Merge pull request #3863 from karttikjangid/feat/gsoc7-optimize-docker-image-remove-pytorch

replace PyTorch with minimal CUDA runtime wheels, reduce image by ~4.7 GB

Author: javizqh

scripts/RADI/Dockerfile.dependencies_humble

@@ -28,6 +28,12 @@ ENV PYTHONPATH=/opt/ros/${ROS_DISTRO}/lib/python3.10/site-packages
 ENV ROS_PYTHON_VERSION=3
 ENV ROS_VERSION=2
 
+# gz.msgs10 Python bindings (installed by gz-harmonic apt package) are compiled
+# with protoc 3.x and cannot create descriptors under protobuf 4.x / 5.x C++
+# extension. Force pure-Python protobuf parser (backward-compatible with all
+# .pb2 files) so gz.msgs10 and onnxruntime-gpu coexist without version pinning.
+ENV PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
 # Install common tools
 RUN apt-get update && apt-get install -y \
   software-properties-common \
@@ -162,9 +168,42 @@ RUN python3.10 -m pip install --no-cache-dir \
   pyulog==1.0.1 pyyaml==5.4.1 requests==2.31.0 serial==0.0.97 six==1.16.0 toml==0.10.2 psutil==5.9.0 \
   onnxruntime-gpu==1.22.0 Pillow==9.0.1 opencv-python==4.5.5.64 watchdog==2.1.5 utm==0.7.0 psycopg2 jedi pyapriltags
 
-# install pytorch with CUDA 12.8 support 
-RUN python3.10 -m pip install --no-cache-dir --ignore-installed sympy==1.13.3
-RUN python3.10 -m pip install --no-cache-dir torch --extra-index-url https://download.pytorch.org/whl/cu128
+# CUDA runtime wheels for onnxruntime-gpu==1.22.0 CUDA Execution Provider.
+#
+# All seven packages below must be upgraded together whenever onnxruntime-gpu
+# is bumped. Compatibility ranges declared by the wheel (via ~= operator):
+#   nvidia-cuda-runtime-cu12  ~= 12.0   nvidia-cuda-nvrtc-cu12 ~= 12.0
+#   nvidia-cublas-cu12        (runtime dep of cudnn)
+#   nvidia-cudnn-cu12         ~=  9.0
+#   nvidia-curand-cu12        ~= 10.0   nvidia-cufft-cu12      ~= 11.0
+#
+# Upgrade checklist:
+#   1. Check the new ORT release notes for CUDA / cuDNN major version changes.
+#   2. Run: pip show onnxruntime-gpu | grep Requires  — to see declared ~= ranges.
+#   3. Pick PyPI versions whose MAJOR matches those ranges (stay on CUDA 12.x,
+#      cuDNN 9.x, curand 10.x, cufft 11.x until ORT itself changes the range).
+#   4. After build, verify inside the container:
+#        python3 -c "import onnxruntime as ort; print(ort.get_available_providers())"
+#      CUDAExecutionProvider must appear first.
+#   5. Run: ldd $(python3 -c "import onnxruntime, os; print(os.path.dirname(onnxruntime.__file__))")/capi/libonnxruntime_providers_cuda.so
+#      Every .so must resolve (no "not found" lines).
+RUN python3.10 -m pip install --no-cache-dir \
+    nvidia-cuda-runtime-cu12==12.8.90 \
+    nvidia-cublas-cu12==12.8.4.1 \
+    nvidia-cudnn-cu12==9.10.2.21 \
+    nvidia-nvjitlink-cu12==12.8.93 \
+    nvidia-cuda-nvrtc-cu12==12.8.93 \
+    nvidia-curand-cu12==10.3.10.19 \
+    nvidia-cufft-cu12==11.4.1.4
+
+# Register all pip-installed nvidia .so files with ldconfig so onnxruntime-gpu
+# finds them regardless of whether pip used site-packages or dist-packages
+RUN python3.10 -c "\
+import sysconfig, pathlib; \
+site = pathlib.Path(sysconfig.get_path('purelib')); \
+conf = '\n'.join(str(d) for d in site.glob('nvidia/*/lib') if d.is_dir()); \
+pathlib.Path('/etc/ld.so.conf.d/nvidia-pip.conf').write_text(conf + '\n')" \
+    && ldconfig
 
 # monaco editor
 RUN python3.10 -m pip install --no-cache-dir black==24.10.0