rlmesh.torch.RemoteModel

classExperimentalfrom rlmesh.torch import RemoteModel[source]

Experimental Torch-backed handle to a served model (policy).

RemoteModel(
    address: str | None = None,
    *,
    host: str | None = None,
    port: int | None = None,
    path: str | None = None,
    transport: Transport | None = None,
    connect_timeout_seconds: float | None = None,
    request_timeout_seconds: float | None = None,
) -> None

Bind it to an env with rlmesh.session(model, env) to get a rlmesh.Session whose predict accepts and returns Torch values, driven symmetrically with the env.

Parameters

Name	Type	Description
`address`	`str \| None`	Model endpoint address such as `"tcp://127.0.0.1:5556"`.
`host`	`str \| None`	TCP host helper used when `address` is omitted.
`port`	`int \| None`	TCP port helper used when `address` is omitted.
`path`	`str \| None`	Unix socket path helper used when `address` is omitted.
`transport`	`Transport \| None`	Explicit transport selector.

Examples

>>> import rlmesh
>>> from rlmesh.torch import RemoteEnv, RemoteModel
>>> env = RemoteEnv("127.0.0.1:5555")
>>> sess = rlmesh.session(RemoteModel("127.0.0.1:5556"), env)
>>> obs, _ = sess.reset(seed=0)
>>> action = sess.predict(obs)
>>> obs, reward, terminated, truncated, _ = sess.step(action)

Attributes

addressModel endpoint address this handle dials.

Methods

session()Bind this served policy to env and return a rlmesh.Session.

address

property[source]

property address: str

Model endpoint address this handle dials.

session

function[source]

session(
    env: EnvTarget,
    *,
    instruction: str | None = None,
    close_env: bool = False,
    trust_entrypoints: bool | None = None,
    execution_horizon: int = 1,
    view: ViewArg = None,
    workflow_edition: str | None = None,
) -> Session[Any, Any]

Bind this served policy to env and return a rlmesh.Session.

env is an env client (e.g. RemoteEnv/SandboxEnv) exposing an env_contract; that contract – including the env’s adapter tags – is sent to the model server, which resolves its adapter for this env. instruction and trust_entrypoints apply to local models and are rejected here (the served model owns its adapter). view opts into the built-in live viewer over this client’s own env loop (see rlmesh.run()).

execution_horizon (> 1) opts a chunk-capable served model into action chunking: the model emits its native chunk, and this client replays a prefix of it one action per step open-loop (skipping the RPC), re-planning every execution_horizon steps. The model must define predict_chunk; otherwise it re-plans every step.

workflow_edition declares this runtime’s edition for the session, ahead of the process and project declarations. It also overrides the runtime declaration used to connect the env handle; the env server’s own declaration still constrains the session floor.