Upload model distil-whisper_distil-large-v3_turbo_600MB - QLoRA compressed encoder variant
95db02e
verified
| program(1.0) | |
| [buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}, {"coremltools-component-torch", "2.2.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "7.1"}})] | |
| { | |
| func main<ios16>(tensor<int32, [1]> cache_length, tensor<fp16, [1, 224]> decoder_key_padding_mask, tensor<fp16, [1, 1280, 1, 1500]> encoder_output_embeds, tensor<int32, [1]> input_ids, tensor<fp16, [1, 2560, 1, 224]> key_cache, tensor<fp16, [1, 224]> kv_cache_update_mask, tensor<fp16, [1, 2560, 1, 224]> value_cache) { | |
| tensor<int32, []> var_20_axis_0 = const()[name = tensor<string, []>("op_20_axis_0"), val = tensor<int32, []>(0)]; | |
| tensor<int32, []> var_20_batch_dims_0 = const()[name = tensor<string, []>("op_20_batch_dims_0"), val = tensor<int32, []>(0)]; | |
| 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)))]; | |
| tensor<fp16, [1, 1280]> var_20_cast_fp16 = gather(axis = var_20_axis_0, batch_dims = var_20_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor<string, []>("op_20_cast_fp16")]; | |
| tensor<int32, []> var_24_axis_0 = const()[name = tensor<string, []>("op_24_axis_0"), val = tensor<int32, []>(0)]; | |
| tensor<int32, []> var_24_batch_dims_0 = const()[name = tensor<string, []>("op_24_batch_dims_0"), val = tensor<int32, []>(0)]; | |
| 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)))]; | |
| tensor<fp16, [1, 1280]> var_24_cast_fp16 = gather(axis = var_24_axis_0, batch_dims = var_24_batch_dims_0, indices = cache_length, x = embed_positions_weight_to_fp16)[name = tensor<string, []>("op_24_cast_fp16")]; | |
| tensor<fp16, [1, 1280]> hidden_states_1_cast_fp16 = add(x = var_20_cast_fp16, y = var_24_cast_fp16)[name = tensor<string, []>("hidden_states_1_cast_fp16")]; | |
| tensor<int32, [1]> var_38_axes_0 = const()[name = tensor<string, []>("op_38_axes_0"), val = tensor<int32, [1]>([2])]; | |
| tensor<fp16, [1, 1280, 1]> var_38_cast_fp16 = expand_dims(axes = var_38_axes_0, x = hidden_states_1_cast_fp16)[name = tensor<string, []>("op_38_cast_fp16")]; | |
| tensor<int32, [1]> inputs_1_axes_0 = const()[name = tensor<string, []>("inputs_1_axes_0"), val = tensor<int32, [1]>([3])]; | |
| tensor<fp16, [1, 1280, 1, 1]> inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_38_cast_fp16)[name = tensor<string, []>("inputs_1_cast_fp16")]; | |
| tensor<int32, [2]> tile_0 = const()[name = tensor<string, []>("tile_0"), val = tensor<int32, [2]>([1280, 1280])]; | |
| tensor<int32, []> var_43_axis_0 = const()[name = tensor<string, []>("op_43_axis_0"), val = tensor<int32, []>(1)]; | |
| tensor<fp16, [1, 1280, 1, 224]> var_43_cast_fp16_0, tensor<fp16, [1, 1280, 1, 224]> var_43_cast_fp16_1 = split(axis = var_43_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor<string, []>("op_43_cast_fp16")]; | |
| tensor<int32, [2]> tile_1 = const()[name = tensor<string, []>("tile_1"), val = tensor<int32, [2]>([1280, 1280])]; | |
| tensor<int32, []> var_48_axis_0 = const()[name = tensor<string, []>("op_48_axis_0"), val = tensor<int32, []>(1)]; | |
| tensor<fp16, [1, 1280, 1, 224]> var_48_cast_fp16_0, tensor<fp16, [1, 1280, 1, 224]> var_48_cast_fp16_1 = split(axis = var_48_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor<string, []>("op_48_cast_fp16")]; | |
| tensor<int32, []> var_56 = const()[name = tensor<string, []>("op_56"), val = tensor<int32, []>(3)]; | |
| tensor<int32, []> var_63 = const()[name = tensor<string, []>("op_63"), val = tensor<int32, []>(1)]; | |
| tensor<bool, []> var_64 = const()[name = tensor<string, []>("op_64"), val = tensor<bool, []>(true)]; | |
| tensor<int32, [1]> var_76 = const()[name = tensor<string, []>("op_76"), val = tensor<int32, [1]>([1])]; | |
| tensor<fp16, [1, 1, 1, 1]> channels_mean_1_cast_fp16 = reduce_mean(axes = var_76, keep_dims = var_64, x = inputs_1_cast_fp16)[name = tensor<string, []>("channels_mean_1_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> zero_mean_1_cast_fp16 = sub(x = inputs_1_cast_fp16, y = channels_mean_1_cast_fp16)[name = tensor<string, []>("zero_mean_1_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = zero_mean_1_cast_fp16)[name = tensor<string, []>("zero_mean_sq_1_cast_fp16")]; | |
| tensor<int32, [1]> var_80 = const()[name = tensor<string, []>("op_80"), val = tensor<int32, [1]>([1])]; | |
| tensor<fp16, [1, 1, 1, 1]> var_81_cast_fp16 = reduce_mean(axes = var_80, keep_dims = var_64, x = zero_mean_sq_1_cast_fp16)[name = tensor<string, []>("op_81_cast_fp16")]; | |
| tensor<fp16, []> var_82_to_fp16 = const()[name = tensor<string, []>("op_82_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; | |
| tensor<fp16, [1, 1, 1, 1]> var_83_cast_fp16 = add(x = var_81_cast_fp16, y = var_82_to_fp16)[name = tensor<string, []>("op_83_cast_fp16")]; | |
| tensor<fp16, []> denom_1_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_1_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)]; | |
| tensor<fp16, [1, 1, 1, 1]> denom_1_cast_fp16 = rsqrt(epsilon = denom_1_epsilon_0_to_fp16, x = var_83_cast_fp16)[name = tensor<string, []>("denom_1_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> out_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = denom_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")]; | |
| 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)))]; | |
| 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)))]; | |
| 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)))]; | |
| 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)))]; | |
| tensor<fp16, []> obj_1_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_1_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; | |
| 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")]; | |
| tensor<int32, [2]> var_98 = const()[name = tensor<string, []>("op_98"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [2]> var_100 = const()[name = tensor<string, []>("op_100"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<string, []> query_1_pad_type_0 = const()[name = tensor<string, []>("query_1_pad_type_0"), val = tensor<string, []>("custom")]; | |
| tensor<int32, [4]> query_1_pad_0 = const()[name = tensor<string, []>("query_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| 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)))]; | |
| 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)))]; | |
| tensor<fp16, [1, 1280, 1, 1]> query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = var_100, groups = var_63, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = var_98, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("query_1_cast_fp16")]; | |
