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

Fix docs for clip max new tokens (#9082)

Browse Source
This commit is contained in:
Liangsheng Yin
2025-08-11 13:15:21 -07:00
committed by GitHub
parent 0d9e89ec69
commit f9afa7dceb
4 changed files with 10 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
python/sglang/srt/disaggregation/decode.py
View File

@@ -59,9 +59,7 @@ if TYPE_CHECKING:
from sglang.srt.managers.schedule_batch import Req
from sglang.srt.managers.scheduler import Scheduler
DECODE_CLIP_MAX_NEW_TOKEN = get_int_env_var(
"SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION", 4096
)
CLIP_MAX_NEW_TOKEN = get_int_env_var("SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION", 4096)
class DecodeReqToTokenPool:
@@ -390,7 +388,7 @@ class DecodePreallocQueue:
origin_input_len
+ min(
decode_req.req.sampling_params.max_new_tokens,
DECODE_CLIP_MAX_NEW_TOKEN,
CLIP_MAX_NEW_TOKEN,
)
- retractable_tokens,
)
@@ -440,7 +438,7 @@ class DecodePreallocQueue:
need_space_for_single_req = (
max(
[
min(x.sampling_params.max_new_tokens, DECODE_CLIP_MAX_NEW_TOKEN)
min(x.sampling_params.max_new_tokens, CLIP_MAX_NEW_TOKEN)
+ len(x.origin_input_ids)
- retractable_tokens
for x in self.scheduler.running_batch.reqs