Post-training quantization (PTQ) techniques applied to weights, activations, and the KV cache greatly reduce memory usage, latency, and power consumption of Large Language Models (LLMs), but may lead to large quantization errors when outliers are present. Rotating activation or weight matrices helps remove outliers and benefits quantization. In this work, we identify a collection of applicable rotation parameterizations that lead to identical outputs in full-precision Transformer architectures while enhancing quantization accuracy. In addition, we find that some random rotations lead to much better quantization than others, with an up to 13 points difference in downstream zero-shot reasoning performance. As a result, we propose SpinQuant, a novel approach that incorporates learned rotation matrices for optimal quantized network accuracy. With 4-bit quantization of weight, activation, and KV-cache, SpinQuant narrows the accuracy gap on zero-shot reasoning tasks with full precision to merely 2.9 points on the LLaMA-2 7B model, surpassing LLM-QAT by 19.1 points and SmoothQuant by 25.0 points. Furthermore, SpinQuant also outperforms concurrent work QuaRot, which applies random rotations to remove outliers. In particular, for LLaMA-3 8B models that are hard to quantize, SpinQuant reduces the gap to full precision by up to 45.1% relative to QuaRot. Code is available at https://github.com/facebookresearch/SpinQuant.

Very deep neural networks lead to significantly better performance on various real tasks. However, it usually causes slow inference and is hard to be deployed on real-world devices. How to reduce the number of layers to save memory and to accelerate the inference is an eye-catching topic. In this work, we introduce an intermediate objective, a continuous-time network, before distilling deep networks into shallow networks. First, we distill a given deep network into a continuous-time neural flow model, which can be discretized with an ODE solver and the inference requires passing through the network multiple times. By forcing the flow transport trajectory to be straight lines, we find that it is easier to compress the infinite step model into a one-step neural flow model, which only requires passing through the flow model once. Secondly, we refine the one-step flow model together with the final head layer with knowledge distillation and finally, we can replace the given deep network with this one-step flow network. Empirically, we demonstrate that our method outperforms direct distillation and other baselines on different model architectures (e.g. ResNet, ViT) on image classification and semantic segmentation tasks. We also manifest that our distilled model naturally serves as an early-exit dynamic inference model.

Embedding learning is an important technique in deep recommendation models to map categorical features to dense vectors. However, the embedding tables often demand an extremely large number of parameters, which become the storage and efficiency bottlenecks. Distributed training solutions have been adopted to partition the embedding tables into multiple devices. However, the embedding tables can easily lead to imbalances if not carefully partitioned. This is a significant design challenge of distributed systems named embedding table sharding, i.e., how we should partition the embedding tables to balance the costs across devices, which is a non-trivial task because 1) it is hard to efficiently and precisely measure the cost, and 2) the partition problem is known to be NP-hard. In this work, we introduce our novel practice in Meta, namely AutoShard, which uses a neural cost model to directly predict the multi-table costs and leverages deep reinforcement learning to solve the partition problem. Experimental results on an open-sourced large-scale synthetic dataset and Meta's production dataset demonstrate the superiority of AutoShard over the heuristics. Moreover, the learned policy of AutoShard can transfer to sharding tasks with various numbers of tables and different ratios of the unseen tables without any fine-tuning. Furthermore, AutoShard can efficiently shard hundreds of tables in seconds. The effectiveness, transferability, and efficiency of AutoShard make it desirable for production use. Our algorithms have been deployed in Meta production environment. A prototype is available at https://github.com/daochenzha/autoshard

One of the key challenges of learning an online recommendation model is the temporal domain shift, which causes the mismatch between the training and testing data distribution and hence domain generalization error. To overcome, we propose to learn a meta future gradient generator that forecasts the gradient information of the future data distribution for training so that the recommendation model can be trained as if we were able to look ahead at the future of its deployment. Compared with Batch Update, a widely used paradigm, our theory suggests that the proposed algorithm achieves smaller temporal domain generalization error measured by a gradient variation term in a local regret. We demonstrate the empirical advantage by comparing with various representative baselines.

Deep learning recommendation systems at scale have provided remarkable gains through increasing model capacity (i.e. wider and deeper neural networks), but it comes at significant training cost and infrastructure cost. Model pruning is an effective technique to reduce computation overhead for deep neural networks by removing redundant parameters. However, modern recommendation systems are still thirsty for model capacity due to the demand for handling big data. Thus, pruning a recommendation model at scale results in a smaller model capacity and consequently lower accuracy. To reduce computation cost without sacrificing model capacity, we propose a dynamic training scheme, namely alternate model growth and pruning, to alternatively construct and prune weights in the course of training. Our method leverages structured sparsification to reduce computational cost without hurting the model capacity at the end of offline training so that a full-size model is available in the recurring training stage to learn new data in real time. To the best of our knowledge, this is the first work to provide in-depth experiments and discussion of applying structural dynamics to recommendation systems at scale to reduce training cost. The proposed method is validated with an open-source deep learning recommendation model (DLRM) and state-of-the-art industrial-scale production models.

In this paper, we consider hybrid parallelism---a paradigm that employs both Data Parallelism (DP) and Model Parallelism (MP)---to scale distributed training of large recommendation models. We propose a compression framework called Dynamic Communication Thresholding (DCT) for communication-efficient hybrid training. DCT filters the entities to be communicated across the network through a simple hard-thresholding function, allowing only the most relevant information to pass through. For communication efficient DP, DCT compresses the parameter gradients sent to the parameter server during model synchronization. The threshold is updated only once every few thousand iterations to reduce the computational overhead of compression. For communication efficient MP, DCT incorporates a novel technique to compress the activations and gradients sent across the network during the forward and backward propagation, respectively. This is done by identifying and updating only the most relevant neurons of the neural network for each training sample in the data. We evaluate DCT on publicly available natural language processing and recommender models and datasets, as well as recommendation systems used in production at Facebook. DCT reduces communication by at least 100x and 20x during DP and MP, respectively. The algorithm has been deployed in production, and it improves end-to-end training time for a state-of-the-art industrial recommender model by 37%, without any loss in performance.

Neural network based recommendation models are widely used to power many internet-scale applications including product recommendation and feed ranking. As the models become more complex and more training data is required during training, improving the training scalability of these recommendation models becomes an urgent need. However, improving the scalability without sacrificing the model quality is challenging. In this paper, we conduct an in-depth analysis of the scalability bottleneck in existing training architecture on large scale CPU clusters. Based on these observations, we propose a new training architecture called Hierarchical Training, which exploits both data parallelism and model parallelism for the neural network part of the model within a group. We implement hierarchical training with a two-layer design: a tagging system that decides the operator placement and a net transformation system that materializes the training plans, and integrate hierarchical training into existing training stack. We propose several optimizations to improve the scalability of hierarchical training including model architecture optimization, communication compression, and various system-level improvements. Extensive experiments at massive scale demonstrate that hierarchical training can speed up distributed recommendation model training by 1.9x without model quality drop.

The presence of multiple MCs and their integration into the on-chip network fabric creates a highly concurrent system that can support significant levels of memory level parallelism (MLP) across cores. This work exposes the trade-off between DRAM parameters, bank level parallelism (BLP), and row buffer hit rate that exposes the amount of effective BLP that is necessary to approximate a 100% hit rate. We further study how this trade-off can be controlled and propose a class of global (system) and local (within an MC) address mappings that can be tuned to optimize the performance across a set of multiprogrammed benchmarks.

The on-chip memory is a dominant source of power and energy consumption in modern and future processors. This paper explores the use of a new emerging non-volatile memory technology as a replacement for SRAM based lower level caches - Spin Torque Transfer(STT) RAM. While STTRAM achieves a reduction in leakage energy of 90% compared to SRAM, the dynamic energy for a write operation is 2X that of SRAM. Consequently, we propose additional microarchitectural optimizations to reduce overall dynamic energy which achieve an average reduction in dynamic energy over the base case of 30% with a range of 16% to 60% across 10 benchmarks.