`nemo_rl.models.generation.dynamo.config`#

Validated configuration for the managed Dynamo generation backend.

Module Contents#

Classes#

`DynamoWorkerArgs`	Structured arguments passed to every managed `dynamo.vllm` worker.
`DynamoFrontendArgs`	Structured arguments passed to the managed Dynamo frontend.
`DynamoCfg`	Driver-owned Dynamo service and worker-fleet configuration.
`DynamoVllmConfig`	Known vLLM settings consumed by `dynamo.vllm`.
`DynamoConfig`	Validated boundary for `policy.generation.backend=dynamo`.

Functions#

_require_nonempty_vllm_config

Data#

`VLLM_PACKED_BUFFER_SIZE_BYTES`
`VLLM_PACKED_NUM_BUFFERS`
`DYNAMO_VLLM_FLAGS`
`_VLLM_CFG_STRUCTURAL`
`_VLLM_CFG_MOVED`
`_VLLM_CFG_UNSUPPORTED`
`_VLLM_CFG_MANAGED_RUNTIME`
`_VLLM_CFG_INAPPLICABLE`
`_VLLM_SINGLE_RANK_ONLY_FIELDS`

API#

nemo_rl.models.generation.dynamo.config.VLLM_PACKED_BUFFER_SIZE_BYTES#: None

nemo_rl.models.generation.dynamo.config.VLLM_PACKED_NUM_BUFFERS#: 2

nemo_rl.models.generation.dynamo.config.DYNAMO_VLLM_FLAGS: dict[str, str]#: None

nemo_rl.models.generation.dynamo.config._VLLM_CFG_STRUCTURAL#: None

nemo_rl.models.generation.dynamo.config._VLLM_CFG_MOVED#: None

nemo_rl.models.generation.dynamo.config._VLLM_CFG_UNSUPPORTED#: None

nemo_rl.models.generation.dynamo.config._VLLM_CFG_MANAGED_RUNTIME#: None

nemo_rl.models.generation.dynamo.config._VLLM_CFG_INAPPLICABLE#: None

nemo_rl.models.generation.dynamo.config._VLLM_SINGLE_RANK_ONLY_FIELDS#: None

class nemo_rl.models.generation.dynamo.config.DynamoWorkerArgs#

Bases: pydantic.BaseModel

Structured arguments passed to every managed dynamo.vllm worker.

tool_call_parser: str | None#: None

reasoning_parser: str | None#: None

exclude_tools_when_tool_choice_none: bool#: None

enable_structural_tag: bool#: None

structural_tag_scope: Literal[auto, always]#: None

structural_tag_schema: Literal[auto, strict]#: None

custom_jinja_template: str | None#: None

endpoint_types: list[Literal[chat, completions]]#: None

extra_cli_args: list[str]#: None

_validate_endpoint_types() → nemo_rl.models.generation.dynamo.config.DynamoWorkerArgs#

class nemo_rl.models.generation.dynamo.config.DynamoFrontendArgs#

Bases: pydantic.BaseModel

Structured arguments passed to the managed Dynamo frontend.

tokenizer: Literal[default, fastokens]#: None

tokenizer_cache: bool#: None

tokenizer_cache_bytes: pydantic.PositiveInt#: None

router_mode: Literal[round-robin, random, power-of-two, kv, direct, least-loaded, device-aware-weighted]#: None

router_reset_states: bool#: None

extra_cli_args: list[str]#: None

class nemo_rl.models.generation.dynamo.config.DynamoCfg#

Bases: pydantic.BaseModel

Driver-owned Dynamo service and worker-fleet configuration.

engine: Literal[vllm]#: None

startup_timeout_s: pydantic.PositiveFloat#: None

request_timeout_s: pydantic.PositiveFloat#: None

control_timeout_s: pydantic.PositiveFloat#: None

worker_args: nemo_rl.models.generation.dynamo.config.DynamoWorkerArgs#: None

frontend_args: nemo_rl.models.generation.dynamo.config.DynamoFrontendArgs#: None

metrics_include_prefixes: list[str] | None#: None

metrics_exclude_prefixes: list[str] | None#: None

class nemo_rl.models.generation.dynamo.config.DynamoVllmConfig#

Bases: pydantic.BaseModel

Known vLLM settings consumed by dynamo.vllm.

Additional fields remain visible in model_extra so argument construction can warn rather than silently dropping an inherited vLLM setting.

async_engine: bool#: None

tensor_parallel_size: pydantic.PositiveInt#: None

pipeline_parallel_size: pydantic.PositiveInt#: None

expert_parallel_size: pydantic.PositiveInt#: None

gpu_memory_utilization: float#: None

max_model_len: pydantic.PositiveInt#: None

kv_cache_dtype: str#: None

load_format: str#: None

precision: str#: None

enforce_eager: bool#: None

expose_http_server: bool#: None

enable_vllm_metrics_logger: bool#: None

vllm_metrics_logger_interval: pydantic.PositiveFloat#: None

env_vars: dict[str, str] | None#: None

_validate_parallelism_and_precision() → nemo_rl.models.generation.dynamo.config.DynamoVllmConfig#

nemo_rl.models.generation.dynamo.config._require_nonempty_vllm_config(value: Any) → Any#

class nemo_rl.models.generation.dynamo.config.DynamoConfig#

Bases: pydantic.BaseModel

Validated boundary for policy.generation.backend=dynamo.

backend: Literal[dynamo]#: None

dynamo_cfg: nemo_rl.models.generation.dynamo.config.DynamoCfg#: None

vllm_cfg: Annotated[nemo_rl.models.generation.dynamo.config.DynamoVllmConfig, BeforeValidator(_require_nonempty_vllm_config)]#: None

vllm_kwargs: dict[str, Any]#: None

property engine_world_size: int#: Return the derived ranks in each single-node vLLM engine.

_validate_backend_boundary() → nemo_rl.models.generation.dynamo.config.DynamoConfig#

nemo_rl.models.generation.dynamo.config#

Module Contents#

Classes#

Functions#

Data#

API#

`nemo_rl.models.generation.dynamo.config`#