[core] fix sleep mode in pytorch 2.6 (vllm-project#13456)

Signed-off-by: youkaichao <youkaichao@gmail.com>

Author: youkaichao

vllm/device_allocator/cumem.py

@@ -9,7 +9,7 @@
 # the only successful approach is to call cuda driver API in C.
 import dataclasses
 from contextlib import contextmanager
-from typing import Callable, Dict, Optional, Tuple, Union
+from typing import Any, Callable, Dict, Optional, Tuple, Union
 
 import torch
 
@@ -97,7 +97,7 @@ def use_memory_pool_with_allocator(
     new_alloc = get_pluggable_allocator(python_malloc_fn, python_free_func)
     mem_pool = torch.cuda.memory.MemPool(new_alloc._allocator)
     with torch.cuda.memory.use_mem_pool(mem_pool):
-        yield mem_pool
+        yield mem_pool, new_alloc
 
 
 class CuMemAllocator:
@@ -142,6 +142,7 @@ def get_instance() -> "CuMemAllocator":
     def __init__(self):
         self.pointer_to_data: Dict[int, AllocationData] = {}
         self.current_tag: str = CuMemAllocator.default_tag
+        self.allocator_and_pools: Dict[str, Any] = {}
 
     def python_malloc_callback(self, allocation_handle: HandleType) -> None:
         """
@@ -231,7 +232,13 @@ def use_memory_pool(self, tag: Optional[str] = None):
         old_tag = self.current_tag
         self.current_tag = tag
         with use_memory_pool_with_allocator(self.python_malloc_callback,
-                                            self.python_free_callback):
+                                            self.python_free_callback) as data:
+            # start to hit another PyTorch bug in PyTorch 2.6,
+            # possibly because of gc-related issue w.r.t. the allocator and
+            # the memory pool.
+            # to avoid the issue, we keep a reference of the data.
+            # see https://github.com/pytorch/pytorch/issues/146431 .
+            self.allocator_and_pools[tag] = data
             yield
             # PyTorch's bug, calling torch.cuda.empty_cache() will error
             # when using pluggable allocator, see