Load MiniMax H3 Diffusion Model (SGLang)

If you've run MiniMax H3 natively in ComfyUI, you know the pain. It's a 33B omni-modal video model - text, image, video and audio in one context - and the full weights land around 42.5 GB. A single GPU chewing through a 20-second 1080p clip is a four-hour affair. This node is the pack's whole pitch in one box: it keeps H3 denoising in your ComfyUI graph but ships each denoiser evaluation to a local SGLang worker pool spread across your GPUs. The README's paired benchmark claims up to 8.21× faster on warm heavy workloads (with the honest footnote that's eight GPUs versus one native GPU - hardware is the whole story here).

It's a drop-in replacement for the Load Diffusion Model node in Comfy-Org's official H3 text-to-video, first/last-frame image-to-video, and reference-to-video workflows. Same MODEL output, same sockets, so you swap it in, and the existing conditioning, guider, sampler, scheduler, latent and VAE decode nodes never notice. If you've got the H3 templates running at all, this is a five-minute substitution.

How it works

ComfyUI keeps the sampling loop and does prompt encoding, scheduling, VAE decoding and output encoding in the host process. What changes is the middle: the current H3 latent streams are sent to SGLang workers for each denoiser step, using the tensor-parallel (TP) and Ulysses layouts you pick. The first generation starts the workers - that's your "cold" run. Consecutive compatible generations reuse the runtime and the loaded checkpoint, which is where the real speedup lives. Before a queued workflow that doesn't use this pack starts, the pack unloads its workers so native ComfyUI models can reclaim the VRAM. Change the checkpoint or the topology and the runtime is rebuilt.

The inputs that matter

There are exactly two, and neither is a free-text field:

• model_name - any compatible H3 .safetensors in ComfyUI/models/diffusion_models. The filename must contain fl2va or ref2va; the pack checks and refuses anything else. Full BF16 and ComfyUI pruned INT8 ConvRot exports both work. Pick FL2VA for text-to-video and first/last-frame i2v, Ref2VA for the image/video/audio reference workflows.
• topology - a TP/Ulysses layout generated from how many accelerators ComfyUI can see. H3's sharding constraints mean only certain splits are legal, so the node only offers valid ones. On a single GPU you'll still get a TP1 option - the parallelism is optional, not mandatory.

The only output is model (a standard MODEL), which wires straight into the rest of the official graph.

Installing it

Install the pack once - via ComfyUI Manager (search "H3 SGLang Pack") or:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/TensorClay/ComfyUI-H3-SGLang-Pack.git

Then restart ComfyUI. Note the pack itself declares zero pip dependencies. The heavy lifting - SGLang Diffusion - has to be in the same Python environment as ComfyUI, installed separately via the official SGLang Diffusion docs. That's the real friction point of this whole pack, not the node install.

Gotchas

Where people get burned:

• "No compatible CUDA/ROCm GPUs detected" on the topology field means the worker can't see your GPUs - check the SGLang Diffusion install and that the accelerators are visible to ComfyUI's PyTorch process.
• Worker-local ControlNet tensors and arbitrary transformer patch callbacks don't survive the trip through the standard MODEL interface. For attention, LoRA and caching, use the pack's own adapters (that's what its other five nodes are for).
• This is a young, small pack and the author says so in the README: "If you couldn't tell from the tone of the README, this whole thing was vibe-coded. Use at your own risk!" Treat the first update after a ComfyUI or SGLang version bump as mandatory.
• And the elephant in the room: the H3 weights themselves are geofenced. The MiniMax H3 Community License excludes the US, EU, UK and South Korea from its Applicable Territory - running or even displaying output from the local weights there isn't licensed. The hosted Hailuo API is the only licensed path in those regions. Worth knowing before you build a workflow around this.