Skip to content

Conversation

@anmyachev
Copy link
Contributor

@anmyachev anmyachev commented Dec 2, 2025

This PR changes the Triton base from e4e68d1 to dcad270 (Nov 23).

Pass rate: 95.42%->95.86%

ptillet and others added 3 commits November 23, 2025 13:27
If building from root `CMakeLists.txt`, we will have an empty
`CMAKE_LIBRARY_OUTPUT_DIRECTORY`, then this will try
to create `/plugins` and see permission failures.
This PR fixes test failures in `test_dot_multidim` that occur on AMD
RDNA3 GPUs due to overly strict floating-point comparisons.
The test currently uses `torch.equal()` which requires exact bit-for-bit
equality, but different GPU architectures and drivers can produce
slightly different results at the machine epsilon level
while still being mathematically correct. This change replaces the exact
equality check with `torch.allclose()` using appropriate tolerances.
 
 
Triton --> tensor([[[[ -3.0000, -44.0000, 8.0000, ..., 1.0000, -28.0000,
-12.0000],
[ -6.0000, 46.0000, 10.0000, ..., -13.0000, 35.0000, 50.0000],
[ 83.0000, 5.0000, 59.0000, ..., -32.0000, -19.0000, 96.0000],
          ...,
		  
		  
 Torch ---> tensor([[[[ -3., -44.,   8.,  ...,   1., -28., -12.],
          [ -6.,  46.,  10.,  ..., -13.,  35.,  50.],
          [ 83.,   5.,  59.,  ..., -32., -19.,  96.],
          ...,


The printed values appear identical, but `torch.equal()` detects
differences at the least significant bits level.
Copy link
Contributor

@whitneywhtsang whitneywhtsang left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why 1479afd changes only test_matmul.py, but e33b846 changes more?

@anmyachev
Copy link
Contributor Author

Why 1479afd changes only test_matmul.py, but e33b846 changes more?

I took the commit via cherry-pick only to indicate the author. Will add all the changes necessary for CI to it so as not to make an additional commit (many commits during a merge are quite difficult to sort out later).

@anmyachev anmyachev marked this pull request as ready for review December 2, 2025 18:17
…t None and do_gamma is set

Signed-off-by: Witold Dziurdz <[email protected]>
(cherry picked from commit 1479afd)
Signed-off-by: Anatoly Myachev <[email protected]>
@whitneywhtsang whitneywhtsang merged commit e8efee9 into main Dec 3, 2025
25 checks passed
@whitneywhtsang whitneywhtsang deleted the amyachev/merge118 branch December 3, 2025 00:15
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

7 participants