Fix wrong inplace acc grad

oneflow/core/autograd/autograd_engine.cpp

@@ -48,8 +48,15 @@ Maybe<void> CopyOrAccGrad(AutogradMeta* autograd_meta, bool autograd_mode) {
   }
   if (autograd_meta->acc_grad()) {
     DevVmDepObjectConsumeModeGuard guard(DevVmDepObjectConsumeMode::NONE);
+    // Should not inplace accumulate grad. For example,
+    // >>> z = x + y
+    // >>> p = x / z
+    // >>> p.sum().backward()
+    //
+    // As we know that dx = dz + dp / z and dy = dz, so it will lead to wrong value
+    // for dy if dx is shared with dz.
     const auto& output =
-        JUST(functional::Add(autograd_meta->acc_grad(), current_grad, /*inplace=*/true));
+        JUST(functional::Add(autograd_meta->acc_grad(), current_grad, /*inplace=*/false));
     JUST(autograd_meta->set_acc_grad(output));
   } else {
     JUST(autograd_meta->set_acc_grad(current_grad));

oneflow/core/framework/tensor_arg.cpp

@@ -29,7 +29,14 @@ Maybe<void> TensorArg::PushPartialTensor(const std::shared_ptr<Tensor>& partial_
   if (!acc_tensor_) {
     acc_tensor_ = partial_tensor;
   } else {
-    acc_tensor_ = JUST(functional::Add(partial_tensor, acc_tensor_, /*inplace=*/true));
+    // Should not inplace accumulate grad. For example,
+    // >>> z = x + y
+    // >>> p = x / z
+    // >>> p.sum().backward()
+    //
+    // As we know that dx = dz + dp / z and dy = dz, so it will lead to wrong value
+    // for dy if dx is shared with dz.
+    acc_tensor_ = JUST(functional::Add(partial_tensor, acc_tensor_, /*inplace=*/false));
   }
   return Maybe<void>::Ok();
 }

python/oneflow/test/modules/test_autograd.py

@@ -20,6 +20,7 @@
 import numpy as np
 from test_util import GenArgList
 
+from automated_test_util import *
 import oneflow as flow
 import oneflow.unittest
 
@@ -85,6 +86,14 @@ def test_autograd_interface(test_case):
         for arg in GenArgList(arg_dict):
             arg[0](test_case, *arg[1:])
 
+    @autotest(n=10, auto_backward=True, rtol=1e-3, atol=1e-3)
+    def test_accumulate_grad(test_case):
+        device = random_device()
+        ndim = random(1, 4).to(int)
+        x = random_pytorch_tensor(ndim=ndim, requires_grad=True).to(device)
+        y = random_pytorch_tensor(ndim=ndim, requires_grad=True).to(device)
+        return x / (x + y)
+
 
 if __name__ == "__main__":
     unittest.main()