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

[CI] upgrade vllm to 0.8.5 (#715)

Browse Source 
1. Upgrade vllm to 0.8.5
2. Drop 0.8.4 support
3. Keep doc to 0.8.4rc2 until we release 0.8.5

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2025-04-30 09:15:50 +08:00
committed by GitHub
parent 95e7aa4736
commit f8350569e6
20 changed files with 48 additions and 579 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm_ascend/patch/worker/__init__.py
View File

@@ -18,8 +18,8 @@
from vllm_ascend.utils import vllm_version_is
# Import specific patches for different versions
if vllm_version_is("0.8.4"):
from vllm_ascend.patch.worker import patch_0_8_4 # noqa: F401
if vllm_version_is("0.8.5"):
from vllm_ascend.patch.worker import patch_0_8_5 # noqa: F401
from vllm_ascend.patch.worker import patch_common # noqa: F401
else:
from vllm_ascend.patch.worker import patch_common # noqa: F401

59
vllm_ascend/patch/worker/patch_0_8_4/patch_metrics.py
View File

@@ -1,59 +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, Optional, Union
import torch
from vllm.spec_decode.metrics import (AsyncMetricsCollector,
SpecDecodeWorkerMetrics)
Timer = Callable[[], float]
# TODO: revert this patch when the cuda hard code is removed in vllm
# init_tensors: Modified the hard-coded cuda judgment logic to npu;
# maybe_collect_rejsample_metrics: Removed the check for current_platform.is_cuda_alike()
def init_tensors(self,
rank: int,
device_type: Union[torch.device, str] = 'npu') -> None:
self._rank = rank
if isinstance(device_type, torch.device):
device_type = device_type.type
if device_type == 'npu':
self._copy_stream = torch.npu.Stream()
def maybe_collect_rejsample_metrics(
self, k: int) -> Optional[SpecDecodeWorkerMetrics]:
# If a copy was initiated in the previous call, collect and return.
if self._in_flight_copy is not None:
ready_event = self._in_flight_copy
self._in_flight_copy = None
return self._collect_rejsample_metrics(k, ready_event)
# Otherwise, check if we should start a new copy.
if self._should_collect_rejsample_metrics(self._timer()):
assert self._in_flight_copy is None
self._in_flight_copy = self._copy_rejsample_metrics_async()
return None
AsyncMetricsCollector.init_tensors = init_tensors
AsyncMetricsCollector.maybe_collect_rejsample_metrics = maybe_collect_rejsample_metrics

30
vllm_ascend/patch/worker/patch_0_8_4/patch_spec_decode_worker.py
View File

@@ -1,30 +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 vllm.spec_decode.spec_decode_worker import SpecDecodeWorker
def _configure_model_sampler_for_spec_decode(self):
(self.scorer_worker.model_runner.model.sampler.include_gpu_probs_tensor
) = True
(self.scorer_worker.model_runner.model.sampler.
should_modify_greedy_probs_inplace) = True
self.proposer_worker.set_include_gpu_probs_tensor()
self.proposer_worker.set_should_modify_greedy_probs_inplace()
SpecDecodeWorker._configure_model_sampler_for_spec_decode = _configure_model_sampler_for_spec_decode

71
vllm_ascend/patch/worker/patch_0_8_4/patch_tritonplaceholder.py
View File

@@ -1,71 +0,0 @@
#
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
# Copyright 2023 The vLLM team.
#
# 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.
# This file is a part of the vllm-ascend project.
# Adapted from vllm/triton_utils/importing.py
#
import importlib
import sys
import types
from importlib.util import find_spec
from vllm.logger import logger
HAS_TRITON = (
find_spec("triton") is not None
or find_spec("pytorch-triton-xpu") is not None # Not compatible
)
if not HAS_TRITON:
logger.info("Triton not installed or not compatible; certain GPU-related"
" functions will not be available.")
class TritonPlaceholder(types.ModuleType):
def __init__(self):
super().__init__("triton")
self.jit = self._dummy_decorator("jit")
self.autotune = self._dummy_decorator("autotune")
self.heuristics = self._dummy_decorator("heuristics")
self.language = TritonLanguagePlaceholder()
self.__spec__ = importlib.machinery.ModuleSpec(
name="triton", loader=None, origin="placeholder")
logger.warning_once(
"Triton is not installed. Using dummy decorators. "
"Install it via `pip install triton` to enable kernel"
" compilation.")
def _dummy_decorator(self, name):
def decorator(func=None, **kwargs):
if func is None:
return lambda f: f
return func
return decorator
class TritonLanguagePlaceholder(types.ModuleType):
def __init__(self):
super().__init__("triton.language")
self.constexpr = None
self.dtype = None
sys.modules['triton'] = TritonPlaceholder()
sys.modules['triton.language'] = TritonLanguagePlaceholder()
if 'triton' in sys.modules:
logger.info("Triton module has been replaced with a placeholder.")

3
vllm_ascend/patch/worker/patch_0_8_4/__init__.py → vllm_ascend/patch/worker/patch_0_8_5/__init__.py
View File

@@ -14,6 +14,3 @@
# See the License for the specific language governing permissions and
# limitations under the License.
#
import vllm_ascend.patch.worker.patch_0_8_4.patch_metrics # noqa
import vllm_ascend.patch.worker.patch_0_8_4.patch_tritonplaceholder # noqa

7
vllm_ascend/patch/worker/patch_common/patch_multi_step_worker.py
View File

@@ -22,7 +22,6 @@ from vllm.model_executor.layers.sampler import SamplerOutput
from vllm.sequence import ExecuteModelRequest
from vllm.spec_decode.multi_step_worker import MultiStepWorker
from vllm_ascend.utils import vllm_version_is
from vllm_ascend.worker.draft_model_runner import TP1DraftModelRunner
@@ -93,16 +92,14 @@ def set_include_gpu_probs_tensor(self) -> None:
# Need include_gpu_probs_tensor for MultiSteoWorker
if hasattr(self.model_runner.model, "sampler"):
self.model_runner.model.sampler.include_gpu_probs_tensor = True
if not vllm_version_is("0.8.4"):
self.model_runner.sampler.include_gpu_probs_tensor = True
self.model_runner.sampler.include_gpu_probs_tensor = True
def set_should_modify_greedy_probs_inplace(self) -> None:
if hasattr(self.model_runner.model, "sampler"):
self.model_runner.model.sampler.should_modify_greedy_probs_inplace = (
True)
if not vllm_version_is("0.8.4"):
self.model_runner.sampler.should_modify_greedy_probs_inplace = True
self.model_runner.sampler.should_modify_greedy_probs_inplace = True
MultiStepWorker.sampler_output = torch.inference_mode()(sampler_output)