File size: 8,593 Bytes
5472531 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 |
"""
A model worker using Apple MLX
https://github.com/ml-explore/mlx-examples/tree/main/llms
Code based on vllm_worker https://github.com/lm-sys/FastChat/blob/main/fastchat/serve/vllm_worker.py
You must install MLX python:
pip install mlx-lm
"""
import argparse
import asyncio
import atexit
import json
from typing import List
import uuid
from fastapi import FastAPI, Request, BackgroundTasks
from fastapi.concurrency import run_in_threadpool
from fastapi.responses import StreamingResponse, JSONResponse
import uvicorn
from fastchat.serve.base_model_worker import BaseModelWorker
from fastchat.serve.model_worker import (
logger,
worker_id,
)
from fastchat.utils import get_context_length, is_partial_stop
import mlx.core as mx
from mlx_lm import load, generate
from mlx_lm.utils import generate_step
app = FastAPI()
class MLXWorker(BaseModelWorker):
def __init__(
self,
controller_addr: str,
worker_addr: str,
worker_id: str,
model_path: str,
model_names: List[str],
limit_worker_concurrency: int,
no_register: bool,
llm_engine: "MLX",
conv_template: str,
):
super().__init__(
controller_addr,
worker_addr,
worker_id,
model_path,
model_names,
limit_worker_concurrency,
conv_template,
)
logger.info(
f"Loading the model {self.model_names} on worker {worker_id}, worker type: MLX worker..."
)
self.model_name = model_path
self.mlx_model, self.mlx_tokenizer = load(model_path)
self.tokenizer = self.mlx_tokenizer
# self.context_len = get_context_length(
# llm_engine.engine.model_config.hf_config)
self.context_len = 2048 # hard code for now -- not sure how to get in MLX
if not no_register:
self.init_heart_beat()
async def generate_stream(self, params):
self.call_ct += 1
context = params.pop("prompt")
request_id = params.pop("request_id")
temperature = float(params.get("temperature", 1.0))
top_p = float(params.get("top_p", 1.0))
top_k = params.get("top_k", -1.0)
presence_penalty = float(params.get("presence_penalty", 0.0))
frequency_penalty = float(params.get("frequency_penalty", 0.0))
max_new_tokens = params.get("max_new_tokens", 256)
stop_str = params.get("stop", None)
stop_token_ids = params.get("stop_token_ids", None) or []
if self.tokenizer.eos_token_id is not None:
stop_token_ids.append(self.tokenizer.eos_token_id)
echo = params.get("echo", True)
use_beam_search = params.get("use_beam_search", False)
best_of = params.get("best_of", None)
# Handle stop_str
stop = set()
if isinstance(stop_str, str) and stop_str != "":
stop.add(stop_str)
elif isinstance(stop_str, list) and stop_str != []:
stop.update(stop_str)
for tid in stop_token_ids:
if tid is not None:
s = self.tokenizer.decode(tid)
if s != "":
stop.add(s)
print("Stop patterns: ", stop)
top_p = max(top_p, 1e-5)
if temperature <= 1e-5:
top_p = 1.0
tokens = []
skip = 0
context_mlx = mx.array(self.tokenizer.encode(context))
finish_reason = "length"
iterator = await run_in_threadpool(
generate_step, context_mlx, self.mlx_model, temperature
)
for i in range(max_new_tokens):
token = await run_in_threadpool(next, iterator)
if token == self.mlx_tokenizer.eos_token_id:
finish_reason = "stop"
break
tokens.append(token.item())
tokens_decoded = self.mlx_tokenizer.decode(tokens)
last_token_decoded = self.mlx_tokenizer.decode([token.item()])
skip = len(tokens_decoded)
partial_stop = any(is_partial_stop(tokens_decoded, i) for i in stop)
if partial_stop:
finish_reason = "stop"
break
ret = {
"text": tokens_decoded,
"error_code": 0,
"usage": {
"prompt_tokens": len(context),
"completion_tokens": len(tokens),
"total_tokens": len(context) + len(tokens),
},
"cumulative_logprob": [],
"finish_reason": None, # hard code for now
}
# print(ret)
yield (json.dumps(ret) + "\0").encode()
ret = {
"text": self.mlx_tokenizer.decode(tokens),
"error_code": 0,
"usage": {},
"cumulative_logprob": [],
"finish_reason": finish_reason,
}
yield (json.dumps(obj={**ret, **{"finish_reason": None}}) + "\0").encode()
yield (json.dumps(ret) + "\0").encode()
async def generate(self, params):
async for x in self.generate_stream(params):
pass
return json.loads(x[:-1].decode())
def release_worker_semaphore():
worker.semaphore.release()
def acquire_worker_semaphore():
if worker.semaphore is None:
worker.semaphore = asyncio.Semaphore(worker.limit_worker_concurrency)
return worker.semaphore.acquire()
def create_background_tasks(request_id):
async def abort_request() -> None:
print("trying to abort but not implemented")
background_tasks = BackgroundTasks()
background_tasks.add_task(release_worker_semaphore)
background_tasks.add_task(abort_request)
return background_tasks
@app.post("/worker_generate_stream")
async def api_generate_stream(request: Request):
params = await request.json()
await acquire_worker_semaphore()
request_id = uuid.uuid4()
params["request_id"] = str(request_id)
generator = worker.generate_stream(params)
background_tasks = create_background_tasks(request_id)
return StreamingResponse(generator, background=background_tasks)
@app.post("/worker_generate")
async def api_generate(request: Request):
params = await request.json()
await acquire_worker_semaphore()
request_id = uuid.uuid4()
params["request_id"] = str(request_id)
output = await worker.generate(params)
release_worker_semaphore()
# await engine.abort(request_id)
print("Trying to abort but not implemented")
return JSONResponse(output)
@app.post("/worker_get_status")
async def api_get_status(request: Request):
return worker.get_status()
@app.post("/count_token")
async def api_count_token(request: Request):
params = await request.json()
return worker.count_token(params)
@app.post("/worker_get_conv_template")
async def api_get_conv(request: Request):
return worker.get_conv_template()
@app.post("/model_details")
async def api_model_details(request: Request):
return {"context_length": worker.context_len}
worker = None
def cleanup_at_exit():
global worker
print("Cleaning up...")
del worker
atexit.register(cleanup_at_exit)
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="microsoft/phi-2")
parser.add_argument(
"--model-names",
type=lambda s: s.split(","),
help="Optional display comma separated names",
)
parser.add_argument(
"--conv-template", type=str, default=None, help="Conversation prompt template."
)
parser.add_argument(
"--trust_remote_code",
action="store_false",
default=True,
help="Trust remote code (e.g., from HuggingFace) when"
"downloading the model and tokenizer.",
)
args, unknown = parser.parse_known_args()
if args.model_path:
args.model = args.model_path
worker = MLXWorker(
args.controller_address,
args.worker_address,
worker_id,
args.model_path,
args.model_names,
1024,
False,
"MLX",
args.conv_template,
)
uvicorn.run(app, host=args.host, port=args.port, log_level="info")
|