[MLU] fix copy error (#45194)

paddle/fluid/operators/gaussian_random_op_mlu.cc

@@ -45,6 +45,7 @@ class MLUGaussianRandomKernel : public framework::OpKernel<T> {
     auto& dev_ctx =
         context.template device_context<paddle::platform::MLUDeviceContext>();
     framework::TensorCopy(cpu_tensor, context.GetPlace(), dev_ctx, tensor);
+    dev_ctx.Wait();
   }
 };
 

paddle/fluid/operators/optimizers/adam_op_mlu.cc

@@ -68,6 +68,7 @@ class AdamMLUKernel : public framework::OpKernel<T> {
       std::vector<bool> skip_update_vec;
       paddle::framework::TensorToVector(
           *skip_update_tensor, ctx.device_context(), &skip_update_vec);
+      ctx.device_context().Wait();
       skip_update = skip_update_vec[0];
     }
     // skip_update=true, just copy input to output, and TensorCopy will call
@@ -286,6 +287,7 @@ class AdamWMLUKernel : public AdamMLUKernel<T> {
       std::vector<bool> skip_update_vec;
       paddle::framework::TensorToVector(
           *skip_update_tensor, ctx.device_context(), &skip_update_vec);
+      ctx.device_context().Wait();
       skip_update = skip_update_vec[0];
     }
     bool with_decay = ctx.Attr<bool>("with_decay");

paddle/fluid/pybind/tensor_py.h

@@ -439,7 +439,11 @@ void SetTensorFromPyArrayT(
     platform::Place tmp_place = place;
     platform::MLUDeviceGuard guard(tmp_place.device);
     auto dst = self->mutable_data<T>(place);
-    paddle::platform::MLUMemcpyH2DSync(dst, array.data(), array.nbytes());
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
+    auto dev_ctx = static_cast<platform::MLUDeviceContext *>(pool.Get(place));
+    paddle::platform::MLUMemcpyH2DAsync(
+        dst, array.data(), array.nbytes(), dev_ctx->stream());
+    dev_ctx->Wait();
 #else
     PADDLE_THROW(platform::errors::PermissionDenied(
         "Cannot use MLUPlace in CPU/GPU version, "