0

I'm using this code as a baseline. But in this xla_client, I can only use the GPU resources. However I want to use the CPU swap(unified) memory as a GPU resources. How could I create the xla_client with CPU memory support version? Is there any way..? I really want to create the GPU client with CPU memory available.

self.distributed_client = (
            xla_client._xla.get_distributed_runtime_client(
            server_address, host_id, use_coordination_service=False))
logger.debug(
        f"{host_id}: Trying to connect to xla runtime at {server_address}")
self.distributed_client.connect()
logger.debug(
        f"{host_id}: Success to connect to xla runtime at {server_address}")
if global_config.backend == "gpu":
    self.backend = xla_client.make_gpu_client(self.distributed_client,
                                                  node_id=host_id) 
else:
    raise NotImplementedError(
            f"backend {global_config.backend} is not supported")
self.local_devices = self.backend.local_devices()
YuGyoung Yun
  • 25
  • 4
  • I'm having trouble understanding your question, but if you're asking whether you can set up a GPU such that it has direct runtime access to CPU memory resources, I think the answer is No. – jakevdp Mar 19 '23 at 15:53

0 Answers0