nv-one-logger-pytorch-lightning-integration

What it is and what it does

This package wraps the Trainer class to inject telemetry collection hooks that automatically capture training lifecycle events—training loops, validation iterations, checkpointing, and model/optimizer initialization. It supplements the Lightning callback mechanism with async checkpoint tracking and app lifecycle events that Lightning's native callbacks don't expose.

You integrate it by calling `hook_trainer_cls()` on your Trainer class, then instantiate the hooked version as you normally would. Many training events (training loop, validation, checkpoint save) are captured implicitly when you call `trainer.fit()`, while others (model init, dataloader init, testing, checkpoint load) require explicit callback method calls. The package depends on nv-one-logger-core and nv-one-logger-training-telemetry to handle configuration and export.

Use it for:

• Automatically instrument training jobs to collect performance metrics and training timelines without modifying training code.
• Track model initialization, optimizer setup, and dataloader creation times alongside training loop metrics.
• Export training telemetry to external systems via configurable exporters for centralized monitoring.
• Measure checkpoint save/load performance and async checkpoint operations during distributed training.
• Correlate training events with system performance data for end-to-end training job analysis.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

Adds training telemetry collection to Lightning applications via callback hooks, capturing training loop events, checkpointing, and lifecycle metrics automatically or on-demand.

Yes, if you use lightning and need built-in training telemetry without manual instrumentation. The package is permissively licensed, has low install friction, and carries no known vulnerabilities. The aging status (289 days since release) is a minor signal to verify that the telemetry output format and exporter ecosystem meet your needs before committing to it in production.

Install

nv-one-logger-pytorch-lightning-integration on PyPI

pip install nv-one-logger-pytorch-lightning-integration

uv add nv-one-logger-pytorch-lightning-integration

poetry add nv-one-logger-pytorch-lightning-integration

Installing nv-one-logger-pytorch-lightning-integration

Before you install

Low install friction with a pure-Python wheel. The package is aging (289 days since release) but carries no known vulnerabilities and depends on stable libraries (lightning, nv-one-logger-core, nv-one-logger-training-telemetry, setuptools).

License in practice

Licensed under Apache-2.0 (permissive), allowing commercial and private use with minimal restrictions—suitable for most production and research contexts.

Quickstart

from nv_one_logger.training_telemetry.api.training_telemetry_provider import TrainingTelemetryProvider
from nv_one_logger.training_telemetry.integration.pytorch_lightning import hook_trainer_cls

TrainingTelemetryProvider.instance().with_base_config(config).with_exporter(exporter).configure_provider()
HookedTrainer, callback = hook_trainer_cls(Trainer, TrainingTelemetryProvider.instance())
trainer = HookedTrainer(max_epochs=10, devices=1)
trainer.fit(model, train_loader)
callback.on_app_end()

Requires torch >= 2.8.0 and lightning >= 2.5.3; Python >= 3.9, < 3.14.

Verify before relying

• Whether the telemetry output format and exporter ecosystem are documented or stable across versions.
• Performance overhead of telemetry collection during training runs.
• Compatibility with custom callbacks and trainer subclasses.
• Whether async checkpoint operations work with all distributed training backends.

Package facts

Evidence: nv_one_logger_pytorch_lightning_integration-2.3.1-py3-none-any.whl

Tags

More Artificial Intelligence packages