Llama-2-7b-pruned70-retrained

This repo contains model files for a Llama 2 7B model that has had 50% of the parameters pruned in one-shot with SparseGPT, then retrained by Cerebras with 50B tokens from SlimPajama while maintaining sparsity. It was then one-shot pruned to 70% sparsity and trained for another 100B tokens.

Official model weights from Enabling High-Sparsity Foundational Llama Models with Efficient Pretraining and Deployment.

Authors: Neural Magic, Cerebras

Usage

Below we share some code snippets on how to get quickly started with running the model.

Sparse Transfer

By leveraging a pre-sparsified model's structure, you can efficiently fine-tune on new data, leading to reduced hyperparameter tuning, training times, and computational costs. Learn about this process here.

Running the model

This model has not been fine-tuned for instruction-following but may be run with the transformers library. For accelerated inference with sparsity, deploy with nm-vllm or deepsparse.

# pip install transformers accelerate
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("neuralmagic/Llama-2-7b-pruned70-retrained")
model = AutoModelForCausalLM.from_pretrained("neuralmagic/Llama-2-7b-pruned70-retrained", device_map="auto")

input_text = "Write me a poem about Machine Learning."
input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")

outputs = model.generate(**input_ids)
print(tokenizer.decode(outputs[0]))

Evaluation Benchmark Results

Model evaluation metrics and results. [UPDATE]

Benchmark	Metric	Llama-2-7b	Llama-2-7b-pruned70-retrained
MMLU	5-shot	46.9%	36.5%
HellaSwag	0-shot	78.6%	74.1%
WinoGrande	5-shot	74.0%	69.5%
ARC-c	25-shot	53.1%	45.4%
TruthfulQA	5-shot	38.8%	36.7%
GSM8K	5-shot	14.5%	8.0%
HumanEval	pass@1	13.4%	14.4%

Model Training Details

[UPDATE]

Help

For further support, and discussions on these models and AI in general, join Neural Magic's Slack Community