Need workaround to support multiprocess CUDA tensor sharing on Jetson Platforms #60401
Labels
feature
A request for a proper, new feature.
module: multiprocessing
Related to torch.multiprocessing
triaged
This issue has been looked at a team member, and triaged and prioritized into an appropriate module
馃殌 Feature
A different method for sharing CUDA Tensors across processes on Jetson platforms is needed.
CUDA unified addressing based IPC functionality isn't yet supported on Tegra platforms
https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__DEVICE.html#group__CUDART__DEVICE_1g8a37f7dfafaca652391d0758b3667539
and current implementation of THPStorage depends on it, making existing dataloader implementations not compatible on Jetson.
Motivation
Current data loader test cannot pass without the requested feature. Setting
multiprocessing.sharing_strategy
only affects CPU tensors and doesn't go around this problem.gives
cuda runtime error (801) : operation not supported
The text was updated successfully, but these errors were encountered: