File size: 2,794 Bytes
27009c8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: other
pretty_name: CUDA Extension Wheelhouse
tags:
- wheels
- cuda
- pytorch
- prebuilt
viewer: false
---

# CUDA Extension Wheelhouse

Prebuilt Python **wheels** for CUDA / PyTorch extensions
(`cumesh`, `flash_attn`, `flex_gemm`, `natten`, `nvdiffrast`, `nvdiffrec_render`,
`o_voxel`). Compiling these from source dominates install time — often **hours**
on a single machine. This dataset is a shared cache so machines and CI can
**build once, then reuse the wheel** and skip the compile entirely.

Built for GPU archs `sm75`, `sm80`, `sm89`, `sm120` against `torch2.7`/`torch2.9`
(CUDA 12.8), for CPython `cp310` and `cp312`.

## Why a wheel can't just be cached by filename

A compiled-extension wheel's runnability depends on **four** axes, but the wheel
filename only encodes **two** (dist name + Python ABI tag). The other two — the
**GPU compute architecture** and the **torch version it was ABI-linked against** —
are invisible in the name. A wheel built for `sm120` / `torch2.9` installs fine on
`sm89` / `torch2.7` and then dies at the first kernel launch
(`no kernel image available`). So wheels are stored under a path that carries
arch + torch.

## Layout

```
<package>/sm<arch>/torch<major.minor>-cu<cuda>/<wheel>.whl
```

Examples:

```
flash_attn/sm89/torch2.9-cu128/flash_attn-2.8.3+cu128torch2.9-cp312-cp312-linux_x86_64.whl
nvdiffrast/sm89/torch2.9-cu128/nvdiffrast-0.4.0-cp312-cp312-linux_x86_64.whl
natten/sm120/torch2.7-cu126/natten-0.21.0-cp310-cp310-linux_x86_64.whl
```

- `sm<arch>` — GPU compute capability of the card the wheel runs on (e.g. `sm89`
  for Ada / RTX 40xx, `sm120` for Blackwell). Keyed off the **real GPU**, not a
  multi-arch build list.
- `torch<mm>-cu<cuda>` — the torch + CUDA the extension was linked against
  (e.g. `torch2.9-cu128`).
- The Python ABI tag (`cp310`, `cp312`, …) lives in the wheel filename itself.

## Usage

Download a specific wheel:

```python
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="DoctorPopi/python_wheelhouse",
    repo_type="dataset",
    filename="flash_attn/sm89/torch2.9-cu128/flash_attn-2.8.3+cu128torch2.9-cp312-cp312-linux_x86_64.whl",
)
# then: pip install --force-reinstall --no-deps <path>
```

> Install compiled wheels with `--force-reinstall --no-deps`: several declare an
> unpinned `torch` requirement and a bare reinstall would clobber the env's pinned
> torch, breaking every other C-extension built against the original ABI.

## Licensing

Each wheel is a build of its **upstream project** and remains under that project's
own license (e.g. natten, nvdiffrast, flash-attention each have their own terms).
This dataset only redistributes compiled artifacts for convenience; refer to each
package's upstream repository for license terms.