Update app.py

app.py

@@ -7,29 +7,32 @@ from huggingface_hub import InferenceClient
 import os
 
 # Initialize Qwen client
-qwen_client = InferenceClient(api_key=os.environ.get("HF_TOKEN"))
+client = InferenceClient("meta-llama/Llama-3.2-3B-Instruct")
 
 # Model and ONNX setup
 HG_MODEL = "livekit/turn-detector"
 ONNX_FILENAME = "model_quantized.onnx"
 PUNCS = string.punctuation.replace("'", "")
-MAX_HISTORY = 4
+MAX_HISTORY = 4  # Adjusted to use the last 4 messages
 MAX_HISTORY_TOKENS = 512
-EOU_THRESHOLD = 0.5
+EOU_THRESHOLD = 0.5  # Updated threshold to match original
 
 # Initialize ONNX model
 tokenizer = AutoTokenizer.from_pretrained(HG_MODEL)
 onnx_session = ort.InferenceSession(ONNX_FILENAME, providers=["CPUExecutionProvider"])
 
+# Softmax function
 def softmax(logits):
     exp_logits = np.exp(logits - np.max(logits))
     return exp_logits / np.sum(exp_logits)
 
+# Normalize text
 def normalize_text(text):
     def strip_puncs(text):
         return text.translate(str.maketrans("", "", PUNCS))
     return " ".join(strip_puncs(text).lower().split())
 
+# Format chat context
 def format_chat_ctx(chat_ctx):
     new_chat_ctx = []
     for msg in chat_ctx:
@@ -38,14 +41,19 @@ def format_chat_ctx(chat_ctx):
             if content:
                 msg["content"] = content
                 new_chat_ctx.append(msg)
+
+    # Tokenize with chat template
     convo_text = tokenizer.apply_chat_template(
         new_chat_ctx, add_generation_prompt=False, add_special_tokens=False, tokenize=False
     )
+
+    # Remove EOU token from the current utterance
     ix = convo_text.rfind("<|im_end|>")
     return convo_text[:ix]
 
+# Calculate EOU probability
 def calculate_eou(chat_ctx, session):
-    formatted_text = format_chat_ctx(chat_ctx[-MAX_HISTORY:])
+    formatted_text = format_chat_ctx(chat_ctx[-MAX_HISTORY:])  # Use the last 4 messages
     inputs = tokenizer(
         formatted_text,
         return_tensors="np",
@@ -59,50 +67,66 @@ def calculate_eou(chat_ctx, session):
     eou_token_id = tokenizer.encode("<|im_end|>")[-1]
     return probs[eou_token_id]
 
+
+# Respond function
 def respond(
     message,
     history: list[tuple[str, str]],
-	max_tokens=2048,
-    temperature=0.6,
-    top_p=0.95,
+    max_tokens,
+    temperature,
+    top_p,
 ):
-    messages = [{"role": "system", "content": os.environ.get("CHARACTER_DESC", "You are a helpful assistant.")}]
-    
-    for val in history[-MAX_HISTORY:]:
+    # Keep the last 4 conversation pairs (user-assistant)
+    messages = [{"role": "system", "content": os.environ.get("CHARACTER_DESC")}]
+
+    for val in history[-10:]:  # Only use the last 4 pairs
         if val[0]:
             messages.append({"role": "user", "content": val[0]})
         if val[1]:
             messages.append({"role": "assistant", "content": val[1]})
-    
+
+    # Add the new user message to the context
     messages.append({"role": "user", "content": message})
 
+    # Calculate EOU probability
     eou_prob = calculate_eou(messages, onnx_session)
+    print(f"EOU Probability: {eou_prob}")  # Debug output
+
+    # If EOU is below the threshold, ask for more input
     if eou_prob < EOU_THRESHOLD:
-        yield "[Wait... Keep typing...]"
+        yield "[Waiting for user to continue input...]"
         return
 
-
-    stream = qwen_client.chat.completions.create(
-        model="deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B", 
-        messages=messages,
+    # Generate response with Qwen
+    response = ""
+    for message in client.chat_completion(
+        messages,
         max_tokens=max_tokens,
+        stream=True,
         temperature=temperature,
         top_p=top_p,
-    )
+    ):
+        token = message.choices[0].delta.content
+        response += token
+        yield response
+    
+    print(f"Generated response: {response}")
+
 
-    for chunk in stream:
-        new_token = chunk.choices[0].delta.content
-        yield "".join(new_token)
-        
-# Create Gradio interface
+# Gradio interface
 demo = gr.ChatInterface(
     respond,
-    additional_inputs=[
-        gr.Slider(1, 4096, value=512, label="Max Tokens"),
-        gr.Slider(0.1, 4.0, value=0.66, label="Temperature"),
-        gr.Slider(0.1, 1.0, value=0.95, label="Top-p"),
-    ]
+    # additional_inputs=[
+    #     # Commented out to disable user modification of the system message
+    #     # gr.Textbox(value="You are an assistant.", label="System message"),
+    #     gr.Slider(minimum=1, maximum=4096, value=256, step=1, label="Max new tokens"),
+    #     gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
+    #     gr.Slider(
+    #         minimum=0.1,
+    #         maximum=1.0, value=0.95, step=0.05, label="Top-p (nucleus sampling)"
+    #     ),
+    # ],
 )
 
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch()