MiniMax H3 Adaptive Cache:
MiniMaxH3AdaptiveCache adds a content-aware cache to ComfyUI’s native MiniMax H3 model during sampling. It reuses suitable work from later model blocks, called the “tail,” so those blocks can be skipped when the cache controller allows it. This can reduce repeated computation, with a tradeoff between speed and fidelity controlled by the selected preset. The node works through model block replacements rather than changes to MiniMax H3’s model file. Its description also states that it is compatible with Sol-Attn when Morton is disabled and safely bypasses Morton calls.
MiniMax H3 Adaptive Cache Input Parameters:
model
The required MODEL to patch. Connect a ComfyUI native MiniMax H3 model; the node checks for that model’s DiT blocks and for the block-replacement capability it needs. Other model types, or a model connection without the required patching support, cannot be processed. This parameter has no numeric minimum, maximum, options, or default.
preset
Selects how the adaptive cache is configured. The options are safe, balanced, fast, and sigma_only; the default is balanced. According to the node’s tooltip, safe favors fidelity, fast skips more tail blocks, and sigma_only disables the content guard. Use balanced as a starting point, then compare results if you change presets. The source does not provide numeric thresholds or promise a particular speedup for any option. There is no numeric minimum or maximum.
cache_device
Chooses where cached residuals are held. The options are auto, gpu, and cpu; the default is auto. The tooltip states that auto keeps residuals on the GPU only when enough free VRAM remains and otherwise uses pinned CPU memory. Choosing gpu or cpu requests that location directly. This setting concerns cache storage, not a change to the model’s output type; the source does not specify a fixed VRAM threshold or a guaranteed performance difference. There is no numeric minimum or maximum.
MiniMax H3 Adaptive Cache Output Parameters:
model
The single MODEL output is a patched clone of the connected model, ready to pass to the next compatible step in your sampling workflow. The node registers cache-aware replacements for the MiniMax H3 DiT blocks and adds a sampling lifecycle wrapper to manage the cache. It does not output an image, a cache file, or a performance measurement.
MiniMax H3 Adaptive Cache Usage Tips:
- Connect the native MiniMax H3
MODELand pass this node’smodeloutput onward in your sampling workflow. The cache is applied to the output model, not returned as a separate object. - Start with
balancedandauto, the provided defaults. They let you try the cache without first choosing a more aggressive preset or a fixed storage device. - Choose
safewhen preserving fidelity is your priority. Tryfastwhen you want to test whether skipping more tail blocks helps your workflow, and compare the resulting images with your usual output. - Use
sigma_onlyonly when you intentionally want the content guard disabled. Because that removes one of the checks named in the interface, compare results before relying on it for a final render. - Keep
autoif available VRAM varies between runs. If you choosegpuorcpudirectly, check performance on your own system; the supplied source does not establish which will be faster for every setup. - When combining this node with Sol-Attn, leave Morton disabled as specified in the node description.
MiniMax H3 Adaptive Cache Common Errors and Solutions:
This node only supports ComfyUI's native MiniMax H3 model
- Explanation: The node could not find the expected native MiniMax H3 model inside the connected
MODEL. - Solution: Check that the upstream loader supplies ComfyUI’s native MiniMax H3 model, rather than a different model architecture or an incompatible implementation.
The input MODEL does not expose ModelPatcher block replacement
- Explanation: The connected model does not provide the block-replacement method this node uses to install the cache.
- Solution: Connect a compatible ComfyUI model output with ModelPatcher block-replacement support. If another custom node wraps or converts the model first, try connecting MiniMaxH3AdaptiveCache before that step.
MiniMax H3 DiT blocks were not found
- Explanation: The node found the expected model location but could not find at least two DiT blocks to patch.
- Solution: Confirm that the connected model is a correctly loaded native MiniMax H3 model. If it is, check whether another node or an incompatible model implementation has changed its internal structure.
This ComfyUI build does not expose the OUTER_SAMPLE wrapper API required for safe cache lifecycle management
- Explanation: The node could not register the sampling wrapper it requires to manage the cache safely. The error can also report an underlying exception from that registration attempt.
- Solution: Use a ComfyUI build that supports the required
OUTER_SAMPLEwrapper API. If your build should support it, inspect the underlying exception for the more specific cause.
KeyError for a preset name
- Explanation: The node looks up the selected preset by name. A workflow containing a preset name unavailable in the installed version may fail at that lookup.
- Solution: Re-select one of the options shown by the node:
safe,balanced,fast, orsigma_only.
