kinetic-manager

name	kinetic-manager
description	Use this skill to manage Keras and JAX remote jobs using Kinetic. It provides expertise in infrastructure provisioning, job submission, monitoring, and result retrieval on GCP.

Kinetic Manager

This skill guides the agent in using Kinetic to run Keras/JAX workloads on cloud TPUs and GPUs.

Workflow

1. Infrastructure Setup

If infrastructure is not yet provisioned:

• Use kinetic up to create the GKE cluster and required resources.
• Use kinetic status to check the current state.

2. Job Submission

To run a workload:

• Use @kinetic.run() for synchronous execution (blocks until completion).
• Call func.run_async() on a @kinetic.run() decorated function for asynchronous execution (returns a JobHandle).

Example:

import kinetic

@kinetic.run(accelerator="tpu-v5e-1")
def train():
  import keras
  # ... training code ...
  return history.history

job = train.run_async()

3. Job Management

For jobs submitted with run_async():

• List jobs: kinetic jobs list
• Check status: kinetic jobs status <job-id>
• Stream logs: kinetic jobs logs <job-id> --follow
• Collect result: kinetic jobs result <job-id>
• Cancel job: kinetic jobs cancel <job-id>

4. Data Management

Use kinetic.Data to pass local or GCS data to your remote functions.

@kinetic.run(accelerator="tpu-v5e-1")
def train(data_dir):
  # ...
  pass

train(kinetic.Data("./my_dataset/"))

Best Practices

• Prefer run_async() for long-running jobs to avoid blocking the local session.
• Always use kinetic down when finished to avoid ongoing cloud costs.
• Use kinetic doctor to diagnose environment or credential issues.