[Feature Request] Add DiCache for sample-adaptive caching in FLUX transformers

未關閉
#15,134 0 則留言 0 個 reaction 已指派 0 人 在 GitHub 檢視

維護者通常 1 天內回覆

還沒有人認領這個 Issue。

評估

難度
5/5
預估耗時
一週以上
新手友好度
35/100
Issue 類型
功能
描述清晰度
基本清楚
活躍度
活躍
技術堆疊
python, pytorch

研究方向

Start by reading Diffusers’ existing caching API, cache-hook and state-management architecture, then inspect the FLUX transformer integration and its cache tests. The issue proposes adding a compatible DiCacheConfig with sample-adaptive probe decisions and trajectory-aligned residual reuse. Done means a FLUX-first integration without pipeline-specific monkey-patching, with regression tests and documentation; the issue provides no file paths or test names.

由索引模型根據 Issue 內容生成。

描述

Diffusers already supports caching techniques such as FirstBlockCache, but additional speed–quality trade-offs may be possible through sample-adaptive cache decisions and trajectory-aligned residual reuse.

I would like to propose integrating [DiCache (ICLR 2026)](https://proceedings.iclr.cc/paper_files/paper/2026/hash/78288ef33b18a351c3cd679dc9a15c8d-Abstract-Conference.html), which uses shallow transformer probes to determine whether a full forward pass is necessary and applies trajectory-aligned reuse when skipping computation.

Describe the solution you'd like.

Add a DiCacheConfig compatible with Diffusers' existing caching API:

from diffusers import DiCacheConfig

pipe.transformer.enable_cache(
    DiCacheConfig(threshold=0.2, probe_depth=1)
)

# Generate images

pipe.transformer.disable_cache()

This is a proposed API, not an existing implementation. I propose an initial FLUX integration using the existing cache-hook and state-management architecture, without pipeline-specific monkey-patching.

Describe alternatives you've considered.

I benchmarked DiCache against Diffusers' FirstBlockCache and two DiCache ablations on pretrained FLUX.1-dev.

Method Speedup SSIM ↑ LPIPS ↓
FirstBlockCache (0.05) 1.23× 0.9851 0.0107
DiCache (0.2) 2.14× 0.9419 0.0515
DiCache (0.3) 2.42× 0.9102 0.0834
FirstBlockCache (0.2) 2.69× 0.8779 0.1301
Fixed-schedule DiCache 2.12× 0.9348 0.0557
Last-residual-only DiCache 2.13× 0.9274 0.0686

These results suggest that DiCache provides an additional operating region in the speed–fidelity trade-off. They do not establish superiority over a fully threshold-tuned FirstBlockCache Pareto frontier.

Additional context.

Experimental setup:

  • Model: aniketppanchal/flux.1-dev-nf4-pkg, revision dde39613d8f3d7abe59d7c59ff16838b14bcebf8
  • Hardware: Tesla T4
  • Resolution: 512×512, 30 denoising steps
  • Evaluation: 10 prompts × 3 seeds × 7 variants, with 2 timing repetitions each
  • Versions: Diffusers 0.35.1, Transformers 4.56.1
  • Measurements: denoising-only runtime, SSIM, PSNR, LPIPS, CLIP alignment, latent error, and GPU memory

Validation results:

All 210 method–prompt–seed configurations completed with zero recorded failures. Cache-disabled inference achieved exact tensor parity with native FLUX inference. Deterministic cache-reset and interleaved-generation tests also passed.

Per-timestep instrumentation demonstrated sample-dependent scheduling: DiCache produced 6 distinct cache schedules at threshold 0.2 and 3 at threshold 0.3 across 30 trajectories, compared with 1 schedule for the fixed-schedule control.

Ablations indicated improvements from both adaptive decisions and residual-history reuse.

Limitations: The experiments use an NF4-quantized checkpoint, not the paper's original full-precision setup. Fidelity can vary significantly between samples, with minimum SSIM values of 0.8120 and 0.7362 for DiCache thresholds 0.2 and 0.3, respectively. Additional matched-fidelity comparisons and broader model coverage would be valuable.

References:

I have reproducible benchmark scripts, paired images, cache trajectories, and validation records available.

I can prepare a minimal FLUX-first integration with the existing Diffusers caching infrastructure, together with regression tests and documentation. I would be happy to align the implementation with the maintainers' preferred design.

Would the maintainers be interested in this contribution? Any guidance on the preferred integration approach would be greatly appreciated.

If approved, I am willing to implement this feature and submit a pull request. I would also request this issue be assigned to me if possible.

主要語言
Python
星號
34.9k
分支
7.5k
平均合併
3 天 16 小時
30 天內合併 PR
65

環境準備

從這裡開始

  1. 先讀完整個 Issue,再讀專案的貢獻指南。
  2. 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
  3. Fork 儲存庫,在一個分支上完成修改。
  4. 送出 Pull Request,並在描述裡引用這個 Issue 編號。

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 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。