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

[NVIDIA] disable chunked prefix cache when dp and blackwell is used (#9861)

Browse Source
This commit is contained in:
Kaixi Hou
2025-09-05 23:05:16 -07:00
committed by GitHub
parent 9a719b7afc
commit 90dfe3de4c
1 changed files with 11 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
python/sglang/srt/model_executor/model_runner.py
View File

@@ -525,6 +525,17 @@ class ModelRunner:
if not self.use_mla_backend: if not self.use_mla_backend:
server_args.disable_chunked_prefix_cache = True server_args.disable_chunked_prefix_cache = True
# TODO(kaixih@nvidia): remove this once we have a better solution for DP attention.
# For more details, see: https://github.com/sgl-project/sglang/issues/8616
elif (
self.dp_size > 1
and is_sm100_supported()
and server_args.attention_backend != "triton"
):
logger.info(
"Disable chunked prefix cache when dp size > 1 and attention backend is not triton."
)
server_args.disable_chunked_prefix_cache = True
if not server_args.disable_chunked_prefix_cache: if not server_args.disable_chunked_prefix_cache:
logger.info("Chunked prefix cache is turned on.") logger.info("Chunked prefix cache is turned on.")