LLaVA-Plus-Codebase/serve/ram_worker.py

"""
A model worker executes the model.
"""
import sys, os
sys.path.append(os.path.join(os.path.dirname(__file__), ".."))

import argparse
import asyncio
import dataclasses
import logging
import json
import os
import sys
import time
from typing import List, Tuple, Union
import threading
import uuid

from io import BytesIO
import base64

from fastapi import FastAPI, Request, BackgroundTasks
from fastapi.responses import StreamingResponse, JSONResponse
import numpy as np
import requests
from PIL import Image

sys.path.append('Tag2Text')
from Tag2Text.models import tag2text
from Tag2Text import inference_ram
import torchvision.transforms as TS

try:
    from transformers import (
        AutoTokenizer,
        AutoModelForCausalLM,
        LlamaTokenizer,
        AutoModel,
    )
except ImportError:
    from transformers import (
        AutoTokenizer,
        AutoModelForCausalLM,
        LLaMATokenizer,
        AutoModel,
    )
from transformers import AutoProcessor, Blip2ForConditionalGeneration

import torch
import torch.nn.functional as F
import uvicorn

from serve.constants import WORKER_HEART_BEAT_INTERVAL, ErrorCode, SERVER_ERROR_MSG
from serve.utils import build_logger, pretty_print_semaphore

GB = 1 << 30


now_file_name = os.__file__
logdir = "logs/workers/"
os.makedirs(logdir, exist_ok=True)
logfile = os.path.join(logdir, f"{now_file_name}.log")

worker_id = str(uuid.uuid4())[:6]
logger = build_logger(now_file_name, logfile)
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_names,
        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]
        self.model_names = model_names or [model_path.split("/")[-1]]
        self.device = device

        # load model
        logger.info(f"Loading the model {self.model_names} on worker {worker_id} ...")
        self.ram_model = tag2text.ram(pretrained=model_path,
                                        image_size=384,
                                        vit='swin_l')
        self.ram_model.to(device)
        self.ram_model.eval()

        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_names}. "
            f"Semaphore: {pretty_print_semaphore(model_semaphore)}. "
            f"global_counter: {global_counter}. "
            f"worker_id: {worker_id}. "
        )

        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
            or model_semaphore._value is None
            or model_semaphore._waiters is None
        ):
            return 0
        else:
            return (
                args.limit_model_concurrency
                - model_semaphore._value
                + len(model_semaphore._waiters)
            )

    def get_status(self):
        return {
            "model_names": self.model_names,
            "speed": 1,
            "queue_length": self.get_queue_length(),
        }

    def load_image(self, image_path: str) -> Tuple[np.array, torch.Tensor]:
        

        if os.path.exists(image_path):
            image_source = Image.open(image_path).convert("RGB")
        else:
            # base64 coding
            image_source = Image.open(BytesIO(base64.b64decode(image_path))).convert("RGB")

        normalize = TS.Normalize(mean=[0.485, 0.456, 0.406],
                                     std=[0.229, 0.224, 0.225])
        transform = TS.Compose([
                    TS.Resize((384, 384)),
                    TS.ToTensor(), normalize
                ])
        image_source = transform(image_source)
        return image_source

    def generate_stream_func(self, model, params, device):
        # get inputs
        image_path = params["image"]

        # load image and run models
        image = self.load_image(image_path).unsqueeze(0).to(device)

        # tagging:
        res = inference_ram.inference(image , self.ram_model)

        tags=res[0].replace(' |', ',')
        tags_chinese=res[1].replace(' |', ',')

        _, __, h, w = image.shape
        pred_dict = {
            "tags": tags,
            "tags_chinese": tags_chinese,
            "size": [h, w],  # H,W
        }

        return pred_dict

    def generate_gate(self, params):
        try:

            ret = {"text": "", "error_code": 0}
            ret = self.generate_stream_func(
                self.ram_model,
                params,
                self.device,
            )
        except torch.cuda.OutOfMemoryError as e:
            ret = {
                "text": f"{SERVER_ERROR_MSG}\n\n({e})",
                "error_code": ErrorCode.CUDA_OUT_OF_MEMORY,
            }
        except (ValueError, RuntimeError) as e:
            ret = {
                "text": f"{SERVER_ERROR_MSG}\n\n({e})",
                "error_code": ErrorCode.INTERNAL_ERROR,
            }
        return ret

