Instructions to use chimbiwide/Gemma3NPC-it with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use chimbiwide/Gemma3NPC-it with Transformers:

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

pipe = pipeline("image-text-to-text", model="chimbiwide/Gemma3NPC-it")

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("chimbiwide/Gemma3NPC-it", dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use chimbiwide/Gemma3NPC-it with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "chimbiwide/Gemma3NPC-it"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "chimbiwide/Gemma3NPC-it",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/chimbiwide/Gemma3NPC-it

SGLang

How to use chimbiwide/Gemma3NPC-it 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 "chimbiwide/Gemma3NPC-it" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "chimbiwide/Gemma3NPC-it",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "chimbiwide/Gemma3NPC-it" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "chimbiwide/Gemma3NPC-it",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Unsloth Studio

How to use chimbiwide/Gemma3NPC-it with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for chimbiwide/Gemma3NPC-it to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for chimbiwide/Gemma3NPC-it to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for chimbiwide/Gemma3NPC-it to start chatting

Load model with FastModel

pip install unsloth
from unsloth import FastModel
model, tokenizer = FastModel.from_pretrained(
    model_name="chimbiwide/Gemma3NPC-it",
    max_seq_length=2048,
)

Docker Model Runner
How to use chimbiwide/Gemma3NPC-it with Docker Model Runner:
```
docker model run hf.co/chimbiwide/Gemma3NPC-it
```

Game model

by ncsta - opened Apr 6

Discussion

ncsta

Apr 6

This comment has been hidden (marked as Resolved)

chimbiwide

Owner Apr 6

sure this sounds interesting, can you elaborate a little more on what you are doing?
What is the parameter count you want to target for the 15mb model?
How does the character architecture work? Are the character card/system prompts tbaked into the model or is it injected during inference?
I'm also curious what inference speed you're getting on CPU, and what device did you run it?
I have a private NPC eval dataset that I would be willing to test if your are interested in seeing the model's performance beyond TinyStories.

ncsta

Apr 7

This comment has been hidden (marked as Resolved)

chimbiwide

Owner Apr 7

Thanks for the update, this sounds pretty interesting, especially the real-time CPU inference aspect. I would be happy to test it out once it's further along.
A couple questions:
Is the 13000 characters the full training dataset?
What is the context length you are targeting?
Again, are the character cards baked into the weights or injected during inference?

ncsta

Apr 8

This comment has been hidden (marked as Resolved)

chimbiwide

Owner Apr 8

Thank you for your interest and response! We are interested in seeing what the final ternary Mamba model would look like, we would be interested in testing the final model.
As for PIPPA, we used it to train our Gemma3NPC models as well, and I think the licence for PIPPA (Apache 2.0) allows commercial use.
Is there any other way we could communicate other than here?

ncsta

Apr 10

This comment has been hidden (marked as Resolved)

ncsta changed discussion title from Ternary Mumba to Game model Apr 12

ncsta changed discussion status to closed Apr 12

chimbiwide changed discussion status to open Apr 12

chimbiwide

Owner Apr 12

Sorry can you please leave ur email again, I did not get a chance to save it.

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