Tech Stack Overview

May 5, 2026 · View on GitHub

How the toolchain fits together — from build.bat to a running app with loaded extensions.

Components

Component	What it is	Where it lives
repoman	NVIDIA repo orchestration tool. Coordinates dependencies, building, testing, packaging, and publishing. All `repo.bat` subcommands dispatch through it.	`tools/repoman/repoman.py`
packman	Binary package manager. Downloads and caches build dependencies (Kit SDK, premake, runtime assets) from NVIDIA servers. Packages are cached in `C:\packman-repo\` (Windows) or `~/.cache/packman/` (Linux) and symlinked into `_build/`.	`tools/packman/`
Kit SDK	The Omniverse application framework. Provides the extension system, UI framework (`omni.ui`), USD integration, and an embedded Python 3.10 interpreter. Pulled in by packman via `deps/kit-sdk.packman.xml`.	`_build/windows-x86_64/release/kit/`
premake	Build system generator. In this repo, primarily used to discover extensions and create symlinks from `source/` to `_build/exts/` so Kit can load source code directly without copying.	`_build/host-deps/premake/`
Python 3.10	Bundled with Kit — no system Python required. All extensions, tools, and scripts run under Kit's embedded interpreter.	`_build/windows-x86_64/release/kit/python/`

Dependency Files (`deps/`)

File	What it pulls
`kit-sdk.packman.xml`	Kit SDK (kit-kernel) — the application framework + embedded Python
`repo-deps.packman.xml`	Repo tools (repo_build, repo_test, repo_format, repo_kit_tools, etc.)
`host-deps.packman.xml`	Build-time tools (premake)
`target-deps.packman.xml`	Runtime dependencies (RTX Remix runtime, HDRemix models, AI tools models)
`pip_flux.toml`	Open-source Python packages (frozen pip archive — see below)
`pip_internal.toml`	NVIDIA internal Python packages (frozen pip archive — see below)

Third-party Python packages (e.g., pydantic, numpy, torch) are not installed via pip install at runtime. They are pre-bundled into frozen archives during the build and loaded by a dedicated extension at startup.

Pip dependencies

File	Target folder	Contains
`deps/pip_flux.toml`	`_build/target-deps/flux_pip_prebundle/`	Open-source packages: `pydantic`, `numpy`, `fastapi`, `pillow`, `pygit2`, `sentry-sdk`, `huggingface-hub`, etc.

Repoman installs them during the build (configured in repo.toml under [repo_build.fetch.pip]).

How packages reach extensions at runtime

During the build, repoman runs pip install from the TOML file into the prebundle folder using Kit's embedded Python
The omni.flux.pip_archive extension loads very early (order -2000) and adds the prebundle folder to sys.path
Any extension that imports a third-party package must declare "omni.flux.pip_archive" = {} in its [dependencies] — without it, the import will fail even though the package exists on disk
After merging to main, the prebundle is published to NVIDIA's packman servers so subsequent builds download the pre-built archive instead of running pip

Why two files?

License separation: open-source packages vs. proprietary NVIDIA packages require different OSRB tracking
Dependency layering: internal packages declare install_dependencies = false because pip_flux.toml already provides all their transitive dependencies
Access control: pip_internal.toml uses NVIDIA's internal PyPI registry, which requires VPN access

For the full procedure to add a new package, see Adding Pip Package Dependencies.

Build Pipeline

What happens when you run .\build.bat:

build.bat
  → repo.bat build
    → repoman.py (via Kit's Python)

      1. Resolve dependencies (packman)
         Download/cache Kit SDK, premake, runtime deps → symlink into _build/

      2. Install pip dependencies
         Read deps/pip_flux.toml → pip install into Kit's Python environment

      3. Discover extensions
         Scan source/extensions/ → find ~190 extension.toml files

      4. Generate symlinks (premake)
         For each extension: source/extensions/<name>/ → _build/exts/<name>/
         Kit reads source directly — no compile step needed for Python

      5. Run pre-build commands
         Type stub generation, CLI wrapper generation

      6. Precache extensions (optional)
         Run Kit once per app to pre-download registry dependencies into extscache/

Build Output

_build/windows-x86_64/release/
├── kit/                    ← Kit SDK (from packman)
│   ├── kit.exe             ← The application
│   ├── python/             ← Embedded Python 3.10 (+ pip packages)
│   └── exts/               ← Built-in Kit extensions
├── exts/                   ← Symlinked extensions from this repo (~190)
├── extscache/              ← Downloaded extension dependencies
├── apps/                   ← .kit app definitions
├── tests-*.bat             ← Generated test runners (one per extension)
└── lightspeed.app.*.bat    ← App launchers

Runtime

When you launch lightspeed.app.trex_dev.bat:

Kit loads the .kit app file (e.g., lightspeed.app.trex.kit)
Resolves [dependencies] — each dependency is an extension name
Searches for extensions in exts/, extscache/, and kit/exts/
For each extension: loads extension.toml, imports the Python module, calls on_startup()
App is running with all extensions loaded

Configuration

File	Purpose
`repo.toml`	Central config for repoman — controls build, test, format, lint, package, publish, and docs phases
`source/apps/*.kit`	App definitions — list which extensions to load and their settings
`source/extensions/*/config/extension.toml`	Extension metadata — version, dependencies, settings, test config
`source/extensions/*/premake5.lua`	Symlink setup — maps source directories into the build output
`deps/pip_flux.toml`	Open-source Python packages — installed into Kit's Python during build
`deps/pip_internal.toml`	NVIDIA internal Python packages — installed after `pip_flux.toml`