vllm-vs-atom-decode-trace-comparison SKILL.md skillA
vllm-vs-atom-decode-trace-comparison from shen-shanshan/vllm-dev-skills is an agent skill: a SKILL.md folder an AI agent loads when a task calls for it.
mdr add shen-shanshan/vllm-dev-skills/vllm-vs-atom-decode-trace-comparison@git:20260905.9600717Pins git:20260905.9600717 in mdr.lock. Pin the fingerprint instead.
- Audit grade
- A, 17 of 17 checks passed
- Kind
- SKILL.md skill
- Repository
- shen-shanshan/vllm-dev-skills
- Path
skills/vllm-vs-atom-decode-trace-comparison/SKILL.md- Versions recorded
- 1
- Latest version
- git:20260905.9600717, committed Sep 5, 2026
- Size
- 5,582 bytes
- Repository stars
- 30
- License
- Apache-2.0
Compare decode-phase kernel implementations of vLLM vs ATOM (ROCm/ATOM) running the same model with the same config, using torch profiler Chrome-trace JSONs from both engines. Samples one decode step, extracts one layer per layer type, produces per-layer-type comparison tables (which kernels each engine uses, fused or split, multi-stream, quantization differences, timing), derives a vLLM optimization TODO list, and writes a Markdown report (Chinese by default, English, or both) to the skill's…
Collected from public GitHub. Every version is saved exactly as it was and never changes. Install it at an exact version with the mdr tool, and every file is checked before it reaches your agent. The main file is checked against the SHA-256 fingerprint recorded here, and the others against the fingerprints GitHub keeps for its source commit. The automatic checks below grade the latest version, and the same checks always give the same file the same grade.
What the file says
# vLLM vs ATOM Decode Trace 对比分析 ## 目标与输入 产出一份 Markdown 对比报告：同一模型、同一配置下 vLLM 与 ATOM 的 decode 阶段 GPU kernel 实现差异， 并对每类层生成对比表 + vLLM 优化 TODO 清单。 需要 4 个输入（缺失则询问用户）： 1. 模型名（如 `MiniMaxAI/MiniMax-M3`） 2. vllm trace 路径（torch profiler 导出的 Chrome-trace JSON，可 `.gz`） 3. atom trace 路径（同上） 4. 报告语言：`zh`（默认）/ `en` / `both` 第 1 步先向用户确认"两份 trace 是同模型同配置跑的"这一前提；若用户不确认， 在报告 §1 中显式标注风险。配置明显不同（GPU 数量、引擎版本差异过大）时提醒用户，不拒绝执行。 ## 运行提取脚本 对两份 trace 各执行一次（产物写入本 skill 的 outputs/）： ```bash python3 <skill_dir>/scripts/extract_trace.py --trace <vllm_trace.json> --output-dir <skill_dir>/outputs python3 <skill_dir>/scripts/extract_trace.py --trace <atom_trace.json> --output-dir <skill_dir>/outputs ``` stdout 会给出 `digest:` 与 `slice:` 两个产物路径，以及 `engine/steps/chosen_step/kernels/distinct/replay_stable/busy_pct` 摘要。 - 脚本自动：识别引擎（顶层 key）、用 gpu 通道 `execute_context_*` 注解切分 decode step、跳过首尾窗口选中间 step、 对每个 kernel 名做统计（count/mean/median/sum/streams）、取中间 occurrence 作单层样本、输出 digest(.md) + slice(.json)。 - 一般无需调参。仅当 digest 异常（如 `replay_stable: False` 且回退窗口可疑）时才用 `--step-index N` / `--occurrence N` 覆盖；trace 无 gpu 注解时用 `--bracket cpu-annot`（结果标注为近似）。 - 退出码 2 = trace 没有 step 注解（需用带 profiler 注解的方式重新采集），此时停止并向用户说明。 …
Read the whole file at its exact version.
How to install
mdr add shen-shanshan/vllm-dev-skills/vllm-vs-atom-decode-trace-comparison@sha256:b39d80a73ea4f319The line at the top of this page pins the label git:20260905.9600717, which is what most people want. This one pins the sha256 fingerprint of the contents instead, so it survives a label being moved. Either way the exact version is written to mdr.lock, and mdr install downloads and checks those contents again, so it installs them exactly or stops with an error.
[](https://markdownregistry.com/a/art_uzquolaiznqt7mxu)
0 badge image fetches in 30 days. What the badge shows.
Versions
The registry has recorded 1 version of this file, committed on Sep 5, 2026.
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 (5582 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
shen-shanshan/vllm-dev-skills · 30 stars · license Apache-2.0 · pushed 2026-11-05 · branch master
API
GET https://markdownregistry.com/api/v1/artifacts/art_uzquolaiznqt7mxu GET https://markdownregistry.com/api/v1/resolve?ref=shen-shanshan/vllm-dev-skills/vllm-vs-atom-decode-trace-comparison GET https://markdownregistry.com/api/v1/blob/b39d80a73ea4f31922570ebfea903c85e4ef12ae29398e84909ee9860d55d273
Your AI assistant 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.
Guides: the SKILL.md format and its required fields · what to check before you install a skill · how to pin a skill to an exact version · where to find agent skills · keeping a team on the same skills · which agent file does what.