Juntao Zhao

Juntao Zhao, Borui Wan, Yanghua Peng et al.

A number of production deep learning clusters have attempted to explore inference hardware for DNN training, at the off-peak serving hours with many inference GPUs idling. Conducting DNN training with a combination of heterogeneous training and inference GPUs, known as hybrid device training, presents considerable challenges due to disparities in compute capability and significant differences in memory capacity. We propose QSync, a training system that enables efficient synchronous data-parallel DNN training over hybrid devices by strategically exploiting quantized operators. According to each device's available resource capacity, QSync selects a quantization-minimized setting for operators in the distributed DNN training graph, minimizing model accuracy degradation but keeping the training efficiency brought by quantization. We carefully design a predictor with a bi-directional mixed-precision indicator to reflect the sensitivity of DNN layers on fixed-point and floating-point low-precision operators, a replayer with a neighborhood-aware cost mapper to accurately estimate the latency of distributed hybrid mixed-precision training, and then an allocator that efficiently synchronizes workers with minimized model accuracy degradation. QSync bridges the computational graph on PyTorch to an optimized backend for quantization kernel performance and flexible support for various GPU architectures. Extensive experiments show that QSync's predictor can accurately simulate distributed mixed-precision training with <5% error, with a consistent 0.27-1.03% accuracy improvement over the from-scratch training tasks compared to uniform precision.

Juntao Zhao, Wenhao Lu, Sheng Wang et al.

Quantization is widely adopted to accelerate inference and reduce memory consumption in large language models (LLMs). While activation-weight joint quantization enables efficient low-precision decoding, it suffers from substantial performance degradation on multi-step reasoning tasks. We propose QSpec, a novel quantization paradigm that decouples efficiency from quality by integrating two complementary schemes via speculative decoding: low-precision joint quantization for fast drafting and high-precision weight-only quantization for accurate verification. QSpec reuses both weights and KV cache across stages, enabling near-zero-cost switching without retraining or auxiliary models. Compared to high-precision baselines, QSpec achieves up to 1.64x speedup without quality degradation, and outperforms state-of-the-art speculative decoding methods by up to 1.55x in batched settings. Furthermore, QSpec supports plug-and-play deployment and generalizes well across model scales, quantization methods, and workloads. These properties make QSpec a practical and scalable solution for high-fidelity quantized LLM serving under memory-constrained scenarios. Our code is available at https://github.com/hku-netexplo-lab/QSpec.

Juntao Zhao, Borui Wan, Yanghua Peng et al.

Recent breakthroughs in Large-scale language models (LLMs) have demonstrated impressive performance on various tasks. The immense sizes of LLMs have led to very high resource demand and cost for running the models. Though the models are largely served using uniform high-caliber GPUs nowadays, utilizing a heterogeneous cluster with a mix of available high- and low-capacity GPUs can potentially substantially reduce the serving cost. There is a lack of designs to support efficient LLM serving using a heterogeneous cluster, while the current solutions focus on model partition and uniform compression among homogeneous devices. This paper proposes LLM-PQ, a system that advocates adaptive model quantization and phase-aware partition to improve LLM serving efficiency on heterogeneous GPU clusters. We carefully decide on mixed-precision model quantization together with phase-aware model partition and micro-batch sizing in distributed LLM serving with an efficient algorithm, to greatly enhance inference throughput while fulfilling user-specified model quality targets. Extensive experiments on production inference workloads in 11 different clusters demonstrate that LLM-PQ achieves up to 2.88x (2.26x on average) throughput improvement in inference, showing great advantages over state-of-the-art works.

Wan Borui, Zhao Juntao, Jiang Chenyu et al.

Recent breakthroughs in large Language Models (LLMs) have enabled various generative tasks on a single model. Real-world services (e.g., OpenAI's ChatGPT [27]) powered by an LLM often concurrently support latency-critical requests for interactive applications (e.g., question-answering systems, referred to as real-time or RT requests) and throughput-oriented requests for back-of-house processing (e.g., documents batch processing [28], referred to best-effort or BE requests), with complex hybrid inference workloads to the underlying model. State-of-the-art (SOTA) LLM serving systems dedicate machines to each type of request, towards either low inference latency or high serving throughput, respectively. This practice simplifies request scheduling and management but suffers from poor resource utilization. We propose BROS, a hybrid LLM serving system that aims to collocate RT/BE requests, meeting RT requests' latency requirements while maintaining BE requests' throughput. BROS formulates the problem of hybrid RT/BE request scheduling and solves it with a dynamic priority-based algorithm. BROS designs a bidirectional KV cache management mechanism, allowing RT requests to share KV memory with BE requests to remove the scheduling restrictions caused by insufficient KV memory and improve utilization. Extensive experiments validate that BROS achieves a good trade-off when serving hybrid RT and BE requests. It significantly reduces the latency of RT requests (up to 74.20%), improving their fine-grained service level objectives (SLOs) attainments (up to 36.38x), with negligible throughput reduction for BE requests, showing significant advantages over SOTA systems like vLLM and TGI.

Juntao Zhao, Jiuru Li, Chuan Wu

Utilizing CPUs to serve large language models (LLMs) is a resource-friendly alternative to GPU serving. Existing CPU-based solutions ignore workload differences between the prefill and the decode phases of LLM inference, applying a static per-NUMA (Non-Uniform Memory Access) node model partition and utilizing vendor libraries for operator-level execution, which is suboptimal. We propose Sandwich, a hardware-centric CPU-based LLM serving engine that uses different execution plans for the prefill and decode phases and optimizes them separately. We evaluate Sandwich across diverse baselines and datasets on five CPU platforms, including x86 with AVX-2 and AVX-512, as well as ARM with NEON. Sandwich achieves an average 2.01x throughput improvement and 90% satisfactory time-to-first-token (TTFT) and time-per-output-token (TPOT) latencies with up to 3.40x lower requirements in single sequence serving, and significant improvement in Goodput in continuous-batching serving. The GEMM kernels generated by Sandwich outperform representative vendor kernels and other dynamic shape solutions, achieving performance comparable to static compilers with three orders of magnitude less kernel tuning costs.

Juntao Zhao, Qi Lu, Wei Jia et al.

Modern frameworks for training large foundation models (LFMs) employ dataloaders in a data-parallel manner, with each loader processing a disjoint subset of training data. Under multisource preprocessing, two fundamental challenges exist. First, due to the quadratic computational complexity of the attention operator, the non-uniform sample distribution over data-parallel ranks leads to significant workload imbalance among dataloaders, degrading the training efficiency. Second, supporting diverse data sources requires per-dataset file access states that are redundantly replicated across parallel loaders, consuming excessive memory. This also hinders dynamic data mixing (e.g., curriculum learning) and causes redundant access/memory overhead in hybrid parallelism. We present Omniload, an industrial-grade distributed data loading architecture for LFMs, with four innovations: (1) Disaggregated data preprocessing via role-specific actors (Source Loaders/Data Constructors) to eliminate source and parallelism redundant data access and ensure multisource scalability. (2) Centralized and declarative data plane for elastic multisource orchestration, such as long-short context, multimodality, and curriculum learning. (3) Multi-level auto-partitioning and scaling mechanism for source loaders under heterogeneous preprocessing costs. (4) Shadow loaders with differential checkpointing for fault recovery without workflow interruption. Deployed on production clusters scaling to multi-thousand GPUs, Omniload achieves: (1) 4.5x end-to-end training throughput improvement, (2) 13.5x reduction in CPU memory usage.