FlashAttention 2.8.4 (CUDA 13.0 / Torch 2.12 / Triton 3.7)

Prebuilt FlashAttention wheel for modern NVIDIA GPUs (Blackwell / Ada / Hopper).

---

🚀 Download

👉 pip install --no-deps https://github.com/CozzyTyler/flash-attn-cu130-wheel/releases/download/v2.8.4-cu130-torch2.12-triton3.7/flash_attn-2.8.4+cu130.torch2.12.triton3.7-cp310-cp310-linux_x86_64.whl

---

⚙️ Build info

• CUDA: 13.0
• PyTorch: 2.12.0.dev+cu130
• Triton: 3.7.0
• Python: 3.10
• ABI: CXX11 (True)

Built on RTX 5090 Laptop GPU (Blackwell).

---

⚠️ Important

This wheel is NOT universal.

It is compiled for a specific stack:

• Torch 2.12 (nightly)
• CUDA 13.0
• Triton 3.7

👉 May NOT work on:

• stable PyTorch
• older CUDA
• different Triton

---

📦 Install

Download .whl from Releases and install:

pip install --no-deps https://github.com/CozzyTyler/flash-attn-cu130-wheel/releases/download/v2.8.4-cu130-torch2.12-triton3.7/flash_attn-2.8.4+cu130.torch2.12.triton3.7-cp310-cp310-linux_x86_64.whl