clearml/clearml-agent

What it solves

ClearML Agent solves the complexity of MLOps and LLMOps by providing a zero-configuration orchestration and scheduling solution. It eliminates the need for manual environment setup, YAML/JSON configuration, or deep DevOps knowledge to run machine learning experiments across a hybrid cluster of local and cloud-based GPU/CPU resources.

How it works

The agent acts as a job scheduler that listens to specific queues on a ClearML Server. When an experiment is enqueued via the UI or API, the agent:

1. Creates a virtual environment or launches a specified Docker container.
2. Clones the required code from a Git repository.
3. Installs the necessary Python packages (automatically selecting PyTorch versions based on the machine's CUDA version).
4. Executes the code and monitors its progress, logging all output (stdout/stderr) back to the ClearML UI.

It can run on bare-metal, VMs, or within Kubernetes (including a "Glue mode" that maps ClearML jobs to K8s jobs). A "Bootstrap" system is also available to speed up cold starts by providing precompiled binaries for Python, Git, and UV.

Who it’s for

It is designed for ML researchers and engineers who need to scale their experiments from a single machine to a dynamic cluster without spending significant time on infrastructure management or container orchestration.

Highlights

• Hybrid Resource Management: Combine GPU resources across on-premises and multiple cloud providers.
• Zero-Config Execution: No need for manual template or configuration files to launch jobs.
• Flexible Deployment: Supports virtualenv, Docker, Kubernetes, and SLURM.
• Priority Scheduling: Supports priority queues to ensure critical jobs are executed first.
• Rapid Bootstrapping: Uses a self-contained bootstrap system to launch agents up to 10x faster.
• Services Mode: Ability to launch long-running service containers for autoscaling, controllers, or optimizers.