fix(ml): better model unloading (#3340)

* restart process on inactivity

* formatting

* always update `last_called`

* load models sequentially

* renamed variable, updated docs

* formatting

* made poll env name consistent with model ttl env
This commit is contained in:
Mert
2023-11-16 21:42:44 -05:00
committed by GitHub
parent 98f87c6548
commit a6af4892e3
3 changed files with 42 additions and 11 deletions

View File

@@ -1,5 +1,9 @@
import asyncio
import gc
import os
import sys
import threading
import time
from concurrent.futures import ThreadPoolExecutor
from typing import Any
from zipfile import BadZipFile
@@ -34,7 +38,10 @@ def init_state() -> None:
)
# asyncio is a huge bottleneck for performance, so we use a thread pool to run blocking code
app.state.thread_pool = ThreadPoolExecutor(settings.request_threads) if settings.request_threads > 0 else None
app.state.locks = {model_type: threading.Lock() for model_type in ModelType}
app.state.lock = threading.Lock()
app.state.last_called = None
if settings.model_ttl > 0 and settings.model_ttl_poll_s > 0:
asyncio.ensure_future(idle_shutdown_task())
log.info(f"Initialized request thread pool with {settings.request_threads} threads.")
@@ -79,9 +86,9 @@ async def predict(
async def run(model: InferenceModel, inputs: Any) -> Any:
app.state.last_called = time.time()
if app.state.thread_pool is None:
return model.predict(inputs)
return await asyncio.get_running_loop().run_in_executor(app.state.thread_pool, model.predict, inputs)
@@ -90,7 +97,7 @@ async def load(model: InferenceModel) -> InferenceModel:
return model
def _load() -> None:
with app.state.locks[model.model_type]:
with app.state.lock:
model.load()
loop = asyncio.get_running_loop()
@@ -113,3 +120,27 @@ async def load(model: InferenceModel) -> InferenceModel:
else:
await loop.run_in_executor(app.state.thread_pool, _load)
return model
async def idle_shutdown_task() -> None:
while True:
log.debug("Checking for inactivity...")
if app.state.last_called is not None and time.time() - app.state.last_called > settings.model_ttl:
log.info("Shutting down due to inactivity.")
loop = asyncio.get_running_loop()
for task in asyncio.all_tasks(loop):
if task is not asyncio.current_task():
try:
task.cancel()
except asyncio.CancelledError:
pass
sys.stderr.close()
sys.stdout.close()
sys.stdout = sys.stderr = open(os.devnull, "w")
try:
await app.state.model_cache.cache.clear()
gc.collect()
loop.stop()
except asyncio.CancelledError:
pass
await asyncio.sleep(settings.model_ttl_poll_s)