openai_whisper-large-v3-v20240930/TextDecoder.mlmodelc/model.mil
| 1 | program(1.0) |
| 2 | [buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}, {"coremltools-component-torch", "2.4.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.0"}})] |
| 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_28_axis_0 = const()[name = tensor<string, []>("op_28_axis_0"), val = tensor<int32, []>(0)]; |
| 10 | tensor<int32, []> var_28_batch_dims_0 = const()[name = tensor<string, []>("op_28_batch_dims_0"), val = tensor<int32, []>(0)]; |
| 11 | tensor<fp16, [448, 1280]> embed_positions_weight_to_fp16 = const()[name = tensor<string, []>("embed_positions_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_28_cast_fp16 = gather(axis = var_28_axis_0, batch_dims = var_28_batch_dims_0, indices = cache_length, x = embed_positions_weight_to_fp16)[name = tensor<string, []>("op_28_cast_fp16")]; |
| 13 | tensor<fp16, [1, 1280]> hidden_states_1_cast_fp16 = add(x = var_24_cast_fp16, y = var_28_cast_fp16)[name = tensor<string, []>("hidden_states_1_cast_fp16")]; |
| 14 | tensor<int32, [1]> var_42_axes_0 = const()[name = tensor<string, []>("op_42_axes_0"), val = tensor<int32, [1]>([2])]; |
| 15 | tensor<fp16, [1, 1280, 1]> var_42_cast_fp16 = expand_dims(axes = var_42_axes_0, x = hidden_states_1_cast_fp16)[name = tensor<string, []>("op_42_cast_fp16")]; |
| 16 | tensor<int32, [1]> inputs_1_axes_0 = const()[name = tensor<string, []>("inputs_1_axes_0"), val = tensor<int32, [1]>([3])]; |
| 17 | tensor<fp16, [1, 1280, 1, 1]> inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_42_cast_fp16)[name = tensor<string, []>("inputs_1_cast_fp16")]; |
| 18 | tensor<int32, [4]> tile_0 = const()[name = tensor<string, []>("tile_0"), val = tensor<int32, [4]>([1280, 1280, 1280, 1280])]; |
| 19 | tensor<int32, []> var_47_axis_0 = const()[name = tensor<string, []>("op_47_axis_0"), val = tensor<int32, []>(1)]; |
| 20 | tensor<fp16, [1, 1280, 1, 448]> var_47_cast_fp16_0, tensor<fp16, [1, 1280, 1, 448]> var_47_cast_fp16_1, tensor<fp16, [1, 1280, 1, 448]> var_47_cast_fp16_2, tensor<fp16, [1, 1280, 1, 448]> var_47_cast_fp16_3 = split(axis = var_47_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor<string, []>("op_47_cast_fp16")]; |
| 21 | tensor<int32, [4]> tile_1 = const()[name = tensor<string, []>("tile_1"), val = tensor<int32, [4]>([1280, 1280, 1280, 1280])]; |
| 22 | tensor<int32, []> var_54_axis_0 = const()[name = tensor<string, []>("op_54_axis_0"), val = tensor<int32, []>(1)]; |
| 23 | tensor<fp16, [1, 1280, 1, 448]> var_54_cast_fp16_0, tensor<fp16, [1, 1280, 1, 448]> var_54_cast_fp16_1, tensor<fp16, [1, 1280, 1, 448]> var_54_cast_fp16_2, tensor<fp16, [1, 1280, 1, 448]> var_54_cast_fp16_3 = split(axis = var_54_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor<string, []>("op_54_cast_fp16")]; |
| 24 | tensor<int32, []> var_64 = const()[name = tensor<string, []>("op_64"), val = tensor<int32, []>(3)]; |
| 25 | tensor<int32, [1]> out_1_axes_0 = const()[name = tensor<string, []>("out_1_axes_0"), val = tensor<int32, [1]>([1])]; |
| 26 | tensor<fp16, []> var_90_to_fp16 = const()[name = tensor<string, []>("op_90_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 27 | tensor<fp16, [1, 1280, 1, 1]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_90_to_fp16, x = inputs_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")]; |
| 28 | 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, []>(133924032)))]; |
| 29 | 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, []>(133926656)))]; |
| 30 | 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, []>(133929280)))]; |
| 31 | 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, []>(133931904)))]; |
| 32 | tensor<fp16, []> obj_1_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_1_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 33 | 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")]; |
| 34 | tensor<string, []> query_1_pad_type_0 = const()[name = tensor<string, []>("query_1_pad_type_0"), val = tensor<string, []>("valid")]; |
| 35 | tensor<int32, [2]> query_1_strides_0 = const()[name = tensor<string, []>("query_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 36 | tensor<int32, [4]> query_1_pad_0 = const()[name = tensor<string, []>("query_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 37 | tensor<int32, [2]> query_1_dilations_0 = const()[name = tensor<string, []>("query_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 38 | tensor<int32, []> query_1_groups_0 = const()[name = tensor<string, []>("query_1_groups_0"), val = tensor<int32, []>(1)]; |
| 39 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133934528)))]; |
| 40 | tensor<fp16, [1280]> layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137211392)))]; |
| 41 | tensor<fp16, [1, 1280, 1, 1]> query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = query_1_dilations_0, groups = query_1_groups_0, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = query_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("query_1_cast_fp16")]; |
| 42 | tensor<string, []> current_key_1_pad_type_0 = const()[name = tensor<string, []>("current_key_1_pad_type_0"), val = tensor<string, []>("valid")]; |
| 43 | tensor<int32, [2]> current_key_1_strides_0 = const()[name = tensor<string, []>("current_key_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 44 | tensor<int32, [4]> current_key_1_pad_0 = const()[name = tensor<string, []>("current_key_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 45 | tensor<int32, [2]> current_key_1_dilations_0 = const()[name = tensor<string, []>("current_key_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 46 | tensor<int32, []> current_key_1_groups_0 = const()[name = tensor<string, []>("current_key_1_groups_0"), val = tensor<int32, []>(1)]; |
| 47 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137214016)))]; |
| 48 | tensor<fp16, [1, 1280, 1, 1]> current_key_1_cast_fp16 = conv(dilations = current_key_1_dilations_0, groups = current_key_1_groups_0, pad = current_key_1_pad_0, pad_type = current_key_1_pad_type_0, strides = current_key_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("current_key_1_cast_fp16")]; |
| 49 | tensor<string, []> current_value_1_pad_type_0 = const()[name = tensor<string, []>("current_value_1_pad_type_0"), val = tensor<string, []>("valid")]; |
| 50 | tensor<int32, [2]> current_value_1_strides_0 = const()[name = tensor<string, []>("current_value_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 51 | tensor<int32, [4]> current_value_1_pad_0 = const()[name = tensor<string, []>("current_value_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 52 | tensor<int32, [2]> current_value_1_dilations_0 = const()[name = tensor<string, []>("current_value_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 53 | tensor<int32, []> current_value_1_groups_0 = const()[name = tensor<string, []>("current_value_1_groups_0"), val = tensor<int32, []>(1)]; |
| 54 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140490880)))]; |
| 55 | tensor<fp16, [1280]> layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143767744)))]; |
| 56 | tensor<fp16, [1, 1280, 1, 1]> current_value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = current_value_1_dilations_0, groups = current_value_1_groups_0, pad = current_value_1_pad_0, pad_type = current_value_1_pad_type_0, strides = current_value_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("current_value_1_cast_fp16")]; |
| 57 | tensor<int32, [1]> var_125_axes_0 = const()[name = tensor<string, []>("op_125_axes_0"), val = tensor<int32, [1]>([1])]; |
| 58 | tensor<fp16, [1, 1, 448]> var_125_cast_fp16 = expand_dims(axes = var_125_axes_0, x = kv_cache_update_mask)[name = tensor<string, []>("op_125_cast_fp16")]; |
| 59 | tensor<int32, [1]> var_126_axes_0 = const()[name = tensor<string, []>("op_126_axes_0"), val = tensor<int32, [1]>([2])]; |
| 60 | tensor<fp16, [1, 1, 1, 448]> var_126_cast_fp16 = expand_dims(axes = var_126_axes_0, x = var_125_cast_fp16)[name = tensor<string, []>("op_126_cast_fp16")]; |
| 61 | tensor<fp16, [1, 1280, 1, 448]> var_128_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_126_cast_fp16)[name = tensor<string, []>("op_128_cast_fp16")]; |
| 62 | tensor<fp16, []> var_65_to_fp16 = const()[name = tensor<string, []>("op_65_to_fp16"), val = tensor<fp16, []>(0x1p+0)]; |
| 63 | tensor<fp16, [1, 1, 1, 448]> var_129_cast_fp16 = sub(x = var_65_to_fp16, y = var_126_cast_fp16)[name = tensor<string, []>("op_129_cast_fp16")]; |
| 64 | tensor<fp16, [1, 1280, 1, 448]> var_130_cast_fp16 = mul(x = var_47_cast_fp16_0, y = var_129_cast_fp16)[name = tensor<string, []>("op_130_cast_fp16")]; |
| 65 | tensor<fp16, [1, 1280, 1, 448]> key_1_cast_fp16 = add(x = var_128_cast_fp16, y = var_130_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")]; |
| 66 | tensor<fp16, [1, 1280, 1, 448]> var_132_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_126_cast_fp16)[name = tensor<string, []>("op_132_cast_fp16")]; |
| 67 | tensor<fp16, [1, 1280, 1, 448]> var_134_cast_fp16 = mul(x = var_54_cast_fp16_0, y = var_129_cast_fp16)[name = tensor<string, []>("op_134_cast_fp16")]; |
| 68 | tensor<fp16, [1, 1280, 1, 448]> value_1_cast_fp16 = add(x = var_132_cast_fp16, y = var_134_cast_fp16)[name = tensor<string, []>("value_1_cast_fp16")]; |
| 69 | tensor<int32, [4]> var_137 = const()[name = tensor<string, []>("op_137"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 70 | tensor<fp16, [1, 20, 64, 1]> mh_q_1_cast_fp16 = reshape(shape = var_137, x = query_1_cast_fp16)[name = tensor<string, []>("mh_q_1_cast_fp16")]; |
| 71 | tensor<fp16, []> var_139_to_fp16 = const()[name = tensor<string, []>("op_139_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| 72 | tensor<fp16, [1, 20, 64, 1]> var_140_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_139_to_fp16)[name = tensor<string, []>("op_140_cast_fp16")]; |
| 73 | tensor<int32, [4]> var_141 = const()[name = tensor<string, []>("op_141"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 74 | tensor<fp16, [1, 20, 64, 448]> var_142_cast_fp16 = reshape(shape = var_141, x = key_1_cast_fp16)[name = tensor<string, []>("op_142_cast_fp16")]; |
| 75 | tensor<bool, []> mh_w_1_transpose_x_0 = const()[name = tensor<string, []>("mh_w_1_transpose_x_0"), val = tensor<bool, []>(true)]; |
| 76 | tensor<bool, []> mh_w_1_transpose_y_0 = const()[name = tensor<string, []>("mh_w_1_transpose_y_0"), val = tensor<bool, []>(false)]; |
| 77 | 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_140_cast_fp16, y = var_142_cast_fp16)[name = tensor<string, []>("mh_w_1_cast_fp16")]; |
| 78 | tensor<int32, [1]> var_146_axes_0 = const()[name = tensor<string, []>("op_146_axes_0"), val = tensor<int32, [1]>([1])]; |
| 79 | tensor<fp16, [1, 1, 448]> var_146_cast_fp16 = expand_dims(axes = var_146_axes_0, x = decoder_key_padding_mask)[name = tensor<string, []>("op_146_cast_fp16")]; |
| 80 | tensor<int32, [1]> var_147_axes_0 = const()[name = tensor<string, []>("op_147_axes_0"), val = tensor<int32, [1]>([2])]; |
| 81 | tensor<fp16, [1, 1, 1, 448]> var_147_cast_fp16 = expand_dims(axes = var_147_axes_0, x = var_146_cast_fp16)[name = tensor<string, []>("op_147_cast_fp16")]; |
| 82 | tensor<fp16, [1, 20, 1, 448]> mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_147_cast_fp16)[name = tensor<string, []>("mh_w_3_cast_fp16")]; |
| 83 | tensor<fp16, [1, 20, 1, 448]> var_150_cast_fp16 = softmax(axis = var_64, x = mh_w_3_cast_fp16)[name = tensor<string, []>("op_150_cast_fp16")]; |
| 84 | tensor<int32, [4]> var_151 = const()[name = tensor<string, []>("op_151"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 85 | tensor<fp16, [1, 20, 64, 448]> var_152_cast_fp16 = reshape(shape = var_151, x = value_1_cast_fp16)[name = tensor<string, []>("op_152_cast_fp16")]; |
| 86 | tensor<bool, []> attn_1_transpose_x_0 = const()[name = tensor<string, []>("attn_1_transpose_x_0"), val = tensor<bool, []>(false)]; |
| 87 | tensor<bool, []> attn_1_transpose_y_0 = const()[name = tensor<string, []>("attn_1_transpose_y_0"), val = tensor<bool, []>(true)]; |
| 88 | 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_152_cast_fp16, y = var_150_cast_fp16)[name = tensor<string, []>("attn_1_cast_fp16")]; |
| 89 | tensor<int32, [4]> var_155 = const()[name = tensor<string, []>("op_155"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; |
| 90 | tensor<fp16, [1, 1280, 1, 1]> input_1_cast_fp16 = reshape(shape = var_155, x = attn_1_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")]; |
