`apply_group_offloading(use_stream=True)` raises on Ascend NPU: no stream branch for `torch.npu`

[li-lizhe]

Describe the bug

apply_group_offloading(..., use_stream=True) in src/diffusers/hooks/group_offloading.py (~line 675) only builds a stream for CUDA (torch.cuda.Stream()) and Intel XPU (torch.Stream()). On every other accelerator it raises:

ValueError: Using streams for data transfer requires a CUDA device, or an Intel XPU device.

Ascend NPU exposes torch.npu.Stream() through torch_npu, so stream-based group offloading is implementable there — but the dispatch never reaches it, making the feature unusable on NPU even though the backend does support streams.

Reproduction

import torch, torch_npu  # Ascend 910B
from diffusers.hooks import apply_group_offloading

print(torch.cuda.is_available(), torch.xpu.is_available(), torch.npu.is_available())
# False False True

apply_group_offloading(module, onload_device=torch.device("npu"), use_stream=True)
# ValueError: Using streams for data transfer requires a CUDA device, or an Intel XPU device.

Environment

• diffusers main
• Ascend 910B NPU, torch 2.14 + torch_npu
• torch.cuda.is_available() == False, torch.xpu.is_available() == False, torch.npu.is_available() == True

Expected behavior

Stream-based onload/offload works on any accelerator that exposes a stream type, or the error message lists the accelerator that is missing.

Fix

Add an npu branch mirroring the existing CUDA/XPU checks and mention Ascend NPU in the error message. PR: #14785