Instructions to use cyankiwi/GLM-4.7-Flash-AWQ-4bit with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use cyankiwi/GLM-4.7-Flash-AWQ-4bit with Transformers:

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

pipe = pipeline("text-generation", model="cyankiwi/GLM-4.7-Flash-AWQ-4bit")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("cyankiwi/GLM-4.7-Flash-AWQ-4bit")
model = AutoModelForCausalLM.from_pretrained("cyankiwi/GLM-4.7-Flash-AWQ-4bit")
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-4.7-Flash-AWQ-4bit with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "cyankiwi/GLM-4.7-Flash-AWQ-4bit"
# 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-4.7-Flash-AWQ-4bit",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/cyankiwi/GLM-4.7-Flash-AWQ-4bit

SGLang

How to use cyankiwi/GLM-4.7-Flash-AWQ-4bit 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-4.7-Flash-AWQ-4bit" \
    --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-4.7-Flash-AWQ-4bit",
		"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-4.7-Flash-AWQ-4bit" \
        --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-4.7-Flash-AWQ-4bit",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use cyankiwi/GLM-4.7-Flash-AWQ-4bit with Docker Model Runner:
```
docker model run hf.co/cyankiwi/GLM-4.7-Flash-AWQ-4bit
```

Could you show the benchmark loss for this quantization?

by TomLucidor - opened Feb 3

Discussion

TomLucidor

Feb 3

I am curious how the perplexity loss might cause some of the agentic tasks to fail more often.

cpatonn

cyankiwi org Feb 3

Thanks for raising the problem to me. The perplexity measurement was made on wikitext dataset, and therefore, not covering tool usage and agentic task.

I am recently aware that the quantized models struggle in tool usage and agentic tasks, which might be a result from the lack of tool calling and agentic calibration dataset. The model was calibrated using nvidia/Nemotron-Post-Training-Dataset-v2, and nvidia/Nemotron-Post-Training-Dataset-v2 does not have tool calling and agentic calibration data.

I will look into this in more detail, and solve this problem soon.

TomLucidor

Feb 3

Ideally include benchmarks for most 4bit quants cus it is easier to see what might get broken by accident

cpatonn

cyankiwi org Feb 3

Yes, I intend to complete full evaluations of my models, but currently I'm limited by my resources.

whadupapp

Feb 11

how many samples of the post-training-dataset do you use for the calibration?

cpatonn

cyankiwi org Feb 11

@whadupapp For calibration, I use the 256 samples from nvidia/Nemotron-Post-Training-Dataset-v2 dataset, with tokens routed to every experts for calibration. Do you also get this problem?

@TomLucidor Could you tell me in more details of the failed agentic tasks? Do the tool-calling and agentic outputs get mixed in its thinking traces, and ultimately lead to agentic task failure?

whadupapp

Feb 11

I was just wondering how the number or the selecton of samples impacts the downstream performance of the model

whadupapp

Feb 11

I did notice the mixing of thinking and toolcalling for the glm4.5 awq quant, though

TomLucidor

Feb 12

@cpatonn there are many failure modes e.g. </think> tag parsing, failing to understand the prompt, thought/action loops, mojibake/jibberish due to quantization/artifacts, etc...
(I am also looking to see if Qwen3-Coder-Next-REAP and Kimi-Linear has similar issues, since liner models are likely faster)

cpatonn

cyankiwi org Feb 19

Thanks for the info. I stumble across this and I think it might be beneficial in your case.

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