cache-dit-cu13 1.3.12

Cache-DiT: A PyTorch-native Inference Engine with Cache, Parallelism, Quantization and CPU Offload for DiTs.

⚡️🎉A PyTorch-native Inference Engine with Cache,
Parallelism, Quantization and CPU Offload for DiTs

**🤗Why Cache-DiT❓❓**Cache-DiT is built on top of the 🤗Diffusers library and now supports nearly ALL DiTs from Diffusers. It provides hybrid cache acceleration (DBCache, TaylorSeer, SCM, etc.) and comprehensive parallelism optimizations, including Context Parallelism, Tensor Parallelism, hybrid 2D or 3D parallelism, and dedicated extra parallelism support for Text Encoder, VAE, and ControlNet.

Cache-DiT is compatible with compilation, CPU Offloading, and quantization, fully integrates with SGLang Diffusion, vLLM-Omni, TensorRT-LLM, ComfyUI, and runs natively on NVIDIA GPUs, Ascend NPUs and AMD GPUs. Cache-DiT is fast, easy to use, and flexible for various DiTs (online docs at 📘cache-dit.io).

⚡️9x speedup by Cache-DiT with Cache, Context Parallelism and Compilation

🚀Quick Start: Cache, Parallelism and Quantization

First, you can install the cache-dit from PyPI or install from source:

uv pip install -U cache-dit # PyPI, stable release.
uv pip install git+https://github.com/vipshop/cache-dit.git # latest.

Then, try to accelerate your DiTs with just ♥️one line♥️ of code ~

>>> import cache_dit
>>> from diffusers import DiffusionPipeline
>>> pipe = DiffusionPipeline.from_pretrained(...).to("cuda")
>>> cache_dit.enable_cache(pipe) # Cache Acceleration with One-line code.
>>> from cache_dit import DBCacheConfig, ParallelismConfig
>>> cache_dit.enable_cache( # Or, Hybrid Cache Acceleration + Parallelism.
...   pipe, cache_config=DBCacheConfig(), # w/ default
...   parallelism_config=ParallelismConfig(ulysses_size=2))
>>> from cache_dit import DBCacheConfig, ParallelismConfig, QuantizeConfig
>>> cache_dit.enable_cache( # Or, Hybrid Cache + Parallelism + Quantization.
...   pipe, cache_config=DBCacheConfig(), # w/ default
...   parallelism_config=ParallelismConfig(ulysses_size=2),
...   quantize_config=QuantizeConfig(quant_type=...))
>>> output = pipe(...) # Then, just call the pipe as normal.

🚀Quick Start: SVDQuant (W4A4) PTQ/DQ workflow

First, install Cache-DiT with SVDQuant support (Experimental):

# Required: CUDA 13.0+, PyTorch 2.11+, Ubuntu 22.04+.
uv pip install -U cache-dit-cu13 # PyPI, stable release.
CACHE_DIT_BUILD_SVDQUANT=1 uv pip install -e ".[quantization]" # latest.

Then, try to quantize your model with just ♥️a few lines♥️ of code ~

>>> from cache_dit import QuantizeConfig
>>> pipe = DiffusionPipeline.from_pretrained(...).to("cuda")
>>> # DQ: "...{dtype}_r{rank}_dq", PTQ: "...{dtype}_r{rank}"
>>> pipe.transformer = cache_dit.quantize(
...   pipe.transformer, quant_config=QuantizeConfig(
...   # ✅ NO calibration needed for SVDQ DQ in Cache-DiT! 🎉
...   quant_type="svdq_{int4|nvfp4}_r{32|64|128|256|...}_dq",
...   svdq_kwargs={"smooth_strategy": "few_shot"})) 
>>> output = pipe(...) # Then, just call the pipe as normal.

🚀Quick Start: Bucket-style Layerwise CPU Offload

Bucket-style Layerwise Offload w/ nearly zero (<5%🎉) latency overhead ~

>>> import cache_dit
>>> cache_dit.layerwise_offload(
...   pipe, # nn.Module: pipe, transformer, text_encoder, etc.
...   onload_device="cuda",
...   offload_device="cpu",
...   async_transfer=True,
...   transfer_buckets=4,
...   persistent_buckets=64,
...   persistent_bins=8,
...   prefetch_limit=True,
...   max_copy_streams=4,
...   max_inflight_prefetch_bytes="8gib")
>>> output = pipe(...) # Then, just call the pipe as normal.

For more advanced features, please refer to our online documentation at 📘cache-dit.io.

🌐Community Integration

• 🎉ComfyUI x Cache-DiT
• 🎉(Intel) llm-scaler x Cache-DiT
• 🎉Diffusers x Cache-DiT
• 🎉TensorRT-LLM x Cache-DiT
• 🎉SGLang Diffusion x Cache-DiT
• 🎉vLLM-Omni x Cache-DiT
• 🎉Nunchaku x Cache-DiT
• 🎉SD.Next x Cache-DiT
• 🎉stable-diffusion.cpp x Cache-DiT
• 🎉jetson-containers x Cache-DiT

©️Acknowledgements

Special thanks to vipshop's Computer Vision AI Team for supporting testing and deployment of this project. We learned and reused codes from: Diffusers, SGLang, vLLM-Omni, Nunchaku, xDiT and TaylorSeer.

©️Citations

@misc{cache-dit@2025,
  title={Cache-DiT: A PyTorch-native Inference Engine with Cache, Parallelism, Quantization and CPU Offload for DiTs.},
  url={https://github.com/vipshop/cache-dit.git},
  note={Open-source software available at https://github.com/vipshop/cache-dit.git},
  author={DefTruth, vipshop.com, etc.},
  year={2025}
}