ComfyUI Node
H3 Depth Skip
Skips named transformer blocks, or just their attention half. Use MiniMax H3 Block Contribution first to find out which blocks are idle on your model — on the released checkpoints that is roughly 3-17, worth about 1.4x. Empty is an exact no-op.
H3 Depth Skip
- model
- MODEL
- report
◄blocks►
◄start_percent0.00►
◄end_percent1.00►
◄modeattention_only►
CategoryNynxz/H3/Sampling
Inputs (5)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | A MiniMax H3 model. | |
| blocks | STRING | Block indices to skip: '3-17', '3-17, 22', '10'. Ranges are inclusive. Empty changes nothing. Read them off MiniMax H3 Block Contribution rather than guessing — the quiet band is a property of the checkpoint, and it is identical across the fl2va and ref2va releases. | |
| start_percent | FLOAT | 0.000–1 | Where in the schedule the skip switches ON, as a fraction of sampling. 0 is the first step. |
| end_percent | FLOAT | 1.000–1 | Where it switches OFF. 1 runs to the last step; the default pair (0, 1) is always-on, which is the behaviour this node had before. Ending early is the useful setting: skip the quiet band while structure is being decided, then run the full stack for refinement, where the blocks you dropped may matter more than their average contribution suggests. |
| mode | COMBO | attention_only | attention_only keeps the MLP, which is the conservative choice and costs about a fifth of the saving — attention is ~83% of block time at these sequence lengths. It is also the better-evidenced one: the MLP's gate varies 8x across depth against attention's 68x, so the MLP is much less obviously idle in any band. whole_block drops everything and is the thing to reach for once an A/B says the band really is inert. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |
| report | STRING | Which blocks were skipped, what fraction of the forward that removes, and what the number does and does not mean. |