[None][perf] Eliminate ~6s FMHA JIT recompile in eager generation by aligning kernel selection with CUDA graph warmup - #13312

[None][perf] Eliminate ~6s FMHA JIT recompile in eager generation by aligning kernel selection with CUDA graph warmup#13312

yunruis wants to merge 3 commits into

NVIDIA:mainNVIDIA/TensorRT-LLM:mainfrom

yunruis:user/yunruis/fmha_narrow_pathsel_cache_missyunruis/TensorRT-LLM:user/yunruis/fmha_narrow_pathsel_cache_missCopy head branch name to clipboard