ComfyUI Node
FlashVSR Sparge Attention
Uses the separately installed SpargeAttn library to execute FlashVSR's logical 128-query by 128-key LCSA mask during streaming. The mask is converted to the GPU kernel's physical block geometry only after routing. It does not replace ModelAttentionBackend: dense self-attention, cross-attention, and full_video_dense continue through the model's selected ComfyUI backend. Patch order relative to Configure FlashVSR Upscaling and ModelAttentionBackend does not matter. A matching SpargeAttn wheel and CUDA GPU are required.
FlashVSR Sparge Attention
- model
- MODEL
CategoryFlashVSR/model_patches
Inputs (1)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |