MiniMax H3 SLA Attention Config:
This node creates a configuration for block-selected sparse attention in MiniMax H3. Sparse attention selects a limited set of attention blocks rather than using the existing dense attention path throughout a sampling run. The configuration lets you decide when sparse attention is appropriate, preserve important text and audio information, and adjust settings related to numerical precision and routing stability. It does not patch a model by itself; you pass its configuration to a compatible attention patcher.
MiniMax H3 SLA Attention Config Input Parameters:
minimum_sequence_length
The sequence-length threshold below which the existing dense attention path is kept. The default is 8192; the allowed range is 0 to 1000000, in steps of 1024 in the node interface. Raising this value keeps more shorter sequences on the dense path, while lowering it allows sparse attention to be considered for shorter sequences. This setting defines a threshold, not a guaranteed speed improvement.
dense_steps
Sampling steps that should keep dense attention instead of using sparse attention. Enter zero-based step numbers separated by commas, with optional inclusive ranges: 0,3-5 means steps 0, 3, 4, and 5. The default is 0, so the first sampling step is kept dense. This is a single-line text field; the node interface does not specify a numeric minimum or maximum for its contents. Use it when you want selected steps to retain the dense path while other eligible steps use the sparse configuration.
protect_audio
Controls whether text and audio token ranges are kept in every sparse attention selection. The options are True and False, and the default is True. Leave it enabled when preserving those ranges matters to your workflow. Turning it off removes this particular protection; the provided source does not establish what visual or audio change that will produce in a given render.
disable_fp16_accumulation
Controls whether FP16 and BF16 reduced-precision matrix-multiplication accumulation is disabled for this SLA sampling run. The options are True and False, and the default is True. This is a numerical-computation setting rather than a direct artistic control. If you change it, compare results in your own workflow; the provided source does not guarantee a particular quality, speed, or memory outcome.
stabilize_routing
Controls whether near-cutoff routing is biased toward the prior sampling step. The options are True and False, and the default is False. Enabling it is intended to reduce unstable motion detail by making borderline attention selections more consistent from one step to the next. It is most relevant when you observe that kind of instability; it does not guarantee that all motion artifacts will disappear.
MiniMax H3 SLA Attention Config Output Parameters:
SLA Config
A MINIMAX_H3_SLA_ATTENTION_CONFIG value containing the five settings supplied to this node, after validation by the configuration implementation. Connect it to the Unified Attention Patcher when using Sparse / MiniMax H3 SLA. This output is a configuration for the patcher to use, not a patched model or a generated image.
MiniMax H3 SLA Attention Config Usage Tips:
- Start with the defaults, including
minimum_sequence_lengthset to8192anddense_stepsset to0, then change one setting at a time so you can identify its effect. - If you want additional sampling steps to remain dense, enter them in
dense_stepsas zero-based numbers or inclusive ranges, such as0,3-5. Check that the step numbers fit your sampling run. - Keep
protect_audioenabled when your workflow includes text or audio token ranges that you want included in every sparse selection. - If motion detail appears unstable, try enabling
stabilize_routingand compare it with the default result. If you are testing precision behavior, changedisable_fp16_accumulationseparately rather than changing both settings at once. - Connect
SLA Configto the compatible Unified Attention Patcher and select Sparse / MiniMax H3 SLA there; this configuration node alone does not activate sparse attention.
MiniMax H3 SLA Attention Config Common Errors and Solutions:
Configuration validation error
- Explanation: The node calls a configuration validation method before returning its output. The provided source does not show that method’s checks or its exact error messages, so a validation failure may have implementation-specific wording.
- Solution: Recheck all five inputs. Start with the documented defaults, confirm that
minimum_sequence_lengthis within0to1000000, and use simple zero-based entries such as0indense_stepsbefore adding ranges.
Invalid value for minimum_sequence_length
- Explanation: The node interface accepts an integer from
0to1000000, with a displayed step of1024. A value outside the allowed range may be rejected by the interface. - Solution: Enter an integer within the stated range; use the default
8192if you are unsure where to start.
Invalid dense_steps entry
- Explanation: The field is intended for comma-separated zero-based step numbers or inclusive ranges. The provided source does not expose the parser or an exact error string, but an entry that does not follow that format may fail validation.
- Solution: Use a format such as
0,3-5, avoid extra wording, and make each range run from a lower step number to a higher one.
Type mismatch when connecting SLA Config
- Explanation:
SLA Confighas the dedicatedMINIMAX_H3_SLA_ATTENTION_CONFIGtype. It is not a model, image, or generic attention output, so an incompatible socket may refuse the connection. - Solution: Connect it to the configuration input of a compatible Unified Attention Patcher when using Sparse / MiniMax H3 SLA.
Sparse attention does not appear to be active
- Explanation: This node only creates a configuration. Also, sequences below
minimum_sequence_lengthretain the existing dense path, and steps listed indense_stepsremain dense. - Solution: Confirm that the configuration is connected to the patcher with the appropriate mode selected. Then check whether your sequence length meets the threshold and whether the steps you are examining are listed in
dense_steps.
