From 168458fd90357063221b605bf55a1dd816e2de23 Mon Sep 17 00:00:00 2001 From: Xinyi Song Date: Tue, 6 Oct 2026 21:35:51 -0700 Subject: [PATCH] [AMD][V4.1] Fix wo_b input flatten for packed fp8 activations on gfx950 (#42796) --- 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 c76e2bab06..7efef24885 100644 --- a/python/sglang/srt/models/deepseek_v4.py +++ b/python/sglang/srt/models/deepseek_v4.py @@ -2621,7 +2621,7 @@ class MQALayer(MqaAttentionBase): and not should_skip_mlp_all_reduce() ) o, _ = self.wo_b( - o if isinstance(o, Mxfp8SwizzledInput) else o.flatten(1), + o.flatten(1) if isinstance(o, torch.Tensor) else o, skip_all_reduce=defer_all_reduce, ) if defer_all_reduce: