Instructions to use cmarkea/Qwen2.5-32B-Instruct-4bit with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use cmarkea/Qwen2.5-32B-Instruct-4bit with Transformers:

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

pipe = pipeline("text-generation", model="cmarkea/Qwen2.5-32B-Instruct-4bit")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("cmarkea/Qwen2.5-32B-Instruct-4bit")
model = AutoModelForCausalLM.from_pretrained("cmarkea/Qwen2.5-32B-Instruct-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

vLLM

How to use cmarkea/Qwen2.5-32B-Instruct-4bit with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "cmarkea/Qwen2.5-32B-Instruct-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": "cmarkea/Qwen2.5-32B-Instruct-4bit",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/cmarkea/Qwen2.5-32B-Instruct-4bit

SGLang

How to use cmarkea/Qwen2.5-32B-Instruct-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 "cmarkea/Qwen2.5-32B-Instruct-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": "cmarkea/Qwen2.5-32B-Instruct-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 "cmarkea/Qwen2.5-32B-Instruct-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": "cmarkea/Qwen2.5-32B-Instruct-4bit",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use cmarkea/Qwen2.5-32B-Instruct-4bit with Docker Model Runner:
```
docker model run hf.co/cmarkea/Qwen2.5-32B-Instruct-4bit
```

Converted version of Qwen2.5-32B-Instruct to 4-bit using bitsandbytes. For more information about the model, refer to the model's page.

Impact on performance

Impact of quantization on a set of models.

Evaluation of the model was conducted using the PoLL (Pool of LLM) technique, assessing performance on 100 French questions with scores aggregated from six evaluations (two per evaluator). The evaluators included GPT-4o, Gemini-1.5-pro, and Claude3.5-sonnet.

Performance Scores (on a scale of 5):

Model	Score	# params (Billion)	size (GB)
gpt-4o	4.13	N/A	N/A
gpt-4o-mini	4.02	N/A	N/A
Qwen/Qwen2.5-32B-Instruct	3.99	32.8	65.6
cmarkea/Qwen2.5-32B-Instruct-4bit	3.98	32.8	16.4
mistralai/Mixtral-8x7B-Instruct-v0.1	3.71	46.7	93.4
cmarkea/Mixtral-8x7B-Instruct-v0.1-4bit	3.68	46.7	23.35
meta-llama/Meta-Llama-3.1-70B-Instruct	3.68	70.06	140.12
gpt-3.5-turbo	3.66	175	350
cmarkea/Meta-Llama-3.1-70B-Instruct-4bit	3.64	70.06	35.3
TheBloke/Mixtral-8x7B-Instruct-v0.1-GPTQ	3.56	46.7	46.7
meta-llama/Meta-Llama-3.1-8B-Instruct	3.25	8.03	16.06
mistralai/Mistral-7B-Instruct-v0.2	1.98	7.25	14.5
cmarkea/bloomz-7b1-mt-sft-chat	1.69	7.07	14.14
cmarkea/bloomz-3b-dpo-chat	1.68	3	6
cmarkea/bloomz-3b-sft-chat	1.51	3	6
croissantllm/CroissantLLMChat-v0.1	1.19	1.3	2.7
cmarkea/bloomz-560m-sft-chat	1.04	0.56	1.12
OpenLLM-France/Claire-Mistral-7B-0.1	0.38	7.25	14.5

The impact of quantization is negligible.

Prompt Pattern

Here is a reminder of the command pattern to interact with the model:

<|im_start|>user\n{user_prompt_1}<|im_end|>\n<|im_start|>assistant\n{model_answer_1}...

Downloads last month: 34

Safetensors

Model size

34B params

Tensor type

F32

BF16

Model tree for cmarkea/Qwen2.5-32B-Instruct-4bit

Base model

Qwen/Qwen2.5-32B

Finetuned

Qwen/Qwen2.5-32B-Instruct

Quantized

(139)

this model

Collection including cmarkea/Qwen2.5-32B-Instruct-4bit

Quantized 4-bit models

Collection

Large model quantized with post-quantization performance very close to the original models, allowing it to run on reasonable infrastructure. • 9 items • Updated Mar 2 • 1