Instructions to use Naphula/Goetia-8B-v1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Naphula/Goetia-8B-v1 with Transformers:

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

pipe = pipeline("text-generation", model="Naphula/Goetia-8B-v1")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Naphula/Goetia-8B-v1")
model = AutoModelForCausalLM.from_pretrained("Naphula/Goetia-8B-v1", 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]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Naphula/Goetia-8B-v1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Naphula/Goetia-8B-v1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Naphula/Goetia-8B-v1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Naphula/Goetia-8B-v1

SGLang

How to use Naphula/Goetia-8B-v1 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 "Naphula/Goetia-8B-v1" \
    --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": "Naphula/Goetia-8B-v1",
		"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 "Naphula/Goetia-8B-v1" \
        --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": "Naphula/Goetia-8B-v1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Naphula/Goetia-8B-v1 with Docker Model Runner:
```
docker model run hf.co/Naphula/Goetia-8B-v1
```

Endless loop

by redaihf - opened Apr 22

Discussion

redaihf

Apr 22

•

edited Apr 22

This model goes into a loop after about 1000 tokens that continues until the context exhausts. Looping happens for both safe and unsafe content with a temperature of 0.75 and a repeat penalty of 1.08. The repeated content is usually a few short paragraphs of text.

Naphula

Owner Apr 22

This could be something caused by dare_linear or maybe some components are bugged I'll have to look into it further.

Naphula

Owner May 9

•

edited May 9

Did you try with lower temps and rep pen 1.12?

I suspect this might be caused by mixing llama with nemotron models. In my recent model_stock tests, it reported too much cosine dissimilarity, much like with the 2501 and 2509 Mistral 24B models.

I'm working on a Nemotron-only merge now with (mostly) my finetunes, and noticed some early terminations there as well when using the della method. It might have something to do with Assistant Pepe using "content": "<|end_of_text|>",instead of "content": "<|eot_id|>", in the special_tokens_map.json.

I'm comparing a version with and without Pepe to inspect this further. Also comparing karcher to della, SCE and model_stock

Update

Surprisingly, the version without Pepe had more early terminations than the one with Pepe. So it must be something else

Naphula

Owner May 9

I think the problem might be related to tokenizer_source: union because when i switched to morpheus tokenizer and swapped pepe's special token map json the early terminations went away (at least from initial tests)

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