From bdc1e46e5ac93dd7ee2d53e5d8c2d3de7dc03f69 Mon Sep 17 00:00:00 2001 From: AlfredYong Date: Thu, 26 Feb 2026 20:11:29 +0800 Subject: [PATCH] [Qwen3.5] Qwen3.5-27B inference repeat bug fix (#19411) --- python/sglang/srt/models/qwen3_5.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/python/sglang/srt/models/qwen3_5.py b/python/sglang/srt/models/qwen3_5.py index 580973865..306adbccb 100644 --- a/python/sglang/srt/models/qwen3_5.py +++ b/python/sglang/srt/models/qwen3_5.py @@ -352,6 +352,7 @@ class Qwen3_5LinearDecoderLayer(nn.Module): input_layernorm=self.input_layernorm, post_attention_layernorm=self.post_attention_layernorm, allow_reduce_scatter=True, + is_last_layer=(layer_id == config.num_hidden_layers - 1), ) def forward( @@ -542,6 +543,7 @@ class Qwen3_5AttentionDecoderLayer(nn.Module): input_layernorm=self.input_layernorm, post_attention_layernorm=self.post_attention_layernorm, allow_reduce_scatter=True, + is_last_layer=(layer_id == config.num_hidden_layers - 1), ) self.alt_stream = alt_stream