Quota-friendlier chunking (12/chunk) and lower table/chart token caps

app.py

@@ -59,13 +59,14 @@ PROMPTS = {
 }
 
 # Max new tokens per task (tables/charts can be long).
-MAX_TOKENS = {"ocr": 1024, "table": 2048, "formula": 512, "chart": 1536}
+MAX_TOKENS = {"ocr": 1024, "table": 1536, "formula": 512, "chart": 1024}
 
-# Free ZeroGPU caps per-call duration, and the scheduler scales the declared
-# duration up by a safety factor. Keep each GPU allocation short and process the
-# page's zones in small chunks across several allocations.
+# Free ZeroGPU caps per-call duration (the scheduler scales the declared duration
+# up by a safety factor, so keep it modest) AND caps total GPU-seconds/day per
+# visitor. Process zones in chunks: few enough allocations to be quota-friendly,
+# short enough per call to stay under the duration cap.
 GPU_DURATION = 55
-CHUNK = 8
+CHUNK = 12
 
 # Map a layout label to a recognition task. Anything not listed -> "ocr".
 LABEL_TO_TASK = {