Add backward pass for softmax kernel #744
Open
+192
−3
Conversation
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
karthickai
added a commit
that referenced
this pull request
Oct 1, 2025
stack-info: PR: #744, branch: karthickai/stack/4
karthickai
force-pushed
the
karthickai/stack/4
branch
from
656fba4 to
3e31a31
Compare
meta-cla
bot
added
the
CLA Signed
oulgen
reviewed
Oct 1, 2025
examples/softmax.py
Outdated
| m, n = grad_output.size() | ||
| grad_input = torch.empty_like(grad_output) | ||
|
|
||
| block_size_m = hl.register_block_size(m) |
There was a problem hiding this comment.
Yeah it is not need, I followed the implementation from softmax two pass, now I've updated the code.
karthickai
added a commit
that referenced
this pull request
Oct 1, 2025
stack-info: PR: #744, branch: karthickai/stack/4
karthickai
force-pushed
the
karthickai/stack/4
branch
from
3e31a31 to
d5f0d11
Compare
74 tasks
yf225
reviewed
Oct 7, 2025
| x: torch.Tensor, | ||
| ) -> torch.Tensor: | ||
| """Softmax with forward + backward support.""" | ||
| return SoftmaxFunction.apply(x) # type: ignore[no-any-return] |
There was a problem hiding this comment.
Would you like to also add integration in benchmarks/run.py (similar to rms_norm-bwd), and test accuracy via tritonbench --metrics accuracy?
There was a problem hiding this comment.
backward pass for softmax is disabled in tritonbench. I enabled it in this PR meta-pytorch/tritonbench#528 and added softmax-bwd in benchmarks/run.py
There was a problem hiding this comment.
HELION_USE_DEFAULT_CONFIG=1 python benchmarks/run.py --op softmax-bwd --num-inputs 3 --metrics accuracy
(M, N) triton_softmax-accuracy quack-accuracy torch_compile_softmax-accuracy helion_softmax_tritonbench-accuracy
----------- ------------------------- ---------------- -------------------------------- -------------------------------------
(4096, 256) 1 1 1 1
(4096, 384) 1 1 1 1
(4096, 512) 1 1 1 1
average 1 1 1 1HELION_USE_DEFAULT_CONFIG=1 python benchmarks/run.py --op softmax --num-inputs 3 --metrics accuracy
(M, N) triton_softmax-accuracy quack-accuracy torch_compile_softmax-accuracy helion_softmax_tritonbench-accuracy
----------- ------------------------- ---------------- -------------------------------- -------------------------------------
(4096, 256) 1 1 1 1
(4096, 384) 1 1 1 1
(4096, 512) 1 1 1 1
average 1 1 1 1There was a problem hiding this comment.
update: tirtonbench softmax-bwd is landed meta-pytorch/tritonbench#528
karthickai
added a commit
that referenced
this pull request
Oct 8, 2025
stack-info: PR: #744, branch: karthickai/stack/4
karthickai
force-pushed
the
karthickai/stack/4
branch
from
d5f0d11 to
2c155fd
Compare
stack-info: PR: #744, branch: karthickai/stack/4
karthickai
force-pushed
the
karthickai/stack/4
branch
from
2c155fd to
62bbc0e
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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.
Stacked PRs:
Add backward pass for softmax kernel