diff --git a/imperative/python/megengine/distributed/__init__.py b/imperative/python/megengine/distributed/__init__.py
index 55ee93c6047f26ea5412623c17b01a3a5e6b4d81..df50e1f01c12a2aa33fe0a7844de9d006e88a1d3 100644
--- a/imperative/python/megengine/distributed/__init__.py
+++ b/imperative/python/megengine/distributed/__init__.py
@@ -26,7 +26,7 @@ from .server import Client, Server
 @mproperty
 def backend(mod):
     r"""Get or set backend of collective communication.
-    Available backends are ['nccl', 'shm', 'rccl']
+    Available backends are ['nccl', 'rccl']
 
     Examples:
 
diff --git a/imperative/python/megengine/distributed/group.py b/imperative/python/megengine/distributed/group.py
index dd961293bedaecc3c1f80d5002a2326223f3425d..e3a7cac88cb5eec6cf4e3b9b6ffe6e85e98d9119 100644
--- a/imperative/python/megengine/distributed/group.py
+++ b/imperative/python/megengine/distributed/group.py
@@ -95,7 +95,7 @@ class Group:
 WORLD = Group([])
 
 _devices = {"gpu", "cuda", "rocm"}
-_backends = {"nccl", "rccl", "shm", "auto"}
+_backends = {"nccl", "rccl", "auto"}
 
 
 def init_process_group(
@@ -115,7 +115,7 @@ def init_process_group(
         world_size: total number of processes participating in the job.
         rank: rank of the current process.
         device: the GPU device id to bind this process to.
-        backend: communicator backend, currently support 'nccl' and 'shm'.
+        backend: communicator backend, currently support 'nccl' and 'rccl'.
     """
     physical_device_type = what_is_xpu() if device_type == "xpu" else device_type
     if not isinstance(master_ip, str):
diff --git a/imperative/python/megengine/distributed/helper.py b/imperative/python/megengine/distributed/helper.py
index 2530438b247bb7ffcd971338979458efd9172d1b..8683e58abf2bf62924ced117f71275975dfa12a7 100644
--- a/imperative/python/megengine/distributed/helper.py
+++ b/imperative/python/megengine/distributed/helper.py
@@ -205,10 +205,7 @@ class AllreduceCallback:
             assert _group._sd, "please call init_process_group first"
             backend = _group._sd.backend
         if backend == "auto":
-            if group.is_single_machine and not _check_enable_p2p():
-                backend = "shm"
-            else:
-                backend = "nccl"
+            backend = "nccl"
         self._backend = backend
 
     def _reset(self):
diff --git a/src/opr-mm/impl/megray_helper.cpp b/src/opr-mm/impl/megray_helper.cpp
index aa8f8c10c43b976125357f158b4c4178def3a35f..8e65e14f5681e1dabea381c51263644747b75c88 100644
--- a/src/opr-mm/impl/megray_helper.cpp
+++ b/src/opr-mm/impl/megray_helper.cpp
@@ -31,10 +31,8 @@ MegRay::Backend mgb::opr::get_megray_backend(const std::string& backend) {
         return MegRay::MEGRAY_RCCL;
     } else if (backend == "ucx") {
         return MegRay::MEGRAY_UCX;
-    } else if (backend == "shm") {
-        return MegRay::MEGRAY_SHM;
     } else {
-        mgb_throw(MegBrainError, "back CollectiveComm backend");
+        mgb_throw(MegBrainError, "bad CollectiveComm backend");
     }
 }