Create and manage multiple runs in a single process

Manage multiple runs in a single Python process using W&B’s reinit functionality

Manage multiple runs in a single Python process. This is useful for workflows where you want to keep a primary process active while creating short-lived secondary processes for sub-tasks. Some use cases include:

• Keeping a single “primary” run active throughout a script while spinning up short-lived “secondary” runs for evaluations or sub-tasks.
• Orchestrating sub-experiments in a single file.
• Logging from one “main” process to several runs that represent different tasks or time periods.

By default, W&B assumes each Python process has only one active run at a time when you call wandb.init(). If you call wandb.init() again, W&B will either return the same run or finish the old run before starting a new one, depending on the configuration. The content in this guide explains how to use reinit to modify the wandb.init() behavior to enable multiple runs in a single Python process.

Requirements

To manage multiple runs in a single Python process, you must have W&B Python SDK version v0.19.10 or newer.

reinit options

Use the reinit parameter to configure how W&B handles multiple calls to wandb.init(). The following table describes valid arguments and their effects:

	Description	Creates a run?	Example use case
create_new	Create a new run with wandb.init() without finishing existing, active runs. W&B does not automatically switch the global wandb.Run to new runs. You must hold onto each run object yourself. See the multiple runs in one process example below for details.	Yes	Ideal for creating and managing concurrent processes. For example, a “primary” run that remains active while you start or end “secondary” runs.
finish_previous	Finish all active runs with run.finish() before creating a new one run with wandb.init(). Default behavior for non notebook environments.	Yes	Ideal when you want to break sequential sub-processes into separate individual runs.
return_previous	Return the most recent, unfinished run. Default behavior for notebook environments.	No

W&B does not support create_new mode for W&B Integrations that assume a single global run, such as Hugging Face Trainer, Keras callbacks, and PyTorch Lightning. If you use these integrations, you should run each sub-experiment in a separate process.

Specifying reinit

• Use wandb.init() with the reinit argument directly:

  import wandb
  wandb.init(reinit="<create_new|finish_previous|return_previous>")
  

• Use wandb.init() and pass a wandb.Settings object to the settings parameter. Specify reinit in the Settings object:

  import wandb
  wandb.init(settings=wandb.Settings(reinit="<create_new|finish_previous|return_previous>"))
  

• Use wandb.setup() to set the reinit option globally for all runs in the current process. This is useful if you want to configure the behavior once and have it apply to all subsequent wandb.init() calls in that process.

  import wandb
  wandb.setup(wandb.Settings(reinit="<create_new|finish_previous|return_previous>"))
  

• Specify the desired value for reinit in the environment variable WANDB_REINIT. Defining an environment variable applies the reinit option to wandb.init() calls.

  export WANDB_REINIT="<create_new|finish_previous|return_previous>"
  

The following code snippet shows a high level overview how to set up W&B to create a new run each time you call wandb.init():

import wandb

wandb.setup(wandb.Settings(reinit="create_new"))

with wandb.init() as experiment_results_run:
    # This run will be used to log the results of each experiment.
    # You can think of this as a parent run that collects results
      with wandb.init() as run:
         # The do_experiment() function logs fine-grained metrics
         # to the given run and returns result metrics that
         # you want to track separately.
         experiment_results = do_experiment(run)

         # After each experiment, log its results to a parent
         # run. Each point in the parent run's charts corresponds
         # to one experiment's results.
         experiment_results_run.log(experiment_results)

Example: Concurrent processes

Suppose you want to create a primary process that remains open for the script’s entire lifespan, while periodically spawning short-lived secondary processes without finishing the primary process. For example, this pattern can be useful if you want to train a model in the primary run, but compute evaluations or do other work in separate runs.

To achieve this, use reinit="create_new" and initialize multiple runs. For this example, suppose “Run A” is the primary process that remains open throughout the script, while “Run B1”, “Run B2”, are short-lived secondary runs for tasks like evaluation.

The high level workflow might look like this:

1. Initialize the primary process Run A with wandb.init() and log training metrics.
2. Initialize Run B1 (with wandb.init()), log data, then finish it.
3. Log more data to Run A.
4. Initialize Run B2, log data, then finish it.
5. Continue logging to Run A.
6. Finally finish Run A at the end.

The following Python code example demonstrates this workflow:

import wandb

def train(name: str) -> None:
    """Perform one training iteration in its own W&B run.

    Using a 'with wandb.init()' block with `reinit="create_new"` ensures that
    this training sub-run can be created even if another run (like our primary
    tracking run) is already active.
    """
    with wandb.init(
        project="my_project",
        name=name,
        reinit="create_new"
    ) as run:
        # In a real script, you'd run your training steps inside this block.
        run.log({"train_loss": 0.42})  # Replace with your real metric(s)

def evaluate_loss_accuracy() -> (float, float):
    """Returns the current model's loss and accuracy.
    
    Replace this placeholder with your real evaluation logic.
    """
    return 0.27, 0.91  # Example metric values

# Create a 'primary' run that remains active throughout multiple train/eval steps.
with wandb.init(
    project="my_project",
    name="tracking_run",
    reinit="create_new"
) as tracking_run:
    # 1) Train once under a sub-run named 'training_1'
    train("training_1")
    loss, accuracy = evaluate_loss_accuracy()
    tracking_run.log({"eval_loss": loss, "eval_accuracy": accuracy})

    # 2) Train again under a sub-run named 'training_2'
    train("training_2")
    loss, accuracy = evaluate_loss_accuracy()
    tracking_run.log({"eval_loss": loss, "eval_accuracy": accuracy})
    
    # The 'tracking_run' finishes automatically when this 'with' block ends.

Note three key points from the previous example:

1. reinit="create_new" creates a new run each time you call wandb.init().
2. You keep references of each run. wandb.run does not automatically point to the new run created with reinit="create_new". Store new runs in variables like run_a, run_b1, etc., and call .log() or .finish() on those objects as needed.
3. You can finish sub-runs whenever you want while keeping the primary run open until.
4. Finish your runs with run.finish() when you are done logging to them. This ensures that all data is uploaded and the run is properly closed.