LLM Pruning and Distillation in Practice: The Minitron Approach

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.

Compact Language Models via Pruning and Knowledge Distillation

Large language models (LLMs) targeting different deployment scales and sizes are currently produced by training each variant from scratch; this is extremely compute-intensive. In this paper, we investigate if pruning an existing LLM and then re-training it with a fraction (<3%) of the original training data can be a suitable alternative to repeated, full retraining. To this end, we develop a set of practical and effective compression best practices for LLMs that combine depth, width, attention and MLP pruning with knowledge distillation-based retraining; we arrive at these best practices through a detailed empirical exploration of pruning strategies for each axis, methods to combine axes, distillation strategies, and search techniques for arriving at optimal compressed architectures. We use this guide to compress the Nemotron-4 family of LLMs by a factor of 2-4x, and compare their performance to similarly-sized models on a variety of language modeling tasks. Deriving 8B and 4B models from an already pretrained 15B model using our approach requires up to 40x fewer training tokens per model compared to training from scratch; this results in compute cost savings of 1.8x for training the full model family (15B, 8B, and 4B). Minitron models exhibit up to a 16% improvement in MMLU scores compared to training from scratch, perform comparably to other community models such as Mistral 7B, Gemma 7B and Llama-3 8B, and outperform state-of-the-art compression techniques from the literature. We have open-sourced Minitron model weights on Huggingface, with corresponding supplementary material including example code available on GitHub.

GANDALF: Generative Adversarial Networks with Discriminator-Adaptive Loss Fine-tuning for Alzheimer's Disease Diagnosis from MRI

Positron Emission Tomography (PET) is now regarded as the gold standard for the diagnosis of Alzheimer's Disease (AD). However, PET imaging can be prohibitive in terms of cost and planning, and is also among the imaging techniques with the highest dosage of radiation. Magnetic Resonance Imaging (MRI), in contrast, is more widely available and provides more flexibility when setting the desired image resolution. Unfortunately, the diagnosis of AD using MRI is difficult due to the very subtle physiological differences between healthy and AD subjects visible on MRI. As a result, many attempts have been made to synthesize PET images from MR images using generative adversarial networks (GANs) in the interest of enabling the diagnosis of AD from MR. Existing work on PET synthesis from MRI has largely focused on Conditional GANs, where MR images are used to generate PET images and subsequently used for AD diagnosis. There is no end-to-end training goal. This paper proposes an alternative approach to the aforementioned, where AD diagnosis is incorporated in the GAN training objective to achieve the best AD classification performance. Different GAN lossesare fine-tuned based on the discriminator performance, and the overall training is stabilized. The proposed network architecture and training regime show state-of-the-art performance for three- and four- class AD classification tasks.

GANBERT: Generative Adversarial Networks with Bidirectional Encoder Representations from Transformers for MRI to PET synthesis

Synthesizing medical images, such as PET, is a challenging task due to the fact that the intensity range is much wider and denser than those in photographs and digital renderings and are often heavily biased toward zero. Above all, intensity values in PET have absolute significance, and are used to compute parameters that are reproducible across the population. Yet, usually much manual adjustment has to be made in pre-/post- processing when synthesizing PET images, because its intensity ranges can vary a lot, e.g., between -100 to 1000 in floating point values. To overcome these challenges, we adopt the Bidirectional Encoder Representations from Transformers (BERT) algorithm that has had great success in natural language processing (NLP), where wide-range floating point intensity values are represented as integers ranging between 0 to 10000 that resemble a dictionary of natural language vocabularies. BERT is then trained to predict a proportion of masked values images, where its "next sentence prediction (NSP)" acts as GAN discriminator. Our proposed approach, is able to generate PET images from MRI images in wide intensity range, with no manual adjustments in pre-/post- processing. It is a method that can scale and ready to deploy.