Skip to content

--enforce-eager option slows down token generation #5

Description

@mingxin-zheng

Description

The current script suggests to use --enforce-eager in vllm server

However, this disables the cuda graph acceleration, and as a result, the token generation on A6000 (Holoscan IGX, arm64) is very slow (11.2 tokes/s)

Possible solution

We can add documentation about the flag.

Also, removing the flag will increase the speed 4x (45.9 tokens/s )

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions