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

[PD] Allow customizing reserved tokens to avoid KV cache waste (#6002)

Browse Source
This commit is contained in:
fzyzcjy
2025-05-05 11:23:15 +08:00
committed by GitHub
parent 357fb2dba5
commit 3008db9c1a
1 changed files with 3 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/disaggregation/decode.py
View File

@@ -97,7 +97,9 @@ class DecodePreallocQueue:
self.tp_size = tp_size self.tp_size = tp_size
self.bootstrap_port = bootstrap_port self.bootstrap_port = bootstrap_port
self.num_reserved_decode_tokens = 512 self.num_reserved_decode_tokens = int(
os.environ.get("SGLANG_NUM_RESERVED_DECODE_TOKENS", "512")
)
# Queue for requests pending pre-allocation # Queue for requests pending pre-allocation
self.queue: List[DecodeRequest] = [] self.queue: List[DecodeRequest] = []