administrator/frigate
1
0
Fork 0
mirror of https://github.com/blakeblackshear/frigate.git synced 2026-11-02 03:40:31 +03:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

build(tensorrt): JetPack 7 / L4T R39 image with GPU-accelerated ONNX detector

Browse Source 
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.
This commit is contained in:
Jorge Ramirez
2026-06-27 17:14:01 -07:00
parent 1ec511b66c
commit 243172fbf4
19 changed files with 486 additions and 73 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docker/tensorrt/Dockerfile.arm64
+64 -14
View File
@@ -4,6 +4,12 @@
ARG DEBIAN_FRONTEND=noninteractive
ARG BASE_IMAGE
ARG TRT_BASE=nvcr.io/nvidia/tensorrt:23.12-py3
ARG BUILD_ONNXRUNTIME_FROM_SOURCE=0
ARG ONNXRUNTIME_VERSION=1.25.1
ARG ONNXRUNTIME_BRANCH=rel-1.25.1
ARG TENSORRT_PYTHON_BRANCH=release/8.6
ARG L4T_APT_RELEASE=
ARG JETSON_SOC_REPO=
# Build TensorRT-specific library
FROM ${TRT_BASE} AS trt-deps
@@ -12,7 +18,9 @@ ARG TARGETARCH
ARG COMPUTE_LEVEL
RUN apt-get update \
&& apt-get install -y git build-essential cuda-nvcc-* cuda-nvtx-* libnvinfer-dev libnvinfer-plugin-dev libnvparsers-dev libnvonnxparsers-dev \
&& TRT_DEV_PACKAGES="git build-essential cuda-nvcc-* cuda-nvtx-* libnvinfer-dev libnvinfer-plugin-dev libnvonnxparsers-dev" \
&& if apt-cache show libnvparsers-dev > /dev/null 2>&1; then TRT_DEV_PACKAGES="${TRT_DEV_PACKAGES} libnvparsers-dev"; fi \
&& apt-get install -y ${TRT_DEV_PACKAGES} \
&& rm -rf /var/lib/apt/lists/*
RUN --mount=type=bind,source=docker/tensorrt/detector/tensorrt_libyolo.sh,target=/tensorrt_libyolo.sh \
/tensorrt_libyolo.sh
@@ -58,6 +66,12 @@ HEALTHCHECK --start-period=600s --start-interval=5s --interval=15s --timeout=5s
FROM ${BASE_IMAGE} AS build-wheels
ARG DEBIAN_FRONTEND
ARG BUILD_ONNXRUNTIME_FROM_SOURCE
ARG ONNXRUNTIME_VERSION
ARG ONNXRUNTIME_BRANCH
ARG TENSORRT_PYTHON_BRANCH
ARG L4T_APT_RELEASE
ARG JETSON_SOC_REPO
# Add deadsnakes PPA for python3.11
RUN apt-get -qq update && \
@@ -69,41 +83,66 @@ RUN apt-get -qq update && \
RUN apt-get -qq update \
&& apt-get -qq install -y --no-install-recommends \
python3.11 python3.11-dev \
wget build-essential cmake git \
wget curl build-essential cmake git ninja-build \
&& rm -rf /var/lib/apt/lists/*
# Ensure python3 defaults to python3.11
RUN update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.11 1
# --ignore-installed: on the JP7 Ubuntu 24.04 (noble) base the deadsnakes python3.11 ships a
# distro-packaged pip 24.0 with no RECORD file, so get-pip.py's default reinstall fails with
# "uninstall-no-record-file". Installing over it without uninstalling avoids that; harmless on JP6.
RUN wget -q https://bootstrap.pypa.io/get-pip.py -O get-pip.py \
&& sed -i 's/args.append("setuptools")/args.append("setuptools==77.0.3")/' get-pip.py \
&& python3 get-pip.py "pip"
&& python3 get-pip.py --ignore-installed "pip"
