WanAnimatePipeline.get_i2v_mask() defaults device to "cuda", which raises on non-CUDA accelerators (NPU/XPU/MPS)

[li-lizhe]

Describe the bug

WanAnimatePipeline.get_i2v_mask() (src/diffusers/pipelines/wan/pipeline_wan_animate.py, line ~471) declares its device parameter with a hard-coded CUDA default:

def get_i2v_mask(
    self,
    batch_size: int,
    latent_t: int,
    latent_h: int,
    latent_w: int,
    mask_len: int = 1,
    mask_pixel_values: torch.Tensor | None = None,
    dtype: torch.dtype | None = None,
    device: str | torch.device = "cuda",   # <-- here
) -> torch.Tensor:

The mask is allocated with torch.zeros(..., device=device) / .to(device=device), so any call that does not pass device explicitly allocates on CUDA. On builds without CUDA (Ascend NPU via torch_npu, Intel XPU, Apple MPS) that raises immediately:

AssertionError: Torch not compiled with CUDA enabled

Every sibling helper on the same pipeline uses the device-agnostic convention instead (prepare_reference_image_latents, encode_image, ... carry device: torch.device | None = None and resolve it with device = device or self._execution_device), so the "cuda" default is inconsistent with the rest of the file and is a trap for any device-agnostic caller.

Reproduction

On a non-CUDA accelerator (verified on Ascend 910B2, torch 2.14.0a0/2.15.0.dev+cpu build with torch_npu):

import torch

torch.zeros(1, device="cuda")
# AssertionError: Torch not compiled with CUDA enabled

Calling the pipeline method without an explicit device hits the same path:

pipe.get_i2v_mask(batch_size=1, latent_t=1, latent_h=8, latent_w=8)

Expected behaviour

The default should resolve to the pipeline's execution device instead of assuming CUDA, matching the convention already used by the neighbouring helpers: device: str | torch.device | None = None plus device = device or self._execution_device at the top of the method.

Scope (for transparency)

The two in-tree call sites in the same file (prepare_reference_image_latents, prepare_prev_segment_cond_latents) both pass device explicitly, so nothing in the current in-tree path crashes today - this is a latent/API-hygiene defect for external callers and for future call sites. The module-level helper get_i2v_mask(lat_t, lat_h, lat_w, mask_len=1, device="cuda") in src/diffusers/modular_pipelines/wan_animate_2/encoders.py carries the same CUDA default; happy to cover it in the linked PR or in a follow-up, whichever the maintainers prefer.

System Info

• diffusers: main
• hardware: Ascend 910B2 NPU (torch 2.14.0a0 / 2.15.0.dev20260917+cpu + torch_npu); the same failure applies to any CUDA-less build (XPU, MPS).