Instructions to use chargoddard/llama3-42b-v0 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use chargoddard/llama3-42b-v0 with Transformers:

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

pipe = pipeline("text-generation", model="chargoddard/llama3-42b-v0")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("chargoddard/llama3-42b-v0")
model = AutoModelForCausalLM.from_pretrained("chargoddard/llama3-42b-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]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use chargoddard/llama3-42b-v0 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "chargoddard/llama3-42b-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": "chargoddard/llama3-42b-v0",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/chargoddard/llama3-42b-v0

SGLang

How to use chargoddard/llama3-42b-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 "chargoddard/llama3-42b-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": "chargoddard/llama3-42b-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 "chargoddard/llama3-42b-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": "chargoddard/llama3-42b-v0",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use chargoddard/llama3-42b-v0 with Docker Model Runner:
```
docker model run hf.co/chargoddard/llama3-42b-v0
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

🚨 THIS IS A BASE MODEL 🚨

This model is pruned from the base Llama 3 70B, which has no instruction tuning and randomly initialized special tokens.

Using this with the Llama 3 instruction format is injecting random noise into latent space and will give you deranged results. (It's pretty funny actually.) Treat this as the untrained foundation model this is and use appropriate prompts.

Meta's Llama 3 70B pruned to 42B parameters using the methodology described in The Unreasonable Ineffectiveness of the Deeper Layers. Post-pruning trained using QLoRA for ~100M tokens from JeanKaddour/minipile.

Layers to prune selected using PruneMe.

Still evaluating, don't get too excited! Might be incredibly dumb. Check out these numbers though:

Groups	Version	Filter	n-shot	Metric	Value		Stderr
mmlu	N/A	none	0	acc	0.7669	±	0.0034
- humanities	N/A	none	5	acc	0.7296	±	0.0062
- other	N/A	none	5	acc	0.8101	±	0.0067
- social_sciences	N/A	none	5	acc	0.8668	±	0.0060
- stem	N/A	none	5	acc	0.6825	±	0.0079
winogrande	1	none	5	acc	0.8027	±	0.0112
hellaswag	1	none	10	acc_norm	0.8025	±	0.0040

Downloads last month: 16

Safetensors

Model size

43B params

Tensor type

BF16

Model tree for chargoddard/llama3-42b-v0

Quantizations

2 models

Dataset used to train chargoddard/llama3-42b-v0

Paper for chargoddard/llama3-42b-v0

The Unreasonable Ineffectiveness of the Deeper Layers

Paper • 2403.17887 • Published Mar 26, 2024 • 83