wip

Author: q10

fbgemm_gpu/src/memory_utils/memory_utils.cu

@@ -157,6 +157,13 @@ inline gpuMemLocation new_mem_location_from_device(const int device_id) {
   return deviceLoc;
 }
 
+inline gpuMemLocation new_mem_location_cpu() {
+  gpuMemLocation deviceLoc;
+  deviceLoc.type = cudaMemLocationTypeHost;
+  deviceLoc.id = cudaCpuDeviceId;
+  return deviceLoc;
+}
+
 } // namespace
 
 Tensor new_managed_tensor(
@@ -177,7 +184,7 @@ Tensor new_managed_tensor(
 #if CUDART_VERSION >= 13000
       // Starting with CUDA 13, the deviceId arg (int) is replaced with
       // cudaMemLocation (struct)
-      new_mem_location_from_device(cudaCpuDeviceId)
+      new_mem_location_cpu()
 #else
       cudaCpuDeviceId
 #endif