Instructions to use cyankiwi/GLM-5.2-AWQ-INT4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use cyankiwi/GLM-5.2-AWQ-INT4 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="cyankiwi/GLM-5.2-AWQ-INT4")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("cyankiwi/GLM-5.2-AWQ-INT4")
model = AutoModelForCausalLM.from_pretrained("cyankiwi/GLM-5.2-AWQ-INT4")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use cyankiwi/GLM-5.2-AWQ-INT4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "cyankiwi/GLM-5.2-AWQ-INT4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "cyankiwi/GLM-5.2-AWQ-INT4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/cyankiwi/GLM-5.2-AWQ-INT4

SGLang

How to use cyankiwi/GLM-5.2-AWQ-INT4 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "cyankiwi/GLM-5.2-AWQ-INT4" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "cyankiwi/GLM-5.2-AWQ-INT4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "cyankiwi/GLM-5.2-AWQ-INT4" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "cyankiwi/GLM-5.2-AWQ-INT4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use cyankiwi/GLM-5.2-AWQ-INT4 with Docker Model Runner:
```
docker model run hf.co/cyankiwi/GLM-5.2-AWQ-INT4
```

Has anyone run GLM-5.2-AWQ-INT4 with vLLM on 4×A800 80GB? Any AWQ-INT2 plan?

by ming616 - opened 11 days ago

Discussion

ming616

11 days ago

Hi everyone,

I am evaluating whether GLM-5.2 can be practically deployed as a local long-context coding / agent model on an NVIDIA A800 server.

My hardware and target use case:

GPUs: 4 × NVIDIA A800-SXM4-80GB
GPU architecture: Ampere / sm80
CPU RAM: 512 GB
OS: Ubuntu 22.04
Target serving stack: vLLM + OpenAI-compatible API + LiteLLM / OpenCode
Main use case: coding agent, long-context repository analysis, log analysis, and multi-turn software development assistance
Concurrency requirement: low concurrency, usually 1–2 active sessions

I have already tested a GLM-5.2 GGUF quantized version with llama.cpp. It can run, and I can allocate 2 slots with 128K context each. However, the practical performance is not ideal for coding-agent workflows.

One of my tests showed:

n_slots = 2
n_ctx_slot = 131072
Prompt eval: around 20.5K tokens in about 50 seconds
Decode speed: around 14–15 tokens/s

This is acceptable for long-context reading, but too slow for interactive coding-agent loops.

Now I am considering cyankiwi/GLM-5.2-AWQ-INT4 with vLLM.

My questions:

Has anyone successfully run cyankiwi/GLM-5.2-AWQ-INT4 on 4 × A800 80GB with vLLM?
If yes, could you share your working configuration?
- vLLM version / commit
- CUDA version
- PyTorch version
- tensor parallel size
- --cpu-offload-gb
- --gpu-memory-utilization
- maximum stable --max-model-len
- whether prefix caching works
- whether reasoning / tool-call parsing works
What decoding speed and prefill speed should I realistically expect on 4×A800?
Are there known issues with GLM-5.2 DSA / sparse MLA on sm80 GPUs such as A100/A800?
Is there any plan or community work for a vLLM-compatible AWQ-INT2 / GPTQ-INT2 / compressed-tensors 2-bit checkpoint of GLM-5.2?
If AWQ-INT2 is not realistic, what is the most practical vLLM-compatible quantized version for 4×A800?

My current understanding is:

The full AWQ-INT4 checkpoint is around 410 GB, so 4×80GB cannot hold all weights plus KV cache purely in GPU memory.
Some CPU offload is probably required.
However, too much CPU offload may make it unsuitable for OpenCode-style coding-agent workflows.
GGUF 2-bit / 3-bit versions are usable for long-context reading, but the decode speed is not good enough for my main use case.

I would appreciate any successful deployment reports, startup scripts, benchmark numbers, or advice on whether this path is worth pursuing.

Thanks!

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment