Skip to content

[BUG]: HybridParallelPlugin does not support customized wrapped LLaMAv2 model  #4613

@eric8607242

Description

@eric8607242

🐛 Describe the bug

Hi,
in my use case, I have to wrap LLaMAv2 with an additional class.

There are two issues when I try to apply HybridParallelPlugin on my customized model:

  1. HybridParallelPlugin can not assign the customized policy for ShardFormer, which makes HybridParallePlugin incompatible with the customized model.
  2. Policy LlamaPolicy lacks the attribute self_attn.num_key_value_heads, which is a new property for LLaMAv2.

Environment

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions