[Relax] NDArray Cache Update with DLTensor Support (#16464)

As NDArray on RPC devices only returns a DLTensor, we add support
for DLTensor in NDArray Cache.

It's not easy to add test cases as we cannot create a raw DLTensor
in Python interface.

src/runtime/relax_vm/ndarray_cache_support.cc

@@ -267,7 +267,28 @@ class NDArrayCache {
 };
 
 TVM_REGISTER_GLOBAL("vm.builtin.ndarray_cache.get").set_body_typed(NDArrayCache::Get);
-TVM_REGISTER_GLOBAL("vm.builtin.ndarray_cache.update").set_body_typed(NDArrayCache::Update);
+TVM_REGISTER_GLOBAL("vm.builtin.ndarray_cache.update").set_body([](TVMArgs args, TVMRetValue* rv) {
+  CHECK(args.size() == 2 || args.size() == 3);
+  String name = args[0];
+  bool is_override = args.size() == 2 ? false : args[2];
+
+  NDArray arr;
+  if (args[1].type_code() == kTVMNDArrayHandle) {
+    arr = args[1];
+  } else {
+    // We support converting DLTensors to NDArrays as RPC references are always DLTensors
+    DLTensor* tensor = args[1];
+    std::vector<int64_t> shape;
+    for (int64_t i = 0; i < tensor->ndim; i++) {
+      shape.push_back(tensor->shape[i]);
+    }
+    NDArray arr = NDArray::Empty(shape, tensor->dtype, tensor->device);
+    arr.CopyFrom(tensor);
+    TVMSynchronize(arr->device.device_type, arr->device.device_id, nullptr);
+  }
+
+  NDArrayCache::Update(name, arr, is_override);
+});
 TVM_REGISTER_GLOBAL("vm.builtin.ndarray_cache.remove").set_body_typed(NDArrayCache::Remove);
 TVM_REGISTER_GLOBAL("vm.builtin.ndarray_cache.clear").set_body_typed(NDArrayCache::Clear);
 TVM_REGISTER_GLOBAL("vm.builtin.ndarray_cache.load").set_body_typed(NDArrayCache::Load);