Optimize the Cuda Kernel performance of Paddle rms_norm #77098
Merged
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
|
你的PR提交成功，感谢你对开源项目的贡献! |
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## develop #77098 +/- ##
===========================================
Coverage ? 100.00%
===========================================
Files ? 4
Lines ? 37
Branches ? 0
===========================================
Hits ? 37
Misses ? 0
Partials ? 0 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
zrr1999
force-pushed
the
opt_rms_norm
branch
2 times, most recently
from
0604b69 to
65aff3e
Compare
This reverts commit ccaaa1b.
zrr1999
force-pushed
the
opt_rms_norm
branch
from
65aff3e to
19513e8
Compare
This reverts commit 19513e8.
zhengshengning
changed the title
A-nnonymous
approved these changes
Jan 5, 2026
Contributor
A-nnonymous
left a comment
A-nnonymous
left a comment
There was a problem hiding this comment.
LGTM in kernel in most case, but need polish afterwards.
| } | ||
| inline __device__ res_t project(acc_t acc) const { | ||
| const auto mean = static_cast<scalar_t>(acc.mean); | ||
| const auto divisor = acc.nf > correction ? acc.nf - correction : 0; |
| if (threadIdx.x == 0) { | ||
| T_ACC m1; // mean | ||
| T_ACC m2; // var | ||
| thrust::pair<T_ACC, T_ACC> res = welford_op.project(val); |
Contributor
There was a problem hiding this comment.
在将来的修改中，尽量避免引入thrust相关的数据结构与算法，方便移植
| int64_t thread_x = static_cast<int64_t>(blockIdx.x) * block_dim_x + | ||
| static_cast<int64_t>(threadIdx.x); | ||
|
|
||
| int lane_id = (threadIdx.y * blockDim.x + threadIdx.x) & (kWarpSize - 1); |
Contributor
There was a problem hiding this comment.
lane_id有更高效的求取方式，优化时需要注意这部分开销
zrr1999
previously approved these changes
Jan 5, 2026
python/paddle/nn/functional/norm.py
Outdated
| def rms_norm( | ||
| input: Tensor, | ||
| normalized_shape: int | Sequence[int], | ||
| normalized_shape: list[int], |
zyfncg
reviewed
Jan 5, 2026
paddle/phi/ops/yaml/ops.yaml
Outdated
|
|
||
| - op: rms_norm | ||
| args: (Tensor x, Tensor scale, float epsilon) | ||
| args: (Tensor x, Tensor scale, IntArray normalized_shape={}, double epsilon= 1e-5) |
Contributor
There was a problem hiding this comment.
normalized_shape 如果 python api 明确是 int 类型，这里可以使用 int64_t[]
zhangbo9674
previously approved these changes
Jan 5, 2026
zyfncg
approved these changes
Jan 6, 2026
zrr1999
approved these changes
Jan 6, 2026
wanghuancoder
approved these changes
Jan 6, 2026
zhangbo9674
approved these changes
Jan 6, 2026
heavengate
approved these changes
Jan 6, 2026
zrr1999
pushed a commit
to zrr1999/Paddle
that referenced
this pull request
Jan 8, 2026
…#77098) * accuracy and Torch alignment * support rms_norm behavior to be the same as torch * fix rms_norm_xpu_kernel * add valueError_test * Revert "add valueError_test" This reverts commit ccaaa1b. * Reapply "add valueError_test" This reverts commit 19513e8. * optimize performance * add vectorization * fix * fix dtype of normalized_shape
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
9 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.
PR Category
Operator Mechanism
PR Types
Improvements
Description
优化rms_norm
a. 【完成】精度：已与torch逐位对齐（fp16、bf16、fp32、fp64）
b. 【完成】功能：原来Paddle rms_norm缺少normalized_shape参数，且不支持weight为空，已与Torch对齐
c. 【完成】性能：目前前向、反向基本与Torch持平
d. 【完成】性能2：API前处理可优化，kernel内可进行向量化加长、分支和融合等优化，目前性能全面优于torch
rms_norm前向：（优化前后，相比Paddle旧的实现平均提升2.2倍，相比Torch平均提升6%）
rms_norm反向：（优化前后，相比Paddle旧的实现平均提升1.6倍，相比Torch平均提升31%）