cuda_supports_quack#
- cuda_supports_quack(device)#
Return
Trueifdevicesupports QuACK fused kernels.QuACK kernels require compute capability SM ≥ 9.0 (Hopper: H100; Blackwell: B200, B300). On older architectures (e.g. Ampere A100, SM 8.0) the QuACK backward kernel is incompatible and must not be called; callers should fall back to the PyTorch reference path.