def create_pytorch_experiment_jobs(experiment, spawner):
master_job_uuid = spawner.job_uuids[TaskType.MASTER][0]
create_job(job_uuid=master_job_uuid,
experiment=experiment,
resources=spawner.spec.master_resources,
node_selector=spawner.spec.master_node_selector,
affinity=spawner.spec.master_affinity,
tolerations=spawner.spec.master_tolerations)
cluster, is_distributed = spawner.spec.cluster_def
environment = spawner.spec.config.pytorch
worker_resources = PytorchSpecification.get_worker_resources(
environment=environment,
cluster=cluster,
is_distributed=is_distributed
)
worker_node_selectors = PytorchSpecification.get_worker_node_selectors(
environment=environment,
cluster=cluster,
is_distributed=is_distributed
)
worker_affinities = PytorchSpecification.get_worker_affinities(
environment=environment,
cluster=cluster,
is_distributed=is_distributed
)
worker_tolerations = PytorchSpecification.get_worker_tolerations(
environment=environment,
cluster=cluster,
is_distributed=is_distributed
)
for i, worker_job_uuid in enumerate(spawner.job_uuids[TaskType.WORKER]):
create_job(job_uuid=worker_job_uuid,
experiment=experiment,
role=TaskType.WORKER,
sequence=i,
resources=worker_resources.get(i),
node_selector=worker_node_selectors.get(i),
affinity=worker_affinities.get(i),
tolerations=worker_tolerations.get(i))
def tolerations(self):
cluster, is_distributed, = self.spec.cluster_def
worker_tolerations = PytorchSpecification.get_worker_tolerations(
environment=self.spec.config.pytorch,
cluster=cluster,
is_distributed=is_distributed
)
return {
TaskType.MASTER: {0: self.spec.master_tolerations},
TaskType.WORKER: worker_tolerations,
}