Update app.py

app.py

@@ -94,7 +94,8 @@ def chat_simple(message, history):
                 max_new_tokens=256,
                 pad_token_id=tokenizer.pad_token_id,
                 do_sample=False,
-                temperature=0.7
+                temperature=0.7,
+                use_cache=False
             )
 
         reply = tokenizer.decode(output[0][input_ids.shape[1]:], skip_special_tokens=True)
@@ -136,7 +137,8 @@ def run_vision(image, text_input, model_id):
                 max_new_tokens=400, 
                 do_sample=False,
                 pad_token_id=processor.tokenizer.pad_token_id or processor.tokenizer.eos_token_id,
-                temperature=0.7
+                temperature=0.7,
+                use_cache=False
             )
             
         output = output[:, inputs["input_ids"].shape[1]:]
@@ -178,7 +180,8 @@ async def api_chat(message: str = Form(...)):
             output = model.generate(
                 input_ids, 
                 max_new_tokens=256,
-                pad_token_id=tokenizer.pad_token_id
+                pad_token_id=tokenizer.pad_token_id,
+                use_cache=False
             )
             
         reply = tokenizer.decode(output[0][input_ids.shape[1]:], skip_special_tokens=True)