rlmesh.specs

Public RLMesh-native spec classes backed by the Rust extension.

import rlmesh.specs

A contract is the handshake an endpoint hands a client: env id, spaces, render mode, metadata, and instance count. You read one off a server or client (env_contract); you rarely build one. SpaceSpec is the wire-safe description of one space. ServeOptions and Tensor, exported from the top-level package, complete the native value boundary.

EnvContract

For Gymnasium envs, id comes from env.spec.id and falls back to "UnknownEnv-v1". Spaces come from observation_space / action_space, or the single_* variants on a vector env.

contract = env.env_contract              # from a server or a remote client
print(contract.id, contract.num_envs)
print(contract.observation_space.kind)
payload = contract.to_dict()             # JSON-serializable, nested specs included

SpaceSpec

SpaceSpec describes a space on the wire; a Space wrapper is the Python object that samples, seeds, and validates values from it. Wrap one with space_from_spec(), or convert it with to_gymnasium_space() where code expects a Gymnasium space.

ServeOptions

The constructor is keyword-only; pass the result to EnvServer(..., options=...).

options = rlmesh.ServeOptions(allow_remote_shutdown=True, idle_timeout_seconds=300.0)
server = rlmesh.EnvServer(env, "127.0.0.1:5555", options=options)

The bearer token that gates an endpoint is not a Python ServeOptions field; it is reachable only from Rust (rlmesh::ServeOptions.token) and C (RlmeshServeOptions.token).

Tensor

Tensor is a validated transport container, not an ndarray: immutable element bytes with shape, dtype, and strides, plus DLPack and buffer-protocol edges. Compute, slicing, and broadcasting belong to the frameworks. device is always "cpu" today.

To leave the native boundary, use rlmesh.numpy.asarray() for a writable NumPy copy (or numpy.from_dlpack(tensor) for a zero-copy read-only view), rlmesh.torch.as_tensor() for a Torch tensor, or rlmesh.jax.asarray() for a JAX array.

Navigation