| 91 | tensor<string, []> obj_7_pad_type_0 = const()[name = tensor<string, []>("obj_7_pad_type_0"), val = tensor<string, []>("valid")]; |
| 92 | tensor<int32, [2]> obj_7_strides_0 = const()[name = tensor<string, []>("obj_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 93 | tensor<int32, [4]> obj_7_pad_0 = const()[name = tensor<string, []>("obj_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 94 | tensor<int32, [2]> obj_7_dilations_0 = const()[name = tensor<string, []>("obj_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 95 | tensor<int32, []> obj_7_groups_0 = const()[name = tensor<string, []>("obj_7_groups_0"), val = tensor<int32, []>(1)]; |
| 96 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143770368)))]; |
| 97 | tensor<fp16, [1280]> layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147047232)))]; |
| 98 | tensor<fp16, [1, 1280, 1, 1]> obj_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = obj_7_dilations_0, groups = obj_7_groups_0, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = obj_7_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("obj_7_cast_fp16")]; |
| 99 | 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")]; |
| 100 | tensor<int32, [1]> out_3_axes_0 = const()[name = tensor<string, []>("out_3_axes_0"), val = tensor<int32, [1]>([1])]; |
| 101 | tensor<fp16, []> var_177_to_fp16 = const()[name = tensor<string, []>("op_177_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 102 | tensor<fp16, [1, 1280, 1, 1]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_177_to_fp16, x = inputs_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")]; |
| 103 | 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, []>(147049856)))]; |
| 104 | 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, []>(147052480)))]; |
| 105 | tensor<fp16, []> obj_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 106 | 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")]; |
| 107 | tensor<string, []> query_3_pad_type_0 = const()[name = tensor<string, []>("query_3_pad_type_0"), val = tensor<string, []>("valid")]; |
| 108 | tensor<int32, [2]> query_3_strides_0 = const()[name = tensor<string, []>("query_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 109 | tensor<int32, [4]> query_3_pad_0 = const()[name = tensor<string, []>("query_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 110 | tensor<int32, [2]> query_3_dilations_0 = const()[name = tensor<string, []>("query_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 111 | tensor<int32, []> query_3_groups_0 = const()[name = tensor<string, []>("query_3_groups_0"), val = tensor<int32, []>(1)]; |
| 112 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147055104)))]; |
| 113 | tensor<fp16, [1280]> layers_0_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150331968)))]; |
| 114 | tensor<fp16, [1, 1280, 1, 1]> query_3_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_bias_to_fp16, dilations = query_3_dilations_0, groups = query_3_groups_0, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = query_3_strides_0, weight = layers_0_encoder_attn_q_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor<string, []>("query_3_cast_fp16")]; |
| 115 | tensor<string, []> key_3_pad_type_0 = const()[name = tensor<string, []>("key_3_pad_type_0"), val = tensor<string, []>("valid")]; |
| 116 | tensor<int32, [2]> key_3_strides_0 = const()[name = tensor<string, []>("key_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 117 | tensor<int32, [4]> key_3_pad_0 = const()[name = tensor<string, []>("key_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 118 | tensor<int32, [2]> key_3_dilations_0 = const()[name = tensor<string, []>("key_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 119 | tensor<int32, []> key_3_groups_0 = const()[name = tensor<string, []>("key_3_groups_0"), val = tensor<int32, []>(1)]; |
| 120 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150334592)))]; |
| 121 | tensor<fp16, [1, 1280, 1, 1500]> key_3_cast_fp16 = conv(dilations = key_3_dilations_0, groups = key_3_groups_0, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = key_3_strides_0, weight = layers_0_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_3_cast_fp16")]; |
| 122 | tensor<string, []> value_3_pad_type_0 = const()[name = tensor<string, []>("value_3_pad_type_0"), val = tensor<string, []>("valid")]; |
| 123 | tensor<int32, [2]> value_3_strides_0 = const()[name = tensor<string, []>("value_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 124 | tensor<int32, [4]> value_3_pad_0 = const()[name = tensor<string, []>("value_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 125 | tensor<int32, [2]> value_3_dilations_0 = const()[name = tensor<string, []>("value_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 126 | tensor<int32, []> value_3_groups_0 = const()[name = tensor<string, []>("value_3_groups_0"), val = tensor<int32, []>(1)]; |
| 127 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153611456)))]; |
| 128 | tensor<fp16, [1280]> layers_0_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156888320)))]; |
| 129 | tensor<fp16, [1, 1280, 1, 1500]> value_3_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_bias_to_fp16, dilations = value_3_dilations_0, groups = value_3_groups_0, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = value_3_strides_0, weight = layers_0_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_3_cast_fp16")]; |
| 130 | tensor<int32, [4]> var_212 = const()[name = tensor<string, []>("op_212"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 131 | tensor<fp16, [1, 20, 64, 1]> mh_q_3_cast_fp16 = reshape(shape = var_212, x = query_3_cast_fp16)[name = tensor<string, []>("mh_q_3_cast_fp16")]; |
| 132 | tensor<fp16, []> var_214_to_fp16 = const()[name = tensor<string, []>("op_214_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| 133 | tensor<fp16, [1, 20, 64, 1]> var_215_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_214_to_fp16)[name = tensor<string, []>("op_215_cast_fp16")]; |
| 134 | tensor<int32, [4]> var_216 = const()[name = tensor<string, []>("op_216"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 135 | tensor<fp16, [1, 20, 64, 1500]> var_217_cast_fp16 = reshape(shape = var_216, x = key_3_cast_fp16)[name = tensor<string, []>("op_217_cast_fp16")]; |
| 136 | tensor<bool, []> mh_w_5_transpose_x_0 = const()[name = tensor<string, []>("mh_w_5_transpose_x_0"), val = tensor<bool, []>(true)]; |
| 137 | tensor<bool, []> mh_w_5_transpose_y_0 = const()[name = tensor<string, []>("mh_w_5_transpose_y_0"), val = tensor<bool, []>(false)]; |
| 138 | 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_215_cast_fp16, y = var_217_cast_fp16)[name = tensor<string, []>("mh_w_5_cast_fp16")]; |
| 139 | tensor<fp16, [1, 20, 1, 1500]> obj_13_cast_fp16 = softmax(axis = var_64, x = mh_w_5_cast_fp16)[name = tensor<string, []>("obj_13_cast_fp16")]; |
| 140 | tensor<int32, [4]> var_221 = const()[name = tensor<string, []>("op_221"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 141 | tensor<fp16, [1, 20, 64, 1500]> var_222_cast_fp16 = reshape(shape = var_221, x = value_3_cast_fp16)[name = tensor<string, []>("op_222_cast_fp16")]; |
| 142 | tensor<bool, []> attn_3_transpose_x_0 = const()[name = tensor<string, []>("attn_3_transpose_x_0"), val = tensor<bool, []>(false)]; |
| 143 | tensor<bool, []> attn_3_transpose_y_0 = const()[name = tensor<string, []>("attn_3_transpose_y_0"), val = tensor<bool, []>(true)]; |
| 144 | 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_222_cast_fp16, y = obj_13_cast_fp16)[name = tensor<string, []>("attn_3_cast_fp16")]; |
| 145 | tensor<int32, [4]> var_225 = const()[name = tensor<string, []>("op_225"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; |
| 146 | tensor<fp16, [1, 1280, 1, 1]> input_3_cast_fp16 = reshape(shape = var_225, x = attn_3_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")]; |
| 147 | tensor<string, []> obj_11_pad_type_0 = const()[name = tensor<string, []>("obj_11_pad_type_0"), val = tensor<string, []>("valid")]; |
| 148 | tensor<int32, [2]> obj_11_strides_0 = const()[name = tensor<string, []>("obj_11_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 149 | tensor<int32, [4]> obj_11_pad_0 = const()[name = tensor<string, []>("obj_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 150 | tensor<int32, [2]> obj_11_dilations_0 = const()[name = tensor<string, []>("obj_11_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 151 | tensor<int32, []> obj_11_groups_0 = const()[name = tensor<string, []>("obj_11_groups_0"), val = tensor<int32, []>(1)]; |
| 152 | tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156890944)))]; |
| 153 | tensor<fp16, [1280]> layers_0_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160167808)))]; |
| 154 | tensor<fp16, [1, 1280, 1, 1]> obj_11_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = obj_11_dilations_0, groups = obj_11_groups_0, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = obj_11_strides_0, weight = layers_0_encoder_attn_o_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("obj_11_cast_fp16")]; |
| 155 | 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")]; |
| 156 | tensor<int32, [1]> out_5_axes_0 = const()[name = tensor<string, []>("out_5_axes_0"), val = tensor<int32, [1]>([1])]; |
| 157 | tensor<fp16, []> var_243_to_fp16 = const()[name = tensor<string, []>("op_243_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 158 | tensor<fp16, [1, 1280, 1, 1]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_243_to_fp16, x = inputs_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")]; |
| 159 | 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, []>(160170432)))]; |
| 160 | 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, []>(160173056)))]; |
| 161 | tensor<fp16, []> input_5_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_5_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 162 | 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")]; |
| 163 | tensor<string, []> input_7_pad_type_0 = const()[name = tensor<string, []>("input_7_pad_type_0"), val = tensor<string, []>("valid")]; |
| 164 | tensor<int32, [2]> input_7_strides_0 = const()[name = tensor<string, []>("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 165 | tensor<int32, [4]> input_7_pad_0 = const()[name = tensor<string, []>("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 166 | tensor<int32, [2]> input_7_dilations_0 = const()[name = tensor<string, []>("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 167 | tensor<int32, []> input_7_groups_0 = const()[name = tensor<string, []>("input_7_groups_0"), val = tensor<int32, []>(1)]; |
| 168 | tensor<fp16, [5120, 1280, 1, 1]> layers_0_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160175680)))]; |
| 169 | tensor<fp16, [5120]> layers_0_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(173282944)))]; |
| 170 | tensor<fp16, [1, 5120, 1, 1]> input_7_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_0_fc1_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")]; |
| 171 | tensor<string, []> input_9_mode_0 = const()[name = tensor<string, []>("input_9_mode_0"), val = tensor<string, []>("EXACT")]; |
| 172 | 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")]; |
| 173 | tensor<string, []> hidden_states_3_pad_type_0 = const()[name = tensor<string, []>("hidden_states_3_pad_type_0"), val = tensor<string, []>("valid")]; |
| 174 | tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = tensor<string, []>("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 175 | tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = tensor<string, []>("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 176 | tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = tensor<string, []>("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 177 | tensor<int32, []> hidden_states_3_groups_0 = const()[name = tensor<string, []>("hidden_states_3_groups_0"), val = tensor<int32, []>(1)]; |
| 178 | tensor<fp16, [1280, 5120, 1, 1]> layers_0_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(173293248)))]; |
| 179 | tensor<fp16, [1280]> layers_0_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186400512)))]; |
| 180 | tensor<fp16, [1, 1280, 1, 1]> hidden_states_3_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_fc2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("hidden_states_3_cast_fp16")]; |
| 181 | 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")]; |
| 182 | tensor<int32, []> var_278 = const()[name = tensor<string, []>("op_278"), val = tensor<int32, []>(3)]; |
