MinusZone - deprecated - CLIPTextEncode(BaseLLamaCPP)

deprecated - CLIPTextEncode(BaseLLamaCPP) is the "base" variant of the pack's prompt beautifier: everything the v1 Llama-3/Phi-3 nodes do, except instead of a preset dropdown you type a raw path to your own GGUF. That's the entire differentiation, and it's also why it got deprecated - the current ModelConfigManualSelect(LLamaCPP) gives you the same "my own model file" flexibility with a dropdown, plus a config object that interchanges with the downloader.

The README describes these base nodes as "可以直接传入模型路径 / you can directly pass in the model path" - they were the escape hatch before the model-config system existed. The llama_cpp_model input is a plain STRING field (placeholder model_path) where you drop the path to any GGUF llama.cpp can load. The rest of the UI is the standard v1 common input: prompt_version (v1, fixed), style_presets (none / high_quality / photography / illustration), text, keep_device, seed, plus optional clip, llama_cpp_options, and customize_instruct. Outputs: text and conditioning.

Why it exists and why it's gone

The value was always "bring your own model," and the mechanism is simple: it calls the same beautify pipeline as the Llama-3 node but skips the downloader logic, so you point at whatever file you already have - a Chinese-tuned Llama, a different quant, a specialist fine-tune. The modern equivalent does this better: ModelConfigManualSelect(LLamaCPP) scans ComfyUI/models/gguf, gives you a real dropdown, and passes a config object forward. There's no reason to type paths when the successor exists - unless you're debugging an old graph that references it, or you genuinely want to hand a file path from some node upstream. Legacy compat is the honest use case.

Installing

It ships in the v1 subpackage, registered on startup:

cd ComfyUI/custom_nodes
git clone https://github.com/MinusZoneAI/ComfyUI-Prompt-MZ

Restart. Depends on llama-cpp-python + diskcache; the GGUF is yours to provide, no auto-download.

Troubleshooting

Because it's a raw path, the number-one failure is a wrong path - the model must actually exist where you typed it, and relative paths resolve oddly. Absolute paths are the safe move. The pack's usual llama-cpp-python friction applies: LLAMA_SPLIT_MODE_LAYER missing → upgrade llama-cpp-python; LLama.dll load failure → CUDA 12.1 PyTorch; wheel build failure → network. And remember this is a text LLM node - pointing it at a vision GGUF (or anything non-instruct) produces nonsense prompts, not an error. If you find yourself reaching for this node, check whether ModelConfigManualSelect(LLamaCPP) already covers the case with less typing.