Update handler.py

handler.py

@@ -3,14 +3,13 @@ from typing import Dict, List, Any
 import torch
 
 class EndpointHandler:
-     def __init__(self, path: str = "Qwen/Qwen3-VL-8B-Instruct"):
+    def __init__(self, path: str = "Qwen/Qwen3-VL-8B-Instruct"):
         # Load tokenizer and model
         self.processor = AutoProcessor.from_pretrained(path)
         self.model = Qwen3VLForConditionalGeneration.from_pretrained(path, device_map="auto")
         self.model.eval()
-         
-    def __call__(self, data: Dict[str, Any]):
-
+    
+    def __call__(self, data: Dict[str, Any]) -> str:
         # Prepare your messages with image and text
         messages = data.get("messages")
 
@@ -24,6 +23,6 @@ class EndpointHandler:
         )
         inputs = inputs.to(self.model.device)
 
-        generated_ids = self.model.generate(**inputs, max_new_tokens=2048)
+        generated_ids = self.model.generate(**inputs, max_new_tokens=128)
         output_text = self.processor.batch_decode(generated_ids, skip_special_tokens=True)
         return output_text[0]