Skip to main content

First-run Troubleshooting

Do not change several model flags at once. Identify the first boundary that fails in the Quick Start.

FailureCheck firstNext action
nvidia-smi or GPU accessHost driver/container runtimeFix GPU visibility before installing or building.
trtmc not foundActive wheel environment or source build pathRun trtmc version in the same shell.
Hugging Face 401/403/not foundExact hf_id, network, auth, gated accessVerify the checkpoint and cache; do not substitute a same-family model silently.
CMake cannot find CUDA/TensorRTDevelopment environmentReturn to the supported container or provide the matching headers/libraries.
Build runs out of memory/diskConfigured cache/profile exceeds available memory or supported tacticsUse the documented Quick Start profile or free capacity.
Bundle inspection failsIncomplete/corrupt outputRebuild and retain the first builder error.
Native plugin not registeredModel DSO/search pathConfirm the owning runtime DSO and pass --model-plugin-dir when needed.
TensorRT or DSO ABI errorMixed build/runtime cohortRun with the same compatible environment used to build/package the artifact.
Output differsPrompt, revision, precision, decodingReproduce the exact manifest and deterministic settings before changing thresholds.

Collect this receipt when asking for help:

Git commit:
Install path: wheel | source
Model ID and revision:
Build command:
Run command:
Bundle checksum:
GPU / driver / CUDA / TensorRT:
First failing command:
Complete error:

For failures after the first smoke test, use Troubleshooting and the task-specific User Guide.