Build XCCL as default and make XCCL the default distributed backend for XPU

Author: Chao1Han

CMakeLists.txt

@@ -259,7 +259,7 @@ option(USE_NATIVE_ARCH "Use -march=native" OFF)
 cmake_dependent_option(USE_MPS "Use MPS for macOS build" ON "MPS_FOUND" OFF)
 cmake_dependent_option(USE_NCCL "Use NCCL" ON
                        "USE_CUDA OR USE_ROCM;UNIX;NOT APPLE" OFF)
-cmake_dependent_option(USE_XCCL "Use XCCL" OFF
+cmake_dependent_option(USE_XCCL "Use XCCL" ON
                        "USE_XPU;UNIX;NOT APPLE" OFF)
 cmake_dependent_option(USE_RCCL "Use RCCL" ON USE_NCCL OFF)
 cmake_dependent_option(USE_STATIC_NCCL "Use static NCCL" OFF "USE_NCCL" OFF)

torch/distributed/distributed_c10d.py

@@ -340,7 +340,7 @@ def register_backend(
 
         if devices is not None:
             for device in devices:
-                if device != "cpu" and device != "cuda":
+                if device != "cpu" and device != "cuda" and device != "xpu":
                     Backend.default_device_backend_map[device] = name.lower()
         Backend.backend_type_map[name.lower()] = ProcessGroup.BackendType.CUSTOM