MiniMax H3 Mem Eff Sage Attention Patch (KJ Alternative)

MiniMax's new H3 video model dropped at the end of July 2026, and the two things everyone said were "wow" and "how does this fit on my card." It's huge - the top comment on the open release was literally "it should not need a supercomputer" - and its first bottleneck in ComfyUI is peak VRAM. This node exists for that second question.

What it is

A one-in, one-out patch: MODEL in, MODEL out. It rewrites the attention forward pass inside MiniMax H3's transformer blocks to use the KJ-style SageAttention kernel, which is what trims the peak memory spike during generation. You load your H3 model, run it through this node, and feed the result to the sampler. There are no other inputs - no kernel picker, no sliders, no modes. It is aggressively simple, because all the complexity lives in the requirements instead.

The requirements are the story

The README is blunt about it: this needs a ComfyUI build with MiniMax H3 support, the latest sageattention, Triton, CUDA, and a supported NVIDIA GPU architecture. Read that as a shopping list, not boilerplate. If your ComfyUI can't load H3 natively, this node has nothing to patch. If your sageattention is stale, the kernel swap fails at load time. This is not an install-one-thing-and-win situation - it's the last mile of a bleeding-edge stack.

What it buys you

Lower peak VRAM during generation - the difference between OOMing on a scene and actually getting it out, or between tiling and not tiling. It is not a magic shrinking ray. It trims the attention memory spike; it does not make an 8GB card comfortable with a model this size. And the node is marked EXPERIMENTAL, which is the honest label for "the model is weeks old and so is this patch."

Install

ComfyUI Manager, search "ComfyUI_Swwan", or:

cd ComfyUI/custom_nodes
git clone https://github.com/aining2022/ComfyUI_Swwan
pip install -r ComfyUI_Swwan/requirements.txt

Then the SageAttention stack - pip install sageattention plus Triton and CUDA matching your ComfyUI's torch - exactly as with the pack's plain Patch Sage Attention node. All of the Windows compile pain transfers over; all of the "no module named sageattention" troubleshooting does too.

Troubleshooting, briefly

• Node errors on load → missing or incompatible sageattention/Triton. Version-check them against your torch before changing anything else.
• Node loads but sampling errors → your ComfyUI build probably doesn't actually have H3 wired up, or your GPU architecture isn't covered by the kernel.
• A ComfyUI update breaks it → that's the model being three weeks old, not you doing it wrong. The H3 ecosystem is moving in real time; expect patches to chase it.