dual-obsidian-client icon indicating copy to clipboard operation
dual-obsidian-client copied to clipboard

Published 20 hours ago • verified publisher icon paulbricman

feat: add gpt-neo model handling

Open onlurking opened this issue 4 years ago • 4 comments

This PR enables neo-gpt model loading using the same API from transformers from huggingface.

Related: #40

onlurking avatar Apr 18 '21 21:04 onlurking

Google Collab with Neo-GPT models support:

https://colab.research.google.com/drive/1xqEZeZY3aYl4w859Ej4sCsX-2LxBGU1l?usp=sharing

onlurking avatar Apr 18 '21 22:04 onlurking

Thanks for the PR! As mentioned on Discord (https://discord.com/channels/817119487999606794/825717174257319974/833584636533407745), I think using the AutoModel class from transformers would make the implementation somewhat simpler, as you can simply give it the local path to the model and it can figure out what's in there. What do you think? Not sure about the tokenizer, though, but I think both GPT-2 models and GPT-Neo use similar tokenizers?

paulbricman avatar Apr 19 '21 08:04 paulbricman

Hi @paulbricman!

I've followed the transformers docs and both models use the same GPT2Tokenizer function, but i´ts totally possible to replace the specific code to use AutoTokenizer, AutoConfig and AutoModel instead.

After work, I’ll take a look on this.

onlurking avatar Apr 19 '21 16:04 onlurking

@paulbricman done!

onlurking avatar Apr 20 '21 01:04 onlurking