def test_ulimit_invalid_type(self):
with pytest.raises(ValueError):
Ulimit(name=None)
with pytest.raises(ValueError):
Ulimit(name='hello', soft='123')
with pytest.raises(ValueError):
Ulimit(name='hello', hard='456')
def create_ulimits(limits):
ulimits = []
if limits['cputime']:
cpu = limits['cputime']
ulimits.append(Ulimit(name='cpu', soft=cpu, hard=cpu))
if 'file_size' in limits:
fsize = limits['file_size']
ulimits.append(Ulimit(name='fsize', soft=fsize, hard=fsize))
return ulimits or None
def test_compare_ulimits_neg(self):
self.fake_data['params']['dimensions'] = {
'ulimits': {
'nofile': {
'soft': 131072,
'hard': 131072
}
}
}
ulimits_nofile = Ulimit(name='nofile', soft=131072, hard=131072)
container_info = dict()
container_info['HostConfig'] = {
'CpuPeriod': 0,
'KernelMemory': 0,
'Memory': 0,
'CpuQuota': 0,
'CpusetCpus': '',
'CpuShares': 0,
'BlkioWeight': 0,
'CpusetMems': '',
'MemorySwap': 0,
'MemoryReservation': 0,
'Ulimits': [ulimits_nofile]
}
self.dw = get_DockerWorker(self.fake_data['params'])
self.assertFalse(self.dw.compare_dimensions(container_info))
def test_create_host_config_obj_ulimit(self):
ulimit_dct = Ulimit(name='nofile', soft=8096)
config = create_host_config(ulimits=[ulimit_dct],
version=DEFAULT_DOCKER_API_VERSION)
assert 'Ulimits' in config
assert len(config['Ulimits']) == 1
ulimit_obj = config['Ulimits'][0]
assert isinstance(ulimit_obj, Ulimit)
assert ulimit_obj == ulimit_dct
def test_create_host_config_obj_ulimit(self):
ulimit_dct = Ulimit(name='nofile', soft=8096)
config = create_host_config(ulimits=[ulimit_dct],
version=DEFAULT_DOCKER_API_VERSION)
self.assertIn('Ulimits', config)
self.assertEqual(len(config['Ulimits']), 1)
ulimit_obj = config['Ulimits'][0]
self.assertTrue(isinstance(ulimit_obj, Ulimit))
self.assertEqual(ulimit_obj, ulimit_dct)
def start(self) -> None:
"""
Start Triton Server Container
"""
devices = [
DeviceRequest(capabilities=[["gpu"]], device_ids=self._devices),
]
LOGGER.info(
f"Triton environment: {json.dumps(self._environment, indent=4)}")
LOGGER.info(f"Starting Triton container {self.name}.")
self._container = self._docker_client.containers.run(
image=self._image,
name=self.name,
device_requests=devices,
detach=True,
tty=True,
shm_size=self._shm_size,
ulimits=[
Ulimit(name="memlock", soft=-1, hard=-1),
Ulimit(name="stack", soft=67108864, hard=67108864),
],
volumes=self._volumes,
environment=self._environment,
network_mode=self._network,
auto_remove=True,
ipc_mode="host",
)
LOGGER.info(f"Triton command:")
LOGGER.info(f" {self._command}")
LOGGER.info(f"Starting Triton Server {self.name}.")
self._triton_exec = self._docker_api_client.exec_create(
container=self._container.id,
cmd=self._command,
)
stream_generator = self._docker_api_client.exec_start(
exec_id=self._triton_exec["Id"], stream=True)
self._logging_thread = Thread(target=TritonServerContainer._logging,
args=(self, stream_generator),
daemon=True)
self._logging_thread.start()
def test_ulimit_invalid_type(self):
self.assertRaises(ValueError, lambda: Ulimit(name=None))
self.assertRaises(ValueError, lambda: Ulimit(name='hello', soft='123'))
self.assertRaises(ValueError, lambda: Ulimit(name='hello', hard='456'))
def serve(
save_path: Union[Path, str],
device: str = 'cpu',
name: str = None,
batch_size: int = 16,
) -> Container:
"""Serve the given model save path in a Docker container.
Args:
save_path (Union[Path, str]): Saved path to the model.
device (str): Device name. E.g.: cpu, cuda, cuda:1.
name (str): Container name. Default to None.
batch_size (int): Batch size for passing to serving containers.
Returns:
Container: Docker container object created.
"""
info = parse_path(Path(save_path))
architecture: str = info['architecture']
engine: Engine = info['engine']
cuda, device_num = get_device(device)
docker_client = docker.from_env()
# set mount
mounts = [
Mount(target=f'/models/{architecture}',
source=str(info['base_dir']),
type='bind',
read_only=True)
]
common_kwargs = remove_dict_null({
'detach': True,
'auto_remove': True,
'mounts': mounts,
'name': name
})
environment = dict()
if cuda:
common_kwargs['runtime'] = 'nvidia'
environment['CUDA_DEVICE_ORDER'] = 'PCI_BUS_ID'
environment['CUDA_VISIBLE_DEVICES'] = device_num
if engine == Engine.TFS:
# Tensorflow Serving 2.2.0 has the issue: https://github.com/tensorflow/serving/issues/1663
docker_tag = '2.1.0-gpu' if cuda else '2.1.0'
ports = {'8501': config.TFS_HTTP_PORT, '8500': config.TFS_GRPC_PORT}
environment['MODEL_NAME'] = architecture
container = docker_client.containers.run(
f'tensorflow/serving:{docker_tag}',
environment=environment,
ports=ports,
**common_kwargs)
elif engine == Engine.TORCHSCRIPT:
docker_tag = 'latest-gpu' if cuda else 'latest'
ports = {
'8000': config.TORCHSCRIPT_HTTP_PORT,
'8001': config.TORCHSCRIPT_GRPC_PORT
}
environment['MODEL_NAME'] = architecture
container = docker_client.containers.run(
f'mlmodelci/pytorch-serving:{docker_tag}',
environment=environment,
ports=ports,
**common_kwargs)
elif engine == Engine.ONNX:
docker_tag = 'latest-gpu' if cuda else 'latest'
ports = {'8000': config.ONNX_HTTP_PORT, '8001': config.ONNX_GRPC_PORT}
environment['MODEL_NAME'] = architecture
container = docker_client.containers.run(
f'mlmodelci/onnx-serving:{docker_tag}',
environment=environment,
ports=ports,
**common_kwargs)
elif engine == Engine.TRT:
if not cuda:
raise RuntimeError(
'TensorRT cannot be run without CUDA. Please specify a CUDA device.'
)
ports = {
'8000': config.TRT_HTTP_PORT,
'8001': config.TRT_GRPC_PORT,
'8002': config.TRT_PROMETHEUS_PORT
}
ulimits = [
Ulimit(name='memlock', soft=-1, hard=-1),
Ulimit(name='stack', soft=67100864, hard=67100864)
]
trt_kwargs = {'ulimits': ulimits, 'shm_size': '1G'}
container = docker_client.containers.run(
f'nvcr.io/nvidia/tensorrtserver:19.10-py3',
'trtserver --model-repository=/models',
environment=environment,
ports=ports,
**common_kwargs,
**trt_kwargs,
)
else:
raise RuntimeError(
f'Not able to serve model with path `{str(save_path)}`.')
return container
DOCKER_PARAMETERS = {
"image": "pl:latest",
"auto_remove": True,
"cpu_period": 1000,
"cpu_shares": 1024,
"cpu_quota": 0,
"cpuset_cpus": "0",
"detach": True,
"environment": {},
"mem_limit": "100m",
"memswap_limit": "200m",
"network_mode": "none",
"network_disabled": True,
# "storage_opt": {},
"tty": True,
"ulimits": [Ulimit(name="core", soft=0, hard=0)],
}
# Check if any of the above settings are override by a config.py file.
logger = logging.getLogger(__name__)
try:
from config import * # noqa
logger.info("Using config.py...")
