Support for Salesforce/CodeGen about deepspeed-mii

[mrwyattii]

Hi @andrePankraz , you are correct in that we don't have these models in our SUPPORTED_MODEL_TYPES here

However, I did just do a test with a locally modified MII that skips this check and the model you linked to does work. @jeffra @RezaYazdaniAminabadi any reason not to add the codegen models to our supported models? At the very least, adding an option to skip MII compatibility checks may be worthwhile.

[jeffra]

From a kernel acceleration perspective we support GPT-J, I recall seeing there’s a way to covert this codegen model to a GPT-J model. In short term you might try that? However, medium term I think it’s possible for us to adapt our injection policy to support this model. We’ll need to investigate this further though.

https://twitter.com/moyix/status/1551608934091341828?s=46&t=00FUkiOvSQ9s7sPwtFbDwA

if you don’t need kernel acceleration you can turn off the model checking via a new Mii config option: skip_model_check: true

DeepSpeed-MII/mii/config.py

Line 17 in e6b4847

skip_model_check: bool = False