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

drop ascend scheduler (#4498)

Browse Source 
Ascend scheduler was added for non chunk prefill case before, since that
the npu ops didn't work well with chunked prefill.

Now the ops with chunked prefill work better, it's time to remove the
ascend scheduler to use vLLM default scheduler.

- vLLM version: v0.11.2

---------

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2025-11-29 16:18:34 +08:00
committed by GitHub
parent 53a52d6614
commit f10acddb78
52 changed files with 85 additions and 2948 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
vllm_ascend/platform.py
View File

@@ -153,7 +153,6 @@ class NPUPlatform(Platform):
model_config = vllm_config.model_config
parallel_config = vllm_config.parallel_config
cache_config = vllm_config.cache_config
ascend_scheduler_config = ascend_config.ascend_scheduler_config
kv_cache_dtype = vllm_config.additional_config.get(
"kv_cache_dtype", None)
@@ -291,35 +290,23 @@ class NPUPlatform(Platform):
if cache_config:
if cache_config.block_size is None:
cache_config.block_size = 128
if cache_config.enable_prefix_caching or \
not ascend_scheduler_config.enabled or \
getattr(ascend_scheduler_config, "enable_chunked_prefill", False):
logger.warning(
"If chunked prefill or prefix caching is enabled, block size must be set to 128."
)
origin_block_size = cache_config.block_size
cache_config.block_size = 128
# TODO(MengqingCao): Remove the model_type check, after resolving the hidden error in get_kv_cache_groups.
if model_config and model_config.hf_config.model_type == "qwen3_next":
logger.warning(
"When running qwen3-next model, block_size needs to be restored to its original value."
)
cache_config.block_size = origin_block_size
# ignore block size check if model is qwen3-next
# TODO(MengqingCao): Remove the model_type check, after resolving the hidden error in get_kv_cache_groups.
if not (model_config
and model_config.hf_config.model_type == "qwen3_next"):
# we must set block size to 128 if prefix caching is enabled or chunked prefill is enabled
if cache_config.enable_prefix_caching or \
(vllm_config.scheduler_config and vllm_config.scheduler_config.enable_chunked_prefill):
if cache_config.block_size != 128:
logger.warning(
"block size must be set to 128 on NPU platform.")
cache_config.block_size = 128
# Activate custom ops for v1, except on 310P
if get_ascend_device_type() != AscendDeviceType._310P:
compilation_config.custom_ops = ["all"]
# If ascend_scheduler_config is enabled,
# extents original scheduler_config to use AscendScheduler.
if ascend_config.ascend_scheduler_config.enabled:
from vllm_ascend.core.schedule_config import AscendSchedulerConfig
ascend_scheduler_config = AscendSchedulerConfig.initialize_from_config(
vllm_config.scheduler_config,
ascend_config.ascend_scheduler_config)
vllm_config.scheduler_config = ascend_scheduler_config
elif ascend_config.recompute_scheduler_enable:
if ascend_config.recompute_scheduler_enable:
from vllm_ascend.core.recompute_schedule_config import \
RecomputeSchedulerConfig
recompute_scheduler_config = RecomputeSchedulerConfig.initialize_from_config(