From 7825e5ffcad2030ca375fcca4a85ba0cf171fc7b Mon Sep 17 00:00:00 2001 From: AMD-yanfeiwang Date: Fri, 4 Sep 2026 13:59:53 +0800 Subject: [PATCH] [AMD] Fix DSV4 unified attention sink TP slice (#35092) Co-authored-by: kk <43161300+kkHuang-amd@users.noreply.github.com> --- python/sglang/srt/models/deepseek_v4.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/python/sglang/srt/models/deepseek_v4.py b/python/sglang/srt/models/deepseek_v4.py index 68b4d1b08..7a564232f 100644 --- a/python/sglang/srt/models/deepseek_v4.py +++ b/python/sglang/srt/models/deepseek_v4.py @@ -1699,7 +1699,7 @@ class MQALayer(MqaAttentionBase): layer=self.attn_mqa, forward_batch=forward_batch, compress_ratio=self.compress_ratio, - attn_sink=self.attn_sink, + attn_sink=attn_sink[: self.n_local_heads], save_kv_cache=kv is not None, ) else: