trunk/b6008cd71ce673b385224569217395e09a791763: [CUDA] Limit allowed peak-memory for OOM tests on UVM systems (#193641)
- PyTorch: 2242 events in the last 90 days
- PyTorch: 2213th Release in the last 90 days
- Previous: earlier the same day · trunk/32575604cf23f171d0d5973ee7962e12701fd887
What happened
Systems with shared host and device memory e.g., dgx spark will happily allow pytorch to use all of the system's memory which can crash the process or worse during testing. This PR adds a decorator to limit the allowed max memory if such a device is detected for OOM-probing tests and applies it to relevant tests. authored with codex Pull Request resolved: #193641 Approved by: https://github.com/jeffdaily Co-authored…
Summary assembled by rule from the sources below