| 183 | tensor<int32, [1]> out_7_axes_0 = const()[name = tensor<string, []>("out_7_axes_0"), val = tensor<int32, [1]>([1])]; |
| 184 | tensor<fp16, []> var_304_to_fp16 = const()[name = tensor<string, []>("op_304_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 185 | tensor<fp16, [1, 1280, 1, 1]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_304_to_fp16, x = inputs_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")]; |
| 186 | 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, []>(186403136)))]; |
| 187 | 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, []>(186405760)))]; |
| 188 | tensor<fp16, []> obj_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 189 | 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")]; |
| 190 | tensor<string, []> query_5_pad_type_0 = const()[name = tensor<string, []>("query_5_pad_type_0"), val = tensor<string, []>("valid")]; |
| 191 | tensor<int32, [2]> query_5_strides_0 = const()[name = tensor<string, []>("query_5_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 192 | tensor<int32, [4]> query_5_pad_0 = const()[name = tensor<string, []>("query_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 193 | tensor<int32, [2]> query_5_dilations_0 = const()[name = tensor<string, []>("query_5_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 194 | tensor<int32, []> query_5_groups_0 = const()[name = tensor<string, []>("query_5_groups_0"), val = tensor<int32, []>(1)]; |
| 195 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186408384)))]; |
| 196 | tensor<fp16, [1280]> layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189685248)))]; |
| 197 | tensor<fp16, [1, 1280, 1, 1]> query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = query_5_dilations_0, groups = query_5_groups_0, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = query_5_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")]; |
| 198 | tensor<string, []> current_key_3_pad_type_0 = const()[name = tensor<string, []>("current_key_3_pad_type_0"), val = tensor<string, []>("valid")]; |
| 199 | tensor<int32, [2]> current_key_3_strides_0 = const()[name = tensor<string, []>("current_key_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 200 | tensor<int32, [4]> current_key_3_pad_0 = const()[name = tensor<string, []>("current_key_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 201 | tensor<int32, [2]> current_key_3_dilations_0 = const()[name = tensor<string, []>("current_key_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 202 | tensor<int32, []> current_key_3_groups_0 = const()[name = tensor<string, []>("current_key_3_groups_0"), val = tensor<int32, []>(1)]; |
| 203 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189687872)))]; |
| 204 | tensor<fp16, [1, 1280, 1, 1]> current_key_3_cast_fp16 = conv(dilations = current_key_3_dilations_0, groups = current_key_3_groups_0, pad = current_key_3_pad_0, pad_type = current_key_3_pad_type_0, strides = current_key_3_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("current_key_3_cast_fp16")]; |
| 205 | tensor<string, []> current_value_3_pad_type_0 = const()[name = tensor<string, []>("current_value_3_pad_type_0"), val = tensor<string, []>("valid")]; |
| 206 | tensor<int32, [2]> current_value_3_strides_0 = const()[name = tensor<string, []>("current_value_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 207 | tensor<int32, [4]> current_value_3_pad_0 = const()[name = tensor<string, []>("current_value_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 208 | tensor<int32, [2]> current_value_3_dilations_0 = const()[name = tensor<string, []>("current_value_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 209 | tensor<int32, []> current_value_3_groups_0 = const()[name = tensor<string, []>("current_value_3_groups_0"), val = tensor<int32, []>(1)]; |
| 210 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192964736)))]; |
| 211 | tensor<fp16, [1280]> layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196241600)))]; |
| 212 | tensor<fp16, [1, 1280, 1, 1]> current_value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = current_value_3_dilations_0, groups = current_value_3_groups_0, pad = current_value_3_pad_0, pad_type = current_value_3_pad_type_0, strides = current_value_3_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("current_value_3_cast_fp16")]; |
| 213 | tensor<fp16, [1, 1280, 1, 448]> var_342_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_126_cast_fp16)[name = tensor<string, []>("op_342_cast_fp16")]; |
| 214 | tensor<fp16, [1, 1280, 1, 448]> var_344_cast_fp16 = mul(x = var_47_cast_fp16_1, y = var_129_cast_fp16)[name = tensor<string, []>("op_344_cast_fp16")]; |
| 215 | tensor<fp16, [1, 1280, 1, 448]> key_5_cast_fp16 = add(x = var_342_cast_fp16, y = var_344_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")]; |
| 216 | tensor<fp16, [1, 1280, 1, 448]> var_346_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_126_cast_fp16)[name = tensor<string, []>("op_346_cast_fp16")]; |
| 217 | tensor<fp16, [1, 1280, 1, 448]> var_348_cast_fp16 = mul(x = var_54_cast_fp16_1, y = var_129_cast_fp16)[name = tensor<string, []>("op_348_cast_fp16")]; |
| 218 | tensor<fp16, [1, 1280, 1, 448]> value_5_cast_fp16 = add(x = var_346_cast_fp16, y = var_348_cast_fp16)[name = tensor<string, []>("value_5_cast_fp16")]; |
| 219 | tensor<int32, [4]> var_351 = const()[name = tensor<string, []>("op_351"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 220 | tensor<fp16, [1, 20, 64, 1]> mh_q_5_cast_fp16 = reshape(shape = var_351, x = query_5_cast_fp16)[name = tensor<string, []>("mh_q_5_cast_fp16")]; |
| 221 | tensor<fp16, []> var_353_to_fp16 = const()[name = tensor<string, []>("op_353_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| 222 | tensor<fp16, [1, 20, 64, 1]> var_354_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_353_to_fp16)[name = tensor<string, []>("op_354_cast_fp16")]; |
| 223 | tensor<int32, [4]> var_355 = const()[name = tensor<string, []>("op_355"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 224 | tensor<fp16, [1, 20, 64, 448]> var_356_cast_fp16 = reshape(shape = var_355, x = key_5_cast_fp16)[name = tensor<string, []>("op_356_cast_fp16")]; |
| 225 | tensor<bool, []> mh_w_7_transpose_x_0 = const()[name = tensor<string, []>("mh_w_7_transpose_x_0"), val = tensor<bool, []>(true)]; |
| 226 | tensor<bool, []> mh_w_7_transpose_y_0 = const()[name = tensor<string, []>("mh_w_7_transpose_y_0"), val = tensor<bool, []>(false)]; |
| 227 | 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_354_cast_fp16, y = var_356_cast_fp16)[name = tensor<string, []>("mh_w_7_cast_fp16")]; |
| 228 | tensor<fp16, [1, 20, 1, 448]> mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_147_cast_fp16)[name = tensor<string, []>("mh_w_9_cast_fp16")]; |
| 229 | tensor<fp16, [1, 20, 1, 448]> var_364_cast_fp16 = softmax(axis = var_278, x = mh_w_9_cast_fp16)[name = tensor<string, []>("op_364_cast_fp16")]; |
| 230 | tensor<int32, [4]> var_365 = const()[name = tensor<string, []>("op_365"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 231 | tensor<fp16, [1, 20, 64, 448]> var_366_cast_fp16 = reshape(shape = var_365, x = value_5_cast_fp16)[name = tensor<string, []>("op_366_cast_fp16")]; |
| 232 | tensor<bool, []> attn_5_transpose_x_0 = const()[name = tensor<string, []>("attn_5_transpose_x_0"), val = tensor<bool, []>(false)]; |
| 233 | tensor<bool, []> attn_5_transpose_y_0 = const()[name = tensor<string, []>("attn_5_transpose_y_0"), val = tensor<bool, []>(true)]; |
| 234 | 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_366_cast_fp16, y = var_364_cast_fp16)[name = tensor<string, []>("attn_5_cast_fp16")]; |
| 235 | tensor<int32, [4]> var_369 = const()[name = tensor<string, []>("op_369"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; |
| 236 | tensor<fp16, [1, 1280, 1, 1]> input_11_cast_fp16 = reshape(shape = var_369, x = attn_5_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")]; |
| 237 | tensor<string, []> obj_21_pad_type_0 = const()[name = tensor<string, []>("obj_21_pad_type_0"), val = tensor<string, []>("valid")]; |
| 238 | tensor<int32, [2]> obj_21_strides_0 = const()[name = tensor<string, []>("obj_21_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 239 | tensor<int32, [4]> obj_21_pad_0 = const()[name = tensor<string, []>("obj_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 240 | tensor<int32, [2]> obj_21_dilations_0 = const()[name = tensor<string, []>("obj_21_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 241 | tensor<int32, []> obj_21_groups_0 = const()[name = tensor<string, []>("obj_21_groups_0"), val = tensor<int32, []>(1)]; |
| 242 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196244224)))]; |
| 243 | tensor<fp16, [1280]> layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199521088)))]; |
| 244 | tensor<fp16, [1, 1280, 1, 1]> obj_21_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = obj_21_dilations_0, groups = obj_21_groups_0, pad = obj_21_pad_0, pad_type = obj_21_pad_type_0, strides = obj_21_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("obj_21_cast_fp16")]; |
| 245 | 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")]; |
| 246 | tensor<int32, [1]> out_9_axes_0 = const()[name = tensor<string, []>("out_9_axes_0"), val = tensor<int32, [1]>([1])]; |
| 247 | tensor<fp16, []> var_391_to_fp16 = const()[name = tensor<string, []>("op_391_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 248 | tensor<fp16, [1, 1280, 1, 1]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_391_to_fp16, x = inputs_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")]; |
| 249 | 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, []>(199523712)))]; |
| 250 | 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, []>(199526336)))]; |
| 251 | tensor<fp16, []> obj_23_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_23_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 252 | 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")]; |
| 253 | tensor<string, []> query_7_pad_type_0 = const()[name = tensor<string, []>("query_7_pad_type_0"), val = tensor<string, []>("valid")]; |
| 254 | tensor<int32, [2]> query_7_strides_0 = const()[name = tensor<string, []>("query_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 255 | tensor<int32, [4]> query_7_pad_0 = const()[name = tensor<string, []>("query_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 256 | tensor<int32, [2]> query_7_dilations_0 = const()[name = tensor<string, []>("query_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 257 | tensor<int32, []> query_7_groups_0 = const()[name = tensor<string, []>("query_7_groups_0"), val = tensor<int32, []>(1)]; |
| 258 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199528960)))]; |
| 259 | tensor<fp16, [1280]> layers_1_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202805824)))]; |
| 260 | tensor<fp16, [1, 1280, 1, 1]> query_7_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = query_7_dilations_0, groups = query_7_groups_0, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = query_7_strides_0, weight = layers_1_encoder_attn_q_proj_weight_to_fp16, x = obj_23_cast_fp16)[name = tensor<string, []>("query_7_cast_fp16")]; |
| 261 | tensor<string, []> key_7_pad_type_0 = const()[name = tensor<string, []>("key_7_pad_type_0"), val = tensor<string, []>("valid")]; |
| 262 | tensor<int32, [2]> key_7_strides_0 = const()[name = tensor<string, []>("key_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 263 | tensor<int32, [4]> key_7_pad_0 = const()[name = tensor<string, []>("key_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 264 | tensor<int32, [2]> key_7_dilations_0 = const()[name = tensor<string, []>("key_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 265 | tensor<int32, []> key_7_groups_0 = const()[name = tensor<string, []>("key_7_groups_0"), val = tensor<int32, []>(1)]; |
| 266 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202808448)))]; |
| 267 | tensor<fp16, [1, 1280, 1, 1500]> key_7_cast_fp16 = conv(dilations = key_7_dilations_0, groups = key_7_groups_0, pad = key_7_pad_0, pad_type = key_7_pad_type_0, strides = key_7_strides_0, weight = layers_1_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_7_cast_fp16")]; |
| 268 | tensor<string, []> value_7_pad_type_0 = const()[name = tensor<string, []>("value_7_pad_type_0"), val = tensor<string, []>("valid")]; |
| 269 | tensor<int32, [2]> value_7_strides_0 = const()[name = tensor<string, []>("value_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 270 | tensor<int32, [4]> value_7_pad_0 = const()[name = tensor<string, []>("value_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 271 | tensor<int32, [2]> value_7_dilations_0 = const()[name = tensor<string, []>("value_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 272 | tensor<int32, []> value_7_groups_0 = const()[name = tensor<string, []>("value_7_groups_0"), val = tensor<int32, []>(1)]; |
| 273 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206085312)))]; |
| 274 | tensor<fp16, [1280]> layers_1_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209362176)))]; |
| 275 | tensor<fp16, [1, 1280, 1, 1500]> value_7_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_bias_to_fp16, dilations = value_7_dilations_0, groups = value_7_groups_0, pad = value_7_pad_0, pad_type = value_7_pad_type_0, strides = value_7_strides_0, weight = layers_1_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_7_cast_fp16")]; |
| 276 | tensor<int32, [4]> var_426 = const()[name = tensor<string, []>("op_426"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 277 | tensor<fp16, [1, 20, 64, 1]> mh_q_7_cast_fp16 = reshape(shape = var_426, x = query_7_cast_fp16)[name = tensor<string, []>("mh_q_7_cast_fp16")]; |
| 278 | tensor<fp16, []> var_428_to_fp16 = const()[name = tensor<string, []>("op_428_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| 279 | tensor<fp16, [1, 20, 64, 1]> var_429_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_428_to_fp16)[name = tensor<string, []>("op_429_cast_fp16")]; |
| 280 | tensor<int32, [4]> var_430 = const()[name = tensor<string, []>("op_430"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 281 | tensor<fp16, [1, 20, 64, 1500]> var_431_cast_fp16 = reshape(shape = var_430, x = key_7_cast_fp16)[name = tensor<string, []>("op_431_cast_fp16")]; |
| 282 | tensor<bool, []> mh_w_11_transpose_x_0 = const()[name = tensor<string, []>("mh_w_11_transpose_x_0"), val = tensor<bool, []>(true)]; |
| 283 | tensor<bool, []> mh_w_11_transpose_y_0 = const()[name = tensor<string, []>("mh_w_11_transpose_y_0"), val = tensor<bool, []>(false)]; |
| 284 | 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_429_cast_fp16, y = var_431_cast_fp16)[name = tensor<string, []>("mh_w_11_cast_fp16")]; |
| 285 | tensor<fp16, [1, 20, 1, 1500]> obj_27_cast_fp16 = softmax(axis = var_278, x = mh_w_11_cast_fp16)[name = tensor<string, []>("obj_27_cast_fp16")]; |
| 286 | tensor<int32, [4]> var_435 = const()[name = tensor<string, []>("op_435"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 287 | tensor<fp16, [1, 20, 64, 1500]> var_436_cast_fp16 = reshape(shape = var_435, x = value_7_cast_fp16)[name = tensor<string, []>("op_436_cast_fp16")]; |
| 288 | tensor<bool, []> attn_7_transpose_x_0 = const()[name = tensor<string, []>("attn_7_transpose_x_0"), val = tensor<bool, []>(false)]; |
| 289 | tensor<bool, []> attn_7_transpose_y_0 = const()[name = tensor<string, []>("attn_7_transpose_y_0"), val = tensor<bool, []>(true)]; |
| 290 | 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_436_cast_fp16, y = obj_27_cast_fp16)[name = tensor<string, []>("attn_7_cast_fp16")]; |
| 291 | tensor<int32, [4]> var_439 = const()[name = tensor<string, []>("op_439"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; |
| 292 | tensor<fp16, [1, 1280, 1, 1]> input_13_cast_fp16 = reshape(shape = var_439, x = attn_7_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")]; |
| 293 | tensor<string, []> obj_25_pad_type_0 = const()[name = tensor<string, []>("obj_25_pad_type_0"), val = tensor<string, []>("valid")]; |
| 294 | tensor<int32, [2]> obj_25_strides_0 = const()[name = tensor<string, []>("obj_25_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 295 | tensor<int32, [4]> obj_25_pad_0 = const()[name = tensor<string, []>("obj_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 296 | tensor<int32, [2]> obj_25_dilations_0 = const()[name = tensor<string, []>("obj_25_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 297 | tensor<int32, []> obj_25_groups_0 = const()[name = tensor<string, []>("obj_25_groups_0"), val = tensor<int32, []>(1)]; |
| 298 | tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209364800)))]; |
| 299 | tensor<fp16, [1280]> layers_1_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212641664)))]; |
| 300 | tensor<fp16, [1, 1280, 1, 1]> obj_25_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = obj_25_dilations_0, groups = obj_25_groups_0, pad = obj_25_pad_0, pad_type = obj_25_pad_type_0, strides = obj_25_strides_0, weight = layers_1_encoder_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("obj_25_cast_fp16")]; |
| 301 | 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")]; |
| 302 | tensor<int32, [1]> out_11_axes_0 = const()[name = tensor<string, []>("out_11_axes_0"), val = tensor<int32, [1]>([1])]; |
| 303 | tensor<fp16, []> var_457_to_fp16 = const()[name = tensor<string, []>("op_457_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 304 | tensor<fp16, [1, 1280, 1, 1]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_457_to_fp16, x = inputs_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")]; |
| 305 | 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, []>(212644288)))]; |
| 306 | 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, []>(212646912)))]; |
| 307 | tensor<fp16, []> input_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 308 | 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")]; |
| 309 | tensor<string, []> input_17_pad_type_0 = const()[name = tensor<string, []>("input_17_pad_type_0"), val = tensor<string, []>("valid")]; |
| 310 | tensor<int32, [2]> input_17_strides_0 = const()[name = tensor<string, []>("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 311 | tensor<int32, [4]> input_17_pad_0 = const()[name = tensor<string, []>("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 312 | tensor<int32, [2]> input_17_dilations_0 = const()[name = tensor<string, []>("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 313 | tensor<int32, []> input_17_groups_0 = const()[name = tensor<string, []>("input_17_groups_0"), val = tensor<int32, []>(1)]; |
| 314 | tensor<fp16, [5120, 1280, 1, 1]> layers_1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212649536)))]; |
| 315 | tensor<fp16, [5120]> layers_1_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(225756800)))]; |
| 316 | tensor<fp16, [1, 5120, 1, 1]> input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_1_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")]; |
| 317 | tensor<string, []> input_19_mode_0 = const()[name = tensor<string, []>("input_19_mode_0"), val = tensor<string, []>("EXACT")]; |
| 318 | 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")]; |
| 319 | tensor<string, []> hidden_states_5_pad_type_0 = const()[name = tensor<string, []>("hidden_states_5_pad_type_0"), val = tensor<string, []>("valid")]; |
| 320 | tensor<int32, [2]> hidden_states_5_strides_0 = const()[name = tensor<string, []>("hidden_states_5_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 321 | tensor<int32, [4]> hidden_states_5_pad_0 = const()[name = tensor<string, []>("hidden_states_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 322 | tensor<int32, [2]> hidden_states_5_dilations_0 = const()[name = tensor<string, []>("hidden_states_5_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 323 | tensor<int32, []> hidden_states_5_groups_0 = const()[name = tensor<string, []>("hidden_states_5_groups_0"), val = tensor<int32, []>(1)]; |
| 324 | tensor<fp16, [1280, 5120, 1, 1]> layers_1_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(225767104)))]; |
| 325 | tensor<fp16, [1280]> layers_1_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238874368)))]; |
| 326 | tensor<fp16, [1, 1280, 1, 1]> hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = hidden_states_5_dilations_0, groups = hidden_states_5_groups_0, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = hidden_states_5_strides_0, weight = layers_1_fc2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("hidden_states_5_cast_fp16")]; |
| 327 | 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")]; |
| 328 | tensor<int32, []> var_492 = const()[name = tensor<string, []>("op_492"), val = tensor<int32, []>(3)]; |
| 329 | tensor<int32, [1]> out_13_axes_0 = const()[name = tensor<string, []>("out_13_axes_0"), val = tensor<int32, [1]>([1])]; |
| 330 | tensor<fp16, []> var_518_to_fp16 = const()[name = tensor<string, []>("op_518_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 331 | tensor<fp16, [1, 1280, 1, 1]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_518_to_fp16, x = inputs_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")]; |
| 332 | 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, []>(238876992)))]; |
| 333 | 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, []>(238879616)))]; |
| 334 | tensor<fp16, []> obj_29_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_29_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 335 | 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")]; |
| 336 | tensor<string, []> query_9_pad_type_0 = const()[name = tensor<string, []>("query_9_pad_type_0"), val = tensor<string, []>("valid")]; |
| 337 | tensor<int32, [2]> query_9_strides_0 = const()[name = tensor<string, []>("query_9_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 338 | tensor<int32, [4]> query_9_pad_0 = const()[name = tensor<string, []>("query_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 339 | tensor<int32, [2]> query_9_dilations_0 = const()[name = tensor<string, []>("query_9_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 340 | tensor<int32, []> query_9_groups_0 = const()[name = tensor<string, []>("query_9_groups_0"), val = tensor<int32, []>(1)]; |
| 341 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238882240)))]; |
| 342 | tensor<fp16, [1280]> layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(242159104)))]; |
| 343 | tensor<fp16, [1, 1280, 1, 1]> query_9_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = query_9_dilations_0, groups = query_9_groups_0, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = query_9_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("query_9_cast_fp16")]; |
| 344 | tensor<string, []> current_key_5_pad_type_0 = const()[name = tensor<string, []>("current_key_5_pad_type_0"), val = tensor<string, []>("valid")]; |
| 345 | tensor<int32, [2]> current_key_5_strides_0 = const()[name = tensor<string, []>("current_key_5_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 346 | tensor<int32, [4]> current_key_5_pad_0 = const()[name = tensor<string, []>("current_key_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 347 | tensor<int32, [2]> current_key_5_dilations_0 = const()[name = tensor<string, []>("current_key_5_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 348 | tensor<int32, []> current_key_5_groups_0 = const()[name = tensor<string, []>("current_key_5_groups_0"), val = tensor<int32, []>(1)]; |
| 349 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(242161728)))]; |
| 350 | tensor<fp16, [1, 1280, 1, 1]> current_key_5_cast_fp16 = conv(dilations = current_key_5_dilations_0, groups = current_key_5_groups_0, pad = current_key_5_pad_0, pad_type = current_key_5_pad_type_0, strides = current_key_5_strides_0, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("current_key_5_cast_fp16")]; |
| 351 | tensor<string, []> current_value_5_pad_type_0 = const()[name = tensor<string, []>("current_value_5_pad_type_0"), val = tensor<string, []>("valid")]; |
| 352 | tensor<int32, [2]> current_value_5_strides_0 = const()[name = tensor<string, []>("current_value_5_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 353 | tensor<int32, [4]> current_value_5_pad_0 = const()[name = tensor<string, []>("current_value_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 354 | tensor<int32, [2]> current_value_5_dilations_0 = const()[name = tensor<string, []>("current_value_5_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 355 | tensor<int32, []> current_value_5_groups_0 = const()[name = tensor<string, []>("current_value_5_groups_0"), val = tensor<int32, []>(1)]; |
| 356 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(245438592)))]; |
| 357 | tensor<fp16, [1280]> layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248715456)))]; |
