Skip to content
/ cache-dit Public
forked from vipshop/cache-dit

An Unified Cache Acceleration Toolbox for DiTs: Qwen-Image, Qwen-Image-Edit, FLUX.1, Wan2.1, etc.♥️

License

View license
4 stars 8 forks Branches Tags Activity
Star
Notifications You must be signed in to change notification settings

xlite-dev/cache-dit

BranchesTags
 
 

Repository files navigation

An Unified and Training-free Cache Acceleration Toolbox for Diffusion Transformers
♥️ Cache Acceleration with One-line Code ~ ♥️

🔥Unified Cache APIs | DBCache | Hybrid TaylorSeer | Hybrid Cache CFG🔥

🎉Now, cache-dit covers Most mainstream Diffusers' Pipelines🎉
🔥Qwen-Image | FLUX.1 | Wan 2.1/2.2 | ... | CogVideoX🔥

🔥News

Previous News

📖Contents

⚙️Installation

You can install the stable release of cache-dit from PyPI:

pip3 install -U cache-dit

Or you can install the latest develop version from GitHub:

pip3 install git+https://github.com/vipshop/cache-dit.git

🔥Supported Models

Currently, cache-dit library supports almost Any Diffusion Transformers (with Transformer Blocks that match the specific Input and Output patterns). Please check 🎉Unified Cache APIs for more details. Here are just some of the tested models listed:

More Pipelines

🎉Unified Cache APIs

📚Forward Pattern Matching

Currently, for any Diffusion models with Transformer Blocks that match the specific Input/Output patterns, we can use the Unified Cache APIs from cache-dit, namely, the cache_dit.enable_cache(...) API. The Unified Cache APIs are currently in the experimental phase; please stay tuned for updates. The supported patterns are listed as follows:

♥️Cache Acceleration with One-line Code

In most cases, you only need to call one-line of code, that is cache_dit.enable_cache(...). After this API is called, you just need to call the pipe as normal. The pipe param can be any Diffusion Pipeline. Please refer to Qwen-Image as an example.

import cache_dit
from diffusers import DiffusionPipeline 

# Can be any diffusion pipeline
pipe = DiffusionPipeline.from_pretrained("Qwen/Qwen-Image")

# One-line code with default cache options.
cache_dit.enable_cache(pipe) 

# Just call the pipe as normal.
output = pipe(...)

🔥Automatic Block Adapter

But in some cases, you may have a modified Diffusion Pipeline or Transformer that is not located in the diffusers library or not officially supported by cache-dit at this time. The BlockAdapter can help you solve this problems. Please refer to Qwen-Image w/ BlockAdapter as an example.

from cache_dit import ForwardPattern, BlockAdapter

# Use BlockAdapter with `auto` mode.
cache_dit.enable_cache(
    BlockAdapter(pipe=pipe, auto=True), # Qwen-Image, etc.  
    # Check `📚Forward Pattern Matching` documentation and hack the code of
    # of Qwen-Image, you will find that it has satisfied `FORWARD_PATTERN_1`.
    forward_pattern=ForwardPattern.Pattern_1,  
)

# Or, manually setup transformer configurations.
cache_dit.enable_cache(
    BlockAdapter(
        pipe=pipe, # Qwen-Image, etc.
        transformer=pipe.transformer,
        blocks=pipe.transformer.transformer_blocks,
        blocks_name="transformer_blocks",
    ), 
    forward_pattern=ForwardPattern.Pattern_1,
)

For such situations, BlockAdapter can help you quickly apply various cache acceleration features to your own Diffusion Pipelines and Transformers. Please check the 📚BlockAdapter.md for more details.

🤖Cache Acceleration Stats Summary

After finishing each inference of pipe(...), you can call the cache_dit.summary() API on pipe to get the details of the Cache Acceleration Stats for the current inference.

stats = cache_dit.summary(pipe)

You can set details param as True to show more details of cache stats. (markdown table format) Sometimes, this may help you analyze what values of the residual diff threshold would be better.

