Support separate placement for encoding and decoding with tensor parallelism

未关闭
#15,336 1 条评论 0 个 reaction 已指派 1 人 在 GitHub 查看

维护者通常 1 天内回复

@DN6 已经在做这个了。

开始于 2026年11月18日。

评估

这个 Issue 还没有评估数据。

描述

The tensor-parallel inference example from #13718 loads and runs the text encoders and VAE on every rank. Each rank encodes the same prompt and decodes the same final latents. For FLUX.1-schnell, those components account for about 9.3 GiB of bf16 weights per rank.

Could Diffusers support running encoding and decoding once per request, with configurable placement for those stages?

  • Encode the prompt and any input images once, then send the conditioning tensors to the TP group.
  • Run denoising across the TP group.
  • Decode and postprocess the final latents once.

Keeping the encoders and VAE on a single TP rank would reduce total memory use and redundant computation. If those weights stay resident during denoising, that rank still has less room for its denoiser shard. With equal-sized shards, the rank with the least free memory determines what fits across the group. To free memory on every TP rank during denoising, the pipeline could place encoding and decoding on a separate device or offload their weights between stages.

This came up while testing on Trainium, but the placement is hardware-independent. FLUX already accepts precomputed prompt embeddings and can return latents. Pipeline-level support would let callers use those stage boundaries without wiring them up separately for each pipeline.

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。