Configure Sol Sparse Attention

That video model you're sampling is spending most of its time attending over every one of a few thousand tokens - when, for any given 128-token window, most of those K/V tokens barely matter. Configure Sol Sparse Attention is the "stop doing that" switch. You drop a MODEL in, you get a patched MODEL out, and the sampling that follows computes far fewer attention blocks. It's a pure performance patch from wjie98's ComfyUI Turing Utils pack (the repo is comfyui-svdint4 - the folder name and the branding inside disagree, that's normal here), the same pack that brings you the ConvRot INT8 loaders. If you run long-context video - Wan, Bernini, MiniMax H3 - on a 20- or 30-series card with no fp8 acceleration, attention is where the seconds go, and this is the node that spends them differently.

How it works

Sol is a model-generic sparse-attention backend: no particular loader required. For every 128-token query block, it decides which 64-token K/V blocks to compute exactly and which to approximate. Each block gets a cheap proxy score, computed on the same prequantized INT8 Q/K tensors that the exact path uses, and a block is skipped when its score beats mean + routing_threshold × std. Skipped blocks aren't deleted - each is replaced by one 64-token centroid (the default 1x64 mode) that still contributes an online-softmax correction, so V comes out approximately right rather than missing. 2x32 keeps two residual centroids per skipped block for better fidelity without changing which blocks route. Whether the numeric path is integer (W8A8) or FP16 depends on the dense backend your model inherits from Load ConvRot DiT - w8a8 picks the INT8 Tensor Core path, sage/sdpa the floating one.

It's a "strategy" node, which is a rename and a small upgrade: the old Patch Sol Sparse Attention is still registered so saved workflows don't break, but the Configure version inherits its dense backend from the loader instead of asking you to flip a use_w8a8 checkbox. Use the new one in fresh workflows.

The inputs that matter

• routing_threshold - the dial. Default 1.0 matches the official Sol policy. Lower it and more blocks stay exact (slower, closer to dense); raise it and more get skipped.
• dense_prefix_layers / dense_prefix_steps - defaults 2 layers and 1 step keep the first layers and the first denoising step fully dense. This is the "don't let the first pass mangle the structure" guard; raise it if early frames look off, lower it for max speed.
• sparse_reference_image / _video / _audio - false, true, false by default. Long reference-video blocks get sparsified; keyframes, reference images, and dialogue/audio conditioning stay exact unless you deliberately flip them.
• prefix_policy - auto applies the model's own semantic layout; manual protects just the leading manual_prefix_tokens tokens.

Everything flows out through the single model output, straight into your sampler.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/wjie98/comfyui-svdint4
cd comfyui-svdint4
python -m pip install -v --no-build-isolation -e ./kernel

That last line is the part people miss. requirements.txt deliberately excludes the CUDA kernel so ComfyUI Manager won't silently kick off a compile - the kernel build is an explicit step either way, and it compiles for every visible GPU (a 2080 Ti + 3070 box builds 7.5;8.6 in one pass). GPU-less build hosts fall back to 7.5; set COMFYUI_TURING_UTILS_ARCH_LIST to override. You need a CUDA NVIDIA GPU (Turing/Ampere/Ada or newer), Python 3.10+, and PyTorch with CUDA. ConvRot models additionally want comfy-kitchen>=0.2.26.

Troubleshooting

• No speedup, no errors. The node falls back to the dense backend (eventually SDPA) when the kernel isn't built or the model didn't come through the ConvRot loader. Pair it with Load ConvRot DiT so the W8A8 integer path actually engages, and confirm the kernel installed.
• Short sequences auto-crossover to dense. Sol is pointless below its minimum sequence length - it only pays off on genuinely long contexts.
• Watch the dense guardrails. If dense_prefix_layers + dense_suffix_layers reaches the model's layer count, every layer runs dense and the sparse path silently disables. That's intentional, not a bug.
• Slow at high sparsity with video references? That's sparse_reference_video defaulting to true. If the reference video is the part that matters, flip it off.