Instructions to use maldv/Qwenstein2.5-32B-Instruct with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use maldv/Qwenstein2.5-32B-Instruct with Transformers:

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

pipe = pipeline("text-generation", model="maldv/Qwenstein2.5-32B-Instruct")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("maldv/Qwenstein2.5-32B-Instruct")
model = AutoModelForCausalLM.from_pretrained("maldv/Qwenstein2.5-32B-Instruct")
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

vLLM

How to use maldv/Qwenstein2.5-32B-Instruct with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "maldv/Qwenstein2.5-32B-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": "maldv/Qwenstein2.5-32B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/maldv/Qwenstein2.5-32B-Instruct

SGLang

How to use maldv/Qwenstein2.5-32B-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 "maldv/Qwenstein2.5-32B-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": "maldv/Qwenstein2.5-32B-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 "maldv/Qwenstein2.5-32B-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": "maldv/Qwenstein2.5-32B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use maldv/Qwenstein2.5-32B-Instruct with Docker Model Runner:
```
docker model run hf.co/maldv/Qwenstein2.5-32B-Instruct
```

this model have their original restriction staying on their weight

by AnA202 - opened Jan 15, 2025

Discussion

AnA202

Jan 15, 2025

looks like this merger capture the weight for the restriction more, another merger with model that have negative attitude or even albitred will help this model by a lot

maldv

Owner Jan 15, 2025

Yeah, it's way more restricted. I wanted to focus on getting chain of thought solid before I run an abliteration pass.

Fun fact: If you don't want it to be restricted... Change the one word that you want it to accept to Chinese. I've gotten it to accept some very crazy requests that way.

SubtleOne

Jan 16, 2025

Kobold has a cute but remarkably effective jailbreak. It precedes each reply by the model with: "Sure I can help with that:" and the model then continues this reply assuming, I guess, that it had already accepted to do this.

maldv

Owner Jan 16, 2025

😍 That's pretty cool! That's something you can't get from a chat completion API. Once you inject guidance in to most models it will overwhelm the alignment. Most restrictive alignment is very shallow - which is why the original 'abliteration is one direction' works so well.

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