SageAttention 2.2 + FlashAttention 2.8.3 — Linux wheels for Blackwell (sm_120) on PyTorch 2.11 / CUDA 13.0
Community-built .whl files for a combo that, as of June 2026, isn't shipped by upstream:
- SageAttention 2.2.0 — only Windows wheels in upstream releases; Linux requires source build
- FlashAttention 2.8.3.post1 — Dao-AILab releases don't cover cu130 + torch 2.11 yet
If your hardware/software matches the table below, these install in seconds instead of 5–40 minutes of source compilation.
Compatibility
| Requirement | Value | Why it matters |
|---|---|---|
| Python | 3.12.x (built on 3.12.3) | cp312 ABI tag — won't load on 3.11 or 3.13 |
| PyTorch | 2.11.0+cu130 | C++ ABI must match — won't load with cu126 torch |
| CUDA toolkit | 13.0 (driver compatible with cu130) | required for runtime |
| GPU compute capability | 12.0 (Blackwell) — RTX PRO 6000 Blackwell, RTX 5090/5080/5070 Ti, B100/B200, etc. | wheels compiled with arch=compute_120,code=sm_120 natively (no PTX-JIT fallback) |
| OS | Linux x86_64 | manylinux not applied — built on Ubuntu, untested elsewhere |
If any row doesn't match — these wheels probably won't load and you'll need to build from source.
Files & hashes
| File | Size | SHA256 |
|---|---|---|
sageattention-2.2.0-cp312-cp312-linux_x86_64.whl |
14.8 MB | 4a47a462059bf09b7d10be68f93a0a462dd6184068298fc6ff10e94751fa6dd4 |
flash_attn-2.8.3.post1-cp312-cp312-linux_x86_64.whl |
65.8 MB | a77fe499d1c7244e0549d9d2700def0c1c82a1b629d993204e9be78debd96cb8 |
Verify before installing:
sha256sum sageattention-2.2.0-cp312-cp312-linux_x86_64.whl
sha256sum flash_attn-2.8.3.post1-cp312-cp312-linux_x86_64.whl
Install
pip install sageattention-2.2.0-cp312-cp312-linux_x86_64.whl
pip install flash_attn-2.8.3.post1-cp312-cp312-linux_x86_64.whl
Smoke test:
import torch
import sageattention
from sageattention import sageattn
from flash_attn import flash_attn_func
q = torch.randn(2, 8, 256, 64, dtype=torch.float16, device='cuda')
k = torch.randn(2, 8, 256, 64, dtype=torch.float16, device='cuda')
v = torch.randn(2, 8, 256, 64, dtype=torch.float16, device='cuda')
# sageattention expects (B, H, S, D) by default with tensor_layout='HND'
out_sage = sageattn(q, k, v, tensor_layout='HND', is_causal=False)
# flash_attn_func expects (B, S, H, D)
q2 = q.transpose(1, 2).contiguous()
k2 = k.transpose(1, 2).contiguous()
v2 = v.transpose(1, 2).contiguous()
out_flash = flash_attn_func(q2, k2, v2, causal=False)
print('sageattention:', out_sage.shape, '— OK')
print('flash_attn: ', out_flash.shape, '— OK')
Build environment (exactly what produced these wheels)
OS: Ubuntu 24.04 (kernel 6.14.0-1010-oem)
Python: 3.12.3
PyTorch: 2.11.0+cu130
CUDA: 13.0.r13.0 / V13.0.88 (apt cuda-toolkit-13-0)
ninja: 1.11.1.4
setuptools: 81.0.0
GPU: NVIDIA RTX PRO 6000 Blackwell Workstation Edition (sm_120)
Build flags:
sageattention:
TORCH_CUDA_ARCH_LIST="12.0" MAX_JOBS=4 NVCC_THREADS=2 \
pip install . --no-build-isolation
flash-attn:
TORCH_CUDA_ARCH_LIST="12.0" FLASH_ATTN_CUDA_ARCHS="120" MAX_JOBS=2 NVCC_THREADS=1 \
pip install flash-attn==2.8.3.post1 --no-build-isolation
Backward kernels are included in flash-attn — training works, not just inference.
Licenses
Both upstream projects allow redistribution; redistributing these binary builds preserves their license terms:
- SageAttention — Apache License 2.0
- FlashAttention — BSD 3-Clause
The wheels contain copies of LICENSE files in their *.dist-info/licenses/ directories. Originals are at:
Disclaimers (read these)
- Not official. This is a community build. Authors of upstream packages have no involvement.
- Smoke-tested only. Each wheel was verified to: (1) import without errors, (2) execute a small forward pass on a Blackwell GPU and return a tensor of expected shape and dtype. No accuracy benchmarks, no model-level integration tests were run.
- No warranty. If these wheels misbehave on your system — wrong results, crashes, performance regressions — that's on you to verify against a source build before relying on them.
- Reproducible-ish. Wheel SHA-256 may differ slightly between rebuilds due to embedded timestamps even when binary kernel content is identical.
Attribution
These wheels were built by JahJedi with assistance from Claude Code (Anthropic) — the diagnosis of an earlier system freeze under unbounded ninja, the choice of safe build parameters (MAX_JOBS, NVCC_THREADS, single-arch targeting), and this README itself were Claude-assisted. Verification commands were run by the human; final responsibility for the published artifacts is the human's.
If anything in these wheels turns out to be wrong, ping the human first — they can't claim AI ignorance.