ComfyUI Node
MiniMax H3 FastPath Euler Middle Cache
Stock-Euler-only, LoRA-strength-aware H3 cache. Runs a prefix and suffix exactly; on accepted calls it reuses only the residual of the middle blocks. Cache state is isolated per CFG branch.
MiniMax H3 FastPath Euler Middle Cache
- model
- model
◄enabledtrue►
◄prefix_blocks8►
◄suffix_blocks8►
◄reuse_threshold0.120►
◄max_consecutive_reuses1►
◄cache_device▾►
◄require_fastpath_scheduletrue►
◄strict_branch_identitytrue►
◄suppress_candidate_prefetchtrue►
◄verbosefalse►
CategoryMiniMax H3/FastPath
Inputs (11)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| enabled | BOOLEAN | true | — |
| prefix_blocks | INT | 81–48 | Exact blocks run before the cached middle on every model call. |
| suffix_blocks | INT | 81–48 | Exact final blocks run after the cached middle on every model call. |
| reuse_threshold | FLOAT | 0.1200–2 | Maximum relative change in a sampled prefix feature before reuse is rejected. |
| max_consecutive_reuses | INT | 11–4 | One gives FULL/PART/FULL for three calls at the same LoRA strength. |
| cache_device | COMBO | GPU is fastest. CPU saves VRAM but transfers one full hidden residual on reuse. | |
| require_fastpath_schedule | BOOLEAN | true | Refuse approximation unless the matching FastPath scheduled-LoRA signature is present. |
| strict_branch_identity | BOOLEAN | true | Use the exact path if ComfyUI does not provide CFG branch labels. |
| suppress_candidate_prefetch | BOOLEAN | true | On likely reuse calls, stop dynamic INT8 weight prefetch so skipped blocks are not streamed anyway. |
| verbose | BOOLEAN | false | — |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |