[ROCm][CI] Fix failing FP8 tests on RDNA4 (#174873)#3090
Merged
mstankov-amd merged 1 commit intorelease/2.11from Mar 20, 2026
Merged
[ROCm][CI] Fix failing FP8 tests on RDNA4 (#174873)#3090mstankov-amd merged 1 commit intorelease/2.11from
mstankov-amd merged 1 commit intorelease/2.11from
Conversation
Author
|
!cherry-pick --onto release/2.10 |
Author
|
!cherry-pick --onto release/2.9 |
rocm-repo-management-api-6 bot
pushed a commit
that referenced
this pull request
Mar 20, 2026
## Summary This PR fixes FP8 inductor test failures that occur on AMD RDNA4 GPUs when testing matrix multiplications with small M dimensions (M < 16). ## Problem On gfx120x GPUs, FP8 scaled matrix multiplication tests fail with: - 92.4% NaN outputs when M < BLOCK_M (typically 16) - Large numerical mismatches between eager and compiled results - Only occurs in `max-autotune` mode **Root cause:** Autotuned Triton kernels on gfx120x generate incorrect tensor indexing for small M values, using partial indices instead of full computed indices in load/store operations. ## Solution - Added GPU-specific compile mode selection for small M values - gfx120x with M < 16: use `compile_mode="default"` - All other cases: use `compile_mode="max-autotune"` Pull Request resolved: pytorch#174873 Approved by: https://github.com/jeffdaily (cherry picked from commit d667ffe)
rocm-repo-management-api-6 bot
added a commit
that referenced
this pull request
Mar 20, 2026
Cherry-pick of #3090 Co-authored-by: Milica Stankovic <milica.stankovic@amd.com>
|
Created branch autogenerated/release/2.10_cherry-pick_pr-3090 and #3091 Comment processed by Build |
rocm-repo-management-api-6 bot
pushed a commit
that referenced
this pull request
Mar 20, 2026
## Summary This PR fixes FP8 inductor test failures that occur on AMD RDNA4 GPUs when testing matrix multiplications with small M dimensions (M < 16). ## Problem On gfx120x GPUs, FP8 scaled matrix multiplication tests fail with: - 92.4% NaN outputs when M < BLOCK_M (typically 16) - Large numerical mismatches between eager and compiled results - Only occurs in `max-autotune` mode **Root cause:** Autotuned Triton kernels on gfx120x generate incorrect tensor indexing for small M values, using partial indices instead of full computed indices in load/store operations. ## Solution - Added GPU-specific compile mode selection for small M values - gfx120x with M < 16: use `compile_mode="default"` - All other cases: use `compile_mode="max-autotune"` Pull Request resolved: pytorch#174873 Approved by: https://github.com/jeffdaily (cherry picked from commit d667ffe)
|
Created branch autogenerated/release/2.9_cherry-pick_pr-3090 and #3092 Comment processed by Build |
rocm-repo-management-api-6 bot
added a commit
that referenced
this pull request
Mar 20, 2026
Cherry-pick of #3090 Co-authored-by: Milica Stankovic <milica.stankovic@amd.com>
## Summary This PR fixes FP8 inductor test failures that occur on AMD RDNA4 GPUs when testing matrix multiplications with small M dimensions (M < 16). ## Problem On gfx120x GPUs, FP8 scaled matrix multiplication tests fail with: - 92.4% NaN outputs when M < BLOCK_M (typically 16) - Large numerical mismatches between eager and compiled results - Only occurs in `max-autotune` mode **Root cause:** Autotuned Triton kernels on gfx120x generate incorrect tensor indexing for small M values, using partial indices instead of full computed indices in load/store operations. ## Solution - Added GPU-specific compile mode selection for small M values - gfx120x with M < 16: use `compile_mode="default"` - All other cases: use `compile_mode="max-autotune"` Pull Request resolved: pytorch#174873 Approved by: https://github.com/jeffdaily (cherry picked from commit d667ffe)
jithunnair-amd
pushed a commit
that referenced
this pull request
Mar 25, 2026
## Summary This PR fixes FP8 inductor test failures that occur on AMD RDNA4 GPUs when testing matrix multiplications with small M dimensions (M < 16). ## Problem On gfx120x GPUs, FP8 scaled matrix multiplication tests fail with: - 92.4% NaN outputs when M < BLOCK_M (typically 16) - Large numerical mismatches between eager and compiled results - Only occurs in `max-autotune` mode **Root cause:** Autotuned Triton kernels on gfx120x generate incorrect tensor indexing for small M values, using partial indices instead of full computed indices in load/store operations. ## Solution - Added GPU-specific compile mode selection for small M values - gfx120x with M < 16: use `compile_mode="default"` - All other cases: use `compile_mode="max-autotune"` Pull Request resolved: pytorch#174873 Approved by: https://github.com/jeffdaily (cherry picked from commit d667ffe)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
This PR fixes FP8 inductor test failures that occur on AMD RDNA4 GPUs when testing matrix multiplications with small M dimensions (M < 16).
Problem
On gfx120x GPUs, FP8 scaled matrix multiplication tests fail with:
max-autotunemodeRoot cause: Autotuned Triton kernels on gfx120x generate incorrect tensor indexing for small M values, using partial indices instead of full computed indices in load/store operations.
Solution
compile_mode="default"compile_mode="max-autotune"Pull Request resolved: pytorch#174873
Approved by: https://github.com/jeffdaily
(cherry picked from commit d667ffe)
Cherry-picked to release/2.10 branch via #3091
Cherry-picked to release/2.9 branch via #3092