[TE] fix: fix heterogeneous rdma transport error (kvcache-ai#1657)

---------

Co-authored-by: sunwenhan <sunwenhan@xfusion.com>

Author: 2 people

mooncake-transfer-engine/example/transfer_engine_heterogeneous_ascend_perf_initiator.cpp

@@ -160,6 +160,9 @@ std::string loadNicPriorityMatrix() {
            " \"cpu:1\": [[" +
            device_names +
            "], []], "
+           " \"npu:0\": [[" +
+           device_names +
+           "], []], "
            " \"cuda:0\": [[" +
            device_names +
            "], []], "

mooncake-transfer-engine/src/transport/ascend_transport/heterogeneous_rdma_transport/heterogeneous_rdma_transport.cpp

@@ -4,13 +4,19 @@
 namespace mooncake {
 
 namespace {
+// ACL memory location types for aclrtPtrAttributes::location.type:
+//   0 = ACL host memory, 1 = device memory, 2 = regular CPU memory (malloc)
+static constexpr int kDeviceMemoryLocationType = 1;
+
 bool isCpuMemory(void *addr) {
     aclrtPtrAttributes attributes{};
     if (int ret = aclrtPointerGetAttributes(addr, &attributes)) {
-        LOG(ERROR) << "aclrtPointrtGetAttributes error, ret: " << ret;
-        return false;
+        // If ACL cannot identify the pointer, treat it as regular CPU memory.
+        LOG(WARNING) << "aclrtPointerGetAttributes failed for addr " << addr
+                     << ", ret: " << ret << ". Treating as CPU memory.";
+        return true;
     }
-    return (attributes.location.type == 0);
+    return (attributes.location.type != kDeviceMemoryLocationType);
 }
 }  // namespace