File size: 8,045 Bytes
6ef31de |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 |
"""
A model worker executes the model.
"""
import argparse
import asyncio
import dataclasses
import logging
import json
import os
import time
from typing import List, Union
import threading
import uuid
from fastapi import FastAPI, Request, BackgroundTasks
from fastapi.responses import StreamingResponse
import requests
try:
from transformers import (
AutoTokenizer,
AutoModelForCausalLM,
LlamaTokenizer,
AutoModel,
)
except ImportError:
from transformers import (
AutoTokenizer,
AutoModelForCausalLM,
LLaMATokenizer,
AutoModel,
)
import torch
import uvicorn
from fastchat.constants import WORKER_HEART_BEAT_INTERVAL
from fastchat.serve.inference import load_model, generate_stream
from fastchat.serve.serve_chatglm import chatglm_generate_stream
from fastchat.utils import build_logger, server_error_msg, pretty_print_semaphore
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_name,
device,
num_gpus,
max_gpu_memory,
load_8bit=False,
):
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_name = model_name or model_path.split("/")[-1]
self.device = device
logger.info(f"Loading the model {self.model_name} on worker {worker_id} ...")
self.model, self.tokenizer = load_model(
model_path, device, num_gpus, max_gpu_memory, load_8bit
)
if hasattr(self.model.config, "max_sequence_length"):
self.context_len = self.model.config.max_sequence_length
elif hasattr(self.model.config, "max_position_embeddings"):
self.context_len = self.model.config.max_position_embeddings
else:
self.context_len = 2048
is_chatglm = "chatglm" in str(type(self.model)).lower()
if is_chatglm:
self.generate_stream_func = chatglm_generate_stream
else:
self.generate_stream_func = generate_stream
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
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_name],
"speed": 1,
"queue_length": self.get_queue_length(),
}
def generate_stream_gate(self, params):
try:
for output in self.generate_stream_func(
self.model,
self.tokenizer,
params,
self.device,
self.context_len,
args.stream_interval,
):
ret = {
"text": output,
"error_code": 0,
}
yield json.dumps(ret).encode() + b"\0"
except torch.cuda.OutOfMemoryError:
ret = {
"text": server_error_msg,
"error_code": 1,
}
yield json.dumps(ret).encode() + b"\0"
app = FastAPI()
def release_model_semaphore():
model_semaphore.release()
@app.post("/worker_generate_stream")
async def api_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()
generator = worker.generate_stream_gate(params)
background_tasks = BackgroundTasks()
background_tasks.add_task(release_model_semaphore)
return StreamingResponse(generator, background=background_tasks)
@app.post("/worker_get_status")
async def api_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="facebook/opt-350m",
help="The path to the weights",
)
parser.add_argument("--model-name", type=str, help="Optional name")
parser.add_argument(
"--device", type=str, choices=["cpu", "cuda", "mps"], default="cuda"
)
parser.add_argument("--num-gpus", type=int, default=1)
parser.add_argument(
"--gpus",
type=str,
default=None,
help="A single GPU like 1 or multiple GPUs like 0,2"
)
parser.add_argument(
"--max-gpu-memory",
type=str,
help="The maximum memory per gpu. Use a string like '13Gib'",
)
parser.add_argument("--load-8bit", action="store_true")
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}")
if args.gpus:
if args.num_gpus and len(args.gpus.split(",")) < int(args.num_gpus):
raise ValueError(f"Larger --num-gpus ({args.num_gpus}) than --gpus {args.gpus}!")
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpus
worker = ModelWorker(
args.controller_address,
args.worker_address,
worker_id,
args.no_register,
args.model_path,
args.model_name,
args.device,
args.num_gpus,
args.max_gpu_memory,
args.load_8bit,
)
uvicorn.run(app, host=args.host, port=args.port, log_level="info")
|