Use matmul fwd direclty in autograd for performance #1045
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
meta-cla
bot
added
the
CLA Signed
tianrengao
force-pushed
the
tianren/addmm_bwd_fix_impl
branch
from
7301838 to
ad73fba
Compare
tianrengao
changed the title
jansel
requested changes
Nov 7, 2025
Comment on lines
+97
to
+101
| # grad_mat1 = grad_out @ mat2.T | ||
| grad_mat1 = matmul(grad_out, mat2.T) | ||
|
|
||
| # grad_mat2 = mat1.T @ grad_out | ||
| grad_mat2 = matmul(mat1.T, grad_out) |
Contributor
There was a problem hiding this comment.
You only need to compute these if requires_grad is set on the inputs.
Comment on lines
+158
to
+165
| # grad_bias = beta * grad_out | ||
| grad_bias = beta * grad_out | ||
|
|
||
| # grad_mat1 = alpha * (grad_out @ mat2.T) | ||
| grad_mat1 = alpha * matmul(grad_out, mat2.T) | ||
|
|
||
| # grad_mat2 = alpha * (mat1.T @ grad_out) | ||
| grad_mat2 = alpha * matmul(mat1.T, grad_out) |
Contributor
There was a problem hiding this comment.
This results in extra kernels, you should define an epilogue function to put the scaling into the matmul kernel.
Also same issue as above.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
Resolving previous matmul bwd implementation for performance concern. In previous PR #748, the matmul bwd was implemented with a specific kernel via two passes, while we can directly call matmul fwd twice given matmul fwd is fully optimized, as @ngimel pointed out. In this PR the matmul_autograd and addmm_autograd are updated to use two matmul fwds instead of a specific matmul_bwd and addmm_bwd. The benchmark/run.py now only calls these updated bwd.
However, the
@helion.kernelannotation does not allow calling another function(matmul_fwd) within the kernel def, so the original matmul_bwd and addmm_bwd are still preserved only as examples in examples/matmul.py, but they are not actually used in benchmark run.