Open-Source GPU Observability Tool l9gpu Released

l9gpu is a node-level agent that goes beyond hardware-level metrics, allowing you to see which experiment, team, or job is occupying the GPU.

Key Features:

• Kubernetes Integration: Provides GPU metrics correlated with Pod, Namespace, and Deployment
• Slurm Support: Maps metrics by Job ID, user, and partition
• LLM Inference Optimization: Native metrics support for vLLM, SGLang, and TGI
• Multi-Vendor Hardware Support: Supports NVIDIA, AMD MI300X, and Intel Gaudi
• Standardized Data Output: Exports metrics via OTLP, and includes 17 Prometheus alert rules and a Grafana dashboard

It was developed based on Meta's gcm project, with expanded Kubernetes attribute mapping and multi-vendor support. It is distributed under the MIT license.