Support for the GPT-NeoX family of models

[JD-The-65th]

I think support for the GPT-NeoX models would be nice for a project like this, or as a fork of this project.
Together AI's Open Source Chatbot is built off of NeoX, and the base Pythia family of models is looking pretty promising too, with newer models being built off of Pythia (Databricks' Dolly) AFAIK, the model architecture shouldn't be that far off from GPT-J, but take that with a pinch of salt.

[jploski]

FWIW, I tried converting https://huggingface.co/OpenAssistant/oasst-sft-1-pythia-12b to ggml today (using the commit 3173a62), but no luck:

!python3 llama.cpp/convert-pth-to-ggml.py oasst-sft-1-pythia-12b 1
Loading model file oasst-sft-1-pythia-12b/pytorch_model-00001-of-00003.bin
Traceback (most recent call last):
  File "/content/llama.cpp/convert-pth-to-ggml.py", line 11, in <module>
    convert.main(['--outtype', 'f16' if args.ftype == 1 else 'f32', '--', args.dir_model])
  File "/content/llama.cpp/convert.py", line 1128, in main
    model_plus = load_some_model(args.model)
  File "/content/llama.cpp/convert.py", line 1055, in load_some_model
    models_plus.append(lazy_load_file(path))
  File "/content/llama.cpp/convert.py", line 849, in lazy_load_file
    return lazy_load_torch_file(fp, path)
  File "/content/llama.cpp/convert.py", line 721, in lazy_load_torch_file
    model = unpickler.load()
  File "/content/llama.cpp/convert.py", line 710, in find_class
    return self.CLASSES[(module, name)]
KeyError: ('torch', 'ByteStorage')

[ggerganov]

Minimal GPT-NeoX inference example is now available in the ggml repo:

https://github.com/ggerganov/ggml/tree/master/examples/stablelm

[JD-The-65th]

Nice, thanks a lot man! This project is a really cool one, do you have any way I could financially support you? I'm thinking about giving some money to some cool open source stuff I like.