Maxim Naumov

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri et al. · allen-ai, berkeley

Modern artificial intelligence (AI) systems are powered by foundation models. This paper presents a new set of foundation models, called Llama 3. It is a herd of language models that natively support multilinguality, coding, reasoning, and tool usage. Our largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens. This paper presents an extensive empirical evaluation of Llama 3. We find that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks. We publicly release Llama 3, including pre-trained and post-trained versions of the 405B parameter language model and our Llama Guard 3 model for input and output safety. The paper also presents the results of experiments in which we integrate image, video, and speech capabilities into Llama 3 via a compositional approach. We observe this approach performs competitively with the state-of-the-art on image, video, and speech recognition tasks. The resulting models are not yet being broadly released as they are still under development.

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.

Benjamin Ghaemmaghami, Mustafa Ozdal, Rakesh Komuravelli et al.

A key characteristic of deep recommendation models is the immense memory requirements of their embedding tables. These embedding tables can often reach hundreds of gigabytes which increases hardware requirements and training cost. A common technique to reduce model size is to hash all of the categorical variable identifiers (ids) into a smaller space. This hashing reduces the number of unique representations that must be stored in the embedding table; thus decreasing its size. However, this approach introduces collisions between semantically dissimilar ids that degrade model quality. We introduce an alternative approach, Learned Hash Functions, which instead learns a new mapping function that encourages collisions between semantically similar ids. We derive this learned mapping from historical data and embedding access patterns. We experiment with this technique on a production model and find that a mapping informed by the combination of access frequency and a learned low dimension embedding is the most effective. We demonstrate a small improvement relative to the hashing trick and other collision related compression techniques. This is ongoing work that explores the impact of categorical id collisions on recommendation model quality and how those collisions may be controlled to improve model performance.

Omkar Salpekar, Rohan Varma, Kenny Yu et al.

Large-scale training systems typically use synchronous training, requiring all GPUs to be healthy simultaneously. In our experience training on O(100K) GPUs, synchronous training results in a low efficiency due to frequent failures and long recovery time. To address this problem, we propose a novel training paradigm, Fault Tolerant Hybrid-Shared Data Parallelism (FT-HSDP). FT-HSDP uses data parallel replicas as units of fault tolerance. When failures occur, only a single data-parallel replica containing the failed GPU or server is taken offline and restarted, while the other replicas continue training. To realize this idea at scale, FT-HSDP incorporates several techniques: 1) We introduce a Fault Tolerant All Reduce (FTAR) protocol for gradient exchange across data parallel replicas. FTAR relies on the CPU to drive the complex control logic for tasks like adding or removing participants dynamically, and relies on GPU to perform data transfer for best performance. 2) We introduce a non-blocking catch-up protocol, allowing a recovering replica to join training with minimal stall. Compared with fully synchronous training at O(100K) GPUs, FT-HSDP can reduce the stall time due to failure recovery from 10 minutes to 3 minutes, increasing effective training time from 44\% to 80\%. We further demonstrate that FT-HSDP's asynchronous recovery does not bring any meaning degradation to the accuracy of the result model.

Liang Luo, Yinbin Ma, Quanyu Zhu et al.

Recent GPU generations deliver significantly higher FLOPs using lower-precision arithmetic, such as FP8. While successfully applied to large language models (LLMs), its adoption in large recommendation models (LRMs) has been limited. This is because LRMs are numerically sensitive, dominated by small matrix multiplications (GEMMs) followed by normalization, and trained in communication-intensive environments. Applying FP8 directly to LRMs often degrades model quality and prolongs training time. These challenges are inherent to LRM workloads and cannot be resolved merely by introducing better FP8 kernels. Instead, a system-model co-design approach is needed to successfully integrate FP8. We present LoKA (Low-precision Kernel Applications), a framework that makes FP8 practical for LRMs through three principles: profile under realistic distributions to know where low precision is safe, co-design model components with hardware to expand where it is safe, and orchestrate across kernel libraries to maximize the gains. Concretely, LoKA Probe is a statistically grounded, online benchmarking method that learns activation and weight statistics, and quantifies per-layer errors. This process pinpoints safe and unsafe, fast and slow sites for FP8 adoption. LoKA Mods is a set of reusable model adaptations that improve both numerical stability and execution efficiency with FP8. LoKA Dispatch is a runtime that leverages the statistical insights from LoKA Probe to select the fastest FP8 kernel that satisfies the accuracy requirements.

Udit Gupta, Carole-Jean Wu, Xiaodong Wang et al.

The widespread application of deep learning has changed the landscape of computation in the data center. In particular, personalized recommendation for content ranking is now largely accomplished leveraging deep neural networks. However, despite the importance of these models and the amount of compute cycles they consume, relatively little research attention has been devoted to systems for recommendation. To facilitate research and to advance the understanding of these workloads, this paper presents a set of real-world, production-scale DNNs for personalized recommendation coupled with relevant performance metrics for evaluation. In addition to releasing a set of open-source workloads, we conduct in-depth analysis that underpins future system design and optimization for at-scale recommendation: Inference latency varies by 60% across three Intel server generations, batching and co-location of inferences can drastically improve latency-bounded throughput, and the diverse composition of recommendation models leads to different optimization strategies.

Buyun Zhang, Liang Luo, Yuxin Chen et al.

Scaling laws play an instrumental role in the sustainable improvement in model quality. Unfortunately, recommendation models to date do not exhibit such laws similar to those observed in the domain of large language models, due to the inefficiencies of their upscaling mechanisms. This limitation poses significant challenges in adapting these models to increasingly more complex real-world datasets. In this paper, we propose an effective network architecture based purely on stacked factorization machines, and a synergistic upscaling strategy, collectively dubbed Wukong, to establish a scaling law in the domain of recommendation. Wukong's unique design makes it possible to capture diverse, any-order of interactions simply through taller and wider layers. We conducted extensive evaluations on six public datasets, and our results demonstrate that Wukong consistently outperforms state-of-the-art models quality-wise. Further, we assessed Wukong's scalability on an internal, large-scale dataset. The results show that Wukong retains its superiority in quality over state-of-the-art models, while holding the scaling law across two orders of magnitude in model complexity, extending beyond 100 GFLOP/example, where prior arts fall short.

Liang Luo, Buyun Zhang, Michael Tsang et al.

We study a mismatch between the deep learning recommendation models' flat architecture, common distributed training paradigm and hierarchical data center topology. To address the associated inefficiencies, we propose Disaggregated Multi-Tower (DMT), a modeling technique that consists of (1) Semantic-preserving Tower Transform (SPTT), a novel training paradigm that decomposes the monolithic global embedding lookup process into disjoint towers to exploit data center locality; (2) Tower Module (TM), a synergistic dense component attached to each tower to reduce model complexity and communication volume through hierarchical feature interaction; and (3) Tower Partitioner (TP), a feature partitioner to systematically create towers with meaningful feature interactions and load balanced assignments to preserve model quality and training throughput via learned embeddings. We show that DMT can achieve up to 1.9x speedup compared to the state-of-the-art baselines without losing accuracy across multiple generations of hardware at large data center scales.

Ravi Krishna, Aravind Kalaiah, Bichen Wu et al.

Neural architecture search (NAS) methods aim to automatically find the optimal deep neural network (DNN) architecture as measured by a given objective function, typically some combination of task accuracy and inference efficiency. For many areas, such as computer vision and natural language processing, this is a critical, yet still time consuming process. New NAS methods have recently made progress in improving the efficiency of this process. We implement an extensible and modular framework for Differentiable Neural Architecture Search (DNAS) to help solve this problem. We include an overview of the major components of our codebase and how they interact, as well as a section on implementing extensions to it (including a sample), in order to help users adopt our framework for their applications across different categories of deep learning models. To assess the capabilities of our methodology and implementation, we apply DNAS to the problem of ads click-through rate (CTR) prediction, arguably the highest-value and most worked on AI problem at hyperscalers today. We develop and tailor novel search spaces to a Deep Learning Recommendation Model (DLRM) backbone for CTR prediction, and report state-of-the-art results on the Criteo Kaggle CTR prediction dataset.

