Phantora: Live GPU Cluster Simulation for Machine Learning System Performance Estimation

To accommodate ever-increasing model complexity, modern machine learning (ML) systems have to scale to large GPU clusters. Changes in ML model architecture, ML system implementation, and cluster configuration can significantly affect overall ML system performance. However, quantifying the performance impact before deployment is challenging. Existing performance estimation methods use performance modeling or static workload simulation. These techniques are not general: they requires significant human effort and computation capacity to generate training data or a workload. It is also difficult to adapt ML systems to use these techniques. This paper introduces, Phantora, a live GPU cluster simulator for performance estimation. Phantora runs minimally modified ML models and frameworks, intercepting and simulating GPU-related operations to enable high-fidelity performance estimation. Phantora overcomes several research challenges in integrating an event-driven network simulator with live system execution, and introduces a set of techniques to improve simulation speed, scalability, and accuracy. Our evaluation results show that Phantora can deliver similar estimation accuracy to the state-of-the-art workload simulation approach with only one GPU, while reducing human effort and increasing generalizability.

External Large Foundation Model: How to Efficiently Serve Trillions of Parameters for Online Ads Recommendation

Ads recommendation is a prominent service of online advertising systems and has been actively studied. Recent studies indicate that scaling-up and advanced design of the recommendation model can bring significant performance improvement. However, with a larger model scale, such prior studies have a significantly increasing gap from industry as they often neglect two fundamental challenges in industrial-scale applications. First, training and inference budgets are restricted for the model to be served, exceeding which may incur latency and impair user experience. Second, large-volume data arrive in a streaming mode with data distributions dynamically shifting, as new users/ads join and existing users/ads leave the system. We propose the External Large Foundation Model (ExFM) framework to address the overlooked challenges. Specifically, we develop external distillation and a data augmentation system (DAS) to control the computational cost of training/inference while maintaining high performance. We design the teacher in a way like a foundation model (FM) that can serve multiple students as vertical models (VMs) to amortize its building cost. We propose Auxiliary Head and Student Adapter to mitigate the data distribution gap between FM and VMs caused by the streaming data issue. Comprehensive experiments on internal industrial-scale applications and public datasets demonstrate significant performance gain by ExFM.

The Efficiency vs. Accuracy Trade-off: Optimizing RAG-Enhanced LLM Recommender Systems Using Multi-Head Early Exit

The deployment of Large Language Models (LLMs) in recommender systems for predicting Click-Through Rates (CTR) necessitates a delicate balance between computational efficiency and predictive accuracy. This paper presents an optimization framework that combines Retrieval-Augmented Generation (RAG) with an innovative multi-head early exit architecture to concurrently enhance both aspects. By integrating Graph Convolutional Networks (GCNs) as efficient retrieval mechanisms, we are able to significantly reduce data retrieval times while maintaining high model performance. The early exit strategy employed allows for dynamic termination of model inference, utilizing real-time predictive confidence assessments across multiple heads. This not only quickens the responsiveness of LLMs but also upholds or improves their accuracy, making it ideal for real-time application scenarios. Our experiments demonstrate how this architecture effectively decreases computation time without sacrificing the accuracy needed for reliable recommendation delivery, establishing a new standard for efficient, real-time LLM deployment in commercial systems.

Mixture-of-Transformers: A Sparse and Scalable Architecture for Multi-Modal Foundation Models

The development of large language models (LLMs) has expanded to multi-modal systems capable of processing text, images, and speech within a unified framework. Training these models demands significantly larger datasets and computational resources compared to text-only LLMs. To address the scaling challenges, we introduce Mixture-of-Transformers (MoT), a sparse multi-modal transformer architecture that significantly reduces pretraining computational costs. MoT decouples non-embedding parameters of the model by modality -- including feed-forward networks, attention matrices, and layer normalization -- enabling modality-specific processing with global self-attention over the full input sequence. We evaluate MoT across multiple settings and model scales. In the Chameleon 7B setting (autoregressive text-and-image generation), MoT matches the dense baseline's performance using only 55.8\% of the FLOPs. When extended to include speech, MoT reaches speech performance comparable to the dense baseline with only 37.2\% of the FLOPs. In the Transfusion setting, where text and image are trained with different objectives, a 7B MoT model matches the image modality performance of the dense baseline with one third of the FLOPs, and a 760M MoT model outperforms a 1.4B dense baseline across key image generation metrics. System profiling further highlights MoT's practical benefits, achieving dense baseline image quality in 47.2\% of the wall-clock time and text quality in 75.6\% of the wall-clock time (measured on AWS p4de.24xlarge instances with NVIDIA A100 GPUs).

MoMa: Efficient Early-Fusion Pre-training with Mixture of Modality-Aware Experts

We introduce MoMa, a novel modality-aware mixture-of-experts (MoE) architecture designed for pre-training mixed-modal, early-fusion language models. MoMa processes images and text in arbitrary sequences by dividing expert modules into modality-specific groups. These groups exclusively process designated tokens while employing learned routing within each group to maintain semantically informed adaptivity. Our empirical results reveal substantial pre-training efficiency gains through this modality-specific parameter allocation. Under a 1-trillion-token training budget, the MoMa 1.4B model, featuring 4 text experts and 4 image experts, achieves impressive FLOPs savings: 3.7x overall, with 2.6x for text and 5.2x for image processing compared to a compute-equivalent dense baseline, measured by pre-training loss. This outperforms the standard expert-choice MoE with 8 mixed-modal experts, which achieves 3x overall FLOPs savings (3x for text, 2.8x for image). Combining MoMa with mixture-of-depths (MoD) further improves pre-training FLOPs savings to 4.2x overall (text: 3.4x, image: 5.3x), although this combination hurts performance in causal inference due to increased sensitivity to router accuracy. These results demonstrate MoMa's potential to significantly advance the efficiency of mixed-modal, early-fusion language model pre-training, paving the way for more resource-efficient and capable multimodal AI systems.

Wukong: Towards a Scaling Law for Large-Scale Recommendation

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 or equivalently up to Large Language Model (GPT-3) training compute scale, where prior arts fall short.

Disaggregated Multi-Tower: Topology-aware Modeling Technique for Efficient Large-Scale Recommendation

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.

Pre-train and Search: Efficient Embedding Table Sharding with Pre-trained Neural Cost Models

Sharding a large machine learning model across multiple devices to balance the costs is important in distributed training. This is challenging because partitioning is NP-hard, and estimating the costs accurately and efficiently is difficult. In this work, we explore a "pre-train, and search" paradigm for efficient sharding. The idea is to pre-train a universal and once-for-all neural network to predict the costs of all the possible shards, which serves as an efficient sharding simulator. Built upon this pre-trained cost model, we then perform an online search to identify the best sharding plans given any specific sharding task. We instantiate this idea in deep learning recommendation models (DLRMs) and propose NeuroShard for embedding table sharding. NeuroShard pre-trains neural cost models on augmented tables to cover various sharding scenarios. Then it identifies the best column-wise and table-wise sharding plans with beam search and greedy grid search, respectively. Experiments show that NeuroShard significantly and consistently outperforms the state-of-the-art on the benchmark sharding dataset, achieving up to 23.8% improvement. When deployed in an ultra-large production DLRM with multi-terabyte embedding tables, NeuroShard achieves 11.6% improvement in embedding costs over the state-of-the-art, which translates to 6.6% end-to-end training throughput improvement. To facilitate future research of the "pre-train, and search" paradigm in ML for Systems, we open-source our code at https://github.com/daochenzha/neuroshard

Self-discipline on multiple channels

Self-distillation relies on its own information to improve the generalization ability of the model and has a bright future. Existing self-distillation methods either require additional models, model modification, or batch size expansion for training, which increases the difficulty of use, memory consumption, and computational cost. This paper developed Self-discipline on multiple channels(SMC), which combines consistency regularization with self-distillation using the concept of multiple channels. Conceptually, SMC consists of two steps: 1) each channel data is simultaneously passed through the model to obtain its corresponding soft label, and 2) the soft label saved in the previous step is read together with the soft label obtained from the current channel data through the model to calculate the loss function. SMC uses consistent regularization and self-distillation to improve the generalization ability of the model and the robustness of the model to noisy labels. We named the SMC containing only two channels as SMC-2. Comparative experimental results on both datasets show that SMC-2 outperforms Label Smoothing Regularizaion and Self-distillation From The Last Mini-batch on all models, and outperforms the state-of-the-art Sharpness-Aware Minimization method on 83% of the models.Compatibility of SMC-2 and data augmentation experimental results show that using both SMC-2 and data augmentation improves the generalization ability of the model between 0.28% and 1.80% compared to using only data augmentation. Ultimately, the results of the label noise interference experiments show that SMC-2 curbs the tendency that the model's generalization ability decreases in the late training period due to the interference of label noise. The code is available at https://github.com/JiuTiannn/SMC-Self-discipline-on-multiple-channels.

PyTorch FSDP: Experiences on Scaling Fully Sharded Data Parallel

It is widely acknowledged that large models have the potential to deliver superior performance across a broad range of domains. Despite the remarkable progress made in the field of machine learning systems research, which has enabled the development and exploration of large models, such abilities remain confined to a small group of advanced users and industry leaders, resulting in an implicit technical barrier for the wider community to access and leverage these technologies. In this paper, we introduce PyTorch Fully Sharded Data Parallel (FSDP) as an industry-grade solution for large model training. FSDP has been closely co-designed with several key PyTorch core components including Tensor implementation, dispatcher system, and CUDA memory caching allocator, to provide non-intrusive user experiences and high training efficiency. Additionally, FSDP natively incorporates a range of techniques and settings to optimize resource utilization across a variety of hardware configurations. The experimental results demonstrate that FSDP is capable of achieving comparable performance to Distributed Data Parallel while providing support for significantly larger models with near-linear scalability in terms of TFLOPS.