[Perf][Quant][NVFP4] Prefer W4A4 linear kernels over weight-only ones on SM120/121 (#55170)

Signed-off-by: Stefano Castagnetta <scastagnetta@nvidia.com>
This commit is contained in:
Stefano Castagnetta
2026-09-08 15:06:45 +04:00
committed by GitHub
parent e41a17e606
commit 13cf9e05c1
2 changed files with 34 additions and 1 deletions
@@ -533,10 +533,11 @@ _POSSIBLE_MXFP8_KERNELS: dict[PlatformEnum, list[type[Mxfp8LinearKernel]]] = {
_POSSIBLE_NVFP4_KERNELS: dict[PlatformEnum, list[type[NvFp4LinearKernel]]] = {
PlatformEnum.CUDA: [
FlashInferCuteDslNvFp4LinearKernel,
FlashInferCuteDslNvFp4W4A16LinearKernel,
FlashInferCutlassNvFp4LinearKernel,
FlashInferB12xNvFp4LinearKernel,
CutlassNvFp4LinearKernel,
# Weight-only: a fallback on a W4A4-capable checkpoint, not a preference.
FlashInferCuteDslNvFp4W4A16LinearKernel,
MarlinNvFp4LinearKernel,
FlashInferTrtllmNvFp4LinearKernel,
FlashInferCudnnNvFp4LinearKernel,