def cluster_def(self):
cluster = {
TaskType.MASTER: 1,
}
is_distributed = False
environment = self.environment
if not environment:
return cluster, is_distributed
if environment.tensorflow:
return TensorflowSpecification.get_cluster_def(
cluster=cluster, tensorflow_config=environment.tensorflow)
if environment.horovod:
return HorovodSpecification.get_cluster_def(
cluster=cluster, horovod_config=environment.horovod)
if environment.mxnet:
return MXNetSpecification.get_cluster_def(
cluster=cluster, mxnet_config=environment.mxnet)
if environment.pytorch:
return PytorchSpecification.get_cluster_def(
cluster=cluster, pytorch_config=environment.pytorch)
# No specified framework, It should return default standalone mode cluster definition
return cluster, is_distributed
def handle_pytorch_experiment(experiment, spawner, response):
# Get the number of jobs this experiment started
master = response[TaskType.MASTER]
job_uuid = master['pod']['metadata']['labels']['job_uuid']
job_uuid = uuid.UUID(job_uuid)
create_job(job_uuid=job_uuid,
experiment=experiment,
definition=get_job_definition(master),
resources=spawner.spec.master_resources)
cluster, is_distributed, = spawner.spec.cluster_def
worker_resources = PytorchSpecification.get_worker_resources(
environment=spawner.spec.environment,
cluster=cluster,
is_distributed=is_distributed)
for i, worker in enumerate(response[TaskType.WORKER]):
job_uuid = worker['pod']['metadata']['labels']['job_uuid']
job_uuid = uuid.UUID(job_uuid)
create_job(job_uuid=job_uuid,
experiment=experiment,
definition=get_job_definition(worker),
role=TaskType.WORKER,
resources=worker_resources.get(i))
def handle_pytorch_experiment(experiment, spawner, response):
# Get the number of jobs this experiment started
master = response[TaskType.MASTER]
job_uuid = master['pod']['metadata']['labels']['job_uuid']
job_uuid = uuid.UUID(job_uuid)
create_job(job_uuid=job_uuid,
experiment=experiment,
definition=get_job_definition(master),
resources=spawner.spec.master_resources)
cluster, is_distributed, = spawner.spec.cluster_def
worker_resources = PytorchSpecification.get_worker_resources(
environment=spawner.spec.environment,
cluster=cluster,
is_distributed=is_distributed
)
for i, worker in enumerate(response[TaskType.WORKER]):
job_uuid = worker['pod']['metadata']['labels']['job_uuid']
job_uuid = uuid.UUID(job_uuid)
create_job(job_uuid=job_uuid,
experiment=experiment,
definition=get_job_definition(worker),
role=TaskType.WORKER,
resources=worker_resources.get(i))
def node_selectors(self):
cluster, is_distributed, = self.spec.cluster_def
worker_node_selectors = PytorchSpecification.get_worker_node_selectors(
environment=self.spec.environment,
cluster=cluster,
is_distributed=is_distributed
)
return {
TaskType.MASTER: {0: self.spec.master_node_selectors},
TaskType.WORKER: worker_node_selectors,
}
def node_selectors(self):
cluster, is_distributed, = self.spec.cluster_def
worker_node_selectors = PytorchSpecification.get_worker_node_selectors(
environment=self.spec.environment,
cluster=cluster,
is_distributed=is_distributed)
return {
TaskType.MASTER: {
0: self.spec.master_node_selectors
},
TaskType.WORKER: worker_node_selectors,
}
def affinities(self):
cluster, is_distributed, = self.spec.cluster_def
worker_affinities = PytorchSpecification.get_worker_affinities(
environment=self.spec.environment,
cluster=cluster,
is_distributed=is_distributed)
return {
TaskType.MASTER: {
0: self.spec.master_affinity
},
TaskType.WORKER: worker_affinities,
}
def total_resources(self):
environment = self.environment
if not environment:
return None
cluster, is_distributed = self.cluster_def
# Check if any framework is defined
if environment.tensorflow:
return TensorflowSpecification.get_total_resources(
master_resources=self.master_resources,
environment=environment,
cluster=cluster,
is_distributed=is_distributed)
if environment.horovod:
return HorovodSpecification.get_total_resources(
master_resources=self.master_resources,
environment=environment,
cluster=cluster,
is_distributed=is_distributed)
if environment.mxnet:
return MXNetSpecification.get_total_resources(
master_resources=self.master_resources,
environment=environment,
cluster=cluster,
is_distributed=is_distributed)
if environment.pytorch:
return PytorchSpecification.get_total_resources(
master_resources=self.master_resources,
environment=environment,
cluster=cluster,
is_distributed=is_distributed)
# default value is the master resources
return self.master_resources
def test_distributed_pytorch_passes(self):
plxfile = PolyaxonFile(
os.path.abspath('tests/fixtures/distributed_pytorch_file.yml'))
spec = plxfile.specification
assert spec.version == 1
assert spec.project.name == 'project1'
assert isinstance(spec.settings, SettingsConfig)
assert isinstance(spec.settings.logging, LoggingConfig)
assert spec.settings.matrix is None
assert isinstance(spec.environment, EnvironmentConfig)
assert spec.is_runnable
assert spec.framework == Frameworks.PYTORCH
assert spec.environment.pytorch.n_workers == 5
assert isinstance(spec.environment.resources, PodResourcesConfig)
assert isinstance(spec.environment.resources.cpu, K8SResourcesConfig)
assert spec.environment.resources.cpu.requests == 1
assert spec.environment.resources.cpu.limits == 2
assert isinstance(spec.environment.pytorch.default_worker_resources,
PodResourcesConfig)
assert isinstance(
spec.environment.pytorch.default_worker_resources.cpu,
K8SResourcesConfig)
assert spec.environment.pytorch.default_worker_resources.cpu.requests == 3
assert spec.environment.pytorch.default_worker_resources.cpu.limits == 3
assert isinstance(
spec.environment.pytorch.default_worker_resources.memory,
K8SResourcesConfig)
assert spec.environment.pytorch.default_worker_resources.memory.requests == 256
assert spec.environment.pytorch.default_worker_resources.memory.limits == 256
assert isinstance(spec.environment.pytorch.worker_resources[0],
PodResourcesConfig)
assert isinstance(spec.environment.pytorch.worker_resources[0].memory,
K8SResourcesConfig)
assert spec.environment.pytorch.worker_resources[0].index == 3
assert spec.environment.pytorch.worker_resources[
0].memory.requests == 300
assert spec.environment.pytorch.worker_resources[
0].memory.limits == 300
# check that properties for return list of configs and resources is working
cluster, is_distributed = spec.cluster_def
worker_resources = PytorchSpecification.get_worker_resources(
environment=spec.environment,
cluster=cluster,
is_distributed=is_distributed)
assert len(worker_resources) == spec.environment.pytorch.n_workers
assert set(worker_resources.values()) == {
spec.environment.pytorch.default_worker_resources,
spec.environment.pytorch.worker_resources[0]
}
# Check total resources
assert spec.total_resources == {
'cpu': {
'requests': 1 + 3 * 4,
'limits': 2 + 3 * 4
},
'memory': {
'requests': 300 + 256 * 4,
'limits': 300 + 256 * 4
},
'gpu': None
}
assert spec.cluster_def == ({
TaskType.MASTER: 1,
TaskType.WORKER: 5
}, True)
