openai_whisper-large-v3-v20240930_626MB/TextDecoder.mlmodelc/model.mil
| 1 | program(1.0) |
| 2 | [buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3304.5.2"}, {"coremlc-version", "3304.6.2"}})] |
| 3 | { |
| 4 | func main<ios16>(tensor<int32, [1]> cache_length, tensor<fp16, [1, 448]> decoder_key_padding_mask, tensor<fp16, [1, 1280, 1, 1500]> encoder_output_embeds, tensor<int32, [1]> input_ids, tensor<fp16, [1, 5120, 1, 448]> key_cache, tensor<fp16, [1, 448]> kv_cache_update_mask, tensor<fp16, [1, 5120, 1, 448]> value_cache) { |
| 5 | tensor<int32, []> var_24_axis_0 = const()[name = tensor<string, []>("op_24_axis_0"), val = tensor<int32, []>(0)]; |
| 6 | tensor<int32, []> var_24_batch_dims_0 = const()[name = tensor<string, []>("op_24_batch_dims_0"), val = tensor<int32, []>(0)]; |
| 7 | tensor<fp16, [51866, 1280]> embed_tokens_weight_to_fp16 = const()[name = tensor<string, []>("embed_tokens_weight_to_fp16"), val = tensor<fp16, [51866, 1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))]; |
| 8 | tensor<fp16, [1, 1280]> var_24_cast_fp16 = gather(axis = var_24_axis_0, batch_dims = var_24_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor<string, []>("op_24_cast_fp16")]; |
| 9 | tensor<int32, []> var_31_axis_0 = const()[name = tensor<string, []>("op_31_axis_0"), val = tensor<int32, []>(0)]; |
| 10 | tensor<int32, []> var_31_batch_dims_0 = const()[name = tensor<string, []>("op_31_batch_dims_0"), val = tensor<int32, []>(0)]; |
| 11 | tensor<fp16, [448, 1280]> embed_positions_inlier_module_weight_to_fp16 = const()[name = tensor<string, []>("embed_positions_inlier_module_weight_to_fp16"), val = tensor<fp16, [448, 1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132777088)))]; |
| 12 | tensor<fp16, [1, 1280]> var_31_cast_fp16 = gather(axis = var_31_axis_0, batch_dims = var_31_batch_dims_0, indices = cache_length, x = embed_positions_inlier_module_weight_to_fp16)[name = tensor<string, []>("op_31_cast_fp16")]; |
| 13 | tensor<int32, []> var_33_axis_0 = const()[name = tensor<string, []>("op_33_axis_0"), val = tensor<int32, []>(0)]; |
| 14 | tensor<int32, []> var_33_batch_dims_0 = const()[name = tensor<string, []>("op_33_batch_dims_0"), val = tensor<int32, []>(0)]; |
| 15 | tensor<fp16, [448, 1280]> embed_positions_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [71680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133941312))), name = tensor<string, []>("embed_positions_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [8582]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133924032))), shape = tensor<uint32, [2]>([448, 1280])]; |
| 16 | tensor<fp16, [1, 1280]> var_33_cast_fp16 = gather(axis = var_33_axis_0, batch_dims = var_33_batch_dims_0, indices = cache_length, x = embed_positions_outlier_module_weight_to_fp16_sparsified)[name = tensor<string, []>("op_33_cast_fp16")]; |
| 17 | tensor<fp16, [1, 1280]> var_34_cast_fp16 = add(x = var_31_cast_fp16, y = var_33_cast_fp16)[name = tensor<string, []>("op_34_cast_fp16")]; |
| 18 | tensor<fp16, [1, 1280]> hidden_states_1_cast_fp16 = add(x = var_24_cast_fp16, y = var_34_cast_fp16)[name = tensor<string, []>("hidden_states_1_cast_fp16")]; |
| 19 | tensor<int32, [1]> var_48_axes_0 = const()[name = tensor<string, []>("op_48_axes_0"), val = tensor<int32, [1]>([2])]; |
| 20 | tensor<fp16, [1, 1280, 1]> var_48_cast_fp16 = expand_dims(axes = var_48_axes_0, x = hidden_states_1_cast_fp16)[name = tensor<string, []>("op_48_cast_fp16")]; |
| 21 | tensor<int32, [1]> inputs_1_axes_0 = const()[name = tensor<string, []>("inputs_1_axes_0"), val = tensor<int32, [1]>([3])]; |
| 22 | tensor<fp16, [1, 1280, 1, 1]> inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_48_cast_fp16)[name = tensor<string, []>("inputs_1_cast_fp16")]; |
| 23 | tensor<int32, [4]> tile_0 = const()[name = tensor<string, []>("tile_0"), val = tensor<int32, [4]>([1280, 1280, 1280, 1280])]; |
| 24 | tensor<int32, []> var_53_axis_0 = const()[name = tensor<string, []>("op_53_axis_0"), val = tensor<int32, []>(1)]; |
| 25 | tensor<fp16, [1, 1280, 1, 448]> var_53_cast_fp16_0, tensor<fp16, [1, 1280, 1, 448]> var_53_cast_fp16_1, tensor<fp16, [1, 1280, 1, 448]> var_53_cast_fp16_2, tensor<fp16, [1, 1280, 1, 448]> var_53_cast_fp16_3 = split(axis = var_53_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor<string, []>("op_53_cast_fp16")]; |
| 26 | tensor<int32, [4]> tile_1 = const()[name = tensor<string, []>("tile_1"), val = tensor<int32, [4]>([1280, 1280, 1280, 1280])]; |
| 27 | tensor<int32, []> var_60_axis_0 = const()[name = tensor<string, []>("op_60_axis_0"), val = tensor<int32, []>(1)]; |
| 28 | tensor<fp16, [1, 1280, 1, 448]> var_60_cast_fp16_0, tensor<fp16, [1, 1280, 1, 448]> var_60_cast_fp16_1, tensor<fp16, [1, 1280, 1, 448]> var_60_cast_fp16_2, tensor<fp16, [1, 1280, 1, 448]> var_60_cast_fp16_3 = split(axis = var_60_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor<string, []>("op_60_cast_fp16")]; |
| 29 | tensor<int32, []> var_70 = const()[name = tensor<string, []>("op_70"), val = tensor<int32, []>(3)]; |
| 30 | tensor<int32, [1]> out_1_axes_0 = const()[name = tensor<string, []>("out_1_axes_0"), val = tensor<int32, [1]>([1])]; |
| 31 | tensor<fp16, []> var_96_to_fp16 = const()[name = tensor<string, []>("op_96_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 32 | tensor<fp16, [1, 1280, 1, 1]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_96_to_fp16, x = inputs_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")]; |
| 33 | tensor<fp16, [1280]> obj_1_mean_0_to_fp16 = const()[name = tensor<string, []>("obj_1_mean_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134013056)))]; |
| 34 | tensor<fp16, [1280]> obj_1_variance_0_to_fp16 = const()[name = tensor<string, []>("obj_1_variance_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134015680)))]; |
| 35 | tensor<fp16, [1280]> obj_1_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_1_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134018304)))]; |
| 36 | tensor<fp16, [1280]> obj_1_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_1_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134020928)))]; |
| 37 | tensor<fp16, []> obj_1_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_1_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 38 | tensor<fp16, [1, 1280, 1, 1]> obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor<string, []>("obj_1_cast_fp16")]; |
| 39 | tensor<string, []> var_118_pad_type_0 = const()[name = tensor<string, []>("op_118_pad_type_0"), val = tensor<string, []>("valid")]; |
| 40 | tensor<int32, [2]> var_118_strides_0 = const()[name = tensor<string, []>("op_118_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 41 | tensor<int32, [4]> var_118_pad_0 = const()[name = tensor<string, []>("op_118_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 42 | tensor<int32, [2]> var_118_dilations_0 = const()[name = tensor<string, []>("op_118_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 43 | tensor<int32, []> var_118_groups_0 = const()[name = tensor<string, []>("op_118_groups_0"), val = tensor<int32, []>(1)]; |
| 44 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134023552))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134842816))), name = tensor<string, []>("layers_0_self_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 45 | tensor<fp16, [1280]> layers_0_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134842944)))]; |
| 46 | tensor<fp16, [1, 1280, 1, 1]> var_118_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_118_dilations_0, groups = var_118_groups_0, pad = var_118_pad_0, pad_type = var_118_pad_type_0, strides = var_118_strides_0, weight = layers_0_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor<string, []>("op_118_cast_fp16")]; |
| 47 | tensor<string, []> var_124_pad_type_0 = const()[name = tensor<string, []>("op_124_pad_type_0"), val = tensor<string, []>("valid")]; |
| 48 | tensor<int32, [2]> var_124_strides_0 = const()[name = tensor<string, []>("op_124_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 49 | tensor<int32, [4]> var_124_pad_0 = const()[name = tensor<string, []>("op_124_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 50 | tensor<int32, [2]> var_124_dilations_0 = const()[name = tensor<string, []>("op_124_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 51 | tensor<int32, []> var_124_groups_0 = const()[name = tensor<string, []>("op_124_groups_0"), val = tensor<int32, []>(1)]; |
| 52 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134918592))), name = tensor<string, []>("layers_0_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [36461]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134845568))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 53 | tensor<fp16, [1, 1280, 1, 1]> var_124_cast_fp16 = conv(dilations = var_124_dilations_0, groups = var_124_groups_0, pad = var_124_pad_0, pad_type = var_124_pad_type_0, strides = var_124_strides_0, weight = layers_0_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_1_cast_fp16)[name = tensor<string, []>("op_124_cast_fp16")]; |
| 54 | tensor<fp16, [1, 1280, 1, 1]> query_1_cast_fp16 = add(x = var_118_cast_fp16, y = var_124_cast_fp16)[name = tensor<string, []>("query_1_cast_fp16")]; |
| 55 | tensor<string, []> var_133_pad_type_0 = const()[name = tensor<string, []>("op_133_pad_type_0"), val = tensor<string, []>("valid")]; |
| 56 | tensor<int32, [2]> var_133_strides_0 = const()[name = tensor<string, []>("op_133_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 57 | tensor<int32, [4]> var_133_pad_0 = const()[name = tensor<string, []>("op_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 58 | tensor<int32, [2]> var_133_dilations_0 = const()[name = tensor<string, []>("op_133_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 59 | tensor<int32, []> var_133_groups_0 = const()[name = tensor<string, []>("op_133_groups_0"), val = tensor<int32, []>(1)]; |
| 60 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135123456))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135942720))), name = tensor<string, []>("layers_0_self_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 61 | tensor<fp16, [1, 1280, 1, 1]> var_133_cast_fp16 = conv(dilations = var_133_dilations_0, groups = var_133_groups_0, pad = var_133_pad_0, pad_type = var_133_pad_type_0, strides = var_133_strides_0, weight = layers_0_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor<string, []>("op_133_cast_fp16")]; |
| 62 | tensor<string, []> var_139_pad_type_0 = const()[name = tensor<string, []>("op_139_pad_type_0"), val = tensor<string, []>("valid")]; |
| 63 | tensor<int32, [2]> var_139_strides_0 = const()[name = tensor<string, []>("op_139_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 64 | tensor<int32, [4]> var_139_pad_0 = const()[name = tensor<string, []>("op_139_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 65 | tensor<int32, [2]> var_139_dilations_0 = const()[name = tensor<string, []>("op_139_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 66 | tensor<int32, []> var_139_groups_0 = const()[name = tensor<string, []>("op_139_groups_0"), val = tensor<int32, []>(1)]; |
| 67 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135976320))), name = tensor<string, []>("layers_0_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [16673]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135942848))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 68 | tensor<fp16, [1, 1280, 1, 1]> var_139_cast_fp16 = conv(dilations = var_139_dilations_0, groups = var_139_groups_0, pad = var_139_pad_0, pad_type = var_139_pad_type_0, strides = var_139_strides_0, weight = layers_0_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_1_cast_fp16)[name = tensor<string, []>("op_139_cast_fp16")]; |
| 69 | tensor<fp16, [1, 1280, 1, 1]> current_key_1_cast_fp16 = add(x = var_133_cast_fp16, y = var_139_cast_fp16)[name = tensor<string, []>("current_key_1_cast_fp16")]; |
| 70 | tensor<string, []> var_149_pad_type_0 = const()[name = tensor<string, []>("op_149_pad_type_0"), val = tensor<string, []>("valid")]; |
| 71 | tensor<int32, [2]> var_149_strides_0 = const()[name = tensor<string, []>("op_149_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 72 | tensor<int32, [4]> var_149_pad_0 = const()[name = tensor<string, []>("op_149_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 73 | tensor<int32, [2]> var_149_dilations_0 = const()[name = tensor<string, []>("op_149_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 74 | tensor<int32, []> var_149_groups_0 = const()[name = tensor<string, []>("op_149_groups_0"), val = tensor<int32, []>(1)]; |
| 75 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136181184))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137000448))), name = tensor<string, []>("layers_0_self_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 76 | tensor<fp16, [1280]> layers_0_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137000576)))]; |
| 77 | tensor<fp16, [1, 1280, 1, 1]> var_149_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_149_dilations_0, groups = var_149_groups_0, pad = var_149_pad_0, pad_type = var_149_pad_type_0, strides = var_149_strides_0, weight = layers_0_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor<string, []>("op_149_cast_fp16")]; |
| 78 | tensor<string, []> var_155_pad_type_0 = const()[name = tensor<string, []>("op_155_pad_type_0"), val = tensor<string, []>("valid")]; |
| 79 | tensor<int32, [2]> var_155_strides_0 = const()[name = tensor<string, []>("op_155_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 80 | tensor<int32, [4]> var_155_pad_0 = const()[name = tensor<string, []>("op_155_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 81 | tensor<int32, [2]> var_155_dilations_0 = const()[name = tensor<string, []>("op_155_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 82 | tensor<int32, []> var_155_groups_0 = const()[name = tensor<string, []>("op_155_groups_0"), val = tensor<int32, []>(1)]; |
| 83 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137046720))), name = tensor<string, []>("layers_0_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [21721]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137003200))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 84 | tensor<fp16, [1, 1280, 1, 1]> var_155_cast_fp16 = conv(dilations = var_155_dilations_0, groups = var_155_groups_0, pad = var_155_pad_0, pad_type = var_155_pad_type_0, strides = var_155_strides_0, weight = layers_0_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_1_cast_fp16)[name = tensor<string, []>("op_155_cast_fp16")]; |
| 85 | tensor<fp16, [1, 1280, 1, 1]> current_value_1_cast_fp16 = add(x = var_149_cast_fp16, y = var_155_cast_fp16)[name = tensor<string, []>("current_value_1_cast_fp16")]; |
| 86 | tensor<int32, [1]> var_158_axes_0 = const()[name = tensor<string, []>("op_158_axes_0"), val = tensor<int32, [1]>([1])]; |
| 87 | tensor<fp16, [1, 1, 448]> var_158_cast_fp16 = expand_dims(axes = var_158_axes_0, x = kv_cache_update_mask)[name = tensor<string, []>("op_158_cast_fp16")]; |
| 88 | tensor<int32, [1]> var_159_axes_0 = const()[name = tensor<string, []>("op_159_axes_0"), val = tensor<int32, [1]>([2])]; |
| 89 | tensor<fp16, [1, 1, 1, 448]> var_159_cast_fp16 = expand_dims(axes = var_159_axes_0, x = var_158_cast_fp16)[name = tensor<string, []>("op_159_cast_fp16")]; |
| 90 | tensor<fp16, [1, 1280, 1, 448]> var_161_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_159_cast_fp16)[name = tensor<string, []>("op_161_cast_fp16")]; |
| 91 | tensor<fp16, []> var_71_to_fp16 = const()[name = tensor<string, []>("op_71_to_fp16"), val = tensor<fp16, []>(0x1p+0)]; |
| 92 | tensor<fp16, [1, 1, 1, 448]> var_162_cast_fp16 = sub(x = var_71_to_fp16, y = var_159_cast_fp16)[name = tensor<string, []>("op_162_cast_fp16")]; |
| 93 | tensor<fp16, [1, 1280, 1, 448]> var_163_cast_fp16 = mul(x = var_53_cast_fp16_0, y = var_162_cast_fp16)[name = tensor<string, []>("op_163_cast_fp16")]; |
| 94 | tensor<fp16, [1, 1280, 1, 448]> key_1_cast_fp16 = add(x = var_161_cast_fp16, y = var_163_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")]; |
| 95 | tensor<fp16, [1, 1280, 1, 448]> var_165_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_159_cast_fp16)[name = tensor<string, []>("op_165_cast_fp16")]; |
| 96 | tensor<fp16, [1, 1280, 1, 448]> var_167_cast_fp16 = mul(x = var_60_cast_fp16_0, y = var_162_cast_fp16)[name = tensor<string, []>("op_167_cast_fp16")]; |
| 97 | tensor<fp16, [1, 1280, 1, 448]> value_1_cast_fp16 = add(x = var_165_cast_fp16, y = var_167_cast_fp16)[name = tensor<string, []>("value_1_cast_fp16")]; |
| 98 | tensor<int32, [4]> var_170 = const()[name = tensor<string, []>("op_170"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 99 | tensor<fp16, [1, 20, 64, 1]> mh_q_1_cast_fp16 = reshape(shape = var_170, x = query_1_cast_fp16)[name = tensor<string, []>("mh_q_1_cast_fp16")]; |
| 100 | tensor<fp16, []> var_172_to_fp16 = const()[name = tensor<string, []>("op_172_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| 101 | tensor<fp16, [1, 20, 64, 1]> var_173_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_172_to_fp16)[name = tensor<string, []>("op_173_cast_fp16")]; |
| 102 | tensor<int32, [4]> var_174 = const()[name = tensor<string, []>("op_174"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 103 | tensor<fp16, [1, 20, 64, 448]> var_175_cast_fp16 = reshape(shape = var_174, x = key_1_cast_fp16)[name = tensor<string, []>("op_175_cast_fp16")]; |
| 104 | tensor<bool, []> mh_w_1_transpose_x_0 = const()[name = tensor<string, []>("mh_w_1_transpose_x_0"), val = tensor<bool, []>(true)]; |
| 105 | tensor<bool, []> mh_w_1_transpose_y_0 = const()[name = tensor<string, []>("mh_w_1_transpose_y_0"), val = tensor<bool, []>(false)]; |
| 106 | tensor<fp16, [1, 20, 1, 448]> mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_173_cast_fp16, y = var_175_cast_fp16)[name = tensor<string, []>("mh_w_1_cast_fp16")]; |
| 107 | tensor<int32, [1]> var_179_axes_0 = const()[name = tensor<string, []>("op_179_axes_0"), val = tensor<int32, [1]>([1])]; |
| 108 | tensor<fp16, [1, 1, 448]> var_179_cast_fp16 = expand_dims(axes = var_179_axes_0, x = decoder_key_padding_mask)[name = tensor<string, []>("op_179_cast_fp16")]; |
| 109 | tensor<int32, [1]> var_180_axes_0 = const()[name = tensor<string, []>("op_180_axes_0"), val = tensor<int32, [1]>([2])]; |
| 110 | tensor<fp16, [1, 1, 1, 448]> var_180_cast_fp16 = expand_dims(axes = var_180_axes_0, x = var_179_cast_fp16)[name = tensor<string, []>("op_180_cast_fp16")]; |
| 111 | tensor<fp16, [1, 20, 1, 448]> mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_180_cast_fp16)[name = tensor<string, []>("mh_w_3_cast_fp16")]; |
| 112 | tensor<fp16, [1, 20, 1, 448]> var_183_cast_fp16 = softmax(axis = var_70, x = mh_w_3_cast_fp16)[name = tensor<string, []>("op_183_cast_fp16")]; |
| 113 | tensor<int32, [4]> var_184 = const()[name = tensor<string, []>("op_184"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 114 | tensor<fp16, [1, 20, 64, 448]> var_185_cast_fp16 = reshape(shape = var_184, x = value_1_cast_fp16)[name = tensor<string, []>("op_185_cast_fp16")]; |
| 115 | tensor<bool, []> attn_1_transpose_x_0 = const()[name = tensor<string, []>("attn_1_transpose_x_0"), val = tensor<bool, []>(false)]; |
| 116 | tensor<bool, []> attn_1_transpose_y_0 = const()[name = tensor<string, []>("attn_1_transpose_y_0"), val = tensor<bool, []>(true)]; |
| 117 | tensor<fp16, [1, 20, 64, 1]> attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_185_cast_fp16, y = var_183_cast_fp16)[name = tensor<string, []>("attn_1_cast_fp16")]; |
| 118 | tensor<int32, [4]> var_188 = const()[name = tensor<string, []>("op_188"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; |
| 119 | tensor<fp16, [1, 1280, 1, 1]> input_1_cast_fp16 = reshape(shape = var_188, x = attn_1_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")]; |
| 120 | tensor<string, []> var_198_pad_type_0 = const()[name = tensor<string, []>("op_198_pad_type_0"), val = tensor<string, []>("valid")]; |
| 121 | tensor<int32, [2]> var_198_strides_0 = const()[name = tensor<string, []>("op_198_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 122 | tensor<int32, [4]> var_198_pad_0 = const()[name = tensor<string, []>("op_198_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 123 | tensor<int32, [2]> var_198_dilations_0 = const()[name = tensor<string, []>("op_198_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 124 | tensor<int32, []> var_198_groups_0 = const()[name = tensor<string, []>("op_198_groups_0"), val = tensor<int32, []>(1)]; |
| 125 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137251584))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138070848))), name = tensor<string, []>("layers_0_self_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 126 | tensor<fp16, [1280]> layers_0_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138070976)))]; |
| 127 | tensor<fp16, [1, 1280, 1, 1]> var_198_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_198_dilations_0, groups = var_198_groups_0, pad = var_198_pad_0, pad_type = var_198_pad_type_0, strides = var_198_strides_0, weight = layers_0_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_1_cast_fp16)[name = tensor<string, []>("op_198_cast_fp16")]; |
| 128 | tensor<string, []> var_204_pad_type_0 = const()[name = tensor<string, []>("op_204_pad_type_0"), val = tensor<string, []>("valid")]; |
| 129 | tensor<int32, [2]> var_204_strides_0 = const()[name = tensor<string, []>("op_204_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 130 | tensor<int32, [4]> var_204_pad_0 = const()[name = tensor<string, []>("op_204_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 131 | tensor<int32, [2]> var_204_dilations_0 = const()[name = tensor<string, []>("op_204_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 132 | tensor<int32, []> var_204_groups_0 = const()[name = tensor<string, []>("op_204_groups_0"), val = tensor<int32, []>(1)]; |
