[sm100] Packed FP4×FP4 mega-MoE kernel (W4A4) with per-band BLOCK_K#44

Open

Romaosir wants to merge 22 commits into

sgl-project:devfrom

Romaosir:w4a4-packed-fp4-mega-moe-dev

Commits on Apr 25, 2026

support tvm ffi interfaces

Fridge003
and
rainj-me
committed

Commits on May 2, 2026

rebase on 0426 upstream

rainj-me
authored and
Fridge003
committed
Relax timeout to 180s
Fridge003
committed
pin tvm to 0.1.9
Fridge003
committed

Commits on May 4, 2026

Support wheel compilation (#26)
Fridge003
authored

Commits on May 14, 2026

update readme
Fridge003
committed

Commits on May 20, 2026

Export the PDL utils of DeepGEMM (#34)
b8zhong
authored

Commits on May 29, 2026

Expose BF16 grouped GEMM wrappers

popsiclexu
authored and
Fridge003
committed
Fix version of dev branch to 0.0.0
Fridge003
committed

Commits on Jun 1, 2026

Fix IMA guard in paged MQA logits scheduler (#38)

nvjullin
and
liji-nv
authored

Commits on Jun 9, 2026

Fix various issues in DeepGEMM tests (#39)

authored

Commits on Jun 10, 2026

Change license in pyproject.toml to avoid build and publish failures with setuptools>=77

b8zhong
and
Brayden Zhong
authored

Commits on Jun 11, 2026

Move run tests script for DeepGEMM (#42)
b8zhong
authored
sm100: add packed FP4xFP4 mega-MoE kernel (W4A4) with per-band BLOCK_K

Codex Agent
and
claude
committed
mega-moe: packed-FP4 wrapper, env dispatch, and off-by-default A/B scaffolds

Codex Agent
and
claude
committed
tests: packed-FP4 mega-MoE coverage, shape-sweep bench, A/B harness

Codex Agent
and
claude
committed