Director Strict MiniMax H3 Low-VRAM Sage Patch

This node's whole personality is one word: strict. Feed it a MiniMax H3 model and it swaps every self-attention block onto SageAttention's memory-lean int8 kernels - and if it can't prove that path is going to work, it errors instead of quietly doing something else.

H3 is MiniMax's 33B open-weight video model: big, slow, and hard on VRAM. DirectorDeck, the pack this ships in, is a timeline-based director console that uses H3 to build and stitch long films - per-segment FL2VA/Ref2VA recipes, task orchestration, live preview, the works. Long stitching means running a lot of samples in sequence, and that's exactly where a silent swap of the attention implementation between runs becomes a seam you'd never notice until the film is done. That reproducibility gap is what the "strict" nodes in this pack exist to close.

How it works

Stock ComfyUI's Sage attention adapter is friendly: if the kernel raises, it falls back to PyTorch attention so your job still finishes. This node is the opposite. Before it patches anything it checks, in order, that your model is a genuine ComfyUI MiniMaxH3Model sitting on CUDA, that a compatible SageAttention build is importable, and that the exact kernel for your GPU's architecture is actually compiled - the supported set is Ampere through Blackwell (sm80, sm86, sm89, sm90, sm120, sm121). Only then does it clone the ModelPatcher and install one patch per attention block. There is deliberately no try/except around the kernel call: a Sage failure kills the run rather than swapping algorithms underneath you.

The "Low-VRAM" in the name is the payoff. The kernels it installs are the int8-quantized-QK family - int8 attention scores with fp16 (fp8 on newer archs) value projections - the memory-cheap path that squeezes a 33B video model onto cards it otherwise wouldn't fit. It's the same class of trick that makes Sage the go-to speed/memory lever for Wan and friends.

Inputs and outputs

Boring in the best way: one model (MODEL) socket in, one model (MODEL) out. Wire it from your H3 loader, wire the output into your sampler. No widgets, nothing to tune. It either patches the graph or it throws a descriptive error.

Installing it

This is a DirectorDeck node, so you install the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/JYE-HC/DirectorDeck.git
/path/to/ComfyUI/.venv/bin/python -m pip install -r DirectorDeck/requirements.txt

then restart ComfyUI (ComfyUI Manager works too - search "DirectorDeck"). The README suggests ComfyUI 0.33.0 or newer. One crucial gotcha: the pack does not install SageAttention for you. It's probed at runtime, and it's a host dependency you bring yourself. On Windows that means Triton, which the community knows as "notoriously painful" - a manual install that eats an afternoon when the wheel doesn't match your Python/torch. Make sure you install a SageAttention build that includes kernels for your actual GPU.

Where people get burned

• "sageattention_dependency_missing" - SageAttention isn't importable in the Python env running ComfyUI. Install it there, not in some other venv.
• "sageattention_kernel_unavailable" with an arch code - your SageAttention build lacks the kernel for your GPU. Upgrade or rebuild it; this is the "compatible build" part of the description doing real work.
• Wrong model type - it only accepts a real MiniMax H3 model patcher. Feed it anything else and it errors up front.
• Worth saying out loud: the H3 weights carry a community license that excludes the US, EU, UK and South Korea from running them locally. That's a licensing wall, not a tech one, and no node gets you around it.

And honestly - if you're just sampling a five-second clip, you probably don't need this node. ComfyUI's built-in Sage path or KJNodes will get you the speed with less ceremony. Reach for the strict version when the contract matters: Director long-film runs, reproducible multi-segment work, anything where "the run must be exactly what I asked for" beats "it mostly worked."