Update app.py

app.py

@@ -1,6 +1,7 @@
 import subprocess
 import sys
 import multiprocessing
+import re  # ⚡ ADDED: Regular expressions to clean chat history
 
 # Install prebuilt CPU wheel at runtime
 subprocess.run([
@@ -17,54 +18,57 @@ from huggingface_hub import hf_hub_download
 # Calculate physical cores to prevent hyperthreading contention
 physical_cores = max(1, multiprocessing.cpu_count() // 2)
 
-# Safely fetch the NUMA distribution flag, fallback to boolean True
 NUMA_STRATEGY = getattr(llama_cpp, "GGML_NUMA_STRATEGY_DISTRIBUTE", True)
 
-# Download the highly compressed Q3_K_M GGUF model file
 model_path = hf_hub_download(
     "Jackrong/Qwopus3.5-9B-Coder-GGUF",
     "Qwopus3.5-9B-coder-Exp-Q3_K_M.gguf",
 )
 
-# Load GGUF model with Q4 KV Cache and maximum CPU optimizations
 llm = Llama(
     model_path=model_path,
     n_gpu_layers=0,      
     n_ctx=4096,
-    
-    # Required for Qwen/Qwopus to prevent context cache corruption
     chat_format="chatml",
-    
-    # ⚡ 1. Hyperthreading Avoidance
     n_threads=physical_cores,       
     n_threads_batch=physical_cores, 
-    
-    # ⚡ 2. NUMA Awareness 
     numa=NUMA_STRATEGY,
-    
-    # ⚡ 3. Q4_0 KV Cache Quantization + Flash Attention
-    # Cuts memory usage significantly compared to default or Q8
     flash_attn=True,
     type_k=llama_cpp.GGML_TYPE_Q4_0,    
     type_v=llama_cpp.GGML_TYPE_Q4_0,    
-    
-    # ⚡ 4. CPU Cache Alignment 
     n_batch=512,    
     n_ubatch=512,   
 )
 
+# ⚡ ADDED: Function to strip reasoning blocks from past turns
+def clean_thoughts(text):
+    if not text:
+        return text
+    # Removes everything between <think> and </think>, including the tags
+    cleaned = re.sub(r'<think>.*?</think>\n*', '', text, flags=re.DOTALL)
+    # Also clean up any stray orphan closing tags just in case
+    cleaned = cleaned.replace('</think>', '').strip()
+    return cleaned
+
 def respond(message, history, system_message, max_tokens, temperature, top_p):
     messages = [{"role": "system", "content": system_message}]
     
-    # Dynamically handle history format (dicts in Gradio 5 vs tuples in Gradio 4)
+    # Dynamically handle history format
     for item in history:
         if isinstance(item, dict):
-            messages.append(item)
+            content = item.get("content", "")
+            # ⚡ Strip thoughts if the message is from the assistant
+            if item.get("role") == "assistant":
+                content = clean_thoughts(content)
+            messages.append({"role": item.get("role"), "content": content})
+            
         elif isinstance(item, (list, tuple)):
             if item[0]:
                 messages.append({"role": "user", "content": item[0]})
             if item[1]:
-                messages.append({"role": "assistant", "content": item[1]})
+                # ⚡ Strip thoughts from the assistant's previous reply
+                cleaned_reply = clean_thoughts(item[1])
+                messages.append({"role": "assistant", "content": cleaned_reply})
 
     messages.append({"role": "user", "content": message})
 
@@ -81,7 +85,6 @@ def respond(message, history, system_message, max_tokens, temperature, top_p):
             response += delta["content"]
             yield response
 
-# Gradio interface definition
 chatbot = gr.ChatInterface(
     fn=respond,
     additional_inputs=[
@@ -90,7 +93,7 @@ chatbot = gr.ChatInterface(
         gr.Slider(minimum=0.1, maximum=2.0, 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"),
     ],
-    title="Qwopus 3.5 Coder FREE CHAT BY ARUSH",
+    title="Qwopus 3.5 Coder (Q3_K_M + Q4 KV Cache)",
 )
 
 if __name__ == "__main__":