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_20_axis_0 = const()[name = tensor("op_20_axis_0"), val = tensor(0)]; tensor var_20_batch_dims_0 = const()[name = tensor("op_20_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_20_cast_fp16 = gather(axis = var_20_axis_0, batch_dims = var_20_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor("op_20_cast_fp16")]; 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_positions_weight_to_fp16 = const()[name = tensor("embed_positions_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132777088)))]; tensor var_24_cast_fp16 = gather(axis = var_24_axis_0, batch_dims = var_24_batch_dims_0, indices = cache_length, x = embed_positions_weight_to_fp16)[name = tensor("op_24_cast_fp16")]; tensor hidden_states_1_cast_fp16 = add(x = var_20_cast_fp16, y = var_24_cast_fp16)[name = tensor("hidden_states_1_cast_fp16")]; tensor var_38_axes_0 = const()[name = tensor("op_38_axes_0"), val = tensor([2])]; tensor var_38_cast_fp16 = expand_dims(axes = var_38_axes_0, x = hidden_states_1_cast_fp16)[name = tensor("op_38_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_38_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([1280, 1280])]; tensor var_43_axis_0 = const()[name = tensor("op_43_axis_0"), val = tensor(1)]; tensor var_43_cast_fp16_0, tensor var_43_cast_fp16_1 = split(axis = var_43_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor("op_43_cast_fp16")]; tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([1280, 1280])]; tensor var_48_axis_0 = const()[name = tensor("op_48_axis_0"), val = tensor(1)]; tensor var_48_cast_fp16_0, tensor var_48_cast_fp16_1 = split(axis = var_48_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor("op_48_cast_fp16")]; tensor var_56 = const()[name = tensor("op_56"), val = tensor(3)]; tensor var_63 = const()[name = tensor("op_63"), val = tensor(1)]; tensor var_64 = const()[name = tensor("op_64"), val = tensor(true)]; tensor var_76 = const()[name = tensor("op_76"), val = tensor([1])]; tensor channels_mean_1_cast_fp16 = reduce_mean(axes = var_76, keep_dims = var_64, 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_80 = const()[name = tensor("op_80"), val = tensor([1])]; tensor var_81_cast_fp16 = reduce_mean(axes = var_80, keep_dims = var_64, x = zero_mean_sq_1_cast_fp16)[name = tensor("op_81_cast_fp16")]; tensor var_82_to_fp16 = const()[name = tensor("op_82_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_83_cast_fp16 = add(x = var_81_cast_fp16, y = var_82_to_fp16)[name = tensor("op_83_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_83_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(133924032)))]; 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(133926656)))]; 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(133929280)))]; 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(133931904)))]; 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_98 = const()[name = tensor("op_98"), val = tensor([1, 1])]; tensor var_100 = const()[name = tensor("op_100"), 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(133934528)))]; 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(137211392)))]; tensor query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = var_100, groups = var_63, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = var_98, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("query_1_cast_fp16")]; tensor var_104 = const()[name = tensor("op_104"), val = tensor([1, 1])]; tensor var_106 = const()[name = tensor("op_106"), 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(137214016)))]; tensor current_key_1_cast_fp16 = conv(dilations = var_106, groups = var_63, pad = current_key_1_pad_0, pad_type = current_key_1_pad_type_0, strides = var_104, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("current_key_1_cast_fp16")]; tensor var_111 = const()[name = tensor("op_111"), val = tensor([1, 1])]; tensor var_113 = const()[name = tensor("op_113"), 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(140490880)))]; 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(143767744)))]; tensor current_value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = var_113, groups = var_63, pad = current_value_1_pad_0, pad_type = current_value_1_pad_type_0, strides = var_111, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("current_value_1_cast_fp16")]; tensor var_117_axes_0 = const()[name = tensor("op_117_axes_0"), val = tensor([1])]; tensor var_117_cast_fp16 = expand_dims(axes = var_117_axes_0, x = kv_cache_update_mask)[name = tensor("op_117_cast_fp16")]; tensor var_118_axes_0 = const()[name = tensor("op_118_axes_0"), val = tensor([2])]; tensor var_118_cast_fp16 = expand_dims(axes = var_118_axes_0, x = var_117_cast_fp16)[name = tensor("op_118_cast_fp16")]; tensor var_120_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_118_cast_fp16)[name = tensor("op_120_cast_fp16")]; tensor var_57_to_fp16 = const()[name = tensor("op_57_to_fp16"), val = tensor(0x1p+0)]; tensor var_121_cast_fp16 = sub(x = var_57_to_fp16, y = var_118_cast_fp16)[name = tensor("op_121_cast_fp16")]; tensor var_122_cast_fp16 = mul(x = var_43_cast_fp16_0, y = var_121_cast_fp16)[name = tensor("op_122_cast_fp16")]; tensor key_1_cast_fp16 = add(x = var_120_cast_fp16, y = var_122_cast_fp16)[name = tensor("key_1_cast_fp16")]; tensor var_124_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_118_cast_fp16)[name = tensor("op_124_cast_fp16")]; tensor var_126_cast_fp16 = mul(x = var_48_cast_fp16_0, y = var_121_cast_fp16)[name = tensor("op_126_cast_fp16")]; tensor value_1_cast_fp16 = add(x = var_124_cast_fp16, y = var_126_cast_fp16)[name = tensor("value_1_cast_fp16")]; tensor var_129 = const()[name = tensor("op_129"), val = tensor([1, 20, 64, -1])]; tensor var_130_cast_fp16 = reshape(shape = var_129, x = query_1_cast_fp16)[name = tensor("op_130_cast_fp16")]; tensor var_131_to_fp16 = const()[name = tensor("op_131_to_fp16"), val = tensor(0x1p-3)]; tensor var_132_cast_fp16 = mul(x = var_130_cast_fp16, y = var_131_to_fp16)[name = tensor("op_132_cast_fp16")]; tensor var_133 = const()[name = tensor("op_133"), val = tensor([1, 20, 64, -1])]; tensor var_134_cast_fp16 = reshape(shape = var_133, x = key_1_cast_fp16)[name = tensor("op_134_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_132_cast_fp16, y = var_134_cast_fp16)[name = tensor("mh_w_1_cast_fp16")]; tensor var_138_axes_0 = const()[name = tensor("op_138_axes_0"), val = tensor([1])]; tensor var_138_cast_fp16 = expand_dims(axes = var_138_axes_0, x = decoder_key_padding_mask)[name = tensor("op_138_cast_fp16")]; tensor var_139_axes_0 = const()[name = tensor("op_139_axes_0"), val = tensor([2])]; tensor var_139_cast_fp16 = expand_dims(axes = var_139_axes_0, x = var_138_cast_fp16)[name = tensor("op_139_cast_fp16")]; tensor mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_139_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; tensor var_142_cast_fp16 = softmax(axis = var_56, x = mh_w_3_cast_fp16)[name = tensor("op_142_cast_fp16")]; tensor var_143 = const()[name = tensor("op_143"), val = tensor([1, 20, 64, -1])]; tensor var_144_cast_fp16 = reshape(shape = var_143, x = value_1_cast_fp16)[name = tensor("op_144_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_144_cast_fp16, y = var_142_cast_fp16)[name = tensor("attn_1_cast_fp16")]; tensor var_147 = const()[name = tensor("op_147"), val = tensor([1, 1280, 1, -1])]; tensor input_1_cast_fp16 = reshape(shape = var_147, x = attn_1_cast_fp16)[name = tensor("input_1_cast_fp16")]; tensor var_151 = const()[name = tensor("op_151"), val = tensor([1, 1])]; tensor var_153 = const()[name = tensor("op_153"), 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(143770368)))]; 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(147047232)))]; tensor obj_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = var_153, groups = var_63, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = var_151, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("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_163 = const()[name = tensor("op_163"), val = tensor([1])]; tensor channels_mean_3_cast_fp16 = reduce_mean(axes = var_163, keep_dims = var_64, 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_167 = const()[name = tensor("op_167"), val = tensor([1])]; tensor var_168_cast_fp16 = reduce_mean(axes = var_167, keep_dims = var_64, x = zero_mean_sq_3_cast_fp16)[name = tensor("op_168_cast_fp16")]; tensor var_169_to_fp16 = const()[name = tensor("op_169_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_170_cast_fp16 = add(x = var_168_cast_fp16, y = var_169_to_fp16)[name = tensor("op_170_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_170_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(147049856)))]; 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(147052480)))]; 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_185 = const()[name = tensor("op_185"), val = tensor([1, 1])]; tensor var_187 = const()[name = tensor("op_187"), 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(147055104)))]; 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(150331968)))]; tensor query_3_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_bias_to_fp16, dilations = var_187, groups = var_63, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = var_185, weight = layers_0_encoder_attn_q_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("query_3_cast_fp16")]; tensor var_191 = const()[name = tensor("op_191"), val = tensor([1, 1])]; tensor var_193 = const()[name = tensor("op_193"), 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(150334592)))]; tensor key_3_cast_fp16 = conv(dilations = var_193, groups = var_63, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = var_191, weight = layers_0_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_3_cast_fp16")]; tensor var_198 = const()[name = tensor("op_198"), val = tensor([1, 1])]; tensor var_200 = const()[name = tensor("op_200"), 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(153611456)))]; 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(156888320)))]; tensor value_3_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_bias_to_fp16, dilations = var_200, groups = var_63, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = var_198, weight = layers_0_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_3_cast_fp16")]; tensor var_204 = const()[name = tensor("op_204"), val = tensor([1, 20, 64, -1])]; tensor var_205_cast_fp16 = reshape(shape = var_204, x = query_3_cast_fp16)[name = tensor("op_205_cast_fp16")]; tensor var_206_to_fp16 = const()[name = tensor("op_206_to_fp16"), val = tensor(0x1p-3)]; tensor var_207_cast_fp16 = mul(x = var_205_cast_fp16, y = var_206_to_fp16)[name = tensor("op_207_cast_fp16")]; tensor var_208 = const()[name = tensor("op_208"), val = tensor([1, 20, 64, -1])]; tensor var_209_cast_fp16 = reshape(shape = var_208, x = key_3_cast_fp16)[name = tensor("op_209_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_207_cast_fp16, y = var_209_cast_fp16)[name = tensor("mh_w_5_cast_fp16")]; tensor obj_13_cast_fp16 = softmax(axis = var_56, x = mh_w_5_cast_fp16)[name = tensor("obj_13_cast_fp16")]; tensor var_213 = const()[name = tensor("op_213"), val = tensor([1, 20, 64, -1])]; tensor var_214_cast_fp16 = reshape(shape = var_213, x = value_3_cast_fp16)[name = tensor("op_214_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_214_cast_fp16, y = obj_13_cast_fp16)[name = tensor("attn_3_cast_fp16")]; tensor var_217 = const()[name = tensor("op_217"), val = tensor([1, 1280, 1, -1])]; tensor input_3_cast_fp16 = reshape(shape = var_217, x = attn_3_cast_fp16)[name = tensor("input_3_cast_fp16")]; tensor var_221 = const()[name = tensor("op_221"), val = tensor([1, 1])]; tensor var_223 = const()[name = tensor("op_223"), 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(156890944)))]; 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(160167808)))]; tensor obj_11_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = var_223, groups = var_63, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = var_221, weight = layers_0_encoder_attn_o_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("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_229 = const()[name = tensor("op_229"), val = tensor([1])]; tensor channels_mean_5_cast_fp16 = reduce_mean(axes = var_229, keep_dims = var_64, 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_233 = const()[name = tensor("op_233"), val = tensor([1])]; tensor var_234_cast_fp16 = reduce_mean(axes = var_233, keep_dims = var_64, x = zero_mean_sq_5_cast_fp16)[name = tensor("op_234_cast_fp16")]; tensor var_235_to_fp16 = const()[name = tensor("op_235_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_236_cast_fp16 = add(x = var_234_cast_fp16, y = var_235_to_fp16)[name = tensor("op_236_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_236_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(160170432)))]; 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(160173056)))]; 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_247 = const()[name = tensor("op_247"), val = tensor([1, 1])]; tensor var_249 = const()[name = tensor("op_249"), 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(160175680)))]; 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(173282944)))]; tensor input_7_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = var_249, groups = var_63, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = var_247, weight = layers_0_fc1_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("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_255 = const()[name = tensor("op_255"), val = tensor([1, 1])]; tensor var_257 = const()[name = tensor("op_257"), 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(173293248)))]; 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(186400512)))]; tensor hidden_states_3_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = var_257, groups = var_63, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = var_255, weight = layers_0_fc2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("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_270 = const()[name = tensor("op_270"), val = tensor(3)]; tensor var_277 = const()[name = tensor("op_277"), val = tensor(1)]; tensor var_278 = const()[name = tensor("op_278"), val = tensor(true)]; tensor var_290 = const()[name = tensor("op_290"), val = tensor([1])]; tensor channels_mean_7_cast_fp16 = reduce_mean(axes = var_290, keep_dims = var_278, 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_294 = const()[name = tensor("op_294"), val = tensor([1])]; tensor var_295_cast_fp16 = reduce_mean(axes = var_294, keep_dims = var_278, x = zero_mean_sq_7_cast_fp16)[name = tensor("op_295_cast_fp16")]; tensor var_296_to_fp16 = const()[name = tensor("op_296_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_297_cast_fp16 = add(x = var_295_cast_fp16, y = var_296_to_fp16)[name = tensor("op_297_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_297_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(186403136)))]; 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(186405760)))]; 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_312 = const()[name = tensor("op_312"), val = tensor([1, 1])]; tensor var_314 = const()[name = tensor("op_314"), 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(186408384)))]; 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(189685248)))]; tensor query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = var_314, groups = var_277, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = var_312, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("query_5_cast_fp16")]; tensor var_318 = const()[name = tensor("op_318"), val = tensor([1, 1])]; tensor var_320 = const()[name = tensor("op_320"), 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_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(189687872)))]; tensor current_key_cast_fp16 = conv(dilations = var_320, groups = var_277, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = var_318, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("current_key_cast_fp16")]; tensor var_325 = const()[name = tensor("op_325"), val = tensor([1, 1])]; tensor var_327 = const()[name = tensor("op_327"), 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_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(192964736)))]; 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(196241600)))]; tensor current_value_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = var_327, groups = var_277, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = var_325, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("current_value_cast_fp16")]; tensor var_334_cast_fp16 = mul(x = current_key_cast_fp16, y = var_118_cast_fp16)[name = tensor("op_334_cast_fp16")]; tensor var_336_cast_fp16 = mul(x = var_43_cast_fp16_1, y = var_121_cast_fp16)[name = tensor("op_336_cast_fp16")]; tensor key_5_cast_fp16 = add(x = var_334_cast_fp16, y = var_336_cast_fp16)[name = tensor("key_5_cast_fp16")]; tensor var_338_cast_fp16 = mul(x = current_value_cast_fp16, y = var_118_cast_fp16)[name = tensor("op_338_cast_fp16")]; tensor var_340_cast_fp16 = mul(x = var_48_cast_fp16_1, y = var_121_cast_fp16)[name = tensor("op_340_cast_fp16")]; tensor value_5_cast_fp16 = add(x = var_338_cast_fp16, y = var_340_cast_fp16)[name = tensor("value_5_cast_fp16")]; tensor var_343 = const()[name = tensor("op_343"), val = tensor([1, 20, 64, -1])]; tensor var_344_cast_fp16 = reshape(shape = var_343, x = query_5_cast_fp16)[name = tensor("op_344_cast_fp16")]; tensor var_345_to_fp16 = const()[name = tensor("op_345_to_fp16"), val = tensor(0x1p-3)]; tensor var_346_cast_fp16 = mul(x = var_344_cast_fp16, y = var_345_to_fp16)[name = tensor("op_346_cast_fp16")]; tensor var_347 = const()[name = tensor("op_347"), val = tensor([1, 20, 64, -1])]; tensor var_348_cast_fp16 = reshape(shape = var_347, x = key_5_cast_fp16)[name = tensor("op_348_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_346_cast_fp16, y = var_348_cast_fp16)[name = tensor("mh_w_7_cast_fp16")]; tensor mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_139_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; tensor var_356_cast_fp16 = softmax(axis = var_270, x = mh_w_9_cast_fp16)[name = tensor("op_356_cast_fp16")]; tensor var_357 = const()[name = tensor("op_357"), val = tensor([1, 20, 64, -1])]; tensor var_358_cast_fp16 = reshape(shape = var_357, x = value_5_cast_fp16)[name = tensor("op_358_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_358_cast_fp16, y = var_356_cast_fp16)[name = tensor("attn_5_cast_fp16")]; tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 1280, 1, -1])]; tensor input_11_cast_fp16 = reshape(shape = var_361, x = attn_5_cast_fp16)[name = tensor("input_11_cast_fp16")]; tensor var_365 = const()[name = tensor("op_365"), val = tensor([1, 1])]; tensor var_367 = const()[name = tensor("op_367"), 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(196244224)))]; 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(199521088)))]; tensor obj_21_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = var_367, groups = var_277, pad = obj_21_pad_0, pad_type = obj_21_pad_type_0, strides = var_365, 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_377 = const()[name = tensor("op_377"), val = tensor([1])]; tensor channels_mean_9_cast_fp16 = reduce_mean(axes = var_377, keep_dims = var_278, 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_381 = const()[name = tensor("op_381"), val = tensor([1])]; tensor var_382_cast_fp16 = reduce_mean(axes = var_381, keep_dims = var_278, x = zero_mean_sq_9_cast_fp16)[name = tensor("op_382_cast_fp16")]; tensor var_383_to_fp16 = const()[name = tensor("op_383_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_384_cast_fp16 = add(x = var_382_cast_fp16, y = var_383_to_fp16)[name = tensor("op_384_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_384_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(199523712)))]; 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(199526336)))]; 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_399 = const()[name = tensor("op_399"), val = tensor([1, 1])]; tensor var_401 = const()[name = tensor("op_401"), 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_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(199528960)))]; 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(202805824)))]; tensor query_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = var_401, groups = var_277, pad = query_pad_0, pad_type = query_pad_type_0, strides = var_399, weight = layers_1_encoder_attn_q_proj_weight_to_fp16, x = obj_23_cast_fp16)[name = tensor("query_cast_fp16")]; tensor var_405 = const()[name = tensor("op_405"), val = tensor([1, 1])]; tensor var_407 = const()[name = tensor("op_407"), 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_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(202808448)))]; tensor key_cast_fp16 = conv(dilations = var_407, groups = var_277, pad = key_pad_0, pad_type = key_pad_type_0, strides = var_405, weight = layers_1_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_cast_fp16")]; tensor var_412 = const()[name = tensor("op_412"), val = tensor([1, 1])]; tensor var_414 = const()[name = tensor("op_414"), 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_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(206085312)))]; 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(209362176)))]; tensor value_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_bias_to_fp16, dilations = var_414, groups = var_277, pad = value_pad_0, pad_type = value_pad_type_0, strides = var_412, weight = layers_1_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_cast_fp16")]; tensor var_418 = const()[name = tensor("op_418"), val = tensor([1, 20, 64, -1])]; tensor var_419_cast_fp16 = reshape(shape = var_418, x = query_cast_fp16)[name = tensor("op_419_cast_fp16")]; tensor var_420_to_fp16 = const()[name = tensor("op_420_to_fp16"), val = tensor(0x1p-3)]; tensor var_421_cast_fp16 = mul(x = var_419_cast_fp16, y = var_420_to_fp16)[name = tensor("op_421_cast_fp16")]; tensor var_422 = const()[name = tensor("op_422"), val = tensor([1, 20, 64, -1])]; tensor var_423_cast_fp16 = reshape(shape = var_422, x = key_cast_fp16)[name = tensor("op_423_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_421_cast_fp16, y = var_423_cast_fp16)[name = tensor("mh_w_cast_fp16")]; tensor obj_27_cast_fp16 = softmax(axis = var_270, x = mh_w_cast_fp16)[name = tensor("obj_27_cast_fp16")]; tensor var_427 = const()[name = tensor("op_427"), val = tensor([1, 20, 64, -1])]; tensor var_428_cast_fp16 = reshape(shape = var_427, x = value_cast_fp16)[name = tensor("op_428_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_428_cast_fp16, y = obj_27_cast_fp16)[name = tensor("attn_cast_fp16")]; tensor var_431 = const()[name = tensor("op_431"), val = tensor([1, 1280, 1, -1])]; tensor input_13_cast_fp16 = reshape(shape = var_431, x = attn_cast_fp16)[name = tensor("input_13_cast_fp16")]; tensor var_435 = const()[name = tensor("op_435"), val = tensor([1, 1])]; tensor var_437 = const()[name = tensor("op_437"), 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(209364800)))]; 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(212641664)))]; tensor obj_25_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = var_437, groups = var_277, pad = obj_25_pad_0, pad_type = obj_25_pad_type_0, strides = var_435, 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_446 = const()[name = tensor("op_446"), val = tensor([1])]; tensor channels_mean_11_cast_fp16 = reduce_mean(axes = var_446, keep_dims = var_278, 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_450 = const()[name = tensor("op_450"), val = tensor([1])]; tensor var_451_cast_fp16 = reduce_mean(axes = var_450, keep_dims = var_278, x = zero_mean_sq_11_cast_fp16)[name = tensor("op_451_cast_fp16")]; tensor var_452_to_fp16 = const()[name = tensor("op_452_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_453_cast_fp16 = add(x = var_451_cast_fp16, y = var_452_to_fp16)[name = tensor("op_453_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_453_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(212644288)))]; 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(212646912)))]; 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_464 = const()[name = tensor("op_464"), val = tensor([1, 1])]; tensor var_466 = const()[name = tensor("op_466"), 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(212649536)))]; 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(225756800)))]; tensor input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = var_466, groups = var_277, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = var_464, weight = layers_1_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_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_17_cast_fp16)[name = tensor("input_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 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(225767104)))]; 