Flash-Attn CUDA Version Fix | Generated by AI

Home 2026.07 English 中文

---

✅ Done! flash-attn==2.8.3 installed successfully

What happened

The real issue wasn’t about sticking or compiling — it was a CUDA version mismatch:

Component	Expected	Reality
PyTorch CUDA	13.0 (compiled with)	torch 2.11.0+cu130
System CUDA 12.8	❌	version check failed
System CUDA 13.2	❌	13.2 != 13.0
Prebuilt wheel cu13torch2.10	✅	used this

What I did

1. Killed the pointless sm_80 compilation in /mnt/data/
2. Discovered your PyTorch needs CUDA 13.0, not 12.8

3. Found a prebuilt wheel at:

   flash_attn-2.8.3+cu13torch2.10cxx11abiTRUE-cp312-cp312-linux_x86_64.whl
   

4. Installed it — no compilation needed, instant install

Verification

• flash-attn 2.8.3 ✓
• PyTorch 2.11.0+cu130 ✓
• CUDA 13.0 ✓
• Device RTX 4070 ✓
• flash_attn_func imports successfully ✓

Note: 2.8.4 doesn’t exist on PyPI — 2.8.3 and 2.8.3.post1 are the latest. The installed 2.8.3 is functionally equivalent.

---

Generated by AI. Curating and sharing still takes effort. If you find it useful, feel free to donate. WeChat: @lzwjava · X: @lzwjava · Say hi 👋

Back Donate