_fp16 = layer_norm(axes = x_187_axes_0, beta = model_encoder_layers_19_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_19_ffn_layer_norm_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("x_187_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_19_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257027456))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261221824))), name = tensor<string, []>("model_encoder_layers_19_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_19_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_19_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261222400)))]; |
| tensor<fp16, [1, 249, 4096]> linear_119_cast_fp16 = linear(bias = model_encoder_layers_19_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_19_ffn_inner_proj_weight_to_fp16_palettized, x = x_187_cast_fp16)[name = tensor<string, []>("linear_119_cast_fp16")]; |
| tensor<string, []> input_205_mode_0 = const()[name = tensor<string, []>("input_205_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_205_cast_fp16 = gelu(mode = input_205_mode_0, x = linear_119_cast_fp16)[name = tensor<string, []>("input_205_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_19_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261230656))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(265425024))), name = tensor<string, []>("model_encoder_layers_19_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_19_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_19_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(265425600)))]; |
| tensor<fp16, [1, 249, 1024]> linear_120_cast_fp16 = linear(bias = model_encoder_layers_19_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_19_ffn_output_proj_weight_to_fp16_palettized, x = input_205_cast_fp16)[name = tensor<string, []>("linear_120_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_207_cast_fp16 = add(x = linear_120_cast_fp16, y = input_201_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")]; |
| tensor<int32, [1]> x_191_axes_0 = const()[name = tensor<string, []>("x_191_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_20_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_20_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(265427712)))]; |
| tensor<fp16, [1024]> model_encoder_layers_20_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_20_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(265429824)))]; |
| tensor<fp16, [1, 249, 1024]> x_191_cast_fp16 = layer_norm(axes = x_191_axes_0, beta = model_encoder_layers_20_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_20_self_attn_layer_norm_weight_to_fp16, x = input_207_cast_fp16)[name = tensor<string, []>("x_191_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_20_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(265431936))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(266480576))), name = tensor<string, []>("model_encoder_layers_20_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(266481152)))]; |
| tensor<fp16, [1, 249, 1024]> linear_121_cast_fp16 = linear(bias = model_encoder_layers_20_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_20_self_attn_q_proj_weight_to_fp16_palettized, x = x_191_cast_fp16)[name = tensor<string, []>("linear_121_cast_fp16")]; |
| tensor<int32, [4]> concat_80 = const()[name = tensor<string, []>("concat_80"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_81_cast_fp16 = reshape(shape = concat_80, x = linear_121_cast_fp16)[name = tensor<string, []>("q_81_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_20_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(266483264))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267531904))), name = tensor<string, []>("model_encoder_layers_20_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_20_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_20_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267532480)))]; |
| tensor<fp16, [1, 249, 1024]> linear_122_cast_fp16 = linear(bias = model_encoder_layers_20_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_20_self_attn_k_proj_weight_to_fp16_palettized, x = x_191_cast_fp16)[name = tensor<string, []>("linear_122_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_20_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267534592))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(268583232))), name = tensor<string, []>("model_encoder_layers_20_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(268583808)))]; |
| tensor<fp16, [1, 249, 1024]> linear_123_cast_fp16 = linear(bias = model_encoder_layers_20_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_20_self_attn_v_proj_weight_to_fp16_palettized, x = x_191_cast_fp16)[name = tensor<string, []>("linear_123_cast_fp16")]; |
| tensor<int32, [4]> concat_81 = const()[name = tensor<string, []>("concat_81"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_81_cast_fp16 = reshape(shape = concat_81, x = linear_122_cast_fp16)[name = tensor<string, []>("k_81_cast_fp16")]; |
| tensor<int32, [4]> concat_82 = const()[name = tensor<string, []>("concat_82"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_83_cast_fp16 = reshape(shape = concat_82, x = linear_123_cast_fp16)[name = tensor<string, []>("v_83_cast_fp16")]; |
