Patch Sol-Attn

Patch Sol-Attn is a training-free speed hack for diffusion self-attention, dropped into ComfyUI by kijai - the person who basically is the ComfyUI video scene (WanVideoWrapper, KJNodes, day-one ports of every model that matters). It implements Sol-Attn (arXiv 2607.24027), a sparse-attention method that figures out on the fly which parts of a sequence are actually worth attending to and skips the rest. If you generate long video with MiniMax-H3 or Wan and attention is your wall, this is the kind of node you want to know about.

Fair warning: it's marked experimental, the README says it's been tested on an RTX 4090 and a 5090 with MiniMax-H3, and there are sharp edges. But it's genuinely clever - and unlike a lot of "speed" nodes, the quality dial is one honest slider.

How it works

Attention cost grows with sequence length, and video sequences are enormous. Sol-Attn's trick: chunk the sequence into 64-token blocks, compute a cheap summary per block, and use those summaries to predict which blocks each query actually cares about. Blocks that pass the threshold get full exact attention; the rest get an approximate pass from the block summaries. That's the whole idea - most of the attention matrix is near-zero anyway.

The tau slider is the throttle. It sets the routing threshold as mean + tau × std of the block-score distribution, so higher tau means fewer blocks kept exact: the tooltip says tau 1.0 keeps about 16% of blocks exact, 1.5 about 7%, 2.0 about 2.7%. There's no free lunch - higher tau is faster and slightly worse. 1.3 is the default and a sane starting point.

A few other inputs earn their keep:

• start_percent / end_percent - the paper runs dense before 20% of sampling (early steps set the composition), so start_percent defaults to 0.2. Leave it.
• min_tokens (default 4096) - sequences shorter than this stay dense. Short clips gain nothing here.
• dense_blocks - a string like 0-2,-1 that keeps specific transformer blocks exact. The first and last blocks are the fragile ones, because their error hits the output with no later block to absorb it.
• tau_profile (optional input) - per-block tau, e.g. 0-30=2.0 then 39-42=0.9. Block sensitivity genuinely varies several-fold across depth; the Block Probe exists to find the fragile blocks.
• morton - reorders Wan and MiniMax-H3 video tokens into Z-order so each 64-token block is a compact 3D neighborhood instead of a 2-row strip. Routing gets far more accurate at the same density. Logged and skipped on other models, so it's safe to leave on.

Then the stuff you can mostly ignore: int8_qk and int8_pv (INT8 in the exact branch, Sage-style - on by default, "measured free in quality"), sink_conditioning (MiniMax-H3 only; keeps the packed text/audio/reference rows exact so you don't lose prompt adherence), and use_tma (off by default because it "has not measured faster on any tested GPU" - it wants SM90+ and Triton 3.3+).

The output is a patched MODEL, straight into your KSampler. One input (your loaded model), one output.

Installing it

No model files, no extra checkpoints - it patches models you already have. ComfyUI Manager: search "ComfyUI-SolAttn_triton" and install, or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-SolAttn_triton

then restart ComfyUI. The real dependency is Triton - the kernels are pure Triton and compile on first use, so your first run is slower while it autotunes. On Linux, torch ships with a working Triton. On Windows it famously doesn't, and you'll likely need the community triton-windows wheel (the same pain SageAttention users have been through - if that install defeated you before, this is the same wall).

Where people get burned

• It silently does nothing on most models. The kernel is bf16 and head_dim 128 only. Masked attention, cross-attention, fp32, or a model with a different head dim all fall back to the existing attention backend - no error, no speedup. It's built for MiniMax-H3 and Wan, and the README only claims the H3 path.
• Silent fallbacks, noisy blocks. If a Triton kernel fails mid-run it logs [sol_attn] kernel failed and falls back dense, so watch the console if you're not seeing speedup.
• First run is slow. Kernel compilation, then autotune. Give it a real generation before judging.
• It also composes with KJNodes' memory-efficient Sage patch instead of fighting it - Sol-Attn takes eligible self-attention and hands the rest back.

Honest take: if you're on an NVIDIA card generating long H3 or Wan clips, it's worth 20 minutes. If you're on image models or short clips, it'll add a node and do nothing.