Skip to content

Open nowPosted 66 days ago

大模型Infra工程师 (大模型后训练)

shengshu/index98 open roles

Where
Beijing, Chinese Mainland
Get the CV for this job

From $25 per CV, paid once. No subscription.

Your applicationOpen now大模型Infra工程师 (大模型后训练)shengshu/index · Beijing, Chinese Mainland
  1. YouYes, apply to this one.

  2. CV RocketCV written for this posting.

  3. 25 readersRecruiter, hiring manager, skeptic. Round after round.

  4. CV RocketApplied on shengshu/index's own form.

The reply lands in your private mailbox

3×more interviews than doing it yourself with ChatGPT.

The clock on this job

Early applications get read.

12.3% of postings close within 7 days. Measured by our own scanner across the market. shengshu/index postings stay open a median of 1 days.

Share of postings closed within
  1. 2.1%1 day
  2. 4.5%3 days
  3. 12.3%7 days
  4. 19.3%14 days
  5. 35.6%30 days
  6. 57.8%60 days
This job: posted 66 days ago

shengshu/index median: 1 days open

The posting

负责生成式大模型后训练阶段(SFT 微调 → 偏好对齐 → 强化学习)的训练基础设施设计与优化。面对算法快速迭代与大规模探索式采样的双重挑战,构建高效、灵活、可扩展的训练框架,在算法演进与系统效率之间实现最优平衡。

1. 后训练全链路框架设计与演进:主导设计并持续迭代后训练基础设施,完整覆盖 SFT(监督微调)→ RM(奖励模型训练)/ DPO(离线偏好优化)→ PPO / GRPO / DAPO(在线强化学习) 的对齐链路。面向算法快速演进的需求,构建模块化、可插拔的框架架构,实现新算法的高效接入与快速验证。 2. 大规模 Rollout 采样系统优化:针对在线 RL 训练中 Policy 模型频繁生成(Rollout)带来的海量推理请求,设计高效的采样调度系统。优化推理与训练之间的数据流转管道,实现生成、奖励打分、梯度更新三大环节的流水线并行与无缝衔接。 3. 混合资源调度与弹性训练:解决后训练场景下推理(Rollout)与训练(Gradient Update)对 GPU 资源的竞争问题。设计灵活的资源分配与弹性扩缩容策略,根据算法阶段动态调整推理与训练的算力配比,实现集群利用率的全局最优。 4. 训练稳定性与性能调优:保障大规模 RL 训练的稳定性,解决分布式环境下的探索多样性、奖励尺度震荡及训练崩溃等系统性问题。持续优化全链路吞吐,在极致的迭代速度与训练收敛性之间取得平衡。

1. 算法理解与系统转化能力:深刻理解 SFT、DPO、PPO、GRPO、DAPO 等主流后训练算法的数学原理与计算模式,熟悉 Reward Model 的训练范式及其在 RL 链路中的角色,能够从算法特性出发设计系统架构。理解不同算法阶段对 Infra 的差异化诉求。 2. 框架实战经验:深度使用或精读过 veRL、OpenRLHF、TRL 等开源 RL 训练框架之一的核心源码。有在大规模(百卡/千卡级)集群上落地 RL 训练的实际经验,熟悉其中的分布式采样、经验回放、优势估计等关键模块的实现与瓶颈。 3. 系统架构能力:熟悉推理与训练混合部署的架构模式(如 SGLang/vLLM + 训练框架的混合调度),理解两者之间的显存隔离、通信隔离与任务编排策略。 4. 性能优化方法论:具备从端到端视角进行性能压榨的能力,能够精准定位采样瓶颈、数据传输瓶颈与训练瓶颈,并在迭代速度与训练效果之间进行系统级的权衡决策。 5. 加分项: - 有视觉/多模态生成模型的后训练(SFT/RM/DPO/RL)落地经验 - 具备优秀的 SFT 阶段调参经验与数据配比/构建经验,深刻理解 SFT 质量对后续 RM 训练及 RL 收敛效果的影响 - 熟悉在线 RL 与离线 RL 的 Infra 差异,并有实际调优经验 - 有自研或深度定制 RL 训练框架的经历

From $25, paid onceGet the CV for this job

What happens when you press

One press. We do the rest.

  1. A CV for this posting

    Written against shengshu/index's own wording, from every piece of relevant proof in your profile.

  2. 25 readers review it

    Recruiter, hiring manager, skeptic and more read every draft, round after round. You get the best round.

    The review screen in CV Rocket: how each CV was read, round by round.
  3. We apply on shengshu/index's form

    Our application engine gets through the hardest forms there are. Where a question needs you, AI suggests the best answer. Don't want us applying from our IP addresses? Use our Chrome extension: we apply straight from your own browser.

    An application in CV Rocket: every answer filled in on the employer's form.
  4. Every reply, sorted

    shengshu/index's answer lands in your private mailbox, and we classify it on arrival: interview, question, rejection.

    The CV Rocket inbox: each employer reply classified as an interview, an action or a rejection.
  5. Reply with AI

    AI helps you write the email, checks it and sends it. We show you whether the recruiter read it.

  6. The interview in your calendar

    Full integration with your calendar. The invitation goes straight in.

    An interview invitation in the CV Rocket inbox, added to the candidate's calendar.
Get the CV for this job

From $25 per CV, paid once. No subscription.

Why it works

3×

more interviews than doing it yourself with ChatGPT.

ChatGPT writes a CV and never learns what happened to it. We see every reply. For each CV we know:

  • How it was written, and how the review scored it
  • When we applied, and how long after the posting went up
  • Which posting, which company, which city
  • Who got the interview, and who heard nothing

That is how we know which CVs get called.

Get the CV for this job

From $25 per CV, paid once. No subscription.

The numbers game

More applications. More interviews.

Every application goes out with its own CV, written for that posting and paid once. Send enough of them and the law of large numbers finds you the job.

By hand5–10
With CV Rocket100
applications a day

Nearby

Live postings like this one

Same employer first, then the same role elsewhere.

Before you press

Straight answers

Get the CV for this job

From $25 per CV, paid once. No subscription.

What if my background isn't good enough?

We make the most of the background you have. The CV uses every piece of relevant proof your profile holds, and one of the 25 readers reads your whole profile and flags what the CV left out.

Do you really apply for me?

Yes, on the employer's own form, the hardest ones included. Where a question needs you, you answer it right there and AI suggests the best answer. Don't want us applying from our IP addresses? Use our Chrome extension: we apply straight from your own browser.

Is it a subscription?

No. You pay once per CV, from $25. Every application goes out with its own CV, written for that posting.

One job. One CV.
Paid once.

Pick the posting you want. We write for it, apply for you and catch the reply.

Get the CV for this job

From $25 per CV, paid once. No subscription.