Update app.py

app.py

@@ -22,20 +22,6 @@ dante = Llama(
     n_ctx=4086,
 ) 
 
-karg = {
-    'input_prompt': input_prompt,
-    'temperature': 0.15,
-    'top_p': 0.1,
-    'top_k': 40, 
-    'repeat_penalty': 1.1,
-    'max_tokens': 1024,
-    'stop': [
-        "[|Umano|]",
-        "[|Assistente|]",
-    ],
-    'stream': True
-       }
-
 history = []
 
 def generate_text(message, history):
@@ -49,6 +35,19 @@ def generate_text(message, history):
 
     print(input_prompt)
 
+    karg = {
+    'input_prompt': input_prompt,
+    'temperature': 0.15,
+    'top_p': 0.1,
+    'top_k': 40, 
+    'repeat_penalty': 1.1,
+    'max_tokens': 1024,
+    'stop': [
+        "[|Umano|]",
+        "[|Assistente|]",
+    ],
+    'stream': True
+       }
     output = saiga(**karg)
     
     for out in output:
@@ -68,6 +67,20 @@ def generate_text_Dante(message, history):
     input_prompt += "[|Umano|] " + message + "\n[|Assistente|]"
 
     print(input_prompt)
+
+    karg = {
+    'input_prompt': input_prompt,
+    'temperature': 0.15,
+    'top_p': 0.1,
+    'top_k': 40, 
+    'repeat_penalty': 1.1,
+    'max_tokens': 1024,
+    'stop': [
+        "[|Umano|]",
+        "[|Assistente|]",
+    ],
+    'stream': True
+       }
            
     output = dante(**karg)