⚡️Cache Steps and Residual Diffs Statistics: QwenImagePipeline

| Cache Steps | Diffs Min | Diffs P25 | Diffs P50 | Diffs P75 | Diffs P95 | Diffs Max |
|-------------|-----------|-----------|-----------|-----------|-----------|-----------|
| 23          | 0.045     | 0.084     | 0.114     | 0.147     | 0.241     | 0.297     |

⚡️DBCache: Dual Block Cache

DBCache: Dual Block Caching for Diffusion Transformers. Different configurations of compute blocks (F8B12, etc.) can be customized in DBCache, enabling a balanced trade-off between performance and precision. Moreover, it can be entirely training-free. Please check DBCache.md docs for more design details.

  • Fn: Specifies that DBCache uses the first n Transformer blocks to fit the information at time step t, enabling the calculation of a more stable L1 diff and delivering more accurate information to subsequent blocks.
  • Bn: Further fuses approximate information in the last n Transformer blocks to enhance prediction accuracy. These blocks act as an auto-scaler for approximate hidden states that use residual cache.
import cache_dit
from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    torch_dtype=torch.bfloat16,
).to("cuda")

# Default options, F8B0, 8 warmup steps, and unlimited cached 
# steps for good balance between performance and precision
cache_dit.enable_cache(pipe)

# Custom options, F8B8, higher precision
cache_dit.enable_cache(
    pipe,
    max_warmup_steps=8,  # steps do not cache
    max_cached_steps=-1, # -1 means no limit
    Fn_compute_blocks=8, # Fn, F8, etc.
    Bn_compute_blocks=8, # Bn, B8, etc.
    residual_diff_threshold=0.12,
)

Moreover, users configuring higher Bn values (e.g., F8B16) while aiming to maintain good performance can specify Bn_compute_blocks_ids to work with Bn. DBCache will only compute the specified blocks, with the remaining estimated using the previous step's residual cache.

# Custom options, F8B16, higher precision with good performance.
cache_dit.enable_cache(
    pipe,
    Fn_compute_blocks=8,  # Fn, F8, etc.
    Bn_compute_blocks=16, # Bn, B16, etc.
    # 0, 2, 4, ..., 14, 15, etc. [0,16)
    Bn_compute_blocks_ids=cache_dit.block_range(0, 16, 2),
    # If the L1 difference is below this threshold, skip Bn blocks 
    # not in `Bn_compute_blocks_ids`(1, 3,..., etc), Otherwise, 
    # compute these blocks.
    non_compute_blocks_diff_threshold=0.08,
)

DBCache, L20x1 , Steps: 28, "A cat holding a sign that says hello world with complex background"

Baseline(L20x1) F1B0 (0.08) F1B0 (0.20) F8B8 (0.15) F12B12 (0.20) F16B16 (0.20)
24.85s 15.59s 8.58s 15.41s 15.11s 17.74s

🔥Hybrid TaylorSeer

We have supported the TaylorSeers: From Reusing to Forecasting: Accelerating Diffusion Models with TaylorSeers algorithm to further improve the precision of DBCache in cases where the cached steps are large, namely, Hybrid TaylorSeer + DBCache. At timesteps with significant intervals, the feature similarity in diffusion models decreases substantially, significantly harming the generation quality.

$$ \mathcal{F}_{\text {pred }, m}\left(x_{t-k}^l\right)=\mathcal{F}\left(x_t^l\right)+\sum_{i=1}^m \frac{\Delta^i \mathcal{F}\left(x_t^l\right)}{i!\cdot N^i}(-k)^i $$

TaylorSeer employs a differential method to approximate the higher-order derivatives of features and predict features in future timesteps with Taylor series expansion. The TaylorSeer implemented in cache-dit supports both hidden states and residual cache types. That is $\mathcal{F}_{\text {pred }, m}\left(x_{t-k}^l\right)$ can be a residual cache or a hidden-state cache.

cache_dit.enable_cache(
    pipe,
    enable_taylorseer=True,
    enable_encoder_taylorseer=True,
    # Taylorseer cache type cache be hidden_states or residual.
    taylorseer_cache_type="residual",
    # Higher values of n_derivatives will lead to longer 
    # computation time but may improve precision significantly.
    taylorseer_kwargs={
        "n_derivatives": 2, # default is 2.
    },
    max_warmup_steps=3, # prefer: >= n_derivatives + 1
    residual_diff_threshold=0.12
)

Important

Please note that if you have used TaylorSeer as the calibrator for approximate hidden states, the Bn param of DBCache can be set to 0. In essence, DBCache's Bn is also act as a calibrator, so you can choose either Bn > 0 or TaylorSeer. We recommend using the configuration scheme of TaylorSeer + DBCache FnB0.

