afridi26/h100-fp8-wheels

Prebuilt FA3 + TransformerEngine wheels (H100)

Compiled CUDA extensions. ONLY valid for this exact stack:

Python	3.10
torch	2.6.0
CUDA	12.8
arch	sm90 (H100 80GB HBM3)

On a mismatch you get an import error or, worse, silent misbehaviour. torch version is the binding constraint: flash_attn_3 is tagged abi3 so it tolerates newer Pythons, but it is still linked against torch 2.6.0 C++ ABI. transformer_engine is cp310-locked.

• flash_attn_3-3.0.0 from Dao-AILab/flash-attention @ df61ab6, hopper/ subdir (BSD-3)
• transformer_engine-2.17.0+b82e7442 from NVIDIA/TransformerEngine @ b82e7442, stable branch (Apache-2.0), built with NVTE_WITH_NCCL_EP=0 and CMAKE_BUILD_WITH_INSTALL_RPATH=ON

Consumed automatically by Bria-AI/bria-video-eraser branch ali/fp8 via setup_h100.sh, which discovers these filenames from the HF API. Saves ~45 min of compiling per fresh box (flash-attn-3 alone is ~30 min).