| tensor<int32, [2]> var_104 = const()[name = tensor<string, []>("op_104"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [2]> var_106 = const()[name = tensor<string, []>("op_106"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<string, []> current_key_1_pad_type_0 = const()[name = tensor<string, []>("current_key_1_pad_type_0"), val = tensor<string, []>("custom")]; | |
| 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])]; | |
| 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)))]; | |
| tensor<fp16, [1, 1280, 1, 1]> current_key_1_cast_fp16 = conv(dilations = var_106, groups = var_63, pad = current_key_1_pad_0, pad_type = current_key_1_pad_type_0, strides = var_104, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("current_key_1_cast_fp16")]; | |
| tensor<int32, [2]> var_111 = const()[name = tensor<string, []>("op_111"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [2]> var_113 = const()[name = tensor<string, []>("op_113"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<string, []> current_value_1_pad_type_0 = const()[name = tensor<string, []>("current_value_1_pad_type_0"), val = tensor<string, []>("custom")]; | |
| 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])]; | |
| 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)))]; | |
| 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)))]; | |
| tensor<fp16, [1, 1280, 1, 1]> current_value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = var_113, groups = var_63, pad = current_value_1_pad_0, pad_type = current_value_1_pad_type_0, strides = var_111, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("current_value_1_cast_fp16")]; | |
| tensor<int32, [1]> var_117_axes_0 = const()[name = tensor<string, []>("op_117_axes_0"), val = tensor<int32, [1]>([1])]; | |
| tensor<fp16, [1, 1, 224]> var_117_cast_fp16 = expand_dims(axes = var_117_axes_0, x = kv_cache_update_mask)[name = tensor<string, []>("op_117_cast_fp16")]; | |
| tensor<int32, [1]> var_118_axes_0 = const()[name = tensor<string, []>("op_118_axes_0"), val = tensor<int32, [1]>([2])]; | |
| tensor<fp16, [1, 1, 1, 224]> var_118_cast_fp16 = expand_dims(axes = var_118_axes_0, x = var_117_cast_fp16)[name = tensor<string, []>("op_118_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 224]> var_120_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_118_cast_fp16)[name = tensor<string, []>("op_120_cast_fp16")]; | |
| tensor<fp16, []> var_57_to_fp16 = const()[name = tensor<string, []>("op_57_to_fp16"), val = tensor<fp16, []>(0x1p+0)]; | |
| tensor<fp16, [1, 1, 1, 224]> var_121_cast_fp16 = sub(x = var_57_to_fp16, y = var_118_cast_fp16)[name = tensor<string, []>("op_121_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 224]> var_122_cast_fp16 = mul(x = var_43_cast_fp16_0, y = var_121_cast_fp16)[name = tensor<string, []>("op_122_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 224]> key_1_cast_fp16 = add(x = var_120_cast_fp16, y = var_122_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 224]> var_124_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_118_cast_fp16)[name = tensor<string, []>("op_124_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 224]> var_126_cast_fp16 = mul(x = var_48_cast_fp16_0, y = var_121_cast_fp16)[name = tensor<string, []>("op_126_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 224]> value_1_cast_fp16 = add(x = var_124_cast_fp16, y = var_126_cast_fp16)[name = tensor<string, []>("value_1_cast_fp16")]; | |
| tensor<int32, [4]> var_129 = const()[name = tensor<string, []>("op_129"), val = tensor<int32, [4]>([1, 20, 64, -1])]; | |
| tensor<fp16, [1, 20, 64, 1]> var_130_cast_fp16 = reshape(shape = var_129, x = query_1_cast_fp16)[name = tensor<string, []>("op_130_cast_fp16")]; | |
| tensor<fp16, []> var_131_to_fp16 = const()[name = tensor<string, []>("op_131_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 20, 64, 1]> var_132_cast_fp16 = mul(x = var_130_cast_fp16, y = var_131_to_fp16)[name = tensor<string, []>("op_132_cast_fp16")]; | |
| tensor<int32, [4]> var_133 = const()[name = tensor<string, []>("op_133"), val = tensor<int32, [4]>([1, 20, 64, -1])]; | |
| tensor<fp16, [1, 20, 64, 224]> var_134_cast_fp16 = reshape(shape = var_133, x = key_1_cast_fp16)[name = tensor<string, []>("op_134_cast_fp16")]; | |
| tensor<bool, []> mh_w_1_transpose_x_0 = const()[name = tensor<string, []>("mh_w_1_transpose_x_0"), val = tensor<bool, []>(true)]; | |
| tensor<bool, []> mh_w_1_transpose_y_0 = const()[name = tensor<string, []>("mh_w_1_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 20, 1, 224]> mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_132_cast_fp16, y = var_134_cast_fp16)[name = tensor<string, []>("mh_w_1_cast_fp16")]; | |
| tensor<int32, [1]> var_138_axes_0 = const()[name = tensor<string, []>("op_138_axes_0"), val = tensor<int32, [1]>([1])]; | |
| tensor<fp16, [1, 1, 224]> var_138_cast_fp16 = expand_dims(axes = var_138_axes_0, x = decoder_key_padding_mask)[name = tensor<string, []>("op_138_cast_fp16")]; | |
| tensor<int32, [1]> var_139_axes_0 = const()[name = tensor<string, []>("op_139_axes_0"), val = tensor<int32, [1]>([2])]; | |
| tensor<fp16, [1, 1, 1, 224]> var_139_cast_fp16 = expand_dims(axes = var_139_axes_0, x = var_138_cast_fp16)[name = tensor<string, []>("op_139_cast_fp16")]; | |
| tensor<fp16, [1, 20, 1, 224]> mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_139_cast_fp16)[name = tensor<string, []>("mh_w_3_cast_fp16")]; | |
| tensor<fp16, [1, 20, 1, 224]> var_142_cast_fp16 = softmax(axis = var_56, x = mh_w_3_cast_fp16)[name = tensor<string, []>("op_142_cast_fp16")]; | |
| tensor<int32, [4]> var_143 = const()[name = tensor<string, []>("op_143"), val = tensor<int32, [4]>([1, 20, 64, -1])]; | |
| tensor<fp16, [1, 20, 64, 224]> var_144_cast_fp16 = reshape(shape = var_143, x = value_1_cast_fp16)[name = tensor<string, []>("op_144_cast_fp16")]; | |
| tensor<bool, []> attn_1_transpose_x_0 = const()[name = tensor<string, []>("attn_1_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_1_transpose_y_0 = const()[name = tensor<string, []>("attn_1_transpose_y_0"), val = tensor<bool, []>(true)]; | |
| 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_144_cast_fp16, y = var_142_cast_fp16)[name = tensor<string, []>("attn_1_cast_fp16")]; | |
| tensor<int32, [4]> var_147 = const()[name = tensor<string, []>("op_147"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; | |
