ComfyUI Node
MiniMax H3 Long-Sequence Attention (Experimental)
H3-only Sol-Attn candidate from 12K tokens. Builds an explicit accepted-block mask with 90% estimated mass coverage and a 50% exact floor, then uses the stable grouped INT8 QK/PV kernel. Conditioning query/KV rows and sensitive transformer blocks stay dense.
MiniMax H3 Long-Sequence Attention (Experimental)
- model
- MODEL
◄enabledfalse►
◄verbosefalse►
Categorysol_attn/minimax_h3
Inputs (3)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| enabled | BOOLEAN | false | Explicit opt-in; disabled preserves dense attention. |
| verbose | BOOLEAN | false | — |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |