trunk/3a96a1cabff36a3fd760468c7ac7ed095ca6561e: [Test] Make TestSWAUtils device-generic in test_swa_utils.py (#183874)

Summary

Convert test/optim/test_swa_utils.py from manual torch.cuda.is_available() guards to instantiate_device_type_tests so SWA utility tests run on all registered device backends (CPU, CUDA, MPS, OpenReg, XPU).

• Split test_averaged_model_all_devices into test_averaged_model_same_device (device-generic) and test_averaged_model_cross_device (@onlyAccelerator)


• Gate test_averaged_model_mixed_device with @onlyAccelerator, replace .cuda() with .to(device)


• Refactor _test_update_bn helper: replace cuda boolean with device parameter


• Make test_update_bn_dnn/test_update_bn_cnn accept device, use .to(device) for model placement


• Add @onlyCPU to 4 tests that don't need multi-device coverage (state_dict, picklable, exponential, eval_momentum)


• Add instantiate_device_type_tests(TestSWAUtils, globals(), allow_mps=True) in test_optim.py

Test plan

• Full SWA test suite passed on CUDA: 25 passed, 11 skipped, 0 failed


• test_averaged_model_same_device runs and passes on both CPU and CUDA


• test_averaged_model_cross_device runs on CUDA, correctly skips on CPU


• test_averaged_model_mixed_device runs on CUDA, correctly skips on CPU


• test_update_bn_dnn / test_update_bn_cnn run on both CPU and CUDA


• CPU-only tests (@onlyCPU) correctly skip on CUDA

Pull Request resolved: #183874

Approved by: https://github.com/fffrog, https://github.com/jbschlosser