Tanmay PRO

Agreed, and my own data backs you up. LaunchReady scales the KV cache with context and users, but it folds activation peaks and runtime overhead into a flat 2 GiB reserve. On granite-4.0-h-tiny the measured overhead was 3.70 GiB, which was the confirmed part of a 133% overestimate of users. A better estimate would size the activation peak from the engine’s batch settings instead of using a constant. Thanks for pushing on this.