diff --git a/python/sglang/srt/mem_cache/pool_host/mla.py b/python/sglang/srt/mem_cache/pool_host/mla.py
index ed00e5c8a..41c651c3b 100644
--- a/python/sglang/srt/mem_cache/pool_host/mla.py
+++ b/python/sglang/srt/mem_cache/pool_host/mla.py
@@ -441,9 +441,14 @@ class MLATokenToKVPoolHost(HiSparseHostPoolMixin, HostKVCache):
                 )
             return
 
-        device_data_ptrs, device_kv_buffers = self._resolve_device_transfer_buffers(
-            device_pool
-        )
+        if io_backend == "kernel_ascend":
+            # NPU pools use contiguous multi-layer tensors and intentionally do
+            # not build the CUDA-style data_ptrs array.
+            device_data_ptrs, device_kv_buffers = None, None
+        else:
+            device_data_ptrs, device_kv_buffers = self._resolve_device_transfer_buffers(
+                device_pool
+            )
 
         if io_backend == "kernel":
             if self.layout == "layer_first":