minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[Chore] Remove deadcode in prefill delayer (#23389)

Browse Source 
Co-authored-by: Jonah Bernard <96398205+Jonahcb@users.noreply.github.com>
This commit is contained in:
Jonah Bernard
2026-04-27 09:59:36 -07:00
committed by GitHub
co-authored by Jonah Bernard
parent 47b8eadbc4
commit 4fc5ebf0b7
1 changed files with 28 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/managers/prefill_delayer.py
+28 -19
View File
@@ -78,13 +78,17 @@ class PrefillDelayer:
self,
local_prefillable: bool,
token_usage: float,
**kwargs,
running_batch: int = 0,
max_prefill_bs: int = 0,
max_running_requests: int = 0,
) -> _NegotiateOutput:
out = self._negotiate_should_allow_prefill_pure(
prev_state=self._curr_state,
local_prefillable=local_prefillable,
token_usage=token_usage,
**kwargs,
running_batch=running_batch,
max_prefill_bs=max_prefill_bs,
max_running_requests=max_running_requests,
)
self._curr_state = out.next_state
return out
@@ -95,7 +99,9 @@ class PrefillDelayer:
prev_state: Optional[_State],
local_prefillable: bool,
token_usage: float,
**kwargs,
running_batch: int = 0,
max_prefill_bs: int = 0,
max_running_requests: int = 0,
) -> _NegotiateOutput:
# Compute local states
local_token_watermark_force_allow = (
@@ -108,7 +114,8 @@ class PrefillDelayer:
tp0_info = self._gather_info(
local_prefillable=local_prefillable,
local_token_watermark_force_allow=local_token_watermark_force_allow,
**kwargs,
running_batch=running_batch,
max_prefill_bs=max_prefill_bs,
)
global_prefillable = tp0_info[:, 0]
global_token_watermark_force_allow = tp0_info[:, 1]
@@ -133,16 +140,6 @@ class PrefillDelayer:
# Compute outputs
if prefillable_status == "all":
if kwargs is None:
exist_previous_wait = prev_state is not None
return _NegotiateOutput(
next_state=None,
output_allow=True,
output_reason="wait_success" if exist_previous_wait else "no_wait",
**debug_info,
)
max_running_requests = kwargs.get("max_running_requests", 0)
if not self.enable_dp_attention:
max_running_requests = (
max_running_requests + self.dp_size - 1
@@ -210,14 +207,18 @@ class PrefillDelayer:
raise NotImplementedError
def _gather_info(
self, local_prefillable: bool, local_token_watermark_force_allow: bool, **kwargs
self,
local_prefillable: bool,
local_token_watermark_force_allow: bool,
running_batch: int = 0,
max_prefill_bs: int = 0,
):
local_info = torch.tensor(
[
int(local_prefillable),
int(local_token_watermark_force_allow),
kwargs.get("running_batch", 0),
kwargs.get("max_prefill_bs", 0),
running_batch,
max_prefill_bs,
],
device="cpu",
dtype=torch.int64,
@@ -251,12 +252,20 @@ class PrefillDelayerSinglePassExecutor:
metrics_collector=self._prefill_delayer._metrics_collector,
)
def negotiate_should_allow_prefill(self, local_prefillable: bool, **kwargs) -> bool:
def negotiate_should_allow_prefill(
self,
local_prefillable: bool,
running_batch: int = 0,
max_prefill_bs: int = 0,
max_running_requests: int = 0,
) -> bool:
if not self._called:
self._result = self._prefill_delayer._negotiate_should_allow_prefill(
local_prefillable=local_prefillable,
token_usage=self._token_usage,
**kwargs,
running_batch=running_batch,
max_prefill_bs=max_prefill_bs,
max_running_requests=max_running_requests,
)
return self._result.output_allow