[dist] XPU distributed backend selection (XCCL / oneCCL) is unverified — needs XPU hardware #3777

Context

#3771 binds distributed training to the active accelerator instead of assuming CUDA. Writing that change surfaced a second, independent defect on XPU.

get_default_dist_backend() returns "nccl" for every non-NPU device. Since get_accelerator_device_type() does select XPU, an XPU run reaches init_process_group(backend="nccl") and cannot start at all on a build without NCCL — the same class of failure #3771 fixes for Ascend NPU.

Why this is not in #3771

The correct backend for XPU is XCCL (oneCCL; the legacy name on older builds is "ccl"), so the fix is small:

_ACCELERATOR_DIST_BACKENDS = {"npu": ("hccl",), "xpu": ("xccl", "ccl")}

but it advertises a runtime path that nobody here can exercise. We have no Intel XPU hardware in our environment, so under the repository's hardware-validation requirement the XPU arm was the only path #3771 changed without real-device evidence. #3771 therefore keeps the historical "nccl" value for XPU — i.e. exactly the pre-PR behaviour, unchanged — and the selection is tracked here for whoever has the hardware.

What a fix needs

1. Add the XPU entry above (keep "ccl" only for builds that positively confirm it through torch.distributed.is_ccl_available()).
2. Run an exact-head 2-rank torchrun on Intel XPU with XCCL, showing per-rank xpu:0 / xpu:1 binding, an all_reduce, DDP forward/backward, the end-of-epoch cache release, plus the exact PyTorch / oneAPI / device versions.
3. Update tests/test_distributed_device_setup.py accordingly — the current test pins the narrowed behaviour (XPU stays "nccl" even when a build reports is_xccl_available()), so it needs to be flipped.

Related

• fix(train): bind distributed training to the active accelerator #3771 — the Ascend NPU path, verified end to end on 910B2 hardware.
• Failure mode this follows up on: init_process_group(backend="nccl") -> RuntimeError: Distributed package doesn't have NCCL built in.