| 358 | tensor<fp16, [1, 1280, 1, 1]> current_value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = current_value_5_dilations_0, groups = current_value_5_groups_0, pad = current_value_5_pad_0, pad_type = current_value_5_pad_type_0, strides = current_value_5_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("current_value_5_cast_fp16")]; |
| 359 | tensor<fp16, [1, 1280, 1, 448]> var_556_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_126_cast_fp16)[name = tensor<string, []>("op_556_cast_fp16")]; |
| 360 | tensor<fp16, [1, 1280, 1, 448]> var_558_cast_fp16 = mul(x = var_47_cast_fp16_2, y = var_129_cast_fp16)[name = tensor<string, []>("op_558_cast_fp16")]; |
| 361 | tensor<fp16, [1, 1280, 1, 448]> key_9_cast_fp16 = add(x = var_556_cast_fp16, y = var_558_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")]; |
| 362 | tensor<fp16, [1, 1280, 1, 448]> var_560_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_126_cast_fp16)[name = tensor<string, []>("op_560_cast_fp16")]; |
| 363 | tensor<fp16, [1, 1280, 1, 448]> var_562_cast_fp16 = mul(x = var_54_cast_fp16_2, y = var_129_cast_fp16)[name = tensor<string, []>("op_562_cast_fp16")]; |
| 364 | tensor<fp16, [1, 1280, 1, 448]> value_9_cast_fp16 = add(x = var_560_cast_fp16, y = var_562_cast_fp16)[name = tensor<string, []>("value_9_cast_fp16")]; |
| 365 | tensor<int32, [4]> var_565 = const()[name = tensor<string, []>("op_565"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 366 | tensor<fp16, [1, 20, 64, 1]> mh_q_9_cast_fp16 = reshape(shape = var_565, x = query_9_cast_fp16)[name = tensor<string, []>("mh_q_9_cast_fp16")]; |
| 367 | tensor<fp16, []> var_567_to_fp16 = const()[name = tensor<string, []>("op_567_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| 368 | tensor<fp16, [1, 20, 64, 1]> var_568_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_567_to_fp16)[name = tensor<string, []>("op_568_cast_fp16")]; |
| 369 | tensor<int32, [4]> var_569 = const()[name = tensor<string, []>("op_569"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 370 | tensor<fp16, [1, 20, 64, 448]> var_570_cast_fp16 = reshape(shape = var_569, x = key_9_cast_fp16)[name = tensor<string, []>("op_570_cast_fp16")]; |
| 371 | tensor<bool, []> mh_w_13_transpose_x_0 = const()[name = tensor<string, []>("mh_w_13_transpose_x_0"), val = tensor<bool, []>(true)]; |
| 372 | tensor<bool, []> mh_w_13_transpose_y_0 = const()[name = tensor<string, []>("mh_w_13_transpose_y_0"), val = tensor<bool, []>(false)]; |
| 373 | 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_568_cast_fp16, y = var_570_cast_fp16)[name = tensor<string, []>("mh_w_13_cast_fp16")]; |
| 374 | tensor<fp16, [1, 20, 1, 448]> mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_147_cast_fp16)[name = tensor<string, []>("mh_w_15_cast_fp16")]; |
| 375 | tensor<fp16, [1, 20, 1, 448]> var_578_cast_fp16 = softmax(axis = var_492, x = mh_w_15_cast_fp16)[name = tensor<string, []>("op_578_cast_fp16")]; |
| 376 | tensor<int32, [4]> var_579 = const()[name = tensor<string, []>("op_579"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 377 | tensor<fp16, [1, 20, 64, 448]> var_580_cast_fp16 = reshape(shape = var_579, x = value_9_cast_fp16)[name = tensor<string, []>("op_580_cast_fp16")]; |
| 378 | tensor<bool, []> attn_9_transpose_x_0 = const()[name = tensor<string, []>("attn_9_transpose_x_0"), val = tensor<bool, []>(false)]; |
| 379 | tensor<bool, []> attn_9_transpose_y_0 = const()[name = tensor<string, []>("attn_9_transpose_y_0"), val = tensor<bool, []>(true)]; |
| 380 | 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_580_cast_fp16, y = var_578_cast_fp16)[name = tensor<string, []>("attn_9_cast_fp16")]; |
| 381 | tensor<int32, [4]> var_583 = const()[name = tensor<string, []>("op_583"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; |
| 382 | tensor<fp16, [1, 1280, 1, 1]> input_21_cast_fp16 = reshape(shape = var_583, x = attn_9_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")]; |
| 383 | tensor<string, []> obj_35_pad_type_0 = const()[name = tensor<string, []>("obj_35_pad_type_0"), val = tensor<string, []>("valid")]; |
| 384 | tensor<int32, [2]> obj_35_strides_0 = const()[name = tensor<string, []>("obj_35_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 385 | tensor<int32, [4]> obj_35_pad_0 = const()[name = tensor<string, []>("obj_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 386 | tensor<int32, [2]> obj_35_dilations_0 = const()[name = tensor<string, []>("obj_35_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 387 | tensor<int32, []> obj_35_groups_0 = const()[name = tensor<string, []>("obj_35_groups_0"), val = tensor<int32, []>(1)]; |
| 388 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248718080)))]; |
| 389 | tensor<fp16, [1280]> layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251994944)))]; |
| 390 | tensor<fp16, [1, 1280, 1, 1]> obj_35_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = obj_35_dilations_0, groups = obj_35_groups_0, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = obj_35_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("obj_35_cast_fp16")]; |
| 391 | 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")]; |
| 392 | tensor<int32, [1]> out_15_axes_0 = const()[name = tensor<string, []>("out_15_axes_0"), val = tensor<int32, [1]>([1])]; |
| 393 | tensor<fp16, []> var_605_to_fp16 = const()[name = tensor<string, []>("op_605_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 394 | tensor<fp16, [1, 1280, 1, 1]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_605_to_fp16, x = inputs_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")]; |
| 395 | 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, []>(251997568)))]; |
| 396 | 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, []>(252000192)))]; |
| 397 | tensor<fp16, []> obj_37_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_37_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 398 | 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")]; |
| 399 | tensor<string, []> query_11_pad_type_0 = const()[name = tensor<string, []>("query_11_pad_type_0"), val = tensor<string, []>("valid")]; |
| 400 | tensor<int32, [2]> query_11_strides_0 = const()[name = tensor<string, []>("query_11_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 401 | tensor<int32, [4]> query_11_pad_0 = const()[name = tensor<string, []>("query_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 402 | tensor<int32, [2]> query_11_dilations_0 = const()[name = tensor<string, []>("query_11_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 403 | tensor<int32, []> query_11_groups_0 = const()[name = tensor<string, []>("query_11_groups_0"), val = tensor<int32, []>(1)]; |
| 404 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252002816)))]; |
| 405 | tensor<fp16, [1280]> layers_2_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255279680)))]; |
| 406 | tensor<fp16, [1, 1280, 1, 1]> query_11_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_bias_to_fp16, dilations = query_11_dilations_0, groups = query_11_groups_0, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = query_11_strides_0, weight = layers_2_encoder_attn_q_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor<string, []>("query_11_cast_fp16")]; |
| 407 | tensor<string, []> key_11_pad_type_0 = const()[name = tensor<string, []>("key_11_pad_type_0"), val = tensor<string, []>("valid")]; |
| 408 | tensor<int32, [2]> key_11_strides_0 = const()[name = tensor<string, []>("key_11_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 409 | tensor<int32, [4]> key_11_pad_0 = const()[name = tensor<string, []>("key_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 410 | tensor<int32, [2]> key_11_dilations_0 = const()[name = tensor<string, []>("key_11_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 411 | tensor<int32, []> key_11_groups_0 = const()[name = tensor<string, []>("key_11_groups_0"), val = tensor<int32, []>(1)]; |
| 412 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255282304)))]; |
| 413 | tensor<fp16, [1, 1280, 1, 1500]> key_11_cast_fp16 = conv(dilations = key_11_dilations_0, groups = key_11_groups_0, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = key_11_strides_0, weight = layers_2_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_11_cast_fp16")]; |
| 414 | tensor<string, []> value_11_pad_type_0 = const()[name = tensor<string, []>("value_11_pad_type_0"), val = tensor<string, []>("valid")]; |
| 415 | tensor<int32, [2]> value_11_strides_0 = const()[name = tensor<string, []>("value_11_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 416 | tensor<int32, [4]> value_11_pad_0 = const()[name = tensor<string, []>("value_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 417 | tensor<int32, [2]> value_11_dilations_0 = const()[name = tensor<string, []>("value_11_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 418 | tensor<int32, []> value_11_groups_0 = const()[name = tensor<string, []>("value_11_groups_0"), val = tensor<int32, []>(1)]; |
| 419 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258559168)))]; |
| 420 | tensor<fp16, [1280]> layers_2_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261836032)))]; |
| 421 | tensor<fp16, [1, 1280, 1, 1500]> value_11_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_bias_to_fp16, dilations = value_11_dilations_0, groups = value_11_groups_0, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = value_11_strides_0, weight = layers_2_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_11_cast_fp16")]; |
| 422 | tensor<int32, [4]> var_640 = const()[name = tensor<string, []>("op_640"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 423 | tensor<fp16, [1, 20, 64, 1]> mh_q_11_cast_fp16 = reshape(shape = var_640, x = query_11_cast_fp16)[name = tensor<string, []>("mh_q_11_cast_fp16")]; |
| 424 | tensor<fp16, []> var_642_to_fp16 = const()[name = tensor<string, []>("op_642_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| 425 | tensor<fp16, [1, 20, 64, 1]> var_643_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_642_to_fp16)[name = tensor<string, []>("op_643_cast_fp16")]; |
| 426 | tensor<int32, [4]> var_644 = const()[name = tensor<string, []>("op_644"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 427 | tensor<fp16, [1, 20, 64, 1500]> var_645_cast_fp16 = reshape(shape = var_644, x = key_11_cast_fp16)[name = tensor<string, []>("op_645_cast_fp16")]; |
| 428 | tensor<bool, []> mh_w_17_transpose_x_0 = const()[name = tensor<string, []>("mh_w_17_transpose_x_0"), val = tensor<bool, []>(true)]; |
| 429 | tensor<bool, []> mh_w_17_transpose_y_0 = const()[name = tensor<string, []>("mh_w_17_transpose_y_0"), val = tensor<bool, []>(false)]; |
| 430 | 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_643_cast_fp16, y = var_645_cast_fp16)[name = tensor<string, []>("mh_w_17_cast_fp16")]; |
| 431 | tensor<fp16, [1, 20, 1, 1500]> obj_41_cast_fp16 = softmax(axis = var_492, x = mh_w_17_cast_fp16)[name = tensor<string, []>("obj_41_cast_fp16")]; |
| 432 | tensor<int32, [4]> var_649 = const()[name = tensor<string, []>("op_649"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 433 | tensor<fp16, [1, 20, 64, 1500]> var_650_cast_fp16 = reshape(shape = var_649, x = value_11_cast_fp16)[name = tensor<string, []>("op_650_cast_fp16")]; |
| 434 | tensor<bool, []> attn_11_transpose_x_0 = const()[name = tensor<string, []>("attn_11_transpose_x_0"), val = tensor<bool, []>(false)]; |
| 435 | tensor<bool, []> attn_11_transpose_y_0 = const()[name = tensor<string, []>("attn_11_transpose_y_0"), val = tensor<bool, []>(true)]; |
| 436 | 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_650_cast_fp16, y = obj_41_cast_fp16)[name = tensor<string, []>("attn_11_cast_fp16")]; |
| 437 | tensor<int32, [4]> var_653 = const()[name = tensor<string, []>("op_653"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; |
| 438 | tensor<fp16, [1, 1280, 1, 1]> input_23_cast_fp16 = reshape(shape = var_653, x = attn_11_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")]; |
| 439 | tensor<string, []> obj_39_pad_type_0 = const()[name = tensor<string, []>("obj_39_pad_type_0"), val = tensor<string, []>("valid")]; |
| 440 | tensor<int32, [2]> obj_39_strides_0 = const()[name = tensor<string, []>("obj_39_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 441 | tensor<int32, [4]> obj_39_pad_0 = const()[name = tensor<string, []>("obj_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 442 | tensor<int32, [2]> obj_39_dilations_0 = const()[name = tensor<string, []>("obj_39_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 443 | tensor<int32, []> obj_39_groups_0 = const()[name = tensor<string, []>("obj_39_groups_0"), val = tensor<int32, []>(1)]; |
