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

Remove VLLM_ASCEND_ENABLE_TOPK_TOPP_OPTIMIZATION (#4860)

Browse Source 
VLLM_ASCEND_ENABLE_TOPK_TOPP_OPTIMIZATION is enabled by default for long
time. Let's remove it now.

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2025-12-10 23:50:18 +08:00
committed by GitHub
parent 37db0844f5
commit 08441baedd
2 changed files with 2 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
vllm_ascend/envs.py
View File

@@ -92,11 +92,6 @@ env_variables: Dict[str, Callable[[], Any]] = {
"VLLM_ASCEND_KV_CACHE_MEGABYTES_FLOATING_TOLERANCE": "VLLM_ASCEND_KV_CACHE_MEGABYTES_FLOATING_TOLERANCE":
lambda: int( lambda: int(
os.getenv("VLLM_ASCEND_KV_CACHE_MEGABYTES_FLOATING_TOLERANCE", 64)), os.getenv("VLLM_ASCEND_KV_CACHE_MEGABYTES_FLOATING_TOLERANCE", 64)),
# Whether to enable the topk optimization. It's enabled by default. Please set to False if you hit any issue.
# We'll remove this flag in the future once it's stable enough.
"VLLM_ASCEND_ENABLE_TOPK_TOPP_OPTIMIZATION":
lambda: bool(
int(os.getenv("VLLM_ASCEND_ENABLE_TOPK_TOPP_OPTIMIZATION", '1'))),
# Whether to enable mla_pa for deepseek mla decode, this flag will be removed after its available torch_npu is public accessible # Whether to enable mla_pa for deepseek mla decode, this flag will be removed after its available torch_npu is public accessible
# and the mla_pa will be the default path of deepseek decode path. # and the mla_pa will be the default path of deepseek decode path.
"VLLM_ASCEND_MLA_PA": "VLLM_ASCEND_MLA_PA":

9
vllm_ascend/worker/model_runner_v1.py
View File

@@ -141,6 +141,7 @@ from vllm_ascend.patch.worker.patch_module import patch_torch_npu_argsort
from vllm_ascend.platform import NPUPlatform from vllm_ascend.platform import NPUPlatform
from vllm_ascend.sample.logits_processor import build_logitsprocs from vllm_ascend.sample.logits_processor import build_logitsprocs
from vllm_ascend.sample.rejection_sampler import AscendRejectionSampler from vllm_ascend.sample.rejection_sampler import AscendRejectionSampler
from vllm_ascend.sample.sampler import AscendSampler
from vllm_ascend.spec_decode import get_spec_decode_method from vllm_ascend.spec_decode import get_spec_decode_method
from vllm_ascend.spec_decode.eagle_proposer import EagleProposer from vllm_ascend.spec_decode.eagle_proposer import EagleProposer
from vllm_ascend.spec_decode.interface import SpecDcodeType from vllm_ascend.spec_decode.interface import SpecDcodeType
@@ -312,15 +313,7 @@ class NPUModelRunner(LoRAModelRunnerMixin, ECConnectorModelRunnerMixin):
else: else:
self.prefetch_stream = None self.prefetch_stream = None
self.dtype = self.model_config.dtype self.dtype = self.model_config.dtype
if envs_ascend.VLLM_ASCEND_ENABLE_TOPK_TOPP_OPTIMIZATION:
# TODO: drop the env config to use ascend sampler by default
from vllm_ascend.sample.sampler import AscendSampler
self.sampler = AscendSampler() self.sampler = AscendSampler()
else:
from vllm.v1.sample.sampler import Sampler
self.sampler = Sampler()
self.reorder_batch_threshold: Optional[int] = None self.reorder_batch_threshold: Optional[int] = None
# Lazy initialization, these will be set after __init__ # Lazy initialization, these will be set after __init__