Update app.py

app.py

@@ -32,7 +32,7 @@ def generate_text_streaming(prompt, max_new_tokens=128):
             inputs.input_ids.tolist(),  # ctranslate2 needs list of token ids
             max_length=1, # Generate one token at a time
             beam_size=1, # Greedy decoding
-        ).eval()
+        )
 
         new_token_id = outputs[0][0][-1]  # Extract the generated token ID
         new_token = tokenizer.decode(new_token_id, skip_special_tokens=True)