Skip to content

[bugfix]add synchronize on ascend platform #485

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
ygwpz merged 3 commits into from
Dec 9, 2025
Merged

[bugfix]add synchronize on ascend platform #485

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
2fad1b9
add synchronize on ascend platform
qyh111 Dec 8, 2025
f75ba8d
use self.synchronize
qyh111 Dec 8, 2025
03a8992
move synchronize
qyh111 Dec 8, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 10 additions & 7 deletions ucm/integration/vllm/ucm_connector.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,6 @@
from vllm.distributed.parallel_state import get_tp_group, get_world_group
from vllm.platforms import current_platform
from vllm.v1.core.sched.output import SchedulerOutput
from vllm.v1.request import Request

from ucm.logger import init_logger
from ucm.shared.metrics import ucmmonitor
Expand All @@ -29,6 +28,7 @@
from vllm.attention.backends.abstract import AttentionMetadata
from vllm.forward_context import ForwardContext
from vllm.v1.core.kv_cache_manager import KVCacheBlocks
from vllm.v1.request import Request

logger = init_logger(__name__)

Expand Down Expand Up @@ -178,11 +178,12 @@ def __init__(self, vllm_config: "VllmConfig", role: KVConnectorRole):
self.metrics_config,
)
self.monitor = ucmmonitor.StatsMonitor.get_instance()
self.synchronize = (
torch.cuda.synchronize
if current_platform.is_cuda_alike()
else torch.npu.synchronize
)

self.synchronize = (
torch.cuda.synchronize
if current_platform.is_cuda_alike()
else torch.npu.synchronize
)

# invlalid block ids due to load errors
self._invalid_block_ids: set[int] = set()
Expand Down Expand Up @@ -558,7 +559,9 @@ def wait_for_save(self) -> None:
# TODO support PP
if (self.is_mla or self.is_dsa) and self.global_rank != 0:
return
if self.metrics_config:
if self.metrics_config or current_platform.device_type == "npu":
# When use vllm_ascend, we should add synchronize here, otherwise accuracy problem will raise
# This has already been fixed in the latest main branch of vllm_ascend, so synchronize will no longer be needed in future versions.
self.synchronize()

metadata = self._get_connector_metadata()
Expand Down