Heuristic:Triton inference server Server Documentation Standards
| Knowledge Sources | |
|---|---|
| Domains | Documentation, Infrastructure |
| Last Updated | 2026-02-13 17:00 GMT |
Overview
Reference for the documentation conventions and tooling used in the Triton Inference Server project, including copyright header enforcement, Sphinx-based documentation builds with the NVIDIA theme, multi-repo link rewriting, and Markdown support alongside reStructuredText.
Description
The Triton Inference Server enforces a consistent set of documentation standards across all source files and published documentation. Copyright headers are required on every source file and are managed by two complementary scripts: add_copyright.py (which inserts or updates headers) and check_copyright.py (which validates their presence and correctness in CI). Year ranges within copyright headers are automatically updated when files are modified.
Published documentation is built with Sphinx using the NVIDIA theme (docs/conf.py), which provides consistent branding across NVIDIA open-source projects. The build pipeline supports both reStructuredText (.rst) and Markdown (.md) via myst-parser, allowing contributors to author in either format. A dedicated docs/generate_docs.py script handles multi-repo link rewriting so that cross-references between Triton sub-projects resolve correctly in the unified documentation site.
Usage
Use this reference when adding new source files (to ensure correct copyright headers), contributing documentation (to understand the build pipeline and supported formats), or debugging broken cross-repo links in the generated documentation.
The Insight (Rule of Thumb)
Copyright Header Enforcement:
- All source files must include a copyright header with the correct NVIDIA license text
- add_copyright.py inserts or updates copyright headers automatically across the repository
- check_copyright.py validates headers in CI and fails the build if any file is non-compliant
- Year ranges are auto-updated on modification (e.g., Copyright (c) 2020-2026, NVIDIA CORPORATION)
Sphinx Documentation Build:
- Theme: NVIDIA Sphinx theme for consistent branding across NVIDIA projects
- Configuration: docs/conf.py defines extensions, theme settings, and build options
- Formats supported: reStructuredText (.rst) natively; Markdown (.md) via myst-parser extension
- Build command: standard make html within the docs/ directory
Multi-Repo Link Rewriting:
- generate_docs.py rewrites relative links so cross-references between Triton sub-projects (server, client, backend, common) resolve correctly in unified documentation
- Link targets are rewritten from local file paths to published documentation URLs during the build
Trade-off: Supporting both Markdown and RST lowers the contribution barrier but introduces two syntaxes that contributors must be aware of. RST remains the primary format for complex content (directives, cross-references), while Markdown is preferred for simpler pages and README files.
Reasoning
Enforcing copyright headers at the CI level prevents legal compliance gaps that could accumulate if left to manual review. The two-script approach (add + check) separates the remediation workflow from the validation workflow, allowing developers to fix headers locally before CI rejects their pull request.
The NVIDIA Sphinx theme ensures that Triton documentation is visually consistent with other NVIDIA projects (TensorRT, RAPIDS, etc.), reducing cognitive load for users who work across multiple NVIDIA tools. Using myst-parser to support Markdown alongside RST reflects the reality that many contributors are more comfortable with Markdown, while RST remains necessary for advanced Sphinx directives.
Multi-repo link rewriting via generate_docs.py is essential because Triton is split across multiple repositories. Without link rewriting, cross-references between the server, client, backend, and common repositories would break in the published documentation.
Source evidence from tools/add_copyright.py:
# Inserts or updates copyright headers in source files
# Supports multiple file types: .py, .cc, .h, .proto, .sh, etc.
# Automatically adjusts year ranges based on file modification date
Source evidence from docs/conf.py:
# Sphinx configuration using NVIDIA theme
# Extensions include myst-parser for Markdown support
# Configured for multi-version and multi-repo documentation
Related Pages
- Implementation:Triton_inference_server_Server_AddCopyright
- Implementation:Triton_inference_server_Server_GenerateDocs
- Implementation:Triton_inference_server_Server_SphinxDocConf
- Principle:Triton_inference_server_Server_Copyright_Management
- Principle:Triton_inference_server_Server_Documentation_Build