[Bugfix][Core] Stop zero-progress preemption cascades for deferred KV frees (#49675)
Signed-off-by: LearningMachine621 <147347705+LearningMachine621@users.noreply.github.com>
This commit is contained in:
@@ -17,6 +17,7 @@ from vllm.config import (
|
||||
SpeculativeConfig,
|
||||
VllmConfig,
|
||||
)
|
||||
from vllm.config.scheduler import SchedulerPolicy
|
||||
from vllm.multimodal.inputs import (
|
||||
MultiModalFeatureSpec,
|
||||
MultiModalKwargsItem,
|
||||
@@ -74,6 +75,7 @@ def create_scheduler(
|
||||
use_v2_model_runner: bool | None = None,
|
||||
kv_cache_spec: KVCacheSpec | None = None,
|
||||
per_request_spec_decode_metrics: str = "none",
|
||||
scheduling_policy: SchedulerPolicy = "fcfs",
|
||||
) -> Scheduler | AsyncScheduler:
|
||||
"""Create scheduler under test.
|
||||
|
||||
@@ -113,6 +115,7 @@ def create_scheduler(
|
||||
is_encoder_decoder=model_config.is_encoder_decoder,
|
||||
# Ensure admission/preemption mechanics are deterministic
|
||||
watermark=0.0,
|
||||
policy=scheduling_policy,
|
||||
)
|
||||
# Cache config, optionally force APC
|
||||
cache_config = CacheConfig(
|
||||
|
||||
Reference in New Issue
Block a user