Instructions to use trashpanda-org/QwQ-32B-Snowdrop-v0 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use trashpanda-org/QwQ-32B-Snowdrop-v0 with Transformers:

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

pipe = pipeline("text-generation", model="trashpanda-org/QwQ-32B-Snowdrop-v0")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("trashpanda-org/QwQ-32B-Snowdrop-v0")
model = AutoModelForCausalLM.from_pretrained("trashpanda-org/QwQ-32B-Snowdrop-v0")
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 trashpanda-org/QwQ-32B-Snowdrop-v0 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "trashpanda-org/QwQ-32B-Snowdrop-v0"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "trashpanda-org/QwQ-32B-Snowdrop-v0",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/trashpanda-org/QwQ-32B-Snowdrop-v0

SGLang

How to use trashpanda-org/QwQ-32B-Snowdrop-v0 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 "trashpanda-org/QwQ-32B-Snowdrop-v0" \
    --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": "trashpanda-org/QwQ-32B-Snowdrop-v0",
		"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 "trashpanda-org/QwQ-32B-Snowdrop-v0" \
        --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": "trashpanda-org/QwQ-32B-Snowdrop-v0",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use trashpanda-org/QwQ-32B-Snowdrop-v0 with Docker Model Runner:
```
docker model run hf.co/trashpanda-org/QwQ-32B-Snowdrop-v0
```

I guess "Wrap Sequences with Newline" just doesn't work in this finetune?

by Slaaaaaau - opened Mar 9, 2025

Discussion

Slaaaaaau

Mar 9, 2025

Unlike the original model, new lines (/n) are not auto inserted here, and therefore it is necessary to make edits to the chatml format in the form of manually placed line breaks, and in the reasonong block, just remove these breaks.

in the master profile attached to the main page everything is already done, it just confused me when I started comparing it with the original model.

Hasnonname

trashpanda org Mar 9, 2025

To be fair, this isn’t really a finetune, just a really cursed merge I decided to do on a whim and never expected to work half as well as it did. But thanks for reporting this—to clarify, is this an ST issue (ie., an issue with the preset linked on the model card) or a formatting issue by the model itself?

Slaaaaaau

Mar 10, 2025

•

edited Mar 10, 2025

isn’t really a finetune

yeah, sorry - a merge

formatting issue by the model itself

yes, I believe that it is, (coz original qwq32b doesnt do this) not that it breaks anything VERY much, I just had to shovel the prompts to make it work properly. I have more questions about the samplers, for now I use the ones on the snowdrop main page and this of course gives a variety of answers, but their... ahem, coherence/logic - is a bit lame, I have to swipe a lot to avoid so much shizo. I didn't feel this in the original model, although I didn't play with these models much, and the samplers there are different, recommended ones. I don't rule out that I have an overcomplicated prompt thanks to the group chat and the huge lorebook hehe

In any case, this is an interesting merge and with an abundance of free time it would be worth playing around with for longer.

Hasnonname

trashpanda org Mar 10, 2025

I’ve definitely noticed that it can have weird hallucinations, I think the strongest part of this model is how it writes. But it does get confused a lot when there’s a lot going on in the scene. I usually prefer using this one with smaller one-on-one scenarios.

This is also the first time I’ve heard this newline thing happening with any model. Just your usual trashpanda cursedness, I guess.

Regardless though, thanks for trying it out.

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