holoscan 3.4.0

The Holoscan SDK: building high-performance AI streaming applications

Holoscan SDK

The Holoscan SDK Python Wheel is part of NVIDIA Holoscan, the AI sensor processing platform that combines hardware systems for low-latency sensor and network connectivity, optimized libraries for data processing and AI, and core microservices to run streaming, imaging, and other applications, from embedded to edge to cloud. It can be used to build streaming AI pipelines for a variety of domains, including Medical Devices, High Performance Computing at the Edge, Industrial Inspection and more.

Getting Started

Visit the Holoscan User Guide to get started with the Holoscan SDK.

Prerequisites

• Prerequisites for each supported platform are documented in the user guide. Note that the python wheels have a lot of optional dependencies which you may install manually based on your needs (see compatibility matrix at the bottom).
• The Holoscan SDK python wheels are only formally tested on Ubuntu 22.04. They are, however, expected to work on any Linux distribution with glibc 2.35 or above (see output of ldd --version) and CUDA Runtime 12.2 or above.
• Python: 3.9 to 3.12

Troubleshooting

Version 0.6.0 gets installed instead of the latest version

The latest version of the wheels were built and tested on Ubuntu 22.04 with glibc 2.35. You'll need to switch to a Linux distribution with a more recent version of glibc to use the Holoscan SDK python wheels 1.0 or above (check your version with ldd --version), or use the Holoscan SDK NGC container instead.

ERROR: Could not find a version that satisfies the requirement holoscan==<version>
ERROR: No matching distribution found for holoscan==<version>

Same as above, OR incompatible python version.

libc.so.6: version 'GLIBC_2.32 not found
libstdc++.so.6: version `GLIBCXX_3.4.29` not found

Same as above.

ImportError: libcudart.so.12: cannot open shared object file: No such file or directory

CUDA runtime is missing from your system (required even for CPU only pipelines).

• x86_64: two options
  • A) System Installation: Follow the official installation steps for installing the whole toolkit, which should at least include apt install cuda-runtime-dev-12-6.
  • B) PIP installation:

    python3 -m pip install nvidia-cuda-runtime-cu12
    export CUDA_WHL_LIB_DIR=$(python3 -c 'import nvidia.cuda_runtime; print(nvidia.cuda_runtime.__path__[0])')/lib
    export LD_LIBRARY_PATH="$CUDA_WHL_LIB_DIR:$LD_LIBRARY_PATH"
    

• IGX Orin: Ensure the compute stack is installed.
• Jetson: Re-install JetPack 6.0.

catastrophic error: cannot open source file "vector_types.h"

CUDA Runtime headers are missing from your system.

Resolution: same as above.

Reference: https://docs.cupy.dev/en/latest/install.html#cupy-always-raises-nvrtc-error-compilation-6

Error: libnvinfer.so.8: cannot open shared object file: No such file or directory
...
Error: libnvonnxparser.so.8: cannot open shared object file: No such file or directory

TensorRT is missing from your system (note that it is only needed by the holoscan.operators.InferenceOp operator.).

• x86_64: Follow the official installation steps.
  • A) System Installation: Follow the official installation steps.
  • B) PIP installation:

    python3 -m pip install tensorrt-libs~=8.6.1 --index-url https://pypi.nvidia.com
    export TRT_WHL_LIB_DIR=$(python3 -c 'import tensorrt_libs; print(tensorrt_libs.__path__[0])')
    export CUDNN_WHL_LIB_DIR=$(python3 -c 'import nvidia.cudnn; print(nvidia.cudnn.__path__[0])')/lib
    export CUBLAS_WHL_LIB_DIR=$(python3 -c 'import nvidia.cublas; print(nvidia.cublas.__path__[0])')/lib
    export LD_LIBRARY_PATH="$TRT_WHL_LIB_DIR:$CUDNN_WHL_LIB_DIR:$CUBLAS_WHL_LIB_DIR:$LD_LIBRARY_PATH"
    

• IGX Orin: Ensure the compute stack is installed.
• Jetson: Re-install JetPack 6.0.