From 8322fe09a7b6931faedb00582a86174f39c4843b Mon Sep 17 00:00:00 2001 From: ybyang <10629930+whybeyoung@users.noreply.github.com> Date: Wed, 20 May 2026 04:52:00 +0800 Subject: [PATCH] fix(dsv4): upgrade forward metadata on main stream for large PP size (#25729) --- python/sglang/srt/models/deepseek_v4.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/python/sglang/srt/models/deepseek_v4.py b/python/sglang/srt/models/deepseek_v4.py index 425261ff9..d31370da9 100644 --- a/python/sglang/srt/models/deepseek_v4.py +++ b/python/sglang/srt/models/deepseek_v4.py @@ -1045,6 +1045,10 @@ class DeepseekV4Model(nn.Module): hidden_states = cp_split_and_rebuild_data(forward_batch, hidden_states) positions = cp_split_and_rebuild_position(forward_batch, positions) + # Upgrade lazy raw metadata on the main stream once before any layer + # forks alt-streams; later per-layer calls become no-ops. + forward_batch.attn_backend._maybe_upgrade_forward_metadata() + for i in range(self.start_layer, self.end_layer): layer = self.layers[i] hidden_states = layer(