| tensor<int32, [4]> v_85_perm_0 = const()[name = tensor<string, []>("v_85_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_20_y_0_to_fp16 = const()[name = tensor<string, []>("mul_20_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_20_cast_fp16 = mul(x = q_81_cast_fp16, y = mul_20_y_0_to_fp16)[name = tensor<string, []>("mul_20_cast_fp16")]; |
| tensor<bool, []> matmul_20_transpose_y_0 = const()[name = tensor<string, []>("matmul_20_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_20_transpose_x_0 = const()[name = tensor<string, []>("matmul_20_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_136_perm_0 = const()[name = tensor<string, []>("transpose_136_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_137_perm_0 = const()[name = tensor<string, []>("transpose_137_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_137 = transpose(perm = transpose_137_perm_0, x = k_81_cast_fp16)[name = tensor<string, []>("transpose_158")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_136 = transpose(perm = transpose_136_perm_0, x = mul_20_cast_fp16)[name = tensor<string, []>("transpose_159")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_20_cast_fp16 = matmul(transpose_x = matmul_20_transpose_x_0, transpose_y = matmul_20_transpose_y_0, x = transpose_136, y = transpose_137)[name = tensor<string, []>("matmul_20_cast_fp16")]; |
| tensor<int32, []> softmax_20_axis_0 = const()[name = tensor<string, []>("softmax_20_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_20_cast_fp16 = softmax(axis = softmax_20_axis_0, x = matmul_20_cast_fp16)[name = tensor<string, []>("softmax_20_cast_fp16")]; |
| tensor<bool, []> attns_81_transpose_x_0 = const()[name = tensor<string, []>("attns_81_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_81_transpose_y_0 = const()[name = tensor<string, []>("attns_81_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_85_cast_fp16 = transpose(perm = v_85_perm_0, x = v_83_cast_fp16)[name = tensor<string, []>("transpose_157")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_81_cast_fp16 = matmul(transpose_x = attns_81_transpose_x_0, transpose_y = attns_81_transpose_y_0, x = softmax_20_cast_fp16, y = v_85_cast_fp16)[name = tensor<string, []>("attns_81_cast_fp16")]; |
| tensor<int32, [4]> attns_83_perm_0 = const()[name = tensor<string, []>("attns_83_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_83 = const()[name = tensor<string, []>("concat_83"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_83_cast_fp16 = transpose(perm = attns_83_perm_0, x = attns_81_cast_fp16)[name = tensor<string, []>("transpose_156")]; |
| tensor<fp16, [1, 249, 1024]> x_193_cast_fp16 = reshape(shape = concat_83, x = attns_83_cast_fp16)[name = tensor<string, []>("x_193_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_20_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(268585920))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269634560))), name = tensor<string, []>("model_encoder_layers_20_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_20_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_20_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269635136)))]; |
| tensor<fp16, [1, 249, 1024]> linear_124_cast_fp16 = linear(bias = model_encoder_layers_20_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_20_self_attn_output_proj_weight_to_fp16_palettized, x = x_193_cast_fp16)[name = tensor<string, []>("linear_124_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_209_cast_fp16 = add(x = linear_124_cast_fp16, y = input_207_cast_fp16)[name = tensor<string, []>("input_209_cast_fp16")]; |
| tensor<int32, [1]> x_195_axes_0 = const()[name = tensor<string, []>("x_195_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_20_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_20_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269637248)))]; |
| tensor<fp16, [1024]> model_encoder_layers_20_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_20_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269639360)))]; |
| tensor<fp16, [1, 249, 1024]> x_195_cast_fp16 = layer_norm(axes = x_195_axes_0, beta = model_encoder_layers_20_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_20_ffn_layer_norm_weight_to_fp16, x = input_209_cast_fp16)[name = tensor<string, []>("x_195_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_20_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269641472))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273835840))), name = tensor<string, []>("model_encoder_layers_20_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_20_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_20_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273836416)))]; |
| tensor<fp16, [1, 249, 4096]> linear_125_cast_fp16 = linear(bias = model_encoder_layers_20_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_20_ffn_inner_proj_weight_to_fp16_palettized, x = x_195_cast_fp16)[name = tensor<string, []>("linear_125_cast_fp16")]; |
| tensor<string, []> input_213_mode_0 = const()[name = tensor<string, []>("input_213_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_213_cast_fp16 = gelu(mode = input_213_mode_0, x = linear_125_cast_fp16)[name = tensor<string, []>("input_213_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_20_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273844672))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278039040))), name = tensor<string, []>("model_encoder_layers_20_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_20_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_20_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278039616)))]; |
| tensor<fp16, [1, 249, 1024]> linear_126_cast_fp16 = linear(bias = model_encoder_layers_20_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_20_ffn_output_proj_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = tensor<string, []>("linear_126_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_215_cast_fp16 = add(x = linear_126_cast_fp16, y = input_209_cast_fp16)[name = tensor<string, []>("input_215_cast_fp16")]; |
| tensor<int32, [1]> x_199_axes_0 = const()[name = tensor<string, []>("x_199_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_21_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_21_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278041728)))]; |
| tensor<fp16, [1024]> model_encoder_layers_21_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_21_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278043840)))]; |
| tensor<fp16, [1, 249, 1024]> x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = model_encoder_layers_21_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_21_self_attn_layer_norm_weight_to_fp16, x = input_215_cast_fp16)[name = tensor<string, []>("x_199_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_21_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278045952))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279094592))), name = tensor<string, []>("model_encoder_layers_21_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279095168)))]; |
| tensor<fp16, [1, 249, 1024]> linear_127_cast_fp16 = linear(bias = model_encoder_layers_21_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_21_self_attn_q_proj_weight_to_fp16_palettized, x = x_199_cast_fp16)[name = tensor<string, []>("linear_127_cast_fp16")]; |
| tensor<int32, [4]> concat_84 = const()[name = tensor<string, []>("concat_84"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_85_cast_fp16 = reshape(shape = concat_84, x = linear_127_cast_fp16)[name = tensor<string, []>("q_85_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_21_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279097280))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(280145920))), name = tensor<string, []>("model_encoder_layers_21_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_21_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_21_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(280146496)))]; |
| tensor<fp16, [1, 249, 1024]> linear_128_cast_fp16 = linear(bias = model_encoder_layers_21_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_21_self_attn_k_proj_weight_to_fp16_palettized, x = x_199_cast_fp16)[name = tensor<string, []>("linear_128_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_21_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(280148608))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281197248))), name = tensor<string, []>("model_encoder_layers_21_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281197824)))]; |
| tensor<fp16, [1, 249, 1024]> linear_129_cast_fp16 = linear(bias = model_encoder_layers_21_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_21_self_attn_v_proj_weight_to_fp16_palettized, x = x_199_cast_fp16)[name = tensor<string, []>("linear_129_cast_fp16")]; |
| tensor<int32, [4]> concat_85 = const()[name = tensor<string, []>("concat_85"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_85_cast_fp16 = reshape(shape = concat_85, x = linear_128_cast_fp16)[name = tensor<string, []>("k_85_cast_fp16")]; |
| tensor<int32, [4]> concat_86 = const()[name = tensor<string, []>("concat_86"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_87_cast_fp16 = reshape(shape = concat_86, x = linear_129_cast_fp16)[name = tensor<string, []>("v_87_cast_fp16")]; |
| tensor<int32, [4]> v_89_perm_0 = const()[name = tensor<string, []>("v_89_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_21_y_0_to_fp16 = const()[name = tensor<string, []>("mul_21_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_21_cast_fp16 = mul(x = q_85_cast_fp16, y = mul_21_y_0_to_fp16)[name = tensor<string, []>("mul_21_cast_fp16")]; |
| tensor<bool, []> matmul_21_transpose_y_0 = const()[name = tensor<string, []>("matmul_21_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_21_transpose_x_0 = const()[name = tensor<string, []>("matmul_21_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_138_perm_0 = const()[name = tensor<string, []>("transpose_138_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_139_perm_0 = const()[name = tensor<string, []>("transpose_139_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_139 = transpose(perm = transpose_139_perm_0, x = k_85_cast_fp16)[name = tensor<string, []>("transpose_154")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_138 = transpose(perm = transpose_138_perm_0, x = mul_21_cast_fp16)[name = tensor<string, []>("transpose_155")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_21_cast_fp16 = matmul(transpose_x = matmul_21_transpose_x_0, transpose_y = matmul_21_transpose_y_0, x = transpose_138, y = transpose_139)[name = tensor<string, []>("matmul_21_cast_fp16")]; |
