MiniMax-H3 video decode only enables fp16 autocast on CUDA, so the block silently runs in fp32 on other accelerators

[li-lizhe]

Describe the bug

MiniMaxH3VideoDecodeStep.__call__ (src/diffusers/modular_pipelines/minimax_h3/decoders.py, line ~187) wraps the VAE decode in an fp16 autocast whose gate is CUDA-only:

with torch.autocast(device_type=device.type, dtype=torch.float16, enabled=device.type == "cuda"):
    video = components.vae.decode(latents, return_dict=False)[0]

The block's own description states the intent - "the decode itself runs under float16 autocast even though the VAE weights are float32" - but with enabled=device.type == "cuda" that is only true on CUDA. On every other accelerator (Ascend NPU, Intel XPU, Apple MPS) autocast is silently disabled, so the same workflow decodes with the VAE's fp32 weights: slower, and numerically on a different path than the documented CUDA behaviour.

Suggested change - gate on "not CPU" instead of "CUDA only":

enabled=device.type != "cpu",

CPU keeps the current behaviour (autocast disabled).

Reproduction

On Ascend 910B2 NPU (torch 2.15.0.dev20260917+cpu + torch_npu), the enabled flag is the only thing keeping the NPU on the fp32 path:

import torch

x = torch.randn(8, 8, device="npu", dtype=torch.float32)
with torch.autocast(device_type="npu", dtype=torch.float16, enabled=True):
    print((x @ x).dtype)   # torch.float16
with torch.autocast(device_type="npu", dtype=torch.float16, enabled=False):
    print((x @ x).dtype)   # torch.float32

Not measured

End-to-end NPU decode speed-up of the change is not measured here - no A/B benchmark of the decode block was run, only the dtype-path check above. CUDA behaviour is unchanged by the linked PR.

Related but distinct

#14746 reports a trade-off in the opposite direction on CUDA (fp32 VAE weights + fp16 decode costs VRAM and forces a downcast). This issue is only about the device gate being CUDA-only; it does not argue about whether the fp16 autocast should exist at all.

System Info

• diffusers: main
• hardware: Ascend 910B2 NPU (torch 2.15.0.dev20260917+cpu + torch_npu); applies to any non-CUDA accelerator.

[li-lizhe]

Resolved upstream by #14754 (commit 51a454b): the fix was to drop the fp16 autocast around the video decode block and run it at the dtype you pass, rather than enabling autocast on more accelerators — the PR's error table shows the previous fp32-weights + fp16-autocast path was the inaccurate one. The premise of this issue no longer holds, so I'm closing it.

[li-lizhe]

Resolved upstream by #14754 (commit 51a454b): the fix was to drop the fp16 autocast around the video decode block and run it at the dtype you pass, rather than enabling autocast on more accelerators — the PR's error table shows the previous fp32-weights + fp16-autocast path was the inaccurate one. The premise of this issue no longer holds, so I'm closing it.