pytorch - 💡(How to fix) Fix Add distributed_backend() hook to DeviceTypeTestBase [1 pull requests]

🚀 The feature, motivation and pitch

Motivation

Distributed tests currently contain many hardcoded mappings between device types and distributed backends, for example:

if "cuda" in device:
    return "nccl"
elif "hpu" in device:
    return "hccl"
elif "xpu" in device:
    return "xccl"
else:
    return "gloo"

These hardcoded mappings block out-of-tree backends from reusing distributed test infrastructure and existing distributed test cases.

Default distributed backend selection is fundamentally a device-type-level property, but today the logic is duplicated across distributed test helpers instead of being provided by DeviceTypeTestBase.

Proposal

Add a default distributed backend hook to DeviceTypeTestBase:

class DeviceTypeTestBase(TestCase):
    @classmethod
    def distributed_backend(cls) -> str:
        import torch.distributed as dist
        return dist.get_default_backend_for_device(cls.device_type)

cc @awgu @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @pragupta @msaroufim @dcci @aditvenk @weifengpy @fffrog

Alternatives

No response

Additional context

No response