rlmesh.torch

The Torch backend (experimental): the same clients, models, and sandbox sessions as the top-level package, with tensor leaves decoded to Torch tensors.

pip install "rlmesh[torch]"

Memory sharing and mutation

Decoded observations are owned, writable copies, so a model can normalize in place (img.div_(255)) without touching the wire buffer. as_tensor() is the zero-copy opt-in: the result shares memory with the RLMesh tensor over DLPack.

• Decode uses torch.utils.dlpack.from_dlpack; bool tensors fall back to a buffer copy before Torch 2.2.
• uint16, uint32, and uint64 need Torch 2.3 or newer.
• from_tensor() detaches, moves to CPU, and exports over DLPack; NumPy is not required.

Classes

• EnvFactoryExperimentalTorch-backed EnvFactory: served envs speak torch.
• ModelExperimentalExperimental Torch-backed model: predict works in Torch values.
• RemoteEnvExperimentalExperimental Torch-backed remote client for one environment.
• RemoteModelExperimentalExperimental Torch-backed handle to a served model (policy).
• RemoteVectorEnvExperimentalExperimental Torch-backed remote client for vectorized environments.
• SandboxBuildExperimentalBuild-from-source infrastructure for a sandbox image.
• SandboxEnvExperimentalExperimental Torch-backed owned sandbox session for one environment.
• SandboxInfoExperimentalInformation about a running RLMesh sandbox container.
• SandboxModelExperimentalA model served from an isolated container.
• SandboxRuntimeExperimentalContainer run-time settings for a sandbox -- docker run flags.
• SandboxVectorEnvExperimentalExperimental Torch-backed owned sandbox session for vectorized environments.

Functions

• as_tensor()ExperimentalReturn a Torch tensor view or copy of an RLMesh tensor.
• ensure_available()ExperimentalRaise if Torch is not installed.
• from_tensor()ExperimentalEncode a Torch tensor as an RLMesh value.
• space_from_spec()ExperimentalCreate a Torch-adapted space wrapper for a native space spec.

Type aliases

• TorchValueExperimentalPrimitive, Torch tensor, or nested RLMesh-compatible value.