diff --git a/python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py b/python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py
index f18c57070..33345593f 100644
--- a/python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py
+++ b/python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py
@@ -628,7 +628,7 @@ class AscendAttnBackend(AttentionBackend):
             if self.forward_metadata.actual_seq_lengths_q is not None:
                 actual_seq_qlen = self.forward_metadata.actual_seq_lengths_q
             else:
-                actual_seq_qlen = torch.cumsum(forward_batch.seq_lens, dim=0)
+                actual_seq_qlen = torch.cumsum(forward_batch.extend_seq_lens, dim=0)
         else:
             if self.forward_metadata.actual_seq_lengths_q is None:
                 if (
diff --git a/python/sglang/srt/hardware_backend/npu/memory_pool_npu.py b/python/sglang/srt/hardware_backend/npu/memory_pool_npu.py
index 4968c03e8..82e7ce693 100644
--- a/python/sglang/srt/hardware_backend/npu/memory_pool_npu.py
+++ b/python/sglang/srt/hardware_backend/npu/memory_pool_npu.py
@@ -221,6 +221,7 @@ class NPUMLATokenToKVPool(MLATokenToKVPool):
                 dtype=self.store_dtype,
                 device=self.device,
             )
+            self.index_k_buffer = None
             if self.index_head_dim is not None:
                 self.index_k_buffer = torch.zeros(
                     (
diff --git a/python/sglang/srt/layers/attention/nsa/nsa_indexer.py b/python/sglang/srt/layers/attention/nsa/nsa_indexer.py
index 2fe14c368..cf362e712 100644
--- a/python/sglang/srt/layers/attention/nsa/nsa_indexer.py
+++ b/python/sglang/srt/layers/attention/nsa/nsa_indexer.py
@@ -1244,7 +1244,7 @@ class Indexer(MultiPlatformOp):
                 )
             else:
                 actual_seq_lengths_kv = forward_batch.seq_lens
-                actual_seq_lengths_q = forward_batch.seq_lens.cumsum(dim=0)
+                actual_seq_lengths_q = forward_batch.extend_seq_lens.cumsum(dim=0)
         else:
             if forward_batch.attn_backend.forward_metadata.actual_seq_lengths_q is None:
                 if (
diff --git a/python/sglang/srt/mem_cache/memory_pool_host.py b/python/sglang/srt/mem_cache/memory_pool_host.py
index 548c9d9f1..f40ddb17d 100644
--- a/python/sglang/srt/mem_cache/memory_pool_host.py
+++ b/python/sglang/srt/mem_cache/memory_pool_host.py
@@ -769,6 +769,15 @@ class MLATokenToKVPoolHost(HostKVCache):
                 pin_memory=self.pin_memory,
                 allocator=self.allocator,
             )
+            self.index_k_buffer = None
+            if self.device_pool.index_head_dim is not None:
+                self.index_k_buffer = alloc_func(
+                    (*base_dims, self.device_pool.index_head_dim),
+                    dtype=self.dtype,
+                    device=self.device,
+                    pin_memory=self.pin_memory,
+                    allocator=self.allocator,
+                )
             # Return k_buffer to preserve original kv_buffer and data_refs init logic,
             # though Ascend doesn't use these parameters.
             return self.k_buffer
@@ -844,6 +853,8 @@ class MLATokenToKVPoolHost(HostKVCache):
                         host_k=self.k_buffer,
                         device_v=device_pool.v_buffer,
                         host_v=self.v_buffer,
+                        device_index_k=device_pool.index_k_buffer,
+                        host_index_k=self.index_k_buffer,
                         page_size=self.page_size,
                         direction=TransferDirection.H2D,
                     )
@@ -905,6 +916,8 @@ class MLATokenToKVPoolHost(HostKVCache):
                     host_k=self.k_buffer,
                     device_v=device_pool.v_buffer,
                     host_v=self.v_buffer,
+                    device_index_k=device_pool.index_k_buffer,
+                    host_index_k=self.index_k_buffer,
                     page_size=self.page_size,
                     direction=TransferDirection.D2H,
                 )