microsoft

/

phi-1_5

like 1.28k

isn't it the original model?
model = AutoModelForCausalLM.from_pretrained("microsoft/phi-1_5")
tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-1_5")

correct, but I don't know how to save it into the tokenizer.model file required by the convert.py script.

correct, but I don't know how to save it into the tokenizer.model file required by the convert.py script.

This is not a llama model, so you cannot use llama.cpp or ggml to convert it yet. You will have to add support for MixFormerSequentialForCausalLM model type in the ggml library first.

Thanks a lot for this information and sorry for the newbie question. I appreciate any link or tutorial for doing so. A Google search led me to this page https://github.com/guidance-ai/guidance/issues/58 but I could not get much out of it.

Thanks a lot for this information and sorry for the newbie question. I appreciate any link or tutorial for doing so. A Google search led me to this page https://github.com/guidance-ai/guidance/issues/58 but I could not get much out of it.

It won't be easy porting the model to GGML, but not impossible. You can take a look at how GPT-J is implemented here (refer to the HF implementation here), then try and adapt the Phi model the same way. The modeling code for phi is in this repo.