nvalchemi-reporting

par nvidia

Comment ajouter de l'observabilité aux workflows de dynamique et d'entraînement de nvalchemi en utilisant ReportingOrchestrator, RichReporter, TensorBoardReporter, l'extraction de scalaires, personnalisé…

npx skills add https://github.com/nvidia/nvalchemi-toolkit --skill nvalchemi-reporting

nvalchemi Reporting

Overview

Use reporting for curated workflow summaries and dashboards. Use logging for direct event records such as per-system dynamics rows. See docs/userguide/reporting.md, docs/userguide/training.md, docs/userguide/dynamics.md, and docs/userguide/hooks.md for full details.

from nvalchemi.hooks import (
    ReportingOrchestrator,
    RichReporter,
    TensorBoardReporter,
)
from nvalchemi.dynamics.hooks import LoggingHook

Choose The Layer

Use ReportingOrchestrator when the user wants progress summaries, live Rich dashboards, TensorBoard scalar snapshots, rank reductions, or one observability hook that works across training, dynamics, and custom hook-enabled workflows. ReportingOrchestrator is the hook to register with hooks=[...] or register_hook(...); RichReporter, TensorBoardReporter, and custom Reporter objects are sinks owned by that hook and are not registered directly. Workflow engines enter and close hook context managers automatically during run(), so user code should not wrap reporting hooks manually in normal cases.

Use nvalchemi.dynamics.hooks.LoggingHook when the user wants a durable per-graph dynamics event stream. It computes dynamics observables such as energy, fmax, temperature, status, and graph index, then writes one row per system to CSV, TensorBoard, or a custom writer.

Do not reuse the dynamics LoggingHook as a training logger. For training, prefer reporters unless the task explicitly requires a raw training-event log; then implement a training-specific hook with the same hook protocol.


Training Pattern

Attach the ReportingOrchestrator as a normal training hook. Pick stages by enum name when the code already serializes hook specs or when avoiding imports in config files. Use AFTER_OPTIMIZER_STEP for high-frequency loss and learning-rate progress, and validation stages when summaries should align with validation output.

from nvalchemi.hooks import ReportingOrchestrator, RichReporter, TensorBoardReporter
from nvalchemi.training import CheckpointHook, TrainingStrategy

reporting = ReportingOrchestrator(
    [
        TensorBoardReporter("runs/example/tensorboard"),
        RichReporter(layout="training", refresh_per_second=2.0),
    ],
    stages={"AFTER_OPTIMIZER_STEP"},
    frequency=10,
)

strategy = TrainingStrategy(
    models=model,
    optimizer_configs=optimizer_config,
    loss_fn=loss_fn,
    hooks=[
        reporting,
        CheckpointHook("runs/example/checkpoints", epoch_interval=1),
    ],
    num_epochs=20,
)

strategy.run(train_loader)

Guidelines:

  • Register reporting through hooks=[reporting]; strategy.run(...) enters and closes the reporting hook automatically.
  • Keep checkpoints and reporting separate; reporters observe, checkpoint hooks preserve restart state.
  • Use RichReporter(layout="training") for terminal monitoring during local or interactive runs.
  • Use TensorBoardReporter(...) for durable scalar summaries and later review.
  • Set frequency high enough to avoid excessive terminal refresh or file I/O on large distributed runs.
  • Add custom scalar callbacks when a metric is not already exposed through loss, optimizer, scheduler, validation, or workflow context.

Dynamics Pattern

For live summaries or TensorBoard dashboards, use ReportingOrchestrator with a dynamics layout. The dynamics Rich layout asks the reporter to collect default dynamics scalars such as energy, fmax, temperature, convergence fraction, active count, graduated count, and status counts when available.

from nvalchemi.dynamics import DynamicsStage, NVE
from nvalchemi.hooks import ReportingOrchestrator, RichReporter

reporting = ReportingOrchestrator(
    [RichReporter(layout="dynamics", refresh_per_second=2.0)],
    stages={DynamicsStage.AFTER_STEP},
    frequency=25,
)

dynamics = NVE(
    model=model,
    dt=0.5,
    n_steps=10_000,
    hooks=[reporting],
)

final_batch = dynamics.run(batch)

For a durable per-system trajectory-adjacent log, add LoggingHook instead of or in addition to reporting:

from nvalchemi.dynamics import DynamicsStage, NVTLangevin
from nvalchemi.dynamics.hooks import LoggingHook

logger = LoggingHook(
    backend="csv",
    log_path="runs/md/scalars.csv",
    frequency=100,
    stage=DynamicsStage.AFTER_STEP,
)

dynamics = NVTLangevin(
    model=model,
    dt=0.5,
    temperature=300.0,
    n_steps=50_000,
    hooks=[logger],
)

final_batch = dynamics.run(batch)

Guidelines:

  • Register reporting or logging through hooks=[...]; dynamics run(...) enters and closes hook context managers automatically.
  • Use reporting for dashboards and scalar summaries; use SnapshotHook or data sinks for full batch states and trajectories.
  • Use LoggingHook(backend="custom", writer_fn=...) when rows should go to loguru, a database, or a user-owned writer.
  • Give each distributed rank a unique log_path when using dynamics LoggingHook; it writes directly and does not coordinate file access.
  • Attach observation hooks at DynamicsStage.AFTER_STEP unless the metric needs a more specific stage.

