MinusZone - ChatGLM3Loader

Kolors is unusual among diffusion models because its text encoder isn't CLIP and isn't T5 - it's ChatGLM3, a full Chinese-English LLM from Tsinghua's Zhipu team. That's a genuinely different way to turn your prompt into numbers, and MZ_ChatGLM3Loader is what loads that encoder into ComfyUI. It's the node that makes Kolors understand Chinese prompts properly, and it's also the single heaviest piece of this pack: the model is a ~10GB safetensors that lives in models/LLM/.

Why does it matter that it's an LLM? Because Kolors was an early example of the trend that later took over - using a general-purpose language model as the text encoder instead of a frozen CLIP. The practical payoff: it reads long, natural-language prompts well, and it handles Chinese natively. The community quickly noticed that feeding Kolors Chinese prompts (or having another Chinese-capable LLM draft them) often beats English. That's the whole reason this loader exists rather than a standard CLIP loader.

How it works

The loader reads your chosen safetensors, builds a ChatGLMConfig from a config bundled inside the pack, and constructs the model with the bundled tokenizer. Two details matter:

• It tries to use accelerate's init_empty_weights to load the ~10GB file without ballooning VRAM, falling back to a plain load_state_dict if accelerate isn't installed (you'll see a warning).
• If the filename contains 4bit or 8bit, it quantizes the model accordingly. The 4bit path tries to pip install cpm_kernels on the spot - that package is needed for quantized ChatGLM, and it's the source of one of the FAQ's most common errors (see below).

The output, chatglm3_model, is a CHATGLM3MODEL object - a dict holding the text encoder plus tokenizer. It feeds the MZ_ChatGLM3_V2 (or legacy MZ_ChatGLM3) text encode nodes, never a KSampler directly.

Inputs and outputs

• chatglm3_checkpoint - dropdown of .safetensors files in models/LLM/. The README's FAQ error 'ChatGLMModel' object has no attribute 'transformer' is this dropdown pointing at the wrong file, so double-check what you downloaded.
• Output: chatglm3_model (CHATGLM3MODEL) → text encode node.

Install and models

Install via ComfyUI Manager (search "ComfyUI-Kolors-MZ") or git clone https://github.com/MinusZoneAI/ComfyUI-Kolors-MZ into custom_nodes, then restart. The heavy dependency is the download:

# from Kijai/ChatGLM3-safetensors
# chatglm3-fp16.safetensors -> models/LLM/

Use the fp16 build. The README FAQ is blunt about it: RuntimeError: Only Tensors of floating point dtype can require gradients is fixed by switching to chatglm3-fp16.safetensors. fp32 loads if you insist, but on a consumer card it's a bad time.

Common issues

Three errors from the README's FAQ cover most of what people hit. name 'round_up' is not defined means cpm_kernels is missing - pip install cpm_kernels (only matters if you're on a quantized checkpoint). The gradient dtype error above means you grabbed the wrong model file. And a transformer attribute error means the file in the dropdown isn't actually the right ChatGLM3. Also worth knowing: Mac users should use the yiwangsimple fork of this pack, per the README - the original leans on CUDA assumptions in places.