[ascend]optimize: remove sync 2 (DeepLink-org#898)

* feat: remove sync

Author: yewentao256

impl/ascend_npu/torch_npu/csrc/CopyKernel.cpp

@@ -385,6 +385,10 @@ at::Tensor& NPUNativeFunctions::copy_(at::Tensor& self, const at::Tensor& src, b
         internal_set_names_inplace(self, names);
     }
 
+    // Param `non_blocking`: if True and this copy is between CPU and GPU,
+    // the copy may occur asynchronously with respect to the host.
+    // For other cases, this argument has no effect.
+    // https://pytorch.org/docs/stable/generated/torch.Tensor.copy_.html
     if (at_npu::key::isDeviceTensor(self)) {
         if (at_npu::key::isDeviceTensor(src)) {
             copy_d2d(self, src, non_blocking);
@@ -396,9 +400,6 @@ at::Tensor& NPUNativeFunctions::copy_(at::Tensor& self, const at::Tensor& src, b
             copy_d2h(self, src, non_blocking);
         }
     }
-    if (!non_blocking) {
-        c10_npu::getCurrentNPUStream().synchronize();
-    }
     return self;
 }
 

impl/ascend_npu/torch_npu/csrc/DIOPIAdapter.cpp

@@ -2730,10 +2730,7 @@ NPUStream getCurrentNPUStream(c10::DeviceIndex device_index) {
 
 NPUStream getCurrentSecondaryStream(c10::DeviceIndex device_index) { return getCurrentNPUStream(device_index); }
 
-void NPUStream::synchronize() const {
-    NPU_CHECK_ERROR(aclrtSynchronizeStream(aclStream_));
-    NPU_CHECK_ERROR(aclrtSynchronizeDevice());
-}
+void NPUStream::synchronize() const { NPU_CHECK_ERROR(aclrtSynchronizeStream(aclStream_)); }
 
 aclError queue::LaunchAsyncCopyTask(void* dst, size_t dstLen, void* src, size_t srcLen, aclrtMemcpyKind kind) {
     c10_npu::NPUStream stream = c10_npu::getCurrentNPUStream();