LaunchHelm

Training telemetry and live control for Python and PyTorch — the SDK for LaunchHelm. The SDK runs inside your training process; Mission Control, the dashboard, charts the run and applies your controls.

Requires Python 3.11 or later on macOS, Linux, or Windows.

Install

pip install launchhelm

Quickstart

In Mission Control, open the organization's Projects page and choose New project — it creates the project and a project key in one step. The key is shown once. Save it before you leave:

export LAUNCHHELM_API_KEY=lh_...        # shown once — never commit it
export LAUNCHHELM_ORGANIZATION=...      # organization UUID
export LAUNCHHELM_PROJECT=...           # project UUID
# export LAUNCHHELM_URL=https://launchhelm.com   # optional; this is the default

A .env file in the working directory or a parent works too; exported shell variables take precedence. Then in the training script:

import launchhelm

with launchhelm.init("quickstart", config={"lr": 3e-4, "batch": 64}) as run:
    for step in range(100):
        loss = train_one_step()
        run.log({"loss": loss}, step=step)

The first argument is the project name, which must match the project the key belongs to; LAUNCHHELM_PROJECT is that project's UUID.

Rich values

Beyond scalars, a stream can carry rendered values — text, media, tables and histograms:

run.text("notes", "## epoch 4\nloss **0.25**", format="markdown")
run.image("frames", "confusion.png")
run.log_table("sweep", [("lr", "number")], [[3e-4]])
run.histogram("grad", gradient_samples, bins=64)

Media uploads don't block the loop; flush() waits for them.

PyTorch

pip install "launchhelm[torch]"

import launchhelm.torch as lht

tr = lht.init("quickstart", config={"lr": 3e-4, "batch": 64})
model, optimizer = build()
tr.watch(model, optimizer, scheduler=sched)

for step in range(steps):
    loss = train_one_step()
    tr.clip_grad_norm_()          # current (controllable) max norm
    optimizer.step(); sched.step()
    tr.step(loss=loss, samples=batch_size)   # logs + services control; call on every rank

tr.save_checkpoint(extra={"epoch": epoch})
tr.finish()

tr.step logs the built-in series (train/loss, optim/lr/group<i>, optim/grad_norm, perf/step_time_s, perf/samples_per_s, perf/tokens_per_s) and services controls from the dashboard: lr, weight_decay, grad_clip, dropout, checkpoint, eval, plus tr.tunable(name, ...) for custom knobs. In DDP, call step on every rank at the same cadence — a control applies on every rank at the same step.

save_checkpoint writes a torch.load(weights_only=True)-safe bundle — model, optimizer, scheduler, scaler, step, per-rank RNG, tunables, extra — and a crash-resume finds the latest one. load_checkpoint("latest") restores it. For FSDP models, configure a full state dict before saving.

System metrics without torch: launchhelm.system.start(run) (the system extra) emits CPU, memory, disk/network rates and per-GPU stats under system/....

Framework callbacks: launchhelm.torch.hf.LaunchHelmCallback (hf extra) and launchhelm.torch.lightning.LaunchHelmCallback (lightning extra) wire the same lifecycle through each framework's callback API.

Delivery

Links

• launchhelm.com/docs — getting started, Python guide, concepts
• Changelog: CHANGELOG.md, shipped in the source distribution