FROM build-wheels AS trt-wheels
ARG DEBIAN_FRONTEND
ARG TARGETARCH
ARG BUILD_ONNXRUNTIME_FROM_SOURCE
ARG ONNXRUNTIME_VERSION
ARG ONNXRUNTIME_BRANCH
ARG TENSORRT_PYTHON_BRANCH
# python-tensorrt build deps are 3.4 GB!
RUN apt-get update \
&& apt-get install -y ccache cuda-cudart-dev-* cuda-nvcc-* libnvonnxparsers-dev libnvparsers-dev libnvinfer-plugin-dev \
&& TRT_DEV_PACKAGES="ccache cuda-cudart-dev-* cuda-nvcc-* libnvonnxparsers-dev libnvinfer-plugin-dev" \
&& if apt-cache show libnvparsers-dev > /dev/null 2>&1; then TRT_DEV_PACKAGES="${TRT_DEV_PACKAGES} libnvparsers-dev"; fi \
&& apt-get install -y ${TRT_DEV_PACKAGES} \
&& ([ -e /usr/local/cuda ] || ln -s /usr/local/cuda-* /usr/local/cuda) \
&& rm -rf /var/lib/apt/lists/*;
# The JP7 base image ships a stale cmake 3.24 at /usr/local/bin/cmake that shadows apt's 3.28.
# TensorRT 10.11's python CMakeLists (and the ONNX Runtime source build) require cmake >= 3.27,
# so install a current cmake from pip into /usr/local/bin ahead of the stale one. No-op risk on
# JP6 where the bundled cmake already satisfies the minimum.
RUN pip3 install --no-cache-dir --upgrade "cmake>=3.27"
# Determine version of tensorrt already installed in base image, e.g. "Version: 8.4.1-1+cuda11.4"
RUN NVINFER_VER=$(dpkg -s libnvinfer8 | grep -Po "Version: \K.*") \
&& echo $NVINFER_VER | grep -Po "^\d+\.\d+\.\d+" > /etc/TENSORRT_VER
RUN NVINFER_VER="$(dpkg-query -W -f='${Version}\n' 'libnvinfer*' 2>/dev/null | grep -E '^[0-9]+[.][0-9]+[.][0-9]+' | sort -V | tail -n1)" \
&& test -n "$NVINFER_VER" \
&& TENSORRT_VER="$(echo "$NVINFER_VER" | grep -Eo '^[0-9]+[.][0-9]+[.][0-9]+')" \
&& TENSORRT_MAJOR="$(echo "$TENSORRT_VER" | cut -d. -f1)" \
&& echo "$TENSORRT_VER" > /etc/TENSORRT_VER \
&& echo "$TENSORRT_MAJOR" > /etc/TENSORRT_MAJOR
RUN --mount=type=bind,source=docker/tensorrt/detector/build_python_tensorrt.sh,target=/deps/build_python_tensorrt.sh \
--mount=type=cache,target=/root/.ccache \
export PATH="/usr/lib/ccache:$PATH" CCACHE_DIR=/root/.ccache CCACHE_MAXSIZE=2G \
&& TENSORRT_VER=$(cat /etc/TENSORRT_VER) /deps/build_python_tensorrt.sh
&& TENSORRT_VER=$(cat /etc/TENSORRT_VER) TENSORRT_PYTHON_BRANCH="${TENSORRT_PYTHON_BRANCH}" /deps/build_python_tensorrt.sh
COPY docker/tensorrt/requirements-arm64.txt /requirements-tensorrt.txt
RUN pip3 wheel --wheel-dir=/trt-wheels -r /requirements-tensorrt.txt
# See https://elinux.org/Jetson_Zoo#ONNX_Runtime
ADD https://nvidia.box.com/shared/static/9yvw05k6u343qfnkhdv2x6xhygze0aq1.whl /trt-wheels/onnxruntime_gpu-1.19.0-cp311-cp311-linux_aarch64.whl
RUN --mount=type=bind,source=docker/tensorrt/detector/build_onnxruntime_arm64.sh,target=/deps/build_onnxruntime_arm64.sh \
mkdir -p /trt-wheels \
&& if [ "${BUILD_ONNXRUNTIME_FROM_SOURCE}" = "1" ]; then \
ONNXRUNTIME_VERSION="${ONNXRUNTIME_VERSION}" ONNXRUNTIME_BRANCH="${ONNXRUNTIME_BRANCH}" /deps/build_onnxruntime_arm64.sh; \
else \
wget -q https://nvidia.box.com/shared/static/9yvw05k6u343qfnkhdv2x6xhygze0aq1.whl -O /trt-wheels/onnxruntime_gpu-1.19.0-cp311-cp311-linux_aarch64.whl; \
