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

[Tiny] Enable Full Cuda Graph with Page size = 1 (#30835)

Browse Source 
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Yuwei An
2026-07-14 01:07:02 +08:00
committed by GitHub
co-authored by Claude Fable 5
parent 978bce2063
commit b44ac5d49a
1 changed files with 3 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/arg_groups/overrides.py
+3
View File
@@ -1796,6 +1796,9 @@ def _fa4_page_constraint(view: Any) -> dict:
# CUTLASS kernel aborts on at page_size>1. That path only works at # CUTLASS kernel aborts on at page_size>1. That path only works at
# page_size==1, so skip the 128 auto-force for it and keep the default. # page_size==1, so skip the 128 auto-force for it and keep the default.
and (view.speculative_eagle_topk or 0) <= 1 and (view.speculative_eagle_topk or 0) <= 1
# The full prefill CUDA graph runs the FA backend at page_size==1 only
# (#27988), so skip the 128 auto-force for it and keep the default.
and view.cuda_graph_config.prefill.backend != Backend.FULL
): ):
logger.warning( logger.warning(
f"FA4 backend only supports page size 128 for non-MLA model architectures, changing page_size from {view.page_size} to 128." f"FA4 backend only supports page size 128 for non-MLA model architectures, changing page_size from {view.page_size} to 128."