Yang Jing's  Blog

Jing Yang中文

Senior Engineer

RedNote · Zhihu · GitHub · Email

杨靖EN

高级工程师

小红书 · 知乎 · GitHub · 邮箱

About

From LLM algorithms to infra co-design — Huawei Senior Engineer focused on LLM training & inference acceleration on Ascend, system–algorithm co-design, and landing agents in real business. NUDT M.S. in Software Engineering.

Helped build and deliver 10,000+ Ascend NPU clusters; supports Day 0 adaptation for DeepSeek (V3 → 3.2 → V4) and GLM (GLM-4 → GLM-5.2); shipped AscendBot (department-wide adoption); led CPT & post-training (DPO · GRPO) for China Merchants Bank and PICC.

Experience

From Baidu and IDEA NLP internships to Huawei Ascend infra and agents — a stack built project by project in the LLM era, with plenty of lessons learned the hard way.

  • 06/2023 – now · Senior Engineer · Huawei Infra — Ascend training and inference tuning (CLOS · HCCL · Megatron-LM · MindSpeed · vLLM-Ascend · VeRL · CPT / SFT · GRPO / DPO); production-grade agent systems and harnesses (Evaluation · Harness · DSPy · LangGraph · RAG · MCP · Agent · Skill · Vibe Coding)
  • 11/2022 – 03/2023 · NLP Research & Engineering Intern · Baidu NLP — Unified information extraction and knowledge graph construction (UIE · USM · Knowledge Graph · PaddlePaddle); contributed to UTC universal text classification — 1st on both ZeroCLUE & FewCLUE, open-sourced in PaddleNLP
  • 04/2022 – 08/2022 · NLP Research Intern · IDEA / CCNL — BERT / T5 pre-training, multi-task instruction tuning, and information extraction (Randeng-T5 · UBERT · UniEX); continued training of Randeng beat open-source UIE — full-data F1 +1–3%, few-shot F1 +5–10%

Skills & Experience

7 years in AI, riding the LLM wave — full-stack from algorithms to infra; plenty of projects and hard-won lessons, from 10K-card cluster buildout and train/infer tuning all the way to agents.

LLM Agent

Builds production-grade agent systems and harnesses for real business scenarios (LangGraph · MCP · Skill · RAG); Langfuse for trajectory tracing and failure localization, DSPy + MLflow for evaluation, closing the data → training → evaluation loop. Shipped Ascend Bot — agent-driven experiential auto-tuning and troubleshooting — adopted department-wide and spun into its own business group; heavy daily Vibe Coding (Claude Code / Codex / Cursor).

Projects: Ascend Bot · Ascend Optim Agent · RAG diagnostics assistant · Agent evaluation platform · ClaudePad · skill-adapter · Synapse · w3slides

LLM Algorithm

CPT and post-training on 1,000–10,000+ card clusters (Megatron-LM · LLaMA-Factory), multi-round SFT · DPO · GRPO alignment. Data-Juicer / DataRover for cleaning and mixture experiments; multi-agent pipelines for post-training data synthesis; OPD distillation over agent evaluation and trajectory pipelines. PICC post-training: domain benchmarks from real business scenarios, distillation-based synthetic data, multi-round SFT iteration; RAG in production, 20K+ daily calls.

Projects: CMB YiZhao continued training (YiZhao-12B) · PICC post-training · internal Metis model

LLM Infra (Ascend)

  • Training: MindSpeed-LLM/RL · VeRL · Megatron-LM · HCCL · 910B / 910C / 950 framework tuning, communication and MFU gains (ChatGLM 32B/130B MFU 0.85×A800 equivalent), RL efficiency optimization.
  • Inference: Day 0 adaptation and delivery for DeepSeek (V3 → 3.2 → V4) (A2 single-node latency ~90ms → <20ms) and Zhipu GLM (GLM-4 → GLM-5.2) (200K quant W8A8 → W4A8) with partners; MTP / MLA · FusedOp · Quant · PD / AFD / EPD · DeepEP · multi-tenant load balancing · vLLM-Ascend scenario-specific tuning.

Projects: DeepSeek / GLM Ascend Day 0 adaptation · 10K-card RL framework optimization · vLLM-Ascend · MindSpeed-RL

IDC Infra

10,000+ card cluster buildout acceptance, stress testing, and long-run training stability; networking evolution (CLOS → ROCE → HPN 2.0 → ZCube · HCCL / UB / RDMA · A3 SuperPod · UB-Mesh · MoonCake). On-call for slow cards, dropped cards, and optical transceiver flaps.

Projects: 10K-card Ascend cluster buildout and stability

Publications

A handful of papers — none of them without excellent co-authors.

关于

从 LLM Algorithm 走向 Infra Co-design 的工程师——华为 高级工程师，聚焦 Ascend 上的 LLM 训练与推理加速、系统–算法协同设计，并探索将 Agent 落地到具体业务。NUDT 软件工程硕士。

参与万卡级（10,000+）Ascend NPU 集群建设与交付；保障 DeepSeek（V3 → 3.2 → V4）、GLM（GLM-4 → GLM-5.2） Day 0 适配与交付；主导 AscendBot 调优助手（部门级采用并孵化独立业务小组）；服务招行、人保等企业客户 CPT / 后训练（DPO · GRPO）。

工作经历

从百度、IDEA 的 NLP 实习，到华为 Ascend Infra 训推加速与 Agent 落地——LLM 时代技术栈是项目里一点点 debug 出来的，有着丰富踩坑经验。

  • 06/2023 – 至今 · 高级工程师 · 华为 Infra — Ascend 训练与推理调优加速 (CLOS · HCCL · Megatron-LM · MindSpeed · vLLM-Ascend · VeRL · CPT / SFT · GRPO / DPO)； 生产级 Agent 系统与 Harness 落地 (Evaluation · Harness · DSPy · LangGraph · RAG · MCP · Agent · Skill · Vibe Coding)
  • 11/2022 – 03/2023 · NLP 研究与工程实习生 · 百度 NLP — 统一信息抽取与知识图谱构建 (UIE · USM · Knowledge Graph · PaddlePaddle)；参与 UTC 通用文本分类——ZeroCLUE / FewCLUE 双料第一，开源至 PaddleNLP
  • 04/2022 – 08/2022 · NLP 研究实习生 · IDEA / CCNL — BERT / T5 预训练、多任务指令微调与信息抽取 (Randeng-T5 · UBERT · UniEX)；Randeng 继续训练后超越开源 UIE——全量 F1 +1–3%、Few-Shot F1 +5–10%

技能经验

7 年 AI 从业，赶上 LLM 浪潮——技术栈从算法到 Infra 全链路贯通；项目换了不少，坑也踩得够多，从万卡集群建设交付、模型训推调优，一路修到 Agent。

LLM Agent

面向真实业务场景构建生产级 Agent 系统与 Harness (LangGraph · MCP · Skill · RAG)；以 Langfuse 做 trajectory 追踪与出错定位，结合 DSPy + MLflow 搭建评测平台，形成「数据—训练—评测」闭环。开发 Ascend Bot，基于 Agent 进行经验化自动调优和问题处理，获部门级采用并孵化独立业务小组；日常重度 Vibe Coding（Claude Code / Codex / Cursor），积累了大量手感与心得。

项目： Ascend Bot · Ascend Optim Agent · RAG 诊断助手 · Agent 评测平台 · ClaudePad · skill-adapter · Synapse · w3slides

LLM Algorithm

负责千卡～万卡集群 CPT 与后训练 (Megatron-LM · LLaMA-Factory)，多轮 SFT · DPO · GRPO 对齐迭代。Data-Juicer / DataRover 做清洗与配比，multi-agent 流水线合成后训练数据，基于 Agent 评测与轨迹链路开展 OPD 蒸馏。人保后训练：从真实业务构造领域评测集，模型蒸馏合成训练数据，多轮 SFT 迭代调优；RAG 生产化上线，日均调用 2W+。

项目： 招商银行.一招模型.二次续训练微调（YiZhao-12B） · 中国人保.宸灵模型后训练 · 内部 Metis 模型

LLM Infra（Ascend）

  • 训练： MindSpeed-LLM/RL · VeRL · Megatron-LM · HCCL · 910B / 910C / 950 框架调优，通信与 MFU 提升（ChatGLM 32B/130B MFU 达 0.85×A800 等效），RL 效率优化。
  • 推理： 与 DeepSeek 等合作 DeepSeek（V3 → 3.2 → V4）（A2 单机延迟 ~90ms → <20ms）以及智谱 GLM（GLM-4 → GLM-5.2）Day 0 适配（200K 量化 W8A8 → W4A8）与交付； MTP / MLA · FusedOp · Quant · PD / AFD / EPD · DeepEP · 多租负载均衡 · vLLM-Ascend 等场景化调优。

项目： DeepSeek / GLM Ascend Day 0 适配 · 万卡 RL 框架优化 · vLLM-Ascend · MindSpeed-RL

IDC Infra

万卡级集群建设验收、压测与长稳训练保障；组网演进 (CLOS → ROCE → HPN 2.0 → ZCube · HCCL / UB / RDMA · A3 SuperPod · UB-Mesh · MoonCake)。 线上处理快慢卡、掉卡、光模块闪断等问题。

项目： 万卡级 Ascend 集群建设验收与稳定性保障

论文发表

发过几篇文章，在 ChatGPT 出现前，开始探索生成式统一信息抽取以及 MOE 等模型结构，每一篇都离不开优秀的合作者。