app.py

import os
import gc
import io
from llama_cpp import Llama
from concurrent.futures import ThreadPoolExecutor, as_completed
from fastapi import FastAPI, Request, HTTPException, Lifespan
from fastapi.responses import JSONResponse
from tqdm import tqdm
from dotenv import load_dotenv
from pydantic import BaseModel
from huggingface_hub import hf_hub_download, login
from nltk.tokenize import word_tokenize
from nltk.corpus import stopwords
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity
import nltk
import uvicorn
import psutil
import torch
import io

nltk.download('punkt')
nltk.download('stopwords')

load_dotenv()

app = FastAPI()
HUGGINGFACE_TOKEN = os.getenv("HUGGINGFACE_TOKEN")
if HUGGINGFACE_TOKEN:
    login(token=HUGGINGFACE_TOKEN)

model_configs = [
    # ... (Your model configurations remain the same) ...
]

global_data = {'model_configs': model_configs, 'training_data': io.StringIO()}

class ModelManager:
    def __init__(self):
        self.models = {}
        self.load_models()

    def load_models(self):
        for config in tqdm(global_data['model_configs'], desc="Loading models"):
            model_name = config['name']
            if model_name not in self.models:
                try:
                    model_bytes = hf_hub_download(repo_id=config['repo_id'], filename=config['filename'], use_auth_token=HUGGINGFACE_TOKEN)
                    model = Llama(model_path=io.BytesIO(model_bytes), n_ctx=512, n_gpu=1)  # Correct: Use io.BytesIO
                    self.models[model_name] = model
                    print(f"Model '{model_name}' loaded successfully.")
                except Exception as e:
                    print(f"Error loading model {model_name}: {e}")
                    self.models[model_name] = None
                finally:
                    gc.collect()

    def get_model(self, model_name: str):
        return self.models.get(model_name)


model_manager = ModelManager()

class ChatRequest(BaseModel):
    message: str

async def generate_model_response(model, inputs: str) -> str:
    try:
        if model:
            response = model(inputs, max_tokens=150)
            return response['choices'][0]['text'].strip()
        else:
            return "Model not loaded"
    except Exception as e:
        return f"Error: Could not generate a response. Details: {e}"

async def process_message(message: str) -> dict:
    inputs = message.strip()
    responses = {}
    loaded_models = [model for model in global_data['model_configs'] if model_manager.get_model(model['name'])]

    with ThreadPoolExecutor(max_workers=min(len(loaded_models), 4)) as executor:
        futures = [executor.submit(generate_model_response, model_manager.get_model(config['name']), inputs) for config in loaded_models]
        for i, future in enumerate(tqdm(as_completed(futures), total=len(futures), desc="Generating responses")):
            try:
                model_name = loaded_models[i]['name']
                responses[model_name] = future.result()
            except Exception as e:
                responses[model_name] = f"Error processing {model_name}: {e}"

    stop_words = set(stopwords.words('english'))
    vectorizer = TfidfVectorizer(tokenizer=word_tokenize, stop_words=stop_words)
    reference_text = message
    response_texts = list(responses.values())
    if response_texts:
        tfidf_matrix = vectorizer.fit_transform([reference_text] + response_texts)
        similarities = cosine_similarity(tfidf_matrix[0:1], tfidf_matrix[1:])
        best_response_index = similarities.argmax()
        best_response_model = list(responses.keys())[best_response_index]
        best_response_text = response_texts[best_response_index]
        return {"best_response": {"model": best_response_model, "text": best_response_text}, "all_responses": responses}
    else:
        return {"best_response": {"model": None, "text": "No models loaded successfully."}, "all_responses": responses}


@app.post("/generate_multimodel")
async def api_generate_multimodel(request: Request):
    try:
        data = await request.json()
        message = data.get("message")
        if not message:
            raise HTTPException(status_code=400, detail="Missing message")
        response = await process_message(message)
        return JSONResponse(response)
    except HTTPException as e:
        raise e
    except Exception as e:
        return JSONResponse({"error": str(e)}, status_code=500)


async def startup():
    pass

async def shutdown():
    gc.collect()

app.add_event_handler("startup", startup)
app.add_event_handler("shutdown", shutdown)

def release_resources():
    try:
        torch.cuda.empty_cache()
        gc.collect()
    except Exception as e:
        print(f"Failed to release resources: {e}")

def resource_manager():
    MAX_RAM_PERCENT = 20
    MAX_CPU_PERCENT = 20
    MAX_GPU_PERCENT = 20
    MAX_RAM_MB = 2048

    while True:
        try:
            virtual_mem = psutil.virtual_memory()
            current_ram_percent = virtual_mem.percent
            current_ram_mb = virtual_mem.used / (1024 * 1024)

            if current_ram_percent > MAX_RAM_PERCENT or current_ram_mb > MAX_RAM_MB:
                release_resources()

            current_cpu_percent = psutil.cpu_percent()
            if current_cpu_percent > MAX_CPU_PERCENT:
                psutil.Process(os.getpid()).nice()

            if torch.cuda.is_available():
                gpu = torch.cuda.current_device()
                gpu_mem = torch.cuda.memory_percent(gpu)

                if gpu_mem > MAX_GPU_PERCENT:
                    release_resources()

        except Exception as e:
            print(f"Error in resource manager: {e}")

if __name__ == "__main__":
    import threading
    resource_thread = threading.Thread(target=resource_manager)
    resource_thread.daemon = True
    resource_thread.start()
    port = int(os.environ.get("PORT", 7860))
    uvicorn.run(app, host="0.0.0.0", port=port)