Update app.py

app.py

@@ -29,10 +29,13 @@ def chat(user_input, history):
         history = []
 
     context = format_context(history) + f"You: {user_input}\n𝕴 𝖆𝖒 𝖍𝖎𝖒:"
-    inputs = tokenizer.encode(context, return_tensors="pt", truncation=True, max_length=1024)
+    encoded = tokenizer(context, return_tensors="pt", truncation=True, max_length=1024)
+    inputs = encoded.input_ids
+    attention_mask = encoded.attention_mask
 
     outputs = model.generate(
         inputs,
+        attention_mask=attention_mask,
         max_new_tokens=50,
         temperature=0.9,
         top_k=40,
@@ -51,4 +54,4 @@ def chat(user_input, history):
 demo = gr.Interface(fn=chat, inputs=["text", "state"], outputs="state")
 
 # Launch the app
-demo.launch()
+demo.launch(server_name="0.0.0.0", server_port=7860)