steveimm/tritonserver-mmlab repository overview

⁠CU121

This tag is based on the nvcr.io/nvidia/tritonserver:24.10-pyt-python-py3. This one adds support for more modern devices, e.g. ampere and adalovalace.

The libraries that are upgraded are torch==2.1.0 and torchvision==0.16.0. Additionally, for CUDA12, mmcv needs to be compiled with c++17.

⁠CU117

This tag is based on the official tritonserver image by NVIDIA, specifically nvcr.io/nvidia/tritonserver:22.08-pyt-python-py3.

The reason to use that particular image as the base image is because it comes with latest version of CUDA 11.7. That specific CUDA version is required to compile the mmcv library.

Additionally, all the additional libraries have been configured so as not to conflict with one another, specifically:

• torch==2.0.1
• torchvision==0.15.2
• openmim
• mmengine==0.10.5
• mmcv==2.0.1
• mmdet==3.0.0rc6

mmrotate needs to be compiled from source and it has to be from dev-1.x branch.

Finally, there is still problem with opencv which is ImportError: libGL.so.1: cannot open shared object file: No such file or directory. To fix that, I added opencv-python-headless library.