Instructions to use QuantTrio/GLM-5.2-Int4-Int8Mix with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use QuantTrio/GLM-5.2-Int4-Int8Mix with Transformers:

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

pipe = pipeline("text-generation", model="QuantTrio/GLM-5.2-Int4-Int8Mix")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("QuantTrio/GLM-5.2-Int4-Int8Mix")
model = AutoModelForCausalLM.from_pretrained("QuantTrio/GLM-5.2-Int4-Int8Mix")
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 QuantTrio/GLM-5.2-Int4-Int8Mix with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "QuantTrio/GLM-5.2-Int4-Int8Mix"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "QuantTrio/GLM-5.2-Int4-Int8Mix",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/QuantTrio/GLM-5.2-Int4-Int8Mix

SGLang

How to use QuantTrio/GLM-5.2-Int4-Int8Mix 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 "QuantTrio/GLM-5.2-Int4-Int8Mix" \
    --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": "QuantTrio/GLM-5.2-Int4-Int8Mix",
		"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 "QuantTrio/GLM-5.2-Int4-Int8Mix" \
        --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": "QuantTrio/GLM-5.2-Int4-Int8Mix",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use QuantTrio/GLM-5.2-Int4-Int8Mix with Docker Model Runner:
```
docker model run hf.co/QuantTrio/GLM-5.2-Int4-Int8Mix
```

Any chances for A100?

by traphix - opened about 17 hours ago

Discussion

traphix

about 17 hours ago

Can A100 run this model?

JunHowie

QuantTrio org about 17 hours ago

We haven't tested it on A100 yet.

traphix

about 13 hours ago

We haven't tested it on A100 yet.

Looking forward to your A100 test results

S1quence

about 12 hours ago

•

edited about 11 hours ago

Can A100 run this model?

YES, 8xA100 CAN run this model with limited context length!

you need to do:

apply this PR, SM80 gpus do not support Sparse MLA. There is a Triton Sparse MLA for sm80. https://github.com/vllm-project/vllm/pull/38476
download the INT4 model.

here is my command for 8*A100 gpus:

        --served-model-name GLM-5.2
        --dtype bfloat16
        --quantization compressed-tensors
        # --kv-cache-dtype fp8 # Triton Sparse MLA do not support fp8
        --tensor-parallel-size 8
        --enable-expert-parallel
        --max-model-len auto
        --gpu-memory-utilization 0.96
        --max-num-seqs 4
        --tool-call-parser glm47
        --reasoning-parser glm45
        # --speculative-config '{"method":"mtp","num_speculative_tokens":1}' # do not enable Speculative Decoding, it will reduce the context length to 100k.
        --disable-uvicorn-access-log
        --safetensors-load-strategy prefetch

my actual settings & results:

vllm: 0.23.1rc1.dev255+g435f82d61 (nightly version)
decoding speed: Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 36.6 tokens/s. this speed is okey. but not as fast as Minimax-M3 and Nemotron3 Ultra.
actual context length: [kv_cache_utils.py:1943] Auto-fit max_model_len: reduced from 1048576 to 236736 to fit in available GPU memory (20.43 GiB available for KV cache)
attention backend: (Worker_TP0_EP0 pid=3508478) INFO 06-23 17:11:11 [cuda.py:458] Using TRITON_MLA_SPARSE attention backend out of potential backends: ['TRITON_MLA_SPARSE'].
actual experience: worked on Roo code for a night, all thing is okey.

My full command:

argv: vllm serve model_hub/GLM-5.2-Int4-Int8Mix --trust-remote-code --port 1080 --root-path /aiforward1039931975797833728 --enable-log-requests --enable-auto-tool-choice --served-model-name GLM-5.2 --dtype bfloat16 --quantization compressed-tensors --tensor-parallel-size 8 --enable-expert-parallel --max-model-len auto --gpu-memory-utilization 0.96 --max-num-seqs 4 --tool-call-parser glm47 --reasoning-parser glm45 --disable-uvicorn-access-log --safetensors-load-strategy prefetch

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