Ehsan K. Ardestani, Changkyu Kim, Seung Jae Lee et al.

Deep Learning Recommendation Models (DLRM) are widespread, account for a considerable data center footprint, and grow by more than 1.5x per year. With model size soon to be in terabytes range, leveraging Storage ClassMemory (SCM) for inference enables lower power consumption and cost. This paper evaluates the major challenges in extending the memory hierarchy to SCM for DLRM, and presents different techniques to improve performance through a Software Defined Memory. We show how underlying technologies such as Nand Flash and 3DXP differentiate, and relate to real world scenarios, enabling from 5% to 29% power savings.

Zhaoxia, Deng, Jongsoo Park et al.

Tremendous success of machine learning (ML) and the unabated growth in ML model complexity motivated many ML-specific designs in both CPU and accelerator architectures to speed up the model inference. While these architectures are diverse, highly optimized low-precision arithmetic is a component shared by most. Impressive compute throughputs are indeed often exhibited by these architectures on benchmark ML models. Nevertheless, production models such as recommendation systems important to Facebook's personalization services are demanding and complex: These systems must serve billions of users per month responsively with low latency while maintaining high prediction accuracy, notwithstanding computations with many tens of billions parameters per inference. Do these low-precision architectures work well with our production recommendation systems? They do. But not without significant effort. We share in this paper our search strategies to adapt reference recommendation models to low-precision hardware, our optimization of low-precision compute kernels, and the design and development of tool chain so as to maintain our models' accuracy throughout their lifespan during which topic trends and users' interests inevitably evolve. Practicing these low-precision technologies helped us save datacenter capacities while deploying models with up to 5X complexity that would otherwise not be deployed on traditional general-purpose CPUs. We believe these lessons from the trenches promote better co-design between hardware architecture and software engineering and advance the state of the art of ML in industry.

Dheevatsa Mudigere, Yuchen Hao, Jianyu Huang et al.

Deep learning recommendation models (DLRMs) are used across many business-critical services at Facebook and are the single largest AI application in terms of infrastructure demand in its data-centers. In this paper we discuss the SW/HW co-designed solution for high-performance distributed training of large-scale DLRMs. We introduce a high-performance scalable software stack based on PyTorch and pair it with the new evolution of Zion platform, namely ZionEX. We demonstrate the capability to train very large DLRMs with up to 12 Trillion parameters and show that we can attain 40X speedup in terms of time to solution over previous systems. We achieve this by (i) designing the ZionEX platform with dedicated scale-out network, provisioned with high bandwidth, optimal topology and efficient transport (ii) implementing an optimized PyTorch-based training stack supporting both model and data parallelism (iii) developing sharding algorithms capable of hierarchical partitioning of the embedding tables along row, column dimensions and load balancing them across multiple workers; (iv) adding high-performance core operators while retaining flexibility to support optimizers with fully deterministic updates (v) leveraging reduced precision communications, multi-level memory hierarchy (HBM+DDR+SSD) and pipelining. Furthermore, we develop and briefly comment on distributed data ingestion and other supporting services that are required for the robust and efficient end-to-end training in production environments.

Tigran Ishkhanov, Maxim Naumov, Xianjie Chen et al.

In this paper we develop a novel recommendation model that explicitly incorporates time information. The model relies on an embedding layer and TSL attention-like mechanism with inner products in different vector spaces, that can be thought of as a modification of multi-headed attention. This mechanism allows the model to efficiently treat sequences of user behavior of different length. We study the properties of our state-of-the-art model on statistically designed data set. Also, we show that it outperforms more complex models with longer sequence length on the Taobao User Behavior dataset.

Antonio Ginart, Maxim Naumov, Dheevatsa Mudigere et al.

Embedding representations power machine intelligence in many applications, including recommendation systems, but they are space intensive -- potentially occupying hundreds of gigabytes in large-scale settings. To help manage this outsized memory consumption, we explore mixed dimension embeddings, an embedding layer architecture in which a particular embedding vector's dimension scales with its query frequency. Through theoretical analysis and systematic experiments, we demonstrate that using mixed dimensions can drastically reduce the memory usage, while maintaining and even improving the ML performance. Empirically, we show that the proposed mixed dimension layers improve accuracy by 0.1% using half as many parameters or maintain it using 16X fewer parameters for click-through rate prediction task on the Criteo Kaggle dataset.

