fetch_multiple_tensors#

TensorTransportManager.fetch_multiple_tensors(obj_id: str, tensor_transport_metadata: TensorTransportMetadata, communicator_metadata: CommunicatorMetadata, target_buffers: List[Any] | None = None) → FetchRequest[source]#

Initiate a fetch for multiple tensors without waiting for completion.

The default implementation calls recv_multiple_tensors synchronously and stores the result in a FetchRequest. Transports with true async capability should override both this method and wait_fetch_complete.

Call wait_fetch_complete(fetch_request) afterward to retrieve the tensors.

Parameters:

  • obj_id – The object ID for the related GPU object.

  • tensor_transport_metadata – The tensor transport metadata for the GPU object.

  • communicator_metadata – The communicator metadata for the send/recv operation.

  • target_buffers – Pre-allocated buffers to receive the tensors into if possible.

Returns:

A FetchRequest whose tensors field is already populated.