I’m trying to run some code which requires an older version of pytorch (v1.1.0) on an RTX 3090 card and sending anything to the gpu hang for a really long time. I know this is to be expected since the GPU and software compute capability differ in the major version (7.5 vs 8.6) but is there anything I can do about this or do I just have to patch the code?
The “hang” you are seeing is most likely the JIT compilation for the (new) GPU architecture.
In PyTorch 1.1.0 CUDA>=11.0 wasn’t released yet, so you would need to cherry-pick all CUDA11 commits to this branch and recompile it.
I think the easier way would be to update the PyTorch code base and use the latest release (also for new bug fixes and new features).