-
Notifications
You must be signed in to change notification settings - Fork 21.3k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
DISABLED test_non_contiguous_input_addmm (__main__.TestMaxAutotune) #126176
Comments
Hello there! From the DISABLED prefix in this issue title, it looks like you are attempting to disable a test in PyTorch CI. The information I have parsed is below:
Within ~15 minutes, To modify the platforms list, please include a line in the issue body, like below. The default action will disable the test for all platforms if no platforms list is specified.
We currently support the following platforms: asan, dynamo, inductor, linux, mac, macos, rocm, slow, win, windows. |
This reliably reproduces for me when I run test_max_autotune.py, but I'm not able to detect-test-pollution bisect it. It's possible a smarter bisection algorithm could figure it out. |
Resolving the issue because the test is not flaky anymore after 950 reruns without any failures and the issue hasn't been updated in 14 days. Please reopen the issue to re-disable the test if you think this is a false positive |
Another case of trunk flakiness has been found here. Reopening issue. The list of platforms [linux, rocm, slow] appears to contain all the recently affected platforms [linux, rocm]. |
cc @shunting314 - maybe the atol/rtol is too low for mms |
I'm not able to repro on an A100:
|
hmm, running another test first make it repro-able...
|
…tune.py" Fix #126176 . We should not use torch.empty to generate input data if we are gonna do any accuracy test. torch.empty may return NaN. In that cause both the reference and the actual result may contain NaN at the same index. But `NaN != NaN` so the test fail. Also if torch.empty returns NaN is not deterministic. It may depends on other tests running earlier. Generating random data instead of calling torch.empty fixes the problem. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
Fix #126176 . We should not use torch.empty to generate input data if we are gonna do any accuracy test. torch.empty may return NaN. In that cause both the reference and the actual result may contain NaN at the same index. But `NaN != NaN` so the test fail. Also if torch.empty returns NaN is not deterministic. It may depends on other tests running earlier. Generating random data instead of calling torch.empty fixes the problem. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
Platforms: linux, rocm, slow
This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs.
Over the past 3 hours, it has been determined flaky in 5 workflow(s) with 15 failures and 5 successes.
Debugging instructions (after clicking on the recent samples link):
DO NOT ASSUME THINGS ARE OKAY IF THE CI IS GREEN. We now shield flaky tests from developers so CI will thus be green but it will be harder to parse the logs.
To find relevant log snippets:
test_non_contiguous_input_addmm
Sample error message
Test file path:
inductor/test_max_autotune.py
cc @clee2000 @ezyang @msaroufim @bdhirsh @anijain2305 @chauhang @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @peterbell10 @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire
The text was updated successfully, but these errors were encountered: