[Bugfix][ROCm] Fix cumem sleep and teardown (#46203)

Signed-off-by: pei.zhang <pei.zhang@amd.com>
Signed-off-by: Matthew Wong <Matthew.Wong2@amd.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Matthew Wong <Matthew.Wong2@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
peizhang56
2026-06-24 02:45:31 +08:00
committed by GitHub
co-authored by Cursor Matthew Wong Andreas Karatzas
parent 37a682d392
commit 68afd78897
7 changed files with 130 additions and 18 deletions
+4 -2
View File
@@ -3,14 +3,15 @@
import dataclasses
from contextlib import AbstractContextManager
from typing import Protocol
from typing import Protocol, TypeAlias
import torch
from vllm.platforms import current_platform
# py_device, py_size_or_aligned_size, py_ptr, py_handle
HandleType = tuple[int, int, int, int]
# py_handle has type list[int] on ROCm and int otherwise
HandleType: TypeAlias = tuple[int, int, int, list[int] | int]
@dataclasses.dataclass
@@ -18,6 +19,7 @@ class AllocationData:
handle: HandleType
tag: str
cpu_backup_tensor: torch.Tensor | None = None
is_asleep: bool = False
class MemAllocator(Protocol):