gpu-kernels skillA
gpu-kernels is agent-read markdown (skill) from yuecao365/offercome: GPU 算子深挖：CUDA、访存、roofline、FlashAttention、融合、profiler、量化、通信。.
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 ai-infra 里最底层的一段：算子怎么写、怎么量、怎么知道该不该自己写。面试官要候选人能读 profiler 的时间线、用 roofline 说清一个 kernel 是算力密集还是带宽密集、讲出 FlashAttention 省的是什么、写过至少一个 kernel 并有前后数字。校招考 CUDA 编程模型与手撕分块 softmax，社招考实测优化案例与什么 shape 下反而变慢。 ## 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 CUDA / Triton 算子 → 追替代了哪个算子、瓶颈是算力还是带宽、用了什么手段、profiler 前后对比、哪些 shape 反而变慢 - 简历出现 FlashAttention / 自定义 attention → 追分块与在线 softmax 怎么实现、显存与速度各省多少、和官方实现比 - 简历出现算子融合 / torch.compile → 追融合了什么、消除了什么开销、编译时间与失败回退 - 简历出现量化内核 → 追数据格式、反量化放在哪、精度与速度的实测 - 简历出现 NCCL / 通信优化 → 追通信量、重叠怎么做、掉卡与超时怎么定位 - 简历出现"性能提升 X 倍" → 追基线、测量条件、是 kernel 时间还是端到端时间 ## 常见失守与危险信号 - CUDA 执行模型：说不清 warp 与线程块；不知道 occupancy 是什么 - 内存层次与访存：不知道合并访存与 bank conflict；把所有慢归结为"算力不够" - roofline 与瓶颈判断：不会算算力与带宽比；不知道 decode 为什么带宽密集 - FlashAttention：只会说"更快"；说不出省的是显存读写不是计算 - 算子融合与编译：不知道融合消除的是什么开销；编译失败不知道怎么回退 - profiler：不用 nsys / torch profiler；GPU 利用率低只会"加 batch" - 量化内核：不知道反量化放在哪、什么时候反而变慢 - 通信算子：不知道 all-reduce 与 all-to-all 的差别；掉卡只会重跑 ## 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。 ### CUDA 执行模型 …
Read the whole file at its exact version.
How to install
mdr add yuecao365/offercome/gpu-kernels@git:20260920.a4843f5mdr add yuecao365/offercome/gpu-kernels@sha256:6b9e8fdf243fd397Pin 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_cgzyl4jue6hzw2r7)
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 (6202 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_cgzyl4jue6hzw2r7 GET https://markdownregistry.com/api/v1/resolve?ref=yuecao365/offercome/gpu-kernels GET https://markdownregistry.com/api/v1/blob/6b9e8fdf243fd39734875e7aa01aed5c11a76fd9d99ae7252c3792cc305a5059
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.
More from yuecao365/offercome
Every file in yuecao365/offercome