| tensor<fp16, [1, 1280, 1, 1]> input_1_cast_fp16 = reshape(shape = var_147, x = attn_1_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")]; | |
| tensor<int32, [2]> var_151 = const()[name = tensor<string, []>("op_151"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [2]> var_153 = const()[name = tensor<string, []>("op_153"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<string, []> obj_7_pad_type_0 = const()[name = tensor<string, []>("obj_7_pad_type_0"), val = tensor<string, []>("custom")]; | |
| tensor<int32, [4]> obj_7_pad_0 = const()[name = tensor<string, []>("obj_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| 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)))]; | |
| 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)))]; | |
| tensor<fp16, [1, 1280, 1, 1]> obj_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = var_153, groups = var_63, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = var_151, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("obj_7_cast_fp16")]; | |
| 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")]; | |
| tensor<int32, [1]> var_163 = const()[name = tensor<string, []>("op_163"), val = tensor<int32, [1]>([1])]; | |
| tensor<fp16, [1, 1, 1, 1]> channels_mean_3_cast_fp16 = reduce_mean(axes = var_163, keep_dims = var_64, x = inputs_3_cast_fp16)[name = tensor<string, []>("channels_mean_3_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> zero_mean_3_cast_fp16 = sub(x = inputs_3_cast_fp16, y = channels_mean_3_cast_fp16)[name = tensor<string, []>("zero_mean_3_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = zero_mean_3_cast_fp16)[name = tensor<string, []>("zero_mean_sq_3_cast_fp16")]; | |
| tensor<int32, [1]> var_167 = const()[name = tensor<string, []>("op_167"), val = tensor<int32, [1]>([1])]; | |
| tensor<fp16, [1, 1, 1, 1]> var_168_cast_fp16 = reduce_mean(axes = var_167, keep_dims = var_64, x = zero_mean_sq_3_cast_fp16)[name = tensor<string, []>("op_168_cast_fp16")]; | |
| tensor<fp16, []> var_169_to_fp16 = const()[name = tensor<string, []>("op_169_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; | |
| tensor<fp16, [1, 1, 1, 1]> var_170_cast_fp16 = add(x = var_168_cast_fp16, y = var_169_to_fp16)[name = tensor<string, []>("op_170_cast_fp16")]; | |
| tensor<fp16, []> denom_3_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_3_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)]; | |
| tensor<fp16, [1, 1, 1, 1]> denom_3_cast_fp16 = rsqrt(epsilon = denom_3_epsilon_0_to_fp16, x = var_170_cast_fp16)[name = tensor<string, []>("denom_3_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> out_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = denom_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")]; | |
| 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)))]; | |
| 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)))]; | |
| tensor<fp16, []> obj_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; | |
| 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")]; | |
| tensor<int32, [2]> var_185 = const()[name = tensor<string, []>("op_185"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [2]> var_187 = const()[name = tensor<string, []>("op_187"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<string, []> query_3_pad_type_0 = const()[name = tensor<string, []>("query_3_pad_type_0"), val = tensor<string, []>("custom")]; | |
| tensor<int32, [4]> query_3_pad_0 = const()[name = tensor<string, []>("query_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| 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)))]; | |
| 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)))]; | |
| tensor<fp16, [1, 1280, 1, 1]> query_3_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_bias_to_fp16, dilations = var_187, groups = var_63, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = var_185, weight = layers_0_encoder_attn_q_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor<string, []>("query_3_cast_fp16")]; | |
| tensor<int32, [2]> var_191 = const()[name = tensor<string, []>("op_191"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [2]> var_193 = const()[name = tensor<string, []>("op_193"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<string, []> key_3_pad_type_0 = const()[name = tensor<string, []>("key_3_pad_type_0"), val = tensor<string, []>("custom")]; | |
| tensor<int32, [4]> key_3_pad_0 = const()[name = tensor<string, []>("key_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| 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)))]; | |
| tensor<fp16, [1, 1280, 1, 1500]> key_3_cast_fp16 = conv(dilations = var_193, groups = var_63, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = var_191, weight = layers_0_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_3_cast_fp16")]; | |
| tensor<int32, [2]> var_198 = const()[name = tensor<string, []>("op_198"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [2]> var_200 = const()[name = tensor<string, []>("op_200"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<string, []> value_3_pad_type_0 = const()[name = tensor<string, []>("value_3_pad_type_0"), val = tensor<string, []>("custom")]; | |
| tensor<int32, [4]> value_3_pad_0 = const()[name = tensor<string, []>("value_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| 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)))]; | |
| 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)))]; | |
| tensor<fp16, [1, 1280, 1, 1500]> value_3_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_bias_to_fp16, dilations = var_200, groups = var_63, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = var_198, weight = layers_0_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_3_cast_fp16")]; | |
| tensor<int32, [4]> var_204 = const()[name = tensor<string, []>("op_204"), val = tensor<int32, [4]>([1, 20, 64, -1])]; | |
| tensor<fp16, [1, 20, 64, 1]> var_205_cast_fp16 = reshape(shape = var_204, x = query_3_cast_fp16)[name = tensor<string, []>("op_205_cast_fp16")]; | |
| tensor<fp16, []> var_206_to_fp16 = const()[name = tensor<string, []>("op_206_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 20, 64, 1]> var_207_cast_fp16 = mul(x = var_205_cast_fp16, y = var_206_to_fp16)[name = tensor<string, []>("op_207_cast_fp16")]; | |
| tensor<int32, [4]> var_208 = const()[name = tensor<string, []>("op_208"), val = tensor<int32, [4]>([1, 20, 64, -1])]; | |
| tensor<fp16, [1, 20, 64, 1500]> var_209_cast_fp16 = reshape(shape = var_208, x = key_3_cast_fp16)[name = tensor<string, []>("op_209_cast_fp16")]; | |
