app.py

from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
from llama_cpp import Llama
from concurrent.futures import ThreadPoolExecutor, as_completed
import uvicorn
from dotenv import load_dotenv
import re
import huggingface_hub
import spaces

load_dotenv()

app = FastAPI()

global_data = {
    'models': [],
    'tokens': {
        'eos': 'eos_token',
        'pad': 'pad_token',
        'padding': 'padding_token',
        'unk': 'unk_token',
        'bos': 'bos_token',
        'sep': 'sep_token',
        'cls': 'cls_token',
        'mask': 'mask_token'
    }
}

model_configs = [
    {"repo_id": "Ffftdtd5dtft/gpt2-xl-Q2_K-GGUF", "filename": "gpt2-xl-q2_k.gguf", "name": "GPT-2 XL"},
    {"repo_id": "Ffftdtd5dtft/Meta-Llama-3.1-8B-Instruct-Q2_K-GGUF", "filename": "meta-llama-3.1-8b-instruct-q2_k.gguf", "name": "Meta Llama 3.1-8B Instruct"},
    {"repo_id": "Ffftdtd5dtft/gemma-2-9b-it-Q2_K-GGUF", "filename": "gemma-2-9b-it-q2_k.gguf", "name": "Gemma 2-9B IT"},
    {"repo_id": "Ffftdtd5dtft/gemma-2-27b-Q2_K-GGUF", "filename": "gemma-2-27b-q2_k.gguf", "name": "Gemma 2-27B"},
    {"repo_id": "Ffftdtd5dtft/Phi-3-mini-128k-instruct-Q2_K-GGUF", "filename": "phi-3-mini-128k-instruct-q2_k.gguf", "name": "Phi-3 Mini 128K Instruct"},
    {"repo_id": "Ffftdtd5dtft/Meta-Llama-3.1-8B-Q2_K-GGUF", "filename": "meta-llama-3.1-8b-q2_k.gguf", "name": "Meta Llama 3.1-8B"},
    {"repo_id": "Ffftdtd5dtft/Qwen2-7B-Instruct-Q2_K-GGUF", "filename": "qwen2-7b-instruct-q2_k.gguf", "name": "Qwen2 7B Instruct"},
    {"repo_id": "Ffftdtd5dtft/starcoder2-3b-Q2_K-GGUF", "filename": "starcoder2-3b-q2_k.gguf", "name": "Starcoder2 3B"},
    {"repo_id": "Ffftdtd5dtft/Qwen2-1.5B-Instruct-Q2_K-GGUF", "filename": "qwen2-1.5b-instruct-q2_k.gguf", "name": "Qwen2 1.5B Instruct"},
    {"repo_id": "Ffftdtd5dtft/Meta-Llama-3.1-70B-Q2_K-GGUF", "filename": "meta-llama-3.1-70b-q2_k.gguf", "name": "Meta Llama 3.1-70B"},
    {"repo_id": "Ffftdtd5dtft/Mistral-Nemo-Instruct-2407-Q2_K-GGUF", "filename": "mistral-nemo-instruct-2407-q2_k.gguf", "name": "Mistral Nemo Instruct 2407"},
    {"repo_id": "Ffftdtd5dtft/Hermes-3-Llama-3.1-8B-IQ1_S-GGUF", "filename": "hermes-3-llama-3.1-8b-iq1_s-imat.gguf", "name": "Hermes 3 Llama 3.1-8B"},
    {"repo_id": "Ffftdtd5dtft/Phi-3.5-mini-instruct-Q2_K-GGUF", "filename": "phi-3.5-mini-instruct-q2_k.gguf", "name": "Phi 3.5 Mini Instruct"},
    {"repo_id": "Ffftdtd5dtft/Meta-Llama-3.1-70B-Instruct-Q2_K-GGUF", "filename": "meta-llama-3.1-70b-instruct-q2_k.gguf", "name": "Meta Llama 3.1-70B Instruct"},
    {"repo_id": "Ffftdtd5dtft/codegemma-2b-IQ1_S-GGUF", "filename": "codegemma-2b-iq1_s-imat.gguf", "name": "Codegemma 2B"},
    {"repo_id": "Ffftdtd5dtft/Phi-3-mini-128k-instruct-IQ2_XXS-GGUF", "filename": "phi-3-mini-128k-instruct-iq2_xxs-imat.gguf", "name": "Phi 3 Mini 128K Instruct XXS"},
    {"repo_id": "Ffftdtd5dtft/TinyLlama-1.1B-Chat-v1.0-IQ1_S-GGUF", "filename": "tinyllama-1.1b-chat-v1.0-iq1_s-imat.gguf", "name": "TinyLlama 1.1B Chat"},
    {"repo_id": "Ffftdtd5dtft/Mistral-NeMo-Minitron-8B-Base-IQ1_S-GGUF", "filename": "mistral-nemo-minitron-8b-base-iq1_s-imat.gguf", "name": "Mistral NeMo Minitron 8B Base"},
    {"repo_id": "Ffftdtd5dtft/Mistral-Nemo-Instruct-2407-Q2_K-GGUF", "filename": "mistral-nemo-instruct-2407-q2_k.gguf", "name": "Mistral Nemo Instruct 2407"}
]

