Automatically reusing components

未关闭
#15,248 0 条评论 1 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
5/5
预计耗时
一周以上
新手友好度
35/100
Issue 类型
功能
描述清晰度
描述清楚
活跃度
活跃
技术栈
huggingface, python

调研方向

Read the ComponentsManager guide and the Pika H3 Flex example linked in the issue to understand how components are currently shared. The proposal is to discuss opt-in reuse based on checkpoint weight hashes; this issue explicitly asks for feedback and use cases rather than a code change. Share a concrete concern, alternative, or experience with manual component reuse.

由索引模型根据 Issue 内容生成。

描述

idea-contribution-welcome

Diffusers pipelines often share components: the same text encoder or VAE across model families, a finetune reusing its base model's components, a LoRA-based model reusing the base transformer entirely. Loading each of them twice wastes time, and worse, it keeps two copies of the same model in memory. So we want to reuse the same components as much as possible.

What we have today

In Modular Diffusers, ComponentsManager manages components across pipelines. Load one pipeline through it, and a second pipeline can pick up what it needs from the manager instead of loading its own copy:

pika.update_components(transformer=manager.get_one(name="transformer", collection="minimax-h3"))

See the ComponentsManager guide and the Pika H3 Flex example, where a second pipeline is assembled from a loaded MiniMax-H3.

The reuse itself is still manual: you decide which components are shared and pass them over. What the manager gives you is bookkeeping so you can keep track of it. Each component is listed once with the pipelines that use it (when you register under a collection), offloading works across all of them, and the manager warns when two components were loaded from the same repo and subfolder. That warning is as far as duplicate detection goes today, and it comes after the second copy is already loaded.

The Hub now dedups on disk

With huggingface_hub>=1.32.0, identical Xet-backed files share one cached payload across repositories, shard by shard. Download the same text encoder from two repos and it lands on disk once. We could take advantage of that and go one step further: dedup the models we load into memory too.

Proposal

An opt-in setting on the manager. When a pipeline is about to load a component, the manager checks whether the same weights are already loaded, using the same idea the Hub uses to dedup: a content hash of the weights, stored with the checkpoint and compared before loading. If there is a match, the pipeline reuses the loaded component instead of loading again. No download, no second copy.

How to contribute

This issue carries the new idea-contribution-welcome label. For issues with this label we are not looking for a PR. We are looking for discussion: feedback on the proposal, alternatives, use cases we have not thought of, problems you ran into with the current manual way. We will write the PR ourselves, and anyone whose feedback shaped it will be added as a co-author.


Posted by Claude on behalf of @yiyixuxu. The text was drafted together and approved by yiyixuxu.

主要语言
Python
星标
34.9k
派生
7.5k
平均合并
3 天 16 小时
30 天内合并 PR
65

环境准备

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

huggingface/diffusers 的其他 Issue

查看 huggingface/diffusers 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。