Searcharxiv⌕ Search

arXiv subjects

Entong Li

Publications and source records attributed to Entong Li.

6 recordsLinked to original sources

Adaptive Scheduling Pipeline for Multi-Instance Asynchronous Reinforcement Learning

We present the Adaptive Scheduling Pipeline (ASP), an offline rollout scheduler for multi-instance asynchronous reinforcement learning (RL) of large language models (LLMs). ASP jointly reduces three sources of rollout underutilization - padding bubbles from intra-minibatch length heterogeneity, skewness bubbles from long-tail outputs, and sorting bubbles from fluctuations in training-queue occupancy induced by output-length-sorted dispatch - through a single offline preprocessing pass. This enables efficient multi-instance rollout that overcomes the early scaling plateau of single-instance inference while reducing fluctuations in the supply of generated samples to the trainer. We integrate the scheduler into an asynchronous architecture that decouples rollout from training while keeping policy lag observable and bounded. Within our asynchronous architecture, ASP's scheduling reduces mean iteration time by 10.9%. The complete system improves end-to-end throughput by up to 3.22 times and reduces iteration time by up to 76.8% relative to MindSpeed-RL, while reducing iteration time by up to 62.5% relative to verl. It also scales more robustly with additional generation resources than these baselines and shows no visible convergence degradation in the measured bounded-lag regime. ASP has since been used in an internal telecom-domain RL post-training workflow, reducing the reported training period from more than three months to two months and producing a domain-enhanced model for knowledge answering and tool calling.

cs.DC↗

SR-OPSD: Self-Referenced On-Policy Self-Distillation

On-policy self-distillation (OPSD) converts feedback into dense token-level supervision on student-generated trajectories, complementing reinforcement learning with sparse outcome rewards. Its self-teacher, derived from the student's current or exponentially averaged parameters and conditioned on additional context, evolves alongside the student and its rollout context distribution. The benefit of modifying this moving target depends on how target--student probability mismatches translate into updates. We propose \emph{Self-Referenced On-Policy Self-Distillation (SR-OPSD)}, which constructs a normalized geometric target from the self-teacher and a frozen initial policy, then minimizes the forward Rényi divergence from this target to the student. The interpolation coefficient controls the self-teacher's contribution, while the Rényi order controls the power weighting of target-to-student probability ratios in the gradient. For fixed contexts and target components, we establish a conditional variational characterization and derive the exact token-logit gradient, revealing how anchoring and projection jointly shape the effective update target. Experiments across scientific reasoning, tool use, mathematical reasoning, and code generation demonstrate strong performance across multiple model families and scales. Ablations further show that reference anchoring can improve or degrade performance depending on the projection objective, supporting the joint design of target construction and projection geometry.

cs.LG↗

WindVE: Collaborative CPU-NPU Vector Embedding

Retrieval-Augmented Generation is a technology that enhances large language models by integrating information retrieval. In the industry, inference services based on LLMs are highly sensitive to cost-performance ratio, prompting the need for improving hardware resource utilization in the inference service. Specifically, vector embedding and retrieval processes take up to 20% of the total latency. Therefore, optimizing the utilization of computational resources in vector embeddings is crucial for enhancing the cost-performance ratio of inference processes, which in turn boosts their product competitiveness.In this paper, we analyze the deployment costs of vector embedding technology in inference services, propose a theoretical formula, and determine through the mathematical expression that increasing the capacity to process concurrent queries is the key to reducing the deployment costs of vector embeddings. Therefore, in this paper, we focus on improving the product's capability to process concurrent queries. To optimize concurrency without sacrificing performance, we have designed a queue manager that adeptly offloads CPU peak queries. This manager utilizes a linear regression model to ascertain the optimal queue depths, a critical parameter that significantly influences the efficacy of the system. We further develop a system named WindVE that uses a CPU-NPU heterogeneous architecture to offload peak concurrent queries, which leverages the performance differences between the two processors to effectively manage traffic surges. Through experiments, we compare WindVE to the state-of-the-art vector embedding framework FlagEmbedding, and achieve a concurrency level up to 22.3% higher than the scheme without offloading.

cs.DC↗

SLO-Aware Scheduling for Large Language Model Inferences

Large language models (LLMs) have revolutionized applications such as code completion, chatbots, and online classification. To elevate user experiences, service level objectives (SLOs) serve as crucial benchmarks for assessing inference services capabilities. In practice, an inference service processes multiple types of tasks, each with its own distinct SLO. To ensure satisfactory user experiences, each request's distinct SLOs should be considered in scheduling. However, existing designs lack this consideration, leading to insufficient hardware utility and suboptimal performance. This paper analyzes scenarios to process tasks with varying SLOs, and introduces a simulated annealing-based scheduler to decide request priority sequence based on a request's SLO, input lengths, and possible output lengths. As the first specialized scheduler for multi-SLO scenarios, this work improves SLO attainment by up to 5x and reduces average latency by 31.6% on Python-Code-23k-ShareGPT and ShareGPT_Vicuna_unfiltered datasets, compared to current state-of-the-art framework vLLM and a new framework LMDeploy.

cs.DC↗

High-Throughput LLM inference on Heterogeneous Clusters

Nowadays, many companies possess various types of AI accelerators, forming heterogeneous clusters. Efficiently leveraging these clusters for high-throughput large language model (LLM) inference services can significantly reduce costs and expedite task processing. However, LLM inference on heterogeneous clusters presents two main challenges. Firstly, different deployment configurations can result in vastly different performance. The number of possible configurations is large, and evaluating the effectiveness of a specific setup is complex. Thus, finding an optimal configuration is not an easy task. Secondly, LLM inference instances within a heterogeneous cluster possess varying processing capacities, leading to different processing speeds for handling inference requests. Evaluating these capacities and designing a request scheduling algorithm that fully maximizes the potential of each instance is challenging. In this paper, we propose a high-throughput inference service system on heterogeneous clusters. First, the deployment configuration is optimized by modeling the resource amount and expected throughput and using the exhaustive search method. Second, a novel mechanism is proposed to schedule requests among instances, which fully considers the different processing capabilities of various instances. Extensive experiments show that the proposed scheduler improves throughput by 122.5% and 33.6% on two heterogeneous clusters, respectively.

cs.DC↗

LocMoE: A Low-Overhead MoE for Large Language Model Training

The Mixtures-of-Experts (MoE) model is a widespread distributed and integrated learning method for large language models (LLM), which is favored due to its ability to sparsify and expand models efficiently. However, the performance of MoE is limited by load imbalance and high latency of All-to-All communication, along with relatively redundant computation owing to large expert capacity. Load imbalance may result from existing routing policies that consistently tend to select certain experts. The frequent inter-node communication in the All-to-All procedure also significantly prolongs the training time. To alleviate the above performance problems, we propose a novel routing strategy that combines load balance and locality by converting partial inter-node communication to that of intra-node. Notably, we elucidate that there is a minimum threshold for expert capacity, calculated through the maximal angular deviation between the gating weights of the experts and the assigned tokens. We port these modifications on the PanGu-Sigma model based on the MindSpore framework with multi-level routing and conduct experiments on Ascend clusters. The experiment results demonstrate that the proposed LocMoE reduces training time per epoch by 12.68% to 22.24% compared to classical routers, such as hash router and switch router, without impacting the model accuracy.

cs.LG↗