Force PyTorch Attention (xformers fix)
Force PyTorch attention for a single model. Fixes 'No operator found for memory_efficient_attention_forward' on RTX 50xx with xformers (e.g. Qwen-Image 2.1) without disabling xformers globally.
Force PyTorch Attention (xformers fix)
A tiny ComfyUI node that switches attention to PyTorch SDPA for one model only. Every other model in your workflow keeps the global backend (xformers, sage, ...).
Why
On RTX 50xx (Blackwell) with xformers installed, some models (e.g. Qwen-Image 2.1) crash with:
NotImplementedError: No operator found for `memory_efficient_attention_forward`
... requires device with capability <= (9, 0) but your GPU has capability (12, 0) (too new)
The usual fix is --disable-xformers, which changes attention for all models.
This node fixes only the model that needs it.
Usage
Load Diffusion Model → Force PyTorch Attention (xformers fix) → sampler / guider.
No settings. Category: model_patches.
How it works
Sets ComfyUI's built-in transformer_options["optimized_attention_override"] on a clone of the model,
so attention calls of that model go to attention_pytorch. Nothing global is changed.
Install
ComfyUI Manager → search Force PyTorch Attention, or clone this repo into ComfyUI/custom_nodes.
License
MIT