| tensor<int32, []> softmax_21_axis_0 = const()[name = tensor<string, []>("softmax_21_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_21_cast_fp16 = softmax(axis = softmax_21_axis_0, x = matmul_21_cast_fp16)[name = tensor<string, []>("softmax_21_cast_fp16")]; |
| tensor<bool, []> attns_85_transpose_x_0 = const()[name = tensor<string, []>("attns_85_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_85_transpose_y_0 = const()[name = tensor<string, []>("attns_85_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_89_cast_fp16 = transpose(perm = v_89_perm_0, x = v_87_cast_fp16)[name = tensor<string, []>("transpose_153")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_85_cast_fp16 = matmul(transpose_x = attns_85_transpose_x_0, transpose_y = attns_85_transpose_y_0, x = softmax_21_cast_fp16, y = v_89_cast_fp16)[name = tensor<string, []>("attns_85_cast_fp16")]; |
| tensor<int32, [4]> attns_87_perm_0 = const()[name = tensor<string, []>("attns_87_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_87 = const()[name = tensor<string, []>("concat_87"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_87_cast_fp16 = transpose(perm = attns_87_perm_0, x = attns_85_cast_fp16)[name = tensor<string, []>("transpose_152")]; |
| tensor<fp16, [1, 249, 1024]> x_201_cast_fp16 = reshape(shape = concat_87, x = attns_87_cast_fp16)[name = tensor<string, []>("x_201_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_21_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281199936))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(282248576))), name = tensor<string, []>("model_encoder_layers_21_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_21_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_21_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(282249152)))]; |
| tensor<fp16, [1, 249, 1024]> linear_130_cast_fp16 = linear(bias = model_encoder_layers_21_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_21_self_attn_output_proj_weight_to_fp16_palettized, x = x_201_cast_fp16)[name = tensor<string, []>("linear_130_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_217_cast_fp16 = add(x = linear_130_cast_fp16, y = input_215_cast_fp16)[name = tensor<string, []>("input_217_cast_fp16")]; |
| tensor<int32, [1]> x_203_axes_0 = const()[name = tensor<string, []>("x_203_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_21_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_21_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(282251264)))]; |
| tensor<fp16, [1024]> model_encoder_layers_21_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_21_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(282253376)))]; |
| tensor<fp16, [1, 249, 1024]> x_203_cast_fp16 = layer_norm(axes = x_203_axes_0, beta = model_encoder_layers_21_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_21_ffn_layer_norm_weight_to_fp16, x = input_217_cast_fp16)[name = tensor<string, []>("x_203_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_21_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(282255488))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286449856))), name = tensor<string, []>("model_encoder_layers_21_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_21_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_21_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286450432)))]; |
| tensor<fp16, [1, 249, 4096]> linear_131_cast_fp16 = linear(bias = model_encoder_layers_21_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_21_ffn_inner_proj_weight_to_fp16_palettized, x = x_203_cast_fp16)[name = tensor<string, []>("linear_131_cast_fp16")]; |
| tensor<string, []> input_221_mode_0 = const()[name = tensor<string, []>("input_221_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_221_cast_fp16 = gelu(mode = input_221_mode_0, x = linear_131_cast_fp16)[name = tensor<string, []>("input_221_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_21_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286458688))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(290653056))), name = tensor<string, []>("model_encoder_layers_21_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_21_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_21_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(290653632)))]; |
| tensor<fp16, [1, 249, 1024]> linear_132_cast_fp16 = linear(bias = model_encoder_layers_21_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_21_ffn_output_proj_weight_to_fp16_palettized, x = input_221_cast_fp16)[name = tensor<string, []>("linear_132_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_223_cast_fp16 = add(x = linear_132_cast_fp16, y = input_217_cast_fp16)[name = tensor<string, []>("input_223_cast_fp16")]; |
