Hugging Face and Lightning

Both integrations are single callbacks that wire the launchhelm.torch surface (metrics, controls, checkpoints) into the framework's own training loop.

Hugging Face Trainer

pip install "launchhelm[hf]" (pulls transformers and accelerate).

args = TrainingArguments(
    output_dir="/tmp/lh-hf-example",
    num_train_epochs=1,
    per_device_train_batch_size=4,
    logging_steps=1,
    save_strategy="no",
    report_to=[],  # LaunchHelm is the only logger
    disable_tqdm=True,
    run_name="hf-example",
)
trainer = Trainer(
    model=Tiny(),
    args=args,
    train_dataset=Rows(),
    callbacks=[LaunchHelmCallback("quickstart")],
)
trainer.train()

python-hf_trainer.py · verbatim from sdks/python/examples/hf_trainer.py

The callback: opens the run at train start (project falls back to run_name; the full TrainingArguments dict is stored as run config), logs on every logging_steps boundary, applies dashboard commands at on_step_end — which is the collective control point, safe under DDP — uploads each checkpoint save as an artifact, and finishes the run at train end. LaunchHelmCallback(project, controls=...) narrows which built-in controls register (default: lr, weight_decay, grad_clip, dropout, checkpoint). rollback is not offered here: these trainers keep their own step and data position, which a restore in place would leave behind; resume from a checkpoint instead.

PyTorch Lightning

pip install "launchhelm[lightning]".

trainer = Trainer(
    max_epochs=1,
    callbacks=[LaunchHelmCallback("quickstart", name="lightning-example")],
    enable_progress_bar=False,
    enable_model_summary=False,
    logger=False,
    default_root_dir="/tmp/lh-lightning-example",
)
trainer.fit(Module(), train_dataloaders=DataLoader(Rows(), batch_size=4))

python-lightning_trainer.py · verbatim from sdks/python/examples/lightning_trainer.py

The same protocol applies after each optimizer step (on_train_batch_end), on every rank. self.log calls flow through to LaunchHelm streams.

When to use these vs lht directly

The callbacks are the right default when the framework owns the loop. Use lht.init/tr.step directly when you control the loop and want explicit safe points, tunables, or log_reduced aggregates — see pytorch and distributed.