DBCache F1B0 + TaylorSeer, L20x1, Steps: 28,
"A cat holding a sign that says hello world with complex background"

Baseline(L20x1) F1B0 (0.12) +TaylorSeer F1B0 (0.15) +TaylorSeer +compile
24.85s 12.85s 12.86s 10.27s 10.28s 8.48s

⚡️Hybrid Cache CFG

cache-dit supports caching for CFG (classifier-free guidance). For models that fuse CFG and non-CFG into a single forward step, or models that do not include CFG (classifier-free guidance) in the forward step, please set do_separate_cfg param to False (default). Otherwise, set it to True. For examples:

cache_dit.enable_cache(
    pipe, 
    ...,
    # CFG: classifier free guidance or not
    # For model that fused CFG and non-CFG into single forward step,
    # should set do_separate_cfg as False. For example, set it as True 
    # for Wan 2.1/Qwen-Image and set it as False for FLUX.1, HunyuanVideo, 
    # CogVideoX, Mochi, etc.
    do_separate_cfg=True, # Wan 2.1, Qwen-Image
    # Compute cfg forward first or not, default False, namely, 
    # 0, 2, 4, ..., -> non-CFG step; 1, 3, 5, ... -> CFG step.
    cfg_compute_first=False,
    # Compute spearate diff values for CFG and non-CFG step, 
    # default True. If False, we will use the computed diff from 
    # current non-CFG transformer step for current CFG step.
    cfg_diff_compute_separate=True,
)

⚙️Torch Compile

By the way, cache-dit is designed to work compatibly with torch.compile. You can easily use cache-dit with torch.compile to further achieve a better performance. For example:

cache_dit.enable_cache(pipe)

# Compile the Transformer module
pipe.transformer = torch.compile(pipe.transformer)

However, users intending to use cache-dit for DiT with dynamic input shapes should consider increasing the recompile limit of torch._dynamo. Otherwise, the recompile_limit error may be triggered, causing the module to fall back to eager mode.

torch._dynamo.config.recompile_limit = 96  # default is 8
torch._dynamo.config.accumulated_recompile_limit = 2048  # default is 256

Please check bench.py for more details.

🛠Metrics CLI

You can utilize the APIs provided by cache-dit to quickly evaluate the accuracy losses caused by different cache configurations. For example:

from cache_dit.metrics import compute_psnr
from cache_dit.metrics import compute_video_psnr
from cache_dit.metrics import FrechetInceptionDistance  # FID

FID = FrechetInceptionDistance()
image_psnr, n = compute_psnr("true.png", "test.png") # Num: n
image_fid,  n = FID.compute_fid("true_dir", "test_dir")
video_psnr, n = compute_video_psnr("true.mp4", "test.mp4") # Frames: n

Please check test_metrics.py for more details. Or, you can use cache-dit-metrics-cli tool. For examples:

cache-dit-metrics-cli -h  # show usage
# all: PSNR, FID, SSIM, MSE, ..., etc.
cache-dit-metrics-cli all  -i1 true.png -i2 test.png  # image
cache-dit-metrics-cli all  -i1 true_dir -i2 test_dir  # image dir

👋Contribute

How to contribute? Star ⭐️ this repo to support us or check CONTRIBUTE.md.

©️License

The cache-dit codebase is adapted from FBCache. Special thanks to their excellent work! We have followed the original License from FBCache, please check LICENSE for more details.

©️Citations

@misc{cache-dit@2025,
  title={cache-dit: An Unified and Training-free Cache Acceleration Toolbox for Diffusion Transformers},
  url={https://github.com/vipshop/cache-dit.git},
  note={Open-source software available at https://github.com/vipshop/cache-dit.git},
  author={vipshop.com},
  year={2025}
}

About

An Unified Cache Acceleration Toolbox for DiTs: Qwen-Image, Qwen-Image-Edit, FLUX.1, Wan2.1, etc.♥️

Resources

Readme

License

View license
Activity
Custom properties

Stars

4 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%