[None][feat] Two-level GVR decode top-K dispatch; remove the CUDA heuristic and temporal-only prior state - #18446

[None][feat] Two-level GVR decode top-K dispatch; remove the CUDA heuristic and temporal-only prior state#18446

juney-nvidia merged 13 commits into

NVIDIA:mainNVIDIA/TensorRT-LLM:mainfrom

longcheng-nv:feat/gvr-unified-routerlongcheng-nv/TensorRT-LLM:feat/gvr-unified-routerCopy head branch name to clipboard