GPT Loader Simple

GGUF's home turf, actually - the format started here, in llama.cpp, well before it jumped to diffusion models like Flux and became the standard way to shrink an image model onto a consumer GPU. This node is what lets ComfyUI load a GGUF language model the same way, wrapping llama-cpp-python's Llama() constructor so you get a text-generation model sitting inside your graph as a first-class node output, ready to feed into GPTSampler.

How it works

You point ckpt_name at a .gguf file, and the node spins up a llama.cpp instance around it. Three settings control how that instance runs: gpu_layers (default 27) is how many of the model's transformer layers get offloaded to your GPU - higher means more VRAM used but faster generation, 0 means pure CPU. n_threads (default 8) sets the CPU thread count for whatever isn't running on GPU. max_ctx (default 2048) caps the context window in tokens - how much prompt-plus-history the model can actually see at once.

The outputs that matter

Two: model, the loaded Llama instance itself, which is what GPTSampler needs as its own model input, and model_path, the resolved path string to the checkpoint you loaded - GPTSampler's bare version wants that too, so both outputs from this node feed directly into the sampler downstream.

Installing the pack, and where models go

• ComfyUI Manager - search "ComfyUI-N-Nodes," install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/Nuked88/ComfyUI-N-Nodes.git, then restart ComfyUI.

llama-cpp-python installs automatically as part of the pack setup now - the README is explicit that the old requirement to manually compile it against CUDA is gone, replaced by prebuilt CUDA wheels. Drop your .gguf files in a GPTcheckpoints folder under ComfyUI/models (created automatically), or point extra_model_paths.yaml at wherever you already keep GGUF LLMs if you don't want to duplicate storage.

Common issues & troubleshooting

ckpt_name dropdown is empty. Nothing's been placed in GPTcheckpoints yet, or your extra_model_paths.yaml entry doesn't point where you think it does. Confirm the file's actually a .gguf - this loader doesn't read safetensors or other diffusion-model formats, it's llama.cpp's format specifically.

Slow generation despite a decent GPU. Check gpu_layers - if it's low relative to your model's total layer count, most of the work is still landing on CPU regardless of what card you have. Push it up (the ceiling in this node is 100) until you either run out of VRAM or hit the model's actual layer count, whichever comes first.

Not sure which GGUF quant to grab. The same Q-ladder logic that applies to diffusion GGUF applies here: Q8 is close to indistinguishable from full precision at half the size, Q4-ish tiers are the practical floor before quality drops noticeably. Bigger LLM checkpoints benefit the most from dropping a tier or two if VRAM is tight.

Looking for a general-purpose chat node and don't need the vision-model extras. This loader is fine for that, but it's worth knowing it isn't the only option in the ecosystem - you'll also see people pointed toward newer, narrower LLM packs like Searge-LLM for straightforward text-in, text-out work. This one earns its keep specifically because it also drives the vision models (see GPT-Loader-Simple [n-suite]) in the same node family.