nemo_rl.models.generation.dynamo.managed_runtime#

Driver-owned lifecycle for a fixed Ray-managed Dynamo deployment.

Module Contents#

Classes#

ManagedDynamoRuntime

Own etcd, NATS, frontend, and a fixed Ray actor worker fleet.

Functions#

Data#

API#

nemo_rl.models.generation.dynamo.managed_runtime.LOGGER#

‘getLogger(…)’

nemo_rl.models.generation.dynamo.managed_runtime._managed_namespace() str#
class nemo_rl.models.generation.dynamo.managed_runtime.ManagedDynamoRuntime(
*,
cluster: nemo_rl.distributed.virtual_cluster.RayVirtualCluster,
config: dict[str, Any],
)#

Own etcd, NATS, frontend, and a fixed Ray actor worker fleet.

Initialization

property frontend_url: str#
start() None#

Start the complete managed service fleet.

static _stop_process(
process: subprocess.Popen | None,
label: str,
timeout_s: float = 15,
) None#
_service_env() dict[str, str]#
_frontend_env() dict[str, str]#
_start_etcd() None#
_wait_for_etcd() None#
_start_nats() None#
_wait_for_port(
port: int,
label: str,
process: subprocess.Popen | None = None,
) None#
_start_frontend() None#
_wait_for_frontend(expected_workers: int) None#
refit_workers() list[dict[str, Any]]#
validate_workers(
expected: list[dict[str, Any]],
) list[dict[str, Any]]#
_assert_services_alive() None#
shutdown() None#

Best-effort, idempotent teardown of every owned resource.