Hao-Jun Michael Shi, Dheevatsa Mudigere, Maxim Naumov et al.

Modern deep learning-based recommendation systems exploit hundreds to thousands of different categorical features, each with millions of different categories ranging from clicks to posts. To respect the natural diversity within the categorical data, embeddings map each category to a unique dense representation within an embedded space. Since each categorical feature could take on as many as tens of millions of different possible categories, the embedding tables form the primary memory bottleneck during both training and inference. We propose a novel approach for reducing the embedding size in an end-to-end fashion by exploiting complementary partitions of the category set to produce a unique embedding vector for each category without explicit definition. By storing multiple smaller embedding tables based on each complementary partition and combining embeddings from each table, we define a unique embedding for each category at smaller memory cost. This approach may be interpreted as using a specific fixed codebook to ensure uniqueness of each category's representation. Our experimental results demonstrate the effectiveness of our approach over the hashing trick for reducing the size of the embedding tables in terms of model loss and accuracy, while retaining a similar reduction in the number of parameters.

Maxim Naumov, Dheevatsa Mudigere, Hao-Jun Michael Shi et al.

With the advent of deep learning, neural network-based recommendation models have emerged as an important tool for tackling personalization and recommendation tasks. These networks differ significantly from other deep learning networks due to their need to handle categorical features and are not well studied or understood. In this paper, we develop a state-of-the-art deep learning recommendation model (DLRM) and provide its implementation in both PyTorch and Caffe2 frameworks. In addition, we design a specialized parallelization scheme utilizing model parallelism on the embedding tables to mitigate memory constraints while exploiting data parallelism to scale-out compute from the fully-connected layers. We compare DLRM against existing recommendation models and characterize its performance on the Big Basin AI platform, demonstrating its usefulness as a benchmark for future algorithmic experimentation and system co-design.

Jiecao Yu, Jongsoo Park, Maxim Naumov

Deep convolutional neural networks (CNNs) are deployed in various applications but demand immense computational requirements. Pruning techniques and Winograd convolution are two typical methods to reduce the CNN computation. However, they cannot be directly combined because Winograd transformation fills in the sparsity resulting from pruning. Li et al. (2017) propose sparse Winograd convolution in which weights are directly pruned in the Winograd domain, but this technique is not very practical because Winograd-domain retraining requires low learning rates and hence significantly longer training time. Besides, Liu et al. (2018) move the ReLU function into the Winograd domain, which can help increase the weight sparsity but requires changes in the network structure. To achieve a high Winograd-domain weight sparsity without changing network structures, we propose a new pruning method, spatial-Winograd pruning. As the first step, spatial-domain weights are pruned in a structured way, which efficiently transfers the spatial-domain sparsity into the Winograd domain and avoids Winograd-domain retraining. For the next step, we also perform pruning and retraining directly in the Winograd domain but propose to use an importance factor matrix to adjust weight importance and weight gradients. This adjustment makes it possible to effectively retrain the pruned Winograd-domain network without changing the network structure. For the three models on the datasets of CIFAR10, CIFAR-100, and ImageNet, our proposed method can achieve the Winograd domain sparsities of 63%, 50%, and 74%, respectively.

Maxim Naumov

In this note we discuss a common misconception, namely that embeddings are always used to reduce the dimensionality of the item space. We show that when we measure dimensionality in terms of information entropy then the embedding of sparse probability distributions, that can be used to represent sparse features or data, may or not reduce the dimensionality of the item space. However, the embeddings do provide a different and often more meaningful representation of the items for a particular task at hand. Also, we give upper bounds and more precise guidelines for choosing the embedding dimension.

Jongsoo Park, Maxim Naumov, Protonu Basu et al.