| 444 | tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261838656)))]; |
| 445 | tensor<fp16, [1280]> layers_2_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(265115520)))]; |
| 446 | tensor<fp16, [1, 1280, 1, 1]> obj_39_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_bias_to_fp16, dilations = obj_39_dilations_0, groups = obj_39_groups_0, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = obj_39_strides_0, weight = layers_2_encoder_attn_o_proj_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("obj_39_cast_fp16")]; |
| 447 | 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")]; |
| 448 | tensor<int32, [1]> out_17_axes_0 = const()[name = tensor<string, []>("out_17_axes_0"), val = tensor<int32, [1]>([1])]; |
| 449 | tensor<fp16, []> var_674_to_fp16 = const()[name = tensor<string, []>("op_674_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 450 | tensor<fp16, [1, 1280, 1, 1]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_674_to_fp16, x = inputs_17_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")]; |
| 451 | 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, []>(265118144)))]; |
| 452 | 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, []>(265120768)))]; |
| 453 | tensor<fp16, []> input_25_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_25_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 454 | 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")]; |
| 455 | tensor<string, []> input_27_pad_type_0 = const()[name = tensor<string, []>("input_27_pad_type_0"), val = tensor<string, []>("valid")]; |
| 456 | tensor<int32, [2]> input_27_strides_0 = const()[name = tensor<string, []>("input_27_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 457 | tensor<int32, [4]> input_27_pad_0 = const()[name = tensor<string, []>("input_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 458 | tensor<int32, [2]> input_27_dilations_0 = const()[name = tensor<string, []>("input_27_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 459 | tensor<int32, []> input_27_groups_0 = const()[name = tensor<string, []>("input_27_groups_0"), val = tensor<int32, []>(1)]; |
| 460 | tensor<fp16, [5120, 1280, 1, 1]> layers_2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(265123392)))]; |
| 461 | tensor<fp16, [5120]> layers_2_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278230656)))]; |
| 462 | tensor<fp16, [1, 5120, 1, 1]> input_27_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = input_27_dilations_0, groups = input_27_groups_0, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = input_27_strides_0, weight = layers_2_fc1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")]; |
| 463 | tensor<string, []> input_29_mode_0 = const()[name = tensor<string, []>("input_29_mode_0"), val = tensor<string, []>("EXACT")]; |
| 464 | 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")]; |
| 465 | tensor<string, []> hidden_states_7_pad_type_0 = const()[name = tensor<string, []>("hidden_states_7_pad_type_0"), val = tensor<string, []>("valid")]; |
| 466 | tensor<int32, [2]> hidden_states_7_strides_0 = const()[name = tensor<string, []>("hidden_states_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 467 | tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = tensor<string, []>("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 468 | tensor<int32, [2]> hidden_states_7_dilations_0 = const()[name = tensor<string, []>("hidden_states_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 469 | tensor<int32, []> hidden_states_7_groups_0 = const()[name = tensor<string, []>("hidden_states_7_groups_0"), val = tensor<int32, []>(1)]; |
| 470 | tensor<fp16, [1280, 5120, 1, 1]> layers_2_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278240960)))]; |
| 471 | tensor<fp16, [1280]> layers_2_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291348224)))]; |
| 472 | tensor<fp16, [1, 1280, 1, 1]> hidden_states_7_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_2_fc2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("hidden_states_7_cast_fp16")]; |
| 473 | 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")]; |
| 474 | tensor<int32, []> var_710 = const()[name = tensor<string, []>("op_710"), val = tensor<int32, []>(3)]; |
| 475 | tensor<int32, [1]> out_19_axes_0 = const()[name = tensor<string, []>("out_19_axes_0"), val = tensor<int32, [1]>([1])]; |
| 476 | tensor<fp16, []> var_736_to_fp16 = const()[name = tensor<string, []>("op_736_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 477 | tensor<fp16, [1, 1280, 1, 1]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_736_to_fp16, x = inputs_19_cast_fp16)[name = tensor<string, []>("out_19_cast_fp16")]; |
| 478 | 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, []>(291350848)))]; |
| 479 | 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, []>(291353472)))]; |
| 480 | tensor<fp16, []> obj_43_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_43_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 481 | 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")]; |
| 482 | tensor<string, []> query_13_pad_type_0 = const()[name = tensor<string, []>("query_13_pad_type_0"), val = tensor<string, []>("valid")]; |
| 483 | tensor<int32, [2]> query_13_strides_0 = const()[name = tensor<string, []>("query_13_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 484 | tensor<int32, [4]> query_13_pad_0 = const()[name = tensor<string, []>("query_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 485 | tensor<int32, [2]> query_13_dilations_0 = const()[name = tensor<string, []>("query_13_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 486 | tensor<int32, []> query_13_groups_0 = const()[name = tensor<string, []>("query_13_groups_0"), val = tensor<int32, []>(1)]; |
| 487 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291356096)))]; |
| 488 | tensor<fp16, [1280]> layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294632960)))]; |
| 489 | tensor<fp16, [1, 1280, 1, 1]> query_13_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = query_13_dilations_0, groups = query_13_groups_0, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = query_13_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("query_13_cast_fp16")]; |
| 490 | tensor<string, []> current_key_pad_type_0 = const()[name = tensor<string, []>("current_key_pad_type_0"), val = tensor<string, []>("valid")]; |
| 491 | tensor<int32, [2]> current_key_strides_0 = const()[name = tensor<string, []>("current_key_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 492 | tensor<int32, [4]> current_key_pad_0 = const()[name = tensor<string, []>("current_key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 493 | tensor<int32, [2]> current_key_dilations_0 = const()[name = tensor<string, []>("current_key_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 494 | tensor<int32, []> current_key_groups_0 = const()[name = tensor<string, []>("current_key_groups_0"), val = tensor<int32, []>(1)]; |
| 495 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294635584)))]; |
| 496 | tensor<fp16, [1, 1280, 1, 1]> current_key_cast_fp16 = conv(dilations = current_key_dilations_0, groups = current_key_groups_0, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = current_key_strides_0, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("current_key_cast_fp16")]; |
| 497 | tensor<string, []> current_value_pad_type_0 = const()[name = tensor<string, []>("current_value_pad_type_0"), val = tensor<string, []>("valid")]; |
| 498 | tensor<int32, [2]> current_value_strides_0 = const()[name = tensor<string, []>("current_value_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 499 | tensor<int32, [4]> current_value_pad_0 = const()[name = tensor<string, []>("current_value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 500 | tensor<int32, [2]> current_value_dilations_0 = const()[name = tensor<string, []>("current_value_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 501 | tensor<int32, []> current_value_groups_0 = const()[name = tensor<string, []>("current_value_groups_0"), val = tensor<int32, []>(1)]; |
| 502 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297912448)))]; |
| 503 | tensor<fp16, [1280]> layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(301189312)))]; |
| 504 | tensor<fp16, [1, 1280, 1, 1]> current_value_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = current_value_dilations_0, groups = current_value_groups_0, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = current_value_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("current_value_cast_fp16")]; |
| 505 | tensor<fp16, [1, 1280, 1, 448]> var_774_cast_fp16 = mul(x = current_key_cast_fp16, y = var_126_cast_fp16)[name = tensor<string, []>("op_774_cast_fp16")]; |
| 506 | tensor<fp16, [1, 1280, 1, 448]> var_776_cast_fp16 = mul(x = var_47_cast_fp16_3, y = var_129_cast_fp16)[name = tensor<string, []>("op_776_cast_fp16")]; |
| 507 | tensor<fp16, [1, 1280, 1, 448]> key_13_cast_fp16 = add(x = var_774_cast_fp16, y = var_776_cast_fp16)[name = tensor<string, []>("key_13_cast_fp16")]; |
| 508 | tensor<fp16, [1, 1280, 1, 448]> var_778_cast_fp16 = mul(x = current_value_cast_fp16, y = var_126_cast_fp16)[name = tensor<string, []>("op_778_cast_fp16")]; |
| 509 | tensor<fp16, [1, 1280, 1, 448]> var_780_cast_fp16 = mul(x = var_54_cast_fp16_3, y = var_129_cast_fp16)[name = tensor<string, []>("op_780_cast_fp16")]; |
| 510 | tensor<fp16, [1, 1280, 1, 448]> value_13_cast_fp16 = add(x = var_778_cast_fp16, y = var_780_cast_fp16)[name = tensor<string, []>("value_13_cast_fp16")]; |
| 511 | tensor<int32, [4]> var_783 = const()[name = tensor<string, []>("op_783"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 512 | tensor<fp16, [1, 20, 64, 1]> mh_q_13_cast_fp16 = reshape(shape = var_783, x = query_13_cast_fp16)[name = tensor<string, []>("mh_q_13_cast_fp16")]; |
| 513 | tensor<fp16, []> var_785_to_fp16 = const()[name = tensor<string, []>("op_785_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| 514 | tensor<fp16, [1, 20, 64, 1]> var_786_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_785_to_fp16)[name = tensor<string, []>("op_786_cast_fp16")]; |
| 515 | tensor<int32, [4]> var_787 = const()[name = tensor<string, []>("op_787"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 516 | tensor<fp16, [1, 20, 64, 448]> var_788_cast_fp16 = reshape(shape = var_787, x = key_13_cast_fp16)[name = tensor<string, []>("op_788_cast_fp16")]; |
| 517 | tensor<bool, []> mh_w_19_transpose_x_0 = const()[name = tensor<string, []>("mh_w_19_transpose_x_0"), val = tensor<bool, []>(true)]; |
| 518 | tensor<bool, []> mh_w_19_transpose_y_0 = const()[name = tensor<string, []>("mh_w_19_transpose_y_0"), val = tensor<bool, []>(false)]; |
| 519 | 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_786_cast_fp16, y = var_788_cast_fp16)[name = tensor<string, []>("mh_w_19_cast_fp16")]; |
| 520 | tensor<fp16, [1, 20, 1, 448]> mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_147_cast_fp16)[name = tensor<string, []>("mh_w_21_cast_fp16")]; |
| 521 | tensor<fp16, [1, 20, 1, 448]> var_796_cast_fp16 = softmax(axis = var_710, x = mh_w_21_cast_fp16)[name = tensor<string, []>("op_796_cast_fp16")]; |
| 522 | tensor<int32, [4]> var_797 = const()[name = tensor<string, []>("op_797"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 523 | tensor<fp16, [1, 20, 64, 448]> var_798_cast_fp16 = reshape(shape = var_797, x = value_13_cast_fp16)[name = tensor<string, []>("op_798_cast_fp16")]; |
| 524 | tensor<bool, []> attn_13_transpose_x_0 = const()[name = tensor<string, []>("attn_13_transpose_x_0"), val = tensor<bool, []>(false)]; |
| 525 | tensor<bool, []> attn_13_transpose_y_0 = const()[name = tensor<string, []>("attn_13_transpose_y_0"), val = tensor<bool, []>(true)]; |
| 526 | 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_798_cast_fp16, y = var_796_cast_fp16)[name = tensor<string, []>("attn_13_cast_fp16")]; |
| 527 | tensor<int32, [4]> var_801 = const()[name = tensor<string, []>("op_801"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; |
| 528 | tensor<fp16, [1, 1280, 1, 1]> input_31_cast_fp16 = reshape(shape = var_801, x = attn_13_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")]; |
| 529 | tensor<string, []> obj_49_pad_type_0 = const()[name = tensor<string, []>("obj_49_pad_type_0"), val = tensor<string, []>("valid")]; |
