Update app.py

app.py

@@ -2,49 +2,183 @@ import os
 import time
 import spaces
 import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
+from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer, BitsAndBytesConfig
 import gradio as gr
 from threading import Thread
+import re
 
 HF_TOKEN = os.environ.get("HF_TOKEN", None)
 MODEL = "AGI-0/Art-v0-3B"
 
-device = "cuda"  # Use "cpu" if no GPU available
+TITLE = """<h2>Link to the model: <a href="https://huggingface.co/AGI-0/Art-v0-3B">click here</a></h2>"""
 
-tokenizer = AutoTokenizer.from_pretrained(MODEL)
-model = AutoModelForCausalLM.from_pretrained(
-    MODEL, torch_dtype=torch.bfloat16, device_map="auto"
-)
-end_of_sentence = tokenizer.convert_tokens_to_ids("<|im_end|>")
+PLACEHOLDER = """
+<center>
+<p>Hi! How can I help you today?</p>
+</center>
+"""
+
+CSS = """
+.duplicate-button {
+    margin: auto !important;
+    color: white !important;
+    background: black !important;
+    border-radius: 100vh !important;
+}
+h3 {
+    text-align: center;
+}
+"""
 
 class ConversationManager:
     def __init__(self):
-        self.user_history = []  # User-facing history with formatting
-        self.model_history = []  # Model-facing history without formatting
-
-    def add_exchange(self, user_message, model_response):
-        formatted_response = self.format_response(model_response)
-        self.model_history.append((user_message, model_response))
-        self.user_history.append((user_message, formatted_response))
-        print(f"\nModel History Updated: {self.model_history}")
-        print(f"\nUser History Updated: {self.user_history}")
-
-    def format_response(self, response):
-        """Format response for UI while keeping raw text for model."""
-        if "<|end_reasoning|>" in response:
-            parts = response.split("<|end_reasoning|>")
-            reasoning, rest = parts[0], parts[1] if len(parts) > 1 else ""
-            return f"<details><summary>Click to see reasoning</summary>\n\n{reasoning}\n\n</details>\n\n{rest}"
-        return response
-
-    def get_user_history(self):
-        return self.user_history
-
+        self.user_history = []  # For displaying to user (markdown)
+        self.model_history = []  # For feeding back to model (special tags)
+        self.debug_log = []
+        
+    def log(self, message):
+        """Add timestamped log entry"""
+        timestamp = time.strftime('%Y-%m-%d %H:%M:%S')
+        log_entry = f"[{timestamp}] {message}"
+        print(log_entry)
+        self.debug_log.append(log_entry)
+    
+    def convert_to_markdown(self, model_text):
+        """Convert from model format (with special tags) to markdown"""
+        self.log(f"Converting to markdown - Input length: {len(model_text)}")
+        self.log(f"Input text: {model_text[:200]}..." if len(model_text) > 200 else f"Input text: {model_text}")
+        
+        markdown_text = model_text
+        
+        # Convert special tags to markdown
+        tag_conversions = [
+            # Reasoning blocks
+            ("<|start_reasoning|>", "<details><summary>Reasoning</summary>\n\n"),
+            ("<|end_reasoning|>", "\n\n</details>\n\n"),
+            
+            # Other special tags (add more as needed)
+            ("<|im_start|>", ""),
+            ("<|im_end|>", ""),
+            ("<|assistant|>", ""),
+            ("<|user|>", ""),
+        ]
+        
+        for old, new in tag_conversions:
+            if old in markdown_text:
+                self.log(f"Converting tag: {old} -> {new}")
+                markdown_text = markdown_text.replace(old, new)
+        
+        # Clean up any remaining special tags using regex
+        markdown_text = re.sub(r'<\|[^>]+\|>', '', markdown_text)
+        
+        # Fix common markdown formatting issues
+        markdown_text = re.sub(r'\n{3,}', '\n\n', markdown_text)  # Remove excess newlines
+        markdown_text = markdown_text.strip()
+        
+        self.log(f"Markdown conversion complete - Output length: {len(markdown_text)}")
+        self.log(f"Output text: {markdown_text[:200]}..." if len(markdown_text) > 200 else f"Output text: {markdown_text}")
+        
+        return markdown_text
+    
+    def convert_to_model_format(self, markdown_text):
+        """Convert from markdown to model format (with special tags)"""
+        self.log(f"Converting to model format - Input length: {len(markdown_text)}")
+        self.log(f"Input text: {markdown_text[:200]}..." if len(markdown_text) > 200 else f"Input text: {markdown_text}")
+        
+        model_text = markdown_text
+        
+        # Convert markdown to special tags
+        if "<details>" in markdown_text and "</details>" in markdown_text:
+            try:
+                # Extract content between details tags
+                pattern = r'<details><summary>.*?</summary>\s*(.*?)\s*</details>'
+                matches = re.findall(pattern, markdown_text, re.DOTALL)
+                
+                for match in matches:
+                    original = f"<details><summary>Reasoning</summary>\n\n{match}\n\n</details>"
+                    replacement = f"<|start_reasoning|>{match}<|end_reasoning|>"
+                    model_text = model_text.replace(original, replacement)
+                    self.log(f"Converted details block to reasoning tags")
+            except Exception as e:
+                self.log(f"Warning: Failed to convert details block: {str(e)}")
+        
+        # Clean up formatting
+        model_text = re.sub(r'\n{3,}', '\n\n', model_text)  # Remove excess newlines
+        model_text = model_text.strip()
+        
+        self.log(f"Model format conversion complete - Output length: {len(model_text)}")
+        self.log(f"Output text: {model_text[:200]}..." if len(model_text) > 200 else f"Output text: {model_text}")
+        
+        return model_text
+    
+    def add_exchange(self, user_message, assistant_response):
+        """Add a new exchange to both histories"""
+        self.log(f"\n=== Adding New Exchange ===")
+        self.log(f"User Message: {user_message[:100]}..." if len(user_message) > 100 else f"User Message: {user_message}")
+        self.log(f"Assistant Response: {assistant_response[:100]}..." if len(assistant_response) > 100 else f"Assistant Response: {assistant_response}")
+        
+        # Convert assistant response to markdown for user display
+        markdown_response = self.convert_to_markdown(assistant_response)
+        
+        # Store both versions
+        self.model_history.append((user_message, assistant_response))
+        self.user_history.append((user_message, markdown_response))
+        
+        self.log(f"Current History State:")
+        self.log(f"- Model History: {len(self.model_history)} exchanges")
+        self.log(f"- User History: {len(self.user_history)} exchanges")
+    
+    def sync_with_ui_history(self, ui_history):
+        """Sync our histories with the UI history"""
+        self.log(f"\n=== Syncing with UI History ===")
+        self.log(f"UI History Length: {len(ui_history)}")
+        
+        # Clear current histories
+        self.model_history = []
+        self.user_history = []
+        
+        # Rebuild histories from UI
+        for user_msg, markdown_response in ui_history:
+            model_response = self.convert_to_model_format(markdown_response)
+            self.model_history.append((user_msg, model_response))
+            self.user_history.append((user_msg, markdown_response))
+        
+        self.log(f"Sync Complete:")
+        self.log(f"- Model History: {len(self.model_history)} exchanges")
+        self.log(f"- User History: {len(self.user_history)} exchanges")
+        
+        # Verify sync integrity
+        if len(self.model_history) != len(self.user_history) or len(self.model_history) != len(ui_history):
+            self.log("WARNING: History length mismatch after sync!")
+    
     def get_model_history(self):
+        """Get history in model format"""
+        self.log(f"\nReturning Model History ({len(self.model_history)} exchanges)")
         return self.model_history
+    
+    def get_user_history(self):
+        """Get history in markdown format"""
+        self.log(f"\nReturning User History ({len(self.user_history)} exchanges)")
+        return self.user_history
+    
+    def get_debug_log(self):
+        """Get the full debug log"""
+        return "\n".join(self.debug_log)
 
+# Initialize global conversation manager
 conversation_manager = ConversationManager()
 
+device = "cuda"  # for GPU usage or "cpu" for CPU usage
+
+# Initialize model and tokenizer
+tokenizer = AutoTokenizer.from_pretrained(MODEL)
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL,
+    torch_dtype=torch.bfloat16,
+    device_map="auto"
+)
+end_of_sentence = tokenizer.convert_tokens_to_ids("<|im_end|>")
+
 @spaces.GPU()
 def stream_chat(
     message: str,
@@ -56,29 +190,46 @@ def stream_chat(
     top_k: int = 1,
     penalty: float = 1.1,
 ):
-    print(f'User Message: {message}')
+    conversation_manager.log(f'\n=== New Chat Request ===')
+    conversation_manager.log(f'Message: {message}')
+    conversation_manager.log(f'History Length: {len(history)}')
+    conversation_manager.log(f'System Prompt: {system_prompt}')
+    conversation_manager.log(f'Parameters: temp={temperature}, max_tokens={max_new_tokens}, top_p={top_p}, top_k={top_k}, penalty={penalty}')
+    
+    # Sync with UI history
+    conversation_manager.sync_with_ui_history(history)
     
+    # Get model-formatted history
     model_history = conversation_manager.get_model_history()
-    print(f'Model History: {model_history}')
     
+    # Build conversation for model
     conversation = []
+    if system_prompt:
+        conversation.append({"role": "system", "content": system_prompt})
+    
     for prompt, answer in model_history:
         conversation.extend([
             {"role": "user", "content": prompt},
             {"role": "assistant", "content": answer},
         ])
+    
     conversation.append({"role": "user", "content": message})
-
-    print(f'Formatted Conversation for Model: {conversation}')
+    
+    conversation_manager.log(f'Built conversation with {len(conversation)} messages')
     
     input_ids = tokenizer.apply_chat_template(
-        conversation, add_generation_prompt=True, return_tensors="pt"
+        conversation,
+        add_generation_prompt=True,
+        return_tensors="pt"
     ).to(model.device)
-
+    
     streamer = TextIteratorStreamer(
-        tokenizer, timeout=60.0, skip_prompt=True, skip_special_tokens=True
+        tokenizer,
+        timeout=60.0,
+        skip_prompt=True,
+        skip_special_tokens=True
     )
-
+    
     generate_kwargs = dict(
         input_ids=input_ids,
         max_new_tokens=max_new_tokens,
@@ -90,42 +241,96 @@ def stream_chat(
         eos_token_id=[end_of_sentence],
         streamer=streamer,
     )
-
+    
     buffer = ""
     original_response = ""
-
+    
     with torch.no_grad():
         thread = Thread(target=model.generate, kwargs=generate_kwargs)
         thread.start()
-
+        
         for new_text in streamer:
             buffer += new_text
             original_response += new_text
-            print(f'Streaming: {buffer}')
-            formatted_buffer = conversation_manager.format_response(buffer)
-            yield formatted_buffer, history + [[message, formatted_buffer]]
-
+            
+            # Convert buffer to markdown for display
+            formatted_buffer = conversation_manager.convert_to_markdown(buffer)
+            
+            if thread.is_alive() is False:
+                conversation_manager.log(f'Generation Complete:')
+                conversation_manager.log(f'Final Response Length: {len(original_response)}')
+                
+                conversation_manager.add_exchange(
+                    message,
+                    original_response  # Original for model
+                )
+            
+            yield formatted_buffer
 
-    conversation_manager.add_exchange(message, original_response)
+# Initialize Gradio interface
+chatbot = gr.Chatbot(height=600, placeholder=PLACEHOLDER)
 
-chatbot = gr.Chatbot(height=600, placeholder="<center><p>Hi! How can I help you today?</p></center>")
-
-demo = gr.Blocks()
-with demo:
-    gr.HTML("<h2>Link to the model: <a href='https://huggingface.co/AGI-0/Art-v0-3B'>click here</a></h2>")
-    gr.DuplicateButton(value="Duplicate Space for private use", elem_classes="duplicate-button")
+with gr.Blocks(css=CSS, theme="soft") as demo:
+    gr.HTML(TITLE)
+    gr.DuplicateButton(
+        value="Duplicate Space for private use",
+        elem_classes="duplicate-button"
+    )
     gr.ChatInterface(
         fn=stream_chat,
         chatbot=chatbot,
         fill_height=True,
-        additional_inputs_accordion=gr.Accordion(label="⚙️ Parameters", open=False, render=False),
+        additional_inputs_accordion=gr.Accordion(
+            label="⚙️ Parameters",
+            open=False,
+            render=False
+        ),
         additional_inputs=[
-            gr.Textbox(value="", label="System Prompt", render=False),
-            gr.Slider(minimum=0, maximum=1, step=0.1, value=0.2, label="Temperature", render=False),
-            gr.Slider(minimum=128, maximum=8192, step=1, value=4096, label="Max new tokens", render=False),
-            gr.Slider(minimum=0.0, maximum=1.0, step=0.1, value=1.0, label="top_p", render=False),
-            gr.Slider(minimum=1, maximum=50, step=1, value=1, label="top_k", render=False),
-            gr.Slider(minimum=0.0, maximum=2.0, step=0.1, value=1.1, label="Repetition penalty", render=False),
+            gr.Textbox(
+                value="",
+                label="System Prompt",
+                render=False,
+            ),
+            gr.Slider(
+                minimum=0,
+                maximum=1,
+                step=0.1,
+                value=0.2,
+                label="Temperature",
+                render=False,
+            ),
+            gr.Slider(
+                minimum=128,
+                maximum=8192,
+                step=1,
+                value=4096,
+                label="Max new tokens",
+                render=False,
+            ),
+            gr.Slider(
+                minimum=0.0,
+                maximum=1.0,
+                step=0.1,
+                value=1.0,
+                label="top_p",
+                render=False,
+            ),
+            gr.Slider(
+                minimum=1,
+                maximum=50,
+                step=1,
+                value=1,
+                label="top_k",
+                render=False,
+            ),
+            gr.Slider(
+                minimum=0.0,
+                maximum=2.0,
+                step=0.1,
+                value=1.1,
+                label="Repetition penalty",
+                render=False,
+            ),
         ],
         examples=[
             ["Help me study vocabulary: write a sentence for me to fill in the blank, and I'll try to pick the correct option."],