是否应该修改roll/distributed/scheduler/generate_scheduler.py?
gen_batch = request_data.pop(
batch_keys=["input_ids", "attention_mask", "position_ids"],
non_tensor_batch_keys=["multi_modal_data"] if "multi_modal_data" in request_data.non_tensor_batch else []
)
gen_batch.meta_info = request_data.meta_info
num_return_sequences = generation_config["num_return_sequences"]
is_num_return_sequences_expand = self.pipeline_config.is_num_return_sequences_expand
if is_num_return_sequences_expand:
request_data = request_data.repeat(repeat_times=num_return_sequences)
gen_batch.meta_info["generation_config"]["num_return_sequences"] = 1
num_return_sequences = 1
是否应该修改
roll/distributed/scheduler/generate_scheduler.py?