diff --git "a/openai_whisper-tiny.en/TextDecoder.mlmodelc/model.mil" "b/openai_whisper-tiny.en/TextDecoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/openai_whisper-tiny.en/TextDecoder.mlmodelc/model.mil" @@ -0,0 +1,791 @@ +program(1.0) +[buildInfo = dict, tensor>({{"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(tensor cache_length, tensor decoder_key_padding_mask, tensor encoder_output_embeds, tensor input_ids, tensor key_cache, tensor kv_cache_update_mask, tensor value_cache) { + tensor var_24_axis_0 = const()[name = tensor("op_24_axis_0"), val = tensor(0)]; + tensor var_24_batch_dims_0 = const()[name = tensor("op_24_batch_dims_0"), val = tensor(0)]; + tensor embed_tokens_weight_to_fp16 = const()[name = tensor("embed_tokens_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor 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("op_24_cast_fp16")]; + tensor var_28_axis_0 = const()[name = tensor("op_28_axis_0"), val = tensor(0)]; + tensor var_28_batch_dims_0 = const()[name = tensor("op_28_batch_dims_0"), val = tensor(0)]; + tensor embed_positions_weight_to_fp16 = const()[name = tensor("embed_positions_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39831680)))]; + tensor 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("op_28_cast_fp16")]; + tensor hidden_states_1_cast_fp16 = add(x = var_24_cast_fp16, y = var_28_cast_fp16)[name = tensor("hidden_states_1_cast_fp16")]; + tensor var_42_axes_0 = const()[name = tensor("op_42_axes_0"), val = tensor([2])]; + tensor var_42_cast_fp16 = expand_dims(axes = var_42_axes_0, x = hidden_states_1_cast_fp16)[name = tensor("op_42_cast_fp16")]; + tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([3])]; + tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_42_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([384, 384, 384, 384])]; + tensor var_47_axis_0 = const()[name = tensor("op_47_axis_0"), val = tensor(1)]; + tensor var_47_cast_fp16_0, tensor var_47_cast_fp16_1, tensor var_47_cast_fp16_2, tensor var_47_cast_fp16_3 = split(axis = var_47_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor("op_47_cast_fp16")]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([384, 384, 384, 384])]; + tensor var_54_axis_0 = const()[name = tensor("op_54_axis_0"), val = tensor(1)]; + tensor var_54_cast_fp16_0, tensor var_54_cast_fp16_1, tensor var_54_cast_fp16_2, tensor var_54_cast_fp16_3 = split(axis = var_54_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor("op_54_cast_fp16")]; + tensor var_64 = const()[name = tensor("op_64"), val = tensor(3)]; + tensor var_71 = const()[name = tensor("op_71"), val = tensor(1)]; + tensor var_72 = const()[name = tensor("op_72"), val = tensor(true)]; + tensor var_84 = const()[name = tensor("op_84"), val = tensor([1])]; + tensor channels_mean_1_cast_fp16 = reduce_mean(axes = var_84, keep_dims = var_72, x = inputs_1_cast_fp16)[name = tensor("channels_mean_1_cast_fp16")]; + tensor zero_mean_1_cast_fp16 = sub(x = inputs_1_cast_fp16, y = channels_mean_1_cast_fp16)[name = tensor("zero_mean_1_cast_fp16")]; + tensor zero_mean_sq_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = zero_mean_1_cast_fp16)[name = tensor("zero_mean_sq_1_cast_fp16")]; + tensor var_88 = const()[name = tensor("op_88"), val = tensor([1])]; + tensor var_89_cast_fp16 = reduce_mean(axes = var_88, keep_dims = var_72, x = zero_mean_sq_1_cast_fp16)[name = tensor("op_89_cast_fp16")]; + tensor var_90_to_fp16 = const()[name = tensor("op_90_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_91_cast_fp16 = add(x = var_89_cast_fp16, y = var_90_to_fp16)[name = tensor("op_91_cast_fp16")]; + tensor denom_1_epsilon_0_to_fp16 = const()[name = tensor("denom_1_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_1_cast_fp16 = rsqrt(epsilon = denom_1_epsilon_0_to_fp16, x = var_91_cast_fp16)[name = tensor("denom_1_cast_fp16")]; + tensor out_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = denom_1_cast_fp16)[name = tensor("out_1_cast_fp16")]; + tensor obj_1_mean_0_to_fp16 = const()[name = tensor("obj_1_mean_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40175808)))]; + tensor obj_1_variance_0_to_fp16 = const()[name = tensor("obj_1_variance_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40176640)))]; + tensor obj_1_gamma_0_to_fp16 = const()[name = tensor("obj_1_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40177472)))]; + tensor obj_1_beta_0_to_fp16 = const()[name = tensor("obj_1_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40178304)))]; + tensor obj_1_epsilon_0_to_fp16 = const()[name = tensor("obj_1_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor 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("obj_1_cast_fp16")]; + tensor var_106 = const()[name = tensor("op_106"), val = tensor([1, 1])]; + tensor var_108 = const()[name = tensor("op_108"), val = tensor([1, 1])]; + tensor query_1_pad_type_0 = const()[name = tensor("query_1_pad_type_0"), val = tensor("custom")]; + tensor query_1_pad_0 = const()[name = tensor("query_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40179136)))]; + tensor layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40474112)))]; + tensor query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = var_108, groups = var_71, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = var_106, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("query_1_cast_fp16")]; + tensor var_112 = const()[name = tensor("op_112"), val = tensor([1, 1])]; + tensor var_114 = const()[name = tensor("op_114"), val = tensor([1, 1])]; + tensor current_key_1_pad_type_0 = const()[name = tensor("current_key_1_pad_type_0"), val = tensor("custom")]; + tensor current_key_1_pad_0 = const()[name = tensor("current_key_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40474944)))]; + tensor current_key_1_cast_fp16 = conv(dilations = var_114, groups = var_71, pad = current_key_1_pad_0, pad_type = current_key_1_pad_type_0, strides = var_112, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("current_key_1_cast_fp16")]; + tensor var_119 = const()[name = tensor("op_119"), val = tensor([1, 1])]; + tensor var_121 = const()[name = tensor("op_121"), val = tensor([1, 1])]; + tensor current_value_1_pad_type_0 = const()[name = tensor("current_value_1_pad_type_0"), val = tensor("custom")]; + tensor current_value_1_pad_0 = const()[name = tensor("current_value_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40769920)))]; + tensor layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41064896)))]; + tensor current_value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = var_121, groups = var_71, pad = current_value_1_pad_0, pad_type = current_value_1_pad_type_0, strides = var_119, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("current_value_1_cast_fp16")]; + tensor var_125_axes_0 = const()[name = tensor("op_125_axes_0"), val = tensor([1])]; + tensor var_125_cast_fp16 = expand_dims(axes = var_125_axes_0, x = kv_cache_update_mask)[name = tensor("op_125_cast_fp16")]; + tensor var_126_axes_0 = const()[name = tensor("op_126_axes_0"), val = tensor([2])]; + tensor var_126_cast_fp16 = expand_dims(axes = var_126_axes_0, x = var_125_cast_fp16)[name = tensor("op_126_cast_fp16")]; + tensor var_128_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_126_cast_fp16)[name = tensor("op_128_cast_fp16")]; + tensor var_65_to_fp16 = const()[name = tensor("op_65_to_fp16"), val = tensor(0x1p+0)]; + tensor var_129_cast_fp16 = sub(x = var_65_to_fp16, y = var_126_cast_fp16)[name = tensor("op_129_cast_fp16")]; + tensor var_130_cast_fp16 = mul(x = var_47_cast_fp16_0, y = var_129_cast_fp16)[name = tensor("op_130_cast_fp16")]; + tensor key_1_cast_fp16 = add(x = var_128_cast_fp16, y = var_130_cast_fp16)[name = tensor("key_1_cast_fp16")]; + tensor var_132_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_126_cast_fp16)[name = tensor("op_132_cast_fp16")]; + tensor var_134_cast_fp16 = mul(x = var_54_cast_fp16_0, y = var_129_cast_fp16)[name = tensor("op_134_cast_fp16")]; + tensor value_1_cast_fp16 = add(x = var_132_cast_fp16, y = var_134_cast_fp16)[name = tensor("value_1_cast_fp16")]; + tensor var_137 = const()[name = tensor("op_137"), val = tensor([1, 6, 64, -1])]; + tensor var_138_cast_fp16 = reshape(shape = var_137, x = query_1_cast_fp16)[name = tensor("op_138_cast_fp16")]; + tensor var_139_to_fp16 = const()[name = tensor("op_139_to_fp16"), val = tensor(0x1p-3)]; + tensor var_140_cast_fp16 = mul(x = var_138_cast_fp16, y = var_139_to_fp16)[name = tensor("op_140_cast_fp16")]; + tensor var_141 = const()[name = tensor("op_141"), val = tensor([1, 6, 64, -1])]; + tensor var_142_cast_fp16 = reshape(shape = var_141, x = key_1_cast_fp16)[name = tensor("op_142_cast_fp16")]; + tensor mh_w_1_transpose_x_0 = const()[name = tensor("mh_w_1_transpose_x_0"), val = tensor(true)]; + tensor mh_w_1_transpose_y_0 = const()[name = tensor("mh_w_1_transpose_y_0"), val = tensor(false)]; + tensor 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("mh_w_1_cast_fp16")]; + tensor var_146_axes_0 = const()[name = tensor("op_146_axes_0"), val = tensor([1])]; + tensor var_146_cast_fp16 = expand_dims(axes = var_146_axes_0, x = decoder_key_padding_mask)[name = tensor("op_146_cast_fp16")]; + tensor var_147_axes_0 = const()[name = tensor("op_147_axes_0"), val = tensor([2])]; + tensor var_147_cast_fp16 = expand_dims(axes = var_147_axes_0, x = var_146_cast_fp16)[name = tensor("op_147_cast_fp16")]; + tensor mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_147_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; + tensor var_150_cast_fp16 = softmax(axis = var_64, x = mh_w_3_cast_fp16)[name = tensor("op_150_cast_fp16")]; + tensor var_151 = const()[name = tensor("op_151"), val = tensor([1, 6, 64, -1])]; + tensor var_152_cast_fp16 = reshape(shape = var_151, x = value_1_cast_fp16)[name = tensor("op_152_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(true)]; + tensor 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("attn_1_cast_fp16")]; + tensor var_155 = const()[name = tensor("op_155"), val = tensor([1, 384, 1, -1])]; + tensor input_1_cast_fp16 = reshape(shape = var_155, x = attn_1_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor var_159 = const()[name = tensor("op_159"), val = tensor([1, 1])]; + tensor var_161 = const()[name = tensor("op_161"), val = tensor([1, 1])]; + tensor obj_7_pad_type_0 = const()[name = tensor("obj_7_pad_type_0"), val = tensor("custom")]; + tensor obj_7_pad_0 = const()[name = tensor("obj_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41065728)))]; + tensor layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41360704)))]; + tensor