| tensor<bool, []> mh_w_5_transpose_x_0 = const()[name = tensor<string, []>("mh_w_5_transpose_x_0"), val = tensor<bool, []>(true)]; | |
| tensor<bool, []> mh_w_5_transpose_y_0 = const()[name = tensor<string, []>("mh_w_5_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| 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_207_cast_fp16, y = var_209_cast_fp16)[name = tensor<string, []>("mh_w_5_cast_fp16")]; | |
| tensor<fp16, [1, 20, 1, 1500]> var_212_cast_fp16 = softmax(axis = var_56, x = mh_w_5_cast_fp16)[name = tensor<string, []>("op_212_cast_fp16")]; | |
| tensor<int32, [4]> var_213 = const()[name = tensor<string, []>("op_213"), val = tensor<int32, [4]>([1, 20, 64, -1])]; | |
| tensor<fp16, [1, 20, 64, 1500]> var_214_cast_fp16 = reshape(shape = var_213, x = value_3_cast_fp16)[name = tensor<string, []>("op_214_cast_fp16")]; | |
| tensor<bool, []> attn_3_transpose_x_0 = const()[name = tensor<string, []>("attn_3_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_3_transpose_y_0 = const()[name = tensor<string, []>("attn_3_transpose_y_0"), val = tensor<bool, []>(true)]; | |
| 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_214_cast_fp16, y = var_212_cast_fp16)[name = tensor<string, []>("attn_3_cast_fp16")]; | |
| tensor<int32, [4]> var_217 = const()[name = tensor<string, []>("op_217"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; | |
| tensor<fp16, [1, 1280, 1, 1]> input_3_cast_fp16 = reshape(shape = var_217, x = attn_3_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")]; | |
| tensor<int32, [2]> var_221 = const()[name = tensor<string, []>("op_221"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [2]> var_223 = const()[name = tensor<string, []>("op_223"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<string, []> obj_11_pad_type_0 = const()[name = tensor<string, []>("obj_11_pad_type_0"), val = tensor<string, []>("custom")]; | |
| tensor<int32, [4]> obj_11_pad_0 = const()[name = tensor<string, []>("obj_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| 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)))]; | |
| 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)))]; | |
| tensor<fp16, [1, 1280, 1, 1]> obj_11_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = var_223, groups = var_63, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = var_221, weight = layers_0_encoder_attn_o_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("obj_11_cast_fp16")]; | |
| 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")]; | |
| tensor<int32, [1]> var_229 = const()[name = tensor<string, []>("op_229"), val = tensor<int32, [1]>([1])]; | |
| tensor<fp16, [1, 1, 1, 1]> channels_mean_5_cast_fp16 = reduce_mean(axes = var_229, keep_dims = var_64, x = inputs_5_cast_fp16)[name = tensor<string, []>("channels_mean_5_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> zero_mean_5_cast_fp16 = sub(x = inputs_5_cast_fp16, y = channels_mean_5_cast_fp16)[name = tensor<string, []>("zero_mean_5_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = zero_mean_5_cast_fp16)[name = tensor<string, []>("zero_mean_sq_5_cast_fp16")]; | |
| tensor<int32, [1]> var_233 = const()[name = tensor<string, []>("op_233"), val = tensor<int32, [1]>([1])]; | |
| tensor<fp16, [1, 1, 1, 1]> var_234_cast_fp16 = reduce_mean(axes = var_233, keep_dims = var_64, x = zero_mean_sq_5_cast_fp16)[name = tensor<string, []>("op_234_cast_fp16")]; | |
| tensor<fp16, []> var_235_to_fp16 = const()[name = tensor<string, []>("op_235_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; | |
| tensor<fp16, [1, 1, 1, 1]> var_236_cast_fp16 = add(x = var_234_cast_fp16, y = var_235_to_fp16)[name = tensor<string, []>("op_236_cast_fp16")]; | |
| tensor<fp16, []> denom_5_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_5_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)]; | |
| tensor<fp16, [1, 1, 1, 1]> denom_5_cast_fp16 = rsqrt(epsilon = denom_5_epsilon_0_to_fp16, x = var_236_cast_fp16)[name = tensor<string, []>("denom_5_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> out_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = denom_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")]; | |
| 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)))]; | |
| 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)))]; | |
| tensor<fp16, []> input_5_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_5_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; | |
| 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")]; | |
| tensor<int32, [2]> var_247 = const()[name = tensor<string, []>("op_247"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [2]> var_249 = const()[name = tensor<string, []>("op_249"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<string, []> input_7_pad_type_0 = const()[name = tensor<string, []>("input_7_pad_type_0"), val = tensor<string, []>("custom")]; | |
| tensor<int32, [4]> input_7_pad_0 = const()[name = tensor<string, []>("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| 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)))]; | |
| 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)))]; | |
| tensor<fp16, [1, 5120, 1, 1]> input_7_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = var_249, groups = var_63, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = var_247, weight = layers_0_fc1_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")]; | |
| tensor<string, []> input_9_mode_0 = const()[name = tensor<string, []>("input_9_mode_0"), val = tensor<string, []>("EXACT")]; | |
| 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")]; | |
| tensor<int32, [2]> var_255 = const()[name = tensor<string, []>("op_255"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [2]> var_257 = const()[name = tensor<string, []>("op_257"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<string, []> hidden_states_3_pad_type_0 = const()[name = tensor<string, []>("hidden_states_3_pad_type_0"), val = tensor<string, []>("custom")]; | |
| 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])]; | |
| 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)))]; | |
| 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)))]; | |
| tensor<fp16, [1, 1280, 1, 1]> hidden_states_3_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = var_257, groups = var_63, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = var_255, weight = layers_0_fc2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("hidden_states_3_cast_fp16")]; | |
