Skip to main content

First-run Troubleshooting

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.
Python build module unavailableActive wheel/editable environmentRun python -m tensorrt_model_connect build --help in the same shell.
trtmc unavailableNative build/install output and PATHRun trtmc version or the explicit built binary.
Hugging Face 401/403/not foundExact ID/revision, network, authenticationVerify the checkpoint/cache; never substitute a nearby model silently.
CMake cannot find CUDA/TensorRTDevelopment image and explicit SDK pathsReturn to the documented container/toolchain cohort.
Build OOM or disk failureRequested checkpoint, shape, precision, and cache capacityUse the exact family manifest/profile or free capacity; retain the first error.
No family or multiple families matchRoot checkpoint identity metadataUse a supported exact checkpoint; do not add prefix/fallback matching.
Bundle inspection failsPartial/corrupt bundleRebuild; failed builds must not publish a partial output.
Family/backend DSO missing--runtime-root contentsConfirm libtrtmc_core.so, libtrtmc_runtime.so, selected backend, and exact family DSO are together.
Task mismatchBundle task versus CLI commandUse the Task command named by the family manifest/header.
TensorRT/DSO ABI errorMixed software or hardware cohortRun with a compatible environment and rebuild the bundle when required.
Output differsRevision, input framing, precision, sampling, oracleReproduce the exact family testcase before changing code or thresholds.

Collect the source revision, model ID/revision, complete build/run commands, bundle inspection, GPU/driver/CUDA/TensorRT versions, runtime-root listing, and first complete error when asking for help.