From b44ac5d49aa22264603d38085bd7bd340c3d8517 Mon Sep 17 00:00:00 2001 From: Yuwei An Date: Mon, 13 Jul 2026 10:07:02 -0700 Subject: [PATCH] [Tiny] Enable Full Cuda Graph with Page size = 1 (#30835) Co-authored-by: Claude Fable 5 --- python/sglang/srt/arg_groups/overrides.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/python/sglang/srt/arg_groups/overrides.py b/python/sglang/srt/arg_groups/overrides.py index 4d175a47a..bbb775722 100644 --- a/python/sglang/srt/arg_groups/overrides.py +++ b/python/sglang/srt/arg_groups/overrides.py @@ -1796,6 +1796,9 @@ def _fa4_page_constraint(view: Any) -> dict: # 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. 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( f"FA4 backend only supports page size 128 for non-MLA model architectures, changing page_size from {view.page_size} to 128."