WanVideo Mem Eff Sage Attention Patch(WanVideoMemoryEfficientSageAttentionPatch)

Description

An experimental node that patches the WanVideo self-attention mechanism to use a custom sageattention implementation, reducing peak VRAM usage during inference. This node overrides the attention mode of the model.

Requires the latest version of sageattention – ensure it is installed, compatible, and up to date before using this node.

---

Inputs

---

Outputs

---

Usage Notes

• This node is experimental – behaviour may change in future releases.
• It only affects the self‑attention layers of WanVideo; cross‑attention or other components remain unchanged.
• Patching is applied as a model wrapper; the output MODEL can be used directly in subsequent sampling nodes.
• If sageattention is not installed or is outdated, an error will be raised at runtime.
• Running this patch may alter output quality or behaviour – test carefully with your specific workflow.

Installation (SageAttention)

This node requires the SageAttention library to be installed in your ComfyUI Python environment:

• Official repo (Linux): thu-ml/SageAttention - pip install sageattention (requires python>=3.9, torch>=2.3.0, triton>=3.0.0, CUDA>=12.0; see the README for version requirements)
• Windows: woct0rdho/SageAttention - prebuilt wheels for Windows; install triton-windows first, then pick a wheel from the releases page matching your PyTorch version (see the README)

For ComfyUI's built-in global acceleration (no node needed), start ComfyUI with --use-sage-attention. If WanVideo produces black or noisy output due to quantization overflow, use this node with sageattn_qk_int8_pv_fp16_cuda (least likely to overflow), or adjust pv_accum_dtype.