Skip to content

FSDP2 Support #32

Description

@garrett361

This issue tracks progress on running Bamba models with FSDP2.

WIP branch and PR here.

Top level goals:

  • FSDP2 + torch kernel impl
  • FSDP2 + mamba_ssm kernel impl
  • FSDP2 + TP
  • FSDP2 + CP
  • FSDP2 + FP8
  • FSDP2 + PP

Known Issues/Open Questions

  • Requires custom op registration for FSDP2 compatibility
  • DTensor currently only has minimal/experimental support for convolutions. Because of the depthwise convolution in the mamba layers, this is a blocker for TP/CP support. @garrett361 is working on more robust conv support.

cc @raghukiran1224

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