[TRTLLM-14054][perf] Reduce per-step host preparation overhead in the PyTorch executor decode path#16313

kaiyux merged 5 commits into

NVIDIA:mainNVIDIA/TensorRT-LLM:mainfrom

kaiyux:perf/decode-step-host-prep-20260713kaiyux/TensorRT-LLM:perf/decode-step-host-prep-20260713Copy head branch name to clipboard