Question: Why not padding to the same sequence length within the batch during the sft training phase?

Question: In the SFT training phase in dschat, I found that function create_dataset_split in data_utils.py will pad the samples to the maximum length. Therefore, why not dynamically padding to the maximum length of samples in the batch during training, which can significantly speed up training.
if chosen_sentence is not None:
chosen_sentence += end_of_conversation_token
chosen_token = tokenizer(chosen_sentence,
max_length=max_seq_len,
padding="max_length",
truncation=True,
return_tensors="pt")
chosen_token["input_ids"] = chosen_token["input_ids"].squeeze(0)
chosen_token["attention_mask"] = chosen_token["attention_mask"].squeeze(0)
chosen_dataset.append(chosen_token)

Start by reading data_utils.py and the create_dataset_split function, focusing on how SFT samples are tokenized and padded. Trace how those tensors enter the training batches, then determine whether batch-level padding is supported; done means documenting the reason for the current max-length padding or defining the changes needed for dynamic padding.