[Bugfix] Add strong reference to CUDA pluggable allocator callbacks (#23477)

Signed-off-by: 22quinn <33176974+22quinn@users.noreply.github.com>
Signed-off-by: youkaichao <youkaichao@gmail.com>
Co-authored-by: Eric Marcus <eric.marcus@kaiko.ai>
Co-authored-by: youkaichao <youkaichao@gmail.com>

vllm/device_allocator/cumem.py

@@ -152,8 +152,13 @@ class CuMemAllocator:
         self.pointer_to_data: dict[int, AllocationData] = {}
         self.current_tag: str = CuMemAllocator.default_tag
         self.allocator_and_pools: dict[str, Any] = {}
+        # Creating strong references to the two callbacks here to prevent
+        # these ephemeral bound-method objects being garbage collected.
+        # See discussions in https://github.com/vllm-project/vllm/pull/22724
+        self.python_malloc_callback = self._python_malloc_callback
+        self.python_free_callback = self._python_free_callback
 
-    def python_malloc_callback(self, allocation_handle: HandleType) -> None:
+    def _python_malloc_callback(self, allocation_handle: HandleType) -> None:
         """
         Internal method to store the allocation data
         when memory is allocated in the memory pool."""
@@ -162,7 +167,7 @@ class CuMemAllocator:
             allocation_handle, self.current_tag)
         return
 
-    def python_free_callback(self, ptr: int) -> HandleType:
+    def _python_free_callback(self, ptr: int) -> HandleType:
         """
         Internal method to look up the allocation data
         when memory is freed in the memory pool."""
@@ -212,9 +217,9 @@ class CuMemAllocator:
     def wake_up(self, tags: Optional[list[str]] = None) -> None:
         """
         Wake up the allocator from sleep mode.
-        All data that is previously offloaded will be loaded back to GPU 
+        All data that is previously offloaded will be loaded back to GPU
         memory, and the rest of the data will have empty memory.
-        
+
         :param tags: The tags of the memory allocation that will be loaded
             back to GPU memory. If None, all memory allocation will be loaded
             back to GPU memory.