Abstract:Despite the advanced intelligence abilities of large language models (LLMs) in various applications, they still face significant computational and storage demands. Knowledge Distillation (KD) has emerged as an effective strategy to improve the performance of a smaller LLM (i.e., the student model) by transferring knowledge from a high-performing LLM (i.e., the teacher model). Prevailing techniques in LLM distillation typically use a black-box model API to generate high-quality pretrained and aligned datasets, or utilize white-box distillation by altering the loss function to better transfer knowledge from the teacher LLM. However, these methods ignore the knowledge differences between the student and teacher LLMs across domains. This results in excessive focus on domains with minimal performance gaps and insufficient attention to domains with large gaps, reducing overall performance. In this paper, we introduce a new LLM distillation framework called DDK, which dynamically adjusts the composition of the distillation dataset in a smooth manner according to the domain performance differences between the teacher and student models, making the distillation process more stable and effective. Extensive evaluations show that DDK significantly improves the performance of student models, outperforming both continuously pretrained baselines and existing knowledge distillation methods by a large margin.
Abstract:This paper documents our characterization study and practices for serving text-to-image requests with stable diffusion models in production. We first comprehensively analyze inference request traces for commercial text-to-image applications. It commences with our observation that add-on modules, i.e., ControlNets and LoRAs, that augment the base stable diffusion models, are ubiquitous in generating images for commercial applications. Despite their efficacy, these add-on modules incur high loading overhead, prolong the serving latency, and swallow up expensive GPU resources. Driven by our characterization study, we present SwiftDiffusion, a system that efficiently generates high-quality images using stable diffusion models and add-on modules. To achieve this, SwiftDiffusion reconstructs the existing text-to-image serving workflow by identifying the opportunities for parallel computation and distributing ControlNet computations across multiple GPUs. Further, SwiftDiffusion thoroughly analyzes the dynamics of image generation and develops techniques to eliminate the overhead associated with LoRA loading and patching while preserving the image quality. Last, SwiftDiffusion proposes specialized optimizations in the backbone architecture of the stable diffusion models, which are also compatible with the efficient serving of add-on modules. Compared to state-of-the-art text-to-image serving systems, SwiftDiffusion reduces serving latency by up to 5x and improves serving throughput by up to 2x without compromising image quality.
Abstract:Continual Pre-Training (CPT) on Large Language Models (LLMs) has been widely used to expand the model's fundamental understanding of specific downstream domains (e.g., math and code). For the CPT on domain-specific LLMs, one important question is how to choose the optimal mixture ratio between the general-corpus (e.g., Dolma, Slim-pajama) and the downstream domain-corpus. Existing methods usually adopt laborious human efforts by grid-searching on a set of mixture ratios, which require high GPU training consumption costs. Besides, we cannot guarantee the selected ratio is optimal for the specific domain. To address the limitations of existing methods, inspired by the Scaling Law for performance prediction, we propose to investigate the Scaling Law of the Domain-specific Continual Pre-Training (D-CPT Law) to decide the optimal mixture ratio with acceptable training costs for LLMs of different sizes. Specifically, by fitting the D-CPT Law, we can easily predict the general and downstream performance of arbitrary mixture ratios, model sizes, and dataset sizes using small-scale training costs on limited experiments. Moreover, we also extend our standard D-CPT Law on cross-domain settings and propose the Cross-Domain D-CPT Law to predict the D-CPT law of target domains, where very small training costs (about 1% of the normal training costs) are needed for the target domains. Comprehensive experimental results on six downstream domains demonstrate the effectiveness and generalizability of our proposed D-CPT Law and Cross-Domain D-CPT Law.
Abstract:High-concurrency asynchronous training upon parameter server (PS) architecture and high-performance synchronous training upon all-reduce (AR) architecture are the most commonly deployed distributed training modes for recommender systems. Although the synchronous AR training is designed to have higher training efficiency, the asynchronous PS training would be a better choice on training speed when there are stragglers (slow workers) in the shared cluster, especially under limited computing resources. To take full advantages of these two training modes, an ideal way is to switch between them upon the cluster status. We find two obstacles to a tuning-free approach: the different distribution of the gradient values and the stale gradients from the stragglers. In this paper, we propose Global Batch gradients Aggregation (GBA) over PS, which aggregates and applies gradients with the same global batch size as the synchronous training. A token-control process is implemented to assemble the gradients and decay the gradients with severe staleness. We provide the convergence analysis to demonstrate the robustness of GBA over the recommendation models against the gradient staleness. Experiments on three industrial-scale recommendation tasks show that GBA is an effective tuning-free approach for switching. Compared to the state-of-the-art derived asynchronous training, GBA achieves up to 0.2% improvement on the AUC metric, which is significant for the recommendation models. Meanwhile, under the strained hardware resource, GBA speeds up at least 2.4x compared to the synchronous training.
Abstract:The development of personalized recommendation has significantly improved the accuracy of information matching and the revenue of e-commerce platforms. Recently, it has 2 trends: 1) recommender systems must be trained timely to cope with ever-growing new products and ever-changing user interests from online marketing and social network; 2) SOTA recommendation models introduce DNN modules to improve prediction accuracy. Traditional CPU-based recommender systems cannot meet these two trends, and GPU- centric training has become a trending approach. However, we observe that GPU devices in training recommender systems are underutilized, and they cannot attain an expected throughput improvement as what it has achieved in CV and NLP areas. This issue can be explained by two characteristics of these recommendation models: First, they contain up to a thousand input feature fields, introducing fragmentary and memory-intensive operations; Second, the multiple constituent feature interaction submodules introduce substantial small-sized compute kernels. To remove this roadblock to the development of recommender systems, we propose a novel framework named PICASSO to accelerate the training of recommendation models on commodity hardware. Specifically, we conduct a systematic analysis to reveal the bottlenecks encountered in training recommendation models. We leverage the model structure and data distribution to unleash the potential of hardware through our packing, interleaving, and caching optimization. Experiments show that PICASSO increases the hardware utilization by an order of magnitude on the basis of SOTA baselines and brings up to 6x throughput improvement for a variety of industrial recommendation models. Using the same hardware budget in production, PICASSO on average shortens the walltime of daily training tasks by 7 hours, significantly reducing the delay of continuous delivery.
Abstract:Mixture-of-Experts (MoE) models can achieve promising results with outrageous large amount of parameters but constant computation cost, and thus it has become a trend in model scaling. Still it is a mystery how MoE layers bring quality gains by leveraging the parameters with sparse activation. In this work, we investigate several key factors in sparse expert models. We observe that load imbalance may not be a significant problem affecting model quality, contrary to the perspectives of recent studies, while the number of sparsely activated experts $k$ and expert capacity $C$ in top-$k$ routing can significantly make a difference in this context. Furthermore, we take a step forward to propose a simple method called expert prototyping that splits experts into different prototypes and applies $k$ top-$1$ routing. This strategy improves the model quality but maintains constant computational costs, and our further exploration on extremely large-scale models reflects that it is more effective in training larger models. We push the model scale to over $1$ trillion parameters and implement it on solely $480$ NVIDIA V100-32GB GPUs, in comparison with the recent SOTAs on $2048$ TPU cores. The proposed giant model achieves substantial speedup in convergence over the same-size baseline.