Jiangsu Du

Jiangsu Du, Ziming Liu, Jiarui Fang et al. · berkeley

Large transformer models display promising performance on a wide range of natural language processing (NLP) tasks. Although the AI community has expanded the model scale to the trillion parameter level, the practical deployment of 10-100 billion parameter models is still uncertain due to the latency, throughput, and memory constraints. In this paper, we proposed EnergonAI to solve the challenges of the efficient deployment of 10-100 billion parameter transformer models on single- or multi-GPU systems. EnergonAI adopts a hierarchy-controller system architecture to coordinate multiple devices and efficiently support different parallel patterns. It delegates the execution of sub-models to multiple workers in the single-controller style and applies tensor parallelism and pipeline parallelism among the workers in a multi-controller style. Upon the novel architecture, we propose three techniques, i.e. non-blocking pipeline parallelism, distributed redundant computation elimination, and peer memory pooling. EnergonAI enables the users to program complex parallel code the same as a serial one. Compared with the FasterTransformer, we have proven that EnergonAI has superior performance on latency and throughput. In our experiments, EnergonAI can achieve 37% latency reduction in tensor parallelism, 10% scalability improvement in pipeline parallelism, and it improves the model scale inferred on a single GPU by using a larger heterogeneous memory space at cost of limited performance reduction.

Jiangsu Du, Dongsheng Li, Yingpeng Wen et al.

Novel artificial intelligence (AI) technology has expedited various scientific research, e.g., cosmology, physics and bioinformatics, inevitably becoming a significant category of workload on high performance computing (HPC) systems. Existing AI benchmarks tend to customize well-recognized AI applications, so as to evaluate the AI performance of HPC systems under predefined problem size, in terms of datasets and AI models. Due to lack of scalability on the problem size, static AI benchmarks might be under competent to help understand the performance trend of evolving AI applications on HPC systems, in particular, the scientific AI applications on large-scale systems. In this paper, we propose a scalable evaluation methodology (SAIH) for analyzing the AI performance trend of HPC systems with scaling the problem sizes of customized AI applications. To enable scalability, SAIH builds a set of novel mechanisms for augmenting problem sizes. As the data and model constantly scale, we can investigate the trend and range of AI performance on HPC systems, and further diagnose system bottlenecks. To verify our methodology, we augment a cosmological AI application to evaluate a real HPC system equipped with GPUs as a case study of SAIH.

Shengyuan Ye, Bei Ouyang, Tianyi Qian et al.

Large language models (LLMs) have unlocked a plethora of powerful applications at the network edge, such as intelligent personal assistants. Data privacy and security concerns have prompted a shift towards edge-based fine-tuning of personal LLMs, away from cloud reliance. However, this raises issues of computational intensity and resource scarcity, hindering training efficiency and feasibility. While current studies investigate parameter-efficient fine-tuning (PEFT) techniques to mitigate resource constraints, our analysis indicates that these techniques are not sufficiently resource-efficient for edge devices. To tackle these challenges, we propose Pluto and Charon (PAC), a time and memory efficient collaborative edge AI framework for personal LLMs fine-tuning. PAC breaks the resource wall of personal LLMs fine-tuning with a sophisticated algorithm-system co-design. (1) Algorithmically, PAC implements a personal LLMs fine-tuning technique that is efficient in terms of parameters, time, and memory. It utilizes Parallel Adapters to circumvent the need for a full backward pass through the LLM backbone. Additionally, an activation cache mechanism further streamlining the process by negating the necessity for repeated forward passes across multiple epochs. (2) Systematically, PAC leverages edge devices in close proximity, pooling them as a collective resource for in-situ personal LLMs fine-tuning, utilizing a hybrid data and pipeline parallelism to orchestrate distributed training. The use of the activation cache eliminates the need for forward pass through the LLM backbone,enabling exclusive fine-tuning of the Parallel Adapters using data parallelism. Extensive evaluation based on prototype implementation demonstrates that PAC remarkably outperforms state-of-the-art approaches, achieving up to 8.64x end-to-end speedup and up to 88.16% reduction in memory footprint.

Fengyao Bai, Hongbin Zhang, Zhitao Chen et al.