| tensor<int32, [1]> x_207_axes_0 = const()[name = tensor<string, []>("x_207_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_22_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_22_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(290655744)))]; |
| tensor<fp16, [1024]> model_encoder_layers_22_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_22_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(290657856)))]; |
| tensor<fp16, [1, 249, 1024]> x_207_cast_fp16 = layer_norm(axes = x_207_axes_0, beta = model_encoder_layers_22_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_22_self_attn_layer_norm_weight_to_fp16, x = input_223_cast_fp16)[name = tensor<string, []>("x_207_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_22_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(290659968))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291708608))), name = tensor<string, []>("model_encoder_layers_22_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291709184)))]; |
| tensor<fp16, [1, 249, 1024]> linear_133_cast_fp16 = linear(bias = model_encoder_layers_22_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_22_self_attn_q_proj_weight_to_fp16_palettized, x = x_207_cast_fp16)[name = tensor<string, []>("linear_133_cast_fp16")]; |
| tensor<int32, [4]> concat_88 = const()[name = tensor<string, []>("concat_88"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_89_cast_fp16 = reshape(shape = concat_88, x = linear_133_cast_fp16)[name = tensor<string, []>("q_89_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_22_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291711296))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292759936))), name = tensor<string, []>("model_encoder_layers_22_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_22_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_22_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292760512)))]; |
| tensor<fp16, [1, 249, 1024]> linear_134_cast_fp16 = linear(bias = model_encoder_layers_22_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_22_self_attn_k_proj_weight_to_fp16_palettized, x = x_207_cast_fp16)[name = tensor<string, []>("linear_134_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_22_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292762624))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293811264))), name = tensor<string, []>("model_encoder_layers_22_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293811840)))]; |
| tensor<fp16, [1, 249, 1024]> linear_135_cast_fp16 = linear(bias = model_encoder_layers_22_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_22_self_attn_v_proj_weight_to_fp16_palettized, x = x_207_cast_fp16)[name = tensor<string, []>("linear_135_cast_fp16")]; |
| tensor<int32, [4]> concat_89 = const()[name = tensor<string, []>("concat_89"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_89_cast_fp16 = reshape(shape = concat_89, x = linear_134_cast_fp16)[name = tensor<string, []>("k_89_cast_fp16")]; |
| tensor<int32, [4]> concat_90 = const()[name = tensor<string, []>("concat_90"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_91_cast_fp16 = reshape(shape = concat_90, x = linear_135_cast_fp16)[name = tensor<string, []>("v_91_cast_fp16")]; |
| tensor<int32, [4]> v_93_perm_0 = const()[name = tensor<string, []>("v_93_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_22_y_0_to_fp16 = const()[name = tensor<string, []>("mul_22_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_22_cast_fp16 = mul(x = q_89_cast_fp16, y = mul_22_y_0_to_fp16)[name = tensor<string, []>("mul_22_cast_fp16")]; |
| tensor<bool, []> matmul_22_transpose_y_0 = const()[name = tensor<string, []>("matmul_22_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_22_transpose_x_0 = const()[name = tensor<string, []>("matmul_22_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_140_perm_0 = const()[name = tensor<string, []>("transpose_140_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_141_perm_0 = const()[name = tensor<string, []>("transpose_141_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_141 = transpose(perm = transpose_141_perm_0, x = k_89_cast_fp16)[name = tensor<string, []>("transpose_150")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_140 = transpose(perm = transpose_140_perm_0, x = mul_22_cast_fp16)[name = tensor<string, []>("transpose_151")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_22_cast_fp16 = matmul(transpose_x = matmul_22_transpose_x_0, transpose_y = matmul_22_transpose_y_0, x = transpose_140, y = transpose_141)[name = tensor<string, []>("matmul_22_cast_fp16")]; |
| tensor<int32, []> softmax_22_axis_0 = const()[name = tensor<string, []>("softmax_22_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_22_cast_fp16 = softmax(axis = softmax_22_axis_0, x = matmul_22_cast_fp16)[name = tensor<string, []>("softmax_22_cast_fp16")]; |
| tensor<bool, []> attns_89_transpose_x_0 = const()[name = tensor<string, []>("attns_89_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_89_transpose_y_0 = const()[name = tensor<string, []>("attns_89_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_93_cast_fp16 = transpose(perm = v_93_perm_0, x = v_91_cast_fp16)[name = tensor<string, []>("transpose_149")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_89_cast_fp16 = matmul(transpose_x = attns_89_transpose_x_0, transpose_y = attns_89_transpose_y_0, x = softmax_22_cast_fp16, y = v_93_cast_fp16)[name = tensor<string, []>("attns_89_cast_fp16")]; |
