Enable XPU for blockwise quantization and FSDP tests

[jiqing-feng]

Summary

Replace CUDA-only guards with device-agnostic torch.accelerator APIs so that XPU (and other accelerators) can run the blockwise quantization and FSDP state_dict tests.

Changes

tests/test_functional.py

• Allow XPU to run the full test_dynamic_blockwise_quantization parameter matrix (nested, non-256 blocksizes, fp16/bf16) instead of skipping them as "non-CUDA".

tests/test_linear4bit.py

• Replace torch.cuda.is_available() and NCCL-specific skip guards on test_fsdp_state_dict_save_4bit with a single torch.accelerator.is_available() check.

tests/fsdp_state_dict_save.py

• Auto-detect the accelerator type via torch.accelerator.current_accelerator() and map it to the correct distributed backend (CUDA → nccl, XPU → xccl).
• Replace torch.cuda.set_device() with torch.accelerator.set_device_index() and use the detected device type for model.to().

Hi @matthewdouglas . Would you please review this PR? Thanks!

[matthewdouglas]

Hi @jiqing-feng,

Using torch.accelerator is likely going to break this on older torch versions. I think it was introduced in PyTorch 2.6. This is probably a minimum requirement for XPU (if not even higher), but right now we support PyTorch 2.3+ for CUDA hardware. I will probably soon bump to 2.4, so we would want some backwards compatibility for that.

Otherwise, looks good!

[github-actions]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[jiqing-feng]

Hi @matthewdouglas . I have added backwards compatibility for torch versions without torch.accelerator by falling back to torch.cuda/torch.xpu checks and device setup, while keeping the torch.accelerator path for newer versions.

[jiqing-feng]

The failed tests seems like an env issue. Is this because the change of torch.cuda to torch.accelerator? @matthewdouglas , do you have any suggestions about it?

[matthewdouglas]

@jiqing-feng I think we just need to skip that test on Windows. Otherwise to have a chance at working we'd probably need to switch nccl to gloo and set USE_LIBUV=0. Even then I don't know that it would work, and it's not a use case that's particularly important. We have other tests being skipped when platform.system() == "Windows" too, so it's fine to just do that.

[jiqing-feng]

Hi @matthewdouglas . I have skipped fsdp on Windows for all devices, please check it. Thanks!