Update app.py

app.py

@@ -888,7 +888,7 @@ class ZeroEngine:
                     logger.info(f"[TOKEN] User batch size: {user_batch_size}")
                 
                 # CPU can handle larger batches with quantized models
-                optimal_batch = max(1024, min(8192, optimal_batch))  # 1024-8192 range for CPU
+                optimal_batch = max(256, min(512, optimal_batch))  # 256-512 range for CPU (much more conservative)
                 
                 # Context size
                 optimal_ctx = quant_config["ctx_size"]