| tensor<int32, [4]> attns_91_perm_0 = const()[name = tensor<string, []>("attns_91_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_91 = const()[name = tensor<string, []>("concat_91"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_91_cast_fp16 = transpose(perm = attns_91_perm_0, x = attns_89_cast_fp16)[name = tensor<string, []>("transpose_148")]; |
| tensor<fp16, [1, 249, 1024]> x_209_cast_fp16 = reshape(shape = concat_91, x = attns_91_cast_fp16)[name = tensor<string, []>("x_209_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_22_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293813952))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294862592))), name = tensor<string, []>("model_encoder_layers_22_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_22_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_22_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294863168)))]; |
| tensor<fp16, [1, 249, 1024]> linear_136_cast_fp16 = linear(bias = model_encoder_layers_22_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_22_self_attn_output_proj_weight_to_fp16_palettized, x = x_209_cast_fp16)[name = tensor<string, []>("linear_136_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_225_cast_fp16 = add(x = linear_136_cast_fp16, y = input_223_cast_fp16)[name = tensor<string, []>("input_225_cast_fp16")]; |
| tensor<int32, [1]> x_211_axes_0 = const()[name = tensor<string, []>("x_211_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_22_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_22_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294865280)))]; |
| tensor<fp16, [1024]> model_encoder_layers_22_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_22_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294867392)))]; |
| tensor<fp16, [1, 249, 1024]> x_211_cast_fp16 = layer_norm(axes = x_211_axes_0, beta = model_encoder_layers_22_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_22_ffn_layer_norm_weight_to_fp16, x = input_225_cast_fp16)[name = tensor<string, []>("x_211_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_22_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294869504))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(299063872))), name = tensor<string, []>("model_encoder_layers_22_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_22_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_22_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(299064448)))]; |
| tensor<fp16, [1, 249, 4096]> linear_137_cast_fp16 = linear(bias = model_encoder_layers_22_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_22_ffn_inner_proj_weight_to_fp16_palettized, x = x_211_cast_fp16)[name = tensor<string, []>("linear_137_cast_fp16")]; |
| tensor<string, []> input_229_mode_0 = const()[name = tensor<string, []>("input_229_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = linear_137_cast_fp16)[name = tensor<string, []>("input_229_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_22_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(299072704))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303267072))), name = tensor<string, []>("model_encoder_layers_22_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_22_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_22_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303267648)))]; |
| tensor<fp16, [1, 249, 1024]> linear_138_cast_fp16 = linear(bias = model_encoder_layers_22_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_22_ffn_output_proj_weight_to_fp16_palettized, x = input_229_cast_fp16)[name = tensor<string, []>("linear_138_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_231_cast_fp16 = add(x = linear_138_cast_fp16, y = input_225_cast_fp16)[name = tensor<string, []>("input_231_cast_fp16")]; |
| tensor<int32, [1]> x_215_axes_0 = const()[name = tensor<string, []>("x_215_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_23_self_attn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_23_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303269760)))]; |
| tensor<fp16, [1024]> model_encoder_layers_23_self_attn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_23_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303271872)))]; |
| tensor<fp16, [1, 249, 1024]> x_215_cast_fp16 = layer_norm(axes = x_215_axes_0, beta = model_encoder_layers_23_self_attn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_23_self_attn_layer_norm_weight_to_fp16, x = input_231_cast_fp16)[name = tensor<string, []>("x_215_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_23_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303273984))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304322624))), name = tensor<string, []>("model_encoder_layers_23_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304323200)))]; |
| tensor<fp16, [1, 249, 1024]> linear_139_cast_fp16 = linear(bias = model_encoder_layers_23_self_attn_q_proj_bias_to_fp16, weight = model_encoder_layers_23_self_attn_q_proj_weight_to_fp16_palettized, x = x_215_cast_fp16)[name = tensor<string, []>("linear_139_cast_fp16")]; |
