添加w_scale=1的判断，消除两处elementwise数乘操作。

python/sglang/srt/models/deepseek_v2.py

@@ -1618,11 +1618,18 @@ class DeepseekV2AttentionMLA(nn.Module):
                     torch.bfloat16,
                     q_nope_out,
                 )
-            else:
-                q_nope_out = torch.bmm(
-                    q_nope.to(torch.bfloat16).transpose(0, 1),
-                    self.w_kc.to(torch.bfloat16) * self.w_scale,
+            else:  # TODO: 手写融合算子
+                _q_nope_safe = q_nope.to(torch.bfloat16).transpose(0, 1)
+                _w_kc_safe = self.w_kc.to(torch.bfloat16)
+                if abs(self.w_scale - 1) < 1e-6:
+                    q_nope_out = torch.bmm(_q_nope_safe, _w_kc_safe)
+                else:
+                    q_nope_out = torch.bmm(_q_nope_safe, _w_kc_safe * self.w_scale,
                 )
+                # q_nope_out = torch.bmm(
+                #     q_nope.to(torch.bfloat16).transpose(0, 1),
+                #     self.w_kc.to(torch.bfloat16) * self.w_scale,
+                # )
         elif self.w_kc.dtype == torch.float8_e4m3fn:
             # fix bmm_fp8 error under cublas12.9 caused by bumpallocator, detail in pr#11612
             q_nope_val, q_nope_scale = per_tensor_quant_mla_fp8(
@@ -1763,11 +1770,18 @@ class DeepseekV2AttentionMLA(nn.Module):
                     torch.bfloat16,
                     attn_bmm_output,
                 )
-            else:
-                attn_bmm_output = torch.bmm(
-                    attn_output.to(torch.bfloat16).transpose(0, 1),
-                    self.w_vc.to(torch.bfloat16) * self.w_scale,
+            else:  # TODO: 手写融合算子
+                _attn_output_safe = attn_output.to(torch.bfloat16).transpose(0, 1)
+                _w_vc_safe = self.w_vc.to(torch.bfloat16)
+                if abs(self.w_scale - 1) < 1e-6:
+                    attn_bmm_output = torch.bmm(_attn_output_safe, _w_vc_safe)
+                else:
+                    attn_bmm_output = torch.bmm(_attn_output_safe, _w_vc_safe * self.w_scale,
                 )
+                # attn_bmm_output = torch.bmm(
+                #     attn_output.to(torch.bfloat16).transpose(0, 1),
+                #     self.w_vc.to(torch.bfloat16) * self.w_scale,
+                # )
 
             if self.o_proj.weight.dtype == torch.uint8:
                 attn_bmm_output = attn_bmm_output.transpose(0, 1)