Venmugil Elango

Venmugil Elango, Nidhi Bhatia, Roger Waleffe et al.

Mixture of Experts (MoEs) have become a central component of many state-of-the-art open-source and proprietary large language models. Despite their widespread adoption, it remains unclear how close existing MoE architectures are to optimal with respect to inference cost, as measured by accuracy per floating-point operation and per parameter. In this work, we revisit MoE design from a hardware-software co-design perspective, grounded in empirical and theoretical considerations. We characterize key performance bottlenecks across diverse deployment regimes, spanning offline high-throughput execution and online, latency-critical inference. Guided by these insights, we introduce LatentMoE, a new model architecture resulting from systematic design exploration and optimized for maximal accuracy per unit of compute. Empirical design space exploration at scales of up to 95B parameters and over a 1T-token training horizon, together with supporting theoretical analysis, shows that LatentMoE consistently outperforms standard MoE architectures in terms of accuracy per FLOP and per parameter. Given its strong performance, the LatentMoE architecture has been adopted by the flagship Nemotron-3 Super and Ultra models and scaled to substantially larger regimes, including longer token horizons and larger model sizes, as reported in Nvidia et al. (arXiv:2512.20856).

Bita Rouhani, Ritchie Zhao, Venmugil Elango et al.

This paper introduces Block Data Representations (BDR), a framework for exploring and evaluating a wide spectrum of narrow-precision formats for deep learning. It enables comparison of popular quantization standards, and through BDR, new formats based on shared microexponents (MX) are identified, which outperform other state-of-the-art quantization approaches, including narrow-precision floating-point and block floating-point. MX utilizes multiple levels of quantization scaling with ultra-fine scaling factors based on shared microexponents in the hardware. The effectiveness of MX is demonstrated on real-world models including large-scale generative pretraining and inferencing, and production-scale recommendation systems.

Bita Darvish Rouhani, Ritchie Zhao, Ankit More et al.

Narrow bit-width data formats are key to reducing the computational and storage costs of modern deep learning applications. This paper evaluates Microscaling (MX) data formats that combine a per-block scaling factor with narrow floating-point and integer types for individual elements. MX formats balance the competing needs of hardware efficiency, model accuracy, and user friction. Empirical results on over two dozen benchmarks demonstrate practicality of MX data formats as a drop-in replacement for baseline FP32 for AI inference and training with low user friction. We also show the first instance of training generative language models at sub-8-bit weights, activations, and gradients with minimal accuracy loss and no modifications to the training recipe.

Venmugil Elango

Training a deep neural network (DNN) requires substantial computational and memory requirements. It is common to use multiple devices to train a DNN to reduce the overall training time. There are several choices to parallelize each layer in a DNN. Exhaustively searching this list to find an optimal parallelization strategy is prohibitively time consuming and impractical. The standard practice is to use data parallelism because of its simplicity. However, data parallelism is often sub-optimal, and suffers from poor performance and high memory requirement. Expert-designed strategies have been proposed on a case-by-case basis using domain specific knowledge. These expert-designed strategies do not generalize well to DNNs other than the ones for which they were designed, and are not always necessarily the best choice. In this paper, we propose an approach to automatically find efficient parallelization strategies for DNNs from their computation graphs. We present an efficient algorithm to compute these strategies within a reasonable time in practice. We evaluate the effectiveness of our approach on various DNNs. We also compare the performance of the strategies identified by our approach against data parallelism, expert-designed strategies, and the state-of-the-art approaches. Our results show that the strategies found using our approach outperform the baseline data parallelism strategy in all the cases. In addition, our strategies achieve better performance than the expert-designed strategies and the state-of-the-art approaches.

Hayate Iso, Tiyasa Mitra, Sudipta Mondal et al.

RL post-training of frontier language models is increasingly bottlenecked by autoregressive rollout generation, making rollout acceleration a central systems challenge. Many existing efficiency methods improve throughput by changing the rollout or optimization regime, for example, through off-policy execution, replay, or lower-precision generation. We study speculative decoding as a lossless acceleration primitive for RL rollouts that preserves the target model's output distribution. We implement speculative decoding in NeMo-RL with a vLLM backend, supporting both synchronous and asynchronous pipelines and enabling speculation during RL rollouts. This benefit is realizable across speculation mechanisms, such as pretrained MTP heads, small external draft models or even techniques such as Eagle3, which are traditionally applied after RL phase. This yields a deployment path for state-of-the-art speculative decoding inside RL training. In a reasoning post-training workload at 8B scale under synchronous RL, speculative decoding improves rollout throughput by 1.8x. Using a high-fidelity performance simulator, we project that combining speculative decoding with asynchronous RL yields up to 2.5x end-to-end training speedup at 235B scale.

Venmugil Elango

Transformer-based models have emerged as a leading architecture for natural language processing, natural language generation, and image generation tasks. A fundamental element of the transformer architecture is self-attention, which allows the model to capture intricate dependencies within the data. However, the self-attention mechanism also incurs significant computational and memory costs, particularly for long sequences. In this paper, we introduce ATTENTION2D, a novel approach that exploits parallelism along two dimensions - query and key/value - of the self-attention operation. This method enables efficient distribution and parallelization of computations across multiple devices. Our approach facilitates asymptotically faster training and inference phases compared to previous methods, without relying on approximations or incurring additional computational or memory overheads. Furthermore, unlike existing techniques that struggle to scale with an increasing number of processing units, our approach effectively scales with additional processing units. Our experimental results confirm the effectiveness of our method in improving communication efficiency and scalability. Compared to Ring Attention, our approach demonstrated up to a 5x performance boost on a GPT-3-like model using 64 NVIDIA A100 GPUs across 16 nodes, and up to a 9.4x performance boost on 64 NVIDIA H100 GPUs across 64 nodes.