W&B on GX10: Tracking ML Experiments

01What you'll learn

• What a run, a project and a sweep are in W&B.
• How to create a key and log in without leaving it in code.
• How to record settings and metrics from your own script.
• What data leaves the machine and how to work offline.
• How to start an automatic search for settings (a sweep).

02Before you start

• A machine of the NVIDIA GB10 class (for example the ASUS Ascent GX10 or the DGX Spark) or any computer with Python 3 and pip.
• A Weights & Biases account and an API key. According to the documentation as of 03.10.2026 the account is created through the Forge platform (docs.wandb.ai now redirects to CoreWeave's documentation) — the names may change.
• A training or evaluation script to track. If you have none, the example below simulates training.

03Steps

1. Create an API key

   Log in to the dashboard, open your profile → User settings → API keys → New key, give it a name and copy the key at once. Why at once: the full key is shown only once; afterwards you see only its beginning and if you lose it you create a new one.

   ✅

   The key is a secret

   Do not write it in the script, in Git or on a page. Keep it in an environment variable or a password manager. If it leaks — delete it in the dashboard and create a new one.

2. Install and log in

   A virtual environment keeps the project's packages separate from the system.

   bashcopy

   python3 -m venv .venv
   . .venv/bin/activate
   pip install wandb
   wandb login   # asks for the key

   In automated environments the key is passed with export WANDB_API_KEY=<key> before the script (the variable is described in the documentation). ⚠️ We did not check the installation on Arm64.

3. The first run

   wandb.init() starts a run: you give a project and a dictionary of settings. Everything you record with run.log() becomes a chart in the dashboard. The example is from the W&B quickstart and simulates training:

   python · first_run.pycopy

   import wandb
   import random
   
   wandb.login()
   
   project = "my-awesome-project"
   config = {"epochs": 10, "lr": 0.01}
   
   with wandb.init(project=project, config=config) as run:
       offset = random.random() / 5
       for epoch in range(2, config["epochs"]):
           acc = 1 - 2**-config["epochs"] - random.random() / config["epochs"] - offset
           loss = 2**-config["epochs"] + random.random() / config["epochs"] + offset
           run.log({"accuracy": acc, "loss": loss})

   Run it a few times and the dashboard will show several runs with random names. In your own script put the same run.log({...}) where you get the loss at each step.

4. Compare runs

   In the dashboard open the project: the "Runs" column lists them and the charts overlay the metrics. The point of the settings in config is here — you can see which lr gave a lower loss. The documentation also has a parallel-coordinates chart and a parameter-importance analysis (⚠️ not tried by us).

5. Working without sending: offline mode

   If the script must not reach the internet or the data must not leave, set the variable before running:

   bashcopy

   export WANDB_MODE=offline   # keeps the metadata locally, no sync
   python first_run.py
   # WANDB_MODE=disabled turns W&B off completely

   How offline runs are uploaded later is described in the W&B documentation; we did not try it, so we give no command.

6. Sweep: automatic search for settings

   A sweep tries different values of the settings and records a run for each. You describe the search space, register the sweep and start an "agent" that calls your function. The example is from the W&B tutorial:

   python · sweep.pycopy

   import wandb
   
   def objective(config):
       score = config.x**3 + config.y
       return score
   
   def main():
       with wandb.init(project="my-first-sweep") as run:
           score = objective(run.config)
           run.log({"score": score})
   
   sweep_configuration = {
       "method": "random",
       "metric": {"goal": "minimize", "name": "score"},
       "parameters": {
           "x": {"max": 0.1, "min": 0.01},
           "y": {"values": [1, 3, 7]},
       },
   }
   
   sweep_id = wandb.sweep(sweep=sweep_configuration, project="my-first-sweep")
   wandb.agent(sweep_id, function=main, count=10)

   ⚠️

   One project for the sweep and its runs

   The project name in wandb.init() must match the one in wandb.sweep(). If you use multiprocessing, wrap wandb.sweep() and wandb.agent() in if __name__ == "__main__":. The agent is stopped with Ctrl+C (a second time to end it).

   The method here is random. Other methods and options are in the "Define sweep configuration" section of the documentation (not opened by us).

7. What else there is

   The quickstart also points to: tracking models and datasets with Artifacts, sharing in the Registry and reports. For applications with language models the same vendor has W&B Weave. None of these is run in this lesson.

04Check

• The key is created and is not in code, Git or a page.
• wandb login works and the first run shows in the dashboard with at least two metrics.
• You know what data must not be logged to the cloud service.
• A sweep of several runs has finished and the runs are compared.

Quiz

05What's next

06Sources

1. Weights & Biases: quickstart 🌐 global — key, installation, first run; checked 03.10.2026.
2. W&B: sweeps tutorial — configuration, wandb.sweep, wandb.agent; checked 03.10.2026.
3. W&B: environment variables — WANDB_API_KEY, WANDB_MODE, WANDB_BASE_URL.
4. NVIDIA: AI Observability for Data Flywheel — page of the deprecated blueprint; repository.