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(238874368)))]; tensor hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = var_474, groups = var_277, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = var_472, weight = layers_1_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor("hidden_states_5_cast_fp16")]; tensor inputs_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor("inputs_cast_fp16")]; tensor var_485 = const()[name = tensor("op_485"), val = tensor(true)]; tensor var_489 = const()[name = tensor("op_489"), val = tensor([1])]; tensor channels_mean_cast_fp16 = reduce_mean(axes = var_489, keep_dims = var_485, 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_493 = const()[name = tensor("op_493"), val = tensor([1])]; tensor var_494_cast_fp16 = reduce_mean(axes = var_493, keep_dims = var_485, x = zero_mean_sq_cast_fp16)[name = tensor("op_494_cast_fp16")]; tensor var_495_to_fp16 = const()[name = tensor("op_495_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_496_cast_fp16 = add(x = var_494_cast_fp16, y = var_495_to_fp16)[name = tensor("op_496_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_496_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(238876992)))]; 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(238879616)))]; 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_506_axes_0 = const()[name = tensor("op_506_axes_0"), val = tensor([2])]; tensor var_506_cast_fp16 = squeeze(axes = var_506_axes_0, x = hidden_states_cast_fp16)[name = tensor("op_506_cast_fp16")]; tensor var_509_perm_0 = const()[name = tensor("op_509_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(238882240)))]; tensor transpose_0 = transpose(perm = var_509_perm_0, x = var_506_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_513 = const()[name = tensor("op_513"), val = tensor(1)]; tensor obj_31_interleave_0 = const()[name = tensor("obj_31_interleave_0"), val = tensor(false)]; tensor key_cache_updates = concat(axis = var_513, interleave = obj_31_interleave_0, values = (current_key_1_cast_fp16, current_key_cast_fp16))[name = tensor("obj_31_cast_fp16")]; tensor var_516 = const()[name = tensor("op_516"), val = tensor(1)]; tensor obj_33_interleave_0 = const()[name = tensor("obj_33_interleave_0"), val = tensor(false)]; tensor value_cache_updates = concat(axis = var_516, interleave = obj_33_interleave_0, values = (current_value_1_cast_fp16, current_value_cast_fp16))[name = tensor("obj_33_cast_fp16")]; tensor var_527_begin_0 = const()[name = tensor("op_527_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_527_end_0 = const()[name = tensor("op_527_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_527_end_mask_0 = const()[name = tensor("op_527_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_527_cast_fp16 = slice_by_index(begin = var_527_begin_0, end = var_527_end_0, end_mask = var_527_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_527_cast_fp16")]; tensor var_530_begin_0 = const()[name = tensor("op_530_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_530_end_0 = const()[name = tensor("op_530_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_530_end_mask_0 = const()[name = tensor("op_530_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_530_squeeze_mask_0 = const()[name = tensor("op_530_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_530_cast_fp16 = slice_by_index(begin = var_530_begin_0, end = var_530_end_0, end_mask = var_530_end_mask_0, squeeze_mask = var_530_squeeze_mask_0, x = var_527_cast_fp16)[name = tensor("op_530_cast_fp16")]; tensor var_545_begin_0 = const()[name = tensor("op_545_begin_0"), val = tensor([0, 1, 0, 0])]; tensor var_545_end_0 = const()[name = tensor("op_545_end_0"), val = tensor([1, 2, 1, 1500])]; tensor var_545_end_mask_0 = const()[name = tensor("op_545_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_545_cast_fp16 = slice_by_index(begin = var_545_begin_0, end = var_545_end_0, end_mask = var_545_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_545_cast_fp16")]; tensor var_548_begin_0 = const()[name = tensor("op_548_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_548_end_0 = const()[name = tensor("op_548_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_548_end_mask_0 = const()[name = tensor("op_548_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_548_squeeze_mask_0 = const()[name = tensor("op_548_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_548_cast_fp16 = slice_by_index(begin = var_548_begin_0, end = var_548_end_0, end_mask = var_548_end_mask_0, squeeze_mask = var_548_squeeze_mask_0, x = var_545_cast_fp16)[name = tensor("op_548_cast_fp16")]; tensor var_563_begin_0 = const()[name = tensor("op_563_begin_0"), val = tensor([0, 2, 0, 0])]; tensor var_563_end_0 = const()[name = tensor("op_563_end_0"), val = tensor([1, 3, 1, 1500])]; tensor var_563_end_mask_0 = const()[name = tensor("op_563_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_563_cast_fp16 = slice_by_index(begin = var_563_begin_0, end = var_563_end_0, end_mask = var_563_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_563_cast_fp16")]; tensor var_566_begin_0 = const()[name = tensor("op_566_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_566_end_0 = const()[name = tensor("op_566_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_566_end_mask_0 = const()[name = tensor("op_566_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_566_squeeze_mask_0 = const()[name = tensor("op_566_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_566_cast_fp16 = slice_by_index(begin = var_566_begin_0, end = var_566_end_0, end_mask = var_566_end_mask_0, squeeze_mask = var_566_squeeze_mask_0, x = var_563_cast_fp16)[name = tensor("op_566_cast_fp16")]; tensor var_581_begin_0 = const()[name = tensor("op_581_begin_0"), val = tensor([0, 3, 0, 0])]; tensor var_581_end_0 = const()[name = tensor("op_581_end_0"), val = tensor([1, 4, 1, 1500])]; tensor var_581_end_mask_0 = const()[name = tensor("op_581_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_581_cast_fp16 = slice_by_index(begin = var_581_begin_0, end = var_581_end_0, end_mask = var_581_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_581_cast_fp16")]; tensor var_584_begin_0 = const()[name = tensor("op_584_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_584_end_0 = const()[name = tensor("op_584_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_584_end_mask_0 = const()[name = tensor("op_584_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_584_squeeze_mask_0 = const()[name = tensor("op_584_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_584_cast_fp16 = slice_by_index(begin = var_584_begin_0, end = var_584_end_0, end_mask = var_584_end_mask_0, squeeze_mask = var_584_squeeze_mask_0, x = var_581_cast_fp16)[name = tensor("op_584_cast_fp16")]; tensor var_599_begin_0 = const()[name = tensor("op_599_begin_0"), val = tensor([0, 4, 0, 0])]; tensor var_599_end_0 = const()[name = tensor("op_599_end_0"), val = tensor([1, 5, 1, 1500])]; tensor var_599_end_mask_0 = const()[name = tensor("op_599_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_599_cast_fp16 = slice_by_index(begin = var_599_begin_0, end = var_599_end_0, end_mask = var_599_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_599_cast_fp16")]; tensor var_602_begin_0 = const()[name = tensor("op_602_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_602_end_0 = const()[name = tensor("op_602_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_602_end_mask_0 = const()[name = tensor("op_602_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_602_squeeze_mask_0 = const()[name = tensor("op_602_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_602_cast_fp16 = slice_by_index(begin = var_602_begin_0, end = var_602_end_0, end_mask = var_602_end_mask_0, squeeze_mask = var_602_squeeze_mask_0, x = var_599_cast_fp16)[name = tensor("op_602_cast_fp16")]; tensor var_617_begin_0 = const()[name = tensor("op_617_begin_0"), val = tensor([0, 5, 0, 0])]; tensor var_617_end_0 = const()[name = tensor("op_617_end_0"), val = tensor([1, 6, 1, 1500])]; tensor var_617_end_mask_0 = const()[name = tensor("op_617_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_617_cast_fp16 = slice_by_index(begin = var_617_begin_0, end = var_617_end_0, end_mask = var_617_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_617_cast_fp16")]; tensor var_620_begin_0 = const()[name = tensor("op_620_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_620_end_0 = const()[name = tensor("op_620_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_620_end_mask_0 = const()[name = tensor("op_620_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_620_squeeze_mask_0 = const()[name = tensor("op_620_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_620_cast_fp16 = slice_by_index(begin = var_620_begin_0, end = var_620_end_0, end_mask = var_620_end_mask_0, squeeze_mask = var_620_squeeze_mask_0, x = var_617_cast_fp16)[name = tensor("op_620_cast_fp16")]; tensor var_635_begin_0 = const()[name = tensor("op_635_begin_0"), val = tensor([0, 6, 0, 0])]; tensor var_635_end_0 = const()[name = tensor("op_635_end_0"), val = tensor([1, 7, 1, 1500])]; tensor var_635_end_mask_0 = const()[name = tensor("op_635_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_635_cast_fp16 = slice_by_index(begin = var_635_begin_0, end = var_635_end_0, end_mask = var_635_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_635_cast_fp16")]; tensor var_638_begin_0 = const()[name = tensor("op_638_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_638_end_0 = const()[name = tensor("op_638_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_638_end_mask_0 = const()[name = tensor("op_638_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_638_squeeze_mask_0 = const()[name = tensor("op_638_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_638_cast_fp16 = slice_by_index(begin = var_638_begin_0, end = var_638_end_0, end_mask = var_638_end_mask_0, squeeze_mask = var_638_squeeze_mask_0, x = var_635_cast_fp16)[name = tensor("op_638_cast_fp16")]; tensor var_653_begin_0 = const()[name = tensor("op_653_begin_0"), val = tensor([0, 7, 0, 0])]; tensor var_653_end_0 = const()[name = tensor("op_653_end_0"), val = tensor([1, 8, 1, 1500])]; tensor var_653_end_mask_0 = const()[name = tensor("op_653_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_653_cast_fp16 = slice_by_index(begin = var_653_begin_0, end = var_653_end_0, end_mask = var_653_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_653_cast_fp16")]; tensor var_656_begin_0 = const()[name = tensor("op_656_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_656_end_0 = const()[name = tensor("op_656_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_656_end_mask_0 = const()[name = tensor("op_656_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_656_squeeze_mask_0 = const()[name = tensor("op_656_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_656_cast_fp16 = slice_by_index(begin = var_656_begin_0, end = var_656_end_0, end_mask = var_656_end_mask_0, squeeze_mask = var_656_squeeze_mask_0, x = var_653_cast_fp16)[name = tensor("op_656_cast_fp16")]; tensor var_671_begin_0 = const()[name = tensor("op_671_begin_0"), val = tensor([0, 8, 0, 0])]; tensor var_671_end_0 = const()[name = tensor("op_671_end_0"), val = tensor([1, 9, 1, 1500])]; tensor var_671_end_mask_0 = const()[name = tensor("op_671_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_671_cast_fp16 = slice_by_index(begin = var_671_begin_0, end = var_671_end_0, end_mask = var_671_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_671_cast_fp16")]; tensor var_674_begin_0 = const()[name = tensor("op_674_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_674_end_0 = const()[name = tensor("op_674_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_674_end_mask_0 = const()[name = tensor("op_674_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_674_squeeze_mask_0 = const()[name = tensor("op_674_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_674_cast_fp16 = slice_by_index(begin = var_674_begin_0, end = var_674_end_0, end_mask = var_674_end_mask_0, squeeze_mask = var_674_squeeze_mask_0, x = var_671_cast_fp16)[name = tensor("op_674_cast_fp16")]; tensor var_689_begin_0 = const()[name = tensor("op_689_begin_0"), val = tensor([0, 9, 0, 0])]; tensor var_689_end_0 = const()[name = tensor("op_689_end_0"), val = tensor([1, 10, 1, 1500])]; tensor var_689_end_mask_0 = const()[name = tensor("op_689_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_689_cast_fp16 = slice_by_index(begin = var_689_begin_0, end = var_689_end_0, end_mask = var_689_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_689_cast_fp16")]; tensor var_692_begin_0 = const()[name = tensor("op_692_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_692_end_0 = const()[name = tensor("op_692_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_692_end_mask_0 = const()[name = tensor("op_692_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_692_squeeze_mask_0 = const()[name = tensor("op_692_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_692_cast_fp16 = slice_by_index(begin = var_692_begin_0, end = var_692_end_0, end_mask = var_692_end_mask_0, squeeze_mask = var_692_squeeze_mask_0, x = var_689_cast_fp16)[name = tensor("op_692_cast_fp16")]; tensor var_707_begin_0 = const()[name = tensor("op_707_begin_0"), val = tensor([0, 10, 0, 0])]; tensor var_707_end_0 = const()[name = tensor("op_707_end_0"), val = tensor([1, 11, 1, 1500])]; tensor var_707_end_mask_0 = const()[name = tensor("op_707_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_707_cast_fp16 = slice_by_index(begin = var_707_begin_0, end = var_707_end_0, end_mask = var_707_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_707_cast_fp16")]; tensor var_710_begin_0 = const()[name = tensor("op_710_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_710_end_0 = const()[name = tensor("op_710_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_710_end_mask_0 = const()[name = tensor("op_710_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_710_squeeze_mask_0 = const()[name = tensor("op_710_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_710_cast_fp16 = slice_by_index(begin = var_710_begin_0, end = var_710_end_0, end_mask = var_710_end_mask_0, squeeze_mask = var_710_squeeze_mask_0, x = var_707_cast_fp16)[name = tensor("op_710_cast_fp16")]; tensor var_725_begin_0 = const()[name = tensor("op_725_begin_0"), val = tensor([0, 11, 0, 0])]; tensor var_725_end_0 = const()[name = tensor("op_725_end_0"), val = tensor([1, 12, 1, 1500])]; tensor var_725_end_mask_0 = const()[name = tensor("op_725_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_725_cast_fp16 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_725_cast_fp16")]; tensor var_728_begin_0 = const()[name = tensor("op_728_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_728_end_0 = const()[name = tensor("op_728_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_728_end_mask_0 = const()[name = tensor("op_728_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_728_squeeze_mask_0 = const()[name = tensor("op_728_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_728_cast_fp16 = slice_by_index(begin = var_728_begin_0, end = var_728_end_0, end_mask = var_728_end_mask_0, squeeze_mask = var_728_squeeze_mask_0, x = var_725_cast_fp16)[name = tensor("op_728_cast_fp16")]; tensor var_743_begin_0 = const()[name = tensor("op_743_begin_0"), val = tensor([0, 12, 0, 0])]; tensor var_743_end_0 = const()[name = tensor("op_743_end_0"), val = tensor([1, 13, 1, 1500])]; tensor var_743_end_mask_0 = const()[name = tensor("op_743_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_743_cast_fp16 = slice_by_index(begin = var_743_begin_0, end = var_743_end_0, end_mask = var_743_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_743_cast_fp16")]; tensor var_746_begin_0 = const()[name = tensor("op_746_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_746_end_0 = const()[name = tensor("op_746_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_746_end_mask_0 = const()[name = tensor("op_746_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_746_squeeze_mask_0 = const()[name = tensor("op_746_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_746_cast_fp16 = slice_by_index(begin = var_746_begin_0, end = var_746_end_0, end_mask = var_746_end_mask_0, squeeze_mask = var_746_squeeze_mask_0, x = var_743_cast_fp16)[name = tensor("op_746_cast_fp16")]; tensor var_761_begin_0 = const()[name = tensor("op_761_begin_0"), val = tensor([0, 13, 0, 0])]; tensor var_761_end_0 = const()[name = tensor("op_761_end_0"), val = tensor([1, 14, 1, 1500])]; tensor var_761_end_mask_0 = const()[name = tensor("op_761_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_761_cast_fp16 = slice_by_index(begin = var_761_begin_0, end = var_761_end_0, end_mask = var_761_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_761_cast_fp16")]; tensor var_764_begin_0 = const()[name = tensor("op_764_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_764_end_0 = const()[name = tensor("op_764_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_764_end_mask_0 = const()[name = tensor("op_764_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_764_squeeze_mask_0 = const()[name = tensor("op_764_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_764_cast_fp16 = slice_by_index(begin = var_764_begin_0, end = var_764_end_0, end_mask = var_764_end_mask_0, squeeze_mask = var_764_squeeze_mask_0, x = var_761_cast_fp16)[name = tensor("op_764_cast_fp16")]; tensor var_779_begin_0 = const()[name = tensor("op_779_begin_0"), val = tensor([0, 14, 0, 0])]; tensor var_779_end_0 = const()[name = tensor("op_779_end_0"), val = tensor([1, 15, 1, 1500])]; tensor var_779_end_mask_0 = const()[name = tensor("op_779_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_779_cast_fp16 = slice_by_index(begin = var_779_begin_0, end = var_779_end_0, end_mask = var_779_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_779_cast_fp16")]; tensor var_782_begin_0 = const()[name = tensor("op_782_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_782_end_0 = const()[name = tensor("op_782_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_782_end_mask_0 = const()[name = tensor("op_782_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_782_squeeze_mask_0 = const()[name = tensor("op_782_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_782_cast_fp16 = slice_by_index(begin = var_782_begin_0, end = var_782_end_0, end_mask = var_782_end_mask_0, squeeze_mask = var_782_squeeze_mask_0, x = var_779_cast_fp16)[name = tensor("op_782_cast_fp16")]; tensor var_797_begin_0 = const()[name = tensor("op_797_begin_0"), val = tensor([0, 15, 0, 0])]; tensor var_797_end_0 = const()[name = tensor("op_797_end_0"), val = tensor([1, 16, 1, 1500])]; tensor var_797_end_mask_0 = const()[name = tensor("op_797_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_797_cast_fp16 = slice_by_index(begin = var_797_begin_0, end = var_797_end_0, end_mask = var_797_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_797_cast_fp16")]; tensor var_800_begin_0 = const()[name = tensor("op_800_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_800_end_0 = const()[name = tensor("op_800_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_800_end_mask_0 = const()[name = tensor("op_800_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_800_squeeze_mask_0 = const()[name = tensor("op_800_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_800_cast_fp16 = slice_by_index(begin = var_800_begin_0, end = var_800_end_0, end_mask = var_800_end_mask_0, squeeze_mask = var_800_squeeze_mask_0, x = var_797_cast_fp16)[name = tensor("op_800_cast_fp16")]; tensor var_815_begin_0 = const()[name = tensor("op_815_begin_0"), val = tensor([0, 16, 0, 0])]; tensor var_815_end_0 = const()[name = tensor("op_815_end_0"), val = tensor([1, 17, 1, 1500])]; tensor var_815_end_mask_0 = const()[name = tensor("op_815_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_815_cast_fp16 = slice_by_index(begin = var_815_begin_0, end = var_815_end_0, end_mask = var_815_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_815_cast_fp16")]; tensor var_818_begin_0 = const()[name = tensor("op_818_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_818_end_0 = const()[name = tensor("op_818_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_818_end_mask_0 = const()[name = tensor("op_818_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_818_squeeze_mask_0 = const()[name = tensor("op_818_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_818_cast_fp16 = slice_by_index(begin = var_818_begin_0, end = var_818_end_0, end_mask = var_818_end_mask_0, squeeze_mask = var_818_squeeze_mask_0, x = var_815_cast_fp16)[name = tensor("op_818_cast_fp16")]; tensor var_833_begin_0 = const()[name = tensor("op_833_begin_0"), val = tensor([0, 17, 0, 0])]; tensor var_833_end_0 = const()[name = tensor("op_833_end_0"), val = tensor([1, 18, 1, 1500])]; tensor var_833_end_mask_0 = const()[name = tensor("op_833_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_833_cast_fp16 = slice_by_index(begin = var_833_begin_0, end = var_833_end_0, end_mask = var_833_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_833_cast_fp16")]; tensor var_836_begin_0 = const()[name = tensor("op_836_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_836_end_0 = const()[name = tensor("op_836_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_836_end_mask_0 = const()[name = tensor("op_836_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_836_squeeze_mask_0 = const()[name = tensor("op_836_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_836_cast_fp16 = slice_by_index(begin = var_836_begin_0, end = var_836_end_0, end_mask = var_836_end_mask_0, squeeze_mask = var_836_squeeze_mask_0, x = var_833_cast_fp16)[name = tensor("op_836_cast_fp16")]; tensor var_851_begin_0 = const()[name = tensor("op_851_begin_0"), val = tensor([0, 18, 0, 0])]; tensor var_851_end_0 = const()[name = tensor("op_851_end_0"), val = tensor([1, 19, 1, 1500])]; tensor var_851_end_mask_0 = const()[name = tensor("op_851_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_851_cast_fp16 = slice_by_index(begin = var_851_begin_0, end = var_851_end_0, end_mask = var_851_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_851_cast_fp16")]; tensor var_854_begin_0 = const()[name = tensor("op_854_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_854_end_0 = const()[name = tensor("op_854_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_854_end_mask_0 = const()[name = tensor("op_854_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_854_squeeze_mask_0 = const()[name = tensor("op_854_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_854_cast_fp16 = slice_by_index(begin = var_854_begin_0, end = var_854_end_0, end_mask = var_854_end_mask_0, squeeze_mask = var_854_squeeze_mask_0, x = var_851_cast_fp16)[name = tensor("op_854_cast_fp16")]; tensor var_869_begin_0 = const()[name = tensor("op_869_begin_0"), val = tensor([0, 19, 0, 0])]; tensor var_869_end_0 = const()[name = tensor("op_869_end_0"), val = tensor([1, 20, 1, 1500])]; tensor var_869_end_mask_0 = const()[name = tensor("op_869_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_869_cast_fp16 = slice_by_index(begin = var_869_begin_0, end = var_869_end_0, end_mask = var_869_end_mask_0, x = obj_27_cast_fp16)[name = tensor("op_869_cast_fp16")]; tensor var_872_begin_0 = const()[name = tensor("op_872_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_872_end_0 = const()[name = tensor("op_872_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_872_end_mask_0 = const()[name = tensor("op_872_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_872_squeeze_mask_0 = const()[name = tensor("op_872_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_872_cast_fp16 = slice_by_index(begin = var_872_begin_0, end = var_872_end_0, end_mask = var_872_end_mask_0, squeeze_mask = var_872_squeeze_mask_0, x = var_869_cast_fp16)[name = tensor("op_872_cast_fp16")]; tensor var_879 = const()[name = tensor("op_879"), val = tensor(1)]; tensor var_880_interleave_0 = const()[name = tensor("op_880_interleave_0"), val = tensor(false)]; tensor var_880_cast_fp16 = concat(axis = var_879, interleave = var_880_interleave_0, values = (var_530_cast_fp16, var_548_cast_fp16, var_566_cast_fp16, var_584_cast_fp16, var_602_cast_fp16, var_620_cast_fp16, var_638_cast_fp16, var_656_cast_fp16, var_674_cast_fp16, var_692_cast_fp16, var_710_cast_fp16, var_728_cast_fp16, var_746_cast_fp16, var_764_cast_fp16, var_782_cast_fp16, var_800_cast_fp16, var_818_cast_fp16, var_836_cast_fp16, var_854_cast_fp16, var_872_cast_fp16))[name = tensor("op_880_cast_fp16")]; tensor var_882 = const()[name = tensor("op_882"), val = tensor([1])]; tensor var_883 = const()[name = tensor("op_883"), val = tensor(false)]; tensor alignment_heads_weights = reduce_mean(axes = var_882, keep_dims = var_883, x = var_880_cast_fp16)[name = tensor("obj_cast_fp16")]; } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights); }