Why custom modeling code?

#1
by nickhugs - opened

The modeling_codegen.py and configuration_codegen.py in this repo appear to just be older versions of those files from the transformers library. In particular they don't include recent fixes to position_ids support (for handling variable left-padding properly).

Can this section be removed from config.json (and the py files can then be removed too)?

  "auto_map": {
    "AutoConfig": "configuration_codegen.CodeGenConfig",
    "AutoModel": "modeling_codegen.CodeGenModel",
    "AutoModelForCausalLM": "modeling_codegen.CodeGenForCausalLM"
  },

I saw some difference in mp_num = 4, but the provision of the custom code is confusing too me as well.

Sign up or log in to comment