Update app.py

app.py

@@ -1,24 +1,24 @@
-from transformers import AutoTokenizer, AutoModelForCausalLM
-
 import gradio as gr
+from transformers import AutoTokenizer
 import onnxruntime as ort
 import numpy as np
 import string
 from huggingface_hub import InferenceClient
-import os
-
 
-model = AutoModelForCausalLM.from_pretrained('Qwen/Qwen2.5-1.5B-Instruct')
-tokenizer = AutoTokenizer.from_pretrained('Qwen/Qwen2.5-1.5B-Instruct')
+# Initialize Qwen client
+qwen_client = InferenceClient("ystemsrx/Qwen2.5-Sex")
 
-# ONNX setup
+# Model and ONNX setup
+HG_MODEL = "livekit/turn-detector"
 ONNX_FILENAME = "model_quantized.onnx"
-onnx_session = ort.InferenceSession(ONNX_FILENAME, providers=["CPUExecutionProvider"])
-
 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):
@@ -32,26 +32,27 @@ def normalize_text(text):
     return " ".join(strip_puncs(text).lower().split())
 
 # Format chat context
-# Update your format_chat_ctx function to ensure proper tokenization
 def format_chat_ctx(chat_ctx):
     new_chat_ctx = []
     for msg in chat_ctx:
         if msg["role"] in ("user", "assistant"):
             content = normalize_text(msg["content"])
             if content:
-                new_chat_ctx.append({"role": msg["role"], "content": content})
-    
-    # Use correct chat template formatting
-    return tokenizer.apply_chat_template(
-        new_chat_ctx,
-        add_generation_prompt=False,
-        tokenize=False,
-        add_special_tokens=True  # Keep special tokens consistent
+                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",
@@ -65,6 +66,10 @@ def calculate_eou(chat_ctx, session):
     eou_token_id = tokenizer.encode("<|im_end|>")[-1]
     return probs[eou_token_id]
 
+# Read system message from file
+with open("character/herta.txt", "r") as f:
+    system_message = f.read()
+
 # Respond function
 def respond(
     message,
@@ -73,18 +78,28 @@ def respond(
     temperature,
     top_p,
 ):
-    messages = [{"role": "system", "content": os.environ.get("CHARACTER_DESC")}]
-    for val in history[-10:]:
+    # Keep the last 4 conversation pairs (user-assistant)
+    messages = [{"role": "system", "content": system_message}]
+
+    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}")
+    print(f"EOU Probability: {eou_prob}")  # Debug output
+
+    # If EOU is below the threshold, ask for more input
     if eou_prob < EOU_THRESHOLD:
         yield "[Waiting for user to continue input...]"
         return
+
+    # Generate response with Qwen
     response = ""
     for message in qwen_client.chat_completion(
         messages,
@@ -96,11 +111,23 @@ def respond(
         token = message.choices[0].delta.content
         response += token
         yield response
+    
     print(f"Generated response: {response}")
 
+
 # Gradio interface
 demo = gr.ChatInterface(
     respond,
+    # 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__":