OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Infrastructure & Serving

Paraattention

Paraattention is a Python library that accelerates diffusion-transformer inference with context-parallel attention and First Block Cache. Its attention approach supports Ulysses-style and Ring-style parallelism, and its repository includes Diffusers adapters and examples for FLUX, HunyuanVideo, Mochi and CogVideoX.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Paraattention — OpenViST project profile in AI Infrastructure & Serving.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use Paraattention when you are integrating a supported diffusion pipeline and want to test multi-GPU context-parallel attention, First Block Cache, or both as inference optimizations.

Who it is for

  • Developers integrating acceleration into Diffusers pipelines.
  • Teams running diffusion-transformer inference across multiple GPUs.
  • Researchers adapting context-parallel attention to custom models.
  • Practitioners seeking a single-GPU caching option for supported example pipelines.

Documented uses

  • Parallelize attention across GPUs by partitioning activations along the sequence dimension.
  • Apply First Block Cache to skip later transformer-block computations when the first-block residual changes little.
  • Use a unified interface to combine Ulysses-style and Ring-style attention parallelism.
  • Parallelize supported VAE models through a Diffusers adapter.
Before you choose

Requirements & limits

  • The README's installation instructions pin PyTorch to 2.6.0.
  • HunyuanVideo requires Diffusers from its latest master branch according to the README.
  • The README suggests at least 48 GB GPU memory for HunyuanVideo and warns of out-of-memory errors or performance loss with frequent reallocation.
  • VAE parallelization is stated to support AutoencoderKL and AutoencoderKLHunyuanVideo.
  • The reported First Block Cache speedup is up to 2x and depends on the workload.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Install Paraattention from PyPI

Install the library with the PyTorch version specified in the README.

Before you start

  • Python package installer pip3.

Steps

  1. Install PyTorch with the command below.

    Documented command
    pip3 install 'torch==2.6.0'

  2. Install Paraattention with the command below.

    Documented command
    pip3 install para-attn

Expected outcome

The README's PyPI installation commands have been run for PyTorch 2.6.0 and Paraattention.

Guide limitations

  • These installation steps do not configure a model pipeline or run inference.
  • The source does not specify hardware requirements for this installation path.
Documentation 1 ↗
Source-based practical guide

Enable First Block Cache for a single-GPU HunyuanVideo example

Use the repository's single-GPU First Block Cache example for HunyuanVideo.

Before you start

  • Paraattention installed and the repository's example files available.
  • Diffusers installed from its latest master branch.
  • A GPU; the README suggests at least 48 GB of memory for HunyuanVideo.

Steps

  1. Open the repository's first_block_cache_examples/run_hunyuan_video.py example.

  2. Run the command below.

    Documented command
    python3 first_block_cache_examples/run_hunyuan_video.py

Expected outcome

The HunyuanVideo First Block Cache example is invoked as documented.

Guide limitations

  • The README warns that lower-memory GPUs may encounter out-of-memory errors and that frequent memory reallocation can reduce performance.
  • The documented command is an example invocation, not a guarantee of a successful run or a particular speedup.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
18 October 2026
Original website title
chengzeyi/ParaAttention · Official homepage title · retained in its source language
How we review sourcesSuggest a correction