Skip to main content
The Infra page shows the hardware and system-level state of the cluster. It has three tabs: Metrics, Topology, and Model.

​
Metrics

Time-series charts for GPU, CPU, and vLLM metrics across the cluster, organized into sections:
  • GPU metrics — utilization, memory usage, temperature, power draw, and PyTorch-level metrics like torch_allocated_gb, per GPU. You can filter by role (trainer vs inference) and group by node
  • CPU metrics — CPU utilization and system memory usage per node
  • vLLM metrics — per-server stats like requests running/waiting, KV cache usage, token throughput, and latencies
A Live mode keeps the view pinned to the most recent data. You can also enable Aggregate smoothing to average metrics over a configurable time window, useful for reducing noise on high-frequency metrics.

​
Topology

A visual map of the cluster showing every node, its GPUs, and which role each GPU is assigned to (trainer or inference). This gives you a quick overview of how resources are distributed across the cluster, including hardware details like GPU model, memory, CUDA/PyTorch/Python versions, and interconnect info.

​
Model

A visualization of the model architecture, showing the layer structure and parameter layout. Useful for understanding the model you’re training at a glance.