---
license: apache-2.0
tags:
  - wheels
  - cuda
  - pytorch
  - flash-attention
  - sageattention
  - xformers
  - windows
  - linux
  - aarch64
---

# Precompiled Wheels

Prebuilt Python wheels for CUDA and PyTorch combinations that are awkward to build locally.

Filename convention:

```text
<package>-<version>+cu<cuda>.torch<torch_ver>-<py_tag>-<abi_tag>-<platform>.whl
```

Example:

```text
sageattention-2.2.0+cu130.torch2.11-cp313-cp313-win_amd64.whl
```

This means:

- CUDA `13.0`
- PyTorch `2.11`
- Python `3.13`
- Windows `x86_64`

## Files

### Windows x86_64

| File | Notes |
| --- | --- |
| `deep_gemm-2.3.0+cu128.torch2.10-cp310-cp310-win_amd64.whl` | DeepGEMM, Python 3.10, CUDA 12.8, Torch 2.10 |
| `flash_attn-2.8.3-cp311-cp311-win_amd64.whl` | FlashAttention 2, Python 3.11 |
| `sageattention-2.2.0+cu128.torch2.10-cp310-cp310-win_amd64.whl` | SageAttention 2.2.0, Python 3.10, CUDA 12.8, Torch 2.10 |
| `sageattention-2.2.0+cu130.torch2.11-cp312-cp312-win_amd64.whl` | SageAttention 2.2.0, Python 3.12, CUDA 13.0, Torch 2.11 |
| `sageattention-2.2.0+cu130.torch2.11-cp313-cp313-win_amd64.whl` | SageAttention 2.2.0, Python 3.13, CUDA 13.0, Torch 2.11 |
| `xformers-0.0.35+cu130.torch2.11-py39-none-win_amd64.whl` | Existing stable-ABI xformers wheel |
| `xformers-0.0.35+cu130.torch2.11-cp312-cp312-win_amd64.whl` | Repacked xformers wheel for Python 3.12 |
| `xformers-0.0.35+cu130.torch2.11-cp313-cp313-win_amd64.whl` | Repacked xformers wheel for Python 3.13 |

### Linux aarch64

| File | Notes |
| --- | --- |
| `flash_attn-2.8.3-cp310-cp310-linux_aarch64.whl` | FlashAttention 2, Python 3.10 |
| `flash_attn-2.8.4+cu130.torch2.11-cp312-cp312-linux_aarch64.whl` | FlashAttention 2, Python 3.12, CUDA 13.0, Torch 2.11 |
| `flash_attn-2.8.4+cu130.torch2.11-cp313-cp313-linux_aarch64.whl` | FlashAttention 2, Python 3.13, CUDA 13.0, Torch 2.11 |
| `sageattention-2.2.0+cu130.torch2.11-cp312-cp312-linux_aarch64.whl` | SageAttention 2.2.0, Python 3.12, CUDA 13.0, Torch 2.11 |
| `sageattention-2.2.0+cu130.torch2.11-cp313-cp313-linux_aarch64.whl` | SageAttention 2.2.0, Python 3.13, CUDA 13.0, Torch 2.11 |
| `xformers-0.0.35+cu130.torch2.11-py39-none-linux_aarch64.whl` | Existing stable-ABI xformers wheel |
| `xformers-0.0.35+cu130.torch2.11-cp312-cp312-linux_aarch64.whl` | Repacked xformers wheel for Python 3.12 |
| `xformers-0.0.35+cu130.torch2.11-cp313-cp313-linux_aarch64.whl` | Repacked xformers wheel for Python 3.13 |

### Pure Python / Any platform

| File | Notes |
| --- | --- |
| `angelslim-0.0.0.dev0+cu128.torch2.10-py3-none-any.whl` | Pure Python wheel |

## Install examples

### Windows, Python 3.12 / Torch 2.11 / CUDA 13.0

```powershell
pip install --index-url https://download.pytorch.org/whl/cu130 torch==2.11.0+cu130
$base = "https://huggingface.co/datasets/tori29umai/PrecompiledWheels/resolve/main"
pip install "$base/sageattention-2.2.0+cu130.torch2.11-cp312-cp312-win_amd64.whl"
pip install "$base/xformers-0.0.35+cu130.torch2.11-cp312-cp312-win_amd64.whl"
```

### Windows, Python 3.13 / Torch 2.11 / CUDA 13.0

```powershell
pip install --index-url https://download.pytorch.org/whl/cu130 torch==2.11.0+cu130
$base = "https://huggingface.co/datasets/tori29umai/PrecompiledWheels/resolve/main"
pip install "$base/sageattention-2.2.0+cu130.torch2.11-cp313-cp313-win_amd64.whl"
pip install "$base/xformers-0.0.35+cu130.torch2.11-cp313-cp313-win_amd64.whl"
```

### Linux aarch64, Python 3.12 / Torch 2.11 / CUDA 13.0

```bash
pip install --index-url https://download.pytorch.org/whl/cu130 torch==2.11.0+cu130
BASE=https://huggingface.co/datasets/tori29umai/PrecompiledWheels/resolve/main
pip install "$BASE/flash_attn-2.8.4+cu130.torch2.11-cp312-cp312-linux_aarch64.whl"
pip install "$BASE/sageattention-2.2.0+cu130.torch2.11-cp312-cp312-linux_aarch64.whl"
pip install "$BASE/xformers-0.0.35+cu130.torch2.11-cp312-cp312-linux_aarch64.whl"
```

### Linux aarch64, Python 3.13 / Torch 2.11 / CUDA 13.0

```bash
pip install --index-url https://download.pytorch.org/whl/cu130 torch==2.11.0+cu130
BASE=https://huggingface.co/datasets/tori29umai/PrecompiledWheels/resolve/main
pip install "$BASE/flash_attn-2.8.4+cu130.torch2.11-cp313-cp313-linux_aarch64.whl"
pip install "$BASE/sageattention-2.2.0+cu130.torch2.11-cp313-cp313-linux_aarch64.whl"
pip install "$BASE/xformers-0.0.35+cu130.torch2.11-cp313-cp313-linux_aarch64.whl"
```

## Notes

- These wheels are redistributed for convenience.
- Please check each upstream project's license before reuse.

## Upstream projects

| Package | Repository |
| --- | --- |
| `flash_attn` | https://github.com/Dao-AILab/flash-attention |
| `sageattention` | https://github.com/thu-ml/SageAttention |
| `xformers` | https://github.com/facebookresearch/xformers |
| `deep_gemm` | https://github.com/deepseek-ai/DeepGEMM |
| `angelslim` | https://github.com/Tencent/AngelSlim |