out ot memory when i use 32GB V100s to fine-tuning Vicuna-7B-v1.5 with Lora

Nobody has claimed this yet.

Here is my Lora fine-tuning script, when `model_max_length=2048` it will be OOM, when setting 1024 it won't, what's the reason for this, I saw other people seem to be able to fine-tune successfully in 2048 case.
```
deepspeed --include localhost:3 --master_port=9901 fastchat/train/train_lora.py \
--model_name_or_path ./models/vicuna-7B \
--lora_r 8 \
--lora_alpha 16 \
--lora_dropout 0.05 \
--data_path ./data/new_sft_train_data.json \
--fp16 True \
--output_dir ./checkpoints \
--num_train_epochs 3 \
--per_device_train_batch_size 1 \
--per_device_eval_batch_size 1 \
--gradient_accumulation_steps 1 \
--evaluation_strategy "no" \
--save_strategy "steps" \
--save_steps 5000 \
--save_total_limit 100 \
--learning_rate 2e-5 \
--weight_decay 0. \
--warmup_ratio 0.03 \
--lr_scheduler_type "cosine" \
--logging_steps 1 \
--model_max_length 2048 \
--q_lora True \
--deepspeed playground/deepspeed_config_s2.json

+-------------------------------+----------------------+----------------------+
| 3 Tesla V100S-PCI... Off | 00000000:00:06.0 Off | 0 |
| N/A 30C P0 25W / 250W | 4MiB / 32768MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+

Start with fastchat/train/train_lora.py and playground/deepspeed_config_s2.json, then reproduce the command with model_max_length set to 1024 and 2048 on the stated V100 setup. Trace the memory use around the LoRA and QLoRA training path; done means documenting the cause of the 2048 OOM and a verified configuration or code change that permits the larger length.

Written by the indexing model from the issue text.