from pydantic import BaseModel from llama_cpp import Llama from concurrent.futures import ThreadPoolExecutor, as_completed import re import httpx from spaces import GPU import asyncio 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 = {} def load_model(self, model_config): if model_config['name'] not in self.models: try: self.models[model_config['name']] = Llama.from_pretrained(repo_id=model_config['repo_id'], filename=model_config['filename']) except Exception as e: print(f"Error loading model {model_config['name']}: {e}") def load_all_models(self): with ThreadPoolExecutor() as executor: for config in model_configs: executor.submit(self.load_model, config) return self.models model_manager = ModelManager() global_data['models'] = model_manager.load_all_models() class ChatRequest(BaseModel): message: str 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 = [] seen_lines = set() for line in lines: if line not in seen_lines: unique_lines.append(line) seen_lines.add(line) return '\n'.join(unique_lines) @GPU(duration=0) def generate_model_response(model, inputs): try: response = model(inputs) return remove_duplicates(response['choices'][0]['text']) except Exception as e: print(f"Error generating model response: {e}") return "" async def handle_request(request): if request.method == "POST" and request.url.path == "/generate": try: chat_request = ChatRequest(**request.json()) inputs = normalize_input(chat_request.message) with ThreadPoolExecutor() as executor: futures = [ executor.submit(generate_model_response, model, inputs) for model in global_data['models'].values() ] responses = [{'model': model_name, 'response': future.result()} for model_name, future in zip(global_data['models'].keys(), as_completed(futures))] unique_responses = remove_repetitive_responses(responses) return httpx.Response(status_code=200, json=unique_responses) except Exception as e: print(f"Error handling request: {e}") return httpx.Response(status_code=500, json={"error": f"Error handling request: {e}"}) else: return httpx.Response(status_code=404, text="Not Found") async def run_server(port: int): async with httpx.AsyncClient(base_url=f"http://localhost:{port}") as client: while True: request = await client.get("/") # You might need to adjust this based on your expected requests response = await handle_request(request) print(f"Received request: {request}") print(f"Sending response: {response}") await asyncio.sleep(1) # Adjust the sleep duration as needed def remove_repetitive_responses(responses): unique_responses = {} for response in responses: if response['model'] not in unique_responses: unique_responses[response['model']] = response['response'] return unique_responses if __name__ == "__main__": port = 7860 asyncio.run(run_server(port))