NVIDIA's TensorRT-LLM cuts DeepSeek V4 tuning time 70% in test NVIDIA merged pull request 21185 into TensorRT-LLM main on November 16th, 2026, adding an optional nvMatmulHeuristics policy that cut DeepSeek V4 Pro cold autotuning from 226.6 seconds to 68.2 seconds in contributor-reported tests. Across 48 FP8 BMM and MXFP8 GEMM shapes, the heuristic path profiled 838 tactics instead of 82,160 with the full sweep, a roughly 98x reduction, and server start-to-ready time fell from 514 seconds to 350 seconds on two nodes with four VR200 GPUs each. The feature stays off by default. NVIDIA's TensorRT-LLM cuts DeepSeek V4 tuning time 70% in test The merged TensorRT-LLM policy profiled 838 tactics instead of 82,160 on DeepSeek V4 Pro, and the feature stays off by default. By RuntimeWire Staff https://runtimewire.com/author/runtimewire-staff · Published NVIDIA merged pull request 21185 into TensorRT-LLM main on November 16th, 2026, adding an optional nvMatmulHeuristics policy that cut DeepSeek V4 Pro cold autotuning from 226.6 seconds to 68.2 seconds in contributor-reported tests. Across 48 FP8 BMM and MXFP8 GEMM shapes, the heuristic path profiled 838 tactics, down from 82,160 with the full sweep, a reduction of about 98 times. Server start-to-ready time fell from 514 seconds to 350 seconds on two nodes with four VR200 GPUs each.…