sugiv's picture
Change of settings again for High CPU
a8700ab
import gradio as gr
from huggingface_hub import hf_hub_download
from llama_cpp import Llama
import re
import random
import logging
import os
import jwt
from typing import Dict, Any
import autopep8
import textwrap
from datasets import load_dataset
import time
from collections import defaultdict
import threading
import hashlib
# Rate limiting data structures
token_usage = defaultdict(int)
last_reset_time = time.time()
rate_limit_lock = threading.Lock()
# Constants
MAX_TOKEN_USAGE = 10
RESET_INTERVAL = 24 * 60 * 60 # 24 hours in seconds
# Set up logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
# Load the dataset
dataset = load_dataset("sugiv/leetmonkey_python_dataset")
train_dataset = dataset["train"]
# 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"
MODEL_NAME= "leetmonkey_peft_super_block_q6.gguf"
REPO_ID = "sugiv/leetmonkey-peft-gguf"
# Load the model
model_path = hf_hub_download(repo_id=REPO_ID, filename=MODEL_NAME, cache_dir="./models")
#llm = Llama(model_path=model_path, n_ctx=2048, n_threads=16, n_gpu_layers=-1, verbose=False, mlock=True) ## TPU
llm = Llama(model_path=model_path, n_ctx=1024, n_threads=2, n_gpu_layers=0, verbose=False, mlock=True) ## CPU only
#llm = Llama(model_path=model_path, n_ctx=1024, n_threads=4, n_gpu_layers=-1, verbose=False, mlock=True) ## Nvidia
logger.info("8-bit model loaded successfully")
# User data storage
token_to_problem_solution = {}
# Generation parameters
generation_kwargs = {
"max_tokens": 512,
"stop": ["```", "### Instruction:", "### Response:"],
"echo": False,
"temperature": 0.05,
"top_k": 10,
"top_p": 0.9,
"repeat_penalty": 1.1
}
def verify_token(token: str) -> bool:
try:
jwt.decode(token, JWT_SECRET, algorithms=[JWT_ALGORITHM])
return True
except jwt.PyJWTError:
return False
def check_rate_limit(token: str):
global last_reset_time
with rate_limit_lock:
current_time = time.time()
if current_time - last_reset_time >= RESET_INTERVAL:
token_usage.clear()
last_reset_time = current_time
if token_usage[token] >= MAX_TOKEN_USAGE:
return False, "Rate limit exceeded. Please try again later."
token_usage[token] += 1
return True, ""
def extract_and_format_code(text):
code_match = re.search(r'```python\s*(.*?)\s*```', text, re.DOTALL)
if code_match:
code = code_match.group(1)
else:
code = text
code = textwrap.dedent(code)
lines = code.split('\n')
indented_lines = []
for line in lines:
if line.strip().startswith('class') or line.strip().startswith('def'):
indented_lines.append(line)
elif line.strip():
indented_lines.append(' ' + line)
else:
indented_lines.append(line)
formatted_code = '\n'.join(indented_lines)
try:
return autopep8.fix_code(formatted_code)
except:
return formatted_code
def generate_explanation(problem: str, solution: str, token: str) -> Dict[str, Any]:
if not verify_token(token):
return {"error": "Invalid token"}
is_allowed, message = check_rate_limit(token)
if not is_allowed:
return {"error": message}
problem_solution_hash = hashlib.sha256(f"{problem}{solution}".encode()).hexdigest()
if token not in token_to_problem_solution or token_to_problem_solution[token] != problem_solution_hash:
return {"error": "No matching problem-solution pair found for this token"}
system_prompt = "You are a Python coding assistant specialized in explaining LeetCode problem solutions. Provide a clear and concise explanation of the given solution."
full_prompt = f"""### Instruction:
{system_prompt}
Problem:
{problem}
Solution:
{solution}
Explain this solution step by step.
### Response:
Here's the explanation of the solution:
"""
generated_text = ""
for chunk in llm(full_prompt, stream=True, **generation_kwargs):
generated_text += chunk["choices"][0]["text"]
return {"explanation": generated_text}
def generate_solution(instruction: str, token: str) -> Dict[str, Any]:
if not verify_token(token):
return {"error": "Invalid token"}
is_allowed, message = check_rate_limit(token)
if not is_allowed:
return {"error": message}
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
"""
generated_text = ""
for chunk in llm(full_prompt, stream=True, **generation_kwargs):
generated_text += chunk["choices"][0]["text"]
formatted_code = extract_and_format_code(generated_text)
problem_solution_hash = hashlib.sha256(f"{instruction}{formatted_code}".encode()).hexdigest()
token_to_problem_solution[token] = problem_solution_hash
return {"solution": formatted_code}
def random_problem(token: str) -> Dict[str, Any]:
if not verify_token(token):
return {"error": "Invalid token"}
is_allowed, message = check_rate_limit(token)
if not is_allowed:
return {"error": message}
random_item = random.choice(train_dataset)
problem = random_item['instruction']
return {"problem": problem}
# Create Gradio interfaces
generate_interface = gr.Interface(
fn=generate_solution,
inputs=[
gr.Textbox(label="Problem Instruction"),
gr.Textbox(label="JWT Token")
],
outputs=gr.JSON(),
title="Generate Solution API",
description="Provide a LeetCode problem instruction and a valid JWT token to generate a solution."
)
random_problem_interface = gr.Interface(
fn=random_problem,
inputs=[gr.Textbox(label="JWT Token")],
outputs=gr.JSON(),
title="Random Problem API",
description="Provide a valid JWT token to get a random LeetCode problem."
)
explain_interface = gr.Interface(
fn=generate_explanation,
inputs=[
gr.Textbox(label="Problem"),
gr.Textbox(label="Solution"),
gr.Textbox(label="JWT Token")
],
outputs=gr.JSON(),
title="Explain Solution API",
description="Provide a problem, solution, and valid JWT token to get an explanation of the solution."
)
demo = gr.TabbedInterface(
[generate_interface, explain_interface, random_problem_interface],
["Generate Solution", "Explain Solution", "Random Problem"]
)
# Launch the Gradio app
demo.launch()