diff --git "a/openai_whisper-tiny/TextDecoder.mlmodelc/model.mil" "b/openai_whisper-tiny/TextDecoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/openai_whisper-tiny/TextDecoder.mlmodelc/model.mil" @@ -0,0 +1,773 @@ +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(39832448)))]; + 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(40176576)))]; + 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(40177408)))]; + 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(40178240)))]; + 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(40179072)))]; + 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(40179904)))]; + 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(40474880)))]; + 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(40475712)))]; + 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(40770688)))]; + 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(41065664)))]; + 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(41066496)))]; + 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(41361472)))]; + 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(41362304)))]; + 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(41363136)))]; + 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(41363968)))]; + 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(41658944)))]; + 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(41659776)))]; + 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(41954752)))]; + 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(42249728)))]; + 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(42250560)))]; + 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(42545536)))]; + 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(42546368)))]; + 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(42547200)))]; + 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(42548032)))]; + 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(43727744)))]; + 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(43730880)))]; + 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(44910592)))]; + 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(44911424)))]; + 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(44912256)))]; + 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(44913088)))]; + 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(45208064)))]; + 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(45208896)))]; + 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(45503872)))]; + 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(45798848)))]; + 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(45799680)))]; + 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(46094656)))]; + 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(46095488)))]; + 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(46096320)))]; + 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(46097152)))]; + 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(46392128)))]; + 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(46392960)))]; + 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(46687936)))]; + 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(46982912)))]; + 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(46983744)))]; + 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(47278720)))]; + 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_451 = const()[name = tensor("op_451"), val = tensor([1])]; + tensor channels_mean_11_cast_fp16 = reduce_mean(axes = var_451, 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_455 = const()[name = tensor("op_455"), val = tensor([1])]; + tensor var_456_cast_fp16 = reduce_mean(axes = var_455, keep_dims = var_286, x = zero_mean_sq_11_cast_fp16)[name = tensor("op_456_cast_fp16")]; + tensor var_457_to_fp16 = const()[name = tensor("op_457_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_458_cast_fp16 = add(x = var_456_cast_fp16, y = var_457_to_fp16)[name = tensor("op_458_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_458_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(47279552)))]; + 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(47280384)))]; + 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_469 = const()[name = tensor("op_469"), val = tensor([1, 1])]; + tensor var_471 = const()[name = tensor("op_471"), 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(47281216)))]; + 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(48460928)))]; + tensor input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = var_471, groups = var_285, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = var_469, 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_477 = const()[name = tensor("op_477"), val = tensor([1, 1])]; + tensor var_479 = const()[name = tensor("op_479"), 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(48464064)))]; + 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(49643776)))]; + tensor hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = var_479, groups = var_285, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = var_477, 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_492 = const()[name = tensor("op_492"), val = tensor(3)]; + tensor var_499 = const()[name = tensor("op_499"), val = tensor(1)]; + tensor var_500 = const()[name = tensor("op_500"), val = tensor(true)]; + tensor var_512 = const()[name = tensor("op_512"), val = tensor([1])]; + tensor channels_mean_13_cast_fp16 = reduce_mean(axes = var_512, keep_dims = var_500, 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_516 = const()[name = tensor("op_516"), val = tensor([1])]; + tensor var_517_cast_fp16 = reduce_mean(axes = var_516, keep_dims = var_500, x = zero_mean_sq_13_cast_fp16)[name = tensor("op_517_cast_fp16")]; + tensor var_518_to_fp16 = const()[name = tensor("op_518_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_519_cast_fp16 = add(x = var_517_cast_fp16, y = var_518_to_fp16)[name = tensor("op_519_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_519_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(49644608)))]; + 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(49645440)))]; + 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_534 = const()[name = tensor("op_534"), val = tensor([1, 1])]; + tensor var_536 = const()[name = tensor("op_536"), 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(49646272)))]; + 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(49941248)))]; + tensor query_9_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = var_536, groups = var_499, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = var_534, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("query_9_cast_fp16")]; + tensor var_540 = const()[name = tensor("op_540"), val = tensor([1, 1])]; + tensor var_542 = const()[name = tensor("op_542"), 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(49942080)))]; + tensor current_key_5_cast_fp16 = conv(dilations = var_542, groups = var_499, pad = current_key_5_pad_0, pad_type = current_key_5_pad_type_0, strides = var_540, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("current_key_5_cast_fp16")]; + tensor var_547 = const()[name = tensor("op_547"), val = tensor([1, 1])]; + tensor var_549 = const()[name = tensor("op_549"), 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(50237056)))]; + 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(50532032)))]; + tensor current_value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = var_549, groups = var_499, pad = current_value_5_pad_0, pad_type = current_value_5_pad_type_0, strides = var_547, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("current_value_5_cast_fp16")]; + tensor var_556_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_126_cast_fp16)[name = tensor("op_556_cast_fp16")]; + tensor var_558_cast_fp16 = mul(x = var_47_cast_fp16_2, y = var_129_cast_fp16)[name = tensor("op_558_cast_fp16")]; + tensor key_9_cast_fp16 = add(x = var_556_cast_fp16, y = var_558_cast_fp16)[name = tensor("key_9_cast_fp16")]; + tensor var_560_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_126_cast_fp16)[name = tensor("op_560_cast_fp16")]; + tensor var_562_cast_fp16 = mul(x = var_54_cast_fp16_2, y = var_129_cast_fp16)[name = tensor("op_562_cast_fp16")]; + tensor value_9_cast_fp16 = add(x = var_560_cast_fp16, y = var_562_cast_fp16)[name = tensor("value_9_cast_fp16")]; + tensor var_565 = const()[name = tensor("op_565"), val = tensor([1, 6, 64, -1])]; + tensor var_566_cast_fp16 = reshape(shape = var_565, x = query_9_cast_fp16)[name = tensor("op_566_cast_fp16")]; + tensor var_567_to_fp16 = const()[name = tensor("op_567_to_fp16"), val = tensor(0x1p-3)]; + tensor var_568_cast_fp16 = mul(x = var_566_cast_fp16, y = var_567_to_fp16)[name = tensor("op_568_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 = key_9_cast_fp16)[name = tensor("op_570_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_568_cast_fp16, y = var_570_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_578_cast_fp16 = softmax(axis = var_492, x = mh_w_15_cast_fp16)[name = tensor("op_578_cast_fp16")]; + tensor var_579 = const()[name = tensor("op_579"), val = tensor([1, 6, 64, -1])]; + tensor var_580_cast_fp16 = reshape(shape = var_579, x = value_9_cast_fp16)[name = tensor("op_580_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_580_cast_fp16, y = var_578_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_583 = const()[name = tensor("op_583"), val = tensor([1, 384, 1, -1])]; + tensor input_21_cast_fp16 = reshape(shape = var_583, x = attn_9_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor var_587 = const()[name = tensor("op_587"), val = tensor([1, 1])]; + tensor var_589 = const()[name = tensor("op_589"), 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(50532864)))]; + 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(50827840)))]; + tensor obj_35_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = var_589, groups = var_499, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = var_587, 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_599 = const()[name = tensor("op_599"), val = tensor([1])]; + tensor channels_mean_15_cast_fp16 = reduce_mean(axes = var_599, keep_dims = var_500, 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_603 = const()[name = tensor("op_603"), val = tensor([1])]; + tensor var_604_cast_fp16 = reduce_mean(axes = var_603, keep_dims = var_500, x = zero_mean_sq_15_cast_fp16)[name = tensor("op_604_cast_fp16")]; + tensor var_605_to_fp16 = const()[name = tensor("op_605_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_606_cast_fp16 = add(x = var_604_cast_fp16, y = var_605_to_fp16)[name = tensor("op_606_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_606_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(50828672)))]; + 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(50829504)))]; + 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_621 = const()[name = tensor("op_621"), val = tensor([1, 1])]; + tensor var_623 = const()[name = tensor("op_623"), 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(50830336)))]; + 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(51125312)))]; + tensor query_11_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_bias_to_fp16, dilations = var_623, groups = var_499, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = var_621, weight = layers_2_encoder_attn_q_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("query_11_cast_fp16")]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor([1, 1])]; + tensor var_629 = const()[name = tensor("op_629"), 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(51126144)))]; + tensor key_11_cast_fp16 = conv(dilations = var_629, groups = var_499, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = var_627, weight = layers_2_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_11_cast_fp16")]; + tensor var_634 = const()[name = tensor("op_634"), val = tensor([1, 1])]; + tensor var_636 = const()[name = tensor("op_636"), 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(51421120)))]; + 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(51716096)))]; + tensor value_11_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_bias_to_fp16, dilations = var_636, groups = var_499, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = var_634, weight = layers_2_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_11_cast_fp16")]; + tensor var_640 = const()[name = tensor("op_640"), val = tensor([1, 6, 64, -1])]; + tensor var_641_cast_fp16 = reshape(shape = var_640, x = query_11_cast_fp16)[name = tensor("op_641_cast_fp16")]; + tensor var_642_to_fp16 = const()[name = tensor("op_642_to_fp16"), val = tensor(0x1p-3)]; + tensor var_643_cast_fp16 = mul(x = var_641_cast_fp16, y = var_642_to_fp16)[name = tensor("op_643_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 = key_11_cast_fp16)[name = tensor("op_645_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_643_cast_fp16, y = var_645_cast_fp16)[name = tensor("mh_w_17_cast_fp16")]; + tensor obj_41_cast_fp16 = softmax(axis = var_492, x = mh_w_17_cast_fp16)[name = tensor("obj_41_cast_fp16")]; + tensor var_649 = const()[name = tensor("op_649"), val = tensor([1, 6, 64, -1])]; + tensor var_650_cast_fp16 = reshape(shape = var_649, x = value_11_cast_fp16)[name = tensor("op_650_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_650_cast_fp16, y = obj_41_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_653 = const()[name = tensor("op_653"), val = tensor([1, 384, 1, -1])]; + tensor input_23_cast_fp16 = reshape(shape = var_653, x = attn_11_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor var_657 = const()[name = tensor("op_657"), val = tensor([1, 1])]; + tensor var_659 = const()[name = tensor("op_659"), 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(51716928)))]; + 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(52011904)))]; + tensor obj_39_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_bias_to_fp16, dilations = var_659, groups = var_499, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = var_657, 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_668 = const()[name = tensor("op_668"), val = tensor([1])]; + tensor channels_mean_17_cast_fp16 = reduce_mean(axes = var_668, keep_dims = var_500, 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_672 = const()[name = tensor("op_672"), val = tensor([1])]; + tensor var_673_cast_fp16 = reduce_mean(axes = var_672, keep_dims = var_500, x = zero_mean_sq_17_cast_fp16)[name = tensor("op_673_cast_fp16")]; + tensor var_674_to_fp16 = const()[name = tensor("op_674_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_675_cast_fp16 = add(x = var_673_cast_fp16, y = var_674_to_fp16)[name = tensor("op_675_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_675_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(52012736)))]; + 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(52013568)))]; + 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_686 = const()[name = tensor("op_686"), val = tensor([1, 1])]; + tensor var_688 = const()[name = tensor("op_688"), 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(52014400)))]; + 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(53194112)))]; + tensor input_27_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = var_688, groups = var_499, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = var_686, 