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 para a instalação

Repositório: inclusionAI/AReno
Última atividade na origem: 28 de julho de 2026 às 03:45
Idioma detectado do SKILL.md: inglês
Estrelas: 311
Forks: 123

Opções de instalação

Por padrão, está selecionado o prompt que primeiro revisa a origem. Você pode mudar para um comando direto ou baixar uma cópia local.

Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.

Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.

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

O comando permanece em uma só linha. Role horizontalmente para revisá-lo antes de copiar.

Prefere uma cópia local? Baixe os arquivos disponíveis atualmente no SkillsMP.

Revise os arquivos de origem

Leia o SKILL.md e os arquivos complementares exibidos pelo SkillsMP antes de decidir se vai instalar.

Explorador de arquivos

10 arquivos

SKILL.md

Instruções da origem · Visualização somente leitura

Voltar à pesquisa Voltar à instalação

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 todas deste repositório

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.