Update vlm.py

vlm.py

@@ -27,7 +27,7 @@ processor = AutoProcessor.from_pretrained(model_id)
 model = Mistral3ForConditionalGeneration.from_pretrained(
     model_id, 
     #_attn_implementation="flash_attention_2",
-    torch_dtype=torch.bfloat16
+    torch_dtype=torch.float16
 ).eval().to(device)
 
 #
@@ -122,7 +122,7 @@ def stream_response(
         tokenize=True,
         return_dict=True,
         return_tensors="pt",
-    ).to(model.device, dtype=torch.bfloat16)
+    ).to(model.device, dtype=torch.float16)
     
     # Generate
     streamer = TextIteratorStreamer(