Magcache support for Diffuser

[quic-amitraj]

Summary

This PR adds MagCache support for WAN text-to-video in QEfficient Diffusers.

MagCache is a training-free, inference-time caching strategy that uses magnitude-aware dynamics to decide when cached denoising residuals can be reused, reducing expensive model calls while preserving generation quality.

In our WAN setup, we observe around ~30% inference-time reduction for 180p video generation.

Without cache 180p video-

180p_without_cache.mp4

With Cache 180p Video-

180p_with_magcache.mp4

Design Notes

• No training required.
• No ONNX/QPC signature change for MagCache path.
• Default behavior remains unchanged when use_magcache=False.

Performance Note

• WAN text-to-video at 180p shows ~30% reduction in inference time in our measurements.

References

• Paper: MagCache: Fast Video Generation with Magnitude-Aware Cache
  https://arxiv.org/abs/2506.09045
• Official repository:
  https://github.com/Zehong-Ma/MagCache
• Huggingface repo support:
  https://github.com/huggingface/diffusers/blob/main/docs/source/en/optimization/cache.md

[quic-amitraj]

Without cache 180p video-

180p_without_cache.mp4

With Cache 180p Video-

180p_with_magcache.mp4