nemo_rl.models.generation.dynamo.dynamo_worker

Ray actors used by the fixed, managed Dynamo vLLM fleet.

Module Contents

Classes

DynamoGpuReservation	Hold one placement-group GPU while a sibling actor owns the engine.
DynamoVllmWorker	Own one dynamo.vllm subprocess for a model-parallel GPU group.

API

class nemo_rl.models.generation.dynamo.dynamo_worker.DynamoGpuReservation

Hold one placement-group GPU while a sibling actor owns the engine.

Initialization

metadata() → dict[str, Any]

select_free_port(

*,

port_range_low: int,

port_range_high: int,

excluded_ports: list[int],

) → int

Select an unused node-local port from a half-open range.

register_process_group(pid: int) → bool

Record the colocated worker process group for failure cleanup.

cleanup_process_group() → bool

Best-effort cleanup if the subprocess-owning actor died first.

class nemo_rl.models.generation.dynamo.dynamo_worker.DynamoVllmWorker(

config: dict[str, Any],

*,

namespace: str,

group_name: str,

cuda_devices: list[int],

system_port: int,

vllm_port: int,

manager_env: dict[str, str],

startup_timeout_s: float,

seed: int,

cleanup_reservation: ray.actor.ActorHandle,

)

Own one dynamo.vllm subprocess for a model-parallel GPU group.

Initialization

property system_url: str

static _validate_argv(

dynamo_python: str,

argv: list[str],

env: dict[str, str],

*,

timeout_s: float,

) → None

_wait_for_system_port(timeout_s: float) → None

metadata() → dict[str, Any]

is_alive() → bool

_stop_process() → None

shutdown() → bool