| title | Contributor Quickstart |
|---|---|
| description | Prepare, validate, and submit a focused TensorRT-Model-Connect contribution. |
Contributions use the same ownership and evidence rules described throughout this site. Start with a small, reviewable change and prove the behavior at the lowest meaningful layer before requesting broader CI.
The repository-root CONTRIBUTING.md is authoritative for the external contribution workflow, licensing, and Developer Certificate of Origin requirements. This page adds project-specific development and validation detail.
External development happens in a personal fork. Use GitHub's Fork button
on the canonical repository, then clone your fork, add the NVIDIA repository as
upstream, and start from its current main:
git clone https://github.com/YOUR-GITHUB-USERNAME/TensorRT-Model-Connect.git
cd TensorRT-Model-Connect
git remote add upstream https://github.com/NVIDIA/TensorRT-Model-Connect.git
git fetch upstream
git switch -c docs/improve-getting-started upstream/mainIn this layout, origin is your writable fork and upstream is the canonical
repository. Do not develop on either repository's main, and preserve
unrelated local work.
For source changes, use the development environment described in System Requirements.
Install the local quality hooks once in the clone:
python3 -m pip install --requirement requirements/community-ci.txt
pre-commit install --install-hooksOn Windows, use py -3 -m pip in place of python3 -m pip.
Commit-time hooks trim trailing whitespace, ensure one final newline, validate YAML, check Ruff, and verify clang-format. Some commit-time hooks modify files; review and stage those fixes before committing again. Pre-commit manages the Ruff and clang-format environments on Linux, macOS, and Windows instead of depending on host-installed binaries.
The local hook intentionally stays lightweight and does not build the CLI or run the complete CPU suite. After pushing, the pull request automatically runs source quality, ownership analysis, and the selected source-only C++ and Python units on GitHub-hosted public CPU runners. The protected suite retains the filesystem-specific cache-reflink contract that public runners cannot portably execute.
Choose the narrowest current owner:
| Change | Start here |
|---|---|
| Native model support | Add a Model Family |
| Exact delegated runtime for an existing family | Add an Optimized Runtime Implementation |
| Native runtime behavior | Add a Runtime Strategy |
| User-facing configuration | Add a Config Schema |
| Model validation | Validate a Model Contribution |
| External family-owned kernel | Bring Your Own Kernel |
| Public API or shared infrastructure | Developer Guide, then the owning architecture or API reference |
Similar model behavior is not automatically shared infrastructure. Family semantics stay with the owning family unless at least two real owners require the same model-independent contract.
- Keep existing copyright, license, and attribution notices.
- Add the repository SPDX header to new source files.
- Identify third-party source, version, license, and required notices in the pull request.
- Sign every commit:
git commit --signoffThe sign-off certifies the contribution under the Developer Certificate of Origin. Unsigned commits are not accepted.
Always start with repository consistency:
PYTHONPATH=python:. python3 tools/model_ci.py validate
PYTHONPATH=python:. python3 tools/test_impact.py --validate
git diff --checkThen run focused tests for the owner you changed. Model changes also need the declared E2E contract and its required model, runtime, GPU, and comparison artifacts. Documentation changes need strict references and a production site build:
python3 tools/check_doc_file_references.py --strict website/docs
npm --prefix website ci
npm --prefix website run buildDo not weaken an acceptance threshold to make a change pass. If a test is wrong, explain the evidence and request maintainer review.
Sync with current upstream, push the short-lived branch to your fork, and open a
pull request targeting NVIDIA/TensorRT-Model-Connect:main:
git fetch upstream
git rebase upstream/main
git push --set-upstream origin docs/improve-getting-startedComplete every pull-request template section. Use
Not applicable: <reason> when a field does not apply instead of deleting it.
Follow the repository's established PR-description structure:
- Background for the problem, motivation, and linked issue;
- Exit Criteria for completion conditions and non-goals;
- Implementation for the approach, affected models/components, and compatibility surface;
- Validation for exact commands and results, tested revisions, hardware and environment, and paths not executed; and
- Notes For Future Readers for remaining risk, provenance, rollout, and follow-up context.
PR Metadata / Required verifies this evidence contract, while trusted triage
automation derives model and component labels from the actual diff and
repository ownership metadata. Risk and compatibility-change labels come from
the corresponding template selections. DCO sign-off is enforced separately by
the repository's DCO check.
Compilation, source tests, model parity, target-hardware execution, performance, and release qualification are different evidence tiers.
Opening the pull request or pushing a new commit automatically starts
contributor-visible, GitHub-hosted Community CPU validation against GitHub's
exact pull-request merge revision. Separate jobs run source quality, ownership
and impact, and source-only C++ and Python units. No comment or maintainer action
is required.
For directly changed Python unit tests, impact analysis also passes the exact
test files to the CPU unit job. This includes non-E2E model-owned tests under
tests/e2e/models/; tests marked gpu, trt, or e2e remain excluded.
The test jobs have read-only repository permission and no access to private
runners, secrets, or GPUs, and every public job uses a GitHub-hosted
ubuntu-24.04 runner. GitHub publishes public Actions logs with the complete
output for every failed command, together with native pull-request checks.
Fix any failures and wait for Community CPU / Required to pass. A new commit
automatically validates the new merge revision and cancels an older in-progress
Community CPU run for the same pull request. If main advances and GitHub asks
for an update, rebase or update the branch to validate the new exact merge.
The repository premerge workflow is one-shot and label-driven. Pushing a branch or creating a pull request alone does not start the gate. After local checks pass and the pull request is ready, add this comment:
@yifeif-nv This PR is ready for CI. Please trigger CI for the current head.
After public CPU validation passes, the maintainer verifies the PR's
headRefOid and applies run-internal-ci. The trusted bridge consumes that
label after authorization succeeds, verifies a successful public CPU run for
the current head, captures the immutable PR head SHA, and dispatches private
premerge validation for that exact revision. The public result is contributor
feedback, not an authorization token; run-internal-ci remains the
protected-resource security boundary.
Wait for the trtmc/premerge/required status on the same head SHA to complete
successfully. If the head changes intentionally, finish the update and local
validation before mentioning @yifeif-nv once to request a new run. Only a
maintainer with repository maintain or admin permission can authorize the
trigger. Private CI repository details, runner information, logs, artifacts,
and URLs are not public documentation.
{/* Collaborative review anchor: batch 2. */}