Skip to content

Scan + flash attention kernel = NaN #8742

@tengyifei

Description

@tengyifei

🐛 Bug

When scanning over modules that contain flash attention pallas kernels, the output contains NaN.

To Reproduce

Run the test in 1. This test will fail and the output of fake_fa_wrapper contains NaN when use_scan=True

Expected behavior

The output of fake_fa_wrapper should be the same regardless if use_scan=True or False.

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't workingpallas

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions