diff --git a/CMakeLists.txt b/CMakeLists.txt
index 9e8ccfa8718..06f267ee53a 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -112,6 +112,8 @@ endif()
 #
 # spinloop extension (pure CXX; must stay above the non-CUDA device branch so
 # CPU builds define the target before the early return)
+# This extension requires SABI 3.11 since it relies on Py_buffer support. Loading
+# failure is handled gracefully on vLLM side for lower Python versions.
 #
 set(VLLM_SPINLOOP_EXT_SRC "csrc/spinloop.cpp")
 set(SPINLOOP_COMPILE_FLAGS "")
diff --git a/vllm/distributed/device_communicators/shm_broadcast.py b/vllm/distributed/device_communicators/shm_broadcast.py
index dc7e6d151a4..9482568461c 100644
--- a/vllm/distributed/device_communicators/shm_broadcast.py
+++ b/vllm/distributed/device_communicators/shm_broadcast.py
@@ -38,9 +38,19 @@ from vllm.utils.network_utils import (
     is_valid_ipv6_address,
 )
 
-if envs.VLLM_USE_SPINLOOP_EXT:
-    from vllm.spinloop import spinloop
+logger = init_logger(__name__)
 
+
+SPINLOOP_EXT_ENABLED = False
+if envs.VLLM_USE_SPINLOOP_EXT:
+    try:
+        from vllm.spinloop import spinloop
+
+        SPINLOOP_EXT_ENABLED = True
+    except ImportError:
+        logger.warning(
+            "spinloop extension could not be loaded, disabling VLLM_USE_SPINLOOP_EXT!"
+        )
 SPINLOOP_TIMEOUT_SECONDS = 0.1
 
 if TYPE_CHECKING:
@@ -82,9 +92,6 @@ def to_bytes_big(value: int, size: int) -> bytes:
     return value.to_bytes(size, byteorder="big")
 
 
-logger = init_logger(__name__)
-
-
 LONG_WAIT_TIME_LOG_MSG = (
     "No available shared memory broadcast block found "
     "in %d seconds. This typically happens "
@@ -552,7 +559,7 @@ class MessageQueue:
                     written_flag = metadata_buffer[0]
                     return not (written_flag and read_count != self.buffer.n_reader)
 
-                if envs.VLLM_USE_SPINLOOP_EXT and not check():
+                if SPINLOOP_EXT_ENABLED and not check():
                     spinloop(metadata_buffer, check, timeout=SPINLOOP_TIMEOUT_SECONDS)
 
                 if not check():
@@ -673,7 +680,7 @@ class MessageQueue:
                     written_flag = metadata_buffer[0]
                     return not (not written_flag or read_flag)
 
-                if envs.VLLM_USE_SPINLOOP_EXT and not check():
+                if SPINLOOP_EXT_ENABLED and not check():
                     spinloop(
                         metadata_buffer[0 : self.local_reader_rank + 1],
                         check,