ray.experimental.TensorTransportManager.extract_tensor_transport_metadata#

abstract TensorTransportManager.extract_tensor_transport_metadata(obj_id: str, rdt_object: List[Any]) → TensorTransportMetadata[source]#

Implement this method to create the TensorTransportMetadata you defined previously. Ray calls this on the source actor immediately after the actor task creates the result tensors. Implement this to:

  1. Record tensor shapes, dtypes, and devices.

  2. Perform any transport-specific tensor registration such as registering memory for RDMA.

  3. Store any handles or identifiers needed for the transfer.

Parameters:

  • obj_id – The ID of the RDT object to extract the tensor transport metadata from.

  • rdt_object – The RDT object to extract the tensor transport metadata from.

Returns:

The tensor transport metadata.

Return type:

TensorTransportMetadata