Ctrl+K

nvalchemi.hooks.RichReporter#

class nvalchemi.hooks.RichReporter(*, custom_scalars=None, include_losses=True, include_optimizer_lrs=True, include_dynamics_scalars=None, rank_reduction=None, title='nvalchemi report', precision=6, max_scalars=None, history_size=200, layout=None, plot_keys=None, max_plots=3, plot_height=8, refresh_per_second=2.0, console=None, screen=False, transient=False, rank_zero_only=True, strict_layout=False)[source]#

Render scalar reporting snapshots as a live Rich dashboard.

Parameters:

  • custom_scalars (Mapping[str, ScalarCallback] | None, optional) – Additional scalar callbacks passed to collect_scalars().

  • include_losses (bool, default True) – When True, include loss scalars from the hook context.

  • include_optimizer_lrs (bool, default True) – When True, include optimizer learning rates from the hook context.

  • include_dynamics_scalars (bool | None, optional) – When True, include default dynamics observables from the hook context. None lets the selected layout choose; the built-in dynamics layout enables them.

  • rank_reduction (torch.distributed.ReduceOp | {"none", "mean", "sum", "min", "max"} | None, default None) – Optional distributed reduction applied to scalars before rendering. String values are normalized to torch.distributed.ReduceOp. Reduction requires every rank to call this reporter; only rank zero renders the reduced dashboard.

  • title (str, default "nvalchemi report") – Dashboard title.

  • precision (int, default 6) – Significant digits used when formatting scalar values.

  • max_scalars (int | None, optional) – Maximum number of scalar rows to show. When omitted, all scalars are shown.

  • history_size (int, default 200) – Maximum history points retained per scalar.

  • layout (RichLayout | {"training", "dynamics"} | None, optional) – Dashboard layout policy. None and "auto" select the first built-in layout that supports the first reported context.

  • plot_keys (Sequence[str] | None, optional) – Scalar keys to plot. When omitted, the selected layout chooses common metrics for that workflow before falling back to alphabetical order.

  • max_plots (int, default 3) – Maximum number of history plots shown in the dashboard.

  • plot_height (int, default 8) – Height in terminal rows for each plotext plot.

  • refresh_per_second (float, default 2.0) – Rich Live refresh rate used while the reporter is entered.

  • console (Console | None, optional) – Rich console used for output. When omitted, a stderr console is created.

  • screen (bool, default False) – Whether Rich Live should use the terminal alternate screen.

  • transient (bool, default False) – Whether Rich Live should clear the dashboard on exit.

  • rank_zero_only (bool, default True) – Request rank-zero-only dispatch from ReportingOrchestrator.

  • strict_layout (bool, default False) – When True, raise if automatic layout selection cannot match the incoming context. When False, unmatched contexts are ignored.

close()[source]#

Stop the live dashboard if it is active.

Return type:

None

property history: dict[str, tuple[tuple[int, float], ...]]#

Return retained scalar history.

Returns:

Mapping from scalar key to (step, value) history tuples.

Return type:

dict[str, tuple[tuple[int, float], …]]

classmethod preview(*, history=None, layout=None, steps=None, console=None, stage=None, step_count=None, epoch=_PREVIEW_DEFAULT, batch_count=_PREVIEW_DEFAULT, **reporter_kwargs)[source]#

Render a synthetic dashboard preview.

Parameters:

  • history (Mapping[str, Sequence[float]] | None, optional) – Metric history used to populate plots and latest values. Defaults to representative curves from the selected layout.

  • layout (RichLayout | {"training", "dynamics"} | None, optional) – Dashboard layout policy. None selects the training layout.

  • steps (Sequence[int] | None, optional) – Step values aligned with each history sequence. Defaults to range(len(series)).

  • console (Console | None, optional) – Rich console used for preview output.

  • stage (str | None, optional) – Stage label shown in the dashboard header. When omitted, the selected layout supplies a workflow-appropriate default.

  • step_count (int | None, optional) – Step shown in the dashboard header. Defaults to the final step.

  • epoch (int | None, optional) – Epoch shown in dashboard metadata. When omitted, the selected layout supplies a workflow-appropriate default.

  • batch_count (int | None, optional) – Batch count shown in dashboard metadata. When omitted, the selected layout supplies a workflow-appropriate default.

  • **reporter_kwargs (object) – Additional keyword arguments forwarded to RichReporter.

Return type:

None

renderable()[source]#

Build the current dashboard renderable.

Returns:

Rich layout containing the header, latest scalar table, and plots.

Return type:

Layout

report(ctx, stage, state)[source]#

Update the dashboard from one scalar snapshot.

Parameters:

  • ctx (HookContext) – Workflow hook context.

  • stage (Enum) – Hook stage being reported.

  • state (ReportingState) – Shared reporting state from the orchestrator.

Return type:

None

seed_history(history, *, steps=None, stage='AFTER_OPTIMIZER_STEP', step_count=None, epoch=None, batch_count=None, global_rank=0)[source]#

Seed dashboard history without running a workflow.

Parameters:

  • history (Mapping[str, Sequence[float]]) – Metric history used to populate plots and latest scalar values.

  • steps (Sequence[int] | None, optional) – Step values aligned with each metric series.

  • stage (str, default "AFTER_OPTIMIZER_STEP") – Stage label for the synthetic snapshot.

  • step_count (int | None, optional) – Step count for the synthetic snapshot. Defaults to the final step.

  • epoch (int | None, optional) – Epoch metadata for the synthetic snapshot.

  • batch_count (int | None, optional) – Batch metadata for the synthetic snapshot.

  • global_rank (int, default 0) – Rank metadata for the synthetic snapshot.

Returns:

Synthetic latest snapshot produced from history.

Return type:

ScalarSnapshot

On this page
Show Source