A brief summary of the potential issues during the replication and corresponding solutons

[puyuanliu]

1. module transformers has no attribute LLaMATokenizer or 'missing key 'llama'.

First, install the SentencePiece then install transformers from huggingface git repo. i.e., pip install sentencepiece, pip install git+https://github.com/huggingface/transformers.git
The installation order matters.

2. CUDA OOM at the beginning of the training.

Use -fp 16 instead of -bp 16. Lower the batch size and gradient accumulation steps.

3. CUDA OOM during model saving.

Assume you are using torch=1.13.0, change python/lib/python3.9/site packages/torch/distributed/fsdp/fully_sharded_data_parallel.py:2224 from state_dict[fqn] = state_dict[fqn].clone().detach() to state_dict[fqn] = state_dict[fqn].cpu().clone().detach()

This usually happens when using GPUs of small memory (e.g., 40GB or 24GB)

4. How to perform inference?

Refer to #35 (comment)

5. Generated tokens are not human-readable at inference time.

Assume your training goes well (e.g., training loss <0.5), it's most likely your model weights are corrupted during model saving. Make sure there is no error message during the saving.

6. Finetuning is slow.

Refer to #32 (comment)

[ZeyuTeng96]

Hello my friend, like finding treasures in this issue. I had a QQ chat group. Are u willing to come in and help all Chinese friends. My QQ chat group number is: 397447632

[datquocnguyen]

Regarding the CUDA OOM during model saving, with python 3.10: we should make the change in python3.10/site-packages/torch/distributed/fsdp/_state_dict_utils.py