GPU wake schedule …
The speech models run on one A10 slice (about USD 1.27 an hour, 30 a day if left on). It switches itself off after 30 minutes without a request, wakes on demand, and stays on during the warm windows below.
Status
Press Wake GPU, or just send a request: the API answers 503 gpu_starting with an ETA and starts the node. A cold start takes about 7 to 10 minutes (node + model downloads; 7 min 7 s measured); a warm start about 2. Inside a warm window the workers are never switched off; outside it the 30-minute idle rule applies.
The day (IST)
Green = the GPU is kept on; amber = the lead time before a window, when it is being started; the blue line is now.
Edit the schedule admins
Keep the GPU on during the hours users are active (IST). Ranges as HH:MM-HH:MM, comma-separated; a range may cross midnight (22:00-02:00). The workers wake lead minutes before each window. Sign in with Google (top right) to save; the same windows should be kept in Saathi's config/voice_routing.yaml so off-window traffic goes to the vendor instead of waking the GPU.