Skip to content

Revisit modular diffusers weight-duplication workaround now that huggingface_hub has a shared blobs cache #14802

Description

@moon-bot-app

huggingface_hub v1.32.0 ships a shared blobs cache: Xet-backed files are now deduplicated locally across repos (best-effort, via symlinks, backward compatible with older versions and opt-out possible). See the release notes and huggingface/huggingface_hub#4498.

Concretely, downloading two repos that share components now only pays the download cost once — e.g. Qwen/Qwen-Image and Qwen/Qwen-Image-Edit share ~29% of their data (16.8G/57.7G), FLUX.1-dev / FLUX.1-schnell ~18% (10.3GB), Kwai-Kolors/Kolors / Kolors-Inpainting ~56% (>20GB).

This is relevant to modular diffusers repositories, which duplicate weights from the original repos. We currently have a workaround so that weights are fetched from the original repositories instead of the modular repo, precisely because cross-repo local dedup was missing. With the shared blobs cache in place, that workaround may no longer be necessary — modular repos would be a particularly striking use case for the new dedup (near-zero incremental download for duplicated weights).

Proposed follow-ups:

  • Verify dedup actually kicks in for modular diffusers repos (Xet-enabled, identical blobs) and measure the gain.
  • If confirmed, plan removal/simplification of the "download from original repo" workaround, bumping the huggingface_hub floor as needed and keeping a fallback for older versions.

Requested by Sayak Paul - Slack thread - Agent trace

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions