trunk/1021ddaa95dbaec8775442af63d998591b50a353: Key sparse Triton tuning tables by the current accelerator (#194176)
- PyTorch: 2633 events in the last 90 days
- PyTorch: 2594th Release in the last 90 days
- Previous: earlier the same day · trunk/41003927e76c4e4254fd45fa83f2ade160d2617a
What happened
Related to this RFC: #194355 Summary _get_device_name() supplies the device component of the key under which _operation_device_version_data stores tuned Triton parameters. It was a two-branch chain over torch.cuda and torch.xpu returning "" for anything else, and it took no arguments -- so the key named whatever the process was configured for, never the device the inputs were actually on. An accelerator outside that…
Summary assembled by rule from the sources below