import os
import sys
import torch
import uvicorn
import redis
import numpy as np
from fastapi import FastAPI, Query, BackgroundTasks
from fastapi.responses import HTMLResponse
from starlette.middleware.cors import CORSMiddleware
from datasets import load_dataset
from transformers import AutoTokenizer, GPT2LMHeadModel, pipeline
from loguru import logger
from dotenv import load_dotenv
from sklearn.metrics.pairwise import cosine_similarity
sys.path.append('..')
load_dotenv()
huggingface_token = os.getenv('HUGGINGFACE_TOKEN')
kaggle_username = os.getenv('KAGGLE_USERNAME')
kaggle_key = os.getenv('KAGGLE_KEY')
redis_host = os.getenv('REDIS_HOST', 'localhost')
redis_port = os.getenv('REDIS_PORT', 6379)
redis_password = os.getenv('REDIS_PASSWORD', 'huggingface_spaces')
redis_client = redis.Redis(host=redis_host, port=redis_port, password=redis_password, decode_responses=True)
MAX_ITEMS_PER_TABLE = 10000
def get_current_table_index():
return int(redis_client.get("current_table_index") or 0)
def increment_table_index():
current_index = get_current_table_index()
redis_client.set("current_table_index", current_index + 1)
def store_to_redis_table(key, content):
current_index = get_current_table_index()
table_name = f"table_{current_index}"
item_count = redis_client.hlen(table_name)
if item_count >= MAX_ITEMS_PER_TABLE:
increment_table_index()
table_name = f"table_{get_current_table_index()}"
redis_client.hset(table_name, key, content)
def load_and_store_models(model_names):
for name in model_names:
try:
model = GPT2LMHeadModel.from_pretrained(name)
tokenizer = AutoTokenizer.from_pretrained(name)
sample_text = "Sample input"
generated_text = model.generate(tokenizer.encode(sample_text, return_tensors="pt"), max_length=50)
decoded_text = tokenizer.decode(generated_text[0], skip_special_tokens=True)
store_to_redis_table(name, decoded_text)
redis_client.hset("models", name, decoded_text)
except Exception as e:
logger.error(f"Error loading model {name}: {e}")
app = FastAPI()
app.add_middleware(
CORSMiddleware,
allow_origins=["*"],
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"]
)
message_history = []
@app.get('/')
async def index():
chat_history = redis_client.hgetall(f"table_{get_current_table_index()}")
chat_history_html = "".join(f"
{msg}
" for msg in chat_history.values())
html_code = f"""
ChatGPT Chatbot
"""
return HTMLResponse(content=html_code, status_code=200)
def calculate_similarity(base_text, candidate_texts):
base_vector = np.array([len(base_text)])
similarities = []
for text in candidate_texts:
candidate_vector = np.array([len(text)])
similarity = cosine_similarity([base_vector], [candidate_vector])
similarities.append(similarity[0][0])
return similarities
@app.get('/autocomplete')
async def autocomplete(q: str = Query(..., title='query'), background_tasks: BackgroundTasks = BackgroundTasks()): # Corrección: Mover background_tasks al final
global message_history
message_history.append(('user', q))
background_tasks.add_task(generate_responses, q)
return {"status": "Processing request, please wait..."}
@app.get('/get_response')
async def get_response(q: str = Query(..., title='query')):
response = redis_client.hget("responses", q)
return {"response": response}
def generate_responses(q):
generated_responses = []
try:
for model_name in redis_client.hkeys("models"):
try:
model_data = redis_client.hget("models", model_name)
model = GPT2LMHeadModel.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)
text_generation_pipeline = pipeline("text-generation", model=model, tokenizer=tokenizer, device=0 if torch.cuda.is_available() else -1)
generated_response = text_generation_pipeline(q, do_sample=True, max_length=50, num_return_sequences=5)
generated_responses.extend([response['generated_text'] for response in generated_response])
except Exception as e:
logger.error(f"Error generating response with model {model_name}: {e}")
if generated_responses:
similarities = calculate_similarity(q, generated_responses)
most_coherent_response = generated_responses[np.argmax(similarities)]
store_to_redis_table(q, "\n".join(generated_responses))
redis_client.hset("responses", q, most_coherent_response)
else:
logger.warning("No valid responses generated.")
except Exception as e:
logger.error(f"General error in autocomplete: {e}")
if __name__ == '__main__':
gpt2_models = [
"gpt2",
"gpt2-medium",
"gpt2-large",
"gpt2-xl"
]
programming_models = [
"google/bert2bert_L-24_uncased",
"microsoft/CodeGPT-small-java",
"microsoft/CodeGPT-small-python",
"Salesforce/codegen-350M-multi"
]
load_and_store_models(gpt2_models + programming_models)
uvicorn.run(app=app, host='0.0.0.0', port=int(os.getenv("PORT", 8001)))