Sergio P. Perez

Sergio P. Perez, Yan Zhang, James Briggs et al.

FP8 formats are gaining popularity to boost the computational efficiency for training and inference of large deep learning models. Their main challenge is that a careful choice of scaling is needed to prevent degradation due to the reduced dynamic range compared to higher-precision formats. Although there exists ample literature about selecting such scalings for INT formats, this critical aspect has yet to be addressed for FP8. This paper presents a methodology to select the scalings for FP8 linear layers, based on dynamically updating per-tensor scales for the weights, gradients and activations. We apply this methodology to train and validate large language models of the type of GPT and Llama 2 using FP8, for model sizes ranging from 111M to 70B. To facilitate the understanding of the FP8 dynamics, our results are accompanied by plots of the per-tensor scale distribution for weights, activations and gradients during both training and inference.

Remigiusz Kinas, Paweł Kiszczak, Sergio P. Perez et al.

This report details the creation of Bielik-Minitron-7B, a compressed 7.35B parameter version of the Bielik-11B-v3.0 model, specifically optimized for European languages. By leveraging a two-stage compression methodology inspired by the NVIDIA Minitron approach, we combined structured hybrid pruning and knowledge distillation to reduce the model's parameter count by 33.4%, from 11.04B to 7.35B. We utilized the NVIDIA Model Optimizer for structural pruning and the NVIDIA NeMo Framework for logit-based distillation for quality recovery. Following distillation, the model underwent a rigorous alignment pipeline consisting of Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO-P), and Reinforcement Learning (GRPO). Our final model successfully recovered approximately 90% of the baseline model's performance while providing up to 50% inference speedup. This approach demonstrates an efficient pathway to create language models for less-represented languages, preserving the original model quality while reducing inference deployment costs.

Mario Michael Krell, Matej Kosec, Sergio P. Perez et al.

Effective training of today's large language models (LLMs) depends on large batches and long sequences for throughput and accuracy. To handle variable-length sequences on hardware accelerators, it is common practice to introduce padding tokens, so that all sequences in a batch have the same length. We show in this paper that the variation in sequence lengths in common NLP datasets is such that up to 50% of all tokens can be padding. In less common, but not extreme, cases (e.g. GLUE-cola with sequence length 128), the ratio is up to 89%. Existing methods to address the resulting inefficiency are complicated by the need to avoid cross-contamination in self-attention, by a reduction in accuracy when sequence ordering information is lost, or by customized kernel implementations only valid for specific accelerators. This paper introduces a new formalization of sequence packing in the context of the well-studied bin packing problem, and presents new algorithms based on this formulation which, for example, confer a 2x speedup for phase 2 pre-training in BERT. We show how existing models can be adapted to ensure mathematical equivalence between the original and packed models, meaning that packed models can be trained with existing pre-training and fine-tuning practices.

José A. Carrillo, Serafim Kalliadasis, Fuyue Liang et al.

We assess the benefit of including an image inpainting filter before passing damaged images into a classification neural network. For this we employ a modified Cahn-Hilliard equation as an image inpainting filter, which is solved via a finite volume scheme with reduced computational cost and adequate properties for energy stability and boundedness. The benchmark dataset employed here is MNIST, which consists of binary images of handwritten digits and is a standard dataset to validate image-processing methodologies. We train a neural network based of dense layers with the training set of MNIST, and subsequently we contaminate the test set with damage of different types and intensities. We then compare the prediction accuracy of the neural network with and without applying the Cahn-Hilliard filter to the damaged images test. Our results quantify the significant improvement of damaged-image prediction due to applying the Cahn-Hilliard filter, which for specific damages can increase up to 50% and is in general advantageous for low to moderate damage.