chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

fix multimodal gen issues (#12765)

Browse Source 
Co-authored-by: Mick <mickjagger19@icloud.com>
This commit is contained in:
Yuhao Yang
2025-11-07 13:19:49 +08:00
committed by GitHub
parent 3b1cc466c0
commit d8736c756a
6 changed files with 56 additions and 42 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
python/sglang/multimodal_gen/runtime/server_args.py
View File

@@ -777,13 +777,6 @@ class ServerArgs:
)
self.sp_degree = self.ulysses_degree = self.ring_degree = 1
if (
self.ring_degree is not None
and self.ring_degree > 1
and self.attention_backend != "fa3"
):
raise ValueError("Ring Attention is only supported for fa3 backend for now")
if self.sp_degree == -1:
# assume we leave all remaining gpus to sp
num_gpus_per_group = self.dp_size * self.tp_size
@@ -815,6 +808,17 @@ class ServerArgs:
f"Ring degree not set, " f"using default value {self.ring_degree}"
)
if self.ring_degree > 1:
if self.attention_backend != None and self.attention_backend != "fa3":
raise ValueError(
"Ring Attention is only supported for fa3 backend for now"
)
else:
self.attention_backend = "fa3"
logger.info(
"Ring Attention is currently only supported for fa3, attention_backend has been automatically set to fa3"
)
if self.sp_degree == -1:
self.sp_degree = self.ring_degree * self.ulysses_degree
logger.info(