Reproducing sentiment finetuning train_lora extremely slow

I am trying to reproduce the finetuning for the fingpt-sentiment_llama2-13b_lora

The table claims we can do this in just a single RTX 3090 within a day.
I am using a L4 GPU instead.

I downloaded the models to base_models and the dataset to data correctly

I used the script like this

```
deepspeed -i train_lora.py \
--run_name sentiment-llama2-13b-20epoch-64batch \
--base_model llama2-13b-nr \
--dataset sentiment-train \
--max_length 512 \
--batch_size 16 \
--learning_rate 1e-4 \
--num_epochs 20 \
```

I got an OOM.

So i set the `load_in_8_bit=True`

But I am getting extremely slow fine tuning speed A single epoch is estimated to take 2 days.