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.

Ir a la instalación

Repositorio: inclusionAI/AReno
Última actividad en el origen: 28 de julio de 2026 a las 03:45
Idioma detectado de SKILL.md: inglés
Estrellas: 311
Forks: 123

Opciones de instalación

De forma predeterminada está seleccionado el prompt que primero revisa el origen. Puedes cambiar a un comando directo o descargar una copia local.

Instalar con Codex o Claude Copia este prompt, pégalo en Codex, Claude u otro asistente, y deja que revise la página de la skill y la instale por ti.

Un comando directo omite el prompt de revisión. Revisa el origen antes de ejecutarlo.

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

El comando permanece en una sola línea. Desplázate horizontalmente para revisarlo antes de copiarlo.

¿Prefieres una copia local? Descarga los archivos que SkillsMP tiene disponibles ahora.

Revisa los archivos de origen

Lee SKILL.md y los archivos complementarios que muestra SkillsMP antes de decidir si quieres instalarlo.

Explorador de archivos

10 archivos

SKILL.md

Instrucciones de origen · Vista previa de solo lectura

Volver a la búsqueda Volver a la instalación

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

Ver todos los de este repositorio

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.