distributed-training skillA
distributed-training is agent-read markdown (skill) from yuecao365/offercome: 分布式训练深挖：并行策略、ZeRO/FSDP、通信重叠、MFU、checkpoint 与故障、RL 训练系统。.
Indexed from public GitHub and served as immutable, content-addressed versions. Install it pinned to an exact SHA-256 with the mdr CLI, and every file is verified against the hash recorded here before it reaches your agent. The deterministic audit below grades the latest version, and the same file always earns the same grade.
What the file says
## 面试官在意什么 这本是训练侧的系统深挖：把一个模型在几十到几千张卡上训起来、训得稳、训得满。面试官算的是 MFU、通信占比、checkpoint 恢复时间；追的是慢节点、掉卡、loss 尖峰这些真实故障的时间线。2026 年多了一块：RL 后训练的系统（rollout 与训练分离、推理引擎与训练框架的权重同步）成了大模型团队的日常。 ## 项目 / 实习怎么深挖 简历上出现下面这类经历时从哪里切、追什么。追到候选人能说出机制、数字的来源与一次真实的故障或取舍才算实；只有框架名与结论、说不出自己那一段的，记为危险信号。通用的追问方法见 project-deep-dive。 - 简历出现千卡训练 / Megatron / DeepSpeed → 追并行配置与理由、MFU、通信占比、checkpoint 间隔与恢复时间 - 简历出现 FSDP / ZeRO → 追用了哪一级、显存省了多少、通信多了多少 - 简历出现训练事故 → 追 loss 尖峰、掉卡、NCCL 超时怎么发现与处理、事后改了什么 - 简历出现混合精度 → 追 BF16 还是 FP16、溢出怎么处理、哪些层保留高精度 - 简历出现 RL 训练系统 → 追 rollout 怎么跑、权重怎么同步到推理引擎、吞吐瓶颈在哪 - 简历出现"训练加速 X%" → 追基线、测量条件、是吞吐还是收敛速度 ## 常见失守与危险信号 - 并行策略选择：只会背名字；说不出各并行的通信量与适用层；配置照抄 - ZeRO 与 FSDP：说不出三级各分了什么；不知道通信换显存的代价 - 通信与重叠：不知道通信占比；没做过重叠 - MFU 与性能分析：不知道 MFU；不会从 profiler 找训练瓶颈 - checkpoint 与故障恢复：间隔拍脑袋；同步写阻塞训练；掉卡只会重跑 - 混合精度与数值稳定：loss 尖峰只答"调小学习率"；不知道溢出与梯度裁剪 - 数据管线与 IO：训练等数据不知道；shuffle 与打包影响不清楚 - RL 训练系统：不知道 rollout 是成本大头；权重同步靠重启 ## 常考主题清单 只列名字、阶梯与答实的标志，作"问到哪一层算实"的参考；问哪些、问几道由这份 JD 与这份简历定，不是配额。 ### 训练稳定性与并行基础 - 阶梯：loss 不降或 NaN 的排查顺序 → 混合精度（BF16 / FP16）溢出、梯度裁剪、学习率与 batch size 的关系 → 数据并行、张量并行、流水并行、序列并行各解决什么瓶颈；ZeRO 的三级 → 千卡训练的故障（慢节点、掉卡、通信）与 checkpoint 策略 - 答实的标志：有排查清单（先查数据再查 lr 再查精度）；能说出各并行方式的通信量与适用层；知道 checkpoint 频率与恢复时间的取舍 …
Read the whole file at its exact version.
How to install
mdr add yuecao365/offercome/distributed-training@git:20260920.a4843f5mdr add yuecao365/offercome/distributed-training@sha256:00261f59985c872ePin to a label to follow the author's releases, or to a sha256 to freeze the exact bytes forever. Either way the resolved hash is written to mdr.lock, and mdr install reproduces it on any machine.
[](https://markdownregistry.com/a/art_637rcjfgqdo4c3h5)
1 badge views in 30 days
Versions
Audit of the latest version
- pass: Frontmatter block present
- pass: Frontmatter declares a name
- pass: Frontmatter declares a description
- pass: Size between 200 bytes and 200 KB (6228 bytes)
- pass: No zero-width or bidi control characters
- pass: No instruction hidden inside an HTML comment
- pass: No link to an exfiltration or paste host
- pass: No credential-shaped string
- pass: No instruction to send local credentials anywhere
- pass: No text hidden with inline styles
- pass: No prompt-injection phrasing
- pass: No curl or wget piped into a shell
- pass: No recursive delete of root, home or parent
- pass: No instruction to read or print local credentials
- pass: No base64 blob over 200 characters
- pass: No link to a raw IP address
- pass: No script tag
Source
yuecao365/offercome · 23 stars · license MIT · pushed 2026-09-23 · branch main
API
GET https://markdownregistry.com/api/v1/artifacts/art_637rcjfgqdo4c3h5 GET https://markdownregistry.com/api/v1/resolve?ref=yuecao365/offercome/distributed-training GET https://markdownregistry.com/api/v1/blob/00261f59985c872e7eeb54e2f55e9768d2cc028b6e024ba78cb322e260b26f66
Your agent does the legwork. You hear about the deals worth your word. Hand yours the standing instructions at modelranch.com and it joins the network that reads files like this one.