// the find
dusty-nv/jetson-containers
Machine Learning Containers for NVIDIA Jetson and JetPack-L4T
A modular container build system for running ML workloads on NVIDIA Jetson edge devices. It handles the painful part of Jetson ML setup — cross-compiling wheels, resolving JetPack/CUDA version compatibility, and wiring up hardware access inside containers. If you have a Jetson Orin and want to run llama.cpp or vLLM without spending a week on toolchain issues, this is the fastest path there.
The autotag mechanism is genuinely useful — it picks a compatible pre-built image for your specific JetPack version instead of making you figure out the matrix yourself. The local devpi + APT cache server is a real time-saver on slow embedded hardware where pip installs from PyPI are painful. Package coverage is unusually wide: everything from vLLM and SGLang to ROS2, Isaac Sim, and Home Assistant in a single coherent system with shared base layers. The move to ARM SBSA in CUDA 13.2 unifies Jetson Orin with GH200/GB200, which means community-built wheels will be more broadly applicable going forward.
Support is explicitly scoped to JetPack 6.2 and JetPack 7 only — if you're on an older Orin or a Nano/Xavier, you're pointed at a legacy branch with no active maintenance. The dependency graph between packages is implicit in the build scripts rather than declared in a lockfile or manifest, so debugging why a composed build broke requires reading Python config files scattered across directories. There's no clear story for reproducibility: `autotag` picks whatever is compatible at runtime, not a pinned digest, so a build that worked last month may silently pull a different image today. The GitHub Actions CI only builds monthly for most packages, so the gap between upstream releases and tested container images can stretch for weeks.