LLM Pruning and Distillation in Practice: The Minitron Approach

TL;DR

This paper demonstrates effective pruning and distillation techniques to compress large language models like Llama 3.1 and Mistral NeMo, resulting in smaller models with maintained performance, and shares the models openly.

Contribution

It introduces practical pruning and distillation methods for large language models, including two strategies and open-sourcing the compressed models.

Findings

Pruning and distillation produce high-quality smaller models.

Slight fine-tuning of teacher models improves distillation results.

Open-source release of the compressed models facilitates further research.

Abstract

We present a comprehensive report on compressing the Llama 3.1 8B and Mistral NeMo 12B models to 4B and 8B parameters, respectively, using pruning and distillation. We explore two distinct pruning strategies: (1) depth pruning and (2) joint hidden/attention/MLP (width) pruning, and evaluate the results on common benchmarks from the LM Evaluation Harness. The models are then aligned with NeMo Aligner and tested in instruct-tuned versions. This approach produces a compelling 4B model from Llama 3.1 8B and a state-of-the-art Mistral-NeMo-Minitron-8B (MN-Minitron-8B for brevity) model from Mistral NeMo 12B. We found that with no access to the original data, it is beneficial to slightly fine-tune teacher models on the distillation dataset. We open-source our base model weights on Hugging Face with a permissive license.