perf: optimize MXFP4xBF16 & INT4xFP8 CUTLASS MoE backend for SM90#3084

Merged

samuellees merged 33 commits into

flashinfer-ai:mainfrom

samuellees:feat/w4a16-moe-kernel

Apr 23, 2026

Commits on Apr 15, 2026

feat: optimize W4A16 (wMXFP4 x BF16) MoE kernel from TRTLLM PR #12451

samuellees
and
claude
committed

Commits on Apr 17, 2026

test: add W4A16 MoE benchmark script

samuellees
and
claude
committed

Commits on Apr 19, 2026

fix: use percent-based accuracy check matching TRTLLM methodology

samuellees
and
claude
committed
fix: revert test scope to h=128 correctness (matches upstream W4A16 coverage)
samuellees
committed
style: apply pre-commit hooks (clang-format, ruff)
samuellees
committed
feat: port TRTLLM PR #12451 mixed-input CUTLASS extensions
samuellees
committed
feat: wire up W4A16 weight + scale interleave for SM90 mixed-input MoE
samuellees
committed
test: extend W4A16 MoE correctness to K=768/2048/4096 with interleave preprocessing
samuellees
committed
test: use 99.9% percent-based check for W4A16 coverage sweep
samuellees
committed
test: speed up W4A16 reference, trim coverage configs, preprocess bench weights
samuellees
committed
test: apply W4A16 interleave preprocessing in upstream MoE test
samuellees
committed
style: apply pre-commit hooks (clang-format, ruff format)
samuellees
committed
feat: align W4A16 heuristic with TRTLLM PR #12451
samuellees
committed

Commits on Apr 20, 2026

Commits on Apr 21, 2026

Commits on Apr 22, 2026