PaddlePaddle
[Feature] add Blackwell arch support for fused moe
PR Category
Inference
PR Types
Bug fixes
Description
This PR adds Blackwell support for the Fused MoE operator. All tests in test_weight_only_linear.py pass except for the backward case, which is due to some existing code-structure issues.
- Added SM100 entry logic.
- Removed unused
USE_FPAINTB_GEMM_WITH_SM90 - Unified the SM80–SM110 cases by appending them all to the SM80 path
- Added backward-path entry points for both
SM90 and SM100
End-to-End Model Validation
Before the Fix
Q: 1+1=?
A: APPLICATION direct relief den- RelationAPPLICATIONrelief moi wet
After the Fix
Q: 1+1=?
A: “In mathematics, when you add 1 and 1 together, the result is 2. That is, 1 + 1 = 2.”
WINT4 and WINT8 outputs look good after the fix, and the BF16 path works correctly as well.
你的PR提交成功，感谢你对开源项目的贡献! 请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。 Your PR has been submitted. Thanks for your contribution! Please wait for the result of CI firstly. See Paddle CI Manual for details.
![paddle-bot[bot] avatar](https://avatars.githubusercontent.com/in/216376?v=4 "Published by a pub.dev verified publisher"){kind=small}
Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.
:white_check_mark: Echo-Nie
:x: luoyi
luoyi seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
You have signed the CLA already but the status is still pending? Let us recheck it.
