I am a Distributed Training Engineer at Beijing Academy of Artificial Intelligence (BAAI). Previously, I was an HPC Engineer at Baidu PaddlePaddle (2023–2024). I obtained my M.E. in Computer Technology from Nankai University in 2023.

My work focuses on large-scale distributed training systems for frontier LLMs. I specialize in hybrid-parallelism strategies, high-performance kernel engineering (CUDA/Triton), and training infrastructure for Mixture-of-Experts and long-context multimodal models. I have been deeply involved in training systems for DeepSeek V3/V4 and BAAI EMU3.5.

🔥 News

  • 2025.04:   DualPipeV open-sourced and adopted by FlagScale for MoE pipeline parallelism.
  • 2025.02:   Ring-Flex-Attention validated at 64K seq length for 32B multimodal models (EMU3.5).
  • 2024.12:   MTP reference implementation cited by Alibaba PAI and upstream Megatron-LM.
  • 2024.11:   Joined BAAI as Distributed Training Engineer.
  • 2024.04:   Hybrid-Memcached accepted by IEEE Transactions on Computers (CCF-A).

📖 Educations

  • 2020.09 – 2023.06: M.E. in Computer Technology, Nankai University. Advisor: Prof. Xiaoli Gong.
  • 2016.09 – 2020.06: B.E. in Computer Science and Technology, China University of Mining and Technology.

👨‍💻 Work Experience

Distributed Training Engineer, Beijing Academy of Artificial Intelligence (BAAI)   Nov 2024 – Present

Core contributor to DeepSeek V3/V4 and EMU3.5 training infrastructure. Leading performance optimization for sparse attention, MoE parallelism, and long-context multimodal training at scale on NVIDIA H100/A800 clusters.

HPC Engineer, Baidu PaddlePaddle   Jul 2023 – Nov 2024

AI Infra engineer for ERNIE LLMs. Designed semi-auto parallelism distributed training, supporting full-chain hybrid parallel strategies (DP/TP/PP/SP) for Pretrain, SFT, and LoRA.

💻 Selected Projects

DeepSeek V3/V4 Performance Optimization & Tuning   Dec 2024 – Present

  • Sparse Attention for Long-Context Training: Led CP tuning for DeepSeek V4’s Hybrid Attention (CSA/HCA/Sliding-Window). Fused fragmented micro-ops into unified GEMMs and custom Triton kernels; overlapped P2P/All-Gather with computation via multi-stream execution; eliminated CPU-GPU sync bottlenecks; designed end-to-end CP infrastructure with cross-rank halo exchange and causal sparse top-K for 4×/128× compressed attention.
  • Multi-Token Predictor (MTP): Architected the inaugural DeepSeek V3 MTP module with custom pipeline parallelism. This reference design was later cited and adopted by Alibaba PAI and upstream Megatron-LM.
  • Engram Implementation: Engineered AllGather-Based and All-to-All-Based Engram Embedding with CPU state offloading and asynchronous scheduling for large-scale training.
  • DualPipeV: Engineered and open-sourced a pipeline parallelism strategy for MoE, achieving seamless overlap of all-to-all communication with forward/backward computation. Adopted by FlagScale.
  • Auxiliary-Loss-Free Load Balancing: Independently reproduced DeepSeek V3’s Sigmoid Router and Auxiliary-Loss-Free balancing from paper specifications; integrated MLA + Shared-Expert MoE for end-to-end training fidelity.
  • Production Infrastructure: Delivered stable months-long training for DeepSeek V3-style MoE models across scales (16B-A3B and beyond) with fault-tolerant distributed checkpointing.

EMU3.5: Efficient Long-Sequence Multimodal Training   Feb 2025 – Oct 2025

  • Ring-Flex-Attention: Designed a novel fusion of Ring-style Context Parallelism with PyTorch FlexAttention from the ground up. Validated at 64K sequence length for 32B models with GPU multi-streaming and dynamic masked-block bypassing.
  • OmniMask Support: Integrated FlexAttention with TP/SP/PP/CP for Emu’s complex attention patterns (causal, bidirectional, region-specific), replacing 4-D attention masks with per-row block-masks.
  • Hybrid Parallelism: Orchestrated TP8/CP4/SP/DP/PP combinations at multi-node scale.
  • Modality-Aware LoRA: Designed tensor-wise sharded, modality-specific LoRA for parameter-efficient fine-tuning on image tokens.
  • Delivered production-grade training for EMU3.5 (1.5B–32B) across 4K–64K sequence lengths.

NVIDIA TransformerEngine Plugin System   Aug 2025 – Mar 2026

Plugin system officially recognized by NVIDIA; mainline integration in progress.

  • Designed a general plugin-based operator management mechanism for TransformerEngine, enabling multi-chip (NVIDIA, Hygon, Metax, Kunlunxin, Iluvatar, Moore), multi-language (CUDA, Triton), and priority-adaptive operator dispatch.
  • Triton Backend: Reimplemented all core TransformerLayer operators via Triton kernel composition (Generic GEMM, Attention, Adam Optimizer), achieving full Triton coverage for Qwen3 Dense model training.
  • Dynamic Sparse Attention: Built from scratch in Triton, achieving 10% latency reduction vs. Flash Attention by skipping fully-masked regions.
  • Co-optimized Triton GEMM with Sequence Parallel AllGather/AllReduce for long-context throughput gains.

DualPipeV Pipeline Parallelism   Apr 2025 – Jul 2025

  • 7-stage pipeline orchestration with 2-chunk forward/backward micro-batches and precise dependency tracking.
  • Decoupled backward pass into backward-for-inputs and backward-for-weights for finer-grained overlap.
  • Profiling-driven MoE kernel reordering for seamless compute-communication disentanglement.
  • Result: 1.12× throughput vs. Megatron-LM 1F1B baseline on 8×H100 (PP4EP2) with DeepSeek 16B.

ERNIE Semi-Auto Distributed Training   Aug 2023 – Oct 2024

  • Semi-auto parallelism framework design with hybrid parallel strategy support.
  • Full-chain distributed training (DP/TP/PP/SP) and verification for ERNIE LLMs: Pretrain, SFT, LoRA.

📝 Publications

Hybrid-Memcached: A Novel Approach for Memcached Persistence Optimization with Hybrid Memory

IEEE Transactions on Computers (TC), 2024   [CCF-A]

Zhang Jiang*, Xianduo Li* (joint first author), Xiaoli Gong, et al.


ABM-SpConv-SIMD: Accelerating Convolutional Neural Network Inference for Industrial IoT Applications on Edge Devices

IEEE Transactions on Network Science and Engineering (TNSE), 2022   [JCR-Q1]

Xianduo Li, Xiaoli Gong, Dong Wang, Jin Zhang, et al.

🎖 Honors and Awards

  • 2024.03   Member of Ernie-Team, Baidu Proud of 2024
  • 2023.10   Best New-comer Award, Baidu PaddlePaddle
  • 2023.06   Outstanding Master Thesis Award, Nankai University
  • 2022.10   First-class (Top 10%) GongNeng Scholarship, Nankai University
  • 2020.06   Outstanding Student, China University of Mining and Technology

💬 Services

  • Teaching Assistant, Operating System course, Nankai University, Fall 2020.