upstream sync#2

Open

LucasWilkinson wants to merge 28 commits intovllm-project:main-archivefrom

deepseek-ai:main

Commits on Apr 28, 2025

Fix synchronization issues
beginlner
committed

Commits on Apr 29, 2025

Commits on Aug 1, 2025

Add more GPU architctures support (#76)

uchihatmtkinu
and
dianzhangchen
authored

Commits on Aug 14, 2025

Drop support for CUDA <12.8
Jiashi Li
committed
Fix accuracy issue in sum_OdO kernel
Jiashi Li
committed

Commits on Aug 25, 2025

Remove tma padding for fwd inputs (#85)
dianzhangchen
authored
Remove cudaMalloc and cudaFree in backward (#87)

SeanLi-OI
and
Jiashi Li
authored

Commits on Aug 27, 2025

fix calc space bug (#91)
uchihatmtkinu
authored

Commits on Sep 22, 2025

Refine handling for q/v sequence length equals zero. (#92)
dianzhangchen
authored

Commits on Sep 24, 2025

Commits on Sep 30, 2025

Update blog and README
interestingLSY
committed
Fix error message
Jiashi Li
committed
Code format
Jiashi Li
committed
Update README
Jiashi Li
committed

Commits on Jan 16, 2026

Multiple updates and refactorings (#150)
interestingLSY
committed

Commits on Jan 19, 2026

Add missing include<span>

Jiashi Li
and
alibaba-miji
committed

Commits on Jan 20, 2026

nits
interestingLSY
committed

Commits on Feb 6, 2026

Add CUDAGuard and device id assignment in sm100 dense fmha (#160)
uchihatmtkinu
authored

Commits on Mar 31, 2026

Change the order of grid dim in bwd convert kernel to avoid overlimit when sequence length is very large(>1M) (#173)
uchihatmtkinu
authored

Commits on Apr 29, 2026

Swap FlashMLA combine grid dimensions (#182)
PerkzZheng
authored