Unable to Fine-Tune

Following the instruction https://github.com/lm-sys/FastChat, I am able to inference but not able to train.

I am following the fine-tuning section "Fine-tuning Vicuna-7B with Local GPUs" with the script

> torchrun --nproc_per_node=2 --master_port=20001 fastchat/train/train_mem.py \
>     --model_name_or_path ~/model_weights/llama-7b  \
>     --data_path data/dummy_conversation.json \
>     --bf16 True \
>     --output_dir output_vicuna \
>     --num_train_epochs 3 \
>     --per_device_train_batch_size 2 \
>     --per_device_eval_batch_size 2 \
>     --gradient_accumulation_steps 16 \
>     --evaluation_strategy "no" \
>     --save_strategy "steps" \
>     --save_steps 1200 \
>     --save_total_limit 10 \
>     --learning_rate 2e-5 \
>     --weight_decay 0. \
>     --warmup_ratio 0.03 \
>     --lr_scheduler_type "cosine" \
>     --logging_steps 1 \
>     --fsdp "full_shard auto_wrap" \
>     --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \
>     --tf32 True \
>     --model_max_length 2048 \
>     --gradient_checkpointing True \
>     --lazy_preprocess True

I first tried to read ~/model_weights/llama-7b, which use the llama-7b parameters. I downloaded the model from the link in official website and got params.json consolidated.00.pth checklist.chk. the script is unable to recognize this three files, so I rename params.json with the name config.json and rename the consolidated.00.pth with the name pytorch_model.bin but it is not working.

I then try to directly train the vicuna model with the script. so I replace ~/model_weights/llama-7b in the script with the directory pointing to my vicuna model, but unable to do so.

Could anyone answer this question? thank you very much.

Start with the fine-tuning section of the FastChat instructions and fastchat/train/train_mem.py, then reproduce the command using the reported model directory while capturing the complete error output. Compare the expected model files with the downloaded LLaMA or Vicuna files; done means the documented command recognizes the model and begins training, or the instructions clearly explain the required conversion.