Skip to content

gpu_fast_attention not passing segment_ids to jax pallas attention mha #30

@Cjkkkk

Description

@Cjkkkk

q, k, v, sm_scale=1.0 / math.sqrt(h), backward_pass_impl=bwd_pass_impl

https://github.com/google/jax/blame/main/jax/experimental/pallas/ops/attention.py#L163
seems like jax has added segment_ids as required argument but praxis has not updated to add the argument

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