minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

fix(dsv4): upgrade forward metadata on main stream for large PP size (#25729)

Browse Source
This commit is contained in:
ybyang
2026-05-19 20:52:00 +00:00
committed by GitHub
parent cd012ada58
commit 8322fe09a7
1 changed files with 4 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/deepseek_v4.py
+4
View File
@@ -1045,6 +1045,10 @@ class DeepseekV4Model(nn.Module):
hidden_states = cp_split_and_rebuild_data(forward_batch, hidden_states) hidden_states = cp_split_and_rebuild_data(forward_batch, hidden_states)
positions = cp_split_and_rebuild_position(forward_batch, positions) 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): for i in range(self.start_layer, self.end_layer):
layer = self.layers[i] layer = self.layers[i]
hidden_states = layer( hidden_states = layer(