Docs: more details about CUDA models with multiprocessing

[Jiaming-Liu]

In pytorch/docs/source/notes/cuda.rst, it is mentioned that:

unless care is taken to meet the data handling requirements exactly, it is likely that your program will have incorrect or undefined behavior.

IMO, more details on the requirements would be helpful for this part.

[apaszke]

The most important thing are the requirements of the CUDA API. I'd discourage its use unless you know some lower level detail of CUDA IPC. The most important limitation is that if you allocate a CUDA tensor X in process A and send it to process B, then X should never go out of scope and be freed in A until it's used in B (NOTE: it's automatically freed upon A's exit). So it might be ok for sharing CUDA parameters across processes, but not e.g. for data loading.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results here

• 📄 Preview Python docs built from this PR
• 📄 Preview C++ docs built from this PR
• ❓ Need help or want to give feedback on the CI? Visit our office hours

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.