-
Notifications
You must be signed in to change notification settings - Fork 24.3k
Add noncontiguous OpInfo tests for MPS #142202
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
Add noncontiguous OpInfo tests for MPS #142202
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/142202
Note: Links to docs will display an error until the docs builds have been completed. ❌ 3 New FailuresAs of commit d9f9e22 with merge base 7eb51e5 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
✅ Deploy Preview for chimerical-cranachan-793287 ready!
To edit notification comments on pull requests, go to your Netlify site configuration. |
3933c71
to
a274133
Compare
4428097
to
cec0350
Compare
@albanD Could I get an override for the LOC limit in the PR sanity checks? I'm not sure how to break down the OpInfo implementation any further than I already have. Thanks! |
3d884b7
to
e1daa64
Compare
Howdy! I'd probably say for any change that's this big we would probably say no to bypassing the LOC limit. Is there absolutely no way of breaking up this change? |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
We should find a way to do this that does not require a 25k+ LOC change
@seemethere the change got this big because I kept having to reformat common_method_invocations.py to pass the lintrunner. Perhaps I had my linter configured wrong, but it seems like common_method_invocations.py was already out of compliance so I ended up having to reformat much of the document. It is mostly just indentation changes though One alternative could be to exempt the OpInfo definitions from the linter, is that possible? |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Do you mind breaking it down into several smaller PRs, and may be have some sort of a summary of how many ops are passing right now vs how many are failing.
From a reviewer side, it looks like thousands of xfail-if-mps
are being added to common_methods_invocation, which IMO too verbose and isn't helpful.
- If one needs to xfail all
test_non_standard_bool
it should be done viafor op in ops: ops.skips.append(xfail-non-standard-bool)
- If majority of non-contiguous ops fail for MPS, again one can codegen it with
MPS_NON_CONTIG_PASS_LIST={add, sum}
for op in op_db:
if op.name not in MPS_NON_CONTIG_PASS_LIST:
op.skip.append(xfail-if-non-contig-on-mps)
7096e93
to
94d1a88
Compare
@pytorchbot rebase |
@pytorchbot started a rebase job onto refs/remotes/origin/viable/strict. Check the current status here |
Successfully rebased |
bbe2916
to
d9f9e22
Compare
Do not merge, I'll be rebasing this on top of a separate PR where the mps op db will be implemented in a more coherent way |
Looks like this PR hasn't been updated in a while so we're going to go ahead and mark this as |
PR to enable noncontiguous_sample OpInfo tests for MPS
This should help avoid regressions as seen in #141836
Will follow up with additional PRs to enable remaining OpInfo tests. It may not be feasible to review and merge the change if I try enabling them all at once, due to the size of the diff