[training_utils] feat: use TMA to load Tiles in linear_cross_entropy kernels - #4576
Merged
wuxibin89 merged 3 commits intoDec 22, 2025
Merged
Conversation
|
|
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Code Review
This pull request introduces the use of Tensor Memory Access (TMA) in Triton kernels to improve performance on Hopper GPUs, which is a great enhancement. The implementation correctly checks for TMA support and adds separate code paths for it. However, this has led to significant code duplication within several Triton kernels (efficient_entropy_kernel_general_mainloop, efficient_entropy_backward_kernel_general_mainloop_MN, efficient_entropy_backward_kernel_general_d_logits, and efficient_entropy_backward_kernel_general_d_logits_split_N). I've provided suggestions to refactor these kernels to remove duplication. Since the conditional logic is based on a compile-time constant (USE_TMA), these changes should improve maintainability and readability without impacting performance.
CtfGo
force-pushed
the
linear_cross_entropy_with_tma
branch
from
December 17, 2025 14:04
cca9cbd to
28d7c76
Compare
wuxibin89
approved these changes
Dec 22, 2025
Collaborator
|
#4676 fixes the support for triton >= 3.5.0 |
jsfanfanfan
pushed a commit
to meituan-search/verl
that referenced
this pull request
Jan 9, 2026
…kernels (verl-project#4576) ### What does this PR do? The triton kernels of linear_cross_entropy iteratively load `hidden` and `weight` block tile by calculating data address explicitly at present, which shows poor performance on Hopper GPU. This PR improve that with the TMA of Hopper to avoid heavy address calculation. ### Test ``` python3 tests/utils/test_linear_cross_entropy.py ``` Latency(ms) improvement of test cases in`test_linear_cross_entropy.py` on H20 |case |linear_cross_entropy Forward| linear_cross_entropy Forward with TMA|linear_cross_entropy Backward| linear_cross_entropy Backward with TMA| |---|---|----|---|----| |0|129.64|18.83|157.85|56.14| |1|35.35|6.02|49.91|22.10| |2|12.28|2.05|16.41|6.63| |3|71.36|10.29|89.49|31.32| |4|405.92|56.49|477.97|173.43| ### Design & Code Changes add `USE_TMA` parameter and use `tl.make_tensor_descriptor` to create desc of TMA in triton kernels if it is enabled
vyomakesh0728
added a commit
to vyomakesh0728/verl
that referenced
this pull request
Jan 22, 2026
…kernels (verl-project#4576) ### What does this PR do? The triton kernels of linear_cross_entropy iteratively load `hidden` and `weight` block tile by calculating data address explicitly at present, which shows poor performance on Hopper GPU. This PR improve that with the TMA of Hopper to avoid heavy address calculation. ### Test ``` python3 tests/utils/test_linear_cross_entropy.py ``` Latency(ms) improvement of test cases in`test_linear_cross_entropy.py` on H20 |case |linear_cross_entropy Forward| linear_cross_entropy Forward with TMA|linear_cross_entropy Backward| linear_cross_entropy Backward with TMA| |---|---|----|---|----| |0|129.64|18.83|157.85|56.14| |1|35.35|6.02|49.91|22.10| |2|12.28|2.05|16.41|6.63| |3|71.36|10.29|89.49|31.32| |4|405.92|56.49|477.97|173.43| ### Design & Code Changes add `USE_TMA` parameter and use `tl.make_tensor_descriptor` to create desc of TMA in triton kernels if it is enabled
sophiayyya
pushed a commit
to sophiayyya/verl
that referenced
this pull request
Jan 25, 2026
…kernels (verl-project#4576) ### What does this PR do? The triton kernels of linear_cross_entropy iteratively load `hidden` and `weight` block tile by calculating data address explicitly at present, which shows poor performance on Hopper GPU. This PR improve that with the TMA of Hopper to avoid heavy address calculation. ### Test ``` python3 tests/utils/test_linear_cross_entropy.py ``` Latency(ms) improvement of test cases in`test_linear_cross_entropy.py` on H20 |case |linear_cross_entropy Forward| linear_cross_entropy Forward with TMA|linear_cross_entropy Backward| linear_cross_entropy Backward with TMA| |---|---|----|---|----| |0|129.64|18.83|157.85|56.14| |1|35.35|6.02|49.91|22.10| |2|12.28|2.05|16.41|6.63| |3|71.36|10.29|89.49|31.32| |4|405.92|56.49|477.97|173.43| ### Design & Code Changes add `USE_TMA` parameter and use `tl.make_tensor_descriptor` to create desc of TMA in triton kernels if it is enabled
8 tasks
DaizeDong
pushed a commit
to DaizeDong/verl
that referenced
this pull request
Apr 19, 2026
…kernels (verl-project#4576) ### What does this PR do? The triton kernels of linear_cross_entropy iteratively load `hidden` and `weight` block tile by calculating data address explicitly at present, which shows poor performance on Hopper GPU. This PR improve that with the TMA of Hopper to avoid heavy address calculation. ### Test ``` python3 tests/utils/test_linear_cross_entropy.py ``` Latency(ms) improvement of test cases in`test_linear_cross_entropy.py` on H20 |case |linear_cross_entropy Forward| linear_cross_entropy Forward with TMA|linear_cross_entropy Backward| linear_cross_entropy Backward with TMA| |---|---|----|---|----| |0|129.64|18.83|157.85|56.14| |1|35.35|6.02|49.91|22.10| |2|12.28|2.05|16.41|6.63| |3|71.36|10.29|89.49|31.32| |4|405.92|56.49|477.97|173.43| ### Design & Code Changes add `USE_TMA` parameter and use `tl.make_tensor_descriptor` to create desc of TMA in triton kernels if it is enabled
zwluestc
pushed a commit
to zwluestc/verl
that referenced
this pull request
May 12, 2026
…kernels (verl-project#4576) ### What does this PR do? The triton kernels of linear_cross_entropy iteratively load `hidden` and `weight` block tile by calculating data address explicitly at present, which shows poor performance on Hopper GPU. This PR improve that with the TMA of Hopper to avoid heavy address calculation. ### Test ``` python3 tests/utils/test_linear_cross_entropy.py ``` Latency(ms) improvement of test cases in`test_linear_cross_entropy.py` on H20 |case |linear_cross_entropy Forward| linear_cross_entropy Forward with TMA|linear_cross_entropy Backward| linear_cross_entropy Backward with TMA| |---|---|----|---|----| |0|129.64|18.83|157.85|56.14| |1|35.35|6.02|49.91|22.10| |2|12.28|2.05|16.41|6.63| |3|71.36|10.29|89.49|31.32| |4|405.92|56.49|477.97|173.43| ### Design & Code Changes add `USE_TMA` parameter and use `tl.make_tensor_descriptor` to create desc of TMA in triton kernels if it is enabled
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
4 participants
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.
What does this PR do?
The triton kernels of linear_cross_entropy iteratively load
hiddenandweightblock tile by calculating data address explicitly at present, which shows poor performance on Hopper GPU. This PR improve that with the TMA of Hopper to avoid heavy address calculation.Test
Latency(ms) improvement of test cases in
test_linear_cross_entropy.pyon H20Design & Code Changes
add
USE_TMAparameter and usetl.make_tensor_descriptorto create desc of TMA in triton kernels if it is enabled