ComfyUI Node
MiniMax H3 Attention Parallel
Split exact Comfy Kitchen attention heads across local GPUs without copying model weights.
MiniMax H3 Attention Parallel
- model
- model
◄devicesdisabled►
◄min_sequence_length32768►
Categorymodel/patch/minimax
Inputs (3)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| devices | COMBO | disabled | 5 options: disabled, auto, 2, 3, 4 |
| min_sequence_length | INT | 327681–1000000 | In auto mode, use one GPU below this packed token count. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |