ComfyUI Node
Model (RotorQuant Attention)
A ComfyUI node in model/patch with 9 inputs and 1 output.
Model (RotorQuant Attention)
- model
- MODEL
◄keep_components3►
◄min_token_product65536►
◄attention_scope▾►
◄layer_start-1►
◄layer_end-1►
◄rotation_seed0►
◄max_head_dim256►
◄force_fp32▾►
Categorymodel/patch
Inputs (9)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | Model to clone and patch with RotorQuant-style attention. | |
| keep_components | INT | 31–3 | How many rotated coordinates to keep from each 3D rotor group. Values below 3 are currently forced back to 3 because the lossy modes degrade image quality too much. |
| min_token_product | INT | 655360–1073741824 | Only patch attention calls where query_tokens * key_tokens meets this threshold. |
| attention_scope | COMBO | Which attention calls to patch. 'self' is usually the most useful for diffusion latents. | |
| layer_start | INT | -1-1–4096 | First transformer block index to patch. -1 disables the lower bound. |
| layer_end | INT | -1-1–4096 | Last transformer block index to patch. -1 disables the upper bound. |
| rotation_seed | INT | 00–18446744073709550000 | Seed used to generate the per-triplet rotor rotations. |
| max_head_dim | INT | 2563–4096 | Skip heads larger than this to avoid excessive projection overhead. |
| force_fp32 | COMBO | Cast q/k/v to fp32 inside the override for extra numerical stability. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |