-
Notifications
You must be signed in to change notification settings - Fork 982
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Support for LLaMA #841
Support for LLaMA #841
Conversation
@zphang where are we on this? |
Other than the above note on LLaMAMLP, I can incorporate the necessary changes and update the PR. |
Addressed comments. Please take another look. For further additions (e.g. guide to tuning a LLaMA model), I can do a separate PR. |
I'm confused by the Otherwise, this looks good to me. @Quentin-Anthony, any thoughts? |
|
Looking good to me. I'm gonna run some tests then merge. |
@zphang We need you to sign the CLA before merging this PR :) |
Signed! |
all comments appear resolved now
Works for all of my tests. Merging. |
@zphang Thanks for your work! It seems that there is no params.json in the llama checkpoint. Where can I get it? Thanks! |
@zphang Hi, thank you for your work! can you provide a |
Hi @DaoD, this |
Thanks so much! |
Yes. I have converted the checkpoint and tried to use the 6-7B's config to load it, but there are some missing keys and unexpected keys in the dict. Could you please provide a readme for using it? |
@zphang I think we need a new convert_sequential_to_hf.py to convert the obatined model into HF style. Have you done something about this? Thanks! |
@zphang in the tools/convert_raw_llama_weights_to_neox.py, how does one convert llama tokenizer?
|
I think you do not need to convert llama tokenizer. Just set
The tokenizer.model can be obtained from this link. |
I find another problem. The eod token's id of SentencePieceTokenizer (eos_token_id=0) is differnet from the orginal LlamaTokenizer (eos_token_id =1), which may cause some problems in training and inference. |
Hi @DaoD, have you tried to do inference with this model, can you generate reasonable text with it? |
Yes, but I do not use the inference code provided by gpt-neox. I just convert the model into Huggingface style, and use the HF function for generation. It seems correct. |
Thx! This helps a lot:) |
Hello, can you convert a gpt-neox llama model to HF style? |
No description provided.