Training Report of TeleChat3-MoE

• URL: http://arxiv.org/abs/2512.24157v1
• Date: Tue, 30 Dec 2025 11:42:14 GMT
• Title: Training Report of TeleChat3-MoE
• Authors: Xinzhang Liu, Chao Wang, Zhihao Yang, Zhuo Jiang, Xuncheng Zhao, Haoran Wang, Lei Li, Dongdong He, Luobin Liu, Kaizhe Yuan, Han Gao, Zihan Wang, Yitong Yao, Sishi Xiong, Wenmin Deng, Haowei He, Kaidong Yu, Yu Zhao, Ruiyu Fang, Yuhao Jiang, Yingyan Li, Xiaohui Hu, Xi Yu, Jingqi Li, Yanwei Liu, Qingli Li, Xinyu Shi, Junhao Niu, Chengnuo Huang, Yao Xiao, Ruiwen Wang, Fengkai Li, Luwen Pu, Kaipeng Jia, Fubei Yao, Yuyao Huang, Xuewei He, Zhuoru Jiang, Ruiting Song, Rui Xue, Qiyi Xie, Jie Zhang, Zilu Huang, Zhaoxi Zhang, Zhilong Lu, Yanhan Zhang, Yin Zhang, Yanlei Xue, Zhu Yuan, Teng Su, Xin Jiang, Shuangyong Song, Yongxiang Li, Xuelong Li,
• Abstract summary: This technical report mainly presents the underlying training infrastructure that enables reliable and efficient scaling to frontier model sizes.<n>We detail systematic methodologies for operator-level and end-to-end numerical verification accuracy, ensuring consistency across hardware platforms.<n>A systematic parallelization framework, leveraging analytical estimation and integer linear programming, is also proposed to optimize multi-dimensional parallelism configurations.
• Score: 77.94641922160359
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: TeleChat3-MoE is the latest series of TeleChat large language models, featuring a Mixture-of-Experts (MoE) architecture with parameter counts ranging from 105 billion to over one trillion,trained end-to-end on Ascend NPU cluster. This technical report mainly presents the underlying training infrastructure that enables reliable and efficient scaling to frontier model sizes. We detail systematic methodologies for operator-level and end-to-end numerical accuracy verification, ensuring consistency across hardware platforms and distributed parallelism strategies. Furthermore, we introduce a suite of performance optimizations, including interleaved pipeline scheduling, attention-aware data scheduling for long-sequence training,hierarchical and overlapped communication for expert parallelism, and DVM-based operator fusion. A systematic parallelization framework, leveraging analytical estimation and integer linear programming, is also proposed to optimize multi-dimensional parallelism configurations. Additionally, we present methodological approaches to cluster-level optimizations, addressing host- and device-bound bottlenecks during large-scale training tasks. These infrastructure advancements yield significant throughput improvements and near-linear scaling on clusters comprising thousands of devices, providing a robust foundation for large-scale language model development on hardware ecosystems.

