Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add shen-shanshan/vllm-dev-skills --skill vllm-vs-atom-decode-trace-comparisongit clone --depth 1 https://github.com/shen-shanshan/vllm-dev-skillsWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/shen-shanshan/vllm-dev-skills/vllm-vs-atom-decode-trace-comparison)<a href="https://agentmods.dev/skills/shen-shanshan/vllm-dev-skills/vllm-vs-atom-decode-trace-comparison"><img src="https://agentmods.dev/badge/skills/shen-shanshan/vllm-dev-skills/vllm-vs-atom-decode-trace-comparison/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/shen-shanshan/vllm-dev-skills/vllm-vs-atom-decode-trace-comparison"><img src="https://agentmods.dev/badge/skills/shen-shanshan/vllm-dev-skills/vllm-vs-atom-decode-trace-comparison.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector pass
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00000 | $0.01582 |
| Opus 5 | $0.00000 | $0.00791 |
| Sonnet 5 | $0.00000 | $0.00316 |
| Haiku 4.5 | $0.00000 | $0.00158 |
Grade A, and why
vllm-vs-atom-decode-trace-comparison scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 5d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 78 lines — stays where its author put it; the contents beside it link to each section on GitHub.
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):
- 模型名(如
MiniMaxAI/MiniMax-M3) - vllm trace 路径(torch profiler 导出的 Chrome-trace JSON,可
.gz) - atom trace 路径(同上)
- 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/):
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。
加载领域知识
必须完整阅读 references/analysis_guide.md 后再开始解释数据。该文件包含:
trace 格式事实、步骤切分原理、layer 识别方法(occurrence-index)、逻辑算子分组规则、
多流检测、对比表与 TODO 编写规则、pitfalls、报告模板、以及 MiniMax-M3 的完整 worked example
(含两引擎真实 kernel 清单,可作方法论参照——其它模型按同样流程处理)。
逐层类型对比流程
- 校验可比性:读两份 digest 的 §1-§3。确认引擎识别正确、窗口可比、
busy/wall在 ~90-100%、 replay_stable 状态一致。不一致时在报告 §6 说明。 - 判定层型与选中层:按 guide §3,用计数族 + kernel 模板变体 + ts 位置判定该模型的层类型分布,
每类层选一个中间层(digest §5 的 middle-occurrence 行即该层样本)。
若模型有 HF config 可查(网络可用时用模型名查
num_hidden_layers、dense/sparse 层分布),与计数族互证。 - 归组逻辑算子:按 guide §4 的标准 forward 序列把每类层的 kernel 归组为逻辑算子;
对不认识的 kernel 名,用 digest §7 的全名在对应引擎仓库 grep(vllm:
csrc/、vllm/_custom_ops.py; ATOM/aiter:csrc/),确认语义后再归组。 - 生成对比表:每类层一张表(模板见 guide §6/§8)。数字只取自 digest/slice:
count×mean归一为单层单次 µs;差异 <5% 持平、5-20% 略、>20% 显著。 - 派生 TODO 清单:按 guide §6 规则,只对 atom 更快且差异可复现的算子写 TODO(写给 vLLM), 按预期收益排序。
What ships with it
7 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
- outputs/atom-minimax-m3-tp4-8k1k-conc8_digest.md 30 KB
- outputs/atom-minimax-m3-tp4-8k1k-conc8_slice.json 54 KB
- outputs/minimax-m3-trace-comparison_zh_20260905_162716.md 16 KB
- outputs/vllm-minimax-m3-tp4-8k1k-conc8_digest.md 33 KB
- outputs/vllm-minimax-m3-tp4-8k1k-conc8_slice.json 60 KB
- references/analysis_guide.md 15 KB
- scripts/extract_trace.py 20 KB runs code
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 5d ago First seen · 78 lines · 0 tokens per session scan A b39d80a73ea4
vllm-vs-atom-decode-trace-comparison is a skill published in the GitHub repository shen-shanshan/vllm-dev-skills (17 stars, last pushed 2d ago), licensed Apache-2.0. It costs nothing until one of its globs matches a file; then it loads 1,582 tokens. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-06.
Other skills, from other repositories
triage-issues
Triage GitHub issues in the googleapis/mcp-toolbox repo: propose the correct labels (type / priority / product / status), check for duplicates, verify a bug has enough info to act on, and draft a triage comment. Use whenever a maintainer asks you to triage, label, categorize, prioritize, or "look at" an issue (or a…
adk-debug
Diagnoses misbehaving ADK agents by inspecting sessions, events, tool calls, and the exact request that reached the model. Covers the adk run CLI and the adk web dev server with its session, trace, and debug HTTP endpoints. Use when an agent returns the wrong answer, ignores a tool or swallows a tool error, hangs…
browser-edge-cases
SOP for debugging browser automation failures on complex websites. Use when browser tools fail on specific sites like LinkedIn, Twitter/X, SPAs, or sites with Shadow DOM.
session-investigator
Investigate fast-agent session and history files to diagnose issues. Use when a session ended unexpectedly, when debugging tool loops, when correlating sub-agent traces with main sessions, or when analyzing conversation flow and timing. Covers session.json metadata, history JSON format, message structure, tool…
rubber-duck
Adversarial "rubber duck" review that turns explaining-out-loud into a hallucination check. The main session is the PRESENTER (it did the work — a design doc, investigation, or analysis — and holds the real reasoning) and reconstructs the topic to a LISTENER — a spawned subagent pinned to a DIFFERENT-vendor model that…
claw-admin
Claw system administration: service management, IM connections, logs, cron, and workspace diagnostics. Use when the user asks to manage claw services, connect/disconnect IM platforms, view logs, or perform system-wide operations.