nanode-cuda-kernels / README.md
pandafood's picture
kernels-v1: TRELLIS.2 Windows cp311 kernels + Linux cp310 extras
c3c440f verified
|
Raw
History Blame Contribute Delete
2.97 kB
---
license: other
license_name: upstream-licenses
license_link: https://github.com/Unrankedbandit/Nanode/blob/main/THIRD_PARTY_NOTICES.md
tags:
- cuda
- wheels
- trellis
- windows
- prebuilt
---
# Nanode CUDA kernels — prebuilt wheels
Prebuilt CUDA extension wheels used by [Nanode](https://github.com/Unrankedbandit/Nanode),
a node-based AI canvas. Hosted here because most of these builds exist nowhere
else — PyPI carries source-only distributions (or nothing at all) for these
packages, and the upstream binary channels don't cover these ABI combinations.
## Windows — Microsoft TRELLIS.2 kernels (cp311)
The custom CUDA extensions behind [TRELLIS.2](https://github.com/microsoft/TRELLIS)
image-to-3D generation, compiled for Windows. To our knowledge these are the
only public **cp311 win_amd64** builds.
| Wheel | ABI | Built against |
|---|---|---|
| `cumesh-1.0-cp311-cp311-win_amd64.whl` | cp311 / win_amd64 | torch 2.7.0, CUDA 12.8 |
| `flex_gemm-0.0.1-cp311-cp311-win_amd64.whl` | cp311 / win_amd64 | torch 2.7.0, CUDA 12.8 |
| `nvdiffrast-0.4.0-cp311-cp311-win_amd64.whl` | cp311 / win_amd64 | torch 2.7.0, CUDA 12.8 |
| `nvdiffrec_render-0.0.0-cp311-cp311-win_amd64.whl` | cp311 / win_amd64 | torch 2.7.0, CUDA 12.8 |
| `o_voxel-0.0.1-cp311-cp311-win_amd64.whl` | cp311 / win_amd64 | torch 2.7.0, CUDA 12.8 |
**GPU floor: RTX 30-series (sm_86).** The binaries embed `sm_86` cubins plus
`sm_86` PTX — native on RTX 30, RTX 40/50 run via the driver's PTX JIT (one-time
compile pause on first use). GTX 10 / RTX 20 cards (sm_75 and below) cannot run
these builds.
## Linux — supporting kernels (cp310)
Built on WSL2 (Ubuntu) for Nanode's Linux-side pipelines.
| Wheel | Notes |
|---|---|
| `flash_attn-2.7.4.post1-cp310-cp310-linux_x86_64.whl` | Local build; official releases cover torch ≤ 2.6 only |
| `natten-0.21.0-cp310-cp310-linux_x86_64.whl` | PyPI is sdist-only; upstream wheel host currently unreachable |
| `gsplat-1.5.3-cp310-cp310-linux_x86_64.whl` | Alternative to the nerfstudio wheel index |
| `transformer_engine_torch-1.12.0-cp310-cp310-linux_x86_64.whl` | No public wheels exist (NVIDIA ships NGC containers) |
| `fused_ssim-0.0.0-cp310-cp310-linux_x86_64.whl` | Source-only upstream; no other wheel exists |
## Verification
`SHA256SUMS.txt` covers every wheel. Verify with:
```
sha256sum -c SHA256SUMS.txt
```
Nanode's provisioner pins these hashes in its pipeline manifests and refuses
mismatched downloads.
## Licenses
These are unmodified compiled builds of third-party source. Each upstream
project's license applies to its wheel: TRELLIS.2 kernels (`cumesh`,
`flex_gemm`, `o_voxel`, `nvdiffrec_render`) — MIT (Microsoft);
`nvdiffrast` — NVIDIA Source Code License;
`flash_attn` — BSD-3 (Dao AI Lab); `natten` — MIT (SHI Labs);
`gsplat` — Apache-2.0 (nerfstudio); `transformer_engine` — Apache-2.0 (NVIDIA);
`fused_ssim` — MIT. See each project's repository for the authoritative text.