| 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")]; | |
| tensor<int32, []> var_270 = const()[name = tensor<string, []>("op_270"), val = tensor<int32, []>(3)]; | |
| tensor<int32, []> var_277 = const()[name = tensor<string, []>("op_277"), val = tensor<int32, []>(1)]; | |
| tensor<bool, []> var_278 = const()[name = tensor<string, []>("op_278"), val = tensor<bool, []>(true)]; | |
| tensor<int32, [1]> var_290 = const()[name = tensor<string, []>("op_290"), val = tensor<int32, [1]>([1])]; | |
| tensor<fp16, [1, 1, 1, 1]> channels_mean_7_cast_fp16 = reduce_mean(axes = var_290, keep_dims = var_278, x = inputs_7_cast_fp16)[name = tensor<string, []>("channels_mean_7_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> zero_mean_7_cast_fp16 = sub(x = inputs_7_cast_fp16, y = channels_mean_7_cast_fp16)[name = tensor<string, []>("zero_mean_7_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = zero_mean_7_cast_fp16)[name = tensor<string, []>("zero_mean_sq_7_cast_fp16")]; | |
| tensor<int32, [1]> var_294 = const()[name = tensor<string, []>("op_294"), val = tensor<int32, [1]>([1])]; | |
| tensor<fp16, [1, 1, 1, 1]> var_295_cast_fp16 = reduce_mean(axes = var_294, keep_dims = var_278, x = zero_mean_sq_7_cast_fp16)[name = tensor<string, []>("op_295_cast_fp16")]; | |
| tensor<fp16, []> var_296_to_fp16 = const()[name = tensor<string, []>("op_296_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; | |
| tensor<fp16, [1, 1, 1, 1]> var_297_cast_fp16 = add(x = var_295_cast_fp16, y = var_296_to_fp16)[name = tensor<string, []>("op_297_cast_fp16")]; | |
| tensor<fp16, []> denom_7_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_7_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)]; | |
| tensor<fp16, [1, 1, 1, 1]> denom_7_cast_fp16 = rsqrt(epsilon = denom_7_epsilon_0_to_fp16, x = var_297_cast_fp16)[name = tensor<string, []>("denom_7_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> out_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = denom_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")]; | |
| tensor<fp16, [1280]> obj_13_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_13_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186403136)))]; | |
| tensor<fp16, [1280]> obj_13_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_13_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186405760)))]; | |
| tensor<fp16, []> obj_13_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_13_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; | |
| tensor<fp16, [1, 1280, 1, 1]> obj_13_cast_fp16 = batch_norm(beta = obj_13_beta_0_to_fp16, epsilon = obj_13_epsilon_0_to_fp16, gamma = obj_13_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_13_cast_fp16")]; | |
| tensor<int32, [2]> var_312 = const()[name = tensor<string, []>("op_312"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [2]> var_314 = const()[name = tensor<string, []>("op_314"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<string, []> query_5_pad_type_0 = const()[name = tensor<string, []>("query_5_pad_type_0"), val = tensor<string, []>("custom")]; | |
| tensor<int32, [4]> query_5_pad_0 = const()[name = tensor<string, []>("query_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| 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)))]; | |
| 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)))]; | |
| tensor<fp16, [1, 1280, 1, 1]> query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = var_314, groups = var_277, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = var_312, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")]; | |
| tensor<int32, [2]> var_318 = const()[name = tensor<string, []>("op_318"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [2]> var_320 = const()[name = tensor<string, []>("op_320"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<string, []> current_key_pad_type_0 = const()[name = tensor<string, []>("current_key_pad_type_0"), val = tensor<string, []>("custom")]; | |
| tensor<int32, [4]> current_key_pad_0 = const()[name = tensor<string, []>("current_key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| 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)))]; | |
| tensor<fp16, [1, 1280, 1, 1]> current_key_cast_fp16 = conv(dilations = var_320, groups = var_277, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = var_318, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor<string, []>("current_key_cast_fp16")]; | |
| tensor<int32, [2]> var_325 = const()[name = tensor<string, []>("op_325"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [2]> var_327 = const()[name = tensor<string, []>("op_327"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<string, []> current_value_pad_type_0 = const()[name = tensor<string, []>("current_value_pad_type_0"), val = tensor<string, []>("custom")]; | |
| tensor<int32, [4]> current_value_pad_0 = const()[name = tensor<string, []>("current_value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| 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)))]; | |
| 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)))]; | |
| tensor<fp16, [1, 1280, 1, 1]> current_value_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = var_327, groups = var_277, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = var_325, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor<string, []>("current_value_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 224]> var_334_cast_fp16 = mul(x = current_key_cast_fp16, y = var_118_cast_fp16)[name = tensor<string, []>("op_334_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 224]> var_336_cast_fp16 = mul(x = var_43_cast_fp16_1, y = var_121_cast_fp16)[name = tensor<string, []>("op_336_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 224]> key_5_cast_fp16 = add(x = var_334_cast_fp16, y = var_336_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 224]> var_338_cast_fp16 = mul(x = current_value_cast_fp16, y = var_118_cast_fp16)[name = tensor<string, []>("op_338_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 224]> var_340_cast_fp16 = mul(x = var_48_cast_fp16_1, y = var_121_cast_fp16)[name = tensor<string, []>("op_340_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 224]> value_5_cast_fp16 = add(x = var_338_cast_fp16, y = var_340_cast_fp16)[name = tensor<string, []>("value_5_cast_fp16")]; | |
| tensor<int32, [4]> var_343 = const()[name = tensor<string, []>("op_343"), val = tensor<int32, [4]>([1, 20, 64, -1])]; | |