fi
FROM build-wheels AS trt-model-wheels
ARG DEBIAN_FRONTEND
@@ -116,16 +155,20 @@ RUN --mount=type=bind,source=docker/tensorrt/requirements-models-arm64.txt,targe
FROM wget AS jetson-ffmpeg
ARG DEBIAN_FRONTEND
ARG L4T_APT_RELEASE
ARG JETSON_SOC_REPO
ENV CCACHE_DIR /root/.ccache
ENV CCACHE_MAXSIZE 2G
RUN --mount=type=bind,source=docker/tensorrt/build_jetson_ffmpeg.sh,target=/deps/build_jetson_ffmpeg.sh \
--mount=type=cache,target=/root/.ccache \
/deps/build_jetson_ffmpeg.sh
L4T_APT_RELEASE="${L4T_APT_RELEASE}" JETSON_SOC_REPO="${JETSON_SOC_REPO}" /deps/build_jetson_ffmpeg.sh
# Frigate w/ TensorRT for NVIDIA Jetson platforms
FROM tensorrt-base AS frigate-tensorrt
RUN apt-get update \
&& apt-get install -y python-is-python3 libprotobuf23 \
&& pick_package() { for package in "$@"; do if apt-cache show "$package" > /dev/null 2>&1; then echo "$package"; return 0; fi; done; return 1; } \
&& PROTOBUF_RUNTIME="$(pick_package libprotobuf23 libprotobuf32t64 libprotobuf32)" \
&& apt-get install -y python-is-python3 "${PROTOBUF_RUNTIME}" \
&& rm -rf /var/lib/apt/lists/*
COPY --from=jetson-ffmpeg /rootfs /
@@ -134,23 +177,30 @@ ENV INCLUDED_FFMPEG_VERSIONS="${DEFAULT_FFMPEG_VERSION}:${INCLUDED_FFMPEG_VERSIO
# ffmpeg runtime dependencies
RUN apt-get -qq update \
&& pick_package() { for package in "$@"; do if apt-cache show "$package" > /dev/null 2>&1; then echo "$package"; return 0; fi; done; return 1; } \
&& X264_RUNTIME="$(pick_package libx264-163 libx264-164)" \
&& X265_RUNTIME="$(pick_package libx265-199 libx265-209)" \
&& apt-get -qq install -y --no-install-recommends \
libx264-163 libx265-199 libegl1 \
"${X264_RUNTIME}" "${X265_RUNTIME}" libegl1 \
&& rm -rf /var/lib/apt/lists/*
# Fixes "Error loading shared libs"
RUN mkdir -p /etc/ld.so.conf.d && echo /usr/lib/ffmpeg/jetson/lib/ > /etc/ld.so.conf.d/ffmpeg.conf
COPY --from=trt-wheels /etc/TENSORRT_VER /etc/TENSORRT_VER
COPY --from=trt-wheels /etc/TENSORRT_MAJOR /etc/TENSORRT_MAJOR
RUN --mount=type=bind,from=trt-wheels,source=/trt-wheels,target=/deps/trt-wheels \
--mount=type=bind,from=trt-model-wheels,source=/trt-model-wheels,target=/deps/trt-model-wheels \
pip3 uninstall -y onnxruntime \
&& pip3 install -U /deps/trt-wheels/*.whl \
&& pip3 install -U /deps/trt-model-wheels/*.whl \
# --ignore-installed: on the noble JP7 base, `-U` would try to upgrade distro-managed deps that
# lack a RECORD file (e.g. wheel/setuptools), failing with "uninstall-no-record-file". Install
# the TRT/ORT GPU wheels over them into /usr/local instead. Harmless on JP6.
&& pip3 install -U --ignore-installed /deps/trt-wheels/*.whl \
&& pip3 install -U --ignore-installed /deps/trt-model-wheels/*.whl \
&& ldconfig
WORKDIR /opt/frigate/
COPY --from=rootfs / /
# Fixes "Error importing detector runtime: /usr/lib/aarch64-linux-gnu/libstdc++.so.6: cannot allocate memory in static TLS block"
ENV LD_PRELOAD /usr/lib/aarch64-linux-gnu/libstdc++.so.6
ENV LD_PRELOAD /usr/lib/aarch64-linux-gnu/libstdc++.so.6