| 133 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138130624))), name = tensor<string, []>("layers_0_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [28455]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138073600))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 134 | tensor<fp16, [1, 1280, 1, 1]> var_204_cast_fp16 = conv(dilations = var_204_dilations_0, groups = var_204_groups_0, pad = var_204_pad_0, pad_type = var_204_pad_type_0, strides = var_204_strides_0, weight = layers_0_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_1_cast_fp16)[name = tensor<string, []>("op_204_cast_fp16")]; |
| 135 | tensor<fp16, [1, 1280, 1, 1]> obj_7_cast_fp16 = add(x = var_198_cast_fp16, y = var_204_cast_fp16)[name = tensor<string, []>("obj_7_cast_fp16")]; |
| 136 | tensor<fp16, [1, 1280, 1, 1]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_7_cast_fp16)[name = tensor<string, []>("inputs_3_cast_fp16")]; |
| 137 | tensor<int32, [1]> out_3_axes_0 = const()[name = tensor<string, []>("out_3_axes_0"), val = tensor<int32, [1]>([1])]; |
| 138 | tensor<fp16, []> var_219_to_fp16 = const()[name = tensor<string, []>("op_219_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 139 | tensor<fp16, [1, 1280, 1, 1]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_219_to_fp16, x = inputs_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")]; |
| 140 | tensor<fp16, [1280]> obj_9_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_9_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138335488)))]; |
| 141 | tensor<fp16, [1280]> obj_9_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_9_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138338112)))]; |
| 142 | tensor<fp16, []> obj_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 143 | tensor<fp16, [1, 1280, 1, 1]> obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor<string, []>("obj_9_cast_fp16")]; |
| 144 | tensor<string, []> var_241_pad_type_0 = const()[name = tensor<string, []>("op_241_pad_type_0"), val = tensor<string, []>("valid")]; |
| 145 | tensor<int32, [2]> var_241_strides_0 = const()[name = tensor<string, []>("op_241_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 146 | tensor<int32, [4]> var_241_pad_0 = const()[name = tensor<string, []>("op_241_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 147 | tensor<int32, [2]> var_241_dilations_0 = const()[name = tensor<string, []>("op_241_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 148 | tensor<int32, []> var_241_groups_0 = const()[name = tensor<string, []>("op_241_groups_0"), val = tensor<int32, []>(1)]; |
| 149 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138340736))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139160000))), name = tensor<string, []>("layers_0_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 150 | tensor<fp16, [1280]> layers_0_encoder_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139160128)))]; |
| 151 | tensor<fp16, [1, 1280, 1, 1]> var_241_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_241_dilations_0, groups = var_241_groups_0, pad = var_241_pad_0, pad_type = var_241_pad_type_0, strides = var_241_strides_0, weight = layers_0_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = tensor<string, []>("op_241_cast_fp16")]; |
| 152 | tensor<string, []> var_247_pad_type_0 = const()[name = tensor<string, []>("op_247_pad_type_0"), val = tensor<string, []>("valid")]; |
| 153 | tensor<int32, [2]> var_247_strides_0 = const()[name = tensor<string, []>("op_247_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 154 | tensor<int32, [4]> var_247_pad_0 = const()[name = tensor<string, []>("op_247_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 155 | tensor<int32, [2]> var_247_dilations_0 = const()[name = tensor<string, []>("op_247_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 156 | tensor<int32, []> var_247_groups_0 = const()[name = tensor<string, []>("op_247_groups_0"), val = tensor<int32, []>(1)]; |
| 157 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139188224))), name = tensor<string, []>("layers_0_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [12701]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139162752))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 158 | tensor<fp16, [1, 1280, 1, 1]> var_247_cast_fp16 = conv(dilations = var_247_dilations_0, groups = var_247_groups_0, pad = var_247_pad_0, pad_type = var_247_pad_type_0, strides = var_247_strides_0, weight = layers_0_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_9_cast_fp16)[name = tensor<string, []>("op_247_cast_fp16")]; |
| 159 | tensor<fp16, [1, 1280, 1, 1]> query_3_cast_fp16 = add(x = var_241_cast_fp16, y = var_247_cast_fp16)[name = tensor<string, []>("query_3_cast_fp16")]; |
| 160 | tensor<string, []> var_256_pad_type_0 = const()[name = tensor<string, []>("op_256_pad_type_0"), val = tensor<string, []>("valid")]; |
| 161 | tensor<int32, [2]> var_256_strides_0 = const()[name = tensor<string, []>("op_256_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 162 | tensor<int32, [4]> var_256_pad_0 = const()[name = tensor<string, []>("op_256_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 163 | tensor<int32, [2]> var_256_dilations_0 = const()[name = tensor<string, []>("op_256_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 164 | tensor<int32, []> var_256_groups_0 = const()[name = tensor<string, []>("op_256_groups_0"), val = tensor<int32, []>(1)]; |
| 165 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139393088))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140212352))), name = tensor<string, []>("layers_0_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 166 | tensor<fp16, [1, 1280, 1, 1500]> var_256_cast_fp16 = conv(dilations = var_256_dilations_0, groups = var_256_groups_0, pad = var_256_pad_0, pad_type = var_256_pad_type_0, strides = var_256_strides_0, weight = layers_0_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("op_256_cast_fp16")]; |
| 167 | tensor<string, []> var_262_pad_type_0 = const()[name = tensor<string, []>("op_262_pad_type_0"), val = tensor<string, []>("valid")]; |
| 168 | tensor<int32, [2]> var_262_strides_0 = const()[name = tensor<string, []>("op_262_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 169 | tensor<int32, [4]> var_262_pad_0 = const()[name = tensor<string, []>("op_262_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 170 | tensor<int32, [2]> var_262_dilations_0 = const()[name = tensor<string, []>("op_262_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 171 | tensor<int32, []> var_262_groups_0 = const()[name = tensor<string, []>("op_262_groups_0"), val = tensor<int32, []>(1)]; |
| 172 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140272448))), name = tensor<string, []>("layers_0_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [29949]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140212480))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 173 | tensor<fp16, [1, 1280, 1, 1500]> var_262_cast_fp16 = conv(dilations = var_262_dilations_0, groups = var_262_groups_0, pad = var_262_pad_0, pad_type = var_262_pad_type_0, strides = var_262_strides_0, weight = layers_0_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = tensor<string, []>("op_262_cast_fp16")]; |
| 174 | tensor<fp16, [1, 1280, 1, 1500]> key_3_cast_fp16 = add(x = var_256_cast_fp16, y = var_262_cast_fp16)[name = tensor<string, []>("key_3_cast_fp16")]; |
| 175 | tensor<string, []> var_272_pad_type_0 = const()[name = tensor<string, []>("op_272_pad_type_0"), val = tensor<string, []>("valid")]; |
| 176 | tensor<int32, [2]> var_272_strides_0 = const()[name = tensor<string, []>("op_272_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 177 | tensor<int32, [4]> var_272_pad_0 = const()[name = tensor<string, []>("op_272_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 178 | tensor<int32, [2]> var_272_dilations_0 = const()[name = tensor<string, []>("op_272_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 179 | tensor<int32, []> var_272_groups_0 = const()[name = tensor<string, []>("op_272_groups_0"), val = tensor<int32, []>(1)]; |
| 180 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140477312))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141296576))), name = tensor<string, []>("layers_0_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 181 | tensor<fp16, [1280]> layers_0_encoder_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141296704)))]; |
| 182 | tensor<fp16, [1, 1280, 1, 1500]> var_272_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_272_dilations_0, groups = var_272_groups_0, pad = var_272_pad_0, pad_type = var_272_pad_type_0, strides = var_272_strides_0, weight = layers_0_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("op_272_cast_fp16")]; |
| 183 | tensor<string, []> var_278_pad_type_0 = const()[name = tensor<string, []>("op_278_pad_type_0"), val = tensor<string, []>("valid")]; |
| 184 | tensor<int32, [2]> var_278_strides_0 = const()[name = tensor<string, []>("op_278_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 185 | tensor<int32, [4]> var_278_pad_0 = const()[name = tensor<string, []>("op_278_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 186 | tensor<int32, [2]> var_278_dilations_0 = const()[name = tensor<string, []>("op_278_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 187 | tensor<int32, []> var_278_groups_0 = const()[name = tensor<string, []>("op_278_groups_0"), val = tensor<int32, []>(1)]; |
| 188 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141310592))), name = tensor<string, []>("layers_0_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [5596]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141299328))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 189 | tensor<fp16, [1, 1280, 1, 1500]> var_278_cast_fp16 = conv(dilations = var_278_dilations_0, groups = var_278_groups_0, pad = var_278_pad_0, pad_type = var_278_pad_type_0, strides = var_278_strides_0, weight = layers_0_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = tensor<string, []>("op_278_cast_fp16")]; |
| 190 | tensor<fp16, [1, 1280, 1, 1500]> value_3_cast_fp16 = add(x = var_272_cast_fp16, y = var_278_cast_fp16)[name = tensor<string, []>("value_3_cast_fp16")]; |
| 191 | tensor<int32, [4]> var_281 = const()[name = tensor<string, []>("op_281"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 192 | tensor<fp16, [1, 20, 64, 1]> mh_q_3_cast_fp16 = reshape(shape = var_281, x = query_3_cast_fp16)[name = tensor<string, []>("mh_q_3_cast_fp16")]; |
| 193 | tensor<fp16, []> var_283_to_fp16 = const()[name = tensor<string, []>("op_283_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| 194 | tensor<fp16, [1, 20, 64, 1]> var_284_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_283_to_fp16)[name = tensor<string, []>("op_284_cast_fp16")]; |
| 195 | tensor<int32, [4]> var_285 = const()[name = tensor<string, []>("op_285"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 196 | tensor<fp16, [1, 20, 64, 1500]> var_286_cast_fp16 = reshape(shape = var_285, x = key_3_cast_fp16)[name = tensor<string, []>("op_286_cast_fp16")]; |
| 197 | tensor<bool, []> mh_w_5_transpose_x_0 = const()[name = tensor<string, []>("mh_w_5_transpose_x_0"), val = tensor<bool, []>(true)]; |
| 198 | tensor<bool, []> mh_w_5_transpose_y_0 = const()[name = tensor<string, []>("mh_w_5_transpose_y_0"), val = tensor<bool, []>(false)]; |
| 199 | tensor<fp16, [1, 20, 1, 1500]> mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_284_cast_fp16, y = var_286_cast_fp16)[name = tensor<string, []>("mh_w_5_cast_fp16")]; |
| 200 | tensor<fp16, [1, 20, 1, 1500]> obj_13_cast_fp16 = softmax(axis = var_70, x = mh_w_5_cast_fp16)[name = tensor<string, []>("obj_13_cast_fp16")]; |
| 201 | tensor<int32, [4]> var_290 = const()[name = tensor<string, []>("op_290"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 202 | tensor<fp16, [1, 20, 64, 1500]> var_291_cast_fp16 = reshape(shape = var_290, x = value_3_cast_fp16)[name = tensor<string, []>("op_291_cast_fp16")]; |
| 203 | tensor<bool, []> attn_3_transpose_x_0 = const()[name = tensor<string, []>("attn_3_transpose_x_0"), val = tensor<bool, []>(false)]; |
| 204 | tensor<bool, []> attn_3_transpose_y_0 = const()[name = tensor<string, []>("attn_3_transpose_y_0"), val = tensor<bool, []>(true)]; |
| 205 | tensor<fp16, [1, 20, 64, 1]> attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_291_cast_fp16, y = obj_13_cast_fp16)[name = tensor<string, []>("attn_3_cast_fp16")]; |
| 206 | tensor<int32, [4]> var_294 = const()[name = tensor<string, []>("op_294"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; |
| 207 | tensor<fp16, [1, 1280, 1, 1]> input_3_cast_fp16 = reshape(shape = var_294, x = attn_3_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")]; |
| 208 | tensor<string, []> var_304_pad_type_0 = const()[name = tensor<string, []>("op_304_pad_type_0"), val = tensor<string, []>("valid")]; |
| 209 | tensor<int32, [2]> var_304_strides_0 = const()[name = tensor<string, []>("op_304_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 210 | tensor<int32, [4]> var_304_pad_0 = const()[name = tensor<string, []>("op_304_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 211 | tensor<int32, [2]> var_304_dilations_0 = const()[name = tensor<string, []>("op_304_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 212 | tensor<int32, []> var_304_groups_0 = const()[name = tensor<string, []>("op_304_groups_0"), val = tensor<int32, []>(1)]; |
| 213 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141515456))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142334720))), name = tensor<string, []>("layers_0_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 214 | tensor<fp16, [1280]> layers_0_encoder_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142334848)))]; |
| 215 | tensor<fp16, [1, 1280, 1, 1]> var_304_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_304_dilations_0, groups = var_304_groups_0, pad = var_304_pad_0, pad_type = var_304_pad_type_0, strides = var_304_strides_0, weight = layers_0_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_3_cast_fp16)[name = tensor<string, []>("op_304_cast_fp16")]; |
| 216 | tensor<string, []> var_310_pad_type_0 = const()[name = tensor<string, []>("op_310_pad_type_0"), val = tensor<string, []>("valid")]; |
| 217 | tensor<int32, [2]> var_310_strides_0 = const()[name = tensor<string, []>("op_310_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 218 | tensor<int32, [4]> var_310_pad_0 = const()[name = tensor<string, []>("op_310_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 219 | tensor<int32, [2]> var_310_dilations_0 = const()[name = tensor<string, []>("op_310_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 220 | tensor<int32, []> var_310_groups_0 = const()[name = tensor<string, []>("op_310_groups_0"), val = tensor<int32, []>(1)]; |
| 221 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142349632))), name = tensor<string, []>("layers_0_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [6041]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142337472))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 222 | tensor<fp16, [1, 1280, 1, 1]> var_310_cast_fp16 = conv(dilations = var_310_dilations_0, groups = var_310_groups_0, pad = var_310_pad_0, pad_type = var_310_pad_type_0, strides = var_310_strides_0, weight = layers_0_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_3_cast_fp16)[name = tensor<string, []>("op_310_cast_fp16")]; |
| 223 | tensor<fp16, [1, 1280, 1, 1]> obj_11_cast_fp16 = add(x = var_304_cast_fp16, y = var_310_cast_fp16)[name = tensor<string, []>("obj_11_cast_fp16")]; |
| 224 | tensor<fp16, [1, 1280, 1, 1]> inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_11_cast_fp16)[name = tensor<string, []>("inputs_5_cast_fp16")]; |
| 225 | tensor<int32, [1]> out_5_axes_0 = const()[name = tensor<string, []>("out_5_axes_0"), val = tensor<int32, [1]>([1])]; |
| 226 | tensor<fp16, []> var_321_to_fp16 = const()[name = tensor<string, []>("op_321_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 227 | tensor<fp16, [1, 1280, 1, 1]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_321_to_fp16, x = inputs_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")]; |
| 228 | tensor<fp16, [1280]> input_5_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_5_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142554496)))]; |
| 229 | tensor<fp16, [1280]> input_5_beta_0_to_fp16 = const()[name = tensor<string, []>("input_5_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142557120)))]; |
| 230 | tensor<fp16, []> input_5_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_5_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 231 | tensor<fp16, [1, 1280, 1, 1]> input_5_cast_fp16 = batch_norm(beta = input_5_beta_0_to_fp16, epsilon = input_5_epsilon_0_to_fp16, gamma = input_5_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")]; |
| 232 | tensor<string, []> var_339_pad_type_0 = const()[name = tensor<string, []>("op_339_pad_type_0"), val = tensor<string, []>("valid")]; |
| 233 | tensor<int32, [2]> var_339_strides_0 = const()[name = tensor<string, []>("op_339_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 234 | tensor<int32, [4]> var_339_pad_0 = const()[name = tensor<string, []>("op_339_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 235 | tensor<int32, [2]> var_339_dilations_0 = const()[name = tensor<string, []>("op_339_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 236 | tensor<int32, []> var_339_groups_0 = const()[name = tensor<string, []>("op_339_groups_0"), val = tensor<int32, []>(1)]; |
| 237 | tensor<fp16, [5120, 1280, 1, 1]> layers_0_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142559744))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145836608))), name = tensor<string, []>("layers_0_fc1_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])]; |
| 238 | tensor<fp16, [5120]> layers_0_fc1_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145836736)))]; |
| 239 | tensor<fp16, [1, 5120, 1, 1]> var_339_cast_fp16 = conv(bias = layers_0_fc1_inlier_module_bias_to_fp16, dilations = var_339_dilations_0, groups = var_339_groups_0, pad = var_339_pad_0, pad_type = var_339_pad_type_0, strides = var_339_strides_0, weight = layers_0_fc1_inlier_module_weight_to_fp16_palettized, x = input_5_cast_fp16)[name = tensor<string, []>("op_339_cast_fp16")]; |
| 240 | tensor<string, []> var_345_pad_type_0 = const()[name = tensor<string, []>("op_345_pad_type_0"), val = tensor<string, []>("valid")]; |
| 241 | tensor<int32, [2]> var_345_strides_0 = const()[name = tensor<string, []>("op_345_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 242 | tensor<int32, [4]> var_345_pad_0 = const()[name = tensor<string, []>("op_345_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 243 | tensor<int32, [2]> var_345_dilations_0 = const()[name = tensor<string, []>("op_345_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 244 | tensor<int32, []> var_345_groups_0 = const()[name = tensor<string, []>("op_345_groups_0"), val = tensor<int32, []>(1)]; |
| 245 | tensor<fp16, [5120, 1280, 1, 1]> layers_0_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145948608))), name = tensor<string, []>("layers_0_fc1_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [50752]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145847040))), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])]; |
| 246 | tensor<fp16, [1, 5120, 1, 1]> var_345_cast_fp16 = conv(dilations = var_345_dilations_0, groups = var_345_groups_0, pad = var_345_pad_0, pad_type = var_345_pad_type_0, strides = var_345_strides_0, weight = layers_0_fc1_outlier_module_weight_to_fp16_sparsified, x = input_5_cast_fp16)[name = tensor<string, []>("op_345_cast_fp16")]; |
| 247 | tensor<fp16, [1, 5120, 1, 1]> input_7_cast_fp16 = add(x = var_339_cast_fp16, y = var_345_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")]; |
| 248 | tensor<string, []> input_9_mode_0 = const()[name = tensor<string, []>("input_9_mode_0"), val = tensor<string, []>("EXACT")]; |
| 249 | tensor<fp16, [1, 5120, 1, 1]> input_9_cast_fp16 = gelu(mode = input_9_mode_0, x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")]; |
| 250 | tensor<string, []> var_356_pad_type_0 = const()[name = tensor<string, []>("op_356_pad_type_0"), val = tensor<string, []>("valid")]; |
| 251 | tensor<int32, [2]> var_356_strides_0 = const()[name = tensor<string, []>("op_356_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 252 | tensor<int32, [4]> var_356_pad_0 = const()[name = tensor<string, []>("op_356_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 253 | tensor<int32, [2]> var_356_dilations_0 = const()[name = tensor<string, []>("op_356_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 254 | tensor<int32, []> var_356_groups_0 = const()[name = tensor<string, []>("op_356_groups_0"), val = tensor<int32, []>(1)]; |
| 255 | tensor<fp16, [1280, 5120, 1, 1]> layers_0_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146767872))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150044736))), name = tensor<string, []>("layers_0_fc2_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])]; |
| 256 | tensor<fp16, [1280]> layers_0_fc2_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150044864)))]; |
| 257 | tensor<fp16, [1, 1280, 1, 1]> var_356_cast_fp16 = conv(bias = layers_0_fc2_inlier_module_bias_to_fp16, dilations = var_356_dilations_0, groups = var_356_groups_0, pad = var_356_pad_0, pad_type = var_356_pad_type_0, strides = var_356_strides_0, weight = layers_0_fc2_inlier_module_weight_to_fp16_palettized, x = input_9_cast_fp16)[name = tensor<string, []>("op_356_cast_fp16")]; |
| 258 | tensor<string, []> var_362_pad_type_0 = const()[name = tensor<string, []>("op_362_pad_type_0"), val = tensor<string, []>("valid")]; |
| 259 | tensor<int32, [2]> var_362_strides_0 = const()[name = tensor<string, []>("op_362_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 260 | tensor<int32, [4]> var_362_pad_0 = const()[name = tensor<string, []>("op_362_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 261 | tensor<int32, [2]> var_362_dilations_0 = const()[name = tensor<string, []>("op_362_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 262 | tensor<int32, []> var_362_groups_0 = const()[name = tensor<string, []>("op_362_groups_0"), val = tensor<int32, []>(1)]; |
| 263 | tensor<fp16, [1280, 5120, 1, 1]> layers_0_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150230016))), name = tensor<string, []>("layers_0_fc2_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [91213]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150047488))), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])]; |
| 264 | tensor<fp16, [1, 1280, 1, 1]> var_362_cast_fp16 = conv(dilations = var_362_dilations_0, groups = var_362_groups_0, pad = var_362_pad_0, pad_type = var_362_pad_type_0, strides = var_362_strides_0, weight = layers_0_fc2_outlier_module_weight_to_fp16_sparsified, x = input_9_cast_fp16)[name = tensor<string, []>("op_362_cast_fp16")]; |
| 265 | tensor<fp16, [1, 1280, 1, 1]> hidden_states_3_cast_fp16 = add(x = var_356_cast_fp16, y = var_362_cast_fp16)[name = tensor<string, []>("hidden_states_3_cast_fp16")]; |
| 266 | tensor<fp16, [1, 1280, 1, 1]> inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = tensor<string, []>("inputs_7_cast_fp16")]; |
| 267 | tensor<int32, []> var_374 = const()[name = tensor<string, []>("op_374"), val = tensor<int32, []>(3)]; |
| 268 | tensor<int32, [1]> out_7_axes_0 = const()[name = tensor<string, []>("out_7_axes_0"), val = tensor<int32, [1]>([1])]; |
| 269 | tensor<fp16, []> var_400_to_fp16 = const()[name = tensor<string, []>("op_400_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 270 | tensor<fp16, [1, 1280, 1, 1]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_400_to_fp16, x = inputs_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")]; |
| 271 | tensor<fp16, [1280]> obj_15_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_15_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151049280)))]; |
| 272 | tensor<fp16, [1280]> obj_15_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_15_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151051904)))]; |
| 273 | tensor<fp16, []> obj_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 274 | tensor<fp16, [1, 1280, 1, 1]> obj_15_cast_fp16 = batch_norm(beta = obj_15_beta_0_to_fp16, epsilon = obj_15_epsilon_0_to_fp16, gamma = obj_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor<string, []>("obj_15_cast_fp16")]; |
| 275 | tensor<string, []> var_422_pad_type_0 = const()[name = tensor<string, []>("op_422_pad_type_0"), val = tensor<string, []>("valid")]; |
| 276 | tensor<int32, [2]> var_422_strides_0 = const()[name = tensor<string, []>("op_422_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 277 | tensor<int32, [4]> var_422_pad_0 = const()[name = tensor<string, []>("op_422_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 278 | tensor<int32, [2]> var_422_dilations_0 = const()[name = tensor<string, []>("op_422_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 279 | tensor<int32, []> var_422_groups_0 = const()[name = tensor<string, []>("op_422_groups_0"), val = tensor<int32, []>(1)]; |
| 280 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151054528))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151873792))), name = tensor<string, []>("layers_1_self_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 281 | tensor<fp16, [1280]> layers_1_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151873920)))]; |
| 282 | tensor<fp16, [1, 1280, 1, 1]> var_422_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_422_dilations_0, groups = var_422_groups_0, pad = var_422_pad_0, pad_type = var_422_pad_type_0, strides = var_422_strides_0, weight = layers_1_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor<string, []>("op_422_cast_fp16")]; |
| 283 | tensor<string, []> var_428_pad_type_0 = const()[name = tensor<string, []>("op_428_pad_type_0"), val = tensor<string, []>("valid")]; |
| 284 | tensor<int32, [2]> var_428_strides_0 = const()[name = tensor<string, []>("op_428_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 285 | tensor<int32, [4]> var_428_pad_0 = const()[name = tensor<string, []>("op_428_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 286 | tensor<int32, [2]> var_428_dilations_0 = const()[name = tensor<string, []>("op_428_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 287 | tensor<int32, []> var_428_groups_0 = const()[name = tensor<string, []>("op_428_groups_0"), val = tensor<int32, []>(1)]; |
| 288 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151936640))), name = tensor<string, []>("layers_1_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [29985]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151876544))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 289 | tensor<fp16, [1, 1280, 1, 1]> var_428_cast_fp16 = conv(dilations = var_428_dilations_0, groups = var_428_groups_0, pad = var_428_pad_0, pad_type = var_428_pad_type_0, strides = var_428_strides_0, weight = layers_1_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_15_cast_fp16)[name = tensor<string, []>("op_428_cast_fp16")]; |
| 290 | tensor<fp16, [1, 1280, 1, 1]> query_5_cast_fp16 = add(x = var_422_cast_fp16, y = var_428_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")]; |
| 291 | tensor<string, []> var_437_pad_type_0 = const()[name = tensor<string, []>("op_437_pad_type_0"), val = tensor<string, []>("valid")]; |
| 292 | tensor<int32, [2]> var_437_strides_0 = const()[name = tensor<string, []>("op_437_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 293 | tensor<int32, [4]> var_437_pad_0 = const()[name = tensor<string, []>("op_437_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 294 | tensor<int32, [2]> var_437_dilations_0 = const()[name = tensor<string, []>("op_437_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 295 | tensor<int32, []> var_437_groups_0 = const()[name = tensor<string, []>("op_437_groups_0"), val = tensor<int32, []>(1)]; |
| 296 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152141504))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152960768))), name = tensor<string, []>("layers_1_self_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 297 | tensor<fp16, [1, 1280, 1, 1]> var_437_cast_fp16 = conv(dilations = var_437_dilations_0, groups = var_437_groups_0, pad = var_437_pad_0, pad_type = var_437_pad_type_0, strides = var_437_strides_0, weight = layers_1_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor<string, []>("op_437_cast_fp16")]; |
| 298 | tensor<string, []> var_443_pad_type_0 = const()[name = tensor<string, []>("op_443_pad_type_0"), val = tensor<string, []>("valid")]; |
| 299 | tensor<int32, [2]> var_443_strides_0 = const()[name = tensor<string, []>("op_443_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 300 | tensor<int32, [4]> var_443_pad_0 = const()[name = tensor<string, []>("op_443_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 301 | tensor<int32, [2]> var_443_dilations_0 = const()[name = tensor<string, []>("op_443_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 302 | tensor<int32, []> var_443_groups_0 = const()[name = tensor<string, []>("op_443_groups_0"), val = tensor<int32, []>(1)]; |
| 303 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153007552))), name = tensor<string, []>("layers_1_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [23287]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152960896))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 304 | tensor<fp16, [1, 1280, 1, 1]> var_443_cast_fp16 = conv(dilations = var_443_dilations_0, groups = var_443_groups_0, pad = var_443_pad_0, pad_type = var_443_pad_type_0, strides = var_443_strides_0, weight = layers_1_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_15_cast_fp16)[name = tensor<string, []>("op_443_cast_fp16")]; |
| 305 | tensor<fp16, [1, 1280, 1, 1]> current_key_3_cast_fp16 = add(x = var_437_cast_fp16, y = var_443_cast_fp16)[name = tensor<string, []>("current_key_3_cast_fp16")]; |
| 306 | tensor<string, []> var_453_pad_type_0 = const()[name = tensor<string, []>("op_453_pad_type_0"), val = tensor<string, []>("valid")]; |
| 307 | tensor<int32, [2]> var_453_strides_0 = const()[name = tensor<string, []>("op_453_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 308 | tensor<int32, [4]> var_453_pad_0 = const()[name = tensor<string, []>("op_453_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 309 | tensor<int32, [2]> var_453_dilations_0 = const()[name = tensor<string, []>("op_453_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 310 | tensor<int32, []> var_453_groups_0 = const()[name = tensor<string, []>("op_453_groups_0"), val = tensor<int32, []>(1)]; |
| 311 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153212416))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154031680))), name = tensor<string, []>("layers_1_self_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 312 | tensor<fp16, [1280]> layers_1_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154031808)))]; |
| 313 | tensor<fp16, [1, 1280, 1, 1]> var_453_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_453_dilations_0, groups = var_453_groups_0, pad = var_453_pad_0, pad_type = var_453_pad_type_0, strides = var_453_strides_0, weight = layers_1_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor<string, []>("op_453_cast_fp16")]; |
| 314 | tensor<string, []> var_459_pad_type_0 = const()[name = tensor<string, []>("op_459_pad_type_0"), val = tensor<string, []>("valid")]; |
| 315 | tensor<int32, [2]> var_459_strides_0 = const()[name = tensor<string, []>("op_459_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 316 | tensor<int32, [4]> var_459_pad_0 = const()[name = tensor<string, []>("op_459_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 317 | tensor<int32, [2]> var_459_dilations_0 = const()[name = tensor<string, []>("op_459_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 318 | tensor<int32, []> var_459_groups_0 = const()[name = tensor<string, []>("op_459_groups_0"), val = tensor<int32, []>(1)]; |
| 319 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154057088))), name = tensor<string, []>("layers_1_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [11267]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154034432))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 320 | tensor<fp16, [1, 1280, 1, 1]> var_459_cast_fp16 = conv(dilations = var_459_dilations_0, groups = var_459_groups_0, pad = var_459_pad_0, pad_type = var_459_pad_type_0, strides = var_459_strides_0, weight = layers_1_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_15_cast_fp16)[name = tensor<string, []>("op_459_cast_fp16")]; |
| 321 | tensor<fp16, [1, 1280, 1, 1]> current_value_3_cast_fp16 = add(x = var_453_cast_fp16, y = var_459_cast_fp16)[name = tensor<string, []>("current_value_3_cast_fp16")]; |
| 322 | tensor<fp16, [1, 1280, 1, 448]> var_465_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_159_cast_fp16)[name = tensor<string, []>("op_465_cast_fp16")]; |
| 323 | tensor<fp16, [1, 1280, 1, 448]> var_467_cast_fp16 = mul(x = var_53_cast_fp16_1, y = var_162_cast_fp16)[name = tensor<string, []>("op_467_cast_fp16")]; |
| 324 | tensor<fp16, [1, 1280, 1, 448]> key_5_cast_fp16 = add(x = var_465_cast_fp16, y = var_467_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")]; |
| 325 | tensor<fp16, [1, 1280, 1, 448]> var_469_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_159_cast_fp16)[name = tensor<string, []>("op_469_cast_fp16")]; |
| 326 | tensor<fp16, [1, 1280, 1, 448]> var_471_cast_fp16 = mul(x = var_60_cast_fp16_1, y = var_162_cast_fp16)[name = tensor<string, []>("op_471_cast_fp16")]; |
| 327 | tensor<fp16, [1, 1280, 1, 448]> value_5_cast_fp16 = add(x = var_469_cast_fp16, y = var_471_cast_fp16)[name = tensor<string, []>("value_5_cast_fp16")]; |
| 328 | tensor<int32, [4]> var_474 = const()[name = tensor<string, []>("op_474"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 329 | tensor<fp16, [1, 20, 64, 1]> mh_q_5_cast_fp16 = reshape(shape = var_474, x = query_5_cast_fp16)[name = tensor<string, []>("mh_q_5_cast_fp16")]; |
| 330 | tensor<fp16, []> var_476_to_fp16 = const()[name = tensor<string, []>("op_476_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| 331 | tensor<fp16, [1, 20, 64, 1]> var_477_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_476_to_fp16)[name = tensor<string, []>("op_477_cast_fp16")]; |
| 332 | tensor<int32, [4]> var_478 = const()[name = tensor<string, []>("op_478"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 333 | tensor<fp16, [1, 20, 64, 448]> var_479_cast_fp16 = reshape(shape = var_478, x = key_5_cast_fp16)[name = tensor<string, []>("op_479_cast_fp16")]; |
| 334 | tensor<bool, []> mh_w_7_transpose_x_0 = const()[name = tensor<string, []>("mh_w_7_transpose_x_0"), val = tensor<bool, []>(true)]; |
| 335 | tensor<bool, []> mh_w_7_transpose_y_0 = const()[name = tensor<string, []>("mh_w_7_transpose_y_0"), val = tensor<bool, []>(false)]; |
| 336 | tensor<fp16, [1, 20, 1, 448]> mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_477_cast_fp16, y = var_479_cast_fp16)[name = tensor<string, []>("mh_w_7_cast_fp16")]; |
| 337 | tensor<fp16, [1, 20, 1, 448]> mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_180_cast_fp16)[name = tensor<string, []>("mh_w_9_cast_fp16")]; |
| 338 | tensor<fp16, [1, 20, 1, 448]> var_487_cast_fp16 = softmax(axis = var_374, x = mh_w_9_cast_fp16)[name = tensor<string, []>("op_487_cast_fp16")]; |
| 339 | tensor<int32, [4]> var_488 = const()[name = tensor<string, []>("op_488"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 340 | tensor<fp16, [1, 20, 64, 448]> var_489_cast_fp16 = reshape(shape = var_488, x = value_5_cast_fp16)[name = tensor<string, []>("op_489_cast_fp16")]; |
| 341 | tensor<bool, []> attn_5_transpose_x_0 = const()[name = tensor<string, []>("attn_5_transpose_x_0"), val = tensor<bool, []>(false)]; |
| 342 | tensor<bool, []> attn_5_transpose_y_0 = const()[name = tensor<string, []>("attn_5_transpose_y_0"), val = tensor<bool, []>(true)]; |
| 343 | tensor<fp16, [1, 20, 64, 1]> attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_489_cast_fp16, y = var_487_cast_fp16)[name = tensor<string, []>("attn_5_cast_fp16")]; |
| 344 | tensor<int32, [4]> var_492 = const()[name = tensor<string, []>("op_492"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; |
| 345 | tensor<fp16, [1, 1280, 1, 1]> input_11_cast_fp16 = reshape(shape = var_492, x = attn_5_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")]; |
| 346 | tensor<string, []> var_502_pad_type_0 = const()[name = tensor<string, []>("op_502_pad_type_0"), val = tensor<string, []>("valid")]; |
| 347 | tensor<int32, [2]> var_502_strides_0 = const()[name = tensor<string, []>("op_502_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 348 | tensor<int32, [4]> var_502_pad_0 = const()[name = tensor<string, []>("op_502_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 349 | tensor<int32, [2]> var_502_dilations_0 = const()[name = tensor<string, []>("op_502_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 350 | tensor<int32, []> var_502_groups_0 = const()[name = tensor<string, []>("op_502_groups_0"), val = tensor<int32, []>(1)]; |
| 351 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154261952))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155081216))), name = tensor<string, []>("layers_1_self_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 352 | tensor<fp16, [1280]> layers_1_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155081344)))]; |
| 353 | tensor<fp16, [1, 1280, 1, 1]> var_502_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_502_dilations_0, groups = var_502_groups_0, pad = var_502_pad_0, pad_type = var_502_pad_type_0, strides = var_502_strides_0, weight = layers_1_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor<string, []>("op_502_cast_fp16")]; |
| 354 | tensor<string, []> var_508_pad_type_0 = const()[name = tensor<string, []>("op_508_pad_type_0"), val = tensor<string, []>("valid")]; |
| 355 | tensor<int32, [2]> var_508_strides_0 = const()[name = tensor<string, []>("op_508_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 356 | tensor<int32, [4]> var_508_pad_0 = const()[name = tensor<string, []>("op_508_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 357 | tensor<int32, [2]> var_508_dilations_0 = const()[name = tensor<string, []>("op_508_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 358 | tensor<int32, []> var_508_groups_0 = const()[name = tensor<string, []>("op_508_groups_0"), val = tensor<int32, []>(1)]; |
| 359 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155108416))), name = tensor<string, []>("layers_1_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [12187]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155083968))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 360 | tensor<fp16, [1, 1280, 1, 1]> var_508_cast_fp16 = conv(dilations = var_508_dilations_0, groups = var_508_groups_0, pad = var_508_pad_0, pad_type = var_508_pad_type_0, strides = var_508_strides_0, weight = layers_1_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_11_cast_fp16)[name = tensor<string, []>("op_508_cast_fp16")]; |
| 361 | tensor<fp16, [1, 1280, 1, 1]> obj_21_cast_fp16 = add(x = var_502_cast_fp16, y = var_508_cast_fp16)[name = tensor<string, []>("obj_21_cast_fp16")]; |
| 362 | tensor<fp16, [1, 1280, 1, 1]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_21_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")]; |
| 363 | tensor<int32, [1]> out_9_axes_0 = const()[name = tensor<string, []>("out_9_axes_0"), val = tensor<int32, [1]>([1])]; |
| 364 | tensor<fp16, []> var_523_to_fp16 = const()[name = tensor<string, []>("op_523_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 365 | tensor<fp16, [1, 1280, 1, 1]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_523_to_fp16, x = inputs_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")]; |
| 366 | tensor<fp16, [1280]> obj_23_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_23_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155313280)))]; |
| 367 | tensor<fp16, [1280]> obj_23_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_23_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155315904)))]; |
| 368 | tensor<fp16, []> obj_23_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_23_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 369 | tensor<fp16, [1, 1280, 1, 1]> obj_23_cast_fp16 = batch_norm(beta = obj_23_beta_0_to_fp16, epsilon = obj_23_epsilon_0_to_fp16, gamma = obj_23_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor<string, []>("obj_23_cast_fp16")]; |
| 370 | tensor<string, []> var_545_pad_type_0 = const()[name = tensor<string, []>("op_545_pad_type_0"), val = tensor<string, []>("valid")]; |
| 371 | tensor<int32, [2]> var_545_strides_0 = const()[name = tensor<string, []>("op_545_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 372 | tensor<int32, [4]> var_545_pad_0 = const()[name = tensor<string, []>("op_545_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 373 | tensor<int32, [2]> var_545_dilations_0 = const()[name = tensor<string, []>("op_545_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 374 | tensor<int32, []> var_545_groups_0 = const()[name = tensor<string, []>("op_545_groups_0"), val = tensor<int32, []>(1)]; |
| 375 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155318528))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156137792))), name = tensor<string, []>("layers_1_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 376 | tensor<fp16, [1280]> layers_1_encoder_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156137920)))]; |
| 377 | tensor<fp16, [1, 1280, 1, 1]> var_545_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_545_dilations_0, groups = var_545_groups_0, pad = var_545_pad_0, pad_type = var_545_pad_type_0, strides = var_545_strides_0, weight = layers_1_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_23_cast_fp16)[name = tensor<string, []>("op_545_cast_fp16")]; |
| 378 | tensor<string, []> var_551_pad_type_0 = const()[name = tensor<string, []>("op_551_pad_type_0"), val = tensor<string, []>("valid")]; |
| 379 | tensor<int32, [2]> var_551_strides_0 = const()[name = tensor<string, []>("op_551_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 380 | tensor<int32, [4]> var_551_pad_0 = const()[name = tensor<string, []>("op_551_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 381 | tensor<int32, [2]> var_551_dilations_0 = const()[name = tensor<string, []>("op_551_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 382 | tensor<int32, []> var_551_groups_0 = const()[name = tensor<string, []>("op_551_groups_0"), val = tensor<int32, []>(1)]; |
| 383 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156183616))), name = tensor<string, []>("layers_1_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [21483]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156140544))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 384 | tensor<fp16, [1, 1280, 1, 1]> var_551_cast_fp16 = conv(dilations = var_551_dilations_0, groups = var_551_groups_0, pad = var_551_pad_0, pad_type = var_551_pad_type_0, strides = var_551_strides_0, weight = layers_1_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_23_cast_fp16)[name = tensor<string, []>("op_551_cast_fp16")]; |
| 385 | tensor<fp16, [1, 1280, 1, 1]> query_7_cast_fp16 = add(x = var_545_cast_fp16, y = var_551_cast_fp16)[name = tensor<string, []>("query_7_cast_fp16")]; |
| 386 | tensor<string, []> var_560_pad_type_0 = const()[name = tensor<string, []>("op_560_pad_type_0"), val = tensor<string, []>("valid")]; |
| 387 | tensor<int32, [2]> var_560_strides_0 = const()[name = tensor<string, []>("op_560_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 388 | tensor<int32, [4]> var_560_pad_0 = const()[name = tensor<string, []>("op_560_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 389 | tensor<int32, [2]> var_560_dilations_0 = const()[name = tensor<string, []>("op_560_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 390 | tensor<int32, []> var_560_groups_0 = const()[name = tensor<string, []>("op_560_groups_0"), val = tensor<int32, []>(1)]; |
| 391 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156388480))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157207744))), name = tensor<string, []>("layers_1_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 392 | tensor<fp16, [1, 1280, 1, 1500]> var_560_cast_fp16 = conv(dilations = var_560_dilations_0, groups = var_560_groups_0, pad = var_560_pad_0, pad_type = var_560_pad_type_0, strides = var_560_strides_0, weight = layers_1_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("op_560_cast_fp16")]; |
| 393 | tensor<string, []> var_566_pad_type_0 = const()[name = tensor<string, []>("op_566_pad_type_0"), val = tensor<string, []>("valid")]; |
| 394 | tensor<int32, [2]> var_566_strides_0 = const()[name = tensor<string, []>("op_566_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 395 | tensor<int32, [4]> var_566_pad_0 = const()[name = tensor<string, []>("op_566_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 396 | tensor<int32, [2]> var_566_dilations_0 = const()[name = tensor<string, []>("op_566_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 397 | tensor<int32, []> var_566_groups_0 = const()[name = tensor<string, []>("op_566_groups_0"), val = tensor<int32, []>(1)]; |
| 398 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157251328))), name = tensor<string, []>("layers_1_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [21667]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157207872))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 399 | tensor<fp16, [1, 1280, 1, 1500]> var_566_cast_fp16 = conv(dilations = var_566_dilations_0, groups = var_566_groups_0, pad = var_566_pad_0, pad_type = var_566_pad_type_0, strides = var_566_strides_0, weight = layers_1_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = tensor<string, []>("op_566_cast_fp16")]; |
| 400 | tensor<fp16, [1, 1280, 1, 1500]> key_7_cast_fp16 = add(x = var_560_cast_fp16, y = var_566_cast_fp16)[name = tensor<string, []>("key_7_cast_fp16")]; |
| 401 | tensor<string, []> var_576_pad_type_0 = const()[name = tensor<string, []>("op_576_pad_type_0"), val = tensor<string, []>("valid")]; |
| 402 | tensor<int32, [2]> var_576_strides_0 = const()[name = tensor<string, []>("op_576_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 403 | tensor<int32, [4]> var_576_pad_0 = const()[name = tensor<string, []>("op_576_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 404 | tensor<int32, [2]> var_576_dilations_0 = const()[name = tensor<string, []>("op_576_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 405 | tensor<int32, []> var_576_groups_0 = const()[name = tensor<string, []>("op_576_groups_0"), val = tensor<int32, []>(1)]; |
| 406 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157456192))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158275456))), name = tensor<string, []>("layers_1_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 407 | tensor<fp16, [1280]> layers_1_encoder_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158275584)))]; |
| 408 | tensor<fp16, [1, 1280, 1, 1500]> var_576_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_576_dilations_0, groups = var_576_groups_0, pad = var_576_pad_0, pad_type = var_576_pad_type_0, strides = var_576_strides_0, weight = layers_1_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("op_576_cast_fp16")]; |
| 409 | tensor<string, []> var_582_pad_type_0 = const()[name = tensor<string, []>("op_582_pad_type_0"), val = tensor<string, []>("valid")]; |
| 410 | tensor<int32, [2]> var_582_strides_0 = const()[name = tensor<string, []>("op_582_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 411 | tensor<int32, [4]> var_582_pad_0 = const()[name = tensor<string, []>("op_582_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 412 | tensor<int32, [2]> var_582_dilations_0 = const()[name = tensor<string, []>("op_582_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 413 | tensor<int32, []> var_582_groups_0 = const()[name = tensor<string, []>("op_582_groups_0"), val = tensor<int32, []>(1)]; |
| 414 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158289408))), name = tensor<string, []>("layers_1_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [5557]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158278208))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 415 | tensor<fp16, [1, 1280, 1, 1500]> var_582_cast_fp16 = conv(dilations = var_582_dilations_0, groups = var_582_groups_0, pad = var_582_pad_0, pad_type = var_582_pad_type_0, strides = var_582_strides_0, weight = layers_1_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = tensor<string, []>("op_582_cast_fp16")]; |
| 416 | tensor<fp16, [1, 1280, 1, 1500]> value_7_cast_fp16 = add(x = var_576_cast_fp16, y = var_582_cast_fp16)[name = tensor<string, []>("value_7_cast_fp16")]; |
| 417 | tensor<int32, [4]> var_585 = const()[name = tensor<string, []>("op_585"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 418 | tensor<fp16, [1, 20, 64, 1]> mh_q_7_cast_fp16 = reshape(shape = var_585, x = query_7_cast_fp16)[name = tensor<string, []>("mh_q_7_cast_fp16")]; |
| 419 | tensor<fp16, []> var_587_to_fp16 = const()[name = tensor<string, []>("op_587_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| 420 | tensor<fp16, [1, 20, 64, 1]> var_588_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_587_to_fp16)[name = tensor<string, []>("op_588_cast_fp16")]; |
| 421 | tensor<int32, [4]> var_589 = const()[name = tensor<string, []>("op_589"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 422 | tensor<fp16, [1, 20, 64, 1500]> var_590_cast_fp16 = reshape(shape = var_589, x = key_7_cast_fp16)[name = tensor<string, []>("op_590_cast_fp16")]; |
| 423 | tensor<bool, []> mh_w_11_transpose_x_0 = const()[name = tensor<string, []>("mh_w_11_transpose_x_0"), val = tensor<bool, []>(true)]; |
| 424 | tensor<bool, []> mh_w_11_transpose_y_0 = const()[name = tensor<string, []>("mh_w_11_transpose_y_0"), val = tensor<bool, []>(false)]; |
| 425 | tensor<fp16, [1, 20, 1, 1500]> mh_w_11_cast_fp16 = matmul(transpose_x = mh_w_11_transpose_x_0, transpose_y = mh_w_11_transpose_y_0, x = var_588_cast_fp16, y = var_590_cast_fp16)[name = tensor<string, []>("mh_w_11_cast_fp16")]; |
| 426 | tensor<fp16, [1, 20, 1, 1500]> obj_27_cast_fp16 = softmax(axis = var_374, x = mh_w_11_cast_fp16)[name = tensor<string, []>("obj_27_cast_fp16")]; |
| 427 | tensor<int32, [4]> var_594 = const()[name = tensor<string, []>("op_594"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 428 | tensor<fp16, [1, 20, 64, 1500]> var_595_cast_fp16 = reshape(shape = var_594, x = value_7_cast_fp16)[name = tensor<string, []>("op_595_cast_fp16")]; |
| 429 | tensor<bool, []> attn_7_transpose_x_0 = const()[name = tensor<string, []>("attn_7_transpose_x_0"), val = tensor<bool, []>(false)]; |
| 430 | tensor<bool, []> attn_7_transpose_y_0 = const()[name = tensor<string, []>("attn_7_transpose_y_0"), val = tensor<bool, []>(true)]; |
| 431 | tensor<fp16, [1, 20, 64, 1]> attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_595_cast_fp16, y = obj_27_cast_fp16)[name = tensor<string, []>("attn_7_cast_fp16")]; |
| 432 | tensor<int32, [4]> var_598 = const()[name = tensor<string, []>("op_598"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; |
| 433 | tensor<fp16, [1, 1280, 1, 1]> input_13_cast_fp16 = reshape(shape = var_598, x = attn_7_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")]; |
| 434 | tensor<string, []> var_608_pad_type_0 = const()[name = tensor<string, []>("op_608_pad_type_0"), val = tensor<string, []>("valid")]; |
| 435 | tensor<int32, [2]> var_608_strides_0 = const()[name = tensor<string, []>("op_608_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 436 | tensor<int32, [4]> var_608_pad_0 = const()[name = tensor<string, []>("op_608_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 437 | tensor<int32, [2]> var_608_dilations_0 = const()[name = tensor<string, []>("op_608_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 438 | tensor<int32, []> var_608_groups_0 = const()[name = tensor<string, []>("op_608_groups_0"), val = tensor<int32, []>(1)]; |
| 439 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158494272))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159313536))), name = tensor<string, []>("layers_1_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 440 | tensor<fp16, [1280]> layers_1_encoder_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159313664)))]; |
| 441 | tensor<fp16, [1, 1280, 1, 1]> var_608_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_608_dilations_0, groups = var_608_groups_0, pad = var_608_pad_0, pad_type = var_608_pad_type_0, strides = var_608_strides_0, weight = layers_1_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_13_cast_fp16)[name = tensor<string, []>("op_608_cast_fp16")]; |
| 442 | tensor<string, []> var_614_pad_type_0 = const()[name = tensor<string, []>("op_614_pad_type_0"), val = tensor<string, []>("valid")]; |
| 443 | tensor<int32, [2]> var_614_strides_0 = const()[name = tensor<string, []>("op_614_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 444 | tensor<int32, [4]> var_614_pad_0 = const()[name = tensor<string, []>("op_614_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 445 | tensor<int32, [2]> var_614_dilations_0 = const()[name = tensor<string, []>("op_614_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 446 | tensor<int32, []> var_614_groups_0 = const()[name = tensor<string, []>("op_614_groups_0"), val = tensor<int32, []>(1)]; |
| 447 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159326656))), name = tensor<string, []>("layers_1_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [5143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159316288))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 448 | tensor<fp16, [1, 1280, 1, 1]> var_614_cast_fp16 = conv(dilations = var_614_dilations_0, groups = var_614_groups_0, pad = var_614_pad_0, pad_type = var_614_pad_type_0, strides = var_614_strides_0, weight = layers_1_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_13_cast_fp16)[name = tensor<string, []>("op_614_cast_fp16")]; |
| 449 | tensor<fp16, [1, 1280, 1, 1]> obj_25_cast_fp16 = add(x = var_608_cast_fp16, y = var_614_cast_fp16)[name = tensor<string, []>("obj_25_cast_fp16")]; |
| 450 | tensor<fp16, [1, 1280, 1, 1]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_25_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")]; |
| 451 | tensor<int32, [1]> out_11_axes_0 = const()[name = tensor<string, []>("out_11_axes_0"), val = tensor<int32, [1]>([1])]; |
| 452 | tensor<fp16, []> var_625_to_fp16 = const()[name = tensor<string, []>("op_625_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 453 | tensor<fp16, [1, 1280, 1, 1]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_625_to_fp16, x = inputs_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")]; |
| 454 | tensor<fp16, [1280]> input_15_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_15_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159531520)))]; |
| 455 | tensor<fp16, [1280]> input_15_beta_0_to_fp16 = const()[name = tensor<string, []>("input_15_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159534144)))]; |
| 456 | tensor<fp16, []> input_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 457 | tensor<fp16, [1, 1280, 1, 1]> input_15_cast_fp16 = batch_norm(beta = input_15_beta_0_to_fp16, epsilon = input_15_epsilon_0_to_fp16, gamma = input_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")]; |
| 458 | tensor<string, []> var_643_pad_type_0 = const()[name = tensor<string, []>("op_643_pad_type_0"), val = tensor<string, []>("valid")]; |
| 459 | tensor<int32, [2]> var_643_strides_0 = const()[name = tensor<string, []>("op_643_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 460 | tensor<int32, [4]> var_643_pad_0 = const()[name = tensor<string, []>("op_643_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 461 | tensor<int32, [2]> var_643_dilations_0 = const()[name = tensor<string, []>("op_643_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 462 | tensor<int32, []> var_643_groups_0 = const()[name = tensor<string, []>("op_643_groups_0"), val = tensor<int32, []>(1)]; |
| 463 | tensor<fp16, [5120, 1280, 1, 1]> layers_1_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159536768))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162813632))), name = tensor<string, []>("layers_1_fc1_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])]; |
| 464 | tensor<fp16, [5120]> layers_1_fc1_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162813760)))]; |
| 465 | tensor<fp16, [1, 5120, 1, 1]> var_643_cast_fp16 = conv(bias = layers_1_fc1_inlier_module_bias_to_fp16, dilations = var_643_dilations_0, groups = var_643_groups_0, pad = var_643_pad_0, pad_type = var_643_pad_type_0, strides = var_643_strides_0, weight = layers_1_fc1_inlier_module_weight_to_fp16_palettized, x = input_15_cast_fp16)[name = tensor<string, []>("op_643_cast_fp16")]; |
| 466 | tensor<string, []> var_649_pad_type_0 = const()[name = tensor<string, []>("op_649_pad_type_0"), val = tensor<string, []>("valid")]; |
| 467 | tensor<int32, [2]> var_649_strides_0 = const()[name = tensor<string, []>("op_649_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 468 | tensor<int32, [4]> var_649_pad_0 = const()[name = tensor<string, []>("op_649_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 469 | tensor<int32, [2]> var_649_dilations_0 = const()[name = tensor<string, []>("op_649_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 470 | tensor<int32, []> var_649_groups_0 = const()[name = tensor<string, []>("op_649_groups_0"), val = tensor<int32, []>(1)]; |
| 471 | tensor<fp16, [5120, 1280, 1, 1]> layers_1_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162909312))), name = tensor<string, []>("layers_1_fc1_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [42562]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162824064))), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])]; |
| 472 | tensor<fp16, [1, 5120, 1, 1]> var_649_cast_fp16 = conv(dilations = var_649_dilations_0, groups = var_649_groups_0, pad = var_649_pad_0, pad_type = var_649_pad_type_0, strides = var_649_strides_0, weight = layers_1_fc1_outlier_module_weight_to_fp16_sparsified, x = input_15_cast_fp16)[name = tensor<string, []>("op_649_cast_fp16")]; |
| 473 | tensor<fp16, [1, 5120, 1, 1]> input_17_cast_fp16 = add(x = var_643_cast_fp16, y = var_649_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")]; |
| 474 | tensor<string, []> input_19_mode_0 = const()[name = tensor<string, []>("input_19_mode_0"), val = tensor<string, []>("EXACT")]; |
| 475 | tensor<fp16, [1, 5120, 1, 1]> input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")]; |
| 476 | tensor<string, []> var_660_pad_type_0 = const()[name = tensor<string, []>("op_660_pad_type_0"), val = tensor<string, []>("valid")]; |
| 477 | tensor<int32, [2]> var_660_strides_0 = const()[name = tensor<string, []>("op_660_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 478 | tensor<int32, [4]> var_660_pad_0 = const()[name = tensor<string, []>("op_660_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 479 | tensor<int32, [2]> var_660_dilations_0 = const()[name = tensor<string, []>("op_660_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 480 | tensor<int32, []> var_660_groups_0 = const()[name = tensor<string, []>("op_660_groups_0"), val = tensor<int32, []>(1)]; |
| 481 | tensor<fp16, [1280, 5120, 1, 1]> layers_1_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163728576))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167005440))), name = tensor<string, []>("layers_1_fc2_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])]; |
| 482 | tensor<fp16, [1280]> layers_1_fc2_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167005568)))]; |
| 483 | tensor<fp16, [1, 1280, 1, 1]> var_660_cast_fp16 = conv(bias = layers_1_fc2_inlier_module_bias_to_fp16, dilations = var_660_dilations_0, groups = var_660_groups_0, pad = var_660_pad_0, pad_type = var_660_pad_type_0, strides = var_660_strides_0, weight = layers_1_fc2_inlier_module_weight_to_fp16_palettized, x = input_19_cast_fp16)[name = tensor<string, []>("op_660_cast_fp16")]; |
| 484 | tensor<string, []> var_666_pad_type_0 = const()[name = tensor<string, []>("op_666_pad_type_0"), val = tensor<string, []>("valid")]; |
| 485 | tensor<int32, [2]> var_666_strides_0 = const()[name = tensor<string, []>("op_666_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 486 | tensor<int32, [4]> var_666_pad_0 = const()[name = tensor<string, []>("op_666_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 487 | tensor<int32, [2]> var_666_dilations_0 = const()[name = tensor<string, []>("op_666_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 488 | tensor<int32, []> var_666_groups_0 = const()[name = tensor<string, []>("op_666_groups_0"), val = tensor<int32, []>(1)]; |
| 489 | tensor<fp16, [1280, 5120, 1, 1]> layers_1_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167096192))), name = tensor<string, []>("layers_1_fc2_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [43939]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167008192))), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])]; |
| 490 | tensor<fp16, [1, 1280, 1, 1]> var_666_cast_fp16 = conv(dilations = var_666_dilations_0, groups = var_666_groups_0, pad = var_666_pad_0, pad_type = var_666_pad_type_0, strides = var_666_strides_0, weight = layers_1_fc2_outlier_module_weight_to_fp16_sparsified, x = input_19_cast_fp16)[name = tensor<string, []>("op_666_cast_fp16")]; |
| 491 | tensor<fp16, [1, 1280, 1, 1]> hidden_states_5_cast_fp16 = add(x = var_660_cast_fp16, y = var_666_cast_fp16)[name = tensor<string, []>("hidden_states_5_cast_fp16")]; |
| 492 | tensor<fp16, [1, 1280, 1, 1]> inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor<string, []>("inputs_13_cast_fp16")]; |
| 493 | tensor<int32, []> var_678 = const()[name = tensor<string, []>("op_678"), val = tensor<int32, []>(3)]; |
| 494 | tensor<int32, [1]> out_13_axes_0 = const()[name = tensor<string, []>("out_13_axes_0"), val = tensor<int32, [1]>([1])]; |
| 495 | tensor<fp16, []> var_704_to_fp16 = const()[name = tensor<string, []>("op_704_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 496 | tensor<fp16, [1, 1280, 1, 1]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_704_to_fp16, x = inputs_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")]; |
| 497 | tensor<fp16, [1280]> obj_29_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_29_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167915456)))]; |
| 498 | tensor<fp16, [1280]> obj_29_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_29_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167918080)))]; |
| 499 | tensor<fp16, []> obj_29_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_29_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 500 | tensor<fp16, [1, 1280, 1, 1]> obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor<string, []>("obj_29_cast_fp16")]; |
| 501 | tensor<string, []> var_726_pad_type_0 = const()[name = tensor<string, []>("op_726_pad_type_0"), val = tensor<string, []>("valid")]; |
| 502 | tensor<int32, [2]> var_726_strides_0 = const()[name = tensor<string, []>("op_726_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 503 | tensor<int32, [4]> var_726_pad_0 = const()[name = tensor<string, []>("op_726_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 504 | tensor<int32, [2]> var_726_dilations_0 = const()[name = tensor<string, []>("op_726_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 505 | tensor<int32, []> var_726_groups_0 = const()[name = tensor<string, []>("op_726_groups_0"), val = tensor<int32, []>(1)]; |
| 506 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167920704))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168739968))), name = tensor<string, []>("layers_2_self_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 507 | tensor<fp16, [1280]> layers_2_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168740096)))]; |
| 508 | tensor<fp16, [1, 1280, 1, 1]> var_726_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_726_dilations_0, groups = var_726_groups_0, pad = var_726_pad_0, pad_type = var_726_pad_type_0, strides = var_726_strides_0, weight = layers_2_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor<string, []>("op_726_cast_fp16")]; |
| 509 | tensor<string, []> var_732_pad_type_0 = const()[name = tensor<string, []>("op_732_pad_type_0"), val = tensor<string, []>("valid")]; |
| 510 | tensor<int32, [2]> var_732_strides_0 = const()[name = tensor<string, []>("op_732_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 511 | tensor<int32, [4]> var_732_pad_0 = const()[name = tensor<string, []>("op_732_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 512 | tensor<int32, [2]> var_732_dilations_0 = const()[name = tensor<string, []>("op_732_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 513 | tensor<int32, []> var_732_groups_0 = const()[name = tensor<string, []>("op_732_groups_0"), val = tensor<int32, []>(1)]; |
| 514 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168774976))), name = tensor<string, []>("layers_2_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [16094]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168742720))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 515 | tensor<fp16, [1, 1280, 1, 1]> var_732_cast_fp16 = conv(dilations = var_732_dilations_0, groups = var_732_groups_0, pad = var_732_pad_0, pad_type = var_732_pad_type_0, strides = var_732_strides_0, weight = layers_2_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_29_cast_fp16)[name = tensor<string, []>("op_732_cast_fp16")]; |
| 516 | tensor<fp16, [1, 1280, 1, 1]> query_9_cast_fp16 = add(x = var_726_cast_fp16, y = var_732_cast_fp16)[name = tensor<string, []>("query_9_cast_fp16")]; |
| 517 | tensor<string, []> var_741_pad_type_0 = const()[name = tensor<string, []>("op_741_pad_type_0"), val = tensor<string, []>("valid")]; |
| 518 | tensor<int32, [2]> var_741_strides_0 = const()[name = tensor<string, []>("op_741_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 519 | tensor<int32, [4]> var_741_pad_0 = const()[name = tensor<string, []>("op_741_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 520 | tensor<int32, [2]> var_741_dilations_0 = const()[name = tensor<string, []>("op_741_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 521 | tensor<int32, []> var_741_groups_0 = const()[name = tensor<string, []>("op_741_groups_0"), val = tensor<int32, []>(1)]; |
| 522 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168979840))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169799104))), name = tensor<string, []>("layers_2_self_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 523 | tensor<fp16, [1, 1280, 1, 1]> var_741_cast_fp16 = conv(dilations = var_741_dilations_0, groups = var_741_groups_0, pad = var_741_pad_0, pad_type = var_741_pad_type_0, strides = var_741_strides_0, weight = layers_2_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor<string, []>("op_741_cast_fp16")]; |
| 524 | tensor<string, []> var_747_pad_type_0 = const()[name = tensor<string, []>("op_747_pad_type_0"), val = tensor<string, []>("valid")]; |
| 525 | tensor<int32, [2]> var_747_strides_0 = const()[name = tensor<string, []>("op_747_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 526 | tensor<int32, [4]> var_747_pad_0 = const()[name = tensor<string, []>("op_747_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 527 | tensor<int32, [2]> var_747_dilations_0 = const()[name = tensor<string, []>("op_747_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 528 | tensor<int32, []> var_747_groups_0 = const()[name = tensor<string, []>("op_747_groups_0"), val = tensor<int32, []>(1)]; |
| 529 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169836736))), name = tensor<string, []>("layers_2_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [18690]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169799232))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 530 | tensor<fp16, [1, 1280, 1, 1]> var_747_cast_fp16 = conv(dilations = var_747_dilations_0, groups = var_747_groups_0, pad = var_747_pad_0, pad_type = var_747_pad_type_0, strides = var_747_strides_0, weight = layers_2_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_29_cast_fp16)[name = tensor<string, []>("op_747_cast_fp16")]; |
| 531 | tensor<fp16, [1, 1280, 1, 1]> current_key_5_cast_fp16 = add(x = var_741_cast_fp16, y = var_747_cast_fp16)[name = tensor<string, []>("current_key_5_cast_fp16")]; |
| 532 | tensor<string, []> var_757_pad_type_0 = const()[name = tensor<string, []>("op_757_pad_type_0"), val = tensor<string, []>("valid")]; |
| 533 | tensor<int32, [2]> var_757_strides_0 = const()[name = tensor<string, []>("op_757_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 534 | tensor<int32, [4]> var_757_pad_0 = const()[name = tensor<string, []>("op_757_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 535 | tensor<int32, [2]> var_757_dilations_0 = const()[name = tensor<string, []>("op_757_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 536 | tensor<int32, []> var_757_groups_0 = const()[name = tensor<string, []>("op_757_groups_0"), val = tensor<int32, []>(1)]; |
| 537 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170041600))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170860864))), name = tensor<string, []>("layers_2_self_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 538 | tensor<fp16, [1280]> layers_2_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170860992)))]; |
| 539 | tensor<fp16, [1, 1280, 1, 1]> var_757_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_757_dilations_0, groups = var_757_groups_0, pad = var_757_pad_0, pad_type = var_757_pad_type_0, strides = var_757_strides_0, weight = layers_2_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor<string, []>("op_757_cast_fp16")]; |
| 540 | tensor<string, []> var_763_pad_type_0 = const()[name = tensor<string, []>("op_763_pad_type_0"), val = tensor<string, []>("valid")]; |
| 541 | tensor<int32, [2]> var_763_strides_0 = const()[name = tensor<string, []>("op_763_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 542 | tensor<int32, [4]> var_763_pad_0 = const()[name = tensor<string, []>("op_763_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 543 | tensor<int32, [2]> var_763_dilations_0 = const()[name = tensor<string, []>("op_763_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 544 | tensor<int32, []> var_763_groups_0 = const()[name = tensor<string, []>("op_763_groups_0"), val = tensor<int32, []>(1)]; |
| 545 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170876544))), name = tensor<string, []>("layers_2_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [6431]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170863616))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 546 | tensor<fp16, [1, 1280, 1, 1]> var_763_cast_fp16 = conv(dilations = var_763_dilations_0, groups = var_763_groups_0, pad = var_763_pad_0, pad_type = var_763_pad_type_0, strides = var_763_strides_0, weight = layers_2_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_29_cast_fp16)[name = tensor<string, []>("op_763_cast_fp16")]; |
| 547 | tensor<fp16, [1, 1280, 1, 1]> current_value_5_cast_fp16 = add(x = var_757_cast_fp16, y = var_763_cast_fp16)[name = tensor<string, []>("current_value_5_cast_fp16")]; |
| 548 | tensor<fp16, [1, 1280, 1, 448]> var_769_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_159_cast_fp16)[name = tensor<string, []>("op_769_cast_fp16")]; |
| 549 | tensor<fp16, [1, 1280, 1, 448]> var_771_cast_fp16 = mul(x = var_53_cast_fp16_2, y = var_162_cast_fp16)[name = tensor<string, []>("op_771_cast_fp16")]; |
| 550 | tensor<fp16, [1, 1280, 1, 448]> key_9_cast_fp16 = add(x = var_769_cast_fp16, y = var_771_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")]; |
| 551 | tensor<fp16, [1, 1280, 1, 448]> var_773_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_159_cast_fp16)[name = tensor<string, []>("op_773_cast_fp16")]; |
| 552 | tensor<fp16, [1, 1280, 1, 448]> var_775_cast_fp16 = mul(x = var_60_cast_fp16_2, y = var_162_cast_fp16)[name = tensor<string, []>("op_775_cast_fp16")]; |
| 553 | tensor<fp16, [1, 1280, 1, 448]> value_9_cast_fp16 = add(x = var_773_cast_fp16, y = var_775_cast_fp16)[name = tensor<string, []>("value_9_cast_fp16")]; |
| 554 | tensor<int32, [4]> var_778 = const()[name = tensor<string, []>("op_778"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 555 | tensor<fp16, [1, 20, 64, 1]> mh_q_9_cast_fp16 = reshape(shape = var_778, x = query_9_cast_fp16)[name = tensor<string, []>("mh_q_9_cast_fp16")]; |
| 556 | tensor<fp16, []> var_780_to_fp16 = const()[name = tensor<string, []>("op_780_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| 557 | tensor<fp16, [1, 20, 64, 1]> var_781_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_780_to_fp16)[name = tensor<string, []>("op_781_cast_fp16")]; |
| 558 | tensor<int32, [4]> var_782 = const()[name = tensor<string, []>("op_782"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 559 | tensor<fp16, [1, 20, 64, 448]> var_783_cast_fp16 = reshape(shape = var_782, x = key_9_cast_fp16)[name = tensor<string, []>("op_783_cast_fp16")]; |
| 560 | tensor<bool, []> mh_w_13_transpose_x_0 = const()[name = tensor<string, []>("mh_w_13_transpose_x_0"), val = tensor<bool, []>(true)]; |
| 561 | tensor<bool, []> mh_w_13_transpose_y_0 = const()[name = tensor<string, []>("mh_w_13_transpose_y_0"), val = tensor<bool, []>(false)]; |
| 562 | tensor<fp16, [1, 20, 1, 448]> mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_781_cast_fp16, y = var_783_cast_fp16)[name = tensor<string, []>("mh_w_13_cast_fp16")]; |
| 563 | tensor<fp16, [1, 20, 1, 448]> mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_180_cast_fp16)[name = tensor<string, []>("mh_w_15_cast_fp16")]; |
| 564 | tensor<fp16, [1, 20, 1, 448]> var_791_cast_fp16 = softmax(axis = var_678, x = mh_w_15_cast_fp16)[name = tensor<string, []>("op_791_cast_fp16")]; |
| 565 | tensor<int32, [4]> var_792 = const()[name = tensor<string, []>("op_792"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 566 | tensor<fp16, [1, 20, 64, 448]> var_793_cast_fp16 = reshape(shape = var_792, x = value_9_cast_fp16)[name = tensor<string, []>("op_793_cast_fp16")]; |
| 567 | tensor<bool, []> attn_9_transpose_x_0 = const()[name = tensor<string, []>("attn_9_transpose_x_0"), val = tensor<bool, []>(false)]; |
| 568 | tensor<bool, []> attn_9_transpose_y_0 = const()[name = tensor<string, []>("attn_9_transpose_y_0"), val = tensor<bool, []>(true)]; |
| 569 | tensor<fp16, [1, 20, 64, 1]> attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_793_cast_fp16, y = var_791_cast_fp16)[name = tensor<string, []>("attn_9_cast_fp16")]; |
| 570 | tensor<int32, [4]> var_796 = const()[name = tensor<string, []>("op_796"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; |
| 571 | tensor<fp16, [1, 1280, 1, 1]> input_21_cast_fp16 = reshape(shape = var_796, x = attn_9_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")]; |
| 572 | tensor<string, []> var_806_pad_type_0 = const()[name = tensor<string, []>("op_806_pad_type_0"), val = tensor<string, []>("valid")]; |
| 573 | tensor<int32, [2]> var_806_strides_0 = const()[name = tensor<string, []>("op_806_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 574 | tensor<int32, [4]> var_806_pad_0 = const()[name = tensor<string, []>("op_806_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 575 | tensor<int32, [2]> var_806_dilations_0 = const()[name = tensor<string, []>("op_806_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 576 | tensor<int32, []> var_806_groups_0 = const()[name = tensor<string, []>("op_806_groups_0"), val = tensor<int32, []>(1)]; |
| 577 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171081408))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171900672))), name = tensor<string, []>("layers_2_self_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 578 | tensor<fp16, [1280]> layers_2_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171900800)))]; |
| 579 | tensor<fp16, [1, 1280, 1, 1]> var_806_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_806_dilations_0, groups = var_806_groups_0, pad = var_806_pad_0, pad_type = var_806_pad_type_0, strides = var_806_strides_0, weight = layers_2_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_21_cast_fp16)[name = tensor<string, []>("op_806_cast_fp16")]; |
| 580 | tensor<string, []> var_812_pad_type_0 = const()[name = tensor<string, []>("op_812_pad_type_0"), val = tensor<string, []>("valid")]; |
| 581 | tensor<int32, [2]> var_812_strides_0 = const()[name = tensor<string, []>("op_812_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 582 | tensor<int32, [4]> var_812_pad_0 = const()[name = tensor<string, []>("op_812_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 583 | tensor<int32, [2]> var_812_dilations_0 = const()[name = tensor<string, []>("op_812_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 584 | tensor<int32, []> var_812_groups_0 = const()[name = tensor<string, []>("op_812_groups_0"), val = tensor<int32, []>(1)]; |
| 585 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171914880))), name = tensor<string, []>("layers_2_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [5678]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171903424))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 586 | tensor<fp16, [1, 1280, 1, 1]> var_812_cast_fp16 = conv(dilations = var_812_dilations_0, groups = var_812_groups_0, pad = var_812_pad_0, pad_type = var_812_pad_type_0, strides = var_812_strides_0, weight = layers_2_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_21_cast_fp16)[name = tensor<string, []>("op_812_cast_fp16")]; |
| 587 | tensor<fp16, [1, 1280, 1, 1]> obj_35_cast_fp16 = add(x = var_806_cast_fp16, y = var_812_cast_fp16)[name = tensor<string, []>("obj_35_cast_fp16")]; |
| 588 | tensor<fp16, [1, 1280, 1, 1]> inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_35_cast_fp16)[name = tensor<string, []>("inputs_15_cast_fp16")]; |
| 589 | tensor<int32, [1]> out_15_axes_0 = const()[name = tensor<string, []>("out_15_axes_0"), val = tensor<int32, [1]>([1])]; |
| 590 | tensor<fp16, []> var_827_to_fp16 = const()[name = tensor<string, []>("op_827_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 591 | tensor<fp16, [1, 1280, 1, 1]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_827_to_fp16, x = inputs_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")]; |
| 592 | tensor<fp16, [1280]> obj_37_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_37_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172119744)))]; |
| 593 | tensor<fp16, [1280]> obj_37_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_37_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172122368)))]; |
| 594 | tensor<fp16, []> obj_37_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_37_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 595 | tensor<fp16, [1, 1280, 1, 1]> obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor<string, []>("obj_37_cast_fp16")]; |
| 596 | tensor<string, []> var_849_pad_type_0 = const()[name = tensor<string, []>("op_849_pad_type_0"), val = tensor<string, []>("valid")]; |
| 597 | tensor<int32, [2]> var_849_strides_0 = const()[name = tensor<string, []>("op_849_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 598 | tensor<int32, [4]> var_849_pad_0 = const()[name = tensor<string, []>("op_849_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 599 | tensor<int32, [2]> var_849_dilations_0 = const()[name = tensor<string, []>("op_849_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 600 | tensor<int32, []> var_849_groups_0 = const()[name = tensor<string, []>("op_849_groups_0"), val = tensor<int32, []>(1)]; |
| 601 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172124992))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172944256))), name = tensor<string, []>("layers_2_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 602 | tensor<fp16, [1280]> layers_2_encoder_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172944384)))]; |
| 603 | tensor<fp16, [1, 1280, 1, 1]> var_849_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_849_dilations_0, groups = var_849_groups_0, pad = var_849_pad_0, pad_type = var_849_pad_type_0, strides = var_849_strides_0, weight = layers_2_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor<string, []>("op_849_cast_fp16")]; |
| 604 | tensor<string, []> var_855_pad_type_0 = const()[name = tensor<string, []>("op_855_pad_type_0"), val = tensor<string, []>("valid")]; |
| 605 | tensor<int32, [2]> var_855_strides_0 = const()[name = tensor<string, []>("op_855_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 606 | tensor<int32, [4]> var_855_pad_0 = const()[name = tensor<string, []>("op_855_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 607 | tensor<int32, [2]> var_855_dilations_0 = const()[name = tensor<string, []>("op_855_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 608 | tensor<int32, []> var_855_groups_0 = const()[name = tensor<string, []>("op_855_groups_0"), val = tensor<int32, []>(1)]; |
| 609 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172974720))), name = tensor<string, []>("layers_2_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [13824]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172947008))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 610 | tensor<fp16, [1, 1280, 1, 1]> var_855_cast_fp16 = conv(dilations = var_855_dilations_0, groups = var_855_groups_0, pad = var_855_pad_0, pad_type = var_855_pad_type_0, strides = var_855_strides_0, weight = layers_2_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_37_cast_fp16)[name = tensor<string, []>("op_855_cast_fp16")]; |
| 611 | tensor<fp16, [1, 1280, 1, 1]> query_11_cast_fp16 = add(x = var_849_cast_fp16, y = var_855_cast_fp16)[name = tensor<string, []>("query_11_cast_fp16")]; |
| 612 | tensor<string, []> var_864_pad_type_0 = const()[name = tensor<string, []>("op_864_pad_type_0"), val = tensor<string, []>("valid")]; |
| 613 | tensor<int32, [2]> var_864_strides_0 = const()[name = tensor<string, []>("op_864_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 614 | tensor<int32, [4]> var_864_pad_0 = const()[name = tensor<string, []>("op_864_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 615 | tensor<int32, [2]> var_864_dilations_0 = const()[name = tensor<string, []>("op_864_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 616 | tensor<int32, []> var_864_groups_0 = const()[name = tensor<string, []>("op_864_groups_0"), val = tensor<int32, []>(1)]; |
| 617 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(173179584))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(173998848))), name = tensor<string, []>("layers_2_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 618 | tensor<fp16, [1, 1280, 1, 1500]> var_864_cast_fp16 = conv(dilations = var_864_dilations_0, groups = var_864_groups_0, pad = var_864_pad_0, pad_type = var_864_pad_type_0, strides = var_864_strides_0, weight = layers_2_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("op_864_cast_fp16")]; |
| 619 | tensor<string, []> var_870_pad_type_0 = const()[name = tensor<string, []>("op_870_pad_type_0"), val = tensor<string, []>("valid")]; |
| 620 | tensor<int32, [2]> var_870_strides_0 = const()[name = tensor<string, []>("op_870_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 621 | tensor<int32, [4]> var_870_pad_0 = const()[name = tensor<string, []>("op_870_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 622 | tensor<int32, [2]> var_870_dilations_0 = const()[name = tensor<string, []>("op_870_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 623 | tensor<int32, []> var_870_groups_0 = const()[name = tensor<string, []>("op_870_groups_0"), val = tensor<int32, []>(1)]; |
| 624 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174026816))), name = tensor<string, []>("layers_2_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [13879]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(173998976))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 625 | tensor<fp16, [1, 1280, 1, 1500]> var_870_cast_fp16 = conv(dilations = var_870_dilations_0, groups = var_870_groups_0, pad = var_870_pad_0, pad_type = var_870_pad_type_0, strides = var_870_strides_0, weight = layers_2_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = tensor<string, []>("op_870_cast_fp16")]; |
| 626 | tensor<fp16, [1, 1280, 1, 1500]> key_11_cast_fp16 = add(x = var_864_cast_fp16, y = var_870_cast_fp16)[name = tensor<string, []>("key_11_cast_fp16")]; |
| 627 | tensor<string, []> var_880_pad_type_0 = const()[name = tensor<string, []>("op_880_pad_type_0"), val = tensor<string, []>("valid")]; |
| 628 | tensor<int32, [2]> var_880_strides_0 = const()[name = tensor<string, []>("op_880_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 629 | tensor<int32, [4]> var_880_pad_0 = const()[name = tensor<string, []>("op_880_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 630 | tensor<int32, [2]> var_880_dilations_0 = const()[name = tensor<string, []>("op_880_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 631 | tensor<int32, []> var_880_groups_0 = const()[name = tensor<string, []>("op_880_groups_0"), val = tensor<int32, []>(1)]; |
| 632 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174231680))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175050944))), name = tensor<string, []>("layers_2_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 633 | tensor<fp16, [1280]> layers_2_encoder_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175051072)))]; |
| 634 | tensor<fp16, [1, 1280, 1, 1500]> var_880_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_880_dilations_0, groups = var_880_groups_0, pad = var_880_pad_0, pad_type = var_880_pad_type_0, strides = var_880_strides_0, weight = layers_2_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("op_880_cast_fp16")]; |
| 635 | tensor<string, []> var_886_pad_type_0 = const()[name = tensor<string, []>("op_886_pad_type_0"), val = tensor<string, []>("valid")]; |
| 636 | tensor<int32, [2]> var_886_strides_0 = const()[name = tensor<string, []>("op_886_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 637 | tensor<int32, [4]> var_886_pad_0 = const()[name = tensor<string, []>("op_886_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 638 | tensor<int32, [2]> var_886_dilations_0 = const()[name = tensor<string, []>("op_886_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 639 | tensor<int32, []> var_886_groups_0 = const()[name = tensor<string, []>("op_886_groups_0"), val = tensor<int32, []>(1)]; |
| 640 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175065280))), name = tensor<string, []>("layers_2_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [5756]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175053696))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 641 | tensor<fp16, [1, 1280, 1, 1500]> var_886_cast_fp16 = conv(dilations = var_886_dilations_0, groups = var_886_groups_0, pad = var_886_pad_0, pad_type = var_886_pad_type_0, strides = var_886_strides_0, weight = layers_2_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = tensor<string, []>("op_886_cast_fp16")]; |
| 642 | tensor<fp16, [1, 1280, 1, 1500]> value_11_cast_fp16 = add(x = var_880_cast_fp16, y = var_886_cast_fp16)[name = tensor<string, []>("value_11_cast_fp16")]; |
| 643 | tensor<int32, [4]> var_889 = const()[name = tensor<string, []>("op_889"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 644 | tensor<fp16, [1, 20, 64, 1]> mh_q_11_cast_fp16 = reshape(shape = var_889, x = query_11_cast_fp16)[name = tensor<string, []>("mh_q_11_cast_fp16")]; |
| 645 | tensor<fp16, []> var_891_to_fp16 = const()[name = tensor<string, []>("op_891_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| 646 | tensor<fp16, [1, 20, 64, 1]> var_892_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_891_to_fp16)[name = tensor<string, []>("op_892_cast_fp16")]; |
| 647 | tensor<int32, [4]> var_893 = const()[name = tensor<string, []>("op_893"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 648 | tensor<fp16, [1, 20, 64, 1500]> var_894_cast_fp16 = reshape(shape = var_893, x = key_11_cast_fp16)[name = tensor<string, []>("op_894_cast_fp16")]; |
| 649 | tensor<bool, []> mh_w_17_transpose_x_0 = const()[name = tensor<string, []>("mh_w_17_transpose_x_0"), val = tensor<bool, []>(true)]; |
| 650 | tensor<bool, []> mh_w_17_transpose_y_0 = const()[name = tensor<string, []>("mh_w_17_transpose_y_0"), val = tensor<bool, []>(false)]; |
| 651 | tensor<fp16, [1, 20, 1, 1500]> mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_892_cast_fp16, y = var_894_cast_fp16)[name = tensor<string, []>("mh_w_17_cast_fp16")]; |
| 652 | tensor<fp16, [1, 20, 1, 1500]> obj_41_cast_fp16 = softmax(axis = var_678, x = mh_w_17_cast_fp16)[name = tensor<string, []>("obj_41_cast_fp16")]; |
| 653 | tensor<int32, [4]> var_898 = const()[name = tensor<string, []>("op_898"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 654 | tensor<fp16, [1, 20, 64, 1500]> var_899_cast_fp16 = reshape(shape = var_898, x = value_11_cast_fp16)[name = tensor<string, []>("op_899_cast_fp16")]; |
| 655 | tensor<bool, []> attn_11_transpose_x_0 = const()[name = tensor<string, []>("attn_11_transpose_x_0"), val = tensor<bool, []>(false)]; |
| 656 | tensor<bool, []> attn_11_transpose_y_0 = const()[name = tensor<string, []>("attn_11_transpose_y_0"), val = tensor<bool, []>(true)]; |
| 657 | tensor<fp16, [1, 20, 64, 1]> attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_899_cast_fp16, y = obj_41_cast_fp16)[name = tensor<string, []>("attn_11_cast_fp16")]; |
| 658 | tensor<int32, [4]> var_902 = const()[name = tensor<string, []>("op_902"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; |
| 659 | tensor<fp16, [1, 1280, 1, 1]> input_23_cast_fp16 = reshape(shape = var_902, x = attn_11_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")]; |
| 660 | tensor<string, []> var_912_pad_type_0 = const()[name = tensor<string, []>("op_912_pad_type_0"), val = tensor<string, []>("valid")]; |
| 661 | tensor<int32, [2]> var_912_strides_0 = const()[name = tensor<string, []>("op_912_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 662 | tensor<int32, [4]> var_912_pad_0 = const()[name = tensor<string, []>("op_912_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 663 | tensor<int32, [2]> var_912_dilations_0 = const()[name = tensor<string, []>("op_912_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 664 | tensor<int32, []> var_912_groups_0 = const()[name = tensor<string, []>("op_912_groups_0"), val = tensor<int32, []>(1)]; |
| 665 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175270144))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176089408))), name = tensor<string, []>("layers_2_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 666 | tensor<fp16, [1280]> layers_2_encoder_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176089536)))]; |
| 667 | tensor<fp16, [1, 1280, 1, 1]> var_912_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_912_dilations_0, groups = var_912_groups_0, pad = var_912_pad_0, pad_type = var_912_pad_type_0, strides = var_912_strides_0, weight = layers_2_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_23_cast_fp16)[name = tensor<string, []>("op_912_cast_fp16")]; |
| 668 | tensor<string, []> var_918_pad_type_0 = const()[name = tensor<string, []>("op_918_pad_type_0"), val = tensor<string, []>("valid")]; |
| 669 | tensor<int32, [2]> var_918_strides_0 = const()[name = tensor<string, []>("op_918_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 670 | tensor<int32, [4]> var_918_pad_0 = const()[name = tensor<string, []>("op_918_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 671 | tensor<int32, [2]> var_918_dilations_0 = const()[name = tensor<string, []>("op_918_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 672 | tensor<int32, []> var_918_groups_0 = const()[name = tensor<string, []>("op_918_groups_0"), val = tensor<int32, []>(1)]; |
| 673 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176105152))), name = tensor<string, []>("layers_2_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [6438]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176092160))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 674 | tensor<fp16, [1, 1280, 1, 1]> var_918_cast_fp16 = conv(dilations = var_918_dilations_0, groups = var_918_groups_0, pad = var_918_pad_0, pad_type = var_918_pad_type_0, strides = var_918_strides_0, weight = layers_2_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_23_cast_fp16)[name = tensor<string, []>("op_918_cast_fp16")]; |
| 675 | tensor<fp16, [1, 1280, 1, 1]> obj_39_cast_fp16 = add(x = var_912_cast_fp16, y = var_918_cast_fp16)[name = tensor<string, []>("obj_39_cast_fp16")]; |
| 676 | tensor<fp16, [1, 1280, 1, 1]> inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_39_cast_fp16)[name = tensor<string, []>("inputs_17_cast_fp16")]; |
| 677 | tensor<int32, [1]> out_17_axes_0 = const()[name = tensor<string, []>("out_17_axes_0"), val = tensor<int32, [1]>([1])]; |
| 678 | tensor<fp16, []> var_932_to_fp16 = const()[name = tensor<string, []>("op_932_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 679 | tensor<fp16, [1, 1280, 1, 1]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_932_to_fp16, x = inputs_17_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")]; |
| 680 | tensor<fp16, [1280]> input_25_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_25_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176310016)))]; |
| 681 | tensor<fp16, [1280]> input_25_beta_0_to_fp16 = const()[name = tensor<string, []>("input_25_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176312640)))]; |
| 682 | tensor<fp16, []> input_25_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_25_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 683 | tensor<fp16, [1, 1280, 1, 1]> input_25_cast_fp16 = batch_norm(beta = input_25_beta_0_to_fp16, epsilon = input_25_epsilon_0_to_fp16, gamma = input_25_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")]; |
| 684 | tensor<string, []> var_950_pad_type_0 = const()[name = tensor<string, []>("op_950_pad_type_0"), val = tensor<string, []>("valid")]; |
| 685 | tensor<int32, [2]> var_950_strides_0 = const()[name = tensor<string, []>("op_950_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 686 | tensor<int32, [4]> var_950_pad_0 = const()[name = tensor<string, []>("op_950_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 687 | tensor<int32, [2]> var_950_dilations_0 = const()[name = tensor<string, []>("op_950_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 688 | tensor<int32, []> var_950_groups_0 = const()[name = tensor<string, []>("op_950_groups_0"), val = tensor<int32, []>(1)]; |
| 689 | tensor<fp16, [5120, 1280, 1, 1]> layers_2_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176315264))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179592128))), name = tensor<string, []>("layers_2_fc1_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])]; |
| 690 | tensor<fp16, [5120]> layers_2_fc1_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179592256)))]; |
| 691 | tensor<fp16, [1, 5120, 1, 1]> var_950_cast_fp16 = conv(bias = layers_2_fc1_inlier_module_bias_to_fp16, dilations = var_950_dilations_0, groups = var_950_groups_0, pad = var_950_pad_0, pad_type = var_950_pad_type_0, strides = var_950_strides_0, weight = layers_2_fc1_inlier_module_weight_to_fp16_palettized, x = input_25_cast_fp16)[name = tensor<string, []>("op_950_cast_fp16")]; |
| 692 | tensor<string, []> var_956_pad_type_0 = const()[name = tensor<string, []>("op_956_pad_type_0"), val = tensor<string, []>("valid")]; |
| 693 | tensor<int32, [2]> var_956_strides_0 = const()[name = tensor<string, []>("op_956_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 694 | tensor<int32, [4]> var_956_pad_0 = const()[name = tensor<string, []>("op_956_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 695 | tensor<int32, [2]> var_956_dilations_0 = const()[name = tensor<string, []>("op_956_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 696 | tensor<int32, []> var_956_groups_0 = const()[name = tensor<string, []>("op_956_groups_0"), val = tensor<int32, []>(1)]; |
| 697 | tensor<fp16, [5120, 1280, 1, 1]> layers_2_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179764480))), name = tensor<string, []>("layers_2_fc1_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [80920]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179602560))), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])]; |
| 698 | tensor<fp16, [1, 5120, 1, 1]> var_956_cast_fp16 = conv(dilations = var_956_dilations_0, groups = var_956_groups_0, pad = var_956_pad_0, pad_type = var_956_pad_type_0, strides = var_956_strides_0, weight = layers_2_fc1_outlier_module_weight_to_fp16_sparsified, x = input_25_cast_fp16)[name = tensor<string, []>("op_956_cast_fp16")]; |
| 699 | tensor<fp16, [1, 5120, 1, 1]> input_27_cast_fp16 = add(x = var_950_cast_fp16, y = var_956_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")]; |
| 700 | tensor<string, []> input_29_mode_0 = const()[name = tensor<string, []>("input_29_mode_0"), val = tensor<string, []>("EXACT")]; |
| 701 | tensor<fp16, [1, 5120, 1, 1]> input_29_cast_fp16 = gelu(mode = input_29_mode_0, x = input_27_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")]; |
| 702 | tensor<string, []> var_967_pad_type_0 = const()[name = tensor<string, []>("op_967_pad_type_0"), val = tensor<string, []>("valid")]; |
| 703 | tensor<int32, [2]> var_967_strides_0 = const()[name = tensor<string, []>("op_967_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 704 | tensor<int32, [4]> var_967_pad_0 = const()[name = tensor<string, []>("op_967_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 705 | tensor<int32, [2]> var_967_dilations_0 = const()[name = tensor<string, []>("op_967_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 706 | tensor<int32, []> var_967_groups_0 = const()[name = tensor<string, []>("op_967_groups_0"), val = tensor<int32, []>(1)]; |
| 707 | tensor<fp16, [1280, 5120, 1, 1]> layers_2_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180583744))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183860608))), name = tensor<string, []>("layers_2_fc2_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])]; |
| 708 | tensor<fp16, [1280]> layers_2_fc2_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183860736)))]; |
| 709 | tensor<fp16, [1, 1280, 1, 1]> var_967_cast_fp16 = conv(bias = layers_2_fc2_inlier_module_bias_to_fp16, dilations = var_967_dilations_0, groups = var_967_groups_0, pad = var_967_pad_0, pad_type = var_967_pad_type_0, strides = var_967_strides_0, weight = layers_2_fc2_inlier_module_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = tensor<string, []>("op_967_cast_fp16")]; |
| 710 | tensor<string, []> var_973_pad_type_0 = const()[name = tensor<string, []>("op_973_pad_type_0"), val = tensor<string, []>("valid")]; |
| 711 | tensor<int32, [2]> var_973_strides_0 = const()[name = tensor<string, []>("op_973_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 712 | tensor<int32, [4]> var_973_pad_0 = const()[name = tensor<string, []>("op_973_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 713 | tensor<int32, [2]> var_973_dilations_0 = const()[name = tensor<string, []>("op_973_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 714 | tensor<int32, []> var_973_groups_0 = const()[name = tensor<string, []>("op_973_groups_0"), val = tensor<int32, []>(1)]; |
| 715 | tensor<fp16, [1280, 5120, 1, 1]> layers_2_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183943552))), name = tensor<string, []>("layers_2_fc2_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [40054]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183863360))), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])]; |
| 716 | tensor<fp16, [1, 1280, 1, 1]> var_973_cast_fp16 = conv(dilations = var_973_dilations_0, groups = var_973_groups_0, pad = var_973_pad_0, pad_type = var_973_pad_type_0, strides = var_973_strides_0, weight = layers_2_fc2_outlier_module_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = tensor<string, []>("op_973_cast_fp16")]; |
| 717 | tensor<fp16, [1, 1280, 1, 1]> hidden_states_7_cast_fp16 = add(x = var_967_cast_fp16, y = var_973_cast_fp16)[name = tensor<string, []>("hidden_states_7_cast_fp16")]; |
| 718 | tensor<fp16, [1, 1280, 1, 1]> inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor<string, []>("inputs_19_cast_fp16")]; |
| 719 | tensor<int32, []> var_986 = const()[name = tensor<string, []>("op_986"), val = tensor<int32, []>(3)]; |
| 720 | tensor<int32, [1]> out_19_axes_0 = const()[name = tensor<string, []>("out_19_axes_0"), val = tensor<int32, [1]>([1])]; |
| 721 | tensor<fp16, []> var_1012_to_fp16 = const()[name = tensor<string, []>("op_1012_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 722 | tensor<fp16, [1, 1280, 1, 1]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_1012_to_fp16, x = inputs_19_cast_fp16)[name = tensor<string, []>("out_19_cast_fp16")]; |
| 723 | tensor<fp16, [1280]> obj_43_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_43_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184762816)))]; |
| 724 | tensor<fp16, [1280]> obj_43_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_43_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184765440)))]; |
| 725 | tensor<fp16, []> obj_43_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_43_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 726 | tensor<fp16, [1, 1280, 1, 1]> obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor<string, []>("obj_43_cast_fp16")]; |
| 727 | tensor<string, []> var_1034_pad_type_0 = const()[name = tensor<string, []>("op_1034_pad_type_0"), val = tensor<string, []>("valid")]; |
| 728 | tensor<int32, [2]> var_1034_strides_0 = const()[name = tensor<string, []>("op_1034_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 729 | tensor<int32, [4]> var_1034_pad_0 = const()[name = tensor<string, []>("op_1034_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 730 | tensor<int32, [2]> var_1034_dilations_0 = const()[name = tensor<string, []>("op_1034_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 731 | tensor<int32, []> var_1034_groups_0 = const()[name = tensor<string, []>("op_1034_groups_0"), val = tensor<int32, []>(1)]; |
| 732 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184768064))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185587328))), name = tensor<string, []>("layers_3_self_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 733 | tensor<fp16, [1280]> layers_3_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185587456)))]; |
| 734 | tensor<fp16, [1, 1280, 1, 1]> var_1034_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_1034_dilations_0, groups = var_1034_groups_0, pad = var_1034_pad_0, pad_type = var_1034_pad_type_0, strides = var_1034_strides_0, weight = layers_3_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor<string, []>("op_1034_cast_fp16")]; |
| 735 | tensor<string, []> var_1040_pad_type_0 = const()[name = tensor<string, []>("op_1040_pad_type_0"), val = tensor<string, []>("valid")]; |
| 736 | tensor<int32, [2]> var_1040_strides_0 = const()[name = tensor<string, []>("op_1040_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 737 | tensor<int32, [4]> var_1040_pad_0 = const()[name = tensor<string, []>("op_1040_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 738 | tensor<int32, [2]> var_1040_dilations_0 = const()[name = tensor<string, []>("op_1040_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 739 | tensor<int32, []> var_1040_groups_0 = const()[name = tensor<string, []>("op_1040_groups_0"), val = tensor<int32, []>(1)]; |
| 740 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185611520))), name = tensor<string, []>("layers_3_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [10664]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185590080))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 741 | tensor<fp16, [1, 1280, 1, 1]> var_1040_cast_fp16 = conv(dilations = var_1040_dilations_0, groups = var_1040_groups_0, pad = var_1040_pad_0, pad_type = var_1040_pad_type_0, strides = var_1040_strides_0, weight = layers_3_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_43_cast_fp16)[name = tensor<string, []>("op_1040_cast_fp16")]; |
| 742 | tensor<fp16, [1, 1280, 1, 1]> query_13_cast_fp16 = add(x = var_1034_cast_fp16, y = var_1040_cast_fp16)[name = tensor<string, []>("query_13_cast_fp16")]; |
| 743 | tensor<string, []> var_1049_pad_type_0 = const()[name = tensor<string, []>("op_1049_pad_type_0"), val = tensor<string, []>("valid")]; |
| 744 | tensor<int32, [2]> var_1049_strides_0 = const()[name = tensor<string, []>("op_1049_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 745 | tensor<int32, [4]> var_1049_pad_0 = const()[name = tensor<string, []>("op_1049_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 746 | tensor<int32, [2]> var_1049_dilations_0 = const()[name = tensor<string, []>("op_1049_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 747 | tensor<int32, []> var_1049_groups_0 = const()[name = tensor<string, []>("op_1049_groups_0"), val = tensor<int32, []>(1)]; |
| 748 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185816384))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186635648))), name = tensor<string, []>("layers_3_self_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 749 | tensor<fp16, [1, 1280, 1, 1]> var_1049_cast_fp16 = conv(dilations = var_1049_dilations_0, groups = var_1049_groups_0, pad = var_1049_pad_0, pad_type = var_1049_pad_type_0, strides = var_1049_strides_0, weight = layers_3_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor<string, []>("op_1049_cast_fp16")]; |
| 750 | tensor<string, []> var_1055_pad_type_0 = const()[name = tensor<string, []>("op_1055_pad_type_0"), val = tensor<string, []>("valid")]; |
| 751 | tensor<int32, [2]> var_1055_strides_0 = const()[name = tensor<string, []>("op_1055_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 752 | tensor<int32, [4]> var_1055_pad_0 = const()[name = tensor<string, []>("op_1055_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 753 | tensor<int32, [2]> var_1055_dilations_0 = const()[name = tensor<string, []>("op_1055_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 754 | tensor<int32, []> var_1055_groups_0 = const()[name = tensor<string, []>("op_1055_groups_0"), val = tensor<int32, []>(1)]; |
| 755 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186656640))), name = tensor<string, []>("layers_3_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [10387]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186635776))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 756 | tensor<fp16, [1, 1280, 1, 1]> var_1055_cast_fp16 = conv(dilations = var_1055_dilations_0, groups = var_1055_groups_0, pad = var_1055_pad_0, pad_type = var_1055_pad_type_0, strides = var_1055_strides_0, weight = layers_3_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_43_cast_fp16)[name = tensor<string, []>("op_1055_cast_fp16")]; |
| 757 | tensor<fp16, [1, 1280, 1, 1]> current_key_cast_fp16 = add(x = var_1049_cast_fp16, y = var_1055_cast_fp16)[name = tensor<string, []>("current_key_cast_fp16")]; |
| 758 | tensor<string, []> var_1065_pad_type_0 = const()[name = tensor<string, []>("op_1065_pad_type_0"), val = tensor<string, []>("valid")]; |
| 759 | tensor<int32, [2]> var_1065_strides_0 = const()[name = tensor<string, []>("op_1065_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 760 | tensor<int32, [4]> var_1065_pad_0 = const()[name = tensor<string, []>("op_1065_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 761 | tensor<int32, [2]> var_1065_dilations_0 = const()[name = tensor<string, []>("op_1065_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 762 | tensor<int32, []> var_1065_groups_0 = const()[name = tensor<string, []>("op_1065_groups_0"), val = tensor<int32, []>(1)]; |
| 763 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186861504))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187680768))), name = tensor<string, []>("layers_3_self_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 764 | tensor<fp16, [1280]> layers_3_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187680896)))]; |
| 765 | tensor<fp16, [1, 1280, 1, 1]> var_1065_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_1065_dilations_0, groups = var_1065_groups_0, pad = var_1065_pad_0, pad_type = var_1065_pad_type_0, strides = var_1065_strides_0, weight = layers_3_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor<string, []>("op_1065_cast_fp16")]; |
| 766 | tensor<string, []> var_1071_pad_type_0 = const()[name = tensor<string, []>("op_1071_pad_type_0"), val = tensor<string, []>("valid")]; |
| 767 | tensor<int32, [2]> var_1071_strides_0 = const()[name = tensor<string, []>("op_1071_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 768 | tensor<int32, [4]> var_1071_pad_0 = const()[name = tensor<string, []>("op_1071_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 769 | tensor<int32, [2]> var_1071_dilations_0 = const()[name = tensor<string, []>("op_1071_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 770 | tensor<int32, []> var_1071_groups_0 = const()[name = tensor<string, []>("op_1071_groups_0"), val = tensor<int32, []>(1)]; |
| 771 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187698304))), name = tensor<string, []>("layers_3_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [7342]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187683520))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 772 | tensor<fp16, [1, 1280, 1, 1]> var_1071_cast_fp16 = conv(dilations = var_1071_dilations_0, groups = var_1071_groups_0, pad = var_1071_pad_0, pad_type = var_1071_pad_type_0, strides = var_1071_strides_0, weight = layers_3_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_43_cast_fp16)[name = tensor<string, []>("op_1071_cast_fp16")]; |
| 773 | tensor<fp16, [1, 1280, 1, 1]> current_value_cast_fp16 = add(x = var_1065_cast_fp16, y = var_1071_cast_fp16)[name = tensor<string, []>("current_value_cast_fp16")]; |
| 774 | tensor<fp16, [1, 1280, 1, 448]> var_1077_cast_fp16 = mul(x = current_key_cast_fp16, y = var_159_cast_fp16)[name = tensor<string, []>("op_1077_cast_fp16")]; |
| 775 | tensor<fp16, [1, 1280, 1, 448]> var_1079_cast_fp16 = mul(x = var_53_cast_fp16_3, y = var_162_cast_fp16)[name = tensor<string, []>("op_1079_cast_fp16")]; |
| 776 | tensor<fp16, [1, 1280, 1, 448]> key_13_cast_fp16 = add(x = var_1077_cast_fp16, y = var_1079_cast_fp16)[name = tensor<string, []>("key_13_cast_fp16")]; |
| 777 | tensor<fp16, [1, 1280, 1, 448]> var_1081_cast_fp16 = mul(x = current_value_cast_fp16, y = var_159_cast_fp16)[name = tensor<string, []>("op_1081_cast_fp16")]; |
| 778 | tensor<fp16, [1, 1280, 1, 448]> var_1083_cast_fp16 = mul(x = var_60_cast_fp16_3, y = var_162_cast_fp16)[name = tensor<string, []>("op_1083_cast_fp16")]; |
| 779 | tensor<fp16, [1, 1280, 1, 448]> value_13_cast_fp16 = add(x = var_1081_cast_fp16, y = var_1083_cast_fp16)[name = tensor<string, []>("value_13_cast_fp16")]; |
| 780 | tensor<int32, [4]> var_1086 = const()[name = tensor<string, []>("op_1086"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 781 | tensor<fp16, [1, 20, 64, 1]> mh_q_13_cast_fp16 = reshape(shape = var_1086, x = query_13_cast_fp16)[name = tensor<string, []>("mh_q_13_cast_fp16")]; |
| 782 | tensor<fp16, []> var_1088_to_fp16 = const()[name = tensor<string, []>("op_1088_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| 783 | tensor<fp16, [1, 20, 64, 1]> var_1089_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_1088_to_fp16)[name = tensor<string, []>("op_1089_cast_fp16")]; |
| 784 | tensor<int32, [4]> var_1090 = const()[name = tensor<string, []>("op_1090"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 785 | tensor<fp16, [1, 20, 64, 448]> var_1091_cast_fp16 = reshape(shape = var_1090, x = key_13_cast_fp16)[name = tensor<string, []>("op_1091_cast_fp16")]; |
| 786 | tensor<bool, []> mh_w_19_transpose_x_0 = const()[name = tensor<string, []>("mh_w_19_transpose_x_0"), val = tensor<bool, []>(true)]; |
| 787 | tensor<bool, []> mh_w_19_transpose_y_0 = const()[name = tensor<string, []>("mh_w_19_transpose_y_0"), val = tensor<bool, []>(false)]; |
| 788 | tensor<fp16, [1, 20, 1, 448]> mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_1089_cast_fp16, y = var_1091_cast_fp16)[name = tensor<string, []>("mh_w_19_cast_fp16")]; |
| 789 | tensor<fp16, [1, 20, 1, 448]> mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_180_cast_fp16)[name = tensor<string, []>("mh_w_21_cast_fp16")]; |
| 790 | tensor<fp16, [1, 20, 1, 448]> var_1099_cast_fp16 = softmax(axis = var_986, x = mh_w_21_cast_fp16)[name = tensor<string, []>("op_1099_cast_fp16")]; |
| 791 | tensor<int32, [4]> var_1100 = const()[name = tensor<string, []>("op_1100"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 792 | tensor<fp16, [1, 20, 64, 448]> var_1101_cast_fp16 = reshape(shape = var_1100, x = value_13_cast_fp16)[name = tensor<string, []>("op_1101_cast_fp16")]; |
| 793 | tensor<bool, []> attn_13_transpose_x_0 = const()[name = tensor<string, []>("attn_13_transpose_x_0"), val = tensor<bool, []>(false)]; |
| 794 | tensor<bool, []> attn_13_transpose_y_0 = const()[name = tensor<string, []>("attn_13_transpose_y_0"), val = tensor<bool, []>(true)]; |
| 795 | tensor<fp16, [1, 20, 64, 1]> attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_1101_cast_fp16, y = var_1099_cast_fp16)[name = tensor<string, []>("attn_13_cast_fp16")]; |
| 796 | tensor<int32, [4]> var_1104 = const()[name = tensor<string, []>("op_1104"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; |
| 797 | tensor<fp16, [1, 1280, 1, 1]> input_31_cast_fp16 = reshape(shape = var_1104, x = attn_13_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")]; |
| 798 | tensor<string, []> var_1114_pad_type_0 = const()[name = tensor<string, []>("op_1114_pad_type_0"), val = tensor<string, []>("valid")]; |
| 799 | tensor<int32, [2]> var_1114_strides_0 = const()[name = tensor<string, []>("op_1114_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 800 | tensor<int32, [4]> var_1114_pad_0 = const()[name = tensor<string, []>("op_1114_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 801 | tensor<int32, [2]> var_1114_dilations_0 = const()[name = tensor<string, []>("op_1114_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 802 | tensor<int32, []> var_1114_groups_0 = const()[name = tensor<string, []>("op_1114_groups_0"), val = tensor<int32, []>(1)]; |
| 803 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187903168))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188722432))), name = tensor<string, []>("layers_3_self_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 804 | tensor<fp16, [1280]> layers_3_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188722560)))]; |
| 805 | tensor<fp16, [1, 1280, 1, 1]> var_1114_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_1114_dilations_0, groups = var_1114_groups_0, pad = var_1114_pad_0, pad_type = var_1114_pad_type_0, strides = var_1114_strides_0, weight = layers_3_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_31_cast_fp16)[name = tensor<string, []>("op_1114_cast_fp16")]; |
| 806 | tensor<string, []> var_1120_pad_type_0 = const()[name = tensor<string, []>("op_1120_pad_type_0"), val = tensor<string, []>("valid")]; |
| 807 | tensor<int32, [2]> var_1120_strides_0 = const()[name = tensor<string, []>("op_1120_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 808 | tensor<int32, [4]> var_1120_pad_0 = const()[name = tensor<string, []>("op_1120_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 809 | tensor<int32, [2]> var_1120_dilations_0 = const()[name = tensor<string, []>("op_1120_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 810 | tensor<int32, []> var_1120_groups_0 = const()[name = tensor<string, []>("op_1120_groups_0"), val = tensor<int32, []>(1)]; |
| 811 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188739712))), name = tensor<string, []>("layers_3_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [7219]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188725184))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 812 | tensor<fp16, [1, 1280, 1, 1]> var_1120_cast_fp16 = conv(dilations = var_1120_dilations_0, groups = var_1120_groups_0, pad = var_1120_pad_0, pad_type = var_1120_pad_type_0, strides = var_1120_strides_0, weight = layers_3_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_31_cast_fp16)[name = tensor<string, []>("op_1120_cast_fp16")]; |
| 813 | tensor<fp16, [1, 1280, 1, 1]> obj_49_cast_fp16 = add(x = var_1114_cast_fp16, y = var_1120_cast_fp16)[name = tensor<string, []>("obj_49_cast_fp16")]; |
| 814 | tensor<fp16, [1, 1280, 1, 1]> inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_49_cast_fp16)[name = tensor<string, []>("inputs_21_cast_fp16")]; |
| 815 | tensor<int32, [1]> out_21_axes_0 = const()[name = tensor<string, []>("out_21_axes_0"), val = tensor<int32, [1]>([1])]; |
| 816 | tensor<fp16, []> var_1135_to_fp16 = const()[name = tensor<string, []>("op_1135_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 817 | tensor<fp16, [1, 1280, 1, 1]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_1135_to_fp16, x = inputs_21_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")]; |
| 818 | tensor<fp16, [1280]> obj_51_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_51_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188944576)))]; |
| 819 | tensor<fp16, [1280]> obj_51_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_51_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188947200)))]; |
| 820 | tensor<fp16, []> obj_51_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_51_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 821 | tensor<fp16, [1, 1280, 1, 1]> obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor<string, []>("obj_51_cast_fp16")]; |
| 822 | tensor<string, []> var_1157_pad_type_0 = const()[name = tensor<string, []>("op_1157_pad_type_0"), val = tensor<string, []>("valid")]; |
| 823 | tensor<int32, [2]> var_1157_strides_0 = const()[name = tensor<string, []>("op_1157_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 824 | tensor<int32, [4]> var_1157_pad_0 = const()[name = tensor<string, []>("op_1157_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 825 | tensor<int32, [2]> var_1157_dilations_0 = const()[name = tensor<string, []>("op_1157_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 826 | tensor<int32, []> var_1157_groups_0 = const()[name = tensor<string, []>("op_1157_groups_0"), val = tensor<int32, []>(1)]; |
| 827 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188949824))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189769088))), name = tensor<string, []>("layers_3_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 828 | tensor<fp16, [1280]> layers_3_encoder_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189769216)))]; |
| 829 | tensor<fp16, [1, 1280, 1, 1]> var_1157_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_1157_dilations_0, groups = var_1157_groups_0, pad = var_1157_pad_0, pad_type = var_1157_pad_type_0, strides = var_1157_strides_0, weight = layers_3_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_51_cast_fp16)[name = tensor<string, []>("op_1157_cast_fp16")]; |
| 830 | tensor<string, []> var_1163_pad_type_0 = const()[name = tensor<string, []>("op_1163_pad_type_0"), val = tensor<string, []>("valid")]; |
| 831 | tensor<int32, [2]> var_1163_strides_0 = const()[name = tensor<string, []>("op_1163_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 832 | tensor<int32, [4]> var_1163_pad_0 = const()[name = tensor<string, []>("op_1163_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 833 | tensor<int32, [2]> var_1163_dilations_0 = const()[name = tensor<string, []>("op_1163_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 834 | tensor<int32, []> var_1163_groups_0 = const()[name = tensor<string, []>("op_1163_groups_0"), val = tensor<int32, []>(1)]; |
| 835 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189787264))), name = tensor<string, []>("layers_3_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [7675]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189771840))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 836 | tensor<fp16, [1, 1280, 1, 1]> var_1163_cast_fp16 = conv(dilations = var_1163_dilations_0, groups = var_1163_groups_0, pad = var_1163_pad_0, pad_type = var_1163_pad_type_0, strides = var_1163_strides_0, weight = layers_3_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_51_cast_fp16)[name = tensor<string, []>("op_1163_cast_fp16")]; |
| 837 | tensor<fp16, [1, 1280, 1, 1]> query_cast_fp16 = add(x = var_1157_cast_fp16, y = var_1163_cast_fp16)[name = tensor<string, []>("query_cast_fp16")]; |
| 838 | tensor<string, []> var_1172_pad_type_0 = const()[name = tensor<string, []>("op_1172_pad_type_0"), val = tensor<string, []>("valid")]; |
| 839 | tensor<int32, [2]> var_1172_strides_0 = const()[name = tensor<string, []>("op_1172_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 840 | tensor<int32, [4]> var_1172_pad_0 = const()[name = tensor<string, []>("op_1172_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 841 | tensor<int32, [2]> var_1172_dilations_0 = const()[name = tensor<string, []>("op_1172_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 842 | tensor<int32, []> var_1172_groups_0 = const()[name = tensor<string, []>("op_1172_groups_0"), val = tensor<int32, []>(1)]; |
| 843 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189992128))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190811392))), name = tensor<string, []>("layers_3_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 844 | tensor<fp16, [1, 1280, 1, 1500]> var_1172_cast_fp16 = conv(dilations = var_1172_dilations_0, groups = var_1172_groups_0, pad = var_1172_pad_0, pad_type = var_1172_pad_type_0, strides = var_1172_strides_0, weight = layers_3_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("op_1172_cast_fp16")]; |
| 845 | tensor<string, []> var_1178_pad_type_0 = const()[name = tensor<string, []>("op_1178_pad_type_0"), val = tensor<string, []>("valid")]; |
| 846 | tensor<int32, [2]> var_1178_strides_0 = const()[name = tensor<string, []>("op_1178_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 847 | tensor<int32, [4]> var_1178_pad_0 = const()[name = tensor<string, []>("op_1178_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 848 | tensor<int32, [2]> var_1178_dilations_0 = const()[name = tensor<string, []>("op_1178_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 849 | tensor<int32, []> var_1178_groups_0 = const()[name = tensor<string, []>("op_1178_groups_0"), val = tensor<int32, []>(1)]; |
| 850 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190834240))), name = tensor<string, []>("layers_3_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [11308]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190811520))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 851 | tensor<fp16, [1, 1280, 1, 1500]> var_1178_cast_fp16 = conv(dilations = var_1178_dilations_0, groups = var_1178_groups_0, pad = var_1178_pad_0, pad_type = var_1178_pad_type_0, strides = var_1178_strides_0, weight = layers_3_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = tensor<string, []>("op_1178_cast_fp16")]; |
| 852 | tensor<fp16, [1, 1280, 1, 1500]> key_cast_fp16 = add(x = var_1172_cast_fp16, y = var_1178_cast_fp16)[name = tensor<string, []>("key_cast_fp16")]; |
| 853 | tensor<string, []> var_1188_pad_type_0 = const()[name = tensor<string, []>("op_1188_pad_type_0"), val = tensor<string, []>("valid")]; |
| 854 | tensor<int32, [2]> var_1188_strides_0 = const()[name = tensor<string, []>("op_1188_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 855 | tensor<int32, [4]> var_1188_pad_0 = const()[name = tensor<string, []>("op_1188_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 856 | tensor<int32, [2]> var_1188_dilations_0 = const()[name = tensor<string, []>("op_1188_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 857 | tensor<int32, []> var_1188_groups_0 = const()[name = tensor<string, []>("op_1188_groups_0"), val = tensor<int32, []>(1)]; |
| 858 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191039104))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191858368))), name = tensor<string, []>("layers_3_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 859 | tensor<fp16, [1280]> layers_3_encoder_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191858496)))]; |
| 860 | tensor<fp16, [1, 1280, 1, 1500]> var_1188_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_1188_dilations_0, groups = var_1188_groups_0, pad = var_1188_pad_0, pad_type = var_1188_pad_type_0, strides = var_1188_strides_0, weight = layers_3_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("op_1188_cast_fp16")]; |
| 861 | tensor<string, []> var_1194_pad_type_0 = const()[name = tensor<string, []>("op_1194_pad_type_0"), val = tensor<string, []>("valid")]; |
| 862 | tensor<int32, [2]> var_1194_strides_0 = const()[name = tensor<string, []>("op_1194_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 863 | tensor<int32, [4]> var_1194_pad_0 = const()[name = tensor<string, []>("op_1194_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 864 | tensor<int32, [2]> var_1194_dilations_0 = const()[name = tensor<string, []>("op_1194_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 865 | tensor<int32, []> var_1194_groups_0 = const()[name = tensor<string, []>("op_1194_groups_0"), val = tensor<int32, []>(1)]; |
| 866 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191874944))), name = tensor<string, []>("layers_3_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [6870]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191861120))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 867 | tensor<fp16, [1, 1280, 1, 1500]> var_1194_cast_fp16 = conv(dilations = var_1194_dilations_0, groups = var_1194_groups_0, pad = var_1194_pad_0, pad_type = var_1194_pad_type_0, strides = var_1194_strides_0, weight = layers_3_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = tensor<string, []>("op_1194_cast_fp16")]; |
| 868 | tensor<fp16, [1, 1280, 1, 1500]> value_cast_fp16 = add(x = var_1188_cast_fp16, y = var_1194_cast_fp16)[name = tensor<string, []>("value_cast_fp16")]; |
| 869 | tensor<int32, [4]> var_1197 = const()[name = tensor<string, []>("op_1197"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 870 | tensor<fp16, [1, 20, 64, 1]> mh_q_cast_fp16 = reshape(shape = var_1197, x = query_cast_fp16)[name = tensor<string, []>("mh_q_cast_fp16")]; |
| 871 | tensor<fp16, []> var_1199_to_fp16 = const()[name = tensor<string, []>("op_1199_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| 872 | tensor<fp16, [1, 20, 64, 1]> var_1200_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_1199_to_fp16)[name = tensor<string, []>("op_1200_cast_fp16")]; |
| 873 | tensor<int32, [4]> var_1201 = const()[name = tensor<string, []>("op_1201"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 874 | tensor<fp16, [1, 20, 64, 1500]> var_1202_cast_fp16 = reshape(shape = var_1201, x = key_cast_fp16)[name = tensor<string, []>("op_1202_cast_fp16")]; |
| 875 | tensor<bool, []> mh_w_transpose_x_0 = const()[name = tensor<string, []>("mh_w_transpose_x_0"), val = tensor<bool, []>(true)]; |
| 876 | tensor<bool, []> mh_w_transpose_y_0 = const()[name = tensor<string, []>("mh_w_transpose_y_0"), val = tensor<bool, []>(false)]; |
| 877 | tensor<fp16, [1, 20, 1, 1500]> mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_1200_cast_fp16, y = var_1202_cast_fp16)[name = tensor<string, []>("mh_w_cast_fp16")]; |
| 878 | tensor<fp16, [1, 20, 1, 1500]> obj_55_cast_fp16 = softmax(axis = var_986, x = mh_w_cast_fp16)[name = tensor<string, []>("obj_55_cast_fp16")]; |
| 879 | tensor<int32, [4]> var_1206 = const()[name = tensor<string, []>("op_1206"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 880 | tensor<fp16, [1, 20, 64, 1500]> var_1207_cast_fp16 = reshape(shape = var_1206, x = value_cast_fp16)[name = tensor<string, []>("op_1207_cast_fp16")]; |
| 881 | tensor<bool, []> attn_transpose_x_0 = const()[name = tensor<string, []>("attn_transpose_x_0"), val = tensor<bool, []>(false)]; |
| 882 | tensor<bool, []> attn_transpose_y_0 = const()[name = tensor<string, []>("attn_transpose_y_0"), val = tensor<bool, []>(true)]; |
| 883 | tensor<fp16, [1, 20, 64, 1]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_1207_cast_fp16, y = obj_55_cast_fp16)[name = tensor<string, []>("attn_cast_fp16")]; |
| 884 | tensor<int32, [4]> var_1210 = const()[name = tensor<string, []>("op_1210"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; |
| 885 | tensor<fp16, [1, 1280, 1, 1]> input_33_cast_fp16 = reshape(shape = var_1210, x = attn_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")]; |
| 886 | tensor<string, []> var_1220_pad_type_0 = const()[name = tensor<string, []>("op_1220_pad_type_0"), val = tensor<string, []>("valid")]; |
| 887 | tensor<int32, [2]> var_1220_strides_0 = const()[name = tensor<string, []>("op_1220_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 888 | tensor<int32, [4]> var_1220_pad_0 = const()[name = tensor<string, []>("op_1220_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 889 | tensor<int32, [2]> var_1220_dilations_0 = const()[name = tensor<string, []>("op_1220_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 890 | tensor<int32, []> var_1220_groups_0 = const()[name = tensor<string, []>("op_1220_groups_0"), val = tensor<int32, []>(1)]; |
| 891 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192079808))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192899072))), name = tensor<string, []>("layers_3_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 892 | tensor<fp16, [1280]> layers_3_encoder_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192899200)))]; |
| 893 | tensor<fp16, [1, 1280, 1, 1]> var_1220_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_1220_dilations_0, groups = var_1220_groups_0, pad = var_1220_pad_0, pad_type = var_1220_pad_type_0, strides = var_1220_strides_0, weight = layers_3_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_33_cast_fp16)[name = tensor<string, []>("op_1220_cast_fp16")]; |
| 894 | tensor<string, []> var_1226_pad_type_0 = const()[name = tensor<string, []>("op_1226_pad_type_0"), val = tensor<string, []>("valid")]; |
| 895 | tensor<int32, [2]> var_1226_strides_0 = const()[name = tensor<string, []>("op_1226_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 896 | tensor<int32, [4]> var_1226_pad_0 = const()[name = tensor<string, []>("op_1226_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 897 | tensor<int32, [2]> var_1226_dilations_0 = const()[name = tensor<string, []>("op_1226_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 898 | tensor<int32, []> var_1226_groups_0 = const()[name = tensor<string, []>("op_1226_groups_0"), val = tensor<int32, []>(1)]; |
| 899 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [204800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192913536))), name = tensor<string, []>("layers_3_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [5809]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192901824))), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; |
| 900 | tensor<fp16, [1, 1280, 1, 1]> var_1226_cast_fp16 = conv(dilations = var_1226_dilations_0, groups = var_1226_groups_0, pad = var_1226_pad_0, pad_type = var_1226_pad_type_0, strides = var_1226_strides_0, weight = layers_3_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_33_cast_fp16)[name = tensor<string, []>("op_1226_cast_fp16")]; |
| 901 | tensor<fp16, [1, 1280, 1, 1]> obj_53_cast_fp16 = add(x = var_1220_cast_fp16, y = var_1226_cast_fp16)[name = tensor<string, []>("obj_53_cast_fp16")]; |
| 902 | tensor<fp16, [1, 1280, 1, 1]> inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_53_cast_fp16)[name = tensor<string, []>("inputs_23_cast_fp16")]; |
| 903 | tensor<int32, [1]> out_23_axes_0 = const()[name = tensor<string, []>("out_23_axes_0"), val = tensor<int32, [1]>([1])]; |
| 904 | tensor<fp16, []> var_1240_to_fp16 = const()[name = tensor<string, []>("op_1240_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 905 | tensor<fp16, [1, 1280, 1, 1]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_1240_to_fp16, x = inputs_23_cast_fp16)[name = tensor<string, []>("out_23_cast_fp16")]; |
| 906 | tensor<fp16, [1280]> input_35_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_35_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193118400)))]; |
| 907 | tensor<fp16, [1280]> input_35_beta_0_to_fp16 = const()[name = tensor<string, []>("input_35_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193121024)))]; |
| 908 | tensor<fp16, []> input_35_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_35_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 909 | tensor<fp16, [1, 1280, 1, 1]> input_35_cast_fp16 = batch_norm(beta = input_35_beta_0_to_fp16, epsilon = input_35_epsilon_0_to_fp16, gamma = input_35_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")]; |
| 910 | tensor<string, []> var_1258_pad_type_0 = const()[name = tensor<string, []>("op_1258_pad_type_0"), val = tensor<string, []>("valid")]; |
| 911 | tensor<int32, [2]> var_1258_strides_0 = const()[name = tensor<string, []>("op_1258_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 912 | tensor<int32, [4]> var_1258_pad_0 = const()[name = tensor<string, []>("op_1258_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 913 | tensor<int32, [2]> var_1258_dilations_0 = const()[name = tensor<string, []>("op_1258_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 914 | tensor<int32, []> var_1258_groups_0 = const()[name = tensor<string, []>("op_1258_groups_0"), val = tensor<int32, []>(1)]; |
| 915 | tensor<fp16, [5120, 1280, 1, 1]> layers_3_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193123648))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196400512))), name = tensor<string, []>("layers_3_fc1_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])]; |
| 916 | tensor<fp16, [5120]> layers_3_fc1_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_inlier_module_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196400640)))]; |
| 917 | tensor<fp16, [1, 5120, 1, 1]> var_1258_cast_fp16 = conv(bias = layers_3_fc1_inlier_module_bias_to_fp16, dilations = var_1258_dilations_0, groups = var_1258_groups_0, pad = var_1258_pad_0, pad_type = var_1258_pad_type_0, strides = var_1258_strides_0, weight = layers_3_fc1_inlier_module_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = tensor<string, []>("op_1258_cast_fp16")]; |
| 918 | tensor<string, []> var_1264_pad_type_0 = const()[name = tensor<string, []>("op_1264_pad_type_0"), val = tensor<string, []>("valid")]; |
| 919 | tensor<int32, [2]> var_1264_strides_0 = const()[name = tensor<string, []>("op_1264_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 920 | tensor<int32, [4]> var_1264_pad_0 = const()[name = tensor<string, []>("op_1264_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 921 | tensor<int32, [2]> var_1264_dilations_0 = const()[name = tensor<string, []>("op_1264_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 922 | tensor<int32, []> var_1264_groups_0 = const()[name = tensor<string, []>("op_1264_groups_0"), val = tensor<int32, []>(1)]; |
| 923 | tensor<fp16, [5120, 1280, 1, 1]> layers_3_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196463680))), name = tensor<string, []>("layers_3_fc1_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [26331]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196410944))), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])]; |
| 924 | tensor<fp16, [1, 5120, 1, 1]> var_1264_cast_fp16 = conv(dilations = var_1264_dilations_0, groups = var_1264_groups_0, pad = var_1264_pad_0, pad_type = var_1264_pad_type_0, strides = var_1264_strides_0, weight = layers_3_fc1_outlier_module_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = tensor<string, []>("op_1264_cast_fp16")]; |
| 925 | tensor<fp16, [1, 5120, 1, 1]> input_37_cast_fp16 = add(x = var_1258_cast_fp16, y = var_1264_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")]; |
| 926 | tensor<string, []> input_mode_0 = const()[name = tensor<string, []>("input_mode_0"), val = tensor<string, []>("EXACT")]; |
| 927 | tensor<fp16, [1, 5120, 1, 1]> input_cast_fp16 = gelu(mode = input_mode_0, x = input_37_cast_fp16)[name = tensor<string, []>("input_cast_fp16")]; |
| 928 | tensor<string, []> var_1275_pad_type_0 = const()[name = tensor<string, []>("op_1275_pad_type_0"), val = tensor<string, []>("valid")]; |
| 929 | tensor<int32, [2]> var_1275_strides_0 = const()[name = tensor<string, []>("op_1275_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 930 | tensor<int32, [4]> var_1275_pad_0 = const()[name = tensor<string, []>("op_1275_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 931 | tensor<int32, [2]> var_1275_dilations_0 = const()[name = tensor<string, []>("op_1275_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 932 | tensor<int32, []> var_1275_groups_0 = const()[name = tensor<string, []>("op_1275_groups_0"), val = tensor<int32, []>(1)]; |
| 933 | tensor<fp16, [1280, 5120, 1, 1]> layers_3_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197282944))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202198208))), name = tensor<string, []>("layers_3_fc2_inlier_module_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])]; |
| 934 | tensor<fp16, [1280]> layers_3_fc2_inlier_module_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_inlier_module_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202198400)))]; |
| 935 | tensor<fp16, [1, 1280, 1, 1]> var_1275_cast_fp16 = conv(bias = layers_3_fc2_inlier_module_bias_to_fp16, dilations = var_1275_dilations_0, groups = var_1275_groups_0, pad = var_1275_pad_0, pad_type = var_1275_pad_type_0, strides = var_1275_strides_0, weight = layers_3_fc2_inlier_module_weight_to_fp16_palettized, x = input_cast_fp16)[name = tensor<string, []>("op_1275_cast_fp16")]; |
| 936 | tensor<string, []> var_1281_pad_type_0 = const()[name = tensor<string, []>("op_1281_pad_type_0"), val = tensor<string, []>("valid")]; |
| 937 | tensor<int32, [2]> var_1281_strides_0 = const()[name = tensor<string, []>("op_1281_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 938 | tensor<int32, [4]> var_1281_pad_0 = const()[name = tensor<string, []>("op_1281_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 939 | tensor<int32, [2]> var_1281_dilations_0 = const()[name = tensor<string, []>("op_1281_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 940 | tensor<int32, []> var_1281_groups_0 = const()[name = tensor<string, []>("op_1281_groups_0"), val = tensor<int32, []>(1)]; |
| 941 | tensor<fp16, [1280, 5120, 1, 1]> layers_3_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202271552))), name = tensor<string, []>("layers_3_fc2_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor<fp16, [35232]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202201024))), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])]; |
| 942 | tensor<fp16, [1, 1280, 1, 1]> var_1281_cast_fp16 = conv(dilations = var_1281_dilations_0, groups = var_1281_groups_0, pad = var_1281_pad_0, pad_type = var_1281_pad_type_0, strides = var_1281_strides_0, weight = layers_3_fc2_outlier_module_weight_to_fp16_sparsified, x = input_cast_fp16)[name = tensor<string, []>("op_1281_cast_fp16")]; |
| 943 | tensor<fp16, [1, 1280, 1, 1]> hidden_states_9_cast_fp16 = add(x = var_1275_cast_fp16, y = var_1281_cast_fp16)[name = tensor<string, []>("hidden_states_9_cast_fp16")]; |
| 944 | tensor<fp16, [1, 1280, 1, 1]> inputs_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor<string, []>("inputs_cast_fp16")]; |
| 945 | tensor<int32, [1]> out_axes_0 = const()[name = tensor<string, []>("out_axes_0"), val = tensor<int32, [1]>([1])]; |
| 946 | tensor<fp16, []> var_1301_to_fp16 = const()[name = tensor<string, []>("op_1301_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 947 | tensor<fp16, [1, 1280, 1, 1]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_1301_to_fp16, x = inputs_cast_fp16)[name = tensor<string, []>("out_cast_fp16")]; |
| 948 | tensor<fp16, [1280]> hidden_states_gamma_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203090816)))]; |
| 949 | tensor<fp16, [1280]> hidden_states_beta_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203093440)))]; |
| 950 | tensor<fp16, []> hidden_states_epsilon_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 951 | tensor<fp16, [1, 1280, 1, 1]> hidden_states_cast_fp16 = batch_norm(beta = hidden_states_beta_0_to_fp16, epsilon = hidden_states_epsilon_0_to_fp16, gamma = hidden_states_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor<string, []>("hidden_states_cast_fp16")]; |
| 952 | tensor<int32, [1]> var_1312_axes_0 = const()[name = tensor<string, []>("op_1312_axes_0"), val = tensor<int32, [1]>([2])]; |
| 953 | tensor<fp16, [1, 1280, 1]> var_1312_cast_fp16 = squeeze(axes = var_1312_axes_0, x = hidden_states_cast_fp16)[name = tensor<string, []>("op_1312_cast_fp16")]; |
| 954 | tensor<int32, [3]> var_1315_perm_0 = const()[name = tensor<string, []>("op_1315_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| 955 | tensor<fp16, [51866]> linear_0_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_0_bias_0_to_fp16"), val = tensor<fp16, [51866]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203096064)))]; |
| 956 | tensor<fp16, [1, 1, 1280]> var_1315_cast_fp16 = transpose(perm = var_1315_perm_0, x = var_1312_cast_fp16)[name = tensor<string, []>("transpose_0")]; |
| 957 | tensor<fp16, [1, 1, 51866]> logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = var_1315_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")]; |
| 958 | tensor<int32, []> var_1319 = const()[name = tensor<string, []>("op_1319"), val = tensor<int32, []>(1)]; |
| 959 | tensor<bool, []> obj_59_interleave_0 = const()[name = tensor<string, []>("obj_59_interleave_0"), val = tensor<bool, []>(false)]; |
| 960 | tensor<fp16, [1, 5120, 1, 1]> key_cache_updates = concat(axis = var_1319, interleave = obj_59_interleave_0, values = (current_key_1_cast_fp16, current_key_3_cast_fp16, current_key_5_cast_fp16, current_key_cast_fp16))[name = tensor<string, []>("obj_59_cast_fp16")]; |
| 961 | tensor<int32, []> var_1322 = const()[name = tensor<string, []>("op_1322"), val = tensor<int32, []>(1)]; |
| 962 | tensor<bool, []> obj_61_interleave_0 = const()[name = tensor<string, []>("obj_61_interleave_0"), val = tensor<bool, []>(false)]; |
| 963 | tensor<fp16, [1, 5120, 1, 1]> value_cache_updates = concat(axis = var_1322, interleave = obj_61_interleave_0, values = (current_value_1_cast_fp16, current_value_3_cast_fp16, current_value_5_cast_fp16, current_value_cast_fp16))[name = tensor<string, []>("obj_61_cast_fp16")]; |
| 964 | tensor<int32, [4]> var_1333_begin_0 = const()[name = tensor<string, []>("op_1333_begin_0"), val = tensor<int32, [4]>([0, 4, 0, 0])]; |
| 965 | tensor<int32, [4]> var_1333_end_0 = const()[name = tensor<string, []>("op_1333_end_0"), val = tensor<int32, [4]>([1, 5, 1, 1500])]; |
| 966 | tensor<bool, [4]> var_1333_end_mask_0 = const()[name = tensor<string, []>("op_1333_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| 967 | tensor<fp16, [1, 1, 1, 1500]> var_1333_cast_fp16 = slice_by_index(begin = var_1333_begin_0, end = var_1333_end_0, end_mask = var_1333_end_mask_0, x = obj_41_cast_fp16)[name = tensor<string, []>("op_1333_cast_fp16")]; |
| 968 | tensor<int32, [4]> var_1336_begin_0 = const()[name = tensor<string, []>("op_1336_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 969 | tensor<int32, [4]> var_1336_end_0 = const()[name = tensor<string, []>("op_1336_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| 970 | tensor<bool, [4]> var_1336_end_mask_0 = const()[name = tensor<string, []>("op_1336_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| 971 | tensor<bool, [4]> var_1336_squeeze_mask_0 = const()[name = tensor<string, []>("op_1336_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| 972 | tensor<fp16, [1, 1, 1500]> var_1336_cast_fp16 = slice_by_index(begin = var_1336_begin_0, end = var_1336_end_0, end_mask = var_1336_end_mask_0, squeeze_mask = var_1336_squeeze_mask_0, x = var_1333_cast_fp16)[name = tensor<string, []>("op_1336_cast_fp16")]; |
| 973 | tensor<int32, [4]> var_1351_begin_0 = const()[name = tensor<string, []>("op_1351_begin_0"), val = tensor<int32, [4]>([0, 11, 0, 0])]; |
| 974 | tensor<int32, [4]> var_1351_end_0 = const()[name = tensor<string, []>("op_1351_end_0"), val = tensor<int32, [4]>([1, 12, 1, 1500])]; |
| 975 | tensor<bool, [4]> var_1351_end_mask_0 = const()[name = tensor<string, []>("op_1351_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| 976 | tensor<fp16, [1, 1, 1, 1500]> var_1351_cast_fp16 = slice_by_index(begin = var_1351_begin_0, end = var_1351_end_0, end_mask = var_1351_end_mask_0, x = obj_41_cast_fp16)[name = tensor<string, []>("op_1351_cast_fp16")]; |
| 977 | tensor<int32, [4]> var_1354_begin_0 = const()[name = tensor<string, []>("op_1354_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 978 | tensor<int32, [4]> var_1354_end_0 = const()[name = tensor<string, []>("op_1354_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| 979 | tensor<bool, [4]> var_1354_end_mask_0 = const()[name = tensor<string, []>("op_1354_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| 980 | tensor<bool, [4]> var_1354_squeeze_mask_0 = const()[name = tensor<string, []>("op_1354_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| 981 | tensor<fp16, [1, 1, 1500]> var_1354_cast_fp16 = slice_by_index(begin = var_1354_begin_0, end = var_1354_end_0, end_mask = var_1354_end_mask_0, squeeze_mask = var_1354_squeeze_mask_0, x = var_1351_cast_fp16)[name = tensor<string, []>("op_1354_cast_fp16")]; |
| 982 | tensor<int32, [4]> var_1369_begin_0 = const()[name = tensor<string, []>("op_1369_begin_0"), val = tensor<int32, [4]>([0, 3, 0, 0])]; |
| 983 | tensor<int32, [4]> var_1369_end_0 = const()[name = tensor<string, []>("op_1369_end_0"), val = tensor<int32, [4]>([1, 4, 1, 1500])]; |
| 984 | tensor<bool, [4]> var_1369_end_mask_0 = const()[name = tensor<string, []>("op_1369_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| 985 | tensor<fp16, [1, 1, 1, 1500]> var_1369_cast_fp16 = slice_by_index(begin = var_1369_begin_0, end = var_1369_end_0, end_mask = var_1369_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1369_cast_fp16")]; |
| 986 | tensor<int32, [4]> var_1372_begin_0 = const()[name = tensor<string, []>("op_1372_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 987 | tensor<int32, [4]> var_1372_end_0 = const()[name = tensor<string, []>("op_1372_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| 988 | tensor<bool, [4]> var_1372_end_mask_0 = const()[name = tensor<string, []>("op_1372_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| 989 | tensor<bool, [4]> var_1372_squeeze_mask_0 = const()[name = tensor<string, []>("op_1372_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| 990 | tensor<fp16, [1, 1, 1500]> var_1372_cast_fp16 = slice_by_index(begin = var_1372_begin_0, end = var_1372_end_0, end_mask = var_1372_end_mask_0, squeeze_mask = var_1372_squeeze_mask_0, x = var_1369_cast_fp16)[name = tensor<string, []>("op_1372_cast_fp16")]; |
| 991 | tensor<int32, [4]> var_1387_begin_0 = const()[name = tensor<string, []>("op_1387_begin_0"), val = tensor<int32, [4]>([0, 6, 0, 0])]; |
| 992 | tensor<int32, [4]> var_1387_end_0 = const()[name = tensor<string, []>("op_1387_end_0"), val = tensor<int32, [4]>([1, 7, 1, 1500])]; |
| 993 | tensor<bool, [4]> var_1387_end_mask_0 = const()[name = tensor<string, []>("op_1387_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| 994 | tensor<fp16, [1, 1, 1, 1500]> var_1387_cast_fp16 = slice_by_index(begin = var_1387_begin_0, end = var_1387_end_0, end_mask = var_1387_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1387_cast_fp16")]; |
| 995 | tensor<int32, [4]> var_1390_begin_0 = const()[name = tensor<string, []>("op_1390_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 996 | tensor<int32, [4]> var_1390_end_0 = const()[name = tensor<string, []>("op_1390_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| 997 | tensor<bool, [4]> var_1390_end_mask_0 = const()[name = tensor<string, []>("op_1390_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| 998 | tensor<bool, [4]> var_1390_squeeze_mask_0 = const()[name = tensor<string, []>("op_1390_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| 999 | tensor<fp16, [1, 1, 1500]> var_1390_cast_fp16 = slice_by_index(begin = var_1390_begin_0, end = var_1390_end_0, end_mask = var_1390_end_mask_0, squeeze_mask = var_1390_squeeze_mask_0, x = var_1387_cast_fp16)[name = tensor<string, []>("op_1390_cast_fp16")]; |
| 1000 | tensor<int32, [4]> var_1405_begin_0 = const()[name = tensor<string, []>("op_1405_begin_0"), val = tensor<int32, [4]>([0, 11, 0, 0])]; |
| 1001 | tensor<int32, [4]> var_1405_end_0 = const()[name = tensor<string, []>("op_1405_end_0"), val = tensor<int32, [4]>([1, 12, 1, 1500])]; |
| 1002 | tensor<bool, [4]> var_1405_end_mask_0 = const()[name = tensor<string, []>("op_1405_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| 1003 | tensor<fp16, [1, 1, 1, 1500]> var_1405_cast_fp16 = slice_by_index(begin = var_1405_begin_0, end = var_1405_end_0, end_mask = var_1405_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1405_cast_fp16")]; |
| 1004 | tensor<int32, [4]> var_1408_begin_0 = const()[name = tensor<string, []>("op_1408_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 1005 | tensor<int32, [4]> var_1408_end_0 = const()[name = tensor<string, []>("op_1408_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| 1006 | tensor<bool, [4]> var_1408_end_mask_0 = const()[name = tensor<string, []>("op_1408_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| 1007 | tensor<bool, [4]> var_1408_squeeze_mask_0 = const()[name = tensor<string, []>("op_1408_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| 1008 | tensor<fp16, [1, 1, 1500]> var_1408_cast_fp16 = slice_by_index(begin = var_1408_begin_0, end = var_1408_end_0, end_mask = var_1408_end_mask_0, squeeze_mask = var_1408_squeeze_mask_0, x = var_1405_cast_fp16)[name = tensor<string, []>("op_1408_cast_fp16")]; |
| 1009 | tensor<int32, [4]> var_1423_begin_0 = const()[name = tensor<string, []>("op_1423_begin_0"), val = tensor<int32, [4]>([0, 14, 0, 0])]; |
| 1010 | tensor<int32, [4]> var_1423_end_0 = const()[name = tensor<string, []>("op_1423_end_0"), val = tensor<int32, [4]>([1, 15, 1, 1500])]; |
| 1011 | tensor<bool, [4]> var_1423_end_mask_0 = const()[name = tensor<string, []>("op_1423_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| 1012 | tensor<fp16, [1, 1, 1, 1500]> var_1423_cast_fp16 = slice_by_index(begin = var_1423_begin_0, end = var_1423_end_0, end_mask = var_1423_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1423_cast_fp16")]; |
| 1013 | tensor<int32, [4]> var_1426_begin_0 = const()[name = tensor<string, []>("op_1426_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 1014 | tensor<int32, [4]> var_1426_end_0 = const()[name = tensor<string, []>("op_1426_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| 1015 | tensor<bool, [4]> var_1426_end_mask_0 = const()[name = tensor<string, []>("op_1426_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| 1016 | tensor<bool, [4]> var_1426_squeeze_mask_0 = const()[name = tensor<string, []>("op_1426_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| 1017 | tensor<fp16, [1, 1, 1500]> var_1426_cast_fp16 = slice_by_index(begin = var_1426_begin_0, end = var_1426_end_0, end_mask = var_1426_end_mask_0, squeeze_mask = var_1426_squeeze_mask_0, x = var_1423_cast_fp16)[name = tensor<string, []>("op_1426_cast_fp16")]; |
| 1018 | tensor<int32, []> var_1433 = const()[name = tensor<string, []>("op_1433"), val = tensor<int32, []>(1)]; |
| 1019 | tensor<bool, []> var_1434_interleave_0 = const()[name = tensor<string, []>("op_1434_interleave_0"), val = tensor<bool, []>(false)]; |
| 1020 | tensor<fp16, [1, 6, 1500]> var_1434_cast_fp16 = concat(axis = var_1433, interleave = var_1434_interleave_0, values = (var_1336_cast_fp16, var_1354_cast_fp16, var_1372_cast_fp16, var_1390_cast_fp16, var_1408_cast_fp16, var_1426_cast_fp16))[name = tensor<string, []>("op_1434_cast_fp16")]; |
| 1021 | tensor<bool, []> var_1437 = const()[name = tensor<string, []>("op_1437"), val = tensor<bool, []>(false)]; |
| 1022 | tensor<int32, [1]> obj_axes_0 = const()[name = tensor<string, []>("obj_axes_0"), val = tensor<int32, [1]>([1])]; |
| 1023 | tensor<fp16, [1, 1500]> alignment_heads_weights = reduce_mean(axes = obj_axes_0, keep_dims = var_1437, x = var_1434_cast_fp16)[name = tensor<string, []>("obj_cast_fp16")]; |
| 1024 | } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights); |
| 1025 | } |