| tensor<int32, [4]> concat_92 = const()[name = tensor<string, []>("concat_92"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> q_93_cast_fp16 = reshape(shape = concat_92, x = linear_139_cast_fp16)[name = tensor<string, []>("q_93_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_23_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304325312))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305373952))), name = tensor<string, []>("model_encoder_layers_23_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_23_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_23_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305374528)))]; |
| tensor<fp16, [1, 249, 1024]> linear_140_cast_fp16 = linear(bias = model_encoder_layers_23_self_attn_k_proj_bias_to_fp16, weight = model_encoder_layers_23_self_attn_k_proj_weight_to_fp16_palettized, x = x_215_cast_fp16)[name = tensor<string, []>("linear_140_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_23_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305376640))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306425280))), name = tensor<string, []>("model_encoder_layers_23_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306425856)))]; |
| tensor<fp16, [1, 249, 1024]> linear_141_cast_fp16 = linear(bias = model_encoder_layers_23_self_attn_v_proj_bias_to_fp16, weight = model_encoder_layers_23_self_attn_v_proj_weight_to_fp16_palettized, x = x_215_cast_fp16)[name = tensor<string, []>("linear_141_cast_fp16")]; |
| tensor<int32, [4]> concat_93 = const()[name = tensor<string, []>("concat_93"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> k_93_cast_fp16 = reshape(shape = concat_93, x = linear_140_cast_fp16)[name = tensor<string, []>("k_93_cast_fp16")]; |
| tensor<int32, [4]> concat_94 = const()[name = tensor<string, []>("concat_94"), val = tensor<int32, [4]>([1, 249, -1, 64])]; |
| tensor<fp16, [1, 249, 16, 64]> v_95_cast_fp16 = reshape(shape = concat_94, x = linear_141_cast_fp16)[name = tensor<string, []>("v_95_cast_fp16")]; |
| tensor<int32, [4]> v_perm_0 = const()[name = tensor<string, []>("v_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, []> mul_23_y_0_to_fp16 = const()[name = tensor<string, []>("mul_23_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| tensor<fp16, [1, 249, 16, 64]> mul_23_cast_fp16 = mul(x = q_93_cast_fp16, y = mul_23_y_0_to_fp16)[name = tensor<string, []>("mul_23_cast_fp16")]; |
| tensor<bool, []> matmul_23_transpose_y_0 = const()[name = tensor<string, []>("matmul_23_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_23_transpose_x_0 = const()[name = tensor<string, []>("matmul_23_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [4]> transpose_142_perm_0 = const()[name = tensor<string, []>("transpose_142_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_143_perm_0 = const()[name = tensor<string, []>("transpose_143_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_143 = transpose(perm = transpose_143_perm_0, x = k_93_cast_fp16)[name = tensor<string, []>("transpose_146")]; |
| tensor<fp16, [1, 16, 249, 64]> transpose_142 = transpose(perm = transpose_142_perm_0, x = mul_23_cast_fp16)[name = tensor<string, []>("transpose_147")]; |
| tensor<fp16, [1, 16, 249, 249]> matmul_23_cast_fp16 = matmul(transpose_x = matmul_23_transpose_x_0, transpose_y = matmul_23_transpose_y_0, x = transpose_142, y = transpose_143)[name = tensor<string, []>("matmul_23_cast_fp16")]; |
| tensor<int32, []> softmax_23_axis_0 = const()[name = tensor<string, []>("softmax_23_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp16, [1, 16, 249, 249]> softmax_23_cast_fp16 = softmax(axis = softmax_23_axis_0, x = matmul_23_cast_fp16)[name = tensor<string, []>("softmax_23_cast_fp16")]; |
| tensor<bool, []> attns_93_transpose_x_0 = const()[name = tensor<string, []>("attns_93_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> attns_93_transpose_y_0 = const()[name = tensor<string, []>("attns_93_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp16, [1, 16, 249, 64]> v_cast_fp16 = transpose(perm = v_perm_0, x = v_95_cast_fp16)[name = tensor<string, []>("transpose_145")]; |
| tensor<fp16, [1, 16, 249, 64]> attns_93_cast_fp16 = matmul(transpose_x = attns_93_transpose_x_0, transpose_y = attns_93_transpose_y_0, x = softmax_23_cast_fp16, y = v_cast_fp16)[name = tensor<string, []>("attns_93_cast_fp16")]; |
| tensor<int32, [4]> attns_perm_0 = const()[name = tensor<string, []>("attns_perm_0"), val = tensor<int32, [4]>([0, -2, -3, 3])]; |
| tensor<int32, [3]> concat_95 = const()[name = tensor<string, []>("concat_95"), val = tensor<int32, [3]>([1, 249, 1024])]; |
| tensor<fp16, [1, 249, 16, 64]> attns_cast_fp16 = transpose(perm = attns_perm_0, x = attns_93_cast_fp16)[name = tensor<string, []>("transpose_144")]; |
| tensor<fp16, [1, 249, 1024]> x_217_cast_fp16 = reshape(shape = concat_95, x = attns_cast_fp16)[name = tensor<string, []>("x_217_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> model_encoder_layers_23_self_attn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [1048576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306427968))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307476608))), name = tensor<string, []>("model_encoder_layers_23_self_attn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])]; |
| tensor<fp16, [1024]> model_encoder_layers_23_self_attn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_23_self_attn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307477184)))]; |
| tensor<fp16, [1, 249, 1024]> linear_142_cast_fp16 = linear(bias = model_encoder_layers_23_self_attn_output_proj_bias_to_fp16, weight = model_encoder_layers_23_self_attn_output_proj_weight_to_fp16_palettized, x = x_217_cast_fp16)[name = tensor<string, []>("linear_142_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_233_cast_fp16 = add(x = linear_142_cast_fp16, y = input_231_cast_fp16)[name = tensor<string, []>("input_233_cast_fp16")]; |
| tensor<int32, [1]> x_219_axes_0 = const()[name = tensor<string, []>("x_219_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layers_23_ffn_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_23_ffn_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307479296)))]; |
| tensor<fp16, [1024]> model_encoder_layers_23_ffn_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_23_ffn_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307481408)))]; |
| tensor<fp16, [1, 249, 1024]> x_219_cast_fp16 = layer_norm(axes = x_219_axes_0, beta = model_encoder_layers_23_ffn_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layers_23_ffn_layer_norm_weight_to_fp16, x = input_233_cast_fp16)[name = tensor<string, []>("x_219_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> model_encoder_layers_23_ffn_inner_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307483520))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311677888))), name = tensor<string, []>("model_encoder_layers_23_ffn_inner_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])]; |
| tensor<fp16, [4096]> model_encoder_layers_23_ffn_inner_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_23_ffn_inner_proj_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311678464)))]; |
| tensor<fp16, [1, 249, 4096]> linear_143_cast_fp16 = linear(bias = model_encoder_layers_23_ffn_inner_proj_bias_to_fp16, weight = model_encoder_layers_23_ffn_inner_proj_weight_to_fp16_palettized, x = x_219_cast_fp16)[name = tensor<string, []>("linear_143_cast_fp16")]; |
| tensor<string, []> input_237_mode_0 = const()[name = tensor<string, []>("input_237_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp16, [1, 249, 4096]> input_237_cast_fp16 = gelu(mode = input_237_mode_0, x = linear_143_cast_fp16)[name = tensor<string, []>("input_237_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> model_encoder_layers_23_ffn_output_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4194304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311686720))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(315881088))), name = tensor<string, []>("model_encoder_layers_23_ffn_output_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])]; |
| tensor<fp16, [1024]> model_encoder_layers_23_ffn_output_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_23_ffn_output_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(315881664)))]; |
| tensor<fp16, [1, 249, 1024]> linear_144_cast_fp16 = linear(bias = model_encoder_layers_23_ffn_output_proj_bias_to_fp16, weight = model_encoder_layers_23_ffn_output_proj_weight_to_fp16_palettized, x = input_237_cast_fp16)[name = tensor<string, []>("linear_144_cast_fp16")]; |
| tensor<fp16, [1, 249, 1024]> input_cast_fp16 = add(x = linear_144_cast_fp16, y = input_233_cast_fp16)[name = tensor<string, []>("input_cast_fp16")]; |
| tensor<int32, [1]> x_axes_0 = const()[name = tensor<string, []>("x_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> model_encoder_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layer_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(315883776)))]; |
| tensor<fp16, [1024]> model_encoder_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layer_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(315885888)))]; |
| tensor<fp16, [1, 249, 1024]> x_cast_fp16 = layer_norm(axes = x_axes_0, beta = model_encoder_layer_norm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = model_encoder_layer_norm_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("x_cast_fp16")]; |
| tensor<fp16, [10288, 1024]> model_final_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [10534912]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(315888000))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(326422976))), name = tensor<string, []>("model_final_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([10288, 1024])]; |
| tensor<fp16, [10288]> model_final_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_final_proj_bias_to_fp16"), val = tensor<fp16, [10288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(326423552)))]; |
| tensor<fp16, [1, 249, 10288]> logits = linear(bias = model_final_proj_bias_to_fp16, weight = model_final_proj_weight_to_fp16_palettized, x = x_cast_fp16)[name = tensor<string, []>("linear_145_cast_fp16")]; |
| } -> (logits); |
| } |
