areno-profile-performance

Measure and diagnose AReno rollout, prefill, decode, training, checkpoint, role-switch, communication, or Python scheduling performance. Use when throughput or step time is slow and evidence from metrics, py-spy, or Nsight is required. Do not optimize before correctness is established.

跳到安装

仓库: inclusionAI/AReno
最近来源活动: 2026年7月28日 03:45
检测到的 SKILL.md 语言: 英语
星标: 311
分支: 123

安装方式

默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

直接命令不会经过审查 Prompt；运行前请先检查来源。

npx skills add https://github.com/inclusionAI/AReno --skill areno-profile-performance

命令会保持在同一行。复制前请横向滚动并检查完整内容。

想先保存到本地？可下载 SkillsMP 当前能够提供的文件。

检查来源文件

决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。

文件资源管理器

10 个文件

SKILL.md

来源说明 · 只读预览

返回搜索返回安装

areno-build-singleturn-game-demo

Autonomously research, implement, train, evaluate, visualize, and submit a genuinely new AReno RLVR mini-game whose mechanically generated samples are independent state-to-single-output tasks. Use when the user requests a complete new game demo with real reward improvement, not for multi-turn agent workflows, SFT-only examples, or design-only proposals.

2026-09-02

areno-model-adaptation

Add or debug an AReno model family, including config conversion, module construction, checkpoint load/save, text or multimodal inference, training backward, tensor parallelism, CUDA graph decode, and checkpoint round trips. Use only when model-specific implementation or compatibility changes are required.

2026-08-11

areno-add-algorithm

Add or modify an AReno algorithm, trainer, loss, advantage calculation, role model, or algorithm-specific configuration. Use for framework-level SFT, DPO, GSPO, GRPO, PPO, or new optimization method development. Do not use merely to run an existing algorithm.

2026-07-28

查看这个仓库的全部 Skills

Profile AReno Performance

Profile a bounded, representative train or serve workload. Collect low-overhead time, GPU, process, and TensorBoard evidence first; use sampling or tracing only after those signals identify the subsystem to inspect.

python .agents/skills/areno-profile-performance/scripts/monitor_gpu.py \ --pid <areno-pid> --duration 60 --output /tmp/areno-gpu.jsonl python .agents/skills/areno-profile-performance/scripts/monitor_process.py \ --pid <areno-pid> --duration 60 --output /tmp/areno-process.jsonl python .agents/skills/areno-profile-performance/scripts/summarize_monitor.py \ /tmp/areno-gpu.jsonl python .agents/skills/areno-profile-performance/scripts/summarize_monitor.py \ /tmp/areno-process.jsonl python .agents/skills/areno-profile-performance/scripts/summarize_events.py \ <metrics-dir> --list python .agents/skills/areno-profile-performance/scripts/summarize_events.py \ <metrics-dir> --pattern 'time/*' --pattern '*throughput*' --drop-first 1 python .agents/skills/areno-profile-performance/scripts/probe_openai_latency.py \ --base-url http://127.0.0.1:8000 --model <model> --requests 16 --concurrency 4 python .agents/skills/areno-profile-performance/scripts/build_nsys_command.py \ --output /tmp/areno-profile -- <bounded-command> [args...]

Workflow

Record workload, commit, model, topology, token lengths, concurrency, GPU, and dependency versions.

Find the parent AReno PID and monitor its process tree. For train jobs, attach during at least two post-warmup steps. For serve jobs, use a fixed request set and concurrency.

Capture per-GPU utilization, memory, power, and target-process memory. Memory capacity, compute utilization, and throughput are separate signals.

For train, list TensorBoard scalar names before selecting series. Summarize stage time, tokens/throughput, communication, optimizer, loss, and memory metrics that actually exist; do not assume fixed tags.

For serve, measure TTFT and total request latency with streaming enabled while the GPU and process monitors run. Record prompt/output lengths and active concurrency.

Exclude initialization, checkpoint load, compilation, and CUDA graph capture from steady-state conclusions, but report them separately when startup is the problem.

Use py-spy record -p <pid> -o /tmp/areno.svg --duration 30 for Python scheduling, data processing, serialization, blocking I/O, or compilation orchestration.

Use a bounded Nsight Systems capture for GPU compute, communication, synchronization, allocator activity, or launch gaps. Read references/profile-policy.md.

Compare baseline and candidate with identical workloads and multiple steady-state observations, then re-run correctness validation after optimization.

Report raw workload metadata, selected window, GPU peak/average memory and utilization, process CPU/RSS, stage breakdown or TTFT/latency, bottleneck evidence, profiler overhead, and before/after metrics. A single warmup step is not a benchmark.