métier
non classé
description
Optimize kernels/fa4_blackwell.cu (CUDA-only, no CUTLASS/cuBLAS/cuDNN deps) to match then beat FA4/cuDNN/FlashInfer/TRT-LLM on user-selected Blackwell scenarios. Use when tuning FA4 parity, NCU-profiling attention, rewriting fa4_blackwell.cu, or running…
Langue du texte source : anglais
mis à jour