From 966ae87d021f1a902240149937939082faef735c Mon Sep 17 00:00:00 2001 From: AMD-yanfeiwang Date: Tue, 17 Mar 2026 17:55:05 +0800 Subject: [PATCH] [AMD] avoid correction_bias_dtype dtype convert (#20692) --- python/sglang/srt/models/deepseek_v2.py | 1 + 1 file changed, 1 insertion(+) diff --git a/python/sglang/srt/models/deepseek_v2.py b/python/sglang/srt/models/deepseek_v2.py index d10ff54c0..13f0498ac 100644 --- a/python/sglang/srt/models/deepseek_v2.py +++ b/python/sglang/srt/models/deepseek_v2.py @@ -281,6 +281,7 @@ class MoEGate(nn.Module): elif _use_aiter and quant_config.get_name() in ( "fp8", "compressed_tensors", + "quark", ): correction_bias_dtype = torch.bfloat16 self.e_score_correction_bias = nn.Parameter(