| tensor<fp16, [1, 20, 64, 1]> var_344_cast_fp16 = reshape(shape = var_343, x = query_5_cast_fp16)[name = tensor<string, []>("op_344_cast_fp16")]; | |
| tensor<fp16, []> var_345_to_fp16 = const()[name = tensor<string, []>("op_345_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 20, 64, 1]> var_346_cast_fp16 = mul(x = var_344_cast_fp16, y = var_345_to_fp16)[name = tensor<string, []>("op_346_cast_fp16")]; | |
| tensor<int32, [4]> var_347 = const()[name = tensor<string, []>("op_347"), val = tensor<int32, [4]>([1, 20, 64, -1])]; | |
| tensor<fp16, [1, 20, 64, 224]> var_348_cast_fp16 = reshape(shape = var_347, x = key_5_cast_fp16)[name = tensor<string, []>("op_348_cast_fp16")]; | |
| tensor<bool, []> mh_w_7_transpose_x_0 = const()[name = tensor<string, []>("mh_w_7_transpose_x_0"), val = tensor<bool, []>(true)]; | |
| tensor<bool, []> mh_w_7_transpose_y_0 = const()[name = tensor<string, []>("mh_w_7_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 20, 1, 224]> mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_346_cast_fp16, y = var_348_cast_fp16)[name = tensor<string, []>("mh_w_7_cast_fp16")]; | |
| tensor<fp16, [1, 20, 1, 224]> mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_139_cast_fp16)[name = tensor<string, []>("mh_w_9_cast_fp16")]; | |
| tensor<fp16, [1, 20, 1, 224]> var_356_cast_fp16 = softmax(axis = var_270, x = mh_w_9_cast_fp16)[name = tensor<string, []>("op_356_cast_fp16")]; | |
| tensor<int32, [4]> var_357 = const()[name = tensor<string, []>("op_357"), val = tensor<int32, [4]>([1, 20, 64, -1])]; | |
| tensor<fp16, [1, 20, 64, 224]> var_358_cast_fp16 = reshape(shape = var_357, x = value_5_cast_fp16)[name = tensor<string, []>("op_358_cast_fp16")]; | |
| tensor<bool, []> attn_5_transpose_x_0 = const()[name = tensor<string, []>("attn_5_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_5_transpose_y_0 = const()[name = tensor<string, []>("attn_5_transpose_y_0"), val = tensor<bool, []>(true)]; | |
| 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_358_cast_fp16, y = var_356_cast_fp16)[name = tensor<string, []>("attn_5_cast_fp16")]; | |
| tensor<int32, [4]> var_361 = const()[name = tensor<string, []>("op_361"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; | |
| tensor<fp16, [1, 1280, 1, 1]> input_11_cast_fp16 = reshape(shape = var_361, x = attn_5_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")]; | |
| tensor<int32, [2]> var_365 = const()[name = tensor<string, []>("op_365"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [2]> var_367 = const()[name = tensor<string, []>("op_367"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<string, []> obj_19_pad_type_0 = const()[name = tensor<string, []>("obj_19_pad_type_0"), val = tensor<string, []>("custom")]; | |
| tensor<int32, [4]> obj_19_pad_0 = const()[name = tensor<string, []>("obj_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| 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)))]; | |
| 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)))]; | |
| tensor<fp16, [1, 1280, 1, 1]> obj_19_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = var_367, groups = var_277, pad = obj_19_pad_0, pad_type = obj_19_pad_type_0, strides = var_365, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("obj_19_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_19_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")]; | |
| tensor<int32, [1]> var_377 = const()[name = tensor<string, []>("op_377"), val = tensor<int32, [1]>([1])]; | |
| tensor<fp16, [1, 1, 1, 1]> channels_mean_9_cast_fp16 = reduce_mean(axes = var_377, keep_dims = var_278, x = inputs_9_cast_fp16)[name = tensor<string, []>("channels_mean_9_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> zero_mean_9_cast_fp16 = sub(x = inputs_9_cast_fp16, y = channels_mean_9_cast_fp16)[name = tensor<string, []>("zero_mean_9_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = zero_mean_9_cast_fp16)[name = tensor<string, []>("zero_mean_sq_9_cast_fp16")]; | |
| tensor<int32, [1]> var_381 = const()[name = tensor<string, []>("op_381"), val = tensor<int32, [1]>([1])]; | |
| tensor<fp16, [1, 1, 1, 1]> var_382_cast_fp16 = reduce_mean(axes = var_381, keep_dims = var_278, x = zero_mean_sq_9_cast_fp16)[name = tensor<string, []>("op_382_cast_fp16")]; | |
| tensor<fp16, []> var_383_to_fp16 = const()[name = tensor<string, []>("op_383_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; | |
| tensor<fp16, [1, 1, 1, 1]> var_384_cast_fp16 = add(x = var_382_cast_fp16, y = var_383_to_fp16)[name = tensor<string, []>("op_384_cast_fp16")]; | |
| tensor<fp16, []> denom_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)]; | |
| tensor<fp16, [1, 1, 1, 1]> denom_9_cast_fp16 = rsqrt(epsilon = denom_9_epsilon_0_to_fp16, x = var_384_cast_fp16)[name = tensor<string, []>("denom_9_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> out_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = denom_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")]; | |
| tensor<fp16, [1280]> obj_21_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_21_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199523712)))]; | |
| tensor<fp16, [1280]> obj_21_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_21_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199526336)))]; | |
| tensor<fp16, []> obj_21_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_21_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; | |
| tensor<fp16, [1, 1280, 1, 1]> obj_21_cast_fp16 = batch_norm(beta = obj_21_beta_0_to_fp16, epsilon = obj_21_epsilon_0_to_fp16, gamma = obj_21_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_21_cast_fp16")]; | |
| tensor<int32, [2]> var_399 = const()[name = tensor<string, []>("op_399"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [2]> var_401 = const()[name = tensor<string, []>("op_401"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<string, []> query_pad_type_0 = const()[name = tensor<string, []>("query_pad_type_0"), val = tensor<string, []>("custom")]; | |
| tensor<int32, [4]> query_pad_0 = const()[name = tensor<string, []>("query_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| 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)))]; | |
| 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)))]; | |
| tensor<fp16, [1, 1280, 1, 1]> query_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = var_401, groups = var_277, pad = query_pad_0, pad_type = query_pad_type_0, strides = var_399, weight = layers_1_encoder_attn_q_proj_weight_to_fp16, x = obj_21_cast_fp16)[name = tensor<string, []>("query_cast_fp16")]; | |
