TensorRT-LLM PR #18542: FMHA NVRTC JIT warmup sits on the cold-start critical path (5-11 s per kernel, 80-160 s total)

Opened 2026-09-01, still OPEN as of 2026-10-04. Vendor-authored and therefore authoritative about its own cost. Exact claims from the description: at engine startup, TRTLLM-Gen FMHA generation kernels are 'NVRTC-compiled synchronously inside the first warmup forward - ~5-11s per kernel variant depending on host CPU, 80-160s total on MLA fp8-KV configs (DeepSeek-family), ~2min for sparse gen kernels. This sits directly on the engine cold-start critical path.' The proposed opt-in fix TRTLLM_GEN_FMHA_ASYNC_WARMUP=1 overlaps the compile with autotuning and graph capture; internal B200 measurement: 'warmup reduction from 163 seconds to approximately 26-30 seconds'. Note this is a DIFFERENT compile from the C++ build: runtime NVRTC codegen of kernels for the specific shapes and configs being served, paid on every cold start of a new engine.