Patch Comfy Kitchen Attention

ComfyUI's attention backend system got a lot more interesting in 2026. Beyond the classic xformers/sageattention choices, there's now "Comfy Kitchen" - a backend library inside recent ComfyUI builds that ships optimized attention implementations, including an INT8 quantized path. If you've seen comfy_kitchen appear in your console logs, that's it. Patch Comfy Kitchen Attention is a one-input node that swaps the connected model's attention over to Comfy Kitchen's INT8 attention at runtime, with a safe fallback if your build doesn't have it.

Why bother? INT8 attention is a memory and speed play - quantized attention kernels can cut attention memory traffic and often speed up the bottleneck on big transformer models. It's the kind of thing that pairs naturally with a cache node on a heavy model like MiniMax H3, which is exactly the pack's intent: the README shows it chaining with MiniMax H3 Cache on the model path before the guider/sampler.

How it works

This is the model-clone pattern done carefully. The node:

1. Checks whether ComfyUI's attention module has Comfy Kitchen INT8 available (COMFY_KITCHEN_INT8_ATTENTION_IS_AVAILABLE).
2. If yes, grabs the comfy_kitchen_int8 attention function and clones the connected MODEL, setting that as the clone's optimized-attention override.
3. If no, it clones the model anyway and sets the ComfyUI default attention, logging the decision. No crash, no monkey-patch of the global model class - your other workflows keep whatever attention they had.

The MODEL socket in, MODEL out. Nothing else. It's as minimal as a node gets.

Installing it

Same as every other node in this pack - ComfyUI Manager (search DaSiWa-Nodes), or:

cd ComfyUI/custom_nodes
git clone https://github.com/darksidewalker/ComfyUI-DaSiWa-Nodes
pip install -r requirements.txt

then restart. One caveat specific to this node: it needs a modern-enough ComfyUI. The source raises a clear error if your ModelPatcher lacks set_model_optimized_attention(), and the INT8 backend only exists in ComfyUI builds that ship Comfy Kitchen. On an old build you get the fallback (default attention) plus a console log line - it silently does nothing fancy rather than breaking.

Where people get burned

• "It didn't speed anything up." Possible - Comfy Kitchen INT8 may be unavailable on your build, in which case the node logs that it fell back to default attention. Check the console for the log_dasiwa line before assuming the patch took effect. Or your model just isn't attention-bound.
• Chaining order confusion. Both this node and MiniMax H3 Cache are model-clone patches, and they compose in either order - but the cache's hit steps bypass attention entirely, so the INT8 patch only pays off on cache-miss steps. If you want to measure the attention patch's effect, test it with the cache bypassed.
• Expecting a global speedup. It's scoped to the one model you feed it. That's the feature - no risk to the rest of your graph - but it also means one wire per model you want patched.

Honest take: for most people on mid-range hardware this is a "try it and measure" node, not an obvious win. It's one wire and a queue - if your H3 (or other heavy) generation is VRAM-pressured or attention-bound, it's worth the five minutes to test. If you're on a build without Comfy Kitchen, you're not missing much: you get the fallback and a log line that tells you so.