Step 9 of 15Fixed

The adapter lagged the model

The fine-tuned adapter loaded 21 s after the model was Ready. A fixed 2-second check now loads it about 1 s before.

Adapter ready after the model
+21 s → −1 s

When the fine-tuned adapter was ready, relative to the model · Seconds after the model reported Ready

Before the fix
+21 s
After the fix
about 1 s before Ready
The adapter used to arrive 21 seconds after the model. Now it is ready first.

Problem

Found while verifying the production change: the LoRA adapter loaded 21 seconds after the model was Ready. Its loader backed off up to 30 seconds between health checks, which was invisible against a 140-second start.

Change

A fixed two-second health check.

Result

On one start after the fix, the adapter loaded about a second before the pod reported Ready.

Decision

Fixed. A request for the adapter no longer waits on a back-off timer after the model is up.