Instructions to use dreamgen/opus-v1.2-llama-3-8b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use dreamgen/opus-v1.2-llama-3-8b with Transformers:

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

pipe = pipeline("text-generation", model="dreamgen/opus-v1.2-llama-3-8b")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("dreamgen/opus-v1.2-llama-3-8b")
model = AutoModelForCausalLM.from_pretrained("dreamgen/opus-v1.2-llama-3-8b", 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=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use dreamgen/opus-v1.2-llama-3-8b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "dreamgen/opus-v1.2-llama-3-8b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dreamgen/opus-v1.2-llama-3-8b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/dreamgen/opus-v1.2-llama-3-8b

SGLang

How to use dreamgen/opus-v1.2-llama-3-8b 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 "dreamgen/opus-v1.2-llama-3-8b" \
    --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": "dreamgen/opus-v1.2-llama-3-8b",
		"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 "dreamgen/opus-v1.2-llama-3-8b" \
        --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": "dreamgen/opus-v1.2-llama-3-8b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Unsloth Desktop
Docker Model Runner
How to use dreamgen/opus-v1.2-llama-3-8b with Docker Model Runner:
```
docker model run hf.co/dreamgen/opus-v1.2-llama-3-8b
```

Curious about Fine-tuning Methods

by ElliottDyson - opened Apr 19, 2024

Discussion

ElliottDyson

Apr 19, 2024

Hello there, just a simple curiosity here, what were the factors that made it so that you chose (or ended up with) training for 2 epochs over your data?

DreamGenX

DreamGen org Apr 19, 2024

Regarding the number of epoch specifically, it's mostly based on my other trains of other models using this same dataset. The eval loss also flatlined (though might go lower for at least 1 more epoch) and the base model is quite amazing and I did not want to overwrite all of its other capabilities so to speak (even though my dataset is diverse, and has data beyond just writing).

So TL;DR: mostly vibes -- I can't afford to test things properly atm (do a sweep of different params, and compare based on end-to-end side-by-side comparison).

ElliottDyson

Apr 19, 2024

Regarding the number of epoch specifically, it's mostly based on my other trains of other models using this same dataset. The eval loss also flatlined (though might go lower for at least 1 more epoch) and the base model is quite amazing and I did not want to overwrite all of its other capabilities so to speak (even though my dataset is diverse, and has data beyond just writing).

So TL;DR: mostly vibes -- I can't afford to test things properly atm (do a sweep of different params, and compare based on end-to-end side-by-side comparison).

Definitely agree on the quality of the base model, the thought had occured that too much fine-tuning might lose some of its best aspects such as instruction following. Looking forward to seeing potential improvements in the future but totally understand the cost of training. Thanks for getting it out so quickly!

Osd111

Jun 6, 2024

Why does train/loss stop decreasing after reaching a certain stage?

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