| tensor<int32, [2]> var_405 = const()[name = tensor<string, []>("op_405"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [2]> var_407 = const()[name = tensor<string, []>("op_407"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<string, []> key_pad_type_0 = const()[name = tensor<string, []>("key_pad_type_0"), val = tensor<string, []>("custom")]; | |
| tensor<int32, [4]> key_pad_0 = const()[name = tensor<string, []>("key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| 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)))]; | |
| tensor<fp16, [1, 1280, 1, 1500]> key_cast_fp16 = conv(dilations = var_407, groups = var_277, pad = key_pad_0, pad_type = key_pad_type_0, strides = var_405, weight = layers_1_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_cast_fp16")]; | |
| tensor<int32, [2]> var_412 = const()[name = tensor<string, []>("op_412"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [2]> var_414 = const()[name = tensor<string, []>("op_414"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<string, []> value_pad_type_0 = const()[name = tensor<string, []>("value_pad_type_0"), val = tensor<string, []>("custom")]; | |
| tensor<int32, [4]> value_pad_0 = const()[name = tensor<string, []>("value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| 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)))]; | |
| 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)))]; | |
| tensor<fp16, [1, 1280, 1, 1500]> value_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_bias_to_fp16, dilations = var_414, groups = var_277, pad = value_pad_0, pad_type = value_pad_type_0, strides = var_412, weight = layers_1_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_cast_fp16")]; | |
| tensor<int32, [4]> var_418 = const()[name = tensor<string, []>("op_418"), val = tensor<int32, [4]>([1, 20, 64, -1])]; | |
| tensor<fp16, [1, 20, 64, 1]> var_419_cast_fp16 = reshape(shape = var_418, x = query_cast_fp16)[name = tensor<string, []>("op_419_cast_fp16")]; | |
| tensor<fp16, []> var_420_to_fp16 = const()[name = tensor<string, []>("op_420_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 20, 64, 1]> var_421_cast_fp16 = mul(x = var_419_cast_fp16, y = var_420_to_fp16)[name = tensor<string, []>("op_421_cast_fp16")]; | |
| tensor<int32, [4]> var_422 = const()[name = tensor<string, []>("op_422"), val = tensor<int32, [4]>([1, 20, 64, -1])]; | |
| tensor<fp16, [1, 20, 64, 1500]> var_423_cast_fp16 = reshape(shape = var_422, x = key_cast_fp16)[name = tensor<string, []>("op_423_cast_fp16")]; | |
| tensor<bool, []> mh_w_transpose_x_0 = const()[name = tensor<string, []>("mh_w_transpose_x_0"), val = tensor<bool, []>(true)]; | |
| tensor<bool, []> mh_w_transpose_y_0 = const()[name = tensor<string, []>("mh_w_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| 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_421_cast_fp16, y = var_423_cast_fp16)[name = tensor<string, []>("mh_w_cast_fp16")]; | |
| tensor<fp16, [1, 20, 1, 1500]> var_426_cast_fp16 = softmax(axis = var_270, x = mh_w_cast_fp16)[name = tensor<string, []>("op_426_cast_fp16")]; | |
| tensor<int32, [4]> var_427 = const()[name = tensor<string, []>("op_427"), val = tensor<int32, [4]>([1, 20, 64, -1])]; | |
| tensor<fp16, [1, 20, 64, 1500]> var_428_cast_fp16 = reshape(shape = var_427, x = value_cast_fp16)[name = tensor<string, []>("op_428_cast_fp16")]; | |
| tensor<bool, []> attn_transpose_x_0 = const()[name = tensor<string, []>("attn_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_transpose_y_0 = const()[name = tensor<string, []>("attn_transpose_y_0"), val = tensor<bool, []>(true)]; | |
| tensor<fp16, [1, 20, 64, 1]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_428_cast_fp16, y = var_426_cast_fp16)[name = tensor<string, []>("attn_cast_fp16")]; | |
| tensor<int32, [4]> var_431 = const()[name = tensor<string, []>("op_431"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; | |
| tensor<fp16, [1, 1280, 1, 1]> input_13_cast_fp16 = reshape(shape = var_431, x = attn_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")]; | |
| tensor<int32, [2]> var_435 = const()[name = tensor<string, []>("op_435"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [2]> var_437 = const()[name = tensor<string, []>("op_437"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<string, []> obj_23_pad_type_0 = const()[name = tensor<string, []>("obj_23_pad_type_0"), val = tensor<string, []>("custom")]; | |
| tensor<int32, [4]> obj_23_pad_0 = const()[name = tensor<string, []>("obj_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| 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)))]; | |
| 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)))]; | |
| tensor<fp16, [1, 1280, 1, 1]> obj_23_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = var_437, groups = var_277, pad = obj_23_pad_0, pad_type = obj_23_pad_type_0, strides = var_435, weight = layers_1_encoder_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("obj_23_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_23_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")]; | |
| tensor<int32, [1]> var_443 = const()[name = tensor<string, []>("op_443"), val = tensor<int32, [1]>([1])]; | |
| tensor<fp16, [1, 1, 1, 1]> channels_mean_11_cast_fp16 = reduce_mean(axes = var_443, keep_dims = var_278, x = inputs_11_cast_fp16)[name = tensor<string, []>("channels_mean_11_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> zero_mean_11_cast_fp16 = sub(x = inputs_11_cast_fp16, y = channels_mean_11_cast_fp16)[name = tensor<string, []>("zero_mean_11_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = zero_mean_11_cast_fp16)[name = tensor<string, []>("zero_mean_sq_11_cast_fp16")]; | |
| tensor<int32, [1]> var_447 = const()[name = tensor<string, []>("op_447"), val = tensor<int32, [1]>([1])]; | |
| tensor<fp16, [1, 1, 1, 1]> var_448_cast_fp16 = reduce_mean(axes = var_447, keep_dims = var_278, x = zero_mean_sq_11_cast_fp16)[name = tensor<string, []>("op_448_cast_fp16")]; | |
| tensor<fp16, []> var_449_to_fp16 = const()[name = tensor<string, []>("op_449_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; | |
| tensor<fp16, [1, 1, 1, 1]> var_450_cast_fp16 = add(x = var_448_cast_fp16, y = var_449_to_fp16)[name = tensor<string, []>("op_450_cast_fp16")]; | |
| tensor<fp16, []> denom_11_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_11_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)]; | |
