Comfyui-Sparse-Attention

Block-sparse (SLA) attention for MiniMax-H3 -- the inference path the lightx2v SLA turbo LoRA was distilled against.

By BetaDoggo·Created 2 months ago·Updated 2 months ago· 5

BetaDoggo/Comfyui-Sparse-SLA-Attention

Nodes1

On cloudLocal install

CategorySparseAttention

Stars5

Updated2 months ago

Nodes (1)

H3 SLA Attention

The missing kernel behind the lightx2v turbo LoRA

SparseAttention

Readme

Comfyui-Sparse-Attention

This is a standalone version of the ComfyUI-H3-SLA-Attention node from ComfyUI-PlagueKind-Nodes pack, all credit goes to the original authors listed below.

Important note

By default this node won't trigger at all on lower resolutions/lengths (ie: 512x512) because those generations fall outside the default 8192 sequence length. You can of course lower this limit, but results might be worse.

Requirements

triton

Credits

The node and wiring are (c) 2026 PlagueKind, extracted from ComfyUI-PlagueKind-Nodes under the MIT license -- see LICENSE.
sla/block_map.py and sla/kernel.py contain code vendored from LightX2V (Apache-2.0), with their original attribution headers intact -- see LICENSE-APACHE-2.0.txt.