[UT]XCCL remains the default backend for XPU

[Chao1Han]

This test is designed to verify that XCCL remains the default backend for XPU, even when other groups are registered as optional backends for XPU.

[Copilot]

Pull Request Overview

Adds a new unit test to confirm that XCCL remains the default distributed backend on XPU even after another backend is registered.

• Introduces test_xccl_priority to register a dummy backend and run an all-reduce call without specifying a backend.
• Leverages existing requires_xccl decorator to skip if XCCL isn’t available.

Comments suppressed due to low confidence (2)

test/xpu/distributed/test_c10d_xccl.py:568

• The test currently only invokes all_reduce but doesn't assert that the default backend is actually XCCL. Consider retrieving the process group (e.g., via dist.distributed_c10d._get_default_group()) and asserting its type or backend name to ensure the priority behavior is verified.

dist.all_reduce(a)

test/xpu/distributed/test_c10d_xccl.py:555

• [nitpick] The test name test_xccl_priority is a bit generic. Consider renaming to test_default_backend_is_xccl_when_fake_registered for clarity on what scenario is covered.

def test_xccl_priority(self):

[Copilot]

After registering the fake backend, consider unregistering it in a finally block or teardown step to avoid side effects on other tests.

Suggested change

	dist.Backend.register_backend(
	"fake",
	lambda store, rank, size, timeout: dist.ProcessGroup(rank, size),
	devices=["xpu"],
	)
	store = dist.FileStore(self.file_name, self.world_size)
	dist.init_process_group(
	world_size=self.world_size,
	rank=self.rank,
	store=store,
	)
	a = torch.randn(2, device="xpu")
	dist.all_reduce(a)
	try:
	dist.Backend.register_backend(
	"fake",
	lambda store, rank, size, timeout: dist.ProcessGroup(rank, size),
	devices=["xpu"],
	)
	store = dist.FileStore(self.file_name, self.world_size)
	dist.init_process_group(
	world_size=self.world_size,
	rank=self.rank,
	store=store,
	)
	a = torch.randn(2, device="xpu")
	dist.all_reduce(a)
	finally:
	dist.Backend.unregister_backend("fake")

[Chao1Han]

Other cases explicit init with backend xccl, it is safely no unregister.

[Chao1Han]

Close due to pytorch/pytorch#155320 merged