Diffusers 0.38.0 adds Ace-Step 1.5 pipelines and Flash Attention 4 support

• Hugging Face shipped Diffusers 0.38.0, and RisingSayak's release thread says the headline additions are new audio and image pipelines plus several backend and library upgrades.
• The new model coverage in RisingSayak's release thread includes Ace-Step 1.5, LongCat-AudioDiT, and Ernie-Image, with the full changelog linked in RisingSayak's release-notes post and published on the v0.38.0 GitHub release page.
• Runtime support expanded with Flash Attention 4, FlashPack loading, and Ring Anything as a backend for context parallelism, according to RisingSayak's release thread.
• The release also adds an example for profiling a DiffusionPipeline, which RisingSayak's release thread framed as a way to find performance wins during adoption.

You can jump straight from RisingSayak's release-notes post to the full GitHub release notes, and RisingSayak's release thread packs most of the useful nouns into one short list: new audio pipelines, Flash Attention 4, FlashPack, and Ring Anything. Even RisingSayak's follow-up post was basically an instruction to go read the changelog twice.

Diffusers 0.38.0 looks like a model-coverage release first. The items called out in the launch thread span audio and image generation rather than a single modality.

• Ace-Step 1.5
• LongCat-AudioDiT
• Ernie-Image

That is a useful signal for Diffusers users because the library keeps widening from image-centric diffusion plumbing into a broader set of generation pipelines, including audio, as RisingSayak's release thread explicitly notes.

The infrastructure side of the release is tighter than the version number suggests. RisingSayak's feature list groups three additions together:

• Flash Attention 4 support
• Loading with FlashPack
• Ring Anything as a new backend for context parallelism

Those are not end-user model names. They are library-level hooks for faster attention, loading, and distributed execution, the kind of release-note items that matter more to people running larger diffusion workloads than to people just trying one new checkpoint. The official details live in the GitHub release notes.

The last item in the launch thread is easy to miss: Diffusers 0.38.0 adds an example showing how to profile a DiffusionPipeline and hunt for performance bottlenecks.

That is new information, not just packaging. Library releases often add support for more models and kernels, but RisingSayak's performance note also points readers toward a workflow for measuring where their own pipeline is slow, with the linked release notes as the place to inspect the implementation details.