class ModelManager:
    def __init__(self):
        self.models = []
        self.loaded = False

    def load_model(self, model_config):
        try:
            return {"model": Llama.from_pretrained(repo_id=model_config['repo_id'], filename=model_config['filename']), "name": model_config['name']}
        except Exception as e:
            print(f"Error loading model {model_config['name']}: {e}")
            return None
    
    def load_all_models(self):
        if self.loaded:
            return self.models
        
        with ThreadPoolExecutor() as executor:
            futures = [executor.submit(self.load_model, config) for config in model_configs]
            models = []
            for future in as_completed(futures):
                model = future.result()
                if model:
                    models.append(model)
        
        self.models = models
        self.loaded = True
        return self.models

model_manager = ModelManager()
global_data['models'] = model_manager.load_all_models()

class ChatRequest(BaseModel):
    message: str
    top_k: int = 50
    top_p: float = 0.95
    temperature: float = 0.7

def generate_chat_response(request, model_data):
    try:
        user_input = normalize_input(request.message)
        llm = model_data['model']
        response = llm.create_chat_completion(
            messages=[{"role": "user", "content": user_input}],
            top_k=request.top_k,
            top_p=request.top_p,
            temperature=request.temperature
        )
        reply = response['choices'][0]['message']['content']
        return {"response": reply, "literal": user_input, "model_name": model_data['name']}
    except Exception as e:
        return {"response": f"Error: {str(e)}", "literal": user_input, "model_name": model_data['name']}

def normalize_input(input_text):
    return input_text.strip()

def remove_duplicates(text):
    text = re.sub(r'(Hello there, how are you\? \[/INST\]){2,}', 'Hello there, how are you? [/INST]', text)
    text = re.sub(r'(How are you\? \[/INST\]){2,}', 'How are you? [/INST]', text)
    text = text.replace('[/INST]', '')
    lines = text.split('\n')
    unique_lines = list(dict.fromkeys(lines))
    return '\n'.join(unique_lines).strip()

def remove_repetitive_responses(responses):
    seen = set()
    unique_responses = []
    for response in responses:
        normalized_response = remove_duplicates(response['response'])
        if normalized_response not in seen:
            unique_responses.append(response)
            seen.add(normalized_response)
    return unique_responses

def select_best_response(responses):
    if not responses:
        return ""
    responses = remove_repetitive_responses(responses)
    return max(set(responses), key=lambda x: x['response'].count("user"))

@app.post("/generate")
def generate_chat(request: ChatRequest):
    responses = []
    num_models = len(global_data['models'])
    
    with ThreadPoolExecutor() as executor:
        futures = [executor.submit(generate_chat_response, request, model_data) for model_data in global_data['models']]
        for future in as_completed(futures):
            try:
                response = future.result()
                responses.append(response)
            except Exception as e:
                print(f"Error in response generation: {e}")
    
    if not responses:
        raise HTTPException(status_code=500, detail="Error: No responses generated.")
    
    best_response = select_best_response(responses)
    
    return {
        "best_response": best_response,
        "all_responses": responses
    }

if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=7860)