diff --git a/python/sglang/srt/layers/attention/nsa/nsa_indexer.py b/python/sglang/srt/layers/attention/nsa/nsa_indexer.py
index f2e659332..d2c146f3f 100644
--- a/python/sglang/srt/layers/attention/nsa/nsa_indexer.py
+++ b/python/sglang/srt/layers/attention/nsa/nsa_indexer.py
@@ -313,8 +313,8 @@ class Indexer(MultiPlatformOp):
 
         q_rope, k_rope = self.rotary_emb(positions, q_rope, k_rope)
 
-        query[..., : self.rope_head_dim] = q_rope
-        key[..., : self.rope_head_dim] = k_rope
+        query[..., : self.rope_head_dim] = q_rope.clone()
+        key[..., : self.rope_head_dim] = k_rope.clone()
 
         if enable_dual_stream:
             current_stream = torch.cuda.current_stream()
@@ -352,7 +352,7 @@ class Indexer(MultiPlatformOp):
         )
 
         _, k_rope = self.rotary_emb(positions, k_rope, k_rope)
-        key[..., : self.rope_head_dim] = k_rope
+        key[..., : self.rope_head_dim] = k_rope.clone()
         key = rotate_activation(key)
 
         return key
diff --git a/python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mha.py b/python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mha.py
index 7c83904f8..5ff13ebbb 100644
--- a/python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mha.py
+++ b/python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mha.py
@@ -424,7 +424,7 @@ class DeepseekMHAForwardMixin:
             )
         else:
             latent_cache[:, :, : self.kv_lora_rank] = kv_a.unsqueeze(1)
-            latent_cache[:, :, self.kv_lora_rank :] = k_pe
+            latent_cache[:, :, self.kv_lora_rank :] = k_pe.clone()
 
             # Save latent cache
             forward_batch.token_to_kv_pool.set_kv_buffer(