[Feature] Optimize DeepSeek's DeepEP on Ascend NPU (#8355)

Co-authored-by: ronnie_zheng <zl19940307@163.com>
Co-authored-by: Hexq0210 <hexq0809521@gmail.com>

python/sglang/srt/layers/moe/topk.py

@@ -250,10 +250,11 @@ class TopK(CustomOp):
 
         # NOTE: now npu_moe_gating_top_k can only support `group_count=256` pattern
         if global_num_experts == 256:
+            router_logits = router_logits.to(torch.float32)
             return torch_npu.npu_moe_gating_top_k(
                 router_logits,
                 k=self.top_k,
-                bias=self.correction_bias,
+                bias=self.correction_bias.to(torch.float32),
                 k_group=self.topk_group,
                 group_count=self.num_expert_group,
                 group_select_mode=1,