    @torch.inference_mode()
    def get_embeddings(self, params):
        try:
            tokenizer = self.tokenizer
            is_llama = "llama" in str(
                type(self.model)
            )  # vicuna support batch inference
            is_chatglm = "chatglm" in str(type(self.model))
            is_t5 = "t5" in str(type(self.model))
            if is_llama:
                encoding = tokenizer.batch_encode_plus(
                    params["input"], padding=True, return_tensors="pt"
                )
                input_ids = encoding["input_ids"].to(self.device)
                attention_mask = encoding["attention_mask"].to(self.device)
                model_output = self.model(
                    input_ids, attention_mask, output_hidden_states=True
                )
                data = model_output.hidden_states[-1]
                mask = attention_mask.unsqueeze(-1).expand(data.size()).float()
                masked_embeddings = data * mask
                sum_embeddings = torch.sum(masked_embeddings, dim=1)
                seq_length = torch.sum(mask, dim=1)
                embedding = sum_embeddings / seq_length
                normalized_embeddings = F.normalize(embedding, p=2, dim=1)
                ret = {
                    "embedding": normalized_embeddings.tolist(),
                    "token_num": torch.sum(attention_mask).item(),
                }
            else:
                embedding = []
                token_num = 0
                for text in params["input"]:
                    input_ids = tokenizer.encode(text, return_tensors="pt").to(
                        self.device
                    )
                    if is_t5:
                        model_output = self.model(
                            input_ids, decoder_input_ids=input_ids
                        )
                    else:
                        model_output = self.model(input_ids, output_hidden_states=True)
                    if is_chatglm:
                        data = (model_output.hidden_states[-1].transpose(0, 1))[0]
                    elif is_t5:
                        data = model_output.encoder_last_hidden_state[0]
                    else:
                        data = model_output.hidden_states[-1][0]
                    data = F.normalize(torch.mean(data, dim=0), p=2, dim=0)
                    embedding.append(data.tolist())
                    token_num += len(input_ids[0])
                ret = {
                    "embedding": embedding,
                    "token_num": token_num,
                }
        except torch.cuda.OutOfMemoryError as e:
            ret = {
                "text": f"{SERVER_ERROR_MSG}\n\n({e})",
                "error_code": ErrorCode.CUDA_OUT_OF_MEMORY,
            }
        except (ValueError, RuntimeError) as e:
            ret = {
                "text": f"{SERVER_ERROR_MSG}\n\n({e})",
                "error_code": ErrorCode.INTERNAL_ERROR,
            }
        return ret


app = FastAPI()


def release_model_semaphore():
    model_semaphore.release()


def acquire_model_semaphore():
    global model_semaphore, global_counter
    global_counter += 1
    if model_semaphore is None:
        model_semaphore = asyncio.Semaphore(args.limit_model_concurrency)
    return model_semaphore.acquire()


def create_background_tasks():
    background_tasks = BackgroundTasks()
    background_tasks.add_task(release_model_semaphore)
    return background_tasks



@app.post("/worker_generate")
async def api_generate(request: Request):
    params = await request.json()
    await acquire_model_semaphore()
    output = worker.generate_gate(params)
    release_model_semaphore()
    return JSONResponse(output)


@app.post("/worker_get_status")
async def api_get_status(request: Request):
    return worker.get_status()





@app.post("/model_details")
async def model_details(request: Request):
    return {"context_length": worker.context_len}


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--host", type=str, default="localhost")
    parser.add_argument("--port", type=int, default=21009)
    parser.add_argument("--worker-address", type=str, default="http://localhost:21009")
    parser.add_argument(
        "--controller-address", type=str, default="http://localhost:21001"
    )

    parser.add_argument(
        "--model-path", type=str, default="ram_swin_large_14m.pth"
    )

    parser.add_argument(
        "--model-names",
        default="ram",
        type=lambda s: s.split(","),
        help="Optional display comma separated names",
    )
    parser.add_argument("--device", type=str, default="cuda")
    parser.add_argument("--limit-model-concurrency", type=int, default=5)
    parser.add_argument("--stream-interval", type=int, default=2)
    parser.add_argument("--no-register", action="store_true")
    args = parser.parse_args()
    logger.info(f"args: {args}")


    worker = ModelWorker(
        args.controller_address,
        args.worker_address,
        worker_id,
        args.no_register,
        args.model_path,
        args.model_names,
        args.device,
    )
    uvicorn.run(app, host=args.host, port=args.port, log_level="info")