Skip to main content
A deployment records every request it serves as a trace: the request, the response, and the telemetry around it. Traces show what the model does in production, and they are the raw material for continuous retraining.

Trace Contents

Live Metrics

The same telemetry, rolled up per version: requests, error rate, time to first token, latency at p50 and p95, and output tokens per second.
The Console shows the metrics over time under Inference, next to the traces.

View Traces in the Console

  1. Click Inference in the sidebar.
  2. Select the deployment, then the Traces tab.
  3. Filter by version, status, time range, latency, or a metadata tag, and click a row to open the trace.

Export Traces

Each line is one trace as a chat messages row, the format of an SFT dataset, with the fields above alongside. Continuous Retraining shows how reviewed traces become training data.

Collection

Collection is on by default. --traces off on a deployment keeps no request or response content; the live metrics stay. With BYOC, traces stay in artifact storage in the team’s account. The Console shows them only when Team → Data Sharing allows it, and monte trace download reads them with the member’s cloud credentials.