Instructions to use lukealonso/GLM-5.2-NVFP4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use lukealonso/GLM-5.2-NVFP4 with Transformers:

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

pipe = pipeline("text-generation", model="lukealonso/GLM-5.2-NVFP4")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("lukealonso/GLM-5.2-NVFP4")
model = AutoModelForCausalLM.from_pretrained("lukealonso/GLM-5.2-NVFP4", device_map="auto")
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 lukealonso/GLM-5.2-NVFP4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "lukealonso/GLM-5.2-NVFP4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "lukealonso/GLM-5.2-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/lukealonso/GLM-5.2-NVFP4

SGLang

How to use lukealonso/GLM-5.2-NVFP4 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 "lukealonso/GLM-5.2-NVFP4" \
    --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": "lukealonso/GLM-5.2-NVFP4",
		"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 "lukealonso/GLM-5.2-NVFP4" \
        --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": "lukealonso/GLM-5.2-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use lukealonso/GLM-5.2-NVFP4 with Docker Model Runner:
```
docker model run hf.co/lukealonso/GLM-5.2-NVFP4
```

TP5 RTX 6000?

by jpsequeira - opened Jun 18

Discussion

jpsequeira

Jun 18

I can maybe squeeze 100k context at FP8. Would it be possible?

Ununnilium

Jun 18

You mean using 6 GPUs in tensor parallel mode for this model? As far as I know, the number of attention heads (64 for GLM-5.2) must be divisible by the number of GPUs in tensor parallel mode. Therefore, 5 doesn't work, but e.g. 4, 8 or 16 work. You can try data parallel mode, but this is much slower and using just llama.cpp with a GGUF quant could work better.

I was thinking to create a mixed size quant (e.g. "W2A16", "W3A16", "W4A16", "W8A16") with Intel AutoRound which would fit in 4 RTX Pro 6000 GPUs. The problem is that I think my hardware is not enough to handle the unquantized GLM-5.2 model. It is unclear for me if such a mixed quant with e.g., "W2A16" or "W3A16" layers works with good performance in vLLM or SGLang. Maybe someone with more experience in quantization could try to make such a quant?

jpsequeira

Jun 18

I did mean 5. I thought maybe with MiniMax craze and TP3 we would get lucky. But I guess not. Able to run on 4x would be wonderful.

archib4

Jun 19

•

edited Jun 19

I did mean 5. I thought maybe with MiniMax craze and TP3 we would get lucky. But I guess not. Able to run on 4x would be wonderful.

This is a common misunderstanding I also believed when I started.

You need a even number for TP like the comment above said regardless of model here on hf.

You can run 4 TP and 1 PP for a total of 5 but never odd number tensor paralism.

PP (Pipeline Parallelism) you loose features with so I wouldn't run 5 PP.

Ununnilium

Jun 19

To my knowledge, first the number of attention heads must be divisible by the number of GPUs for tensor parallelism. So if a model would be designed with 50 attention heads, tensor parallelism should theoretical be possible with 2 or 5 GPUs, but not with 4 (but other limitations could apply that the uneven number of 5 still doesn't work, I don't know it). I have access to 6 RTX 6000 Pro and cannot run GLM-5 with tensor parallel because 64 % 6 != 0. With pipeline parallel (normally only used to run a model on multiple servers) in SGLang, I got less than 30 tokens/s.

I have checked if a model for vLLM or SGLang with support for 4 RTX Pro 6000 Blackwell can just be generated with AutoRound. The not so important experts layers would have to be in W3A16. I think neither vLLM or SGLang have kernels to run so a quant with W3A16 and I don't know how to write CUDA kernels.

Therefore, I now run GLM-5.2 IQ4_XS from unsloth with the latest llama.cpp and 4 RTX 6000 Pro Blackwell, it works good with these settings:

./llama-server --temp 1.0 --top-p 0.95 -t 8 -m ... -fa on --threads-http 4 --fit-target 250 --kv-unified 1 -np 8 -ctk q8_0 -ctv q5_1 -ngl all -c 380000

I get around 38 t/s at start and 35 t/s at 10k and 29 t/s at 40k context.

brandonmusic

Jun 28

B12X (that published of this model does) will allow you to run tp 3 and tp6 if that's of interest for you. If you use it, perhaps experiment with tp5 though no one has tried that yet. but b12x has the ability to do tp when it's not powers of 2.

jpsequeira

Jun 28

@brandonmusic So it auto pads? no work on the model is needed?

brandonmusic

Jun 28

@brandonmusic So it auto pads? no work on the model is needed?

@jpsequeira You can see the image, and launch script and flags here https://github.com/local-inference-lab/rtx6kpro/blob/master/models/glm5.2_v13.md for what's aviable on glm 5.2. It's not model work thats needed for 3 and 6. I'm not honestly sure on tp 5. Likely to be kernel work to get that to work at the very least.

brandonmusic

Jun 28

You would defintely be doing original work tryign to get tp 5 to work. But if you want 3 or 6, b12x allows for it. Local infernce lab (and teh related discord server) have many people running tp 6 for glm 5.2. Seem to be satisfied (all thanks to the GOAT that created the quant we are posting on right now aka lukealonzo). You could perhaps throw codex at what b12x did for tp 3 and tp6, to see if ther is a possibity to have it do the same for a tp5. but you want want to check the various REAPs that have been created, that can fit on tp4, they seem to be pretty solid in my testing, and that of the discord server. (oxzero has made some good ones, full disclosure i did one of those too, and then there are a few others as well. This model is pretty robust to REAP)

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