[CI] Ascend NPU CI enhancement (#8294)

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

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

@@ -398,8 +398,12 @@ def grouped_topk_gpu(
         .reshape(num_token, -1)
     )  # [n, e]
     tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0)  # [n, e]
+    # TODO: NPU can't support directly evaluating a comparison for now
     topk_weights, topk_ids = torch.topk(
-        tmp_scores, k=topk, dim=-1, sorted=num_fused_shared_experts > 0
+        tmp_scores,
+        k=topk,
+        dim=-1,
+        sorted=(True if num_fused_shared_experts > 0 else False),
     )
     if num_fused_shared_experts:
         topk_ids[:, -1] = torch.randint(
@@ -489,8 +493,12 @@ def biased_grouped_topk_impl(
     tmp_scores = scores_for_choice.masked_fill(
         ~score_mask.bool(), float("-inf")
     )  # [n, e]
+    # TODO: NPU can't support directly evaluating a comparison for now
     _, topk_ids = torch.topk(
-        tmp_scores, k=topk, dim=-1, sorted=num_fused_shared_experts > 0
+        tmp_scores,
+        k=topk,
+        dim=-1,
+        sorted=(True if num_fused_shared_experts > 0 else False),
     )
     topk_weights = scores.gather(1, topk_ids)