Previously for Phi models (both 1.5 and 2) in the HF repo there were modeling_*.py files and we had to download them and instantiate a reference class from the file (to tests correctness of the output between LitGPT and HF).
Now we can do this directly from transformers + for Phi-1.5 there is no longer modeling file available.
Hey there 👋
Previously for Phi models (both 1.5 and 2) in the HF repo there were
modeling_*.py
files and we had to download them and instantiate a reference class from the file (to tests correctness of the output between LitGPT and HF). Now we can do this directly fromtransformers
+ for Phi-1.5 there is no longer modeling file available.Fixes failing tests.