🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/96264

• 📄 Preview Python docs built from this PR
• 📄 Preview C++ docs built from this PR
• ❓ Need help or want to give feedback on the CI? Visit the bot commands wiki or our office hours

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit fb23a09:
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

@pytorchbot label "topic: not user facing"

@ngimel could you please review this again?
I somehow messed up merging the previous one
(again, sorry about that: pretty new to open source..)
Thanks a lot

Test failures are real 🤔

Hm, it does @pw_cast_for_opmath, I don't understand accuracy failures.

I might need some help here :(
The accuracy error does not show up on replacing torch.mv with @/matmul/(A*b).sum(dim=1), but I couldn't understand why..

out = alpha * torch.mv(mat1, vec)
return out + beta * self # fails
----
out = alpha * torch.matmul(mat1, vec)
return out + beta * self # success

And I saw this but removing it doesn't affect any mv related tests, so idk why/when did mv itself faced accuracy issues, or if this changes addmv's behaviour(but why would printing something random cause problem?)

Also, I couldn't understand the test_has_decomposition error: shouldn't @register_decomposition add it to the table?

I guess you need to add higher tolerance for addmv to that table

Is the precision difference due to FMA or something?

blas gemv likely compiles to fma, while out + beta * self doesn't

I wonder if we should have an FMA prim for this reason 🤔

@ngimel I just have one more question for clarity: matmul just forwards the call to torch.mv (in this case). So shouldn't matmul also use the same rounding? Why would matmul pass and not mv if they both call the same function? The print thing also went over my head
Thanks a lot

Yes, matmul should call mv. Sorry, I don't quite understand what fails and what passes from your comments, can you post a standalone runnable script demonstrating the differences between matmul and mv?

There isn't any difference in the result but somehow, matmul passes the accuracy test on my system.. ( instead of using torch.mv(mat1, vec) )

Sorry, both things can't be true

1. there's no difference in the results with matmul and mv
2. matmul passes the test and mv doesn't.
   Can you please dig deeper and see what's going on?
   Or leave mv in decomposition and increase tolerance for the tests.

@ngimel
Whenever I go through matmul, both orig and decomp(in DecompCrossRefMode) have low accuracy(~1e-2 to ~1e-4), and mv's test passes(holds for all seeds).. Not sure what is happening yet. If there is something wrong, I'll add an issue separately.
Error in mv's test is due to (self * vec) not using FMA in its decomp(increased its tolerance). No need to increase addmv's tolerance.
I hope this is fine. Am I allowed to merge this using pytorchbot? Or do I need all reviews to do so?
Thanks

I don't understand how orig can have different accuracy depending on decomposition, but you can land PR in its current form if the tests pass.

ahh.. Could you please run again? Sorry for the trouble

@pytorchbot merge

Merge started

Your change will be merged once all checks pass (ETA 0-4 Hours).

Learn more about merging in the wiki.

Questions? Feedback? Please reach out to the PyTorch DevX Team

Merge failed

Reason: 1 mandatory check(s) failed. The first few are:

• pull / win-vs2019-cpu-py3 / test (default, 2, 2, windows.4xlarge)

Dig deeper by viewing the failures on hud

@pytorchbot merge

Merge started

Your change will be merged once all checks pass (ETA 0-4 Hours).

Learn more about merging in the wiki.

Questions? Feedback? Please reach out to the PyTorch DevX Team