triton-sageattention

Set every WanVideoModelLoader attention_mode to sdpa and disconnect the WanVideoTorchCompileSettings node so the workflow renders unaccelerated. Then optionally install SageAttention with a wheel matched to your torch, CUDA, and python versions.

Find the python ComfyUI actually runs (standalone-env or python_embeded), then install triton-windows pinned to your torch minor and the matching woct0rdho SageAttention wheel for your CUDA line. Always use "<that python>" -m pip, never a global pip.

No, Triton and SageAttention are not available on Mac because there is no CUDA. Use PyTorch sdpa attention on the MPS backend and avoid torch.compile or inductor nodes entirely.

A wheel built for a different CUDA line or torch minor can drag in mismatched CUDA DLLs and break torch imports. Match cu128 versus cu130 and the torch minor exactly, record your torch version first, and roll back if torch stops importing with CUDA.

Yes, the Windows wheels use the cp310-abi3 stable ABI tag, so one wheel covers Python 3.10 through 3.13 and later. The absence of a cp313-specific filename is expected, not a missing build.

Use sdpa when SageAttention is not installed, when you are on Mac, or when sageattn produces noise or NaN artifacts on your GPU. sdpa is built into PyTorch, always available, and costs speed but not correctness.