MoonshotAI / MoonshotAI/FlashKDA

[Bug] test_fwd failed with CUDA 13

Open
#12 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Cuda
Stars
1.3k
Forks
122
PR merge metrics
No merged PRs in 30d

Description

env B200

test with CUDA12.9 ok
test with CUDA13.0 fail

root@localhost:/#python3 tests/test_fwd.py
Testing shape: [1, 8192, 96, 128]
67072.0 67072.0
output | avg_rtol: 3.68277053297561e-09, max_rtol: 4.6591722480116005e-07
output | avg_atol: 6.617795200725141e-09, max_atol: 0.03125
Traceback (most recent call last):
File "/home/users/qyp/flash-kda/tests/test_fwd.py", line 428, in
test_fwd()
File "/home/users/qyp/flash-kda/tests/test_fwd.py", line 260, in test_fwd
assert torch.equal(out_kernel, out_ref), "output mismatch between kernel and torch ref"
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AssertionError: output mismatch between kernel and torch ref

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Reproduce the failure on the reported B200 environment by running python3 tests/test_fwd.py with CUDA 13.0, then inspect tests/test_fwd.py around test_fwd and the assertion at line 260. Compare the kernel and torch reference outputs to identify why CUDA 13.0 produces a mismatch; done means the test passes with matching outputs under CUDA 13.0 while retaining the CUDA 12.9 result.

Written by the indexing model from the issue text.

Assessment

Domain
performance
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Quiet
Clarity
Mostly clear
Newbie friendliness
48/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.