小红书大模型训练框架研发工程师/专家
社招全职5-10年引擎地点：上海 | 北京状态：招聘
任职要求
任职资格： 1、精通PyTorch 框架、veRL/OpenRLHF/Llama-Factory等后训练引擎，具有修改框架源码的实战经验 2、深入理解Megatron/DeepSpeed等框架的并行策略，能自主设计混合并行方案解决显存墙问题 3、掌握强化学习训练全流程优化，有基于PPO/DPO等算法的大模型训练调优经验 4、具备模型训练调优分析经验，能够借助Nsight、nvprof等工具分析发现模型训练性能瓶颈，并进行针对性优化； 5、有良好的沟通表达及团队协作能力，有强烈的责任心和使命感。 加分项： 1、熟悉TRL、DeepSpeed-RL等强化学习框架的底层实现机制 2、掌握LLM训练全链路技术栈，包括分布式数据预处理、序列并行、梯度累积策略 3、有千卡规模大模型训练实战经验，成功解决过跨机房通信、容错训练等生产问题 4、发表过SOSP/MLSys等系统顶会论文，或主导过开源训练框架核心模块开发 5、熟悉NCCL/RDMA/IB/ROCE相关知识，有高性能CUDAKernel相关研发经验；
工作职责
我们是小红书中台大模型 Infra 团队，专注打造领先易用的「AI 大模型全链路基础设施」！团队深耕大模型「数-训-压-推-评」技术闭环，在大模型训练加速、模型压缩、推理优化、部署提效等方向积累了深厚的技术优势，基于 RedAccel 训练引擎、RedSlim 压缩工具、RedServing 推理部署引擎、DirectLLM 大模型 API 服务、QuickSilver 大模型生产部署平台等核心产品，持续赋能社区、商业、交易、安全、数平、研效等多个核心业务，实现 AI 技术高效落地！ 1、参与设计实现支持RLHF/DPO等对齐技术的高效训练框架，优化强化学习阶段的Rollout、Reward Model集成、多阶段训练 Pipline； 2、研发支持多机多卡 RL 的分布式训练框架，开发TP/PP/ZeRO-3与RL流程的动态协同机制，解决 RL 算法在超长时序下的显存/通信瓶刭 3、构建端到端后训练工具链，主导框架与 MLOps 平台集成，提供训练可视化、自动超参搜索等生产级能力 4、与公司各算法部门深度合作，参与大语言模型LLM、多模态大模型 MLLM等业务在 SFT/RL领域的算法探索和引擎迭代； 5、参与分析各业务 GPU 利用率与饱和度等指标，结合业务场景持续优化训练框架能力，提升框架领先性。
包括英文材料
Linux+
https://ryanstutorials.net/linuxtutorial/
Ok, so you want to learn how to use the Bash command line interface (terminal) on Unix/Linux.
https://ubuntu.com/tutorials/command-line-for-beginners
The Linux command line is a text interface to your computer.
https://www.youtube.com/watch?v=6WatcfENsOU
In this Linux crash course, you will learn the fundamental skills and tools you need to become a proficient Linux system administrator.
https://www.youtube.com/watch?v=v392lEyM29A
Never fear the command line again, make it fear you.
https://www.youtube.com/watch?v=ZtqBQ68cfJc
C+
https://www.freecodecamp.org/chinese/news/the-c-beginners-handbook/
本手册遵循二八定律。你将在 20% 的时间内学习 80% 的 C 编程语言。
https://www.youtube.com/watch?v=87SH2Cn0s9A
https://www.youtube.com/watch?v=KJgsSFOSQv0
This course will give you a full introduction into all of the core concepts in the C programming language.
https://www.youtube.com/watch?v=PaPN51Mm5qQ
In this complete C programming course, Dr. Charles Severance (aka Dr. Chuck) will help you understand computer architecture and low-level programming with the help of the classic C Programming language book written by Brian Kernighan and Dennis Ritchie.
C+++
https://www.learncpp.com/
LearnCpp.com is a free website devoted to teaching you how to program in modern C++.
https://www.youtube.com/watch?v=ZzaPdXTrSb8
Python+
https://liaoxuefeng.com/books/python/introduction/index.html
中文，免费，零起点，完整示例，基于最新的Python 3版本。
https://www.learnpython.org/
a free interactive Python tutorial for people who want to learn Python, fast.
https://www.youtube.com/watch?v=K5KVEU3aaeQ
Master Python from scratch 🚀 No fluff—just clear, practical coding skills to kickstart your journey!
https://www.youtube.com/watch?v=rfscVS0vtbw
This course will give you a full introduction into all of the core concepts in python.
