Empty input_ids when training use train.py

Nobody has claimed this yet.

I am running train.py using the following parameters:

```
torchrun --nproc_per_node=4 --master_port=20001 fastchat/train/train.py \
--model_name_or_path /home/ubuntu/FastChat/llama_7B \
--data_path playground/data/dummy.json \
--fp16 True \
--output_dir output \
--num_train_epochs 3 \
--per_device_train_batch_size 1 \
--per_device_eval_batch_size 1 \
--gradient_accumulation_steps 16 \
--evaluation_strategy "no" \
--save_strategy "steps" \
--save_steps 1200 \
--save_total_limit 10 \
--learning_rate 1e-4 \
--weight_decay 0. \
--warmup_ratio 0.03 \
--lr_scheduler_type "cosine" \
--logging_steps 1 \
--model_max_length 1028 \
--gradient_checkpointing True \
--lazy_preprocess True
```

Got Error of empty input_ids. I am using the dummy sample set from playground:

Any ideas for the error? Thanks!

The entry point is fastchat/train/train.py and the reported input is playground/data/dummy.json. Reproduce the supplied torchrun command, then inspect how that dataset is handled and where the empty input_ids error is raised; done means the dummy dataset trains without that error.

Written by the indexing model from the issue text.