diff --git a/python/sglang/srt/hardware_backend/npu/modules/deepseek_v2_attention_mla_npu.py b/python/sglang/srt/hardware_backend/npu/modules/deepseek_v2_attention_mla_npu.py
index 24fbb4708..21c6f6e16 100644
--- a/python/sglang/srt/hardware_backend/npu/modules/deepseek_v2_attention_mla_npu.py
+++ b/python/sglang/srt/hardware_backend/npu/modules/deepseek_v2_attention_mla_npu.py
@@ -403,9 +403,7 @@ def forward_dsa_prepare_npu(
                 latent_cache, forward_batch, k_nope, k_pe
             )
 
-    if m.skip_topk:
-        topk_indices = prev_topk_indices
-    else:
+    if not m.skip_topk or (m.is_nextn and prev_topk_indices is None):
         topk_indices = m.indexer(
             hidden_states,
             q_lora,
@@ -415,6 +413,8 @@ def forward_dsa_prepare_npu(
             layer_scatter_modes,
             dynamic_scale,
         )
+    else:
+        topk_indices = prev_topk_indices
 
     return (
         q_pe,