Bor-Yiing Su

Felix Abecassis, Anjulie Agrusa, Dong Ahn et al. · nvidia

Large Language Models (LLMs) today are powerful problem solvers across many domains, and they continue to get stronger as they scale in model size, training set size, and training set quality, as shown by extensive research and experimentation across the industry. Training a frontier model today requires on the order of tens to hundreds of yottaflops, which is a massive investment of time, compute, and energy. Improving pretraining efficiency is therefore essential to enable the next generation of even more capable LLMs. While 8-bit floating point (FP8) training is now widely adopted, transitioning to even narrower precision, such as 4-bit floating point (FP4), could unlock additional improvements in computational speed and resource utilization. However, quantization at this level poses challenges to training stability, convergence, and implementation, notably for large-scale models trained on long token horizons. In this study, we introduce a novel approach for stable and accurate training of large language models (LLMs) using the NVFP4 format. Our method integrates Random Hadamard transforms (RHT) to bound block-level outliers, employs a two-dimensional quantization scheme for consistent representations across both the forward and backward passes, utilizes stochastic rounding for unbiased gradient estimation, and incorporates selective high-precision layers. We validate our approach by training a 12-billion-parameter model on 10 trillion tokens -- the longest publicly documented training run in 4-bit precision to date. Our results show that the model trained with our NVFP4-based pretraining technique achieves training loss and downstream task accuracies comparable to an FP8 baseline. These findings highlight that NVFP4, when combined with our training approach, represents a major step forward in narrow-precision LLM training algorithms.

Bor-Yiing Su, Peter Dykas, Mike Chrzanowski et al.

Mixed-precision training is a crucial technique for scaling deep learning models, but successful mixedprecision training requires identifying and applying the right combination of training methods. This paper presents our preliminary study on Mixture-of-Representations (MoR), a novel, per-tensor and sub-tensor level quantization framework that dynamically analyzes a tensor's numerical properties to select between a variety of different representations. Based on the framework, we have proposed and experimented concrete algorithms that choose dynamically between FP8 and BF16 representations for both per-tensor and sub-tensor level granularities. Our universal approach is designed to preserve model quality across various quantization partition strategies and datasets. Our initial findings show that this approach can achieve state-of-the-art results with 98.38% of tensors quantized to the FP8 format. This work highlights the potential of dynamic, property-aware quantization while preserving model quality. We believe this approach can generally improve the robustness of low precision training, as demonstrated by achieving FP8 accuracies that are on par with existing approaches without the need for fine-grain partitioning, or can be used in combination with other training methods to improve the leverage of even lower precision number formats such as NVFP4.

Akhiad Bercovich, Itay Levy, Izik Golan et al. · nvidia

We introduce the Llama-Nemotron series of models, an open family of heterogeneous reasoning models that deliver exceptional reasoning capabilities, inference efficiency, and an open license for enterprise use. The family comes in three sizes -- Nano (8B), Super (49B), and Ultra (253B) -- and performs competitively with state-of-the-art reasoning models such as DeepSeek-R1 while offering superior inference throughput and memory efficiency. In this report, we discuss the training procedure for these models, which entails using neural architecture search from Llama 3 models for accelerated inference, knowledge distillation, and continued pretraining, followed by a reasoning-focused post-training stage consisting of two main parts: supervised fine-tuning and large scale reinforcement learning. Llama-Nemotron models are the first open-source models to support a dynamic reasoning toggle, allowing users to switch between standard chat and reasoning modes during inference. To further support open research and facilitate model development, we provide the following resources: 1. We release the Llama-Nemotron reasoning models -- LN-Nano, LN-Super, and LN-Ultra -- under the commercially permissive NVIDIA Open Model License Agreement. 2. We release the complete post-training dataset: Llama-Nemotron-Post-Training-Dataset. 3. We also release our training codebases: NeMo, NeMo-Aligner, and Megatron-LM.

Kiwan Maeng, Shivam Bharuka, Isabel Gao et al.

The paper proposes and optimizes a partial recovery training system, CPR, for recommendation models. CPR relaxes the consistency requirement by enabling non-failed nodes to proceed without loading checkpoints when a node fails during training, improving failure-related overheads. The paper is the first to the extent of our knowledge to perform a data-driven, in-depth analysis of applying partial recovery to recommendation models and identified a trade-off between accuracy and performance. Motivated by the analysis, we present CPR, a partial recovery training system that can reduce the training time and maintain the desired level of model accuracy by (1) estimating the benefit of partial recovery, (2) selecting an appropriate checkpoint saving interval, and (3) prioritizing to save updates of more frequently accessed parameters. Two variants of CPR, CPR-MFU and CPR-SSU, reduce the checkpoint-related overhead from 8.2-8.5% to 0.53-0.68% compared to full recovery, on a configuration emulating the failure pattern and overhead of a production-scale cluster. While reducing overhead significantly, CPR achieves model quality on par with the more expensive full recovery scheme, training the state-of-the-art recommendation model using Criteo's Ads CTR dataset. Our preliminary results also suggest that CPR can speed up training on a real production-scale cluster, without notably degrading the accuracy.

Qinqing Zheng, Bor-Yiing Su, Jiyan Yang et al.

Recommendation systems are often trained with a tremendous amount of data, and distributed training is the workhorse to shorten the training time. While the training throughput can be increased by simply adding more workers, it is also increasingly challenging to preserve the model quality. In this paper, we present \shadowsync, a distributed framework specifically tailored to modern scale recommendation system training. In contrast to previous works where synchronization happens as part of the training process, \shadowsync separates the synchronization from training and runs it in the background. Such isolation significantly reduces the synchronization overhead and increases the synchronization frequency, so that we are able to obtain both high throughput and excellent model quality when training at scale. The superiority of our procedure is confirmed by experiments on training deep neural networks for click-through-rate prediction tasks. Our framework is capable to express data parallelism and/or model parallelism, generic to host various types of synchronization algorithms, and readily applicable to large scale problems in other areas.