Training processing when tokenization mismatch

Nobody has claimed this yet.

```python
if cur_len < tokenizer.model_max_length:
if cur_len != total_len:
--> target[:] = IGNORE_TOKEN_ID
rank0_print(
f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
f" #turn = {len(turns) - 1}. (ignored)"
)
```

Line 158 of https://github.com/lm-sys/FastChat/blob/main/fastchat/train/train.py set all the target tokens to ignore_token_id.
Why not mask out mismatch turn (usually the last turn) only? `target[cur_len:] = IGNORE_TOKEN_ID`

Start in fastchat/train/train.py around line 158, where the tokenization-mismatch branch assigns IGNORE_TOKEN_ID to the full target. Review how cur_len and total_len determine the mismatch, then verify that only the mismatched portion is ignored while valid target tokens remain available for training.

Written by the indexing model from the issue text.