generic reliability update 1

app.py

@@ -17,6 +17,16 @@ headers = {
 
 session = requests.Session()
 
+DEFAULT_PARAMS = {
+    "temperature": 0.8,
+    "top_p": 0.95,
+    "top_k": 40,
+    "frequency_penalty": 0,
+    "presence_penalty": 0,
+    "repetition_penalty": 1.1,
+    "max_tokens": 256
+}
+
 def get_timestamp():
     return datetime.datetime.now().strftime("%H:%M:%S")
 
@@ -32,6 +42,23 @@ def predict(message, history, system_prompt, temperature, top_p, top_k, frequenc
     history_format.append({"role": "user", "content": message})
 
     print(f"<|system|> {system_prompt}")
+    print(f"{get_timestamp()} <|user|> {message}")
+
+    current_params = {
+        "temperature": temperature,
+        "top_p": top_p,
+        "top_k": top_k,
+        "frequency_penalty": frequency_penalty,
+        "presence_penalty": presence_penalty,
+        "repetition_penalty": repetition_penalty,
+        "max_tokens": max_tokens
+    }
+
+    non_default_params = {k: v for k, v in current_params.items() if v != DEFAULT_PARAMS[k]}
+    
+    if non_default_params:
+        for param, value in non_default_params.items():
+            print(f"{param}={value}")
 
     data = {
         "model": "meta-llama/Meta-Llama-3.1-405B-Instruct",
@@ -69,7 +96,6 @@ def predict(message, history, system_prompt, temperature, top_p, top_k, frequenc
                             continue  
 
         if partial_message:
-            print(f"<|assistant|> {partial_message}")
             yield partial_message
 
     except RequestException as e:
@@ -139,10 +165,9 @@ with gr.Blocks(theme='gradio/monochrome') as demo:
             frequency_penalty = gr.Slider(-2, 2, value=0, step=0.1, label="Frequency Penalty")
             presence_penalty = gr.Slider(-2, 2, value=0, step=0.1, label="Presence Penalty")
             repetition_penalty = gr.Slider(0.01, 5, value=1.1, step=0.01, label="Repetition Penalty")
-            max_tokens = gr.Slider(1, 4096, value=256, step=1, label="Max Output (max_tokens)")
+            max_tokens = gr.Slider(1, 4096, value=512, step=1, label="Max Output (max_tokens)")
 
     def user(user_message, history):
-        print(f"{get_timestamp()} <|user|> {user_message}")
         history = history or []
         return "", history + [[user_message, None]]