From 770c51b1272e293124c61ad86d080835f51757fd Mon Sep 17 00:00:00 2001 From: Xingyu Liu <38244988+charlotte12l@users.noreply.github.com> Date: Thu, 28 May 2026 00:52:26 -0700 Subject: [PATCH] [Bug] Forward fixed_split_size in SWA / cross-attention paths of FlashInfer backend (#26412) Signed-off-by: Xingyu Liu --- python/sglang/srt/layers/attention/flashinfer_backend.py | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/python/sglang/srt/layers/attention/flashinfer_backend.py b/python/sglang/srt/layers/attention/flashinfer_backend.py index cd8cc2dcd..23ccf51a4 100644 --- a/python/sglang/srt/layers/attention/flashinfer_backend.py +++ b/python/sglang/srt/layers/attention/flashinfer_backend.py @@ -1059,6 +1059,8 @@ class FlashInferIndicesUpdaterDecode: spec_info, seq_lens_cpu=seq_lens_cpu_tmp, use_sliding_window_kv_pool=use_sliding_window_kv_pool, + fixed_split_size=fixed_split_size, + disable_split_kv=disable_split_kv, ) def update_cross_attention( @@ -1096,6 +1098,8 @@ class FlashInferIndicesUpdaterDecode: kv_start_idx, spec_info, seq_lens_cpu=kv_lens_cpu, + fixed_split_size=fixed_split_size, + disable_split_kv=disable_split_kv, ) def call_begin_forward( @@ -1340,6 +1344,7 @@ class FlashInferIndicesUpdaterPrefill: use_ragged, spec_info, use_sliding_window_kv_pool=use_sliding_window_kv_pool, + fixed_split_size=fixed_split_size, multi_item_params=multi_item_params, ) @@ -1383,6 +1388,7 @@ class FlashInferIndicesUpdaterPrefill: self.qo_indptr[wrapper_id], use_ragged, spec_info, + fixed_split_size=fixed_split_size, multi_item_params=multi_item_params, cross_attention_custom_mask=( cross_attention_custom_mask if wrapper_id == 1 else None