hip-attn 1.2.6

HiP Attention

:sunglasses: HiP Attention could extend the model context length training-free and can serve 3 million tokens with a single L40S 48GB GPU while achieving a 7.24 estimated speedup.

| Paper (Arxiv, InfiniteHiP latest) | Paper (ICLR 2025) | SGlang Integration |

[!NOTE] You can try it in our Playground in DeepAuto.ai!

[!IMPORTANT] This is NOT yet free for commercial use. The license is FSL-1.1-MIT, which is free for non-commercial use but will automatically convert to MIT license two years after each release. Please refer to the LICENSE for more details.

News

• 2025.01.26: Version 1.2 is now ready! The preprint is now prepared in arxiv.
• 2025.01.22: HiP Attention is accepted in ICLR 2025!

... More News ...

• 2025.01.03: Version 1.2 will be released soon. The new version fully supports context extension and better controls pruning hierarchy. It will also have better SGlang support (with proper KV offloading!)
• 2024.10.05: Version 1.1 is now ready, check ainl-hip-offload. KV offloading feature in under alpha state.
• 2024.09.09: Version 1.1 will be released soon. Please refer to the ainl-hip-attention2 branch for a preview. It will reduce the latency further and improve the accuracy (and this will fix most of the internal bugs of v1.0). It offers many more experimental options for further research (e.g., key access logs, modular design of masking kernel). As discussed in the Appendix, this release will actually have (hopefully) a KV offloading feature, either UVM or a custom cache management algorithm. Also, SGLang will be supported by this release. Please take a look at our company's fork for a preview.

Usage

hip-attn package is available on PyPI:

pip install hip-attn

or using uv:

uv add hip-attn

After installation, you can access the hip package from any project. hip is the code name of HiP attention.

import torch
from hip_attn import hip_attention_12, HiPAttentionArgs12

device = 'cuda'

batch_size = 1
kv_len = 128 * 1024
q_len = 32 * 1024
num_heads = 32
num_kv_heads = 8
head_dims = 128
dtype = torch.bfloat16

q = torch.randn(
    (batch_size, q_len, num_heads, head_dims),
    dtype=dtype,
    device=device
)
k = torch.randn(
    (batch_size, kv_len, num_kv_heads, head_dims),
    dtype=dtype,
    device=device,
)
v = k.clone()

output, metadata = hip_attention_12(q=q, k=k, v=v, args=HiPAttentionArgs12())
print(output.shape)

# > torch.Size([1, 32768, 32, 128])

Getting Started

Local development

Using uv (Recommended)

It’s recommended to use uv, a very fast Python environment manager, to create and manage Python environments. Please follow the documentation to install uv. After installing uv, you can create a new Python environment and install hip-attention using the following commands:

# Clone this repository
git clone [e-mail]:DeepAuto-AI/hip-attention.git
cd hip-attention

# This install all research dev dependencies in .venv/
uv sync
uv run pre-commit install

Then you can run any python program with uv run. uv run automatically picks up .venv/ virtual environment:

• Script: uv run src/hip_research/main/model_eval.py
• Module: uv run -m src.hip_research.main.model_eval

Using pip and conda

# Clone this repository
git clone [e-mail]:DeepAuto-AI/hip-attention.git
cd hip-attention

# Make new conda environment
conda create --name hip python=3.11
conda activate hip

# Default install
pip install -e "."
# (Optional) For research benchmarks and unit tests
pip install -e "hip-research"

# Optional, depends on your CUDA environment
export CUDACXX=/usr/local/cuda/bin/nvcc

# Install SGLang with support for HiP Attention
pip install -e ".[sglang]" \
"sglang[all] @ git+https://github.com/DeepAuto-AI/sglang.git@deepauto/release#subdirectory=python" \
--no-build-isolation \
--verbose \
--find-links https://flashinfer.ai/whl/cu124/torch2.5/flashinfer-python

Docker

Docker images deepauto/hip-attention are available on Docker Hub. Docker examples are available in Running section.

Running

See the following pages for more details:

• Running OpenAI API server examples (SGlang)

Experiment Reproduce

Check how to reproduce experiment page

Citation

@misc{willette2025_delta_attention,
      title={Delta Attention: Fast and Accurate Sparse Attention Inference by Delta Correction},
      author={Jeffrey Willette and Heejun Lee and Sung Ju Hwang},
      year={2025},
      eprint={2505.11254},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2505.11254},
}

@misc{lee2025_infinite_hip,
      title={InfiniteHiP: Extending Language Model Context Up to 3 Million Tokens on a Single GPU},
      author={Heejun Lee and Geon Park and Jaduk Suh and Sung Ju Hwang},
      year={2025},
      eprint={2502.08910},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.08910},
}

@inproceedings{lee2025_hip_attention,
      title={A Training-Free Sub-quadratic Cost Transformer Model Serving Framework with Hierarchically Pruned Attention},
      author={Heejun Lee and Geon Park and Youngwan Lee and Jaduk Suh and Jina Kim and Wonyong Jeong and Bumsik Kim and Hyemin Lee and Myeongjae Jeon and Sung Ju Hwang},
      booktitle={The Thirteenth International Conference on Learning Representations},
      year={2025},
      url={https://openreview.net/forum?id=PTcMzQgKmn}
}

Contributing

Building and publishing

• PyPI

rm -rf dist
uv build --no-sources
uv publish

• Docker

git clone [e-mail]:DeepAuto-AI/hip-attention.git
cd hip-attention
docker login

docker build -t deepauto/hip-attention:latest -t deepauto/hip-attention:latest-sglang -t deepauto/hip-attention:$(git rev-parse --short HEAD)-sglang -t deepauto/hip-attention:v$(uv run python -c 'import importlib.metadata; print(importlib.metadata.version("hip-attn"))')-sglang -f Dockerfile.sglang .

docker push deepauto/hip-attention:latest
docker push deepauto/hip-attention:latest-sglang
docker push deepauto/hip-attention:$(git rev-parse --short HEAD)-sglang
docker push deepauto/hip-attention:v$(uv run python -c 'import importlib.metadata; print(importlib.metadata.version("hip-attn"))')-sglang