[NPU] [bugfix] Fix import of ggml_moe_a8_vec and Fix NPU MLA HiCache backup accessing missing data_ptrs (#36640)

Co-authored-by: sglang-npu-bot <sglangnpu@163.com>

python/sglang/srt/layers/quantization/mxfp4_flashinfer_trtllm_moe.py

@@ -17,6 +17,7 @@ from sglang.srt.layers.moe.utils import RoutingMethodType
 from sglang.srt.runtime_context import get_exec
 from sglang.srt.utils import (
     is_flashinfer_available,
+    is_npu,
     log_info_on_rank0,
     set_weight_attrs,
 )
@@ -380,7 +381,6 @@ def maybe_fuse_routed_scale_and_shared_add(
     # alpha=scale)`. With no shared output, the missing scale is applied
     # in-place. Otherwise `routed` is already scale-final and we just add
     # `shared` (or pass through if there is none).
-    from sglang.srt.layers.quantization.expert_pack import ExpertPackMoEMethod
     from sglang.srt.layers.quantization.mxfp4_flashinfer_cutlass_moe import (
         Mxfp4FlashinferCutlassMoEMethod,
     )
@@ -388,15 +388,17 @@ def maybe_fuse_routed_scale_and_shared_add(
         Mxfp4MarlinMoEMethod,
     )
 
-    fused = isinstance(
-        experts.quant_method,
-        (
-            Mxfp4FlashinferTrtllmMoEMethod,
-            Mxfp4FlashinferCutlassMoEMethod,
-            Mxfp4MarlinMoEMethod,
-            ExpertPackMoEMethod,
-        ),
-    )
+    fused_methods = [
+        Mxfp4FlashinferTrtllmMoEMethod,
+        Mxfp4FlashinferCutlassMoEMethod,
+        Mxfp4MarlinMoEMethod,
+    ]
+    if not is_npu():
+        from sglang.srt.layers.quantization.expert_pack import ExpertPackMoEMethod
+
+        fused_methods.append(ExpertPackMoEMethod)
+
+    fused = isinstance(experts.quant_method, tuple(fused_methods))
     if fused:
         already_scaled = experts.should_fuse_routed_scaling_factor_in_topk
         if shared is not None:

python/sglang/srt/mem_cache/pool_host/mla.py

@@ -436,9 +436,14 @@ 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":