哔哩哔哩AI Infra模型研发实习生（训练方向）
实习兼职AI类地点：上海状态：招聘♡ 收藏
工作描述
工作职责: 1. 分布式训练底座 (Training Infra) 框架优化： 负责维护和优化基于 Megatron-LM, FSDP, VeRL的分布式训练框架，通过多维并行策略提高训练吞吐。 算子优化： 参与多模态大模型训练核心算子的设计与优化，包括 Attention、MoE、算子融合等方向，持续提升模型训练效率和硬件利用率。 通信优化： 深入优化 H/NCCL通信库，解决 RDMA/RoCE 网络下的通信瓶颈，提升多机多卡并行效率（DP/PP/TP/CP/EP）。 稳定性保障： 构建自动容错与快速恢复系统（Checkpoint 优化、故障自动检测与接续），确保千卡集群在数月跨度的训练中保持极高可用性。 2. 推理加速与工程化 (Inference Infra) 高性能引擎： 负责基于 vLLM、TensorRT-LLM、SGLang、Triton Inferen…
登录查看完整工作描述
微信扫码，1秒登录
包括英文材料
Megatron+
https://www.youtube.com/watch?v=hc0u4avAkuM
FSDP+
https://docs.pytorch.org/tutorials/intermediate/FSDP_tutorial.html
In DistributedDataParallel (DDP) training, each rank owns a model replica and processes a batch of data, finally it uses all-reduce to sync gradients across ranks.
https://www.youtube.com/watch?v=PjEwLgyzuzQ
FSDP provides a comprehensive framework for large model training in PyTorch.
大模型+
https://www.youtube.com/watch?v=xZDB1naRUlk
You will build projects with LLMs that will enable you to create dynamic interfaces, interact with vast amounts of text data, and even empower LLMs with the capability to browse the internet for research papers.
https://www.youtube.com/watch?v=zjkBMFhNj_g
NCCL+
https://developer.nvidia.com/nccl
The NVIDIA Collective Communication Library (NCCL) implements multi-GPU and multi-node communication primitives optimized for NVIDIA GPUs and networking.
高可用+
https://redis.io/blog/high-availability-architecture/
A high available architecture is when there are a number of different components, modules, or services that work together to maintain optimal performance, irrespective of peak-time loads.
https://www.ibm.com/think/topics/high-availability
High availability (HA) is a term that refers to a system’s ability to be accessible and reliable close to 100% of the time.
vLLM+
https://www.newline.co/@zaoyang/ultimate-guide-to-vllm--aad8b65d
vLLM is a framework designed to make large language models faster, more efficient, and better suited for production environments.
https://www.youtube.com/watch?v=Ju2FrqIrdx0
vLLM is a cutting-edge serving engine designed for large language models (LLMs), offering unparalleled performance and efficiency for AI-driven applications.
TensorRT+
https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/quick-start-guide.html
This TensorRT Quick Start Guide is a starting point for developers who want to try out the TensorRT SDK; specifically, it demonstrates how to quickly construct an application to run inference on a TensorRT engine.
SGLang+
[英文] Install SGLang
https://docs.sglang.ai/get_started/install.html
SGLang is a fast serving framework for large language models and vision language models.
https://github.com/sgl-project/sgl-learning-materials
Triton Inference Server+
https://docs.nvidia.com/deeplearning/triton-inference-server/user-guide/docs/index.html
Triton Inference Server is an open source inference serving software that streamlines AI inferencing.
还有更多 •••
相关职位
社招3年以上
● 扎实的编程功底，在以下至少一个方向有深入理解或实践经验：云计算、多模态理解、安全、Agent、智能办公等； ● 自驱力强，持续学习的能力、不断追求卓越； ● 有较强的问题分析和定位能力，能剥丝抽茧
更新于 2026-09-29北京|杭州|上海
社招3年以上
● 熟悉 LLM 与 Agent 的机制和原理，具备丰富的理论和工程实践经验； ● 熟悉业界代表性 Harness 项目，具备优秀的抽象和系统设计能力； ● 良好的跨团队协作能力，扎实的编程功底； ●
更新于 2026-10-16北京|杭州|上海
社招3年以上
● 卓越的代码能力、系统设计能力和工程优化能力； ● 丰富的大规模分布式系统研发经验； ● 积极主动，负责到底，快速迭代和持续演进； ● 优秀的合作协同意识，乐于探索未知领域。 工作职责 ● 构建面
更新于 2026-10-16北京|杭州|上海
社招3年以上技术-基础平台
1、有扎实的工程算法基础，精通数据结构和常用算法，熟练掌握各种编译、调试、性能分析工具； 2、有极佳的工程实现能力，精通Java、Go、Python之一； 3、有调度系统开发经验者优先，有MLOps系
更新于 2026-09-29北京|杭州