From d5564c2a962af23a0085cae5d08a75cf939bc3ba Mon Sep 17 00:00:00 2001 From: shuwenn <47200617+alphabetc1@users.noreply.github.com> Date: Sat, 9 May 2026 18:22:45 +0800 Subject: [PATCH] fix(fa3): translate page table to SWA loc in EAGLE3 topk>1 spec metadata (#24617) --- .../srt/layers/attention/flashattention_backend.py | 14 ++++++++++++-- 1 file changed, 12 insertions(+), 2 deletions(-) diff --git a/python/sglang/srt/layers/attention/flashattention_backend.py b/python/sglang/srt/layers/attention/flashattention_backend.py index bbb0385c8..709f0c51c 100644 --- a/python/sglang/srt/layers/attention/flashattention_backend.py +++ b/python/sglang/srt/layers/attention/flashattention_backend.py @@ -2483,10 +2483,20 @@ class FlashAttentionBackend(AttentionBackend): else metadata_swa.page_table ) + page_table_a = metadata.page_table + page_table_b = metadata_expand.page_table + if self.use_sliding_window_kv_pool: + page_table_a = self.token_to_kv_pool.translate_loc_from_full_to_swa( + page_table_a + ) + page_table_b = self.token_to_kv_pool.translate_loc_from_full_to_swa( + page_table_b + ) + prepare_swa_spec_page_table_triton( page_table, - metadata.page_table, - metadata_expand.page_table, + page_table_a, + page_table_b, metadata.cache_seqlens_int32, metadata_expand.cache_seqlens_int32, self.speculative_num_draft_tokens,