obj_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = var_161, groups = var_71, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = var_159, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("obj_7_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_7_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor var_171 = const()[name = tensor("op_171"), val = tensor([1])]; + tensor channels_mean_3_cast_fp16 = reduce_mean(axes = var_171, keep_dims = var_72, x = inputs_3_cast_fp16)[name = tensor("channels_mean_3_cast_fp16")]; + tensor zero_mean_3_cast_fp16 = sub(x = inputs_3_cast_fp16, y = channels_mean_3_cast_fp16)[name = tensor("zero_mean_3_cast_fp16")]; + tensor zero_mean_sq_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = zero_mean_3_cast_fp16)[name = tensor("zero_mean_sq_3_cast_fp16")]; + tensor var_175 = const()[name = tensor("op_175"), val = tensor([1])]; + tensor var_176_cast_fp16 = reduce_mean(axes = var_175, keep_dims = var_72, x = zero_mean_sq_3_cast_fp16)[name = tensor("op_176_cast_fp16")]; + tensor var_177_to_fp16 = const()[name = tensor("op_177_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_178_cast_fp16 = add(x = var_176_cast_fp16, y = var_177_to_fp16)[name = tensor("op_178_cast_fp16")]; + tensor denom_3_epsilon_0_to_fp16 = const()[name = tensor("denom_3_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_3_cast_fp16 = rsqrt(epsilon = denom_3_epsilon_0_to_fp16, x = var_178_cast_fp16)[name = tensor("denom_3_cast_fp16")]; + tensor out_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = denom_3_cast_fp16)[name = tensor("out_3_cast_fp16")]; + tensor obj_9_gamma_0_to_fp16 = const()[name = tensor("obj_9_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41361536)))]; + tensor obj_9_beta_0_to_fp16 = const()[name = tensor("obj_9_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41362368)))]; + tensor obj_9_epsilon_0_to_fp16 = const()[name = tensor("obj_9_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor 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("obj_9_cast_fp16")]; + tensor var_193 = const()[name = tensor("op_193"), val = tensor([1, 1])]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1])]; + tensor query_3_pad_type_0 = const()[name = tensor("query_3_pad_type_0"), val = tensor("custom")]; + tensor query_3_pad_0 = const()[name = tensor("query_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41363200)))]; + tensor layers_0_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41658176)))]; + tensor query_3_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_bias_to_fp16, dilations = var_195, groups = var_71, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = var_193, weight = layers_0_encoder_attn_q_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("query_3_cast_fp16")]; + tensor var_199 = const()[name = tensor("op_199"), val = tensor([1, 1])]; + tensor var_201 = const()[name = tensor("op_201"), val = tensor([1, 1])]; + tensor key_3_pad_type_0 = const()[name = tensor("key_3_pad_type_0"), val = tensor("custom")]; + tensor key_3_pad_0 = const()[name = tensor("key_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41659008)))]; + tensor key_3_cast_fp16 = conv(dilations = var_201, groups = var_71, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = var_199, weight = layers_0_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_3_cast_fp16")]; + tensor var_206 = const()[name = tensor("op_206"), val = tensor([1, 1])]; + tensor var_208 = const()[name = tensor("op_208"), val = tensor([1, 1])]; + tensor value_3_pad_type_0 = const()[name = tensor("value_3_pad_type_0"), val = tensor("custom")]; + tensor value_3_pad_0 = const()[name = tensor("value_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41953984)))]; + tensor layers_0_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42248960)))]; + tensor value_3_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_bias_to_fp16, dilations = var_208, groups = var_71, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = var_206, weight = layers_0_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_3_cast_fp16")]; + tensor var_212 = const()[name = tensor("op_212"), val = tensor([1, 6, 64, -1])]; + tensor var_213_cast_fp16 = reshape(shape = var_212, x = query_3_cast_fp16)[name = tensor("op_213_cast_fp16")]; + tensor var_214_to_fp16 = const()[name = tensor("op_214_to_fp16"), val = tensor(0x1p-3)]; + tensor var_215_cast_fp16 = mul(x = var_213_cast_fp16, y = var_214_to_fp16)[name = tensor("op_215_cast_fp16")]; + tensor var_216 = const()[name = tensor("op_216"), val = tensor([1, 6, 64, -1])]; + tensor var_217_cast_fp16 = reshape(shape = var_216, x = key_3_cast_fp16)[name = tensor("op_217_cast_fp16")]; + tensor mh_w_5_transpose_x_0 = const()[name = tensor("mh_w_5_transpose_x_0"), val = tensor(true)]; + tensor mh_w_5_transpose_y_0 = const()[name = tensor("mh_w_5_transpose_y_0"), val = tensor(false)]; + tensor 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("mh_w_5_cast_fp16")]; + tensor obj_13_cast_fp16 = softmax(axis = var_64, x = mh_w_5_cast_fp16)[name = tensor("obj_13_cast_fp16")]; + tensor var_221 = const()[name = tensor("op_221"), val = tensor([1, 6, 64, -1])]; + tensor var_222_cast_fp16 = reshape(shape = var_221, x = value_3_cast_fp16)[name = tensor("op_222_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(true)]; + tensor 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("attn_3_cast_fp16")]; + tensor var_225 = const()[name = tensor("op_225"), val = tensor([1, 384, 1, -1])]; + tensor input_3_cast_fp16 = reshape(shape = var_225, x = attn_3_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor var_229 = const()[name = tensor("op_229"), val = tensor([1, 1])]; + tensor var_231 = const()[name = tensor("op_231"), val = tensor([1, 1])]; + tensor obj_11_pad_type_0 = const()[name = tensor("obj_11_pad_type_0"), val = tensor("custom")]; + tensor obj_11_pad_0 = const()[name = tensor("obj_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42249792)))]; + tensor layers_0_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42544768)))]; + tensor obj_11_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = var_231, groups = var_71, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = var_229, weight = layers_0_encoder_attn_o_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("obj_11_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_11_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor var_237 = const()[name = tensor("op_237"), val = tensor([1])]; + tensor channels_mean_5_cast_fp16 = reduce_mean(axes = var_237, keep_dims = var_72, x = inputs_5_cast_fp16)[name = tensor("channels_mean_5_cast_fp16")]; + tensor zero_mean_5_cast_fp16 = sub(x = inputs_5_cast_fp16, y = channels_mean_5_cast_fp16)[name = tensor("zero_mean_5_cast_fp16")]; + tensor zero_mean_sq_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = zero_mean_5_cast_fp16)[name = tensor("zero_mean_sq_5_cast_fp16")]; + tensor var_241 = const()[name = tensor("op_241"), val = tensor([1])]; + tensor var_242_cast_fp16 = reduce_mean(axes = var_241, keep_dims = var_72, x = zero_mean_sq_5_cast_fp16)[name = tensor("op_242_cast_fp16")]; + tensor var_243_to_fp16 = const()[name = tensor("op_243_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_244_cast_fp16 = add(x = var_242_cast_fp16, y = var_243_to_fp16)[name = tensor("op_244_cast_fp16")]; + tensor denom_5_epsilon_0_to_fp16 = const()[name = tensor("denom_5_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_5_cast_fp16 = rsqrt(epsilon = denom_5_epsilon_0_to_fp16, x = var_244_cast_fp16)[name = tensor("denom_5_cast_fp16")]; + tensor out_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = denom_5_cast_fp16)[name = tensor("out_5_cast_fp16")]; + tensor input_5_gamma_0_to_fp16 = const()[name = tensor("input_5_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42545600)))]; + tensor input_5_beta_0_to_fp16 = const()[name = tensor("input_5_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42546432)))]; + tensor input_5_epsilon_0_to_fp16 = const()[name = tensor("input_5_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor 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("input_5_cast_fp16")]; + tensor var_255 = const()[name = tensor("op_255"), val = tensor([1, 1])]; + tensor var_257 = const()[name = tensor("op_257"), val = tensor([1, 1])]; + tensor input_7_pad_type_0 = const()[name = tensor("input_7_pad_type_0"), val = tensor("custom")]; + tensor input_7_pad_0 = const()[name = tensor("input_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_fc1_weight_to_fp16 = const()[name = tensor("layers_0_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42547264)))]; + tensor layers_0_fc1_bias_to_fp16 = const()[name = tensor("layers_0_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43726976)))]; + tensor input_7_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = var_257, groups = var_71, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = var_255, weight = layers_0_fc1_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_mode_0 = const()[name = tensor("input_9_mode_0"), val = tensor("EXACT")]; + tensor input_9_cast_fp16 = gelu(mode = input_9_mode_0, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor var_263 = const()[name = tensor("op_263"), val = tensor([1, 1])]; + tensor var_265 = const()[name = tensor("op_265"), val = tensor([1, 1])]; + tensor hidden_states_3_pad_type_0 = const()[name = tensor("hidden_states_3_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_3_pad_0 = const()[name = tensor("hidden_states_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_fc2_weight_to_fp16 = const()[name = tensor("layers_0_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43730112)))]; + tensor layers_0_fc2_bias_to_fp16 = const()[name = tensor("layers_0_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44909824)))]; + tensor hidden_states_3_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = var_265, groups = var_71, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = var_263, weight = layers_0_fc2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("hidden_states_3_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor var_278 = const()[name = tensor("op_278"), val = tensor(3)]; + tensor var_285 = const()[name = tensor("op_285"), val = tensor(1)]; + tensor var_286 = const()[name = tensor("op_286"), val = tensor(true)]; + tensor var_298 = const()[name = tensor("op_298"), val = tensor([1])]; + tensor channels_mean_7_cast_fp16 = reduce_mean(axes = var_298, keep_dims = var_286, x = inputs_7_cast_fp16)[name = tensor("channels_mean_7_cast_fp16")]; + tensor zero_mean_7_cast_fp16 = sub(x = inputs_7_cast_fp16, y = channels_mean_7_cast_fp16)[name = tensor("zero_mean_7_cast_fp16")]; + tensor zero_mean_sq_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = zero_mean_7_cast_fp16)[name = tensor("zero_mean_sq_7_cast_fp16")]; + tensor var_302 = const()[name = tensor("op_302"), val = tensor([1])]; + tensor var_303_cast_fp16 = reduce_mean(axes = var_302, keep_dims = var_286, x = zero_mean_sq_7_cast_fp16)[name = tensor("op_303_cast_fp16")]; + tensor var_304_to_fp16 = const()[name = tensor("op_304_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_305_cast_fp16 = add(x = var_303_cast_fp16, y = var_304_to_fp16)[name = tensor("op_305_cast_fp16")]; + tensor denom_7_epsilon_0_to_fp16 = const()[name = tensor("denom_7_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_7_cast_fp16 = rsqrt(epsilon = denom_7_epsilon_0_to_fp16, x = var_305_cast_fp16)[name = tensor("denom_7_cast_fp16")]; + tensor out_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = denom_7_cast_fp16)[name = tensor("out_7_cast_fp16")]; + tensor obj_15_gamma_0_to_fp16 = const()[name = tensor("obj_15_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44910656)))]; + tensor obj_15_beta_0_to_fp16 = const()[name = tensor("obj_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44911488)))]; + tensor obj_15_epsilon_0_to_fp16 = const()[name = tensor("obj_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor 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("obj_15_cast_fp16")]; + tensor var_320 = const()[name = tensor("op_320"), val = tensor([1, 1])]; + tensor var_322 = const()[name = tensor("op_322"), val = tensor([1, 1])]; + tensor query_5_pad_type_0 = const()[name = tensor("query_5_pad_type_0"), val = tensor("custom")]; + tensor query_5_pad_0 = const()[name = tensor("query_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44912320)))]; + tensor layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(45207296)))]; + tensor query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = var_322, groups = var_285, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = var_320, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("query_5_cast_fp16")]; + tensor var_326 = const()[name = tensor("op_326"), val = tensor([1, 1])]; + tensor var_328 = const()[name = tensor("op_328"), val = tensor([1, 1])]; + tensor current_key_3_pad_type_0 = const()[name = tensor("current_key_3_pad_type_0"), val = tensor("custom")]; + tensor current_key_3_pad_0 = const()[name = tensor("current_key_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(45208128)))]; + tensor current_key_3_cast_fp16 = conv(dilations = var_328, groups = var_285, pad = current_key_3_pad_0, pad_type = current_key_3_pad_type_0, strides = var_326, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("current_key_3_cast_fp16")]; + tensor var_333 = const()[name = tensor("op_333"), val = tensor([1, 1])]; + tensor var_335 = const()[name = tensor("op_335"), val = tensor([1, 1])]; + tensor current_value_3_pad_type_0 = const()[name = tensor("current_value_3_pad_type_0"), val = tensor("custom")]; + tensor current_value_3_pad_0 = const()[name = tensor("current_value_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(45503104)))]; + tensor layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(45798080)))]; + tensor current_value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = var_335, groups = var_285, pad = current_value_3_pad_0, pad_type = current_value_3_pad_type_0, strides = var_333, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("current_value_3_cast_fp16")]; + tensor var_342_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_126_cast_fp16)[name = tensor("op_342_cast_fp16")]; + tensor var_344_cast_fp16 = mul(x = var_47_cast_fp16_1, y = var_129_cast_fp16)[name = tensor("op_344_cast_fp16")]; + tensor key_5_cast_fp16 = add(x = var_342_cast_fp16, y = var_344_cast_fp16)[name = tensor("key_5_cast_fp16")]; + tensor var_346_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_126_cast_fp16)[name = tensor("op_346_cast_fp16")]; + tensor var_348_cast_fp16 = mul(x = var_54_cast_fp16_1, y = var_129_cast_fp16)[name = tensor("op_348_cast_fp16")]; + tensor value_5_cast_fp16 = add(x = var_346_cast_fp16, y = var_348_cast_fp16)[name = tensor("value_5_cast_fp16")]; + tensor var_351 = const()[name = tensor("op_351"), val = tensor([1, 6, 64, -1])]; + tensor var_352_cast_fp16 = reshape(shape = var_351, x = query_5_cast_fp16)[name = tensor("op_352_cast_fp16")]; + tensor var_353_to_fp16 = const()[name = tensor("op_353_to_fp16"), val = tensor(0x1p-3)]; + tensor var_354_cast_fp16 = mul(x = var_352_cast_fp16, y = var_353_to_fp16)[name = tensor("op_354_cast_fp16")]; + tensor var_355 = const()[name = tensor("op_355"), val = tensor([1, 6, 64, -1])]; + tensor var_356_cast_fp16 = reshape(shape = var_355, x = key_5_cast_fp16)[name = tensor("op_356_cast_fp16")]; + tensor mh_w_7_transpose_x_0 = const()[name = tensor("mh_w_7_transpose_x_0"), val = tensor(true)]; + tensor mh_w_7_transpose_y_0 = const()[name = tensor("mh_w_7_transpose_y_0"), val = tensor(false)]; + tensor 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("mh_w_7_cast_fp16")]; + tensor mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_147_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; + tensor var_364_cast_fp16 = softmax(axis = var_278, x = mh_w_9_cast_fp16)[name = tensor("op_364_cast_fp16")]; + tensor var_365 = const()[name = tensor("op_365"), val = tensor([1, 6, 64, -1])]; + tensor var_366_cast_fp16 = reshape(shape = var_365, x = value_5_cast_fp16)[name = tensor("op_366_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(true)]; + tensor 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("attn_5_cast_fp16")]; + tensor var_369 = const()[name = tensor("op_369"), val = tensor([1, 384, 1, -1])]; + tensor input_11_cast_fp16 = reshape(shape = var_369, x = attn_5_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor var_373 = const()[name = tensor("op_373"), val = tensor([1, 1])]; + tensor var_375 = const()[name = tensor("op_375"), val = tensor([1, 1])]; + tensor obj_21_pad_type_0 = const()[name = tensor("obj_21_pad_type_0"), val = tensor("custom")]; + tensor obj_21_pad_0 = const()[name = tensor("obj_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(45798912)))]; + tensor layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46093888)))]; + tensor obj_21_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = var_375, groups = var_285, pad = obj_21_pad_0, pad_type = obj_21_pad_type_0, strides = var_373, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("obj_21_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_21_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor var_385 = const()[name = tensor("op_385"), val = tensor([1])]; + tensor channels_mean_9_cast_fp16 = reduce_mean(axes = var_385, keep_dims = var_286, x = inputs_9_cast_fp16)[name = tensor("channels_mean_9_cast_fp16")]; + tensor zero_mean_9_cast_fp16 = sub(x = inputs_9_cast_fp16, y = channels_mean_9_cast_fp16)[name = tensor("zero_mean_9_cast_fp16")]; + tensor zero_mean_sq_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = zero_mean_9_cast_fp16)[name = tensor("zero_mean_sq_9_cast_fp16")]; + tensor var_389 = const()[name = tensor("op_389"), val = tensor([1])]; + tensor var_390_cast_fp16 = reduce_mean(axes = var_389, keep_dims = var_286, x = zero_mean_sq_9_cast_fp16)[name = tensor("op_390_cast_fp16")]; + tensor var_391_to_fp16 = const()[name = tensor("op_391_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_392_cast_fp16 = add(x = var_390_cast_fp16, y = var_391_to_fp16)[name = tensor("op_392_cast_fp16")]; + tensor denom_9_epsilon_0_to_fp16 = const()[name = tensor("denom_9_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_9_cast_fp16 = rsqrt(epsilon = denom_9_epsilon_0_to_fp16, x = var_392_cast_fp16)[name = tensor("denom_9_cast_fp16")]; + tensor out_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = denom_9_cast_fp16)[name = tensor("out_9_cast_fp16")]; + tensor obj_23_gamma_0_to_fp16 = const()[name = tensor("obj_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46094720)))]; + tensor obj_23_beta_0_to_fp16 = const()[name = tensor("obj_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46095552)))]; + tensor obj_23_epsilon_0_to_fp16 = const()[name = tensor("obj_23_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor 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("obj_23_cast_fp16")]; + tensor var_407 = const()[name = tensor("op_407"), val = tensor([1, 1])]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 1])]; + tensor query_7_pad_type_0 = const()[name = tensor("query_7_pad_type_0"), val = tensor("custom")]; + tensor query_7_pad_0 = const()[name = tensor("query_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46096384)))]; + tensor layers_1_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46391360)))]; + tensor query_7_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = var_409, groups = var_285, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = var_407, weight = layers_1_encoder_attn_q_proj_weight_to_fp16, x = obj_23_cast_fp16)[name = tensor("query_7_cast_fp16")]; + tensor var_413 = const()[name = tensor("op_413"), val = tensor([1, 1])]; + tensor var_415 = const()[name = tensor("op_415"), val = tensor([1, 1])]; + tensor key_7_pad_type_0 = const()[name = tensor("key_7_pad_type_0"), val = tensor("custom")]; + tensor key_7_pad_0 = const()[name = tensor("key_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46392192)))]; + tensor key_7_cast_fp16 = conv(dilations = var_415, groups = var_285, pad = key_7_pad_0, pad_type = key_7_pad_type_0, strides = var_413, weight = layers_1_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_7_cast_fp16")]; + tensor var_420 = const()[name = tensor("op_420"), val = tensor([1, 1])]; + tensor var_422 = const()[name = tensor("op_422"), val = tensor([1, 1])]; + tensor value_7_pad_type_0 = const()[name = tensor("value_7_pad_type_0"), val = tensor("custom")]; + tensor value_7_pad_0 = const()[name = tensor("value_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46687168)))]; + tensor layers_1_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46982144)))]; + tensor value_7_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_bias_to_fp16, dilations = var_422, groups = var_285, pad = value_7_pad_0, pad_type = value_7_pad_type_0, strides = var_420, weight = layers_1_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_7_cast_fp16")]; + tensor var_426 = const()[name = tensor("op_426"), val = tensor([1, 6, 64, -1])]; + tensor var_427_cast_fp16 = reshape(shape = var_426, x = query_7_cast_fp16)[name = tensor("op_427_cast_fp16")]; + tensor var_428_to_fp16 = const()[name = tensor("op_428_to_fp16"), val = tensor(0x1p-3)]; + tensor var_429_cast_fp16 = mul(x = var_427_cast_fp16, y = var_428_to_fp16)[name = tensor("op_429_cast_fp16")]; + tensor var_430 = const()[name = tensor("op_430"), val = tensor([1, 6, 64, -1])]; + tensor var_431_cast_fp16 = reshape(shape = var_430, x = key_7_cast_fp16)[name = tensor("op_431_cast_fp16")]; + tensor mh_w_11_transpose_x_0 = const()[name = tensor("mh_w_11_transpose_x_0"), val = tensor(true)]; + tensor mh_w_11_transpose_y_0 = const()[name = tensor("mh_w_11_transpose_y_0"), val = tensor(false)]; + tensor 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("mh_w_11_cast_fp16")]; + tensor obj_27_cast_fp16 = softmax(axis = var_278, x = mh_w_11_cast_fp16)[name = tensor("obj_27_cast_fp16")]; + tensor var_435 = const()[name = tensor("op_435"), val = tensor([1, 6, 64, -1])]; + tensor var_436_cast_fp16 = reshape(shape = var_435, x = value_7_cast_fp16)[name = tensor("op_436_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(true)]; + tensor 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("attn_7_cast_fp16")]; + tensor var_439 = const()[name = tensor("op_439"), val = tensor([1, 384, 1, -1])]; + tensor input_13_cast_fp16 = reshape(shape = var_439, x = attn_7_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 1])]; + tensor var_445 = const()[name = tensor("op_445"), val = tensor([1, 1])]; + tensor obj_25_pad_type_0 = const()[name = tensor("obj_25_pad_type_0"), val = tensor("custom")]; + tensor obj_25_pad_0 = const()[name = tensor("obj_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46982976)))]; + tensor layers_1_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47277952)))]; + tensor obj_25_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = var_445, groups = var_285, pad = obj_25_pad_0, pad_type = obj_25_pad_type_0, strides = var_443, weight = layers_1_encoder_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("obj_25_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_25_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor var_454 = const()[name = tensor("op_454"), val = tensor([1])]; + tensor channels_mean_11_cast_fp16 = reduce_mean(axes = var_454, keep_dims = var_286, x = inputs_11_cast_fp16)[name = tensor("channels_mean_11_cast_fp16")]; + tensor zero_mean_11_cast_fp16 = sub(x = inputs_11_cast_fp16, y = channels_mean_11_cast_fp16)[name = tensor("zero_mean_11_cast_fp16")]; + tensor zero_mean_sq_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = zero_mean_11_cast_fp16)[name = tensor("zero_mean_sq_11_cast_fp16")]; + tensor var_458 = const()[name = tensor("op_458"), val = tensor([1])]; + tensor var_459_cast_fp16 = reduce_mean(axes = var_458, keep_dims = var_286, x = zero_mean_sq_11_cast_fp16)[name = tensor("op_459_cast_fp16")]; + tensor var_460_to_fp16 = const()[name = tensor("op_460_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_461_cast_fp16 = add(x = var_459_cast_fp16, y = var_460_to_fp16)[name = tensor("op_461_cast_fp16")]; + tensor denom_11_epsilon_0_to_fp16 = const()[name = tensor("denom_11_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_11_cast_fp16 = rsqrt(epsilon = denom_11_epsilon_0_to_fp16, x = var_461_cast_fp16)[name = tensor("denom_11_cast_fp16")]; + tensor out_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = denom_11_cast_fp16)[name = tensor("out_11_cast_fp16")]; + tensor input_15_gamma_0_to_fp16 = const()[name = tensor("input_15_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47278784)))]; + tensor input_15_beta_0_to_fp16 = const()[name = tensor("input_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47279616)))]; + tensor input_15_epsilon_0_to_fp16 = const()[name = tensor("input_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor 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("input_15_cast_fp16")]; + tensor var_472 = const()[name = tensor("op_472"), val = tensor([1, 1])]; + tensor var_474 = const()[name = tensor("op_474"), val = tensor([1, 1])]; + tensor input_17_pad_type_0 = const()[name = tensor("input_17_pad_type_0"), val = tensor("custom")]; + tensor input_17_pad_0 = const()[name = tensor("input_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_fc1_weight_to_fp16 = const()[name = tensor("layers_1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47280448)))]; + tensor layers_1_fc1_bias_to_fp16 = const()[name = tensor("layers_1_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48460160)))]; + tensor input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = var_474, groups = var_285, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = var_472, weight = layers_1_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_mode_0 = const()[name = tensor("input_19_mode_0"), val = tensor("EXACT")]; + tensor input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor var_480 = const()[name = tensor("op_480"), val = tensor([1, 1])]; + tensor var_482 = const()[name = tensor("op_482"), val = tensor([1, 1])]; + tensor hidden_states_5_pad_type_0 = const()[name = tensor("hidden_states_5_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_5_pad_0 = const()[name = tensor("hidden_states_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_fc2_weight_to_fp16 = const()[name = tensor("layers_1_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48463296)))]; + tensor layers_1_fc2_bias_to_fp16 = const()[name = tensor("layers_1_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49643008)))]; + tensor hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = var_482, groups = var_285, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = var_480, weight = layers_1_fc2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("hidden_states_5_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_496 = const()[name = tensor("op_496"), val = tensor(3)]; + tensor var_503 = const()[name = tensor("op_503"), val = tensor(1)]; + tensor var_504 = const()[name = tensor("op_504"), val = tensor(true)]; + tensor var_516 = const()[name = tensor("op_516"), val = tensor([1])]; + tensor channels_mean_13_cast_fp16 = reduce_mean(axes = var_516, keep_dims = var_504, x = inputs_13_cast_fp16)[name = tensor("channels_mean_13_cast_fp16")]; + tensor zero_mean_13_cast_fp16 = sub(x = inputs_13_cast_fp16, y = channels_mean_13_cast_fp16)[name = tensor("zero_mean_13_cast_fp16")]; + tensor zero_mean_sq_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = zero_mean_13_cast_fp16)[name = tensor("zero_mean_sq_13_cast_fp16")]; + tensor var_520 = const()[name = tensor("op_520"), val = tensor([1])]; + tensor var_521_cast_fp16 = reduce_mean(axes = var_520, keep_dims = var_504, x = zero_mean_sq_13_cast_fp16)[name = tensor("op_521_cast_fp16")]; + tensor var_522_to_fp16 = const()[name = tensor("op_522_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_523_cast_fp16 = add(x = var_521_cast_fp16, y = var_522_to_fp16)[name = tensor("op_523_cast_fp16")]; + tensor denom_13_epsilon_0_to_fp16 = const()[name = tensor("denom_13_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_13_cast_fp16 = rsqrt(epsilon = denom_13_epsilon_0_to_fp16, x = var_523_cast_fp16)[name = tensor("denom_13_cast_fp16")]; + tensor out_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = denom_13_cast_fp16)[name = tensor("out_13_cast_fp16")]; + tensor obj_29_gamma_0_to_fp16 = const()[name = tensor("obj_29_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49643840)))]; + tensor obj_29_beta_0_to_fp16 = const()[name = tensor("obj_29_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49644672)))]; + tensor obj_29_epsilon_0_to_fp16 = const()[name = tensor("obj_29_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor 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("obj_29_cast_fp16")]; + tensor var_538 = const()[name = tensor("op_538"), val = tensor([1, 1])]; + tensor var_540 = const()[name = tensor("op_540"), val = tensor([1, 1])]; + tensor query_9_pad_type_0 = const()[name = tensor("query_9_pad_type_0"), val = tensor("custom")]; + tensor query_9_pad_0 = const()[name = tensor("query_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49645504)))]; + tensor layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49940480)))]; + tensor query_9_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = var_540, groups = var_503, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = var_538, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("query_9_cast_fp16")]; + tensor var_544 = const()[name = tensor("op_544"), val = tensor([1, 1])]; + tensor var_546 = const()[name = tensor("op_546"), val = tensor([1, 1])]; + tensor current_key_5_pad_type_0 = const()[name = tensor("current_key_5_pad_type_0"), val = tensor("custom")]; + tensor current_key_5_pad_0 = const()[name = tensor("current_key_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49941312)))]; + tensor current_key_5_cast_fp16 = conv(dilations = var_546, groups = var_503, pad = current_key_5_pad_0, pad_type = current_key_5_pad_type_0, strides = var_544, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("current_key_5_cast_fp16")]; + tensor var_551 = const()[name = tensor("op_551"), val = tensor([1, 1])]; + tensor var_553 = const()[name = tensor("op_553"), val = tensor([1, 1])]; + tensor current_value_5_pad_type_0 = const()[name = tensor("current_value_5_pad_type_0"), val = tensor("custom")]; + tensor current_value_5_pad_0 = const()[name = tensor("current_value_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50236288)))]; + tensor layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50531264)))]; + tensor current_value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = var_553, groups = var_503, pad = current_value_5_pad_0, pad_type = current_value_5_pad_type_0, strides = var_551, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("current_value_5_cast_fp16")]; + tensor var_560_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_126_cast_fp16)[name = tensor("op_560_cast_fp16")]; + tensor var_562_cast_fp16 = mul(x = var_47_cast_fp16_2, y = var_129_cast_fp16)[name = tensor("op_562_cast_fp16")]; + tensor key_9_cast_fp16 = add(x = var_560_cast_fp16, y = var_562_cast_fp16)[name = tensor("key_9_cast_fp16")]; + tensor var_564_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_126_cast_fp16)[name = tensor("op_564_cast_fp16")]; + tensor var_566_cast_fp16 = mul(x = var_54_cast_fp16_2, y = var_129_cast_fp16)[name = tensor("op_566_cast_fp16")]; + tensor value_9_cast_fp16 = add(x = var_564_cast_fp16, y = var_566_cast_fp16)[name = tensor("value_9_cast_fp16")]; + tensor var_569 = const()[name = tensor("op_569"), val = tensor([1, 6, 64, -1])]; + tensor var_570_cast_fp16 = reshape(shape = var_569, x = query_9_cast_fp16)[name = tensor("op_570_cast_fp16")]; + tensor var_571_to_fp16 = const()[name = tensor("op_571_to_fp16"), val = tensor(0x1p-3)]; + tensor var_572_cast_fp16 = mul(x = var_570_cast_fp16, y = var_571_to_fp16)[name = tensor("op_572_cast_fp16")]; + tensor var_573 = const()[name = tensor("op_573"), val = tensor([1, 6, 64, -1])]; + tensor var_574_cast_fp16 = reshape(shape = var_573, x = key_9_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor mh_w_13_transpose_x_0 = const()[name = tensor("mh_w_13_transpose_x_0"), val = tensor(true)]; + tensor mh_w_13_transpose_y_0 = const()[name = tensor("mh_w_13_transpose_y_0"), val = tensor(false)]; + tensor mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_572_cast_fp16, y = var_574_cast_fp16)[name = tensor("mh_w_13_cast_fp16")]; + tensor mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_147_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; + tensor var_582_cast_fp16 = softmax(axis = var_496, x = mh_w_15_cast_fp16)[name = tensor("op_582_cast_fp16")]; + tensor var_583 = const()[name = tensor("op_583"), val = tensor([1, 6, 64, -1])]; + tensor var_584_cast_fp16 = reshape(shape = var_583, x = value_9_cast_fp16)[name = tensor("op_584_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(true)]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_584_cast_fp16, y = var_582_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_587 = const()[name = tensor("op_587"), val = tensor([1, 384, 1, -1])]; + tensor input_21_cast_fp16 = reshape(shape = var_587, x = attn_9_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 1])]; + tensor var_593 = const()[name = tensor("op_593"), val = tensor([1, 1])]; + tensor obj_35_pad_type_0 = const()[name = tensor("obj_35_pad_type_0"), val = tensor("custom")]; + tensor obj_35_pad_0 = const()[name = tensor("obj_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50532096)))]; + tensor layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50827072)))]; + tensor obj_35_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = var_593, groups = var_503, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = var_591, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("obj_35_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_35_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor var_603 = const()[name = tensor("op_603"), val = tensor([1])]; + tensor channels_mean_15_cast_fp16 = reduce_mean(axes = var_603, keep_dims = var_504, x = inputs_15_cast_fp16)[name = tensor("channels_mean_15_cast_fp16")]; + tensor zero_mean_15_cast_fp16 = sub(x = inputs_15_cast_fp16, y = channels_mean_15_cast_fp16)[name = tensor("zero_mean_15_cast_fp16")]; + tensor zero_mean_sq_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = zero_mean_15_cast_fp16)[name = tensor("zero_mean_sq_15_cast_fp16")]; + tensor var_607 = const()[name = tensor("op_607"), val = tensor([1])]; + tensor var_608_cast_fp16 = reduce_mean(axes = var_607, keep_dims = var_504, x = zero_mean_sq_15_cast_fp16)[name = tensor("op_608_cast_fp16")]; + tensor var_609_to_fp16 = const()[name = tensor("op_609_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_610_cast_fp16 = add(x = var_608_cast_fp16, y = var_609_to_fp16)[name = tensor("op_610_cast_fp16")]; + tensor denom_15_epsilon_0_to_fp16 = const()[name = tensor("denom_15_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_15_cast_fp16 = rsqrt(epsilon = denom_15_epsilon_0_to_fp16, x = var_610_cast_fp16)[name = tensor("denom_15_cast_fp16")]; + tensor out_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = denom_15_cast_fp16)[name = tensor("out_15_cast_fp16")]; + tensor obj_37_gamma_0_to_fp16 = const()[name = tensor("obj_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50827904)))]; + tensor obj_37_beta_0_to_fp16 = const()[name = tensor("obj_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50828736)))]; + tensor obj_37_epsilon_0_to_fp16 = const()[name = tensor("obj_37_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor 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("obj_37_cast_fp16")]; + tensor var_625 = const()[name = tensor("op_625"), val = tensor([1, 1])]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor([1, 1])]; + tensor query_11_pad_type_0 = const()[name = tensor("query_11_pad_type_0"), val = tensor("custom")]; + tensor query_11_pad_0 = const()[name = tensor("query_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50829568)))]; + tensor layers_2_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51124544)))]; + tensor query_11_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_bias_to_fp16, dilations = var_627, groups = var_503, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = var_625, weight = layers_2_encoder_attn_q_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("query_11_cast_fp16")]; + tensor var_631 = const()[name = tensor("op_631"), val = tensor([1, 1])]; + tensor var_633 = const()[name = tensor("op_633"), val = tensor([1, 1])]; + tensor key_11_pad_type_0 = const()[name = tensor("key_11_pad_type_0"), val = tensor("custom")]; + tensor key_11_pad_0 = const()[name = tensor("key_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51125376)))]; + tensor key_11_cast_fp16 = conv(dilations = var_633, groups = var_503, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = var_631, weight = layers_2_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_11_cast_fp16")]; + tensor var_638 = const()[name = tensor("op_638"), val = tensor([1, 1])]; + tensor var_640 = const()[name = tensor("op_640"), val = tensor([1, 1])]; + tensor value_11_pad_type_0 = const()[name = tensor("value_11_pad_type_0"), val = tensor("custom")]; + tensor value_11_pad_0 = const()[name = tensor("value_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51420352)))]; + tensor layers_2_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51715328)))]; + tensor value_11_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_bias_to_fp16, dilations = var_640, groups = var_503, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = var_638, weight = layers_2_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_11_cast_fp16")]; + tensor var_644 = const()[name = tensor("op_644"), val = tensor([1, 6, 64, -1])]; + tensor var_645_cast_fp16 = reshape(shape = var_644, x = query_11_cast_fp16)[name = tensor("op_645_cast_fp16")]; + tensor var_646_to_fp16 = const()[name = tensor("op_646_to_fp16"), val = tensor(0x1p-3)]; + tensor var_647_cast_fp16 = mul(x = var_645_cast_fp16, y = var_646_to_fp16)[name = tensor("op_647_cast_fp16")]; + tensor var_648 = const()[name = tensor("op_648"), val = tensor([1, 6, 64, -1])]; + tensor var_649_cast_fp16 = reshape(shape = var_648, x = key_11_cast_fp16)[name = tensor("op_649_cast_fp16")]; + tensor mh_w_17_transpose_x_0 = const()[name = tensor("mh_w_17_transpose_x_0"), val = tensor(true)]; + tensor mh_w_17_transpose_y_0 = const()[name = tensor("mh_w_17_transpose_y_0"), val = tensor(false)]; + tensor mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_647_cast_fp16, y = var_649_cast_fp16)[name = tensor("mh_w_17_cast_fp16")]; + tensor obj_41_cast_fp16 = softmax(axis = var_496, x = mh_w_17_cast_fp16)[name = tensor("obj_41_cast_fp16")]; + tensor var_653 = const()[name = tensor("op_653"), val = tensor([1, 6, 64, -1])]; + tensor var_654_cast_fp16 = reshape(shape = var_653, x = value_11_cast_fp16)[name = tensor("op_654_cast_fp16")]; + tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; + tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(true)]; + tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_654_cast_fp16, y = obj_41_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_657 = const()[name = tensor("op_657"), val = tensor([1, 384, 1, -1])]; + tensor input_23_cast_fp16 = reshape(shape = var_657, x = attn_11_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor var_661 = const()[name = tensor("op_661"), val = tensor([1, 1])]; + tensor var_663 = const()[name = tensor("op_663"), val = tensor([1, 1])]; + tensor obj_39_pad_type_0 = const()[name = tensor("obj_39_pad_type_0"), val = tensor("custom")]; + tensor obj_39_pad_0 = const()[name = tensor("obj_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51716160)))]; + tensor layers_2_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52011136)))]; + tensor obj_39_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_bias_to_fp16, dilations = var_663, groups = var_503, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = var_661, weight = layers_2_encoder_attn_o_proj_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("obj_39_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_39_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor var_672 = const()[name = tensor("op_672"), val = tensor([1])]; + tensor channels_mean_17_cast_fp16 = reduce_mean(axes = var_672, keep_dims = var_504, x = inputs_17_cast_fp16)[name = tensor("channels_mean_17_cast_fp16")]; + tensor zero_mean_17_cast_fp16 = sub(x = inputs_17_cast_fp16, y = channels_mean_17_cast_fp16)[name = tensor("zero_mean_17_cast_fp16")]; + tensor zero_mean_sq_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = zero_mean_17_cast_fp16)[name = tensor("zero_mean_sq_17_cast_fp16")]; + tensor var_676 = const()[name = tensor("op_676"), val = tensor([1])]; + tensor var_677_cast_fp16 = reduce_mean(axes = var_676, keep_dims = var_504, x = zero_mean_sq_17_cast_fp16)[name = tensor("op_677_cast_fp16")]; + tensor var_678_to_fp16 = const()[name = tensor("op_678_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_679_cast_fp16 = add(x = var_677_cast_fp16, y = var_678_to_fp16)[name = tensor("op_679_cast_fp16")]; + tensor denom_17_epsilon_0_to_fp16 = const()[name = tensor("denom_17_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_17_cast_fp16 = rsqrt(epsilon = denom_17_epsilon_0_to_fp16, x = var_679_cast_fp16)[name = tensor("denom_17_cast_fp16")]; + tensor out_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = denom_17_cast_fp16)[name = tensor("out_17_cast_fp16")]; + tensor input_25_gamma_0_to_fp16 = const()[name = tensor("input_25_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52011968)))]; + tensor input_25_beta_0_to_fp16 = const()[name = tensor("input_25_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52012800)))]; + tensor input_25_epsilon_0_to_fp16 = const()[name = tensor("input_25_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor 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("input_25_cast_fp16")]; + tensor var_690 = const()[name = tensor("op_690"), val = tensor([1, 1])]; + tensor var_692 = const()[name = tensor("op_692"), val = tensor([1, 1])]; + tensor input_27_pad_type_0 = const()[name = tensor("input_27_pad_type_0"), val = tensor("custom")]; + tensor input_27_pad_0 = const()[name = tensor("input_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_fc1_weight_to_fp16 = const()[name = tensor("layers_2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52013632)))]; + tensor layers_2_fc1_bias_to_fp16 = const()[name = tensor("layers_2_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53193344)))]; + tensor input_27_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = var_692, groups = var_503, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = var_690, weight = layers_2_fc1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_mode_0 = const()[name = tensor("input_29_mode_0"), val = tensor("EXACT")]; + tensor input_29_cast_fp16 = gelu(mode = input_29_mode_0, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor var_698 = const()[name = tensor("op_698"), val = tensor([1, 1])]; + tensor var_700 = const()[name = tensor("op_700"), val = tensor([1, 1])]; + tensor hidden_states_7_pad_type_0 = const()[name = tensor("hidden_states_7_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_7_pad_0 = const()[name = tensor("hidden_states_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_fc2_weight_to_fp16 = const()[name = tensor("layers_2_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53196480)))]; + tensor layers_2_fc2_bias_to_fp16 = const()[name = tensor("layers_2_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54376192)))]; + tensor hidden_states_7_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = var_700, groups = var_503, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = var_698, weight = layers_2_fc2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("hidden_states_7_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; + tensor var_714 = const()[name = tensor("op_714"), val = tensor(3)]; + tensor var_721 = const()[name = tensor("op_721"), val = tensor(1)]; + tensor var_722 = const()[name = tensor("op_722"), val = tensor(true)]; + tensor var_734 = const()[name = tensor("op_734"), val = tensor([1])]; + tensor channels_mean_19_cast_fp16 = reduce_mean(axes = var_734, keep_dims = var_722, x = inputs_19_cast_fp16)[name = tensor("channels_mean_19_cast_fp16")]; + tensor zero_mean_19_cast_fp16 = sub(x = inputs_19_cast_fp16, y = channels_mean_19_cast_fp16)[name = tensor("zero_mean_19_cast_fp16")]; + tensor zero_mean_sq_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = zero_mean_19_cast_fp16)[name = tensor("zero_mean_sq_19_cast_fp16")]; + tensor var_738 = const()[name = tensor("op_738"), val = tensor([1])]; + tensor var_739_cast_fp16 = reduce_mean(axes = var_738, keep_dims = var_722, x = zero_mean_sq_19_cast_fp16)[name = tensor("op_739_cast_fp16")]; + tensor var_740_to_fp16 = const()[name = tensor("op_740_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_741_cast_fp16 = add(x = var_739_cast_fp16, y = var_740_to_fp16)[name = tensor("op_741_cast_fp16")]; + tensor denom_19_epsilon_0_to_fp16 = const()[name = tensor("denom_19_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_19_cast_fp16 = rsqrt(epsilon = denom_19_epsilon_0_to_fp16, x = var_741_cast_fp16)[name = tensor("denom_19_cast_fp16")]; + tensor out_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = denom_19_cast_fp16)[name = tensor("out_19_cast_fp16")]; + tensor obj_43_gamma_0_to_fp16 = const()[name = tensor("obj_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54377024)))]; + tensor obj_43_beta_0_to_fp16 = const()[name = tensor("obj_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54377856)))]; + tensor obj_43_epsilon_0_to_fp16 = const()[name = tensor("obj_43_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor 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("obj_43_cast_fp16")]; + tensor var_756 = const()[name = tensor("op_756"), val = tensor([1, 1])]; + tensor var_758 = const()[name = tensor("op_758"), val = tensor([1, 1])]; + tensor query_13_pad_type_0 = const()[name = tensor("query_13_pad_type_0"), val = tensor("custom")]; + tensor query_13_pad_0 = const()[name = tensor("query_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54378688)))]; + tensor layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54673664)))]; + tensor query_13_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = var_758, groups = var_721, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = var_756, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("query_13_cast_fp16")]; + tensor var_762 = const()[name = tensor("op_762"), val = tensor([1, 1])]; + tensor var_764 = const()[name = tensor("op_764"), val = tensor([1, 1])]; + tensor current_key_pad_type_0 = const()[name = tensor("current_key_pad_type_0"), val = tensor("custom")]; + tensor current_key_pad_0 = const()[name = tensor("current_key_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54674496)))]; + tensor current_key_cast_fp16 = conv(dilations = var_764, groups = var_721, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = var_762, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("current_key_cast_fp16")]; + tensor var_769 = const()[name = tensor("op_769"), val = tensor([1, 1])]; + tensor var_771 = const()[name = tensor("op_771"), val = tensor([1, 1])]; + tensor current_value_pad_type_0 = const()[name = tensor("current_value_pad_type_0"), val = tensor("custom")]; + tensor current_value_pad_0 = const()[name = tensor("current_value_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54969472)))]; + tensor layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(55264448)))]; + tensor current_value_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = var_771, groups = var_721, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = var_769, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("current_value_cast_fp16")]; + tensor var_778_cast_fp16 = mul(x = current_key_cast_fp16, y = var_126_cast_fp16)[name = tensor("op_778_cast_fp16")]; + tensor var_780_cast_fp16 = mul(x = var_47_cast_fp16_3, y = var_129_cast_fp16)[name = tensor("op_780_cast_fp16")]; + tensor key_13_cast_fp16 = add(x = var_778_cast_fp16, y = var_780_cast_fp16)[name = tensor("key_13_cast_fp16")]; + tensor var_782_cast_fp16 = mul(x = current_value_cast_fp16, y = var_126_cast_fp16)[name = tensor("op_782_cast_fp16")]; + tensor var_784_cast_fp16 = mul(x = var_54_cast_fp16_3, y = var_129_cast_fp16)[name = tensor("op_784_cast_fp16")]; + tensor value_13_cast_fp16 = add(x = var_782_cast_fp16, y = var_784_cast_fp16)[name = tensor("value_13_cast_fp16")]; + tensor var_787 = const()[name = tensor("op_787"), val = tensor([1, 6, 64, -1])]; + tensor var_788_cast_fp16 = reshape(shape = var_787, x = query_13_cast_fp16)[name = tensor("op_788_cast_fp16")]; + tensor var_789_to_fp16 = const()[name = tensor("op_789_to_fp16"), val = tensor(0x1p-3)]; + tensor var_790_cast_fp16 = mul(x = var_788_cast_fp16, y = var_789_to_fp16)[name = tensor("op_790_cast_fp16")]; + tensor var_791 = const()[name = tensor("op_791"), val = tensor([1, 6, 64, -1])]; + tensor var_792_cast_fp16 = reshape(shape = var_791, x = key_13_cast_fp16)[name = tensor("op_792_cast_fp16")]; + tensor mh_w_19_transpose_x_0 = const()[name = tensor("mh_w_19_transpose_x_0"), val = tensor(true)]; + tensor mh_w_19_transpose_y_0 = const()[name = tensor("mh_w_19_transpose_y_0"), val = tensor(false)]; + tensor mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_790_cast_fp16, y = var_792_cast_fp16)[name = tensor("mh_w_19_cast_fp16")]; + tensor mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_147_cast_fp16)[name = tensor("mh_w_21_cast_fp16")]; + tensor var_800_cast_fp16 = softmax(axis = var_714, x = mh_w_21_cast_fp16)[name = tensor("op_800_cast_fp16")]; + tensor var_801 = const()[name = tensor("op_801"), val = tensor([1, 6, 64, -1])]; + tensor var_802_cast_fp16 = reshape(shape = var_801, x = value_13_cast_fp16)[name = tensor("op_802_cast_fp16")]; + tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; + tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(true)]; + tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_802_cast_fp16, y = var_800_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 384, 1, -1])]; + tensor input_31_cast_fp16 = reshape(shape = var_805, x = attn_13_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor var_809 = const()[name = tensor("op_809"), val = tensor([1, 1])]; + tensor var_811 = const()[name = tensor("op_811"), val = tensor([1, 1])]; + tensor obj_49_pad_type_0 = const()[name = tensor("obj_49_pad_type_0"), val = tensor("custom")]; + tensor obj_49_pad_0 = const()[name = tensor("obj_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(55265280)))]; + tensor layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(55560256)))]; + tensor obj_49_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = var_811, groups = var_721, pad = obj_49_pad_0, pad_type = obj_49_pad_type_0, strides = var_809, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("obj_49_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_49_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor var_821 = const()[name = tensor("op_821"), val = tensor([1])]; + tensor channels_mean_21_cast_fp16 = reduce_mean(axes = var_821, keep_dims = var_722, x = inputs_21_cast_fp16)[name = tensor("channels_mean_21_cast_fp16")]; + tensor zero_mean_21_cast_fp16 = sub(x = inputs_21_cast_fp16, y = channels_mean_21_cast_fp16)[name = tensor("zero_mean_21_cast_fp16")]; + tensor zero_mean_sq_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = zero_mean_21_cast_fp16)[name = tensor("zero_mean_sq_21_cast_fp16")]; + tensor var_825 = const()[name = tensor("op_825"), val = tensor([1])]; + tensor var_826_cast_fp16 = reduce_mean(axes = var_825, keep_dims = var_722, x = zero_mean_sq_21_cast_fp16)[name = tensor("op_826_cast_fp16")]; + tensor var_827_to_fp16 = const()[name = tensor("op_827_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_828_cast_fp16 = add(x = var_826_cast_fp16, y = var_827_to_fp16)[name = tensor("op_828_cast_fp16")]; + tensor denom_21_epsilon_0_to_fp16 = const()[name = tensor("denom_21_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_21_cast_fp16 = rsqrt(epsilon = denom_21_epsilon_0_to_fp16, x = var_828_cast_fp16)[name = tensor("denom_21_cast_fp16")]; + tensor out_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = denom_21_cast_fp16)[name = tensor("out_21_cast_fp16")]; + tensor obj_51_gamma_0_to_fp16 = const()[name = tensor("obj_51_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(55561088)))]; + tensor obj_51_beta_0_to_fp16 = const()[name = tensor("obj_51_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(55561920)))]; + tensor obj_51_epsilon_0_to_fp16 = const()[name = tensor("obj_51_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor 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("obj_51_cast_fp16")]; + tensor var_843 = const()[name = tensor("op_843"), val = tensor([1, 1])]; + tensor var_845 = const()[name = tensor("op_845"), val = tensor([1, 1])]; + tensor query_pad_type_0 = const()[name = tensor("query_pad_type_0"), val = tensor("custom")]; + tensor query_pad_0 = const()[name = tensor("query_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(55562752)))]; + tensor layers_3_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(55857728)))]; + tensor query_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_bias_to_fp16, dilations = var_845, groups = var_721, pad = query_pad_0, pad_type = query_pad_type_0, strides = var_843, weight = layers_3_encoder_attn_q_proj_weight_to_fp16, x = obj_51_cast_fp16)[name = tensor("query_cast_fp16")]; + tensor var_849 = const()[name = tensor("op_849"), val = tensor([1, 1])]; + tensor var_851 = const()[name = tensor("op_851"), val = tensor([1, 1])]; + tensor key_pad_type_0 = const()[name = tensor("key_pad_type_0"), val = tensor("custom")]; + tensor key_pad_0 = const()[name = tensor("key_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(55858560)))]; + tensor key_cast_fp16 = conv(dilations = var_851, groups = var_721, pad = key_pad_0, pad_type = key_pad_type_0, strides = var_849, weight = layers_3_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_cast_fp16")]; + tensor var_856 = const()[name = tensor("op_856"), val = tensor([1, 1])]; + tensor var_858 = const()[name = tensor("op_858"), val = tensor([1, 1])]; + tensor value_pad_type_0 = const()[name = tensor("value_pad_type_0"), val = tensor("custom")]; + tensor value_pad_0 = const()[name = tensor("value_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56153536)))]; + tensor layers_3_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56448512)))]; + tensor value_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_bias_to_fp16, dilations = var_858, groups = var_721, pad = value_pad_0, pad_type = value_pad_type_0, strides = var_856, weight = layers_3_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_cast_fp16")]; + tensor var_862 = const()[name = tensor("op_862"), val = tensor([1, 6, 64, -1])]; + tensor var_863_cast_fp16 = reshape(shape = var_862, x = query_cast_fp16)[name = tensor("op_863_cast_fp16")]; + tensor var_864_to_fp16 = const()[name = tensor("op_864_to_fp16"), val = tensor(0x1p-3)]; + tensor var_865_cast_fp16 = mul(x = var_863_cast_fp16, y = var_864_to_fp16)[name = tensor("op_865_cast_fp16")]; + tensor var_866 = const()[name = tensor("op_866"), val = tensor([1, 6, 64, -1])]; + tensor var_867_cast_fp16 = reshape(shape = var_866, x = key_cast_fp16)[name = tensor("op_867_cast_fp16")]; + tensor mh_w_transpose_x_0 = const()[name = tensor("mh_w_transpose_x_0"), val = tensor(true)]; + tensor mh_w_transpose_y_0 = const()[name = tensor("mh_w_transpose_y_0"), val = tensor(false)]; + tensor mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_865_cast_fp16, y = var_867_cast_fp16)[name = tensor("mh_w_cast_fp16")]; + tensor obj_55_cast_fp16 = softmax(axis = var_714, x = mh_w_cast_fp16)[name = tensor("obj_55_cast_fp16")]; + tensor var_871 = const()[name = tensor("op_871"), val = tensor([1, 6, 64, -1])]; + tensor var_872_cast_fp16 = reshape(shape = var_871, x = value_cast_fp16)[name = tensor("op_872_cast_fp16")]; + tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; + tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(true)]; + tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_872_cast_fp16, y = obj_55_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_875 = const()[name = tensor("op_875"), val = tensor([1, 384, 1, -1])]; + tensor input_33_cast_fp16 = reshape(shape = var_875, x = attn_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor var_879 = const()[name = tensor("op_879"), val = tensor([1, 1])]; + tensor var_881 = const()[name = tensor("op_881"), val = tensor([1, 1])]; + tensor obj_53_pad_type_0 = const()[name = tensor("obj_53_pad_type_0"), val = tensor("custom")]; + tensor obj_53_pad_0 = const()[name = tensor("obj_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56449344)))]; + tensor layers_3_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56744320)))]; + tensor obj_53_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_bias_to_fp16, dilations = var_881, groups = var_721, pad = obj_53_pad_0, pad_type = obj_53_pad_type_0, strides = var_879, weight = layers_3_encoder_attn_o_proj_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("obj_53_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_53_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor var_890 = const()[name = tensor("op_890"), val = tensor([1])]; + tensor channels_mean_23_cast_fp16 = reduce_mean(axes = var_890, keep_dims = var_722, x = inputs_23_cast_fp16)[name = tensor("channels_mean_23_cast_fp16")]; + tensor zero_mean_23_cast_fp16 = sub(x = inputs_23_cast_fp16, y = channels_mean_23_cast_fp16)[name = tensor("zero_mean_23_cast_fp16")]; + tensor zero_mean_sq_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = zero_mean_23_cast_fp16)[name = tensor("zero_mean_sq_23_cast_fp16")]; + tensor var_894 = const()[name = tensor("op_894"), val = tensor([1])]; + tensor var_895_cast_fp16 = reduce_mean(axes = var_894, keep_dims = var_722, x = zero_mean_sq_23_cast_fp16)[name = tensor("op_895_cast_fp16")]; + tensor var_896_to_fp16 = const()[name = tensor("op_896_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_897_cast_fp16 = add(x = var_895_cast_fp16, y = var_896_to_fp16)[name = tensor("op_897_cast_fp16")]; + tensor denom_23_epsilon_0_to_fp16 = const()[name = tensor("denom_23_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_23_cast_fp16 = rsqrt(epsilon = denom_23_epsilon_0_to_fp16, x = var_897_cast_fp16)[name = tensor("denom_23_cast_fp16")]; + tensor out_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = denom_23_cast_fp16)[name = tensor("out_23_cast_fp16")]; + tensor input_35_gamma_0_to_fp16 = const()[name = tensor("input_35_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56745152)))]; + tensor input_35_beta_0_to_fp16 = const()[name = tensor("input_35_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56745984)))]; + tensor input_35_epsilon_0_to_fp16 = const()[name = tensor("input_35_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor 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("input_35_cast_fp16")]; + tensor var_908 = const()[name = tensor("op_908"), val = tensor([1, 1])]; + tensor var_910 = const()[name = tensor("op_910"), val = tensor([1, 1])]; + tensor input_37_pad_type_0 = const()[name = tensor("input_37_pad_type_0"), val = tensor("custom")]; + tensor input_37_pad_0 = const()[name = tensor("input_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_fc1_weight_to_fp16 = const()[name = tensor("layers_3_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56746816)))]; + tensor layers_3_fc1_bias_to_fp16 = const()[name = tensor("layers_3_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57926528)))]; + tensor input_37_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = var_910, groups = var_721, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = var_908, weight = layers_3_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_37_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor var_916 = const()[name = tensor("op_916"), val = tensor([1, 1])]; + tensor var_918 = const()[name = tensor("op_918"), val = tensor([1, 1])]; + tensor hidden_states_9_pad_type_0 = const()[name = tensor("hidden_states_9_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_9_pad_0 = const()[name = tensor("hidden_states_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_fc2_weight_to_fp16 = const()[name = tensor("layers_3_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57929664)))]; + tensor layers_3_fc2_bias_to_fp16 = const()[name = tensor("layers_3_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59109376)))]; + tensor hidden_states_9_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = var_918, groups = var_721, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = var_916, weight = layers_3_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor("hidden_states_9_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor var_929 = const()[name = tensor("op_929"), val = tensor(true)]; + tensor var_933 = const()[name = tensor("op_933"), val = tensor([1])]; + tensor channels_mean_cast_fp16 = reduce_mean(axes = var_933, keep_dims = var_929, x = inputs_cast_fp16)[name = tensor("channels_mean_cast_fp16")]; + tensor zero_mean_cast_fp16 = sub(x = inputs_cast_fp16, y = channels_mean_cast_fp16)[name = tensor("zero_mean_cast_fp16")]; + tensor zero_mean_sq_cast_fp16 = mul(x = zero_mean_cast_fp16, y = zero_mean_cast_fp16)[name = tensor("zero_mean_sq_cast_fp16")]; + tensor var_937 = const()[name = tensor("op_937"), val = tensor([1])]; + tensor var_938_cast_fp16 = reduce_mean(axes = var_937, keep_dims = var_929, x = zero_mean_sq_cast_fp16)[name = tensor("op_938_cast_fp16")]; + tensor var_939_to_fp16 = const()[name = tensor("op_939_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_940_cast_fp16 = add(x = var_938_cast_fp16, y = var_939_to_fp16)[name = tensor("op_940_cast_fp16")]; + tensor denom_epsilon_0_to_fp16 = const()[name = tensor("denom_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_cast_fp16 = rsqrt(epsilon = denom_epsilon_0_to_fp16, x = var_940_cast_fp16)[name = tensor("denom_cast_fp16")]; + tensor out_cast_fp16 = mul(x = zero_mean_cast_fp16, y = denom_cast_fp16)[name = tensor("out_cast_fp16")]; + tensor hidden_states_gamma_0_to_fp16 = const()[name = tensor("hidden_states_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59110208)))]; + tensor hidden_states_beta_0_to_fp16 = const()[name = tensor("hidden_states_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59111040)))]; + tensor hidden_states_epsilon_0_to_fp16 = const()[name = tensor("hidden_states_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor 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("hidden_states_cast_fp16")]; + tensor var_950_axes_0 = const()[name = tensor("op_950_axes_0"), val = tensor([2])]; + tensor var_950_cast_fp16 = squeeze(axes = var_950_axes_0, x = hidden_states_cast_fp16)[name = tensor("op_950_cast_fp16")]; + tensor var_953_perm_0 = const()[name = tensor("op_953_perm_0"), val = tensor([0, 2, 1])]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59111872)))]; + tensor transpose_0 = transpose(perm = var_953_perm_0, x = var_950_cast_fp16)[name = tensor("transpose_0")]; + tensor logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = transpose_0)[name = tensor("linear_0_cast_fp16")]; + tensor var_957 = const()[name = tensor("op_957"), val = tensor(1)]; + tensor obj_59_interleave_0 = const()[name = tensor("obj_59_interleave_0"), val = tensor(false)]; + tensor key_cache_updates = concat(axis = var_957, 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("obj_59_cast_fp16")]; + tensor var_960 = const()[name = tensor("op_960"), val = tensor(1)]; + tensor obj_61_interleave_0 = const()[name = tensor("obj_61_interleave_0"), val = tensor(false)]; + tensor value_cache_updates = concat(axis = var_960, 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("obj_61_cast_fp16")]; + tensor var_971_begin_0 = const()[name = tensor("op_971_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_971_end_0 = const()[name = tensor("op_971_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_971_end_mask_0 = const()[name = tensor("op_971_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_971_cast_fp16 = slice_by_index(begin = var_971_begin_0, end = var_971_end_0, end_mask = var_971_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_971_cast_fp16")]; + tensor var_974_begin_0 = const()[name = tensor("op_974_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_974_end_0 = const()[name = tensor("op_974_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_974_end_mask_0 = const()[name = tensor("op_974_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_974_squeeze_mask_0 = const()[name = tensor("op_974_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_974_cast_fp16 = slice_by_index(begin = var_974_begin_0, end = var_974_end_0, end_mask = var_974_end_mask_0, squeeze_mask = var_974_squeeze_mask_0, x = var_971_cast_fp16)[name = tensor("op_974_cast_fp16")]; + tensor var_989_begin_0 = const()[name = tensor("op_989_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_989_end_0 = const()[name = tensor("op_989_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_989_end_mask_0 = const()[name = tensor("op_989_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_989_cast_fp16 = slice_by_index(begin = var_989_begin_0, end = var_989_end_0, end_mask = var_989_end_mask_0, x = obj_41_cast_fp16)[name = tensor("op_989_cast_fp16")]; + tensor var_992_begin_0 = const()[name = tensor("op_992_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_992_end_0 = const()[name = tensor("op_992_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_992_end_mask_0 = const()[name = tensor("op_992_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_992_squeeze_mask_0 = const()[name = tensor("op_992_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_992_cast_fp16 = slice_by_index(begin = var_992_begin_0, end = var_992_end_0, end_mask = var_992_end_mask_0, squeeze_mask = var_992_squeeze_mask_0, x = var_989_cast_fp16)[name = tensor("op_992_cast_fp16")]; + tensor var_1007_begin_0 = const()[name = tensor("op_1007_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_1007_end_0 = const()[name = tensor("op_1007_end_0"), val = tensor([1, 6, 1, 1500])]; + tensor var_1007_end_mask_0 = const()[name = tensor("op_1007_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1007_cast_fp16 = slice_by_index(begin = var_1007_begin_0, end = var_1007_end_0, end_mask = var_1007_end_mask_0, x = obj_41_cast_fp16)[name = tensor("op_1007_cast_fp16")]; + tensor var_1010_begin_0 = const()[name = tensor("op_1010_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1010_end_0 = const()[name = tensor("op_1010_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1010_end_mask_0 = const()[name = tensor("op_1010_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1010_squeeze_mask_0 = const()[name = tensor("op_1010_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1010_cast_fp16 = slice_by_index(begin = var_1010_begin_0, end = var_1010_end_0, end_mask = var_1010_end_mask_0, squeeze_mask = var_1010_squeeze_mask_0, x = var_1007_cast_fp16)[name = tensor("op_1010_cast_fp16")]; + tensor var_1025_begin_0 = const()[name = tensor("op_1025_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1025_end_0 = const()[name = tensor("op_1025_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1025_end_mask_0 = const()[name = tensor("op_1025_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1025_cast_fp16 = slice_by_index(begin = var_1025_begin_0, end = var_1025_end_0, end_mask = var_1025_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1025_cast_fp16")]; + tensor var_1028_begin_0 = const()[name = tensor("op_1028_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1028_end_0 = const()[name = tensor("op_1028_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1028_end_mask_0 = const()[name = tensor("op_1028_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1028_squeeze_mask_0 = const()[name = tensor("op_1028_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1028_cast_fp16 = slice_by_index(begin = var_1028_begin_0, end = var_1028_end_0, end_mask = var_1028_end_mask_0, squeeze_mask = var_1028_squeeze_mask_0, x = var_1025_cast_fp16)[name = tensor("op_1028_cast_fp16")]; + tensor var_1043_begin_0 = const()[name = tensor("op_1043_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_1043_end_0 = const()[name = tensor("op_1043_end_0"), val = tensor([1, 2, 1, 1500])]; + tensor var_1043_end_mask_0 = const()[name = tensor("op_1043_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1043_cast_fp16 = slice_by_index(begin = var_1043_begin_0, end = var_1043_end_0, end_mask = var_1043_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1043_cast_fp16")]; + tensor var_1046_begin_0 = const()[name = tensor("op_1046_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1046_end_0 = const()[name = tensor("op_1046_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1046_end_mask_0 = const()[name = tensor("op_1046_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1046_squeeze_mask_0 = const()[name = tensor("op_1046_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1046_cast_fp16 = slice_by_index(begin = var_1046_begin_0, end = var_1046_end_0, end_mask = var_1046_end_mask_0, squeeze_mask = var_1046_squeeze_mask_0, x = var_1043_cast_fp16)[name = tensor("op_1046_cast_fp16")]; + tensor var_1061_begin_0 = const()[name = tensor("op_1061_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_1061_end_0 = const()[name = tensor("op_1061_end_0"), val = tensor([1, 3, 1, 1500])]; + tensor var_1061_end_mask_0 = const()[name = tensor("op_1061_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1061_cast_fp16 = slice_by_index(begin = var_1061_begin_0, end = var_1061_end_0, end_mask = var_1061_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1061_cast_fp16")]; + tensor var_1064_begin_0 = const()[name = tensor("op_1064_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1064_end_0 = const()[name = tensor("op_1064_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1064_end_mask_0 = const()[name = tensor("op_1064_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1064_squeeze_mask_0 = const()[name = tensor("op_1064_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1064_cast_fp16 = slice_by_index(begin = var_1064_begin_0, end = var_1064_end_0, end_mask = var_1064_end_mask_0, squeeze_mask = var_1064_squeeze_mask_0, x = var_1061_cast_fp16)[name = tensor("op_1064_cast_fp16")]; + tensor var_1079_begin_0 = const()[name = tensor("op_1079_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_1079_end_0 = const()[name = tensor("op_1079_end_0"), val = tensor([1, 4, 1, 1500])]; + tensor var_1079_end_mask_0 = const()[name = tensor("op_1079_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1079_cast_fp16 = slice_by_index(begin = var_1079_begin_0, end = var_1079_end_0, end_mask = var_1079_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1079_cast_fp16")]; + tensor var_1082_begin_0 = const()[name = tensor("op_1082_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1082_end_0 = const()[name = tensor("op_1082_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1082_end_mask_0 = const()[name = tensor("op_1082_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1082_squeeze_mask_0 = const()[name = tensor("op_1082_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1082_cast_fp16 = slice_by_index(begin = var_1082_begin_0, end = var_1082_end_0, end_mask = var_1082_end_mask_0, squeeze_mask = var_1082_squeeze_mask_0, x = var_1079_cast_fp16)[name = tensor("op_1082_cast_fp16")]; + tensor var_1097_begin_0 = const()[name = tensor("op_1097_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_1097_end_0 = const()[name = tensor("op_1097_end_0"), val = tensor([1, 5, 1, 1500])]; + tensor var_1097_end_mask_0 = const()[name = tensor("op_1097_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1097_cast_fp16 = slice_by_index(begin = var_1097_begin_0, end = var_1097_end_0, end_mask = var_1097_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1097_cast_fp16")]; + tensor var_1100_begin_0 = const()[name = tensor("op_1100_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1100_end_0 = const()[name = tensor("op_1100_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1100_end_mask_0 = const()[name = tensor("op_1100_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1100_squeeze_mask_0 = const()[name = tensor("op_1100_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1100_cast_fp16 = slice_by_index(begin = var_1100_begin_0, end = var_1100_end_0, end_mask = var_1100_end_mask_0, squeeze_mask = var_1100_squeeze_mask_0, x = var_1097_cast_fp16)[name = tensor("op_1100_cast_fp16")]; + tensor var_1107 = const()[name = tensor("op_1107"), val = tensor(1)]; + tensor var_1108_interleave_0 = const()[name = tensor("op_1108_interleave_0"), val = tensor(false)]; + tensor var_1108_cast_fp16 = concat(axis = var_1107, interleave = var_1108_interleave_0, values = (var_974_cast_fp16, var_992_cast_fp16, var_1010_cast_fp16, var_1028_cast_fp16, var_1046_cast_fp16, var_1064_cast_fp16, var_1082_cast_fp16, var_1100_cast_fp16))[name = tensor("op_1108_cast_fp16")]; + tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([1])]; + tensor var_1111 = const()[name = tensor("op_1111"), val = tensor(false)]; + tensor alignment_heads_weights = reduce_mean(axes = var_1110, keep_dims = var_1111, x = var_1108_cast_fp16)[name = tensor("obj_cast_fp16")]; + } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights); +} \ No newline at end of file