From 6f10e17b4a68df3cea7d37d30859409f62f720dc Mon Sep 17 00:00:00 2001 From: billishyahao Date: Fri, 16 Jan 2026 22:40:25 +0800 Subject: [PATCH] [bugfix] fix qwen3-next alt_stream none issue (#17016) --- python/sglang/srt/models/qwen3_next.py | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/python/sglang/srt/models/qwen3_next.py b/python/sglang/srt/models/qwen3_next.py index 6a662edfa..6f11fdaed 100644 --- a/python/sglang/srt/models/qwen3_next.py +++ b/python/sglang/srt/models/qwen3_next.py @@ -358,7 +358,11 @@ class Qwen3GatedDeltaNet(nn.Module): DUAL_STREAM_TOKEN_THRESHOLD = 1024 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() self.alt_stream.wait_stream(current_stream) projected_states_qkvz, _ = self.in_proj_qkvz(hidden_states)