Related papers

• Eliminating Multi-GPU Performance Taxes: A Systems Approach to Efficient Distributed LLMs [61.953548065938385]
  We introduce the ''Three Taxes'' (Bulk Synchronous, Inter- Kernel Data Locality, and Kernel Launch Overhead) as an analytical framework.<n>We propose moving beyond the rigid BSP model to address key inefficiencies in distributed GPU execution.<n>We observe a 10-20% speedup in end-to-end latency over BSP-based approaches.
  arXiv  Detail & Related papers  (2025-11-04T01:15:44Z)
• CollaPipe: Adaptive Segment-Optimized Pipeline Parallelism for Collaborative LLM Training in Heterogeneous Edge Networks [57.95170323315603]
  We introduce CollaPipe, a distributed learning framework that integrates collaborative pipeline parallelism with federated aggregation to support self-evolving networks.<n>In CollaPipe, the encoder part is adaptively partitioned into variable-sized segments and deployed across mobile devices for pipeline-parallel training, while the decoder is deployed on edge servers to handle generative tasks.<n>To enhance training efficiency, we formulate a joint optimization problem that adaptively allocates model segments, micro-batches, bandwidth, and transmission power.
  arXiv  Detail & Related papers  (2025-09-24T07:54:01Z)
• Research on Model Parallelism and Data Parallelism Optimization Methods in Large Language Model-Based Recommendation Systems [6.453224262551299]
  Large language models (LLMs) in recommendation systems have become increasingly prominent.<n>This paper systematically investigates two classes of optimization methods-model parallelism and data parallelism.<n> Experiments conducted on a real-world recommendation dataset in a simulated service environment demonstrate that our proposed hybrid parallelism scheme increases training throughput by over 30%.
  arXiv  Detail & Related papers  (2025-06-21T02:37:25Z)
• Scaling Intelligence: Designing Data Centers for Next-Gen Language Models [0.6168147650666682]
  Large Language Models (LLMs), such as GPT-4 with 1.8 trillion parameters, demand a fundamental rethinking of data center architecture.<n>Our work provides a comprehensive co-design framework that jointly explores FLOPS, bandwidth and capacity, multiple network topologies.<n>We quantify the benefits of overlapping compute and communication, leveraging hardware-accelerated collectives, widening the scale-out domain, and increasing memory capacity.
  arXiv  Detail & Related papers  (2025-06-17T22:29:37Z)
• Automatic Operator-level Parallelism Planning for Distributed Deep Learning -- A Mixed-Integer Programming Approach [6.449961842220686]
  We propose a bi-level solution framework balancing optimality with computational efficiency.<n>Our framework achieves comparable or superior performance, reducing computational bubbles by half under the same memory constraints.<n>Such capabilities position our solution as both a valuable research tool for exploring optimal parallelization strategies and a practical industrial solution for large-scale AI deployment.
  arXiv  Detail & Related papers  (2025-03-12T13:00:29Z)
• EPS-MoE: Expert Pipeline Scheduler for Cost-Efficient MoE Inference [49.94169109038806]
  This paper introduces EPS-MoE, a novel expert pipeline scheduler for MoE that surpasses the existing parallelism schemes.<n>Our results demonstrate at most 52.4% improvement in prefill throughput compared to existing parallel inference methods.
  arXiv  Detail & Related papers  (2024-10-16T05:17:49Z)
• Advancing Enterprise Spatio-Temporal Forecasting Applications: Data Mining Meets Instruction Tuning of Language Models For Multi-modal Time Series Analysis in Low-Resource Settings [0.0]
  patio-temporal forecasting is crucial in transportation, logistics, and supply chain management. We propose a dynamic, multi-modal approach that integrates the strengths of traditional forecasting methods and instruction tuning of small language models. Our framework enables on-premises customization with reduced computational and memory demands, while maintaining inference speed and data privacy/security.
  arXiv  Detail & Related papers  (2024-08-24T16:32:58Z)
• ATOM: Asynchronous Training of Massive Models for Deep Learning in a Decentralized Environment [7.916080032572087]
  atom is a resilient distributed training framework designed for asynchronous training of vast models in a decentralized setting. atom aims to accommodate a complete LLM on one host (peer) through seamlessly model swapping and concurrently trains multiple copies across various peers to optimize training throughput. Our experiments using different GPT-3 model configurations reveal that, in scenarios with suboptimal network connections, atom can enhance training efficiency up to $20 times$ when juxtaposed with the state-of-the-art decentralized pipeline parallelism approaches.
  arXiv  Detail & Related papers  (2024-03-15T17:43:43Z)
• Holmes: Towards Distributed Training Across Clusters with Heterogeneous NIC Environment [8.30319294116657]
  Large language models (LLMs) such as GPT-3, OPT, and LLaMA have demonstrated remarkable accuracy in a wide range of tasks. Training these models can incur significant expenses, often requiring tens of thousands of GPU for months of continuous operation. We introduce Holmes, a training framework for LLMs that employs thoughtfully crafted data and model parallelism strategies over the heterogeneous NIC environment.
  arXiv  Detail & Related papers  (2023-12-06T15:27:26Z)
• TeraPipe: Token-Level Pipeline Parallelism for Training Large-Scale Language Models [60.23234205219347]
  TeraPipe is a high-performance token-level pipeline parallel algorithm for synchronous model-parallel training of Transformer-based language models. We show that TeraPipe can speed up the training by 5.0x for the largest GPT-3 model with 175 billion parameters on an AWS cluster.
  arXiv  Detail & Related papers  (2021-02-16T07:34:32Z)
• Edge-assisted Democratized Learning Towards Federated Analytics [67.44078999945722]
  We show the hierarchical learning structure of the proposed edge-assisted democratized learning mechanism, namely Edge-DemLearn. We also validate Edge-DemLearn as a flexible model training mechanism to build a distributed control and aggregation methodology in regions.
  arXiv  Detail & Related papers  (2020-12-01T11:46:03Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.