Auto CLIP Loader (MultiGPU - Safetensors & GGUF)

The text encoder is the silent VRAM hog of the FLUX era. A T5-XXL in fp16 is around 10GB all on its own, which is brutal when the model you actually want to sample with is also begging for memory on the same card. AutoCLIPLoaderMultiGPU is the fix from the ComfyUI-AutoMultiGPU pack: a single-CLIP loader that loads the encoder and deliberately keeps it on your second GPU, so GPU0's VRAM belongs to the sampler. It handles both .safetensors and .gguf text encoders, and it covers every architecture the pack knows about.

The pitch, honestly: if you have a second card sitting idle, this is the easiest way to make it useful. Offloading the encoder to the spare GPU frees the main card for latents and the model - often the difference between OOMing at the prompt-encoding step and not.

How it works

It scans ComfyUI's text_encoders, clip, and clip_gguf folders (whichever exist), then loads the selected file onto your chosen device. The type field is the interesting part - a 37-option architecture picker (flux, flux2, sdxl, sd15, sd3, wan, ltxv, hunyuan_video, hunyuan_image, minimax, qwen_image, and so on). That's not decoration: it maps to the right ComfyUI CLIPType, which is what tells the loader whether your file is a T5, a CLIP-G, a Qwen3, or something else entirely. Leave it on auto and it guesses from the filename; set it manually when the filename lies.

For GGUF files specifically, it taps into ComfyUI-GGUF's loader and patches the returned ModelPatcher so the quantized encoder loads on your target device. That's the catch - see below.

The inputs that matter

• clip_name - the text encoder file. Both formats appear in the same list.
• type - architecture; auto is fine for most files.
• target_device - defaults to "Auto (Secondary GPU - cuda:1)", which picks cuda:1 when present, else cuda:0, else CPU.
• cpu_offload - auto_smart_offload (default), always_offload_to_cpu, or disabled_vram_only.

Output is a single CLIP, which wires into CLIP Text Encode / DualCLIPTextEncode like any encoder.

Installing it

ComfyUI Manager (search ComfyUI-AutoMultiGPU), or:

cd ComfyUI/custom_nodes
git clone https://github.com/nexusfinancial-dev/ComfyUI-AutoMultiGPU.git
cd ComfyUI-AutoMultiGPU
pip install -r requirements.txt

The pack itself only needs torch, accelerate, and safetensors - nothing exotic. But GGUF files will not load until you also install ComfyUI-GGUF; without it, the node raises "ComfyUI-GGUF is required." That dependency is in the code but not the README, and it's the #1 way this pack surprises people. Plain .safetensors encoders need no extra install.

Where people get burned

Loading a GGUF without ComfyUI-GGUF installed is the big one. Second: don't pick cpu for the target and expect speed - CPU encoding works, but it's a wall you'll feel on every prompt. And remember a 10GB fp16 T5 still needs to live somewhere; moving it to the second card only helps if that card has the free VRAM and the bus bandwidth to talk to the first one (see the DisTorch article in this pack for the full bandwidth lecture).