Use .backward() with in-place grad mutations for the GA API

Author: rpsilva-aws

test/spmd/test_train_spmd_linear_model.py

@@ -74,10 +74,26 @@ def test_gradient_accumulation_matches(self):
     # Verify that the model losses are not zero, and that the runs match.
     assert all(loss != 0 for loss in baseline_grad_acc_losses)
     assert all(
-        torch.allclose(baseline_loss, checkpointing_loss, rtol=1e-4, atol=1e-8)
-        for baseline_loss, checkpointing_loss in zip(baseline_grad_acc_losses,
+        torch.allclose(baseline_loss, loop_grad_acc_loss, rtol=1e-4, atol=1e-8)
+        for baseline_loss, loop_grad_acc_loss in zip(baseline_grad_acc_losses,
                                                      loop_grad_acc_losses))
 
+    if not SKIP_GRADIENT_CHECKPOINTING:
+      print('Training loop with XLA\'s `While` gradient accumulation and '
+            'gradient checkpointing.')
+      with extended_argv(
+          COMMON_GRAD_ACC_ARGS +
+          ["--use_gradient_accumulation_loop", "--use_gradient_checkpointing"]):
+        loop_grad_acc_grad_chkpt_losses = train_and_evaluate_grad_acc()
+      assert all(
+          torch.allclose(
+              baseline_loss,
+              loop_grad_acc_grad_chkpt_loss,
+              rtol=1e-4,
+              atol=1e-8)
+          for baseline_loss, loop_grad_acc_grad_chkpt_loss in zip(
+              baseline_grad_acc_losses, loop_grad_acc_grad_chkpt_losses))
+
 
 if __name__ == '__main__':
   parser = argparse.ArgumentParser()

torch_xla/experimental/gradient_accumulation.py

@@ -179,6 +179,15 @@ def _prepare_fake_tensors(
 
   body_fn_inputs = (init_iterator, init_loss, *fake_iterable_tensors,
                     *fake_carried_tensors, *params, *grads)
+  # TODO - Fake the gradients once we are able to create placeholder tensors.
+  # Since the body is expected to do an in-place mutation of the gradients, we
+  # clone the gradients and use that as an input to the body. This will ensure
+  # that we retain a device data IR node in the graph. The cloned gradient will
+  # be updated to denote an IR operation (e.g. %add), and that can not be
+  # captured as a device data input for the other required computations, namely
+  # the condition and init for the XLA while loop.
+  for param in params:
+    param.grad = param.grad.clone()
   body_result = body_fn(init_iterator, init_loss, tuple(fake_iterable_tensors),
                         tuple(fake_carried_tensors), tuple(params),
                         tuple(