Skip to content

[BUG] Ascend: the flash-attention correctness test uses the same full-size shapes as the benchmark #4380

Open

Description

@cn-sys

Required prerequisites

What version of TileLang are you using?

0.1.15+cuda.git210611d

System information

Ascend 950PR, CANN 9.0.0-beta.2, Linux 6.6.0-13, Python 3.12. Source build of main @ 210611d.

Problem description

// examples/ascend/README.md:45-46
Each top-level example_<name>.py has one test_<name>.py. Correctness tests use
smaller inputs than performance runs.

The two forward tests use the performance shapes verbatim:

// flash_attention/test_mha.py:10-11     SEQ_LEN = 4096                 (= example_mha.py:16 default)
// flash_attention/test_gqa.py:10        S1, G, S2, D = 8192, 32, 8192, 128   (= example_gqa.py perf)

test_gqa_bwd.py:8 and all five top-level tests do use reduced shapes, so the convention is real and only these two break it. The consequence is practical: pytest examples/ascend spends almost all its time in these two tests, and a "correctness" run costs as much as a benchmark.

Activity

  1. cn-sys commented on Nov 15, 2026

    @cn-sys
    Author

    /formal

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

    ascendbugSomething isn't working

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions