import argparse import asyncio import json import time import threading import uuid import requests import torch import uvicorn import transformers from fastapi import FastAPI, Request, BackgroundTasks from fastapi.responses import StreamingResponse from functools import partial from transformers import TextIteratorStreamer from threading import Thread from serve.constants import WORKER_HEART_BEAT_INTERVAL from serve.utils import (build_logger, server_error_msg, pretty_print_semaphore) from serve.builder import load_pretrained_model from serve.mm_utils import process_images, load_image_from_base64, tokenizer_image_token, get_model_name_from_path, \ KeywordsStoppingCriteria from serve.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN GB = 1 << 30 worker_id = str(uuid.uuid4())[:6] logger = build_logger("model_worker", f"model_worker_{worker_id}.log") global_counter = 0 model_semaphore = None def heart_beat_worker(controller): while True: time.sleep(WORKER_HEART_BEAT_INTERVAL) controller.send_heart_beat() class ModelWorker: def __init__(self, controller_addr, worker_addr, worker_id, no_register, model_path, model_base, model_name, model_type, load_8bit, load_4bit, device): self.controller_addr = controller_addr self.worker_addr = worker_addr self.worker_id = worker_id if model_path.endswith("/"): model_path = model_path[:-1] if model_name is None: self.model_name = get_model_name_from_path(model_path) else: self.model_name = model_name self.device = device logger.info(f"Loading the model {self.model_name} on worker {worker_id} ...") transformers.logging.disable_progress_bar() self.tokenizer, self.model, self.image_processor, self.context_len = load_pretrained_model( model_path, model_base, self.model_name, model_type, load_8bit, load_4bit, device=self.device) self.is_multimodal = True if not no_register: self.register_to_controller() self.heart_beat_thread = threading.Thread( target=heart_beat_worker, args=(self,)) self.heart_beat_thread.start() def register_to_controller(self): logger.info("Register to controller") url = self.controller_addr + "/register_worker" data = { "worker_name": self.worker_addr, "check_heart_beat": True, "worker_status": self.get_status() } r = requests.post(url, json=data) assert r.status_code == 200 def send_heart_beat(self): logger.info(f"Send heart beat. Models: {[self.model_name]}. " f"Semaphore: {pretty_print_semaphore(model_semaphore)}. " f"global_counter: {global_counter}") url = self.controller_addr + "/receive_heart_beat" while True: try: ret = requests.post(url, json={ "worker_name": self.worker_addr, "queue_length": self.get_queue_length()}, timeout=5) exist = ret.json()["exist"] break except requests.exceptions.RequestException as e: logger.error(f"heart beat error: {e}") time.sleep(5) if not exist: self.register_to_controller() def get_queue_length(self): if model_semaphore is None: return 0 else: return args.limit_model_concurrency - model_semaphore._value + (len( model_semaphore._waiters) if model_semaphore._waiters is not None else 0) def get_status(self): return { "model_names": [self.model_name], "speed": 1, "queue_length": self.get_queue_length(), } @torch.inference_mode() def generate_stream(self, params): tokenizer, model, image_processor = self.tokenizer, self.model, self.image_processor prompt = params["prompt"] ori_prompt = prompt images = params.get("images", None) num_image_tokens = 0 if images is not None and len(images) > 0 and self.is_multimodal: if len(images) > 0: if len(images) != prompt.count(DEFAULT_IMAGE_TOKEN): raise ValueError("Number of images does not match number of tokens in prompt") images = [load_image_from_base64(image) for image in images] images = process_images(images, image_processor, model.config) print(f"----> process_images {images}") print(f"----> process_images sum {torch.sum(images)}") if type(images) is list: images = [image.to(self.model.device, dtype=model.dtype) for image in images] else: images = images.to(self.model.device, dtype=model.dtype) replace_token = DEFAULT_IMAGE_TOKEN prompt = prompt.replace(DEFAULT_IMAGE_TOKEN, replace_token) num_image_tokens = prompt.count(replace_token) * model.get_vision_tower().num_patches else: images = None image_args = {"images": images} else: images = None image_args = {} temperature = float(params.get("temperature", 1.0)) top_p = float(params.get("top_p", 1.0)) max_context_length = getattr(model.config, 'max_position_embeddings', 2048) max_new_tokens = min(int(params.get("max_new_tokens", 256)), 1024) stop_str = params.get("stop", None) do_sample = True if temperature > 0.001 else False input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).to( self.device) keywords = [stop_str] stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids) streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=15) max_new_tokens = min(max_new_tokens, max_context_length - input_ids.shape[-1] - num_image_tokens) if max_new_tokens < 1: yield json.dumps({"text": ori_prompt + "Exceeds max token length. Please start a new conversation, thanks.", "error_code": 0}).encode() + b"\0" return print("max_new_tokens", max_new_tokens) print("start!") thread = Thread(target=model.generate, kwargs=dict( inputs=input_ids, do_sample=do_sample, temperature=temperature, top_p=top_p, max_new_tokens=max_new_tokens, streamer=streamer, stopping_criteria=[stopping_criteria], use_cache=True, **image_args )) thread.start() generated_text = ori_prompt for new_text in streamer: if generated_text and not generated_text.endswith(' '): generated_text += ' ' generated_text += new_text if generated_text.endswith(stop_str): generated_text = generated_text[:-len(stop_str)] logger.info(f"new_text: {new_text}") yield json.dumps({"text": generated_text, "error_code": 0}).encode() + b"\0" def generate_stream_gate(self, params): try: for x in self.generate_stream(params): yield x except ValueError as e: print("Caught ValueError:", e) ret = { "text": server_error_msg, "error_code": 1, } yield json.dumps(ret).encode() + b"\0" except torch.cuda.CudaError as e: print("Caught torch.cuda.CudaError:", e) ret = { "text": server_error_msg, "error_code": 1, } yield json.dumps(ret).encode() + b"\0" except Exception as e: print("Caught Unknown Error", e) ret = { "text": server_error_msg, "error_code": 1, } yield json.dumps(ret).encode() + b"\0" app = FastAPI() def release_model_semaphore(fn=None): model_semaphore.release() if fn is not None: fn() @app.post("/worker_generate_stream") async def generate_stream(request: Request): global model_semaphore, global_counter global_counter += 1 params = await request.json() if model_semaphore is None: model_semaphore = asyncio.Semaphore(args.limit_model_concurrency) await model_semaphore.acquire() worker.send_heart_beat() generator = worker.generate_stream_gate(params) background_tasks = BackgroundTasks() background_tasks.add_task(partial(release_model_semaphore, fn=worker.send_heart_beat)) return StreamingResponse(generator, background=background_tasks) @app.post("/worker_get_status") async def get_status(request: Request): return worker.get_status() if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument("--host", type=str, default="localhost") parser.add_argument("--port", type=int, default=21002) parser.add_argument("--worker-address", type=str, default="http://localhost:21002") parser.add_argument("--controller-address", type=str, default="http://localhost:21001") parser.add_argument("--model-path", type=str, default=None) parser.add_argument("--model-base", type=str, default=None) parser.add_argument("--model-name", type=str) parser.add_argument("--model-type", type=str, default=None) parser.add_argument("--device", type=str, default="cuda") parser.add_argument("--multi-modal", action="store_true", help="Multimodal mode is automatically detected with model name.") parser.add_argument("--limit-model-concurrency", type=int, default=5) parser.add_argument("--stream-interval", type=int, default=1) parser.add_argument("--no-register", action="store_true") parser.add_argument("--load-8bit", action="store_true") parser.add_argument("--load-4bit", action="store_true") args = parser.parse_args() logger.info(f"args: {args}") if args.multi_modal: logger.warning("Multimodal mode is automatically detected with model name.") worker = ModelWorker(args.controller_address, args.worker_address, worker_id, args.no_register, args.model_path, args.model_base, args.model_name, args.model_type, args.load_8bit, args.load_4bit, args.device) log_config = uvicorn.config.LOGGING_CONFIG log_config['handlers']['default']['stream'] = 'ext://sys.stdout' uvicorn.run(app, host=args.host, port=args.port, log_level="info")