The application of deep learning techniques resulted in remarkable improvement of machine learning models. In this paper provides detailed characterizations of deep learning models used in many Facebook social network services. We present computational characteristics of our models, describe high performance optimizations targeting existing systems, point out their limitations and make suggestions for the future general-purpose/accelerated inference hardware. Also, we highlight the need for better co-design of algorithms, numerics and computing platforms to address the challenges of workloads often run in data centers.

Maxim Naumov, Utku Diril, Jongsoo Park et al.

Deep learning models have been successfully used in computer vision and many other fields. We propose an unorthodox algorithm for performing quantization of the model parameters. In contrast with popular quantization schemes based on thresholds, we use a novel technique based on periodic functions, such as continuous trigonometric sine or cosine as well as non-continuous hat functions. We apply these functions component-wise and add the sum over the model parameters as a regularizer to the model loss during training. The frequency and amplitude hyper-parameters of these functions can be adjusted during training. The regularization pushes the weights into discrete points that can be encoded as integers. We show that using this technique the resulting quantized models exhibit the same accuracy as the original ones on CIFAR-10 and ImageNet datasets.

Assaf Eisenman, Maxim Naumov, Darryl Gardner et al.

Typical large-scale recommender systems use deep learning models that are stored on a large amount of DRAM. These models often rely on embeddings, which consume most of the required memory. We present Bandana, a storage system that reduces the DRAM footprint of embeddings, by using Non-volatile Memory (NVM) as the primary storage medium, with a small amount of DRAM as cache. The main challenge in storing embeddings on NVM is its limited read bandwidth compared to DRAM. Bandana uses two primary techniques to address this limitation: first, it stores embedding vectors that are likely to be read together in the same physical location, using hypergraph partitioning, and second, it decides the number of embedding vectors to cache in DRAM by simulating dozens of small caches. These techniques allow Bandana to increase the effective read bandwidth of NVM by 2-3x and thereby significantly reduce the total cost of ownership.

Maxim Naumov

We show that the forward and backward propagation can be formulated as a solution of lower and upper triangular systems of equations. For standard feedforward (FNNs) and recurrent neural networks (RNNs) the triangular systems are always block bi-diagonal, while for a general computation graph (directed acyclic graph) they can have a more complex triangular sparsity pattern. We discuss direct and iterative parallel algorithms that can be used for their solution and interpreted as different ways of performing model parallelism. Also, we show that for FNNs and RNNs with $k$ layers and $τ$ time steps the backward propagation can be performed in parallel in O($\log k$) and O($\log k \log τ$) steps, respectively. Finally, we outline the generalization of this technique using Jacobians that potentially allows us to handle arbitrary layers.

Aditya Devarakonda, Maxim Naumov, Michael Garland

Training deep neural networks with Stochastic Gradient Descent, or its variants, requires careful choice of both learning rate and batch size. While smaller batch sizes generally converge in fewer training epochs, larger batch sizes offer more parallelism and hence better computational efficiency. We have developed a new training approach that, rather than statically choosing a single batch size for all epochs, adaptively increases the batch size during the training process. Our method delivers the convergence rate of small batch sizes while achieving performance similar to large batch sizes. We analyse our approach using the standard AlexNet, ResNet, and VGG networks operating on the popular CIFAR-10, CIFAR-100, and ImageNet datasets. Our results demonstrate that learning with adaptive batch sizes can improve performance by factors of up to 6.25 on 4 NVIDIA Tesla P100 GPUs while changing accuracy by less than 1% relative to training with fixed batch sizes.

Maxim Naumov

In this paper we focus on the linear algebra theory behind feedforward (FNN) and recurrent (RNN) neural networks. We review backward propagation, including backward propagation through time (BPTT). Also, we obtain a new exact expression for Hessian, which represents second order effects. We show that for $t$ time steps the weight gradient can be expressed as a rank-$t$ matrix, while the weight Hessian is as a sum of $t^{2}$ Kronecker products of rank-$1$ and $W^{T}AW$ matrices, for some matrix $A$ and weight matrix $W$. Also, we show that for a mini-batch of size $r$, the weight update can be expressed as a rank-$rt$ matrix. Finally, we briefly comment on the eigenvalues of the Hessian matrix.