nemo_rl.models.generation.dynamo.worker_pool

Fixed Ray-managed pool of Dynamo vLLM subprocess owners.

Module Contents

Classes

FixedDynamoWorkerPool

Reserve inference GPUs and launch one worker per model-parallel group.

Functions

_vllm_port_for_node_slot

Return vLLM’s node-local scan base with 100 ports of headroom.

Data

_WORKER_FQN

API

nemo_rl.models.generation.dynamo.worker_pool._WORKER_FQN

‘nemo_rl.models.generation.dynamo.dynamo_worker.DynamoVllmWorker’

nemo_rl.models.generation.dynamo.worker_pool._vllm_port_for_node_slot(node_slot: int) → int

Return vLLM’s node-local scan base with 100 ports of headroom.

class nemo_rl.models.generation.dynamo.worker_pool.FixedDynamoWorkerPool(

*,

cluster: nemo_rl.distributed.virtual_cluster.RayVirtualCluster,

config: dict[str, Any],

namespace: str,

engine_world_size: int,

manager_env: dict[str, str],

startup_timeout_s: float,

)

Reserve inference GPUs and launch one worker per model-parallel group.

Initialization

property size: int

is_alive() → bool

Return whether every managed vLLM subprocess is still alive.

start() → None

refit_workers() → list[dict[str, Any]]

validate(

expected: list[dict[str, Any]],

) → list[dict[str, Any]]

shutdown() → None