Hi! I have an environment where I cannot upgrade CUDA from 11.1 to 11.3. Can I still use PyTorch 1.10 despite the fact that it was built with CUDA 11.3? As far as I understand, the answer is “yes” and the biggest risk I take is that torch 1.10 uses some kernels missing in 11.3 (which will cause explicit error that I will see). Is it correct?
If you are installing the pip wheels or conda binaries, then note that they are shipping with their own CUDA runtime (statically linked in the pip wheels, dynamically via
cudatoolkit in the conda binaries).
Your local CUDA toolkit will be used, if you are building from source or custom CUDA extensions.
No, since CUDA, cuDNN, NCCL, etc. are linked to the binaries, no kernels should be missing (there was a cuDNN issue with static linking, but that’s independent from the wheels).