| 530 | tensor<int32, [2]> obj_49_strides_0 = const()[name = tensor<string, []>("obj_49_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 531 | tensor<int32, [4]> obj_49_pad_0 = const()[name = tensor<string, []>("obj_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 532 | tensor<int32, [2]> obj_49_dilations_0 = const()[name = tensor<string, []>("obj_49_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 533 | tensor<int32, []> obj_49_groups_0 = const()[name = tensor<string, []>("obj_49_groups_0"), val = tensor<int32, []>(1)]; |
| 534 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(301191936)))]; |
| 535 | tensor<fp16, [1280]> layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304468800)))]; |
| 536 | tensor<fp16, [1, 1280, 1, 1]> obj_49_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = obj_49_dilations_0, groups = obj_49_groups_0, pad = obj_49_pad_0, pad_type = obj_49_pad_type_0, strides = obj_49_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("obj_49_cast_fp16")]; |
| 537 | 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")]; |
| 538 | tensor<int32, [1]> out_21_axes_0 = const()[name = tensor<string, []>("out_21_axes_0"), val = tensor<int32, [1]>([1])]; |
| 539 | tensor<fp16, []> var_823_to_fp16 = const()[name = tensor<string, []>("op_823_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 540 | tensor<fp16, [1, 1280, 1, 1]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_823_to_fp16, x = inputs_21_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")]; |
| 541 | 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, []>(304471424)))]; |
| 542 | 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, []>(304474048)))]; |
| 543 | tensor<fp16, []> obj_51_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_51_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 544 | 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")]; |
| 545 | tensor<string, []> query_pad_type_0 = const()[name = tensor<string, []>("query_pad_type_0"), val = tensor<string, []>("valid")]; |
| 546 | tensor<int32, [2]> query_strides_0 = const()[name = tensor<string, []>("query_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 547 | tensor<int32, [4]> query_pad_0 = const()[name = tensor<string, []>("query_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 548 | tensor<int32, [2]> query_dilations_0 = const()[name = tensor<string, []>("query_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 549 | tensor<int32, []> query_groups_0 = const()[name = tensor<string, []>("query_groups_0"), val = tensor<int32, []>(1)]; |
| 550 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304476672)))]; |
| 551 | tensor<fp16, [1280]> layers_3_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307753536)))]; |
| 552 | tensor<fp16, [1, 1280, 1, 1]> query_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_bias_to_fp16, dilations = query_dilations_0, groups = query_groups_0, pad = query_pad_0, pad_type = query_pad_type_0, strides = query_strides_0, weight = layers_3_encoder_attn_q_proj_weight_to_fp16, x = obj_51_cast_fp16)[name = tensor<string, []>("query_cast_fp16")]; |
| 553 | tensor<string, []> key_pad_type_0 = const()[name = tensor<string, []>("key_pad_type_0"), val = tensor<string, []>("valid")]; |
| 554 | tensor<int32, [2]> key_strides_0 = const()[name = tensor<string, []>("key_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 555 | tensor<int32, [4]> key_pad_0 = const()[name = tensor<string, []>("key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 556 | tensor<int32, [2]> key_dilations_0 = const()[name = tensor<string, []>("key_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 557 | tensor<int32, []> key_groups_0 = const()[name = tensor<string, []>("key_groups_0"), val = tensor<int32, []>(1)]; |
| 558 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307756160)))]; |
| 559 | tensor<fp16, [1, 1280, 1, 1500]> key_cast_fp16 = conv(dilations = key_dilations_0, groups = key_groups_0, pad = key_pad_0, pad_type = key_pad_type_0, strides = key_strides_0, weight = layers_3_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_cast_fp16")]; |
| 560 | tensor<string, []> value_pad_type_0 = const()[name = tensor<string, []>("value_pad_type_0"), val = tensor<string, []>("valid")]; |
| 561 | tensor<int32, [2]> value_strides_0 = const()[name = tensor<string, []>("value_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 562 | tensor<int32, [4]> value_pad_0 = const()[name = tensor<string, []>("value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 563 | tensor<int32, [2]> value_dilations_0 = const()[name = tensor<string, []>("value_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 564 | tensor<int32, []> value_groups_0 = const()[name = tensor<string, []>("value_groups_0"), val = tensor<int32, []>(1)]; |
| 565 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311033024)))]; |
| 566 | tensor<fp16, [1280]> layers_3_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314309888)))]; |
| 567 | tensor<fp16, [1, 1280, 1, 1500]> value_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_bias_to_fp16, dilations = value_dilations_0, groups = value_groups_0, pad = value_pad_0, pad_type = value_pad_type_0, strides = value_strides_0, weight = layers_3_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_cast_fp16")]; |
| 568 | tensor<int32, [4]> var_858 = const()[name = tensor<string, []>("op_858"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 569 | tensor<fp16, [1, 20, 64, 1]> mh_q_cast_fp16 = reshape(shape = var_858, x = query_cast_fp16)[name = tensor<string, []>("mh_q_cast_fp16")]; |
| 570 | tensor<fp16, []> var_860_to_fp16 = const()[name = tensor<string, []>("op_860_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; |
| 571 | tensor<fp16, [1, 20, 64, 1]> var_861_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_860_to_fp16)[name = tensor<string, []>("op_861_cast_fp16")]; |
| 572 | tensor<int32, [4]> var_862 = const()[name = tensor<string, []>("op_862"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 573 | tensor<fp16, [1, 20, 64, 1500]> var_863_cast_fp16 = reshape(shape = var_862, x = key_cast_fp16)[name = tensor<string, []>("op_863_cast_fp16")]; |
| 574 | tensor<bool, []> mh_w_transpose_x_0 = const()[name = tensor<string, []>("mh_w_transpose_x_0"), val = tensor<bool, []>(true)]; |
| 575 | tensor<bool, []> mh_w_transpose_y_0 = const()[name = tensor<string, []>("mh_w_transpose_y_0"), val = tensor<bool, []>(false)]; |
| 576 | 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_861_cast_fp16, y = var_863_cast_fp16)[name = tensor<string, []>("mh_w_cast_fp16")]; |
| 577 | tensor<fp16, [1, 20, 1, 1500]> obj_55_cast_fp16 = softmax(axis = var_710, x = mh_w_cast_fp16)[name = tensor<string, []>("obj_55_cast_fp16")]; |
| 578 | tensor<int32, [4]> var_867 = const()[name = tensor<string, []>("op_867"), val = tensor<int32, [4]>([1, 20, 64, -1])]; |
| 579 | tensor<fp16, [1, 20, 64, 1500]> var_868_cast_fp16 = reshape(shape = var_867, x = value_cast_fp16)[name = tensor<string, []>("op_868_cast_fp16")]; |
| 580 | tensor<bool, []> attn_transpose_x_0 = const()[name = tensor<string, []>("attn_transpose_x_0"), val = tensor<bool, []>(false)]; |
| 581 | tensor<bool, []> attn_transpose_y_0 = const()[name = tensor<string, []>("attn_transpose_y_0"), val = tensor<bool, []>(true)]; |
| 582 | tensor<fp16, [1, 20, 64, 1]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_868_cast_fp16, y = obj_55_cast_fp16)[name = tensor<string, []>("attn_cast_fp16")]; |
| 583 | tensor<int32, [4]> var_871 = const()[name = tensor<string, []>("op_871"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; |
| 584 | tensor<fp16, [1, 1280, 1, 1]> input_33_cast_fp16 = reshape(shape = var_871, x = attn_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")]; |
| 585 | tensor<string, []> obj_53_pad_type_0 = const()[name = tensor<string, []>("obj_53_pad_type_0"), val = tensor<string, []>("valid")]; |
| 586 | tensor<int32, [2]> obj_53_strides_0 = const()[name = tensor<string, []>("obj_53_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 587 | tensor<int32, [4]> obj_53_pad_0 = const()[name = tensor<string, []>("obj_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 588 | tensor<int32, [2]> obj_53_dilations_0 = const()[name = tensor<string, []>("obj_53_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 589 | tensor<int32, []> obj_53_groups_0 = const()[name = tensor<string, []>("obj_53_groups_0"), val = tensor<int32, []>(1)]; |
| 590 | tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314312512)))]; |
| 591 | tensor<fp16, [1280]> layers_3_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317589376)))]; |
| 592 | tensor<fp16, [1, 1280, 1, 1]> obj_53_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_bias_to_fp16, dilations = obj_53_dilations_0, groups = obj_53_groups_0, pad = obj_53_pad_0, pad_type = obj_53_pad_type_0, strides = obj_53_strides_0, weight = layers_3_encoder_attn_o_proj_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("obj_53_cast_fp16")]; |
| 593 | 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")]; |
| 594 | tensor<int32, [1]> out_23_axes_0 = const()[name = tensor<string, []>("out_23_axes_0"), val = tensor<int32, [1]>([1])]; |
| 595 | tensor<fp16, []> var_892_to_fp16 = const()[name = tensor<string, []>("op_892_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 596 | tensor<fp16, [1, 1280, 1, 1]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_892_to_fp16, x = inputs_23_cast_fp16)[name = tensor<string, []>("out_23_cast_fp16")]; |
| 597 | 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, []>(317592000)))]; |
| 598 | 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, []>(317594624)))]; |
| 599 | tensor<fp16, []> input_35_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_35_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 600 | 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")]; |
| 601 | tensor<string, []> input_37_pad_type_0 = const()[name = tensor<string, []>("input_37_pad_type_0"), val = tensor<string, []>("valid")]; |
| 602 | tensor<int32, [2]> input_37_strides_0 = const()[name = tensor<string, []>("input_37_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 603 | tensor<int32, [4]> input_37_pad_0 = const()[name = tensor<string, []>("input_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 604 | tensor<int32, [2]> input_37_dilations_0 = const()[name = tensor<string, []>("input_37_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 605 | tensor<int32, []> input_37_groups_0 = const()[name = tensor<string, []>("input_37_groups_0"), val = tensor<int32, []>(1)]; |
| 606 | tensor<fp16, [5120, 1280, 1, 1]> layers_3_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317597248)))]; |
| 607 | tensor<fp16, [5120]> layers_3_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330704512)))]; |
| 608 | tensor<fp16, [1, 5120, 1, 1]> input_37_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = input_37_dilations_0, groups = input_37_groups_0, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = input_37_strides_0, weight = layers_3_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")]; |
| 609 | tensor<string, []> input_mode_0 = const()[name = tensor<string, []>("input_mode_0"), val = tensor<string, []>("EXACT")]; |
| 610 | 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")]; |
| 611 | tensor<string, []> hidden_states_9_pad_type_0 = const()[name = tensor<string, []>("hidden_states_9_pad_type_0"), val = tensor<string, []>("valid")]; |
| 612 | tensor<int32, [2]> hidden_states_9_strides_0 = const()[name = tensor<string, []>("hidden_states_9_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| 613 | tensor<int32, [4]> hidden_states_9_pad_0 = const()[name = tensor<string, []>("hidden_states_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 614 | tensor<int32, [2]> hidden_states_9_dilations_0 = const()[name = tensor<string, []>("hidden_states_9_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| 615 | tensor<int32, []> hidden_states_9_groups_0 = const()[name = tensor<string, []>("hidden_states_9_groups_0"), val = tensor<int32, []>(1)]; |
| 616 | tensor<fp16, [1280, 5120, 1, 1]> layers_3_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330714816)))]; |
| 617 | tensor<fp16, [1280]> layers_3_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(343822080)))]; |
| 618 | tensor<fp16, [1, 1280, 1, 1]> hidden_states_9_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = hidden_states_9_dilations_0, groups = hidden_states_9_groups_0, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = hidden_states_9_strides_0, weight = layers_3_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("hidden_states_9_cast_fp16")]; |
