Add SM90 MegaMoE by cooperative - #360

Open

mpdfdfl wants to merge 16 commits into

deepseek-ai:mainfrom

mpdfdfl:sm90-mega-moe-pingpong-coop

Commits on Jun 12, 2026

sm90 megamoe: add pingpong and cooperative FP8 kernels

mpdfdfl
and
OpenDarrenlu
committed
006709a
sm90 megamoe: token-count dispatch and Python bindings

mpdfdfl
and
OpenDarrenlu
committed
4dd2e94
sm90 megamoe: correctness test and performance benchmark

mpdfdfl
and
OpenDarrenlu
committed
1f84636
chore: ignore profiling artifacts (*.ncu-rep, *.nsys-rep, profile_logs/)

mpdfdfl
and
OpenDarrenlu
committed
4cd92c4
fix mul2

mpdfdfl
and
OpenDarrenlu
committed
4146d0e

Commits on Jun 13, 2026

add sm90 megamoe bench

mpdfdfl
and
OpenDarrenlu
committed
3f16c98

Commits on Jun 17, 2026

fix accuracy

mpdfdfl
and
OpenDarrenlu
committed
d7c89bd
fix accuracy

mpdfdfl
and
OpenDarrenlu
committed
183fec5
add tma to HBM double buffer

mpdfdfl
and
OpenDarrenlu
committed
6b5be1d

Commits on Jun 25, 2026

fix accuract

mpdfdfl
and
OpenDarrenlu
committed
5fdbdda

Commits on Jul 31, 2026

rename sm90 megamoe kernel: drop cooperative suffix, converge API, clean comments

mpdfdfl
and
OpenDarrenlu
committed
0e245d6
sm90 megamoe: chunk the dispatch pull (kNumBytesPerPull <= 4096, aligns with SM100)

mpdfdfl
and
OpenDarrenlu
committed
e484200
sm90 megamoe: L2 GEMM pingpong (128-wide L2 tiles owned by alternating math WGs)

mpdfdfl
and
OpenDarrenlu
committed
89913a4
fix sm90 megamoe pingpong deadlock: non-owner must participate in full/empty

mpdfdfl
and
OpenDarrenlu
committed
950f3fd

Commits on Aug 3, 2026

sm90 megamoe: move all weight-SF reads into the B loader via per-stage SMEM SFB slots

mpdfdfl
and
OpenDarrenlu
committed
e5e88b4
sm90 megamoe: remove the L2 pingpong path, keep the per-stage SFB slots

mpdfdfl
and
OpenDarrenlu
committed
e481caf