Load LoRA (SGLang)

The stock Load LoRA node patches the model in the ComfyUI process. Which is a problem here, because with this pack your H3 model isn't in the ComfyUI process - it's inside SGLang workers on however many GPUs you pointed at it. A patch callback in the host never reaches the transformer that's actually denoising. Load LoRA (SGLang) is the worker-native replacement: same MODEL in, MODEL out, but the adapter gets applied on the workers where it can do something.

The inputs that matter

Three inputs, and honestly only two you'll touch:

• model - has to come from Load MiniMax H3 Diffusion Model (SGLang) or another node in this pack that returns one. Feed it a stock loader's model and the node throws a clear TypeError explaining exactly that.
• lora_name - any .safetensors LoRA sitting in ComfyUI/models/loras.
• strength_model - the transformer adapter strength, default 1.0, and unusually permissive: negatives and values above 1 are allowed, so you can push past 1.0 or invert a style. Set it to 0 and the node passes the model through untouched.

The output is a MODEL, and the wiring note is the fun bit: you can chain multiple of these between the loader and the guider to stack adapters, which is exactly how you'd do it with the stock node. Drop them in where the old Load LoRA or Load LoRA Model Only sat and keep the existing links.

How it works

The LoRA's tensors are transferred to the worker runtime and cached there for as long as the runtime is alive, so the first generation pays the transfer cost and subsequent ones don't. Changing or removing the LoRA restores the base transformer weights before the next generation - no silent carryover if you swap adapters mid-session.

There's also a genuinely nice compatibility detail buried in the README: the stock Load LoRA and Load LoRA Model Only nodes still work with H3. If you wire them into an SGLang graph, the pack picks up their linear H3 LoRA, LoHa and LoKr patches and translates them exactly into a temporary worker LoRA. So you're not forced into this node - it's the cleaner path, and the direct one.

What it won't do

DoRA, convolutional or reshaped adapters, and dense diff/set weight patches are not supported - the translation only covers plain linear transformer adapters. If your favorite H3 LoRA refuses to behave through the stock path, that's the likely reason. Nothing to fix on your end; it's a pack limitation.

Installing it

Same pack install as everything else here - ComfyUI Manager, search "H3 SGLang Pack", or git clone https://github.com/TensorClay/ComfyUI-H3-SGLang-Pack.git into ComfyUI/custom_nodes, then restart. No extra pip dependency for LoRA support; the SGLang Diffusion prerequisite is the one that matters, and it has to live in ComfyUI's Python environment. The usual young-pack caveat applies: the author's own README calls it vibe-coded, so expect rough edges and test after any SGLang update.