High-throughput inference serving is essential for applications built on large language models (LLMs). Existing serving frameworks reduce request-level and batch-level bubbles through batching and scheduling, but often overlook bubbles within each decode iteration. Tokens generated in the same iteration may incur different costs because they depend on KV caches of different lengths; tokens with long KV caches can become bottlenecks and delay the next iteration. We propose AlignedServe, an LLM serving framework built around prefix-aware batching. It groups requests with similar KV-cache lengths into the same batch to reduce iteration-level bubbles. To support this policy efficiently, AlignedServe uses large CPU memory to maintain sufficient in-flight requests for batching and applies a batch-level scheduling policy to reduce batch-level bubbles. It also introduces a GPU-Prefetch-For-GPU architecture, where one GPU prefetches KV cache for another to reduce CPU-to-GPU transfer latency. Experiments on synthetic and application workloads show that AlignedServe improves decoding throughput by up to 1.98 times and reduces latency by up to 7.4 times over state-of-the-art systems.

Hongbin Zhang, Taosheng Wei, Jiazhi Jiang et al.

Offline LLM inference seeks to maximize request processing under fixed budgets, making commodity GPU servers a promising choice. However, prior work typically considers offloading and parallelism in isolation, resulting in suboptimal performance. In this paper, we propose PipeMax, a high-throughput LLM inference system that integrates pipeline parallelism with offloading to overcome interconnect and memory constraints on GPU servers. Particularly, pipeline parallelism naturally incurs low communication overhead and keeps only one batch active on each GPU at a time, which enables offloading the KV cache of inactive batches. By coordinating computation with offloading data movement, PipeMax effectively expands GPU memory capacity and sustains large-batch execution. Experiments show that PipeMax achieves up to 2.51x higher throughput than vLLM, and up to 1.42x and 1.38x higher throughput than state-of-the-art high-throughput LLM systems, respectively, on an 8-GPU node.

Hao Liu, Ye Huang, Chenghuan Huang et al.

Video Diffusion Transformer (DiT) models are a dominant approach for high-quality video generation but suffer from high inference cost due to iterative denoising. Existing caching approaches primarily exploit similarity within the diffusion process of a single request to skip redundant denoising steps. In this paper, we introduce Chorus, a caching approach that leverages similarity across requests to accelerate video diffusion model serving. Chorus achieves up to 45\% speedup on industrial 4-step distilled models, where prior intra-request caching approaches are ineffective. Particularly, Chorus employs a three-stage caching strategy along the denoising process. Stage 1 performs full reuse of latent features from similar requests. Stage 2 exploits inter-request caching in specific latent regions during intermediate denoising steps. This stage is combined with Token-Guided Attention Amplification to improve semantic alignment between the generated video and the conditional prompts, thereby extending the applicability of full reuse to later denoising steps.

Shenggan Cheng, Yuanxin Wei, Lansong Diao et al.

Leveraging the diffusion transformer (DiT) architecture, models like Sora, CogVideoX and Wan have achieved remarkable progress in text-to-video, image-to-video, and video editing tasks. Despite these advances, diffusion-based video generation remains computationally intensive, especially for high-resolution, long-duration videos. Prior work accelerates its inference by skipping computation, usually at the cost of severe quality degradation. In this paper, we propose SRDiffusion, a novel framework that leverages collaboration between large and small models to reduce inference cost. The large model handles high-noise steps to ensure semantic and motion fidelity (Sketching), while the smaller model refines visual details in low-noise steps (Rendering). Experimental results demonstrate that our method outperforms existing approaches, over 3$\times$ speedup for Wan with nearly no quality loss for VBench, and 2$\times$ speedup for CogVideoX. Our method is introduced as a new direction orthogonal to existing acceleration strategies, offering a practical solution for scalable video generation.

Yuanxin Wei, Lansong Diao, Bujiao Chen et al.

Leveraging the Transformer architecture and the diffusion process, video DiT models have emerged as a dominant approach for high-quality video generation. However, their multi-step iterative denoising process incurs high computational cost and inference latency. Caching, a widely adopted optimization method in DiT models, leverages the redundancy in the diffusion process to skip computations in different granularities (e.g., step, cfg, block). Nevertheless, existing caching methods are limited to single-granularity strategies, struggling to balance generation quality and inference speed in a flexible manner. In this work, we propose MixCache, a training-free caching-based framework for efficient video DiT inference. It first distinguishes the interference and boundary between different caching strategies, and then introduces a context-aware cache triggering strategy to determine when caching should be enabled, along with an adaptive hybrid cache decision strategy for dynamically selecting the optimal caching granularity. Extensive experiments on diverse models demonstrate that, MixCache can significantly accelerate video generation (e.g., 1.94$\times$ speedup on Wan 14B, 1.97$\times$ speedup on HunyuanVideo) while delivering both superior generation quality and inference efficiency compared to baseline methods.