Calibrate ZeroGPU duration from measured ~8s/turn throughput

app.py

@@ -377,13 +377,17 @@ def stream_turn(messages, max_new_tokens: int, temperature: float, top_p: float,
 
 
 def _estimate_duration(*args, **kwargs) -> int:
-    max_turns = 8
-    if len(args) >= 5:
-        try:
+    """GPU budget, sized from measured throughput (~8 s/turn at 1536 max_new_tokens)."""
+    max_turns, max_new = 8, 1536
+    try:
+        if len(args) >= 5 and args[4] is not None:
             max_turns = int(args[4])
-        except Exception:
-            pass
-    return int(min(280, 40 + max_turns * 22))
+        if len(args) >= 6 and args[5] is not None:
+            max_new = int(args[5])
+    except Exception:
+        pass
+    per_turn = 6.0 + 6.0 * (max_new / 1536.0)
+    return int(min(280, 15 + max_turns * per_turn))
 
 
 # --------------------------------------------------------------------------------------