42Learning rate schedules: what warmup and cosine decay each buy you.▼mediumNVIDIAGoogleOpenAI1 replies◆ premiumToo high and the loss diverges to NaN, too low and it stalls. Warmup ramps up to survive the early steps, decay anneals down to settle. What each fixes.Open full answer →
31What is model cold-start / warmup in serving, and how do you handle it for autoscaling?▼mediumAmazonMicrosoftGoogle1 replies◆ premiumA freshly started replica is slow on its first requests while it loads weights and compiles kernels, and that clashes with reactive autoscaling. The signal is naming the cold-start sources and the warmup and provisioning fixes.Open full answer →