EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

[Patch] Remove spec_decode.metrics patch (#1016)

Browse Source 
### What this PR does / why we need it?
Remove `spec_decode.metrics` patch as this has been resolved in
https://github.com/vllm-project/vllm/pull/16983 (include in vllm
`v0.9.0`).

Returns a CUDA event recording when the copy is complete **--after
modified-->** Returns a device event (NPU Event for vllm-ascend)
recording when the copy is complete.

Signed-off-by: shen-shanshan <467638484@qq.com>
This commit is contained in:
Shanshan Shen
2025-06-09 15:05:11 +08:00
committed by GitHub
parent 6003afa6d2
commit d2f87ed9cc
2 changed files with 0 additions and 54 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
vllm_ascend/patch/worker/patch_common/__init__.py
View File

@@ -20,7 +20,6 @@
import vllm_ascend.patch.worker.patch_common.patch_utils # noqa isort:skip
import vllm_ascend.patch.worker.patch_common.patch_distributed # noqa
import vllm_ascend.patch.worker.patch_common.patch_eagle # noqa
import vllm_ascend.patch.worker.patch_common.patch_metrics # noqa
import vllm_ascend.patch.worker.patch_common.patch_minicpm # noqa
import vllm_ascend.patch.worker.patch_common.patch_multi_step_worker # noqa
import vllm_ascend.patch.worker.patch_common.patch_sampler # noqa

53
vllm_ascend/patch/worker/patch_common/patch_metrics.py
View File

@@ -1,53 +0,0 @@
#
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
# This file is a part of the vllm-ascend project.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
from typing import Callable
import torch
from vllm.spec_decode.metrics import AsyncMetricsCollector
Timer = Callable[[], float]
def _copy_rejsample_metrics_async(self) -> torch.npu.Event:
"""
TODO: torch.cuda.xxx --> torch.npu.xxx
Copy rejection/typical-acceptance sampling metrics
(number of accepted tokens, etc) to CPU asynchronously.
Returns a NPU event recording when the copy is complete.
"""
assert self._copy_stream is not None
self._copy_stream.wait_stream(torch.npu.current_stream())
with torch.npu.stream(self._copy_stream):
self._aggregate_num_accepted_tokens.copy_(
self.spec_decode_sampler.num_accepted_tokens, non_blocking=True)
self._aggregate_num_emitted_tokens.copy_(
self.spec_decode_sampler.num_emitted_tokens, non_blocking=True)
# Number of draft tokens is calculated on CPU, so no copy is
# required.
self._aggregate_num_draft_tokens = (
self.spec_decode_sampler.num_draft_tokens)
aggregate_metrics_ready = torch.npu.Event()
aggregate_metrics_ready.record(self._copy_stream)
return aggregate_metrics_ready
AsyncMetricsCollector._copy_rejsample_metrics_async = _copy_rejsample_metrics_async