app.py

import os
import re
import logging
import textwrap
import autopep8
import gradio as gr
from huggingface_hub import hf_hub_download
from llama_cpp import Llama
import jwt
from typing import Generator
from fastapi import FastAPI, HTTPException, Depends
from fastapi.security import HTTPBearer, HTTPAuthorizationCredentials
from pydantic import BaseModel
import spaces

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

# JWT settings
JWT_SECRET = os.environ.get("JWT_SECRET")
if not JWT_SECRET:
    raise ValueError("JWT_SECRET environment variable is not set")
JWT_ALGORITHM = "HS256"

# Model settings
MODEL_NAME = "leetmonkey_peft__q8_0.gguf"
REPO_ID = "sugiv/leetmonkey-peft-gguf"

# Generation parameters
generation_kwargs = {
    "max_tokens": 2048,
    "stop": ["```", "### Instruction:", "### Response:"],
    "echo": False,
    "temperature": 0.2,
    "top_k": 50,
    "top_p": 0.95,
    "repeat_penalty": 1.1
}

@spaces.GPU
def download_model(model_name: str) -> str:
    logger.info(f"Downloading model: {model_name}")
    model_path = hf_hub_download(
        repo_id=REPO_ID,
        filename=model_name,
        cache_dir="./models",
        force_download=True,
        resume_download=True
    )
    logger.info(f"Model downloaded: {model_path}")
    return model_path

# Download and load the 8-bit model at startup
model_path = download_model(MODEL_NAME)

@spaces.GPU
def load_model(model_path):
    return Llama(
        model_path=model_path,
        n_ctx=2048,
        n_threads=4,
        n_gpu_layers=-1,  # Use all available GPU layers
        verbose=False
    )

llm = load_model(model_path)
logger.info("8-bit model loaded successfully")

@spaces.GPU
def generate_solution(instruction: str) -> str:
    system_prompt = "You are a Python coding assistant specialized in solving LeetCode problems. Provide only the complete implementation of the given function. Ensure proper indentation and formatting. Do not include any explanations or multiple solutions."
    full_prompt = f"""### Instruction:
{system_prompt}

Implement the following function for the LeetCode problem:

{instruction}

### Response:
Here's the complete Python function implementation:

```python
"""
    
    response = llm(full_prompt, **generation_kwargs)
    return response["choices"][0]["text"]

def extract_and_format_code(text: str) -> str:
    # Extract code between triple backticks
    code_match = re.search(r'```python\s*(.*?)\s*```', text, re.DOTALL)
    if code_match:
        code = code_match.group(1)
    else:
        code = text

    # Remove any text before the function definition
    code = re.sub(r'^.*?(?=def\s+\w+\s*\()', '', code, flags=re.DOTALL)

    # Dedent the code to remove any common leading whitespace
    code = textwrap.dedent(code)

    # Split the code into lines
    lines = code.split('\n')

    # Find the function definition line
    func_def_index = next((i for i, line in enumerate(lines) if line.strip().startswith('def ')), 0)

    # Ensure proper indentation
    indented_lines = [lines[func_def_index]]  # Keep the function definition as is
    for line in lines[func_def_index + 1:]:
        if line.strip():  # If the line is not empty
            indented_lines.append('    ' + line)  # Add 4 spaces of indentation
        else:
            indented_lines.append(line)  # Keep empty lines as is

    formatted_code = '\n'.join(indented_lines)

    try:
        return autopep8.fix_code(formatted_code)
    except:
        return formatted_code

security = HTTPBearer()
app = FastAPI()

class ProblemRequest(BaseModel):
    instruction: str

def verify_token(credentials: HTTPAuthorizationCredentials = Depends(security)):
    try:
        jwt.decode(credentials.credentials, JWT_SECRET, algorithms=[JWT_ALGORITHM])
        return True
    except jwt.PyJWTError:
        raise HTTPException(status_code=401, detail="Invalid token")

@app.post("/generate_solution")
@spaces.GPU
async def generate_solution_api(request: ProblemRequest, authorized: bool = Depends(verify_token)):
    logger.info("Generating solution")
    generated_output = generate_solution(request.instruction)
    formatted_code = extract_and_format_code(generated_output)
    logger.info("Solution generated successfully")
    return {"solution": formatted_code}

@app.post("/stream_solution")
@spaces.GPU
async def stream_solution_api(request: ProblemRequest, authorized: bool = Depends(verify_token)):
    async def generate():
        logger.info("Streaming solution")
        system_prompt = "You are a Python coding assistant specialized in solving LeetCode problems. Provide only the complete implementation of the given function. Ensure proper indentation and formatting. Do not include any explanations or multiple solutions."
        full_prompt = f"""### Instruction:
{system_prompt}

Implement the following function for the LeetCode problem:

{request.instruction}

### Response:
Here's the complete Python function implementation:

```python
"""
        
        generated_text = ""
        for chunk in llm(full_prompt, stream=True, **generation_kwargs):
            token = chunk["choices"]["text"]
            generated_text += token
            yield token
        
        formatted_code = extract_and_format_code(generated_text)
        logger.info("Solution generated successfully")
        yield formatted_code

    return generate()

# Gradio wrapper for FastAPI
def gradio_wrapper(app):
    @spaces.GPU
    def inference(instruction, token):
        import requests
        url = "http://localhost:8000/generate_solution"
        headers = {"Authorization": f"Bearer {token}"}
        response = requests.post(url, json={"instruction": instruction}, headers=headers)
        if response.status_code == 200:
            return response.json()["solution"]
        else:
            return f"Error: {response.status_code}, {response.text}"

    iface = gr.Interface(
        fn=inference,
        inputs=[
            gr.Textbox(label="LeetCode Problem Instruction"),
            gr.Textbox(label="JWT Token")
        ],
        outputs=gr.Code(label="Generated Solution"),
        title="LeetCode Problem Solver API",
        description="Enter a LeetCode problem instruction and your JWT token to generate a solution."
    )
    return iface

if __name__ == "__main__":
    import uvicorn
    from threading import Thread

    # Start FastAPI in a separate thread
    Thread(target=lambda: uvicorn.run(app, host="0.0.0.0", port=8000)).start()

    # Launch Gradio interface
    iface = gradio_wrapper(app)
    iface.launch(share=True)