Distributed Reporting

Reporters can be rank-gated. Defaults are conservative for terminal and file outputs: rank zero writes or renders unless a reporter requires all ranks for a collective reduction. For setting up the distributed training run itself, see the nvalchemi-training-api skill's Scaling to multiple GPUs section.

reporting = ReportingOrchestrator(
    [
        RichReporter(
            layout="training",
            rank_reduction="mean",
            rank_zero_only=True,
        ),
        TensorBoardReporter("runs/ddp/tensorboard", rank_zero_only=True),
    ],
    stages={"AFTER_OPTIMIZER_STEP"},
    frequency=20,
)

Guidelines:

  • Use rank_reduction="mean" for loss curves in DDP when every rank reports matching scalar keys.
  • Ensure every rank reaches reporters that perform reductions; do not add local control flow that skips nonzero ranks before the collective.
  • Use rank-specific paths such as "runs/ddp/rank-{rank}" only when every rank intentionally writes its own artifact.
  • Keep Rich dashboards rank-zero-only for normal terminal runs.

Custom Scalars

Pass custom_scalars to reporters when metrics are present in the context or batch but are not part of the default extraction path. Return plain numbers or scalar tensors; keep callbacks cheap because they run at reporting frequency.

def grad_norm(ctx, stage):
    total = 0.0
    for parameter in ctx.model.parameters():
        if parameter.grad is not None:
            total += float(parameter.grad.detach().norm().item())
    return total

reporting = ReportingOrchestrator(
    [
        RichReporter(custom_scalars={"optimizer/grad_norm": grad_norm}),
        TensorBoardReporter(
            "runs/example/tensorboard",
            custom_scalars={"optimizer/grad_norm": grad_norm},
        ),
    ],
    stages={"AFTER_OPTIMIZER_STEP"},
    frequency=25,
)

For dynamics LoggingHook, callbacks receive DynamicsContext and must return either a per-graph tensor with shape (B,) or a scalar that can be broadcast to all graphs:

logger = LoggingHook(
    backend="csv",
    log_path="runs/md/scalars.csv",
    custom_scalars={
        "max_velocity": lambda ctx: ctx.batch.velocities.norm(dim=-1).max(),
    },
    frequency=100,
)

Rich Dashboard Workflows

Use RichReporter.preview(...) when choosing or checking a dashboard surface without running a real workflow:

from nvalchemi.hooks import RichReporter

RichReporter.preview(layout="training", title="training preview")
RichReporter.preview(layout="dynamics", title="dynamics preview")

Use BaseRichLayout for custom table-plus-plot dashboards. Implement the RichLayout protocol directly only when the output needs a custom Rich renderable. Do not create a nested rich.live.Live; RichReporter owns the console, lifecycle, refresh rate, history, and rank filtering.


Checklist

  • Add observability by default for long-running examples, CLI scaffolds, fine-tuning scripts, DDP jobs, and dynamics simulations.
  • Prefer ReportingOrchestrator([RichReporter(...), TensorBoardReporter(...)]) for training progress.
  • Add LoggingHook(backend="csv", ...) for dynamics runs that need analyzable per-system rows.
  • Register ReportingOrchestrator as the hook; let TrainingStrategy.run() and dynamics run() enter and close hook context managers automatically.
  • Do not ask users to wrap reporters or dynamics LoggingHook manually unless they are invoking hook calls outside a workflow engine.
  • Avoid expensive scalar callbacks and .item() calls inside hot training code unless they run at a controlled reporting frequency.
  • Mention generated output paths in examples and tests so users know where to inspect CSV, TensorBoard, checkpoints, and dashboards.

Plus de skills de nvidia

compileiq-debug
nvidia
Utilisez quand quelque chose ne va pas : Search() bloque, toutes les évaluations retournent INVALID_SCORE, les scores ne s'améliorent pas, chaque configuration retourne le même nombre, erreurs ptxas…
create-github-pr
nvidia
Créer des pull requests GitHub en utilisant l'interface en ligne de commande gh. Utiliser lorsque l'utilisateur souhaite créer une nouvelle PR, soumettre du code pour révision, ou ouvrir une pull request. Mots-clés de déclenchement -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
Analyse les autres problèmes ouverts pour trouver ceux qu’une PR donnée pourrait également corriger ou casser accidentellement. Génère des opportunités de correctifs adjacents et des risques de contradiction avec fichier:ligne…
fhir-basics
nvidia
Apprend aux agents comment fonctionnent les API FHIR R4, quelles ressources sont disponibles, comment les interroger avec des paramètres de recherche, et comment analyser correctement tous les formats de réponse…
compileiq-validate-result
nvidia
Utiliser APRÈS qu'une recherche soit terminée et AVANT de réclamer un accélérateur ou d'expédier un ACF. Charge le CSV dump_results, extrait les K meilleurs candidats (mono-objectif)…
changelog-audit
nvidia
Auditer le CHANGELOG.md de Warp avant une publication : récupérer les entrées perdues, trier par impact utilisateur, affiner le langage des entrées, ajuster les retours à la ligne et (en mode branche de publication) mettre à jour la comparaison…
maintain-dynamic-plugins
nvidia
Maintenir les chargeurs de plugins dynamiques NeMo Relay, les manifestes, les SDK natifs Rust, le protocole worker gRPC, le SDK worker Python, la documentation, les tests et la couverture du workflow de publication
dgx-diagnose
nvidia
Diagnostiquer les problèmes courants du DGX Station GB300 — plantages CUDA, ciblage incorrect du GPU, bugs de conteneur vLLM/SGLang, problèmes d'état MIG, erreurs NVLink/Fabric Manager,…