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_694 = const()[name = tensor("op_694"), val = tensor([1, 1])]; + tensor var_696 = const()[name = tensor("op_696"), 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(53197248)))]; + 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(54376960)))]; + tensor hidden_states_7_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = var_696, groups = var_499, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = var_694, 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_710 = const()[name = tensor("op_710"), val = tensor(3)]; + tensor var_717 = const()[name = tensor("op_717"), val = tensor(1)]; + tensor var_718 = const()[name = tensor("op_718"), val = tensor(true)]; + tensor var_730 = const()[name = tensor("op_730"), val = tensor([1])]; + tensor channels_mean_19_cast_fp16 = reduce_mean(axes = var_730, keep_dims = var_718, 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_734 = const()[name = tensor("op_734"), val = tensor([1])]; + tensor var_735_cast_fp16 = reduce_mean(axes = var_734, keep_dims = var_718, x = zero_mean_sq_19_cast_fp16)[name = tensor("op_735_cast_fp16")]; + tensor var_736_to_fp16 = const()[name = tensor("op_736_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_737_cast_fp16 = add(x = var_735_cast_fp16, y = var_736_to_fp16)[name = tensor("op_737_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_737_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(54377792)))]; + 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(54378624)))]; + 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_752 = const()[name = tensor("op_752"), val = tensor([1, 1])]; + tensor var_754 = const()[name = tensor("op_754"), 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(54379456)))]; + 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(54674432)))]; + tensor query_13_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = var_754, groups = var_717, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = var_752, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("query_13_cast_fp16")]; + tensor var_758 = const()[name = tensor("op_758"), val = tensor([1, 1])]; + tensor var_760 = const()[name = tensor("op_760"), 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(54675264)))]; + tensor current_key_cast_fp16 = conv(dilations = var_760, groups = var_717, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = var_758, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("current_key_cast_fp16")]; + tensor var_765 = const()[name = tensor("op_765"), val = tensor([1, 1])]; + tensor var_767 = const()[name = tensor("op_767"), 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(54970240)))]; + 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(55265216)))]; + tensor current_value_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = var_767, groups = var_717, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = var_765, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("current_value_cast_fp16")]; + tensor var_774_cast_fp16 = mul(x = current_key_cast_fp16, y = var_126_cast_fp16)[name = tensor("op_774_cast_fp16")]; + tensor var_776_cast_fp16 = mul(x = var_47_cast_fp16_3, y = var_129_cast_fp16)[name = tensor("op_776_cast_fp16")]; + tensor key_13_cast_fp16 = add(x = var_774_cast_fp16, y = var_776_cast_fp16)[name = tensor("key_13_cast_fp16")]; + tensor var_778_cast_fp16 = mul(x = current_value_cast_fp16, y = var_126_cast_fp16)[name = tensor("op_778_cast_fp16")]; + tensor var_780_cast_fp16 = mul(x = var_54_cast_fp16_3, y = var_129_cast_fp16)[name = tensor("op_780_cast_fp16")]; + tensor value_13_cast_fp16 = add(x = var_778_cast_fp16, y = var_780_cast_fp16)[name = tensor("value_13_cast_fp16")]; + tensor var_783 = const()[name = tensor("op_783"), val = tensor([1, 6, 64, -1])]; + tensor var_784_cast_fp16 = reshape(shape = var_783, x = query_13_cast_fp16)[name = tensor("op_784_cast_fp16")]; + tensor var_785_to_fp16 = const()[name = tensor("op_785_to_fp16"), val = tensor(0x1p-3)]; + tensor var_786_cast_fp16 = mul(x = var_784_cast_fp16, y = var_785_to_fp16)[name = tensor("op_786_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 = key_13_cast_fp16)[name = tensor("op_788_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_786_cast_fp16, y = var_788_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_796_cast_fp16 = softmax(axis = var_710, x = mh_w_21_cast_fp16)[name = tensor("op_796_cast_fp16")]; + tensor var_797 = const()[name = tensor("op_797"), val = tensor([1, 6, 64, -1])]; + tensor var_798_cast_fp16 = reshape(shape = var_797, x = value_13_cast_fp16)[name = tensor("op_798_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_798_cast_fp16, y = var_796_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_801 = const()[name = tensor("op_801"), val = tensor([1, 384, 1, -1])]; + tensor input_31_cast_fp16 = reshape(shape = var_801, x = attn_13_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 1])]; + tensor var_807 = const()[name = tensor("op_807"), 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(55266048)))]; + 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(55561024)))]; + tensor obj_49_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = var_807, groups = var_717, pad = obj_49_pad_0, pad_type = obj_49_pad_type_0, strides = var_805, 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_817 = const()[name = tensor("op_817"), val = tensor([1])]; + tensor channels_mean_21_cast_fp16 = reduce_mean(axes = var_817, keep_dims = var_718, 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_821 = const()[name = tensor("op_821"), val = tensor([1])]; + tensor var_822_cast_fp16 = reduce_mean(axes = var_821, keep_dims = var_718, x = zero_mean_sq_21_cast_fp16)[name = tensor("op_822_cast_fp16")]; + tensor var_823_to_fp16 = const()[name = tensor("op_823_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_824_cast_fp16 = add(x = var_822_cast_fp16, y = var_823_to_fp16)[name = tensor("op_824_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_824_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(55561856)))]; + 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(55562688)))]; + 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_839 = const()[name = tensor("op_839"), val = tensor([1, 1])]; + tensor var_841 = const()[name = tensor("op_841"), 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(55563520)))]; + 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(55858496)))]; + tensor query_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_bias_to_fp16, dilations = var_841, groups = var_717, pad = query_pad_0, pad_type = query_pad_type_0, strides = var_839, weight = layers_3_encoder_attn_q_proj_weight_to_fp16, x = obj_51_cast_fp16)[name = tensor("query_cast_fp16")]; + tensor var_845 = const()[name = tensor("op_845"), val = tensor([1, 1])]; + tensor var_847 = const()[name = tensor("op_847"), 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(55859328)))]; + tensor key_cast_fp16 = conv(dilations = var_847, groups = var_717, pad = key_pad_0, pad_type = key_pad_type_0, strides = var_845, weight = layers_3_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_cast_fp16")]; + tensor var_852 = const()[name = tensor("op_852"), val = tensor([1, 1])]; + tensor var_854 = const()[name = tensor("op_854"), 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(56154304)))]; + 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(56449280)))]; + tensor value_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_bias_to_fp16, dilations = var_854, groups = var_717, pad = value_pad_0, pad_type = value_pad_type_0, strides = var_852, weight = layers_3_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_cast_fp16")]; + tensor var_858 = const()[name = tensor("op_858"), val = tensor([1, 6, 64, -1])]; + tensor var_859_cast_fp16 = reshape(shape = var_858, x = query_cast_fp16)[name = tensor("op_859_cast_fp16")]; + tensor var_860_to_fp16 = const()[name = tensor("op_860_to_fp16"), val = tensor(0x1p-3)]; + tensor var_861_cast_fp16 = mul(x = var_859_cast_fp16, y = var_860_to_fp16)[name = tensor("op_861_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 = key_cast_fp16)[name = tensor("op_863_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_861_cast_fp16, y = var_863_cast_fp16)[name = tensor("mh_w_cast_fp16")]; + tensor obj_55_cast_fp16 = softmax(axis = var_710, x = mh_w_cast_fp16)[name = tensor("obj_55_cast_fp16")]; + tensor var_867 = const()[name = tensor("op_867"), val = tensor([1, 6, 64, -1])]; + tensor var_868_cast_fp16 = reshape(shape = var_867, x = value_cast_fp16)[name = tensor("op_868_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_868_cast_fp16, y = obj_55_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_871 = const()[name = tensor("op_871"), val = tensor([1, 384, 1, -1])]; + tensor input_33_cast_fp16 = reshape(shape = var_871, x = attn_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor var_875 = const()[name = tensor("op_875"), val = tensor([1, 1])]; + tensor var_877 = const()[name = tensor("op_877"), 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(56450112)))]; + 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(56745088)))]; + tensor obj_53_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_bias_to_fp16, dilations = var_877, groups = var_717, pad = obj_53_pad_0, pad_type = obj_53_pad_type_0, strides = var_875, 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_886 = const()[name = tensor("op_886"), val = tensor([1])]; + tensor channels_mean_23_cast_fp16 = reduce_mean(axes = var_886, keep_dims = var_718, 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_890 = const()[name = tensor("op_890"), val = tensor([1])]; + tensor var_891_cast_fp16 = reduce_mean(axes = var_890, keep_dims = var_718, x = zero_mean_sq_23_cast_fp16)[name = tensor("op_891_cast_fp16")]; + tensor var_892_to_fp16 = const()[name = tensor("op_892_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_893_cast_fp16 = add(x = var_891_cast_fp16, y = var_892_to_fp16)[name = tensor("op_893_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_893_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(56745920)))]; + 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(56746752)))]; + 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_904 = const()[name = tensor("op_904"), val = tensor([1, 1])]; + tensor var_906 = const()[name = tensor("op_906"), 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(56747584)))]; + 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(57927296)))]; + tensor input_37_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = var_906, groups = var_717, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = var_904, 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_912 = const()[name = tensor("op_912"), val = tensor([1, 1])]; + tensor var_914 = const()[name = tensor("op_914"), 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(57930432)))]; + 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(59110144)))]; + tensor hidden_states_9_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = var_914, groups = var_717, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = var_912, 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_925 = const()[name = tensor("op_925"), val = tensor(true)]; + tensor var_929 = const()[name = tensor("op_929"), val = tensor([1])]; + tensor channels_mean_cast_fp16 = reduce_mean(axes = var_929, keep_dims = var_925, 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_933 = const()[name = tensor("op_933"), val = tensor([1])]; + tensor var_934_cast_fp16 = reduce_mean(axes = var_933, keep_dims = var_925, x = zero_mean_sq_cast_fp16)[name = tensor("op_934_cast_fp16")]; + tensor var_935_to_fp16 = const()[name = tensor("op_935_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_936_cast_fp16 = add(x = var_934_cast_fp16, y = var_935_to_fp16)[name = tensor("op_936_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_936_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(59110976)))]; + 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(59111808)))]; + 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_946_axes_0 = const()[name = tensor("op_946_axes_0"), val = tensor([2])]; + tensor var_946_cast_fp16 = squeeze(axes = var_946_axes_0, x = hidden_states_cast_fp16)[name = tensor("op_946_cast_fp16")]; + tensor var_949_perm_0 = const()[name = tensor("op_949_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(59112640)))]; + tensor transpose_0 = transpose(perm = var_949_perm_0, x = var_946_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_953 = const()[name = tensor("op_953"), 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_953, interleave = obj_59_interleave_0, values = (current_key_1_cast_fp16, current_key_3_cast_fp16, current_key_5_cast_fp16, current_key_cast_fp16))[name = tensor("obj_59_cast_fp16")]; + tensor var_956 = const()[name = tensor("op_956"), 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_956, interleave = obj_61_interleave_0, values = (current_value_1_cast_fp16, current_value_3_cast_fp16, current_value_5_cast_fp16, current_value_cast_fp16))[name = tensor("obj_61_cast_fp16")]; + tensor var_967_begin_0 = const()[name = tensor("op_967_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_967_end_0 = const()[name = tensor("op_967_end_0"), val = tensor([1, 3, 1, 1500])]; + tensor var_967_end_mask_0 = const()[name = tensor("op_967_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_967_cast_fp16 = slice_by_index(begin = var_967_begin_0, end = var_967_end_0, end_mask = var_967_end_mask_0, x = obj_41_cast_fp16)[name = tensor("op_967_cast_fp16")]; + tensor var_970_begin_0 = const()[name = tensor("op_970_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_970_end_0 = const()[name = tensor("op_970_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_970_end_mask_0 = const()[name = tensor("op_970_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_970_squeeze_mask_0 = const()[name = tensor("op_970_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_970_cast_fp16 = slice_by_index(begin = var_970_begin_0, end = var_970_end_0, end_mask = var_970_end_mask_0, squeeze_mask = var_970_squeeze_mask_0, x = var_967_cast_fp16)[name = tensor("op_970_cast_fp16")]; + tensor var_985_begin_0 = const()[name = tensor("op_985_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_985_end_0 = const()[name = tensor("op_985_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_985_end_mask_0 = const()[name = tensor("op_985_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_985_cast_fp16 = slice_by_index(begin = var_985_begin_0, end = var_985_end_0, end_mask = var_985_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_985_cast_fp16")]; + tensor var_988_begin_0 = const()[name = tensor("op_988_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_988_end_0 = const()[name = tensor("op_988_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_988_end_mask_0 = const()[name = tensor("op_988_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_988_squeeze_mask_0 = const()[name = tensor("op_988_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_988_cast_fp16 = slice_by_index(begin = var_988_begin_0, end = var_988_end_0, end_mask = var_988_end_mask_0, squeeze_mask = var_988_squeeze_mask_0, x = var_985_cast_fp16)[name = tensor("op_988_cast_fp16")]; + tensor var_1003_begin_0 = const()[name = tensor("op_1003_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_1003_end_0 = const()[name = tensor("op_1003_end_0"), val = tensor([1, 3, 1, 1500])]; + tensor var_1003_end_mask_0 = const()[name = tensor("op_1003_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1003_cast_fp16 = slice_by_index(begin = var_1003_begin_0, end = var_1003_end_0, end_mask = var_1003_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1003_cast_fp16")]; + tensor var_1006_begin_0 = const()[name = tensor("op_1006_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1006_end_0 = const()[name = tensor("op_1006_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1006_end_mask_0 = const()[name = tensor("op_1006_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1006_squeeze_mask_0 = const()[name = tensor("op_1006_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1006_cast_fp16 = slice_by_index(begin = var_1006_begin_0, end = var_1006_end_0, end_mask = var_1006_end_mask_0, squeeze_mask = var_1006_squeeze_mask_0, x = var_1003_cast_fp16)[name = tensor("op_1006_cast_fp16")]; + tensor var_1021_begin_0 = const()[name = tensor("op_1021_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_1021_end_0 = const()[name = tensor("op_1021_end_0"), val = tensor([1, 4, 1, 1500])]; + tensor var_1021_end_mask_0 = const()[name = tensor("op_1021_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1021_cast_fp16 = slice_by_index(begin = var_1021_begin_0, end = var_1021_end_0, end_mask = var_1021_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1021_cast_fp16")]; + tensor var_1024_begin_0 = const()[name = tensor("op_1024_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1024_end_0 = const()[name = tensor("op_1024_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1024_end_mask_0 = const()[name = tensor("op_1024_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1024_squeeze_mask_0 = const()[name = tensor("op_1024_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1024_cast_fp16 = slice_by_index(begin = var_1024_begin_0, end = var_1024_end_0, end_mask = var_1024_end_mask_0, squeeze_mask = var_1024_squeeze_mask_0, x = var_1021_cast_fp16)[name = tensor("op_1024_cast_fp16")]; + tensor var_1039_begin_0 = const()[name = tensor("op_1039_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_1039_end_0 = const()[name = tensor("op_1039_end_0"), val = tensor([1, 5, 1, 1500])]; + tensor var_1039_end_mask_0 = const()[name = tensor("op_1039_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1039_cast_fp16 = slice_by_index(begin = var_1039_begin_0, end = var_1039_end_0, end_mask = var_1039_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1039_cast_fp16")]; + tensor var_1042_begin_0 = const()[name = tensor("op_1042_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1042_end_0 = const()[name = tensor("op_1042_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1042_end_mask_0 = const()[name = tensor("op_1042_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1042_squeeze_mask_0 = const()[name = tensor("op_1042_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1042_cast_fp16 = slice_by_index(begin = var_1042_begin_0, end = var_1042_end_0, end_mask = var_1042_end_mask_0, squeeze_mask = var_1042_squeeze_mask_0, x = var_1039_cast_fp16)[name = tensor("op_1042_cast_fp16")]; + tensor var_1057_begin_0 = const()[name = tensor("op_1057_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_1057_end_0 = const()[name = tensor("op_1057_end_0"), val = tensor([1, 6, 1, 1500])]; + tensor var_1057_end_mask_0 = const()[name = tensor("op_1057_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1057_cast_fp16 = slice_by_index(begin = var_1057_begin_0, end = var_1057_end_0, end_mask = var_1057_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1057_cast_fp16")]; + tensor var_1060_begin_0 = const()[name = tensor("op_1060_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1060_end_0 = const()[name = tensor("op_1060_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1060_end_mask_0 = const()[name = tensor("op_1060_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1060_squeeze_mask_0 = const()[name = tensor("op_1060_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1060_cast_fp16 = slice_by_index(begin = var_1060_begin_0, end = var_1060_end_0, end_mask = var_1060_end_mask_0, squeeze_mask = var_1060_squeeze_mask_0, x = var_1057_cast_fp16)[name = tensor("op_1060_cast_fp16")]; + tensor var_1067 = const()[name = tensor("op_1067"), val = tensor(1)]; + tensor var_1068_interleave_0 = const()[name = tensor("op_1068_interleave_0"), val = tensor(false)]; + tensor var_1068_cast_fp16 = concat(axis = var_1067, interleave = var_1068_interleave_0, values = (var_970_cast_fp16, var_988_cast_fp16, var_1006_cast_fp16, var_1024_cast_fp16, var_1042_cast_fp16, var_1060_cast_fp16))[name = tensor("op_1068_cast_fp16")]; + tensor var_1070 = const()[name = tensor("op_1070"), val = tensor([1])]; + tensor var_1071 = const()[name = tensor("op_1071"), val = tensor(false)]; + tensor alignment_heads_weights = reduce_mean(axes = var_1070, keep_dims = var_1071, x = var_1068_cast_fp16)[name = tensor("obj_cast_fp16")]; + } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights); +} \ No newline at end of file