InferenceClient

class torchrl.modules.inference_server.InferenceClient(transport: InferenceTransport)

Actor-side handle for an InferenceServer.

Wraps a transport’s submit() so that calling client(td) looks like a regular synchronous policy call, while the actual computation is batched on the server.

Parameters:

transport (InferenceTransport) – the transport shared with the server.

Example

>>> client = transport.client()
>>> td_out = client(td_in)          # blocking
>>> future = client.submit(td_in)   # non-blocking
>>> td_out = future.result()

submit(td: TensorDictBase) → Future[TensorDictBase]

Submit a request and return a Future immediately.