Adds a `-tensorrt-jp7` Frigate image for JetPack 7.2 / L4T R39.2 Jetson hosts,
built on nvcr.io/nvidia/tensorrt:26.02-py3-igpu (TensorRT 10.11, CUDA 13, py3.12),
keeping the existing JP6 path unchanged. ONNX Runtime GPU is built from source
(no public aarch64 onnxruntime-gpu wheel), TensorRT-Python branch is selected by
the base image, and TensorRT runtime library checks are major-version aware.
VALIDATED on a real AGX Orin (L4T R39.2 / nv_tegra_release R39 rev 2.0):
- image frigate:test-tensorrt-jp7 builds (rc=0, 16.3GB,
sha256:c9e4d382f1603ee130ee4a7315b4f71f9461405e3785707251505e2d6d088df3)
- ONNX Runtime 1.25.1 exposes CUDAExecutionProvider, and a real Add-model
inference RAN on the iGPU CUDA EP (functional, not just listed)
- frigate.util.model.get_ort_providers(False,"AUTO") = [CUDA, CPU] (CUDA first,
CPU last, TensorRT EP excluded) -> the ONNX detector GPU-accelerates on JP7
- /etc/TENSORRT_VER = 10.11.0
Native `type: tensorrt` (.trt gen) + the ORT TensorRT EP stay DRAFT-GATED: the
L4T R39 host ships NO libnvdla_compiler.so (absent from host AND base image), so
`import tensorrt` and libonnxruntime_providers_tensorrt.so fail to load. ONNX
detector GPU acceleration is the supported JP7 path; native trt is deferred.
Build fixes the new noble/CUDA-13 base surfaced (beyond the plan):
- build_nginx.sh: enable deb-src for the deb822 ubuntu.sources (Ubuntu 24.04)
- tensorrt_libyolo.sh: strip -lnvToolsExt (removed in CUDA 13) + -lnvparsers
(dropped in TensorRT 10) when those libs are absent
- docker/main: noble/py3.12 build adjustments (Dockerfile, build_sqlite_vec.sh)
Reproducible: `make -C docker/tensorrt ... local-trt-jp7` (JETPACK7_ARGS) on any
arm64 builder; built on-device only because the GPU smoke test needs the iGPU.
* Non-Jetson changes
Required for later commits:
- Allow base image to be overridden (and don't assume its WORKDIR)
- Ensure python3.9
- Map hwaccel decode presets as strings instead of lists
Not required:
- Fix existing documentation
- Simplify hwaccel scale logic
* Prepare for multi-arch tensorrt build
* Add tensorrt images for Jetson boards
* Add Jetson ffmpeg hwaccel
* Update docs
* Add CODEOWNERS
* CI
* Change default model from yolov7-tiny-416 to yolov7-320
In my experience the tiny models perform markedly worse without being
much faster
* fixup! Update docs
