minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

remove the fa4 page_size hardcode to 128 restriction on mla model arch (#12801)

Browse Source
This commit is contained in:
Rain Jiang
2025-11-07 13:30:30 -08:00
committed by GitHub
parent ae62279071
commit 0f76976c3c
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/server_args.py
+2 -2
View File
@@ -1263,9 +1263,9 @@ class ServerArgs:
raise ValueError( raise ValueError(
"FA4 backend is only supported for prefill. Please use `--prefill-attention-backend fa4` instead." "FA4 backend is only supported for prefill. Please use `--prefill-attention-backend fa4` instead."
) )
if self.prefill_attention_backend == "fa4": if self.prefill_attention_backend == "fa4" and not self.use_mla_backend():
logger.warning( logger.warning(
f"FA4 backend only supports page size 128, changing page_size from {self.page_size} to 128." f"FA4 backend only supports page size 128 for non-MLA model architectures, changing page_size from {self.page_size} to 128."
) )
self.page_size = 128 self.page_size = 128