| tensor<fp16, [1, 1, 1, 1]> denom_11_cast_fp16 = rsqrt(epsilon = denom_11_epsilon_0_to_fp16, x = var_450_cast_fp16)[name = tensor<string, []>("denom_11_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> out_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = denom_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")]; | |
| 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)))]; | |
| 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)))]; | |
| tensor<fp16, []> input_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; | |
| 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")]; | |
| tensor<int32, [2]> var_461 = const()[name = tensor<string, []>("op_461"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [2]> var_463 = const()[name = tensor<string, []>("op_463"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<string, []> input_17_pad_type_0 = const()[name = tensor<string, []>("input_17_pad_type_0"), val = tensor<string, []>("custom")]; | |
| tensor<int32, [4]> input_17_pad_0 = const()[name = tensor<string, []>("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; | |
| 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)))]; | |
| 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)))]; | |
| tensor<fp16, [1, 5120, 1, 1]> input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = var_463, groups = var_277, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = var_461, weight = layers_1_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")]; | |
| tensor<string, []> input_mode_0 = const()[name = tensor<string, []>("input_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 5120, 1, 1]> input_cast_fp16 = gelu(mode = input_mode_0, x = input_17_cast_fp16)[name = tensor<string, []>("input_cast_fp16")]; | |
| tensor<int32, [2]> var_469 = const()[name = tensor<string, []>("op_469"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<int32, [2]> var_471 = const()[name = tensor<string, []>("op_471"), val = tensor<int32, [2]>([1, 1])]; | |
| tensor<string, []> hidden_states_5_pad_type_0 = const()[name = tensor<string, []>("hidden_states_5_pad_type_0"), val = tensor<string, []>("custom")]; | |
| 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])]; | |
| 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)))]; | |
| 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)))]; | |
| tensor<fp16, [1, 1280, 1, 1]> hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = var_471, groups = var_277, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = var_469, weight = layers_1_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("hidden_states_5_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> inputs_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor<string, []>("inputs_cast_fp16")]; | |
| tensor<bool, []> var_481 = const()[name = tensor<string, []>("op_481"), val = tensor<bool, []>(true)]; | |
| tensor<int32, [1]> var_485 = const()[name = tensor<string, []>("op_485"), val = tensor<int32, [1]>([1])]; | |
| tensor<fp16, [1, 1, 1, 1]> channels_mean_cast_fp16 = reduce_mean(axes = var_485, keep_dims = var_481, x = inputs_cast_fp16)[name = tensor<string, []>("channels_mean_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> zero_mean_cast_fp16 = sub(x = inputs_cast_fp16, y = channels_mean_cast_fp16)[name = tensor<string, []>("zero_mean_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_cast_fp16 = mul(x = zero_mean_cast_fp16, y = zero_mean_cast_fp16)[name = tensor<string, []>("zero_mean_sq_cast_fp16")]; | |
| tensor<int32, [1]> var_489 = const()[name = tensor<string, []>("op_489"), val = tensor<int32, [1]>([1])]; | |
| tensor<fp16, [1, 1, 1, 1]> var_490_cast_fp16 = reduce_mean(axes = var_489, keep_dims = var_481, x = zero_mean_sq_cast_fp16)[name = tensor<string, []>("op_490_cast_fp16")]; | |
| tensor<fp16, []> var_491_to_fp16 = const()[name = tensor<string, []>("op_491_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; | |
| tensor<fp16, [1, 1, 1, 1]> var_492_cast_fp16 = add(x = var_490_cast_fp16, y = var_491_to_fp16)[name = tensor<string, []>("op_492_cast_fp16")]; | |
| tensor<fp16, []> denom_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)]; | |
| tensor<fp16, [1, 1, 1, 1]> denom_cast_fp16 = rsqrt(epsilon = denom_epsilon_0_to_fp16, x = var_492_cast_fp16)[name = tensor<string, []>("denom_cast_fp16")]; | |
| tensor<fp16, [1, 1280, 1, 1]> out_cast_fp16 = mul(x = zero_mean_cast_fp16, y = denom_cast_fp16)[name = tensor<string, []>("out_cast_fp16")]; | |
| 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, []>(238876992)))]; | |
| 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, []>(238879616)))]; | |
| tensor<fp16, []> hidden_states_epsilon_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; | |
| 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")]; | |
| tensor<int32, [1]> var_502_axes_0 = const()[name = tensor<string, []>("op_502_axes_0"), val = tensor<int32, [1]>([2])]; | |
| tensor<fp16, [1, 1280, 1]> var_502_cast_fp16 = squeeze(axes = var_502_axes_0, x = hidden_states_cast_fp16)[name = tensor<string, []>("op_502_cast_fp16")]; | |
| tensor<int32, [3]> var_505_perm_0 = const()[name = tensor<string, []>("op_505_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; | |
| 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, []>(238882240)))]; | |
| tensor<fp16, [1, 1, 1280]> transpose_0 = transpose(perm = var_505_perm_0, x = var_502_cast_fp16)[name = tensor<string, []>("transpose_0")]; | |
| tensor<fp16, [1, 1, 51866]> logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = transpose_0)[name = tensor<string, []>("linear_0_cast_fp16")]; | |
| tensor<int32, []> var_509 = const()[name = tensor<string, []>("op_509"), val = tensor<int32, []>(1)]; | |
| tensor<bool, []> obj_27_interleave_0 = const()[name = tensor<string, []>("obj_27_interleave_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 2560, 1, 1]> key_cache_updates = concat(axis = var_509, interleave = obj_27_interleave_0, values = (current_key_1_cast_fp16, current_key_cast_fp16))[name = tensor<string, []>("obj_27_cast_fp16")]; | |
| tensor<int32, []> var_512 = const()[name = tensor<string, []>("op_512"), val = tensor<int32, []>(1)]; | |
| tensor<bool, []> obj_interleave_0 = const()[name = tensor<string, []>("obj_interleave_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 2560, 1, 1]> value_cache_updates = concat(axis = var_512, interleave = obj_interleave_0, values = (current_value_1_cast_fp16, current_value_cast_fp16))[name = tensor<string, []>("obj_cast_fp16")]; | |
| } -> (logits, key_cache_updates, value_cache_updates); | |
| } |