File size: 3,750 Bytes
aa034eb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
---
license: unlicense
tags:
  - wheels
  - pip
  - flash-attention
  - gpu
  - build-cache
pretty_name: GPU Build Wheel Cache
---

# gpu-wheels

Personal cache of prebuilt Python wheels for packages that are slow to compile from source (flash-attn, etc.), so rented GPU instances (vast.ai and similar) don't have to recompile from scratch every time.

Compiling `flash-attn` from source can take 45-90 minutes. If a new instance has the **exact same** `torch` version, CUDA version, Python version, and C++ ABI as a wheel already in this repo, installing from the cached wheel takes seconds instead.

## Repo layout

```
<package>/torch<torch_version>-py<python_version>-cxx11abi<True|False>/<wheel_filename>.whl
```

Example:

```
flash-attn/torch2.12.0+cu130-py3.12-cxx11abiTrue/flash_attn-2.8.3.post1-cp312-cp312-linux_x86_64.whl
```

The folder name is the compatibility key. A wheel only works on an environment matching **all** of: package version, torch version (incl. CUDA suffix), Python version, and cxx11abi flag.

## Available wheels

| Package | Torch | CUDA | Python | cxx11abi | GPU built on | Path |
|---|---|---|---|---|---|---|
| flash-attn 2.8.3.post1 | 2.12.0 | 13.0 | 3.12 | True | RTX 3090 (sm86) | `flash-attn/torch2.12.0+cu130-py3.12-cxx11abiTrue/` |

CUDA kernel wheels are generally GPU-arch-agnostic across NVIDIA GPUs (they embed multiple SM targets), so a wheel built on one GPU normally works on others — the torch/CUDA/Python/ABI match is what matters, not the specific GPU model.

## Usage: install a cached wheel

```bash
pip install huggingface_hub

python3 -c "
from huggingface_hub import hf_hub_download
path = hf_hub_download(
    repo_id='DanielTobi0/gpu-wheels',
    repo_type='dataset',
    filename='flash-attn/torch2.12.0+cu130-py3.12-cxx11abiTrue/flash_attn-2.8.3.post1-cp312-cp312-linux_x86_64.whl',
)
print(path)
"

pip install <path printed above>
```

Or in one line once you know the filename:

```bash
pip install "$(python3 -c "from huggingface_hub import hf_hub_download; print(hf_hub_download(repo_id='DanielTobi0/gpu-wheels', repo_type='dataset', filename='flash-attn/torch2.12.0+cu130-py3.12-cxx11abiTrue/flash_attn-2.8.3.post1-cp312-cp312-linux_x86_64.whl'))")"
```

**Before installing**, check your new instance's versions match the folder name:

```bash
python3 -c "import torch; print(torch.__version__, torch.version.cuda, torch._C._GLIBCXX_USE_CXX11_ABI)"
```

If they don't match, the wheel likely won't install (or worse, may install but be ABI-incompatible) — build fresh instead and add the new combo to this repo (see below).

## Adding a new wheel after a fresh build

1. Build normally (e.g. `pip install flash-attn --no-build-isolation`).
2. Locate the built wheel. With `uv`, it's cached under `~/.cache/uv/sdists-v9/pypi/<package>/<version>/*/​*.whl`. With plain `pip`, add `--no-clean -v` or build explicitly with `pip wheel <package> --no-build-isolation -w /tmp/wheelhouse`.
3. Record your environment's compatibility key:
   ```bash
   python3 -c "import torch; print(f'torch{torch.__version__}-py{__import__(\"platform\").python_version()[:4]}-cxx11abi{torch._C._GLIBCXX_USE_CXX11_ABI}')"
   ```
4. Upload:
   ```python
   from huggingface_hub import HfApi
   api = HfApi()
   api.upload_file(
       path_or_fileobj="/path/to/built.whl",
       path_in_repo="<package>/<compat-key>/<wheel_filename>.whl",
       repo_id="DanielTobi0/gpu-wheels",
       repo_type="dataset",
   )
   ```
5. Add a row to the table above.

## Notes

- This repo is **private** — wheels may be built against specific local paths/configs and aren't intended for public redistribution.
- Wheels are large (100-300MB+ for CUDA extensions); this is a personal cache, not a package index.