[Core] Add kvcache watermark to reduce preemptions (#44594)
Signed-off-by: Nick Hill <nickhill123@gmail.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 4.8
parent
23eb7c8fbb
commit
4085ff7cb4
@@ -600,6 +600,8 @@ class EngineArgs:
|
||||
|
||||
scheduler_reserve_full_isl: bool = SchedulerConfig.scheduler_reserve_full_isl
|
||||
|
||||
watermark: float = SchedulerConfig.watermark
|
||||
|
||||
disable_hybrid_kv_cache_manager: bool | None = (
|
||||
SchedulerConfig.disable_hybrid_kv_cache_manager
|
||||
)
|
||||
@@ -1408,6 +1410,7 @@ class EngineArgs:
|
||||
"--scheduler-reserve-full-isl",
|
||||
**scheduler_kwargs["scheduler_reserve_full_isl"],
|
||||
)
|
||||
scheduler_group.add_argument("--watermark", **scheduler_kwargs["watermark"])
|
||||
scheduler_group.add_argument(
|
||||
"--disable-hybrid-kv-cache-manager",
|
||||
**scheduler_kwargs["disable_hybrid_kv_cache_manager"],
|
||||
@@ -2045,6 +2048,7 @@ class EngineArgs:
|
||||
max_long_partial_prefills=self.max_long_partial_prefills,
|
||||
long_prefill_token_threshold=self.long_prefill_token_threshold,
|
||||
scheduler_reserve_full_isl=self.scheduler_reserve_full_isl,
|
||||
watermark=self.watermark,
|
||||
disable_hybrid_kv_cache_manager=self.disable_hybrid_kv_cache_manager,
|
||||
async_scheduling=self.async_scheduling,
|
||||
stream_interval=self.stream_interval,
|
||||
|
||||
Reference in New Issue
Block a user