arXiv subjects

Rongqian Zhao

Publications and source records attributed to Rongqian Zhao.

4 recordsLinked to original sources

arXiv2025

Expert-Token Resonance MoE: Bidirectional Routing with Efficiency Affinity-Driven Active Selection

Mixture-of-Experts (MoE) architectures enable efficient scaling of large language models by activating only a subset of parameters per input. However, existing MoE models suffer from two critical limitations: (1) inefficient token-to-expert routing that causes excessive communication overhead, and (2) expert homogenization that leads to redundant computations. Current approaches address these challenges separately, failing to achieve simultaneous improvements in both training efficiency and model performance. We present Expert-Token Resonance (ETR), a theoretically-grounded bidirectional routing mechanism that fundamentally reimagines expert-token interactions in MoE architectures. Our key insight is that optimal routing requires adaptive coordination between token-choice routing (TCR) during early training phases and expert-choice routing (ECR) in later stages. We prove that this dynamic approach maximizes training success rate (the probability of correct token-expert assignments) while reducing the expert capacity lower bound by up to 40%. ETR incorporates three technical innovations: (1) an affinity-based routing architecture using Grouped Average Pooling (GrAP) that reduces computational complexity from O(d^2) to O(d^2/D) while maintaining orthogonality to prevent expert homogenization; (2) a bidirectional selection mechanism that enables both tokens and experts to actively participate in the routing process based on cosine similarity scores; and (3) an adaptive capacity strategy that dynamically adjusts expert bounds based on training progress, eliminating communication bubbles in All-to-All operations. Extensive experiments on Ascend NPU clusters demonstrate that ETR achieves 5.4%-46.6% improvements in end-to-end training efficiency compared to baseline MoE implementations, with 9.7%-14.5% performance gains across GDAD, GPQA, HumanEval, and TeleQnA benchmarks.

cs.CL↗

arXiv2024

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↗

arXiv2024

WindGP: Efficient Graph Partitioning on Heterogenous Machines

Graph Partitioning is widely used in many real-world applications such as fraud detection and social network analysis, in order to enable the distributed graph computing on large graphs. However, existing works fail to balance the computation cost and communication cost on machines with different power (including computing capability, network bandwidth and memory size), as they only consider replication factor and neglect the difference of machines in realistic data centers. In this paper, we propose a general graph partitioning algorithm WindGP, which can support fast and high-quality edge partitioning on heterogeneous machines. WindGP designs novel preprocessing techniques to simplify the metric and balance the computation cost according to the characteristics of graphs and machines. Also, best-first search is proposed instead of BFS and DFS, in order to generate clusters with high cohesion. Furthermore, WindGP adaptively tunes the partition results by sophisticated local search methods. Extensive experiments show that WindGP outperforms all state-of-the-art partition methods by 1.35 - 27 times on both dense and sparse distributed graph algorithms, and has good scalability with graph size and machine number.

cs.DC↗

arXiv2023

RaftGP: Random Fast Graph Partitioning

Graph partitioning (GP), a.k.a. community detection, is a classic problem that divides the node set of a graph into densely-connected blocks. Following prior work on the IEEE HPEC Graph Challenge benchmark and recent advances in graph machine learning, we propose a novel RAndom FasT Graph Partitioning (RaftGP) method based on an efficient graph embedding scheme. It uses the Gaussian random projection to extract community-preserving features from classic GP objectives. These features are fed into a graph neural network (GNN) to derive low-dimensional node embeddings. Surprisingly, our experiments demonstrate that a randomly initialized GNN even without training is enough for RaftGP to derive informative community-preserving embeddings and support high-quality GP. To enable the derived embeddings to tackle GP, we introduce a hierarchical model selection algorithm that simultaneously determines the number of blocks and the corresponding GP result. We evaluate RaftGP on the Graph Challenge benchmark and compare the performance with five baselines, where our method can achieve a better trade-off between quality and efficiency. In particular, compared to the baseline algorithm of the IEEE HPEC Graph Challenge, our method is 6.68x -- 23.9x faster on graphs with 1E3 -- 5E4 nodes and at least 64.5x faster on larger (1E5 node) graphs on which the baseline takes more than 1E4 seconds. Our method achieves better accuracy on all test cases. We also develop a new graph generator to address some limitations of the original generator in the benchmark.

cs.SI↗