gradio_app.py

import os
import logging
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import gradio as gr
import json
import re

# Configure logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

class ModelManager:
    def __init__(self):
        self.model = None
        self.tokenizer = None
        self.device = None
        self.model_loaded = False
        self.load_model()

    def load_model(self):
        """Load the model and tokenizer"""
        try:
            logger.info("Starting model loading...")
            
            # Check if CUDA is available
            if torch.cuda.is_available():
                torch.cuda.set_device(0)
                self.device = "cuda:0"
            else:
                self.device = "cpu"
            logger.info(f"Using device: {self.device}")
            
            if self.device == "cuda:0":
                logger.info(f"GPU: {torch.cuda.get_device_name()}")
                logger.info(f"VRAM Available: {torch.cuda.get_device_properties(0).total_memory / 1024**3:.2f} GB")
            
            # Get HF token from environment
            hf_token = os.getenv("HF_TOKEN")
            
            logger.info("Loading Llama-3.1-8B-Instruct model...")
            base_model_name = "meta-llama/Llama-3.1-8B-Instruct"
            
            self.tokenizer = AutoTokenizer.from_pretrained(
                base_model_name,
                use_fast=True,
                trust_remote_code=True,
                token=hf_token
            )
            
            self.model = AutoModelForCausalLM.from_pretrained(
                base_model_name,
                torch_dtype=torch.float16 if self.device == "cuda:0" else torch.float32,
                device_map="auto" if self.device == "cuda:0" else None,
                trust_remote_code=True,
                token=hf_token,
                attn_implementation="eager"  # Use eager attention (compatible)
            )
            
            # Set pad token
            if self.tokenizer.pad_token is None:
                self.tokenizer.pad_token = self.tokenizer.eos_token
            
            self.model_loaded = True
            logger.info("✅ Model loaded successfully!")
            
        except Exception as e:
            logger.error(f"❌ Error loading model: {str(e)}")
            self.model_loaded = False

def generate_response(prompt, temperature=0.8):
    """ZERO TRUNCATION GENERATION - Never cut anything!"""
    global model_manager
    
    if not model_manager or not model_manager.model_loaded:
        return "Model not loaded"

    try:
        # Detect CoT requests
        is_cot = any(phrase in prompt.lower() for phrase in [
            "return exactly this json array",
            "chain of thinking", 
            "verbatim"
        ])
        
        logger.info(f"🎯 Request type: {'CoT' if is_cot else 'Standard'}")
        
        # Simple system message
        if is_cot:
            system = "You are an expert at generating JSON training data exactly as requested."
        else:
            system = "You are a helpful AI assistant."
            
        # Format prompt
        formatted = f"""<|begin_of_text|><|start_header_id|>system<|end_header_id|>

{system}

<|eot_id|><|start_header_id|>user<|end_header_id|>

{prompt}

<|eot_id|><|start_header_id|>assistant<|end_header_id|>

"""
        
        # Optimized token limits for speed
        if is_cot:
            max_new = 1500  # Reduced for speed
            min_new = 400   # Reduced minimum
        else:
            max_new = 800   # Significantly reduced for speed
            min_new = 50    # Lower minimum
            
        max_input = 6000  # Safe input limit
        
        logger.info(f"🔢 Token allocation: Input≤{max_input}, Output={min_new}-{max_new}")

        # Tokenize
        inputs = model_manager.tokenizer(
            formatted,
            return_tensors="pt",
            truncation=True,
            max_length=max_input
        )
        
        # Move to device
        if model_manager.device == "cuda:0":
            inputs = {k: v.to(next(model_manager.model.parameters()).device) for k, v in inputs.items()}
        
        logger.info("🚀 Starting generation...")
        
        # Generate with generous parameters
        with torch.no_grad():
            outputs = model_manager.model.generate(
                **inputs,
                max_new_tokens=max_new,
                min_new_tokens=min_new,
                temperature=temperature,
                top_p=0.9,
                do_sample=True,
                num_beams=1,  # Greedy search for speed
                pad_token_id=model_manager.tokenizer.eos_token_id,
                early_stopping=True,  # Enable early stopping for speed
                repetition_penalty=1.1,
                use_cache=True
            )
        
        # Decode the COMPLETE response
        full_response = model_manager.tokenizer.decode(outputs[0], skip_special_tokens=True)
        
        logger.info(f"📏 Full response length: {len(full_response)} chars")
        logger.info(f"📝 Response preview: {full_response[:200]}...")
        
        # ZERO TRUNCATION EXTRACTION - Find content intelligently but never cut
        response = full_response
        
        # Look for the assistant response marker
        assistant_marker = "<|start_header_id|>assistant<|end_header_id|>"
        if assistant_marker in full_response:
            # Find the position after the marker
            marker_pos = full_response.find(assistant_marker)
            if marker_pos != -1:
                # Start after the marker + some whitespace
                start_pos = marker_pos + len(assistant_marker)
                # Skip any immediate whitespace/newlines
                while start_pos < len(full_response) and full_response[start_pos] in ' \n\r\t':
                    start_pos += 1
                
                if start_pos < len(full_response):
                    response = full_response[start_pos:]
                    logger.info(f"✂️ Extracted after assistant marker: {len(response)} chars")
                else:
                    logger.info("🔄 Marker found but no content after, using full response")
            else:
                logger.info("🔄 Marker search failed, using full response")
        else:
            logger.info("🔄 No assistant marker found, using full response")
        
        # For CoT, if we have a JSON array, extract it cleanly
        if is_cot and '[' in response and ']' in response:
            # Find the outermost JSON array
            first_bracket = response.find('[')
            last_bracket = response.rfind(']')
            
            if first_bracket != -1 and last_bracket != -1 and last_bracket > first_bracket:
                json_candidate = response[first_bracket:last_bracket+1]
                
                # Validate it contains the expected structure
                if '"user"' in json_candidate and '"assistant"' in json_candidate:
                    # Count the objects to make sure we have multiple items
                    user_count = json_candidate.count('"user"')
                    if user_count >= 2:  # Should have at least 2 user/assistant pairs
                        response = json_candidate
                        logger.info(f"🎯 Extracted JSON array with {user_count} items: {len(response)} chars")
                    else:
                        logger.info(f"⚠️ JSON array has only {user_count} items, using full response")
                else:
                    logger.info("⚠️ JSON candidate failed validation, using full response")
        
        # Final response
        response = response.strip()
        
        logger.info(f"✅ FINAL response: {len(response)} chars")
        logger.info(f"🎬 Starts with: {response[:150]}...")
        logger.info(f"🎭 Ends with: ...{response[-150:]}")
        
        return response

    except Exception as e:
        logger.error(f"💥 Generation error: {e}")
        return f"Error: {e}"

# Initialize model ONCE
model_manager = ModelManager()

def api_respond(message, history_str, temperature, json_mode, template):
    """ZERO TRUNCATION API - Pure content, no wrappers"""
    try:
        logger.info(f"📨 API Request: {len(message)} chars, temp={temperature}")
        response = generate_response(message, temperature)
        logger.info(f"📤 API Response: {len(response)} chars")
        return response
        
    except Exception as e:
        logger.error(f"💥 API Error: {e}")
        return f"Error: {e}"

# BULLETPROOF GRADIO INTERFACE
demo = gr.Interface(
    fn=api_respond,
    inputs=[
        gr.Textbox(label="Message", lines=8, placeholder="Enter your prompt here..."),
        gr.Textbox(label="History", value="[]", visible=False),
        gr.Slider(minimum=0.1, maximum=1.0, value=0.8, step=0.1, label="Temperature"),
        gr.Textbox(label="JSON Mode", value="", visible=False),
        gr.Textbox(label="Template", value="", visible=False)
    ],
    outputs=gr.Textbox(label="Response", lines=20, max_lines=50),
    title="🎯 Question Generation API - ZERO TRUNCATION",
    description="Rebuilt from scratch with ZERO text cutting. Generates complete responses every time.",
    api_name="respond"
)

if __name__ == "__main__":
    # Enable queue with concurrency limit of 10
    demo.queue(
        default_concurrency_limit=10,  # Handle 10 concurrent requests
        max_size=100  # Allow up to 100 requests in queue
    ).launch(server_name="0.0.0.0", server_port=7860, share=False)