Fabryka SDK

The official Python client for the hosted Fabryka tracking API.

pip install fabryka

The SDK sends runs, metrics, logs and artifacts to your configured Fabryka API. It does not start a dashboard, run a server, create a local database, or provision GPUs.

Configure

Create an API key in your Fabryka account and set:

export FABRYKA_API_URL=https://track.fabryka.ai
export FABRYKA_API_KEY=your_api_key

Track a run

from fabryka import run

run.init(
    project="sub150run",
    name="lfm-32m",
    experiment="architecture-ablation",
    config={"parameters": 32_793_472, "tokens": 1_000_000_000},
)

for step in range(1000):
    run.log({"train/loss": loss, "throughput/tokens_sec": tokens_per_second}, step=step)

run.log_text("checkpoint saved")
run.artifact("runs/checkpoint-step-001000.pt")
run.finish()

Events are persisted in a local spool before upload. Temporary network failures therefore do not lose metrics; pending events are retried by the client. Set FABRYKA_SPOOL_DIR to choose a different spool location.

The package exposes RunClient for multiple API clients or dependency injection:

from fabryka import RunClient

client = RunClient(api_url="https://track.fabryka.ai", api_key="...")
client.init(project="demo", name="run-1")
client.log({"loss": 1.2}, step=0)
client.finish()

The self-hosted API and dashboard live in the source repository, but are not included in this client distribution.

Neptune-style compatibility

Common Neptune calls can use the Fabryka adapter:

from fabryka.neptune import init_run

run = init_run(project="sub150run", api_token="...", name="baseline")
run["train/loss"].append(2.1, step=0)
run["model/parameters"] = 32_793_472
run["checkpoints/latest"].upload("runs/latest.pt")
run.stop()

The adapter sends data to Fabryka and does not install or contact Neptune.