| 619 | 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")]; |
| 620 | tensor<int32, [1]> out_axes_0 = const()[name = tensor<string, []>("out_axes_0"), val = tensor<int32, [1]>([1])]; |
| 621 | tensor<fp16, []> var_935_to_fp16 = const()[name = tensor<string, []>("op_935_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 622 | tensor<fp16, [1, 1280, 1, 1]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_935_to_fp16, x = inputs_cast_fp16)[name = tensor<string, []>("out_cast_fp16")]; |
| 623 | 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, []>(343824704)))]; |
| 624 | 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, []>(343827328)))]; |
| 625 | tensor<fp16, []> hidden_states_epsilon_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; |
| 626 | 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")]; |
| 627 | tensor<int32, [1]> var_946_axes_0 = const()[name = tensor<string, []>("op_946_axes_0"), val = tensor<int32, [1]>([2])]; |
| 628 | tensor<fp16, [1, 1280, 1]> var_946_cast_fp16 = squeeze(axes = var_946_axes_0, x = hidden_states_cast_fp16)[name = tensor<string, []>("op_946_cast_fp16")]; |
| 629 | tensor<int32, [3]> var_949_perm_0 = const()[name = tensor<string, []>("op_949_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| 630 | 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, []>(343829952)))]; |
| 631 | tensor<fp16, [1, 1, 1280]> var_949_cast_fp16 = transpose(perm = var_949_perm_0, x = var_946_cast_fp16)[name = tensor<string, []>("transpose_0")]; |
| 632 | tensor<fp16, [1, 1, 51866]> logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = var_949_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")]; |
| 633 | tensor<int32, []> var_953 = const()[name = tensor<string, []>("op_953"), val = tensor<int32, []>(1)]; |
| 634 | tensor<bool, []> obj_59_interleave_0 = const()[name = tensor<string, []>("obj_59_interleave_0"), val = tensor<bool, []>(false)]; |
| 635 | tensor<fp16, [1, 5120, 1, 1]> key_cache_updates = concat(axis = var_953, 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")]; |
| 636 | tensor<int32, []> var_956 = const()[name = tensor<string, []>("op_956"), val = tensor<int32, []>(1)]; |
| 637 | tensor<bool, []> obj_61_interleave_0 = const()[name = tensor<string, []>("obj_61_interleave_0"), val = tensor<bool, []>(false)]; |
| 638 | tensor<fp16, [1, 5120, 1, 1]> value_cache_updates = concat(axis = var_956, 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")]; |
| 639 | tensor<int32, [4]> var_967_begin_0 = const()[name = tensor<string, []>("op_967_begin_0"), val = tensor<int32, [4]>([0, 4, 0, 0])]; |
| 640 | tensor<int32, [4]> var_967_end_0 = const()[name = tensor<string, []>("op_967_end_0"), val = tensor<int32, [4]>([1, 5, 1, 1500])]; |
| 641 | tensor<bool, [4]> var_967_end_mask_0 = const()[name = tensor<string, []>("op_967_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| 642 | tensor<fp16, [1, 1, 1, 1500]> var_967_cast_fp16 = slice_by_index(begin = var_967_begin_0, end = var_967_end_0, end_mask = var_967_end_mask_0, x = obj_41_cast_fp16)[name = tensor<string, []>("op_967_cast_fp16")]; |
| 643 | tensor<int32, [4]> var_970_begin_0 = const()[name = tensor<string, []>("op_970_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 644 | tensor<int32, [4]> var_970_end_0 = const()[name = tensor<string, []>("op_970_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| 645 | tensor<bool, [4]> var_970_end_mask_0 = const()[name = tensor<string, []>("op_970_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| 646 | tensor<bool, [4]> var_970_squeeze_mask_0 = const()[name = tensor<string, []>("op_970_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| 647 | tensor<fp16, [1, 1, 1500]> var_970_cast_fp16 = slice_by_index(begin = var_970_begin_0, end = var_970_end_0, end_mask = var_970_end_mask_0, squeeze_mask = var_970_squeeze_mask_0, x = var_967_cast_fp16)[name = tensor<string, []>("op_970_cast_fp16")]; |
| 648 | tensor<int32, [4]> var_985_begin_0 = const()[name = tensor<string, []>("op_985_begin_0"), val = tensor<int32, [4]>([0, 11, 0, 0])]; |
| 649 | tensor<int32, [4]> var_985_end_0 = const()[name = tensor<string, []>("op_985_end_0"), val = tensor<int32, [4]>([1, 12, 1, 1500])]; |
| 650 | tensor<bool, [4]> var_985_end_mask_0 = const()[name = tensor<string, []>("op_985_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| 651 | tensor<fp16, [1, 1, 1, 1500]> var_985_cast_fp16 = slice_by_index(begin = var_985_begin_0, end = var_985_end_0, end_mask = var_985_end_mask_0, x = obj_41_cast_fp16)[name = tensor<string, []>("op_985_cast_fp16")]; |
| 652 | tensor<int32, [4]> var_988_begin_0 = const()[name = tensor<string, []>("op_988_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 653 | tensor<int32, [4]> var_988_end_0 = const()[name = tensor<string, []>("op_988_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| 654 | tensor<bool, [4]> var_988_end_mask_0 = const()[name = tensor<string, []>("op_988_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| 655 | tensor<bool, [4]> var_988_squeeze_mask_0 = const()[name = tensor<string, []>("op_988_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| 656 | tensor<fp16, [1, 1, 1500]> var_988_cast_fp16 = slice_by_index(begin = var_988_begin_0, end = var_988_end_0, end_mask = var_988_end_mask_0, squeeze_mask = var_988_squeeze_mask_0, x = var_985_cast_fp16)[name = tensor<string, []>("op_988_cast_fp16")]; |
| 657 | tensor<int32, [4]> var_1003_begin_0 = const()[name = tensor<string, []>("op_1003_begin_0"), val = tensor<int32, [4]>([0, 3, 0, 0])]; |
| 658 | tensor<int32, [4]> var_1003_end_0 = const()[name = tensor<string, []>("op_1003_end_0"), val = tensor<int32, [4]>([1, 4, 1, 1500])]; |
| 659 | tensor<bool, [4]> var_1003_end_mask_0 = const()[name = tensor<string, []>("op_1003_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| 660 | tensor<fp16, [1, 1, 1, 1500]> var_1003_cast_fp16 = slice_by_index(begin = var_1003_begin_0, end = var_1003_end_0, end_mask = var_1003_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1003_cast_fp16")]; |
| 661 | tensor<int32, [4]> var_1006_begin_0 = const()[name = tensor<string, []>("op_1006_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 662 | tensor<int32, [4]> var_1006_end_0 = const()[name = tensor<string, []>("op_1006_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| 663 | tensor<bool, [4]> var_1006_end_mask_0 = const()[name = tensor<string, []>("op_1006_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| 664 | tensor<bool, [4]> var_1006_squeeze_mask_0 = const()[name = tensor<string, []>("op_1006_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| 665 | tensor<fp16, [1, 1, 1500]> var_1006_cast_fp16 = slice_by_index(begin = var_1006_begin_0, end = var_1006_end_0, end_mask = var_1006_end_mask_0, squeeze_mask = var_1006_squeeze_mask_0, x = var_1003_cast_fp16)[name = tensor<string, []>("op_1006_cast_fp16")]; |
| 666 | tensor<int32, [4]> var_1021_begin_0 = const()[name = tensor<string, []>("op_1021_begin_0"), val = tensor<int32, [4]>([0, 6, 0, 0])]; |
| 667 | tensor<int32, [4]> var_1021_end_0 = const()[name = tensor<string, []>("op_1021_end_0"), val = tensor<int32, [4]>([1, 7, 1, 1500])]; |
| 668 | tensor<bool, [4]> var_1021_end_mask_0 = const()[name = tensor<string, []>("op_1021_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| 669 | tensor<fp16, [1, 1, 1, 1500]> var_1021_cast_fp16 = slice_by_index(begin = var_1021_begin_0, end = var_1021_end_0, end_mask = var_1021_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1021_cast_fp16")]; |
| 670 | tensor<int32, [4]> var_1024_begin_0 = const()[name = tensor<string, []>("op_1024_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 671 | tensor<int32, [4]> var_1024_end_0 = const()[name = tensor<string, []>("op_1024_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| 672 | tensor<bool, [4]> var_1024_end_mask_0 = const()[name = tensor<string, []>("op_1024_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| 673 | tensor<bool, [4]> var_1024_squeeze_mask_0 = const()[name = tensor<string, []>("op_1024_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| 674 | tensor<fp16, [1, 1, 1500]> var_1024_cast_fp16 = slice_by_index(begin = var_1024_begin_0, end = var_1024_end_0, end_mask = var_1024_end_mask_0, squeeze_mask = var_1024_squeeze_mask_0, x = var_1021_cast_fp16)[name = tensor<string, []>("op_1024_cast_fp16")]; |
| 675 | tensor<int32, [4]> var_1039_begin_0 = const()[name = tensor<string, []>("op_1039_begin_0"), val = tensor<int32, [4]>([0, 11, 0, 0])]; |
| 676 | tensor<int32, [4]> var_1039_end_0 = const()[name = tensor<string, []>("op_1039_end_0"), val = tensor<int32, [4]>([1, 12, 1, 1500])]; |
| 677 | tensor<bool, [4]> var_1039_end_mask_0 = const()[name = tensor<string, []>("op_1039_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| 678 | tensor<fp16, [1, 1, 1, 1500]> var_1039_cast_fp16 = slice_by_index(begin = var_1039_begin_0, end = var_1039_end_0, end_mask = var_1039_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1039_cast_fp16")]; |
| 679 | tensor<int32, [4]> var_1042_begin_0 = const()[name = tensor<string, []>("op_1042_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 680 | tensor<int32, [4]> var_1042_end_0 = const()[name = tensor<string, []>("op_1042_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| 681 | tensor<bool, [4]> var_1042_end_mask_0 = const()[name = tensor<string, []>("op_1042_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| 682 | tensor<bool, [4]> var_1042_squeeze_mask_0 = const()[name = tensor<string, []>("op_1042_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| 683 | tensor<fp16, [1, 1, 1500]> var_1042_cast_fp16 = slice_by_index(begin = var_1042_begin_0, end = var_1042_end_0, end_mask = var_1042_end_mask_0, squeeze_mask = var_1042_squeeze_mask_0, x = var_1039_cast_fp16)[name = tensor<string, []>("op_1042_cast_fp16")]; |
| 684 | tensor<int32, [4]> var_1057_begin_0 = const()[name = tensor<string, []>("op_1057_begin_0"), val = tensor<int32, [4]>([0, 14, 0, 0])]; |
| 685 | tensor<int32, [4]> var_1057_end_0 = const()[name = tensor<string, []>("op_1057_end_0"), val = tensor<int32, [4]>([1, 15, 1, 1500])]; |
| 686 | tensor<bool, [4]> var_1057_end_mask_0 = const()[name = tensor<string, []>("op_1057_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; |
| 687 | tensor<fp16, [1, 1, 1, 1500]> var_1057_cast_fp16 = slice_by_index(begin = var_1057_begin_0, end = var_1057_end_0, end_mask = var_1057_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1057_cast_fp16")]; |
| 688 | tensor<int32, [4]> var_1060_begin_0 = const()[name = tensor<string, []>("op_1060_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| 689 | tensor<int32, [4]> var_1060_end_0 = const()[name = tensor<string, []>("op_1060_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; |
| 690 | tensor<bool, [4]> var_1060_end_mask_0 = const()[name = tensor<string, []>("op_1060_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; |
| 691 | tensor<bool, [4]> var_1060_squeeze_mask_0 = const()[name = tensor<string, []>("op_1060_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; |
| 692 | tensor<fp16, [1, 1, 1500]> var_1060_cast_fp16 = slice_by_index(begin = var_1060_begin_0, end = var_1060_end_0, end_mask = var_1060_end_mask_0, squeeze_mask = var_1060_squeeze_mask_0, x = var_1057_cast_fp16)[name = tensor<string, []>("op_1060_cast_fp16")]; |
| 693 | tensor<int32, []> var_1067 = const()[name = tensor<string, []>("op_1067"), val = tensor<int32, []>(1)]; |
| 694 | tensor<bool, []> var_1068_interleave_0 = const()[name = tensor<string, []>("op_1068_interleave_0"), val = tensor<bool, []>(false)]; |
| 695 | tensor<fp16, [1, 6, 1500]> var_1068_cast_fp16 = concat(axis = var_1067, interleave = var_1068_interleave_0, values = (var_970_cast_fp16, var_988_cast_fp16, var_1006_cast_fp16, var_1024_cast_fp16, var_1042_cast_fp16, var_1060_cast_fp16))[name = tensor<string, []>("op_1068_cast_fp16")]; |
| 696 | tensor<bool, []> var_1071 = const()[name = tensor<string, []>("op_1071"), val = tensor<bool, []>(false)]; |
| 697 | tensor<int32, [1]> obj_axes_0 = const()[name = tensor<string, []>("obj_axes_0"), val = tensor<int32, [1]>([1])]; |
| 698 | tensor<fp16, [1, 1500]> alignment_heads_weights = reduce_mean(axes = obj_axes_0, keep_dims = var_1071, x = var_1068_cast_fp16)[name = tensor<string, []>("obj_cast_fp16")]; |
| 699 | } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights); |
| 700 | } |