数据结构+
https://www.youtube.com/watch?v=8hly31xKli0
In this course you will learn about algorithms and data structures, two of the fundamental topics in computer science.
https://www.youtube.com/watch?v=B31LgI4Y4DQ
Learn about data structures in this comprehensive course. We will be implementing these data structures in C or C++.
https://www.youtube.com/watch?v=CBYHwZcbD-s
Data Structures and Algorithms full course tutorial java
算法+
https://roadmap.sh/datastructures-and-algorithms
Step by step guide to learn Data Structures and Algorithms in 2025
https://www.hellointerview.com/learn/code
A visual guide to the most important patterns and approaches for the coding interview.
https://www.w3schools.com/dsa/
深度学习+
https://d2l.ai/
Interactive deep learning book with code, math, and discussions.
PyTorch+
https://datawhalechina.github.io/thorough-pytorch/
PyTorch是利用深度学习进行数据科学研究的重要工具，在灵活性、可读性和性能上都具备相当的优势，近年来已成为学术界实现深度学习算法最常用的框架。
https://www.youtube.com/watch?v=V_xro1bcAuA
Learn PyTorch for deep learning in this comprehensive course for beginners. PyTorch is a machine learning framework written in Python.
TensorFlow+
https://www.youtube.com/watch?v=tpCFfeUEGs8
Ready to learn the fundamentals of TensorFlow and deep learning with Python? Well, you’ve come to the right place.
https://www.youtube.com/watch?v=ZUKz4125WNI
This part continues right where part one left off so get that Google Colab window open and get ready to write plenty more TensorFlow code.
Megatron+
https://www.youtube.com/watch?v=hc0u4avAkuM
DeepSpeed+
https://www.youtube.com/watch?v=pDGI668pNg0
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
性能调优+
https://goperf.dev/
The Go App Optimization Guide is a series of in-depth, technical articles for developers who want to get more performance out of their Go code without relying on guesswork or cargo cult patterns.
https://web.dev/learn/performance
This course is designed for those new to web performance, a vital aspect of the user experience.
https://www.ibm.com/think/insights/application-performance-optimization
Application performance is not just a simple concern for most organizations; it’s a critical factor in their business’s success.
https://www.oreilly.com/library/view/optimizing-java/9781492039259/
Performance tuning is an experimental science, but that doesn’t mean engineers should resort to guesswork and folklore to get the job done.
CUDA+
https://developer.nvidia.com/blog/even-easier-introduction-cuda/
This post is a super simple introduction to CUDA, the popular parallel computing platform and programming model from NVIDIA.
https://www.youtube.com/watch?v=86FAWCzIe_4
Lean how to program with Nvidia CUDA and leverage GPUs for high-performance computing and deep learning.
PaddlePaddle+
https://learnopencv.com/paddlepaddle/
PaddlePaddle (PArallel Distributed Deep LEarning) is an open-source deep learning framework released by Baidu in 2016.
https://www.paddlepaddle.org.cn/tutorials
本课程采用飞桨特色的「横纵式」 教学法，从易到难，学习难度逐层递进，并结合图形和案例进行讲解，力求让刚接触深度学习的读者可以快速理解。
XTuner+
https://xtuner.readthedocs.io/en/docs/get_started/quickstart.html
In this section, we will show you how to use XTuner to fine-tune a model to help you get started quickly.
https://xtuner.readthedocs.io/zh-cn/latest/
XTuner V1 是一个专为超大规模 MoE 模型打造的新一代大模型训练引擎。
Nsight+
https://developer.nvidia.com/tools-tutorials
NVIDIA Nsight™ Developer tools are a suite of tools for building, profiling, and debugging accelerated applications.
https://www.youtube.com/watch?v=aQ1NYoRvp7o
Profile Python for AI and deep learning applications with NVIDIA's suite of Nsight Developer Tools.
https://www.youtube.com/watch?v=Iuy_RAvguBM
Join NVIDIA’s Jackson Marusarz for an introduction to NVIDIA Nsight Compute, a tool for in-depth analysis of CUDA kernel performance on GPUs.
NVIDIA Visual Profiler+
https://developer.nvidia.com/nvidia-visual-profiler
The NVIDIA Visual Profiler is a cross-platform performance profiling tool that delivers developers vital feedback for optimizing CUDA C/C++ applications.
https://docs.nvidia.com/cuda/profiler-users-guide/
The user manual for NVIDIA profiling tools for optimizing performance of CUDA applications.
https://www.youtube.com/watch?v=F_BazucyCMw
https://www.youtube.com/watch?v=SI4UMz430ZU
This video tutorial has been taken from Learning CUDA 10 Programming.
Llama+
https://github.com/LlamaFamily/Llama-Chinese
Llama中文社区，实时汇总最新Llama学习资料，构建最好的中文Llama大模型开源生态，完全开源可商用。
https://www.llama.com/docs/overview/
This guide provides information and resources to help you set up Llama including how to access the model, hosting, how-to and integration guides.
强化学习+
https://cloud.google.com/discover/what-is-reinforcement-learning?hl=en
Reinforcement learning (RL) is a type of machine learning where an "agent" learns optimal behavior through interaction with its environment.
https://huggingface.co/learn/deep-rl-course/unit0/introduction
This course will teach you about Deep Reinforcement Learning from beginner to expert. It’s completely free and open-source!
https://www.kaggle.com/learn/intro-to-game-ai-and-reinforcement-learning
Build your own video game bots, using classic and cutting-edge algorithms.
LLaMA-Factory+
https://llamafactory.readthedocs.io/en/latest/
LLaMA Factory is an easy-to-use and efficient platform for training and fine-tuning large language models.
相关职位
社招A108975
团队介绍：字节跳动豆包大模型团队（Seed）成立于 2023 年，致力于寻找通用智能的新方法，追求智能上限，并探索新的交互。团队研究方向涵盖 LLM、语音、视觉、世界模型、基础架构、AI Infra、下一代 AI 交互等，在中国、新加坡、美国等地设有实验室和岗位。 豆包大模型团队在 AI 领域拥有长期愿景与决心，坚持深耕基础，期望成为世界一流的 AI 研究团队，为科技和社会发展作出贡献。目前团队已推出业界领先的通用大模型以及前沿的多模态能力，支持豆包、扣子、即梦等超过 50 个应用场景。 1、负责字节跳动机器学习训练框架的研究与开发，服务于全公司各个产品； 2、参与机器学习训练框架底层组件的抽象，设计，优化与落地； 3、与全公司算法部门深度合作，为重点项目进行算法与系统的联合优化。
更新于 2024-05-08
社招3年以上CSIG技术
1.框架开发与优化：负责强化学习、模型精调、知识蒸馏等核心模块的设计与开发，提升框架的训练效率与易用性； 2.分布式训练支持：基于Megatron-LM、DeepSpeed等工具，优化大模型分布式训练策略（数据并行/张量并行/流水并行/专家并行等），解决显存、通信与计算瓶颈； 3.工具链构建：参与开发轻量化训练框架（如LLama-Factory、swift），支持快速模型微调、部署及多硬件平台适配； 4.前沿技术探索：跟踪学术动态（如RLHF、MoE架构、FlashMLA、EPLB、DualPipe等），将最新研究成果转化为框架功能，提升产品竞争力； 5.协作与文档：与产品团队紧密配合，提供框架级解决方案；编写技术文档与案例，赋能公有云客户。
更新于 2025-06-17
社招核心本地商业-基
1.负责文本、多模态等大模型后训练（微调/强化学习）能力建设，包括：训练框架研发、开源模型适配、训练成本优化、训练效果调优等。 2.协同产运、算法团队，推动后训练技术落地业务。 3.与团队其他成员密切合作，共同解决技术难题，推动团队整体技术提升。
更新于 2025-04-30
社招5-10年引擎
1、参与/负责研发面向大语言模型(LLM)/多模态大模型(MLLM)等类型模型的推理服务框架； 2、通过并行计算优化、分布式架构优化、异构调度等多种框架技术，打造高效、易用、领先的AI推理框架； 2、深度参与周边深度学习系统多个子方向的工作，包括但不限于模型管理、推理部署、日志/监控、工作流编排等； 3、与全公司各业务算法部门深度合作，为重点项目进行算法与系统的联合优化，支撑业务目标达成。
更新于 2025-08-25