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

[bugfix] fix qwen3-next alt_stream none issue (#17016)

Browse Source
This commit is contained in:
billishyahao
2026-01-16 22:40:25 +08:00
committed by GitHub
parent c771933dc5
commit 6f10e17b4a
1 changed files with 5 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/qwen3_next.py
+5 -1
View File
@@ -358,7 +358,11 @@ class Qwen3GatedDeltaNet(nn.Module):
DUAL_STREAM_TOKEN_THRESHOLD = 1024 DUAL_STREAM_TOKEN_THRESHOLD = 1024
seq_len, _ = hidden_states.shape seq_len, _ = hidden_states.shape
if seq_len < DUAL_STREAM_TOKEN_THRESHOLD: if (
seq_len < DUAL_STREAM_TOKEN_THRESHOLD
and self.alt_stream is not None
and get_is_capture_mode()
):
current_stream = torch.cuda.current_stream() current_stream = torch.cuda.current_stream()
self.alt_stream.wait_stream(current_stream) self.alt_stream.wait_stream(current_stream)
projected_states_qkvz, _ = self.in_proj_qkvz(hidden_states) projected_states_qkvz, _ = self.in_proj_qkvz(hidden_states)