EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[Refactor] Remove Hicache Load & Write threads (#10127)

Browse Source 
Co-authored-by: Zhiqiang Xie <xiezhq@stanford.edu>
This commit is contained in:
DarkSharpness
2025-09-08 22:18:50 -07:00
committed by GitHub
parent cdc56ef6c1
commit 948b01a04c
10 changed files with 215 additions and 204 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/managers/schedule_batch.py
View File

@@ -911,7 +911,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
is_prefill_only: bool = False
# hicache pointer for synchronizing data loading from CPU to GPU
hicache_consumer_index: int = 0
hicache_consumer_index: int = -1
@classmethod
def init_new(
@@ -1897,7 +1897,7 @@ class ModelWorkerBatch:
spec_info: Optional[Union[EagleVerifyInput, EagleDraftInput]] = None
# If set, the output of the batch contains the hidden states of the run.
capture_hidden_mode: CaptureHiddenMode = None
hicache_consumer_index: int = 0
hicache_consumer_index: int = -1
# Overlap event
launch_done: Optional[threading.Event] = None