[Bugfix][CuMem] Make KV-cache wake cleanup tag-safe (#49857)
Signed-off-by: aoshen02 <aoshen02@users.noreply.github.com> Co-authored-by: aoshen02 <aoshen02@users.noreply.github.com>
aoshen02 committed
Jul 27, 2026 at 03:09 UTC
9e50e1037e268b4488bbd472bd5dfee4cc08bb75
1 file changed
+3
vllm/device_allocator/cumem.py
+3
@@ -291,6 +291,9 @@ class CuMemAllocator:
291
back to GPU memory. If None, all memory allocation will be loaded
292
back to GPU memory.
293
"""
294
+ gc.collect()
295
+ torch.accelerator.empty_cache()
296
+
297
for ptr, data in self.pointer_to_data.items():
298
if tags is None or data.tag in tags:
299
handle = data.handle