minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[DLLM] Remove cuda graph batch size limitation (#17458)

Browse Source
This commit is contained in:
Tiwei Bie
2026-01-23 09:52:39 -08:00
committed by GitHub
parent 010c17a133
commit 5438cd20ce
1 changed files with 0 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/server_args.py
-5
View File
@@ -2600,11 +2600,6 @@ class ServerArgs:
) )
self.attention_backend = "triton" self.attention_backend = "triton"
elif not self.disable_cuda_graph: elif not self.disable_cuda_graph:
if self.cuda_graph_bs != [1]:
logger.warning(
"Cuda graph bs is set to [1] because of using diffusion LLM inference"
)
self.cuda_graph_bs = [1]
if self.attention_backend != "flashinfer": if self.attention_backend != "flashinfer":
logger.warning( logger.warning(
"Attention backend is set to flashinfer because of enabling cuda graph in diffusion LLM inference" "Attention backend is set to flashinfer because of enabling cuda graph in diffusion LLM inference"