[AMD] avoid correction_bias_dtype dtype convert (#20692)

2026-03-17 17:55:05 +08:00
parent 5270a06488
commit 966ae87d02
1 changed files with 1 additions and 0 deletions
--- 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(