CUDA 13.0 Python 3.12 PyTorch 2.11/2.12 FA2/FA3 wheels

Prebuilt Linux x86_64 CUDA 13.0 wheels for Python 3.12.
Included artifacts:

• FlashAttention 2.8.4 for PyTorch 2.11.0+cu130, sm80/sm89/sm90/sm100
• FlashAttention 3.0.0 for PyTorch 2.11.0+cu130, sm90 only
• FlashAttention 2.8.4 for PyTorch 2.12.0+cu130, sm80/sm89/sm90/sm100
• FlashAttention 3.0.0 for PyTorch 2.12.0+cu130, sm90 only

CUDA 13.0 Python 3.11 PyTorch 2.12 FA2/FA3 wheels

CUDA 13.0 / Python 3.11 / PyTorch 2.12 Linux x86_64 wheels.

• FA2: sm80/sm89/sm90/sm100
• FA3: sm90 only

CUDA 13.0 Python 3.11 PyTorch 2.11 FA2/FA3 wheels

CUDA 13.0 / Python 3.11 / PyTorch 2.11 Linux x86_64 wheels.

• FA2: sm80/sm89/sm90/sm100
• FA3: sm90 only

Built on mlx worker and verified against torch 2.11.0+cu130.

FA2 wheel for python 3.12 cuda 12.9 torch 2.10

v0.0.2

Merge pull request #8 from windreamer/copilot/fix-corrupt-patch-error

cuda 12.9 python 3.11 torch 2.9

03/13 clean build for cuda 12.9 python 3.11 torch 2.9

• flash attention 3 for Hopper
• flash attention 2 for Ampere
  • no cute/ included which would pollute fa4 install otherwise