Get Help and File an Issue
Use the route that matches what you need. Public issues are visible to everyone, so remove credentials, private URLs, proprietary logs, and license-restricted model artifacts before submitting anything.
Do not open a public issue for a suspected vulnerability. Follow the security policy to report it privately to NVIDIA PSIRT.
Choose an issue type
| Need | Route |
|---|---|
| Help using or understanding TensorRT-Model-Connect | Ask a question |
| Reproducible behavior that differs from the documented contract | Report a bug |
| A new model, capability, or improvement | Request a feature |
| Incorrect, unclear, or missing documentation | Request a documentation change |
| Unsure which route applies | Open the issue chooser |
Questions receive best-effort help from maintainers and the community. The question form does not create a support SLA or establish that an unqualified model, configuration, or target is supported.
Before filing
- Check the supported-model inventory for the exact checkpoint, profile, precision, runtime path, and evidence level.
- Check Known Issues, this section's Troubleshooting, and First-run Troubleshooting.
- Search open and closed issues for the same model, error, or requested behavior. Add evidence to an existing issue instead of opening a duplicate.
- Reduce failures to the smallest command, configuration, and input that still reproduces the behavior. Preserve the first error rather than retrying with unrelated flags.
Information maintainers need
Include the evidence that applies to your request:
- the TensorRT-Model-Connect release, tag, or full commit SHA;
- installation method and exact container image or package version;
- operating system, GPU model, driver, CUDA, and TensorRT versions;
- exact Hugging Face model ID and revision, or a description of the local checkpoint without uploading restricted artifacts;
- exact commands and relevant configuration, with secrets redacted;
- expected behavior, observed behavior, and the first relevant error or log;
- a minimal reproducer and whether the problem also occurs with the documented model-owned manifest; and
- for performance reports, the timing boundary, warmup, sample count, workload, and output-quality gate.
A command being accepted by a parser, a family being registered, and a model being qualified on an exact hardware/software tuple are different evidence levels. State what you verified and what you did not run.
Contributing a fix
If you plan to submit a change, read CONTRIBUTING.md and the Developer Guide. Link the issue from the pull request so reviewers can connect the proposed implementation to the reported behavior and exit criteria.