From 899e2be7d06047aa95b16abf06e553148c1418f5 Mon Sep 17 00:00:00 2001 From: billishyahao Date: Tue, 17 Feb 2026 14:18:57 +0800 Subject: [PATCH] [TBO] fix cuda graph intermittently becomes disabled bug (#18320) --- .../sglang/srt/batch_overlap/two_batch_overlap.py | 15 ++++++++++----- 1 file changed, 10 insertions(+), 5 deletions(-) diff --git a/python/sglang/srt/batch_overlap/two_batch_overlap.py b/python/sglang/srt/batch_overlap/two_batch_overlap.py index 12589a879..cfd2a54ed 100644 --- a/python/sglang/srt/batch_overlap/two_batch_overlap.py +++ b/python/sglang/srt/batch_overlap/two_batch_overlap.py @@ -440,18 +440,23 @@ class TboDPAttentionPreparer: @staticmethod def _compute_global_forward_mode(forward_modes): - forward_modes_excluding_idle = [ - x for x in forward_modes if x != ForwardMode.IDLE.value + forward_modes_excluding_idle_and_prebuilt = [ + x + for x in forward_modes + if x != ForwardMode.IDLE.value and x != ForwardMode.PREBUILT.value ] - if not forward_modes_excluding_idle: + if not forward_modes_excluding_idle_and_prebuilt: return ForwardMode.IDLE, False forward_mode_agree = TboDPAttentionPreparer._is_all_same( - forward_modes_excluding_idle + forward_modes_excluding_idle_and_prebuilt ) + global_forward_mode = ( - ForwardMode(forward_modes_excluding_idle[0]) if forward_mode_agree else None + ForwardMode(forward_modes_excluding_idle_and_prebuilt[0]) + if forward_mode_agree + else None ) return global_forward_mode, forward_mode_agree