ModelAttentionBackend - ComfyUI Built-in Node Documentation - ComfyUI

This node lets you choose which attention backend a model uses for its attention computations. It creates a copy of the model and swaps in the attention function you select, which can affect performance or behavior. If the chosen backend is not available, it automatically falls back to PyTorch attention and logs a warning.

Inputs

Parameter	Description	Data Type	Required	Range
`model`	The model to which the selected attention backend will be applied.	MODEL	Yes
`attention`	The attention backend to use (default: “pytorch attention”). If the selected backend is unavailable, PyTorch attention is used as a fallback.	STRING	Yes	”pytorch attention" "comfy kitchen attention”

Note: The “comfy kitchen attention” option is only listed when the comfy kitchen int8 attention module is available in the current environment.

Outputs

Output Name	Description	Data Type
`MODEL`	A clone of the input model with the selected attention backend applied.	MODEL

This documentation was AI-generated. If you find any errors or have suggestions for improvement, please feel free to contribute! Edit on GitHub

Source fingerprint (SHA-256): 4ba613cc0bf5b3e7f9effa895b98b3a3bd302e5d20e9d7e18d1633906c783244