Instructions to use Qwen/Qwen2.5-0.5B-Instruct with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Qwen/Qwen2.5-0.5B-Instruct with Transformers:

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

pipe = pipeline("text-generation", model="Qwen/Qwen2.5-0.5B-Instruct")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct", 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]:]))

Inference
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Qwen/Qwen2.5-0.5B-Instruct with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Qwen/Qwen2.5-0.5B-Instruct"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Qwen/Qwen2.5-0.5B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Qwen/Qwen2.5-0.5B-Instruct

SGLang

How to use Qwen/Qwen2.5-0.5B-Instruct 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 "Qwen/Qwen2.5-0.5B-Instruct" \
    --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": "Qwen/Qwen2.5-0.5B-Instruct",
		"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 "Qwen/Qwen2.5-0.5B-Instruct" \
        --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": "Qwen/Qwen2.5-0.5B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Qwen/Qwen2.5-0.5B-Instruct with Docker Model Runner:
```
docker model run hf.co/Qwen/Qwen2.5-0.5B-Instruct
```

Mechanistic interpretability study: executable replacement of all Qwen2.5-0.5B attention heads and MLPs

#22

by maxwelhelp - opened 23 days ago

Discussion

maxwelhelp

23 days ago

Hello Qwen team and community,

I am Maxim Zhivotok, an independent researcher from Ukraine. I have released a fully reproducible mechanistic interpretability study based directly on Qwen2.5-0.5B-Instruct:

“Executable Matrix Programs: Faithful Weight-Derived Replacement and Factorized Intervention in Pretrained Transformers.”

The method derives executable component-level matrix programs directly from the pretrained weights and reinserts them into the model’s native forward pass, without retraining the model or training a separate replacement network.

Main replacement results:

all 336 attention heads replaced inside the native forward pass;
all 24 SwiGLU MLP sublayers replaced;
0.258% median full-logit replacement discrepancy;
0.566% p95 discrepancy;
0.923% maximum discrepancy;
100% top-1 preservation across 64 prompts.

The work also provides:

factorized QK routing and VO payload representations;
native SwiGLU gate/read/write decomposition;
factorized causal interventions;
machine-readable attention, MLP, intervention, and downstream propagation results.

Preprint:
https://doi.org/10.5281/zenodo.21312311

Code and complete results:
https://github.com/maxwelhelp/matrix-programs

ORCID:
https://orcid.org/0009-0005-7722-6811

I would appreciate technical feedback from the Qwen team and community, particularly on:

validation on larger Qwen2.5 or newer Qwen architectures;
additional Qwen-specific benchmarks or prompt sets;
the interpretation of the QK routing and VO payload factors;
possible integration with existing Qwen analysis tooling.

Separately, this is my first arXiv cs.LG submission, and I am currently blocked by the endorsement requirement. If an eligible cs.LG endorser considers the public manuscript appropriate for the category after a brief look, I would be grateful if they could contact me privately.

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