Transition a Script from Detached Mode#
Now that you know how to run a workload via detached mode and submit metrics associated with it to Determined, you are ready to transition to submitting the workload to Determined. Submitting the workload to Determined will give you access to features such as distributed training, hyperparameter search, and resource management.
These step-by-step instructions will cover:
Preparing an experiment configuration file to transition a script from detached mode
Submitting the experiment to Determined
By the end of this guide, you’ll:
Understand the process to transition from detached mode
Grasp the benefits of leveraging Determined’s resource management and distributed training capabilities
A Determined cluster
Familiarity with configuration files
Step 1: Prepare the Experiment Configuration File#
To run an experiment, you need, at minimum, a script and an experiment configuration (YAML) file.
We’ll use the script we created in Use Distributed Training with Sharded Checkpointing and create a new experiment configuration file.
Use the following code to create the experiment configuration file:
name: unmanaged-3-torch-distributed # Here we use Determined's PyTorch distributed launcher. # You don't need to specify details about the cluster topology; Determined # takes care of that for you. # Simply specify resource requirements like resources.slots_per_trial. entrypoint: >- python3 -m determined.launch.torch_distributed python3 3_torch_distributed.py resources: slots_per_trial: 2 # Use the single searcher to run just one instance of the training script. searcher: name: single # metric is required but it shouldn't hurt to ignore it at this point. metric: x # max_length is ignored if the training script ignores it. max_length: 1 max_restarts: 0
Step 2: Submit the Experiment to Determined#
Navigate to the directory containing your model code and use the following command to run the experiment:
det e create -m <MASTER_IP:PORT> exp_config_yaml .
To run the experiment on a machine with a single CPU or GPU, set
This command will submit your experiment to Determined, leveraging its managed mode functionalities.
<DET_MASTER_IP:PORT> in your web browser to see the experiment.
Congratulations! You’ve successfully transitioned a script from detached mode to being managed by Determined, tapping into Determined’s resource management and distributed training capabilities. To learn more about Determined, visit the tutorials.