Pan Li works on AI infrastructure. He holds an MSc in Software Engineering from USTC, is an Apache Fory PMC member and contributor to Ray and vLLM, has published 2 papers on LLM inference, and holds 5 patents. Previously at Bilibili, Tencent, ByteDance, and Xiaohongshu.李攀专注于 AI 基础设施。毕业于中国科学技术大学软件工程硕士，他是 Apache Fory 的 PMC 成员，也是 Ray 与 vLLM 的贡献者，发表 2 篇大模型推理论文，拥有 5 项专利。曾在哔哩哔哩、腾讯、字节跳动、小红书工作。
Responsible for large-model inference infrastructure — built video-generation model serving (Self-Forcing, Wan2.2, LTX2.3) with a layerwise-offload engine for consumer GPUs and vLLM extensions for non-standard models. Deployed DeepSeek V4 with PD separation and KV-cache pooling, cutting end-to-end latency, and built a Ray data platform for feature engineering.负责大模型推理基础设施——主导多模态视频生成模型服务（Self-Forcing、Wan2.2、LTX2.3），自研运行于消费级 GPU 的 layerwise offload 引擎，并扩展 vLLM 支持非标模型。部署 DeepSeek V4 实现 PD 分离与 KV Cache 池化，压缩端到端延迟，并基于 Ray 搭建特征工程数据平台。
Responsible for content-understanding infrastructure — built a Ray-based offline+near-line multimodal compute platform for note features and search embeddings, and prototyped a RAG system with vector, keyword, and knowledge-graph recall for technical-document retrieval.负责内容理解基础设施——基于 Ray 搭建离线+近线多模态计算平台用于笔记特征与搜索 embedding，并预研 RAG 系统，融合向量、关键词与知识图谱三路召回，用于技术文档检索。
Responsible for distributed search infrastructure for WeChat Search — built the recall-evaluation system over ClickHouse/Hive covering accuracy, recall, and GSB, the service-mesh search gateway with CGI audit and chunked transfer, a rule engine for multi-role access control, and an Official Account / Mini Program search-index dashboard, all over 100M-scale sharded data.负责微信搜一搜分布式搜索基础设施——搭建基于 ClickHouse/Hive 的召回评估系统，覆盖准确、召回与 GSB 评估；服务网格搜索网关，含 CGI 审计与分片传输；多角色权限规则引擎；以及公众号/小程序搜索指数看板，整体支撑亿级分库分表数据。
Responsible for the commercialization CRM platform's intelligent toolchain — built the Ocean Engine knowledge center powering ad-serving and brand-marketing scenarios.负责商业化 CRM 平台智能工具链——搭建巨量引擎知识中心，支撑广告投放与品牌营销场景。
Joined 2023.09; entered the Apache Incubator 2023.12; graduated to a Top-Level Project 2025.07. Drove multiple version releases and community building, mentored contributors and students through OSPP, and helped grow the committer base.2023.09 加入 Fory 社区；2023.12 进入 Apache Incubator；2025.07 与 Founder 一起推动 Apache Fory 毕业，成为顶级项目。主导多次版本发布与社区建设，通过 OSPP 指导学生参与贡献，帮助发展 Committer 队伍。
Partitions attention heads into chunks and overlaps communication with computation via dual CUDA streams — orthogonal to FlashAttention/SDPA, no kernel changes, numerically equivalent.将注意力头划分为多个 chunk，通过双 CUDA 流实现通信-计算重叠——与 FlashAttention/SDPA 正交兼容，无需修改 kernel，保持数值等价。
Adapts Self-Forcing to sequence-parallel inference with a causal RoPE variant for localized computation and less cross-rank communication.将 Self-Forcing 适配至序列并行推理，通过因果 RoPE 变体支持局部计算并降低跨 rank 通信。
torch.compile full-graph compilation to eliminate graph breaks in Self-Forcing inference and cut Python/dispatch overhead.使用 torch.compile 整图编译，消除 Self-Forcing 推理中的 Graph Break，降低 Python 解释与调度开销。
Self-Forcing's causal-AR paradigm and adding full sequence-parallel support with a global-time-index 3D positional encoding.Self-Forcing 的因果自回归范式，并补充完整的序列并行支持与基于全局时间索引的 3D 位置编码。