DS4 H200: rocmfp2_mix MoE matvec 3.31x + ROCmFP4 dense matvec 1.53x (+42.6% decode) - #31

Open

DeanoC wants to merge 8 commits into

h200-servingfrom

ds4/h200-rocmfp2-mix-opt

Commits on Aug 18, 2026

perf(cuda): stage MoE activations once per block in rocmfp2_mix (1.98x)

DeanoC
and
claude
committed
87f165b
perf(cuda): issue all row loads before any FMA in rocmfp2_mix (1.057x)

DeanoC
and
claude
committed
a5587fe
perf(cuda): stage rocmfp2_mix activations in LDS, coalesced + swizzled (1.398x)

DeanoC
and
claude
committed
85dc3f5
perf(cuda): nibble-keyed, bank-replicated codebook table in rocmfp2_mix (1.135x)

DeanoC
and
claude
committed
0ee7bda
perf(cuda): static_assert the hand-computed nibble shift in rocmfp2_mix

DeanoC
and
claude
committed
683f75d
perf(cuda): table-free ROCmFP4 codebook expansion via prmt (1.399x, +8.9% decode)

DeanoC
and
claude
committed
e6036de
perf(cuda): aligned dword qs loads in ROCmFP4 vec_dot (1.023x kernel, no e2e claim)

DeanoC
and
claude
committed
84a8c6b
perf(cuda): split the ROCmFP4 small-K block into two row groups (1.526x)

DeanoC
and
claude
committed
25b9a75