except ModuleNotFoundError:
logger.info("No config file found")
del logger
# Override some settings from testing purpose
if TESTING:
DOCKER_COUNT = 5
def _run_start_start(self):
"""Start the container and wait for it to finish starting"""
# Get the value of the URL prefix that the app will have to the outside.
url_prefix = reverse(
"dockerapps:docker-proxy",
kwargs={
"project": self.process.project.sodar_uuid,
"image": self.image.sodar_uuid,
"process": self.image.process.sodar_uuid,
"path": "",
},
)
with transaction.atomic():
self.process.refresh_from_db()
if self.process.state in (STATE_IDLE, STATE_FAILED):
self.job.add_log_entry("Starting container for %s:%s..." %
(self.image.repository, self.image.tag))
# Build environment, interpreting placeholders.
environment = {}
for entry in self.process.environment:
if "__KIOSC_URL_PREFIX__" in entry["value"]:
environment[entry["name"]] = entry["value"].replace(
"__KIOSC_URL_PREFIX__", url_prefix)
else:
environment[entry["name"]] = entry["value"]
# Create and start the Docker container, update database record.
host_config = self.cli.create_host_config()
container = self.cli.create_container(
detach=True,
image=self.image.image_id,
environment=environment,
command=shlex.split(self.process.command)
if self.process.command else None,
ports=[self.process.internal_port],
host_config=self.cli.create_host_config(
port_bindings={
self.process.internal_port: self.process.host_port
},
ulimits=[
Ulimit(
name="nofile",
soft=settings.
KIOSC_DOCKER_MAX_ULIMIT_NOFILE_SOFT,
hard=settings.
KIOSC_DOCKER_MAX_ULIMIT_NOFILE_HARD,
)
],
),
)
self.cli.start(container=container.get("Id"))
self.process.container_id = container.get("Id")
self.process.state = STATE_STARTING
self.process.save()
else:
self.job.add_log_entry(
"Process state is %s, not attempting to start" %
self.process.state)
self.job.add_log_entry("Waiting for container to start...")
timeout_start = time.time()
while time.time() < timeout_start + self.timeout:
if (self.cli.inspect_container(self.process.container_id).get(
"State", {}).get("Running")):
self.job.add_log_entry("Container is running...")
with transaction.atomic():
self.process.refresh_from_db()
self.process.state = STATE_RUNNING
self.process.save()
break
time.sleep(self.sleep_time)
else:
raise RuntimeError("Container did not start on time")
def create_container(self,
name,
image,
ram,
working_directory,
gpus=None,
environment=None,
enable_fuse=False):
"""
Creates a docker container with the given arguments. This docker container is running endlessly until
container.stop() is called.
If nvidia gpus are specified, the nvidia runtime is used, if available. Otherwise a device request for nvidia
gpus is added.
:param name: The name of the container
:type name: str
:param image: The image to use for this container
:type image: str
:param ram: The ram limit for this container in megabytes
:type ram: int
:param working_directory: The working directory inside the docker container
:type working_directory: str
:param gpus: A specification of gpus to enable in this docker container
:type gpus: List[GPUDevice]
:param environment: A dictionary containing environment variables, which should be set inside the container
:type environment: Dict[str, Any]
:param enable_fuse: If True, SYS_ADMIN capabilities are granted for this container and /dev/fuse is mounted
:type enable_fuse: bool
:return: The created container
:rtype: Container
:raise RuntimeNotSupportedError: If the specified runtime is not installed on the docker host
"""
if environment is None:
environment = {}
mem_limit = None
if ram is not None:
mem_limit = '{}m'.format(ram)
gpu_ids = None
if gpus:
set_nvidia_environment_variables(
environment, map(lambda gpu: gpu.device_id, gpus))
gpu_ids = [gpu.device_id for gpu in gpus]
# enable fuse
devices = []
capabilities = []
if enable_fuse:
devices.append('/dev/fuse')
capabilities.append('SYS_ADMIN')
container = create_container_with_gpus(
self._client,
image,
command='/bin/sh',
gpus=gpu_ids,
available_runtimes=self._runtimes,
name=name,
user='******',
working_dir=working_directory,
mem_limit=mem_limit,
memswap_limit=mem_limit,
environment=environment,
cap_add=capabilities,
devices=devices,
ulimits=[
Ulimit(name='nofile', soft=NOFILE_LIMIT, hard=NOFILE_LIMIT)
],
# needed to run the container endlessly
tty=True,
stdin_open=True,
auto_remove=False,
)
container.start()
return container
def get_ulimits_config(config):
return list(map(lambda ulimit:
Ulimit(name=ulimit['name'], soft=ulimit['soft'], hard=ulimit['hard']), config))
