Glm_4_9b_Chat

The pack is called ComfyUI_ChatGLM_API, but this node is the name's biggest lie - it calls no API and needs no key. Glm_4_9b_Chat loads Zhipu AI's open THUDM/glm-4-9b-chat model straight off HuggingFace and runs it on your own GPU via transformers. If you want a local LLM writing, translating, or expanding prompts inside your graph without sending text anywhere, this is the one. It's a 2024-era take on what's since become a whole genre of "LLM-assisted prompting" - the idea that if your image model's encoder reads natural language, having an LLM write that language is a translation between two things that already speak the same dialect.

How it works

Simple under the hood: the node takes your user_content, pushes it through AutoTokenizer.apply_chat_template(), loads AutoModelForCausalLM in bfloat16 on CUDA, and runs a greedy-ish generation with do_sample=True plus top_k. The reply is decoded with special tokens stripped, and out comes a plain string. The model is a ~9B-parameter GLM-4, so "local" here means you need serious hardware - figure ~20GB VRAM in bf16.

The inputs that matter

• repo_id - a forced input, meaning it has to be wired, not typed. It's meant to come from the pack's Glm_Lcoal_Or_Repo selector node (or any string you feed it). This is the model's HF repo path.
• user_content - your message. Multiline, so you can paste a chunk of story, a draft prompt, or a whole request.
• max_length (default 2500) - the total generation window. Your output can't exceed this; drop it to 500–1000 for snappier prompt rewrites.
• top_k (default 1) - set to 1 and sampling is effectively greedy; crank it toward 50–100 for varied, looser output.
• reply_language - an enum that quietly appends "answer me in English / 用中文回复我 / …" to your prompt, so you don't have to write the instruction yourself. Original_language leaves your input language alone.

Output: a single STRING named prompt. Wire it into a CLIP text encode as your conditioning, a show-text node, or a Stringify-style utility - that's the whole point of it living in ComfyUI.

Installing it

The node comes with the pack, so install the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/smthemex/ComfyUI_ChatGLM_API

Then restart ComfyUI. ComfyUI Manager works too - search "ComfyUI_ChatGLM_API". The local nodes need no config.json and no API key; that's only for the Zhipu cloud nodes in the same pack.

One trap the README itself warns about: this pack pins transformers==4.41.1 in requirements.txt, but the GLM-4 repos need a newer transformers to load at all. If you're using this node, update transformers - pip install -U transformers - and don't be surprised that the old pin disagrees with what the code actually wants.

Common issues

• First run downloads ~20GB from HuggingFace. It looks frozen; it isn't. Run it once, wait, then it loads locally.
• ImportError on startup - the pack's requirements list misses two imports the code uses (jwt and cv2). If the nodes won't register, pip install PyJWT opencv-python fixes it.
• OOM - hardcoded device = "cuda" with bf16 on a 9B model. Under ~20GB VRAM, you'll want a quantized GLM-4 from a GGUF-style loader instead; this node won't do it for you.
• If the reply language comes out wrong, remember the node only appends a language instruction - GLM-4 will usually follow it, but it's not a hard constraint.

It's an old, low-profile pack (all six nodes hover near zero usage), and this node is the most demanding of the lot. If you have the VRAM and want a fully offline GLM-4 chat that can hand its answer straight to your sampler's text encoder, it works - just budget the download and the RAM.