Load CLIP (Quadruple)

HiDream is the model that asks the most of your text-encoder VRAM budget: four separate files, loaded together, with one of them being a full 8B-parameter Llama. QuadrupleCLIPLoaderDisTorch2MultiGPU is the DisTorch2 family's entry for that exact loader - the naming puts it alongside the pack's other offload-capable nodes (the Dual CLIP and VAE loaders covered elsewhere on comfy.icu), meant to give you the same kind of device/virtual-VRAM control over where those four files sit.

What the recipe actually is

Per the node's own tooltip, HiDream wants long clip-l, long clip-g, t5xxl, and llama_8b_3.1_instruct - two encoders inherited from the SDXL lineage plus a T5-XXL plus an instruction-tuned Llama. That last file alone rivals the size of a full diffusion checkpoint. This is the clearest case in the whole pack of the text encoder stack outweighing the model it's conditioning: if you're VRAM-constrained on HiDream specifically, the encoders - not the UNet - are usually where the budget actually goes.

(Worth flagging so you don't chase a ghost: this is the original HiDream text-encoder setup - HiDream-I1/Full. The newer pixel-space HiDream-O1 has no VAE and no separate text encoder at all, so it never touches this node.)

Inputs and output

• clip_name1 through clip_name4 - your four encoder files from ComfyUI/models/clip, in the order HiDream expects (clip-l, clip-g, t5xxl, llama).

Output is a single CLIP, wired into CLIPTextEncode like any other.

A straight note on what's actually exposed: as registered on comfy.icu today, this node's inputs are the same four filenames as the plain QuadrupleCLIPLoaderMultiGPU - the device / virtual_vram_gb / donor_device / expert_mode_allocations controls you'll see on this pack's Dual-CLIP and VAE DisTorch2 loaders aren't showing up here yet. If you're relying on this node specifically for the offload sliders, check ComfyUI Manager for a pack update, or use the node's own Help panel (right-click → the question-mark icon) to confirm what your installed version exposes before you build a workflow around it.

Installing it

ComfyUI Manager - search "ComfyUI-MultiGPU," install, restart - is the path the maintainer recommends. Manual install:

cd ComfyUI/custom_nodes
git clone https://github.com/pollockjj/ComfyUI-MultiGPU

No extra dependencies beyond ComfyUI itself; this pack wraps existing loader nodes rather than adding a new inference stack.

Common issues & troubleshooting

Out of memory loading the four encoders, before generation even starts. Expected on anything under 24GB - the Llama component is the culprit. Until the offload controls surface on this specific node in your install, your options are running HiDream on a card with more headroom, or dropping to a quantized (GGUF) version of the Llama/T5 components if your workflow supports it.

Files loaded in the wrong slot. Nothing validates that clip_name3 is actually a T5 file rather than a stray CLIP export - a silent mismatch shows up as garbled or ignored conditioning, not an error. Match HiDream's model card exactly.

New files not appearing in the dropdowns. Refresh the ComfyUI browser tab after adding files to models/clip; the node definitions don't pick up new files without a reload.