JZL - 🚀 模型加载Pro

"JZL - 🚀 模型加载Pro" is the local-LLM end of the JZL-MiniMax-H3 pack's "漫剧" (short drama) pipeline. The pack's Script Processor uses a Qwen-class VLM to write storyboards and camera directions, and this node is where you load that GGUF model and set every sampling parameter the writing quality depends on. If you're only doing reference-to-video encoding, you don't need it - but if you want the LLM-written script chain, this is the loader.

How it works

It lists models from ComfyUI's LLM folder (the same folder the built-in GGUF loader uses) into the model dropdown, lists mmproj projector files for vision-language models, and offers 33 chat_handler presets - LLaVA-1.5/1.6, Moondream2, MiniCPM-v2.6, Gemma3, Qwen2.5-VL, Qwen3-VL and more - so the pack knows how to talk to the model you picked.

The interesting part is the backend switch. Default is llama-cpp-python (in-process). The newer option is llama-server: the pack no longer calls llama-cpp-python at all - it spawns a pinned llama.cpp b10436 binary as a separate subprocess, runs inference, then kills it, which isolates crashes and releases VRAM when the pass is done. That path requires running the pack's installer first (see below).

The inputs that matter

• model - your GGUF, from the LLM folder. mmproj and chat_handler follow from that choice.
• backend - llama-cpp-python or llama-server. Try llama-server if the in-process backend keeps wedging your ComfyUI.
• n_ctx - context length, default 32768, up to 262144. The tooltip's rule of thumb: a short story at 32K, a 56-segment epic at 131K–256K.
• temperature - default 0.6, which the pack says keeps [SHOT_START] formatting strict and reduces hallucinations.
• top_k (40), top_p (0.9), min_p (0.05), repeat_penalty (1.05) - the usual suspects, all with sensible defaults.
• max_tokens - 8192 default; scale up for many segments.
• vram_limit - GB cap, -1 = unlimited.
• gpu_device - only for the llama-server backend; auto follows ComfyUI's current card, or pick an index.

Outputs are llama_model and parameters, which plug straight into the pack's Script Processor.

How to install it

Pack install as usual - ComfyUI Manager, search "ComfyUI-JZL-MiniMax-H3", or git clone https://github.com/wjluoxiao/ComfyUI-JZL-MiniMax-H3 into custom_nodes, then restart.

If you want the llama-server backend (recommended if you hit in-process crashes), run the runtime installer once. On Windows double-click install_runtime.bat; otherwise:

cd ComfyUI/custom_nodes/ComfyUI-JZL-MiniMax-H3
python install_runtime.py          # --dry-run to preview, --force to reinstall

It's pure-Python stdlib, auto-detects your OS/arch/GPU, and downloads a pinned, SHA256-verified llama.cpp b10436 build - zero pip installs. If you already have llama-cpp-python, it won't conflict; you can even pip uninstall llama-cpp-python to reclaim disk.

Common issues

The big one: llama-server fails with "no runtime found" if you never ran the installer - it's a subprocess, so there's no pip fallback. Another classic is picking a chat_handler that doesn't match the model (e.g. Qwen3-VL handler with a Qwen2.5-VL file) and getting garbage tokens instead of [SHOT_START] blocks. And remember the model goes in the LLM folder, not models/checkpoints - the dropdown only sees the former.