TensorRT-LLM issue #8286: '[Usage]: how long the build stage takes for deepseek-r1 models' - 2+ hours observed
forum
Fields
id
sources/ev-trtllm-issue-8286-ds-r1-build-2hPublisher
NVIDIA
Kind
forum
Accessed
2026-10-04
notes
Filed 2025-10-11, CLOSED. Conditions fully stated, which is why it is citable: Ubuntu, CUDA 12.8, b200, driver 570.133.20, TensorRT-LLM 1.0.0.rc6, following the DeepSeek-R1 quick-start recipe. The user's observation is that the process sits at high CPU ('looks cpus are busy' per htop) for 2+ hours and 'is this expected?' - i.e. a 2+ hour floor on a single build stage for one MoE model, with the answer being yes. This is the RUNTIME engine build stage (converting a checkpoint into a TensorRT engine), not the C++ compile, and the two are distinct costs that both belong in a reader's model of the build step.
Notes
Filed 2025-10-11, CLOSED. Conditions fully stated, which is why it is citable: Ubuntu, CUDA 12.8, b200, driver 570.133.20, TensorRT-LLM 1.0.0.rc6, following the DeepSeek-R1 quick-start recipe. The user's observation is that the process sits at high CPU ('looks cpus are busy' per htop) for 2+ hours and 'is this expected?' - i.e. a 2+ hour floor on a single build stage for one MoE model, with the answer being yes. This is the RUNTIME engine build stage (converting a checkpoint into a TensorRT engine), not the C++ compile, and the two are distinct costs that both belong in a reader's model of the build step.
Cross-references
This record references no other records.