lossy 0.3.0

Two lines to hatch a life - lightweight training-run logger for Lossy (wandb-shaped, zero deps)

lossy

Two lines to hatch a life. A tiny, zero-dependency training-run logger that feeds the Lossy pet — a wandb-shaped API with none of the weight.

import lossy

run = lossy.init(project="my-project", name="run-1", config={"lr": 3e-4})
for step in range(1, 301):
    loss = train_step()
    lossy.log({"train/loss": loss}, step=step)
run.finish()

Setup

Pair with the Lossy app (Settings → Lossy SDK) to get your token, then:

export LOSSY_TOKEN=lossy_xxxxx        # pairing token from the app
export LOSSY_API=https://your-deployment.vercel.app   # optional; defaults to prod

No token? Your script still runs exactly as before — the SDK goes inert with a single warning. It never raises, never blocks, never slows a training loop: logging is a lock-guarded append to a bounded in-memory queue (oldest points drop past 10k), a background thread ships batches (≥50 points or every 5s) with retries and backoff, and every network failure is a stderr warning, not an exception.

Crash honesty

Uncaught exceptions are reported as crashed with an exception summary (sys.excepthook); a clean finish() — or a clean exit without one (atexit) — reports finished. Hard kills (OOM/SIGKILL) can't run hooks; the server's liveness sweep covers those.

Ctrl-C currently reports crashed too (distinguishable downstream via exitInfo.type == "KeyboardInterrupt"); a distinct "interrupted" state is a recorded product decision for later.

Environment knobs (mostly for tests)

LOSSY_TOKEN, LOSSY_API, LOSSY_FLUSH_INTERVAL, LOSSY_BATCH_SIZE, LOSSY_QUEUE_MAX, LOSSY_MAX_RETRIES, LOSSY_BACKOFF_BASE, LOSSY_HTTP_TIMEOUT, LOSSY_FINISH_TIMEOUT.

Tests

cd sdk && uv run --with pytest pytest

Fully offline — an in-process stub server plays the S1 ingest API.