[NPU] Fix the DeepSeek-V2-Coder model accuracy issue (#29042)

python/sglang/srt/hardware_backend/npu/moe/topk.py

@@ -83,7 +83,8 @@ def fused_topk_npu(
             group_count=topk_config.num_expert_group if use_grouped_topk else 1,
             group_select_mode=(1 if use_grouped_topk else 0),
             renorm=0,
-            norm_type=1,  # 1 for sigmoid, 0 for softmax
+            # 1 for sigmoid, 0 for softmax
+            norm_type=(0 if topk_config.scoring_func == "softmax" else 1),
             routed_scaling_factor=(
                 1 if renormalize else topk_config.routed_scaling_factor
             ),

python/sglang/srt/models/deepseek_v2.py

@@ -658,6 +658,7 @@ class DeepseekV2MoE(nn.Module):
                 num_expert_group=config.n_group,
                 num_fused_shared_experts=self.num_fused_shared_experts,
                 topk_group=config.topk_group,
+                scoring_func=config.scoring_func,
                 correction_bias=self.gate.e_score_correction_bias,
                 quant_config=quant_config,
                 routed_scaling_factor=self.routed_scaling_factor,

python/sglang/srt/models/llada2.py

@@ -262,6 +262,7 @@ class LLaDA2MoeSparseMoeBlock(nn.Module):
             # num_fused_shared_experts=self.num_fused_shared_experts,
             topk_group=self.topk_group,
             correction_bias=self.correction_bias,
+            scoring_func=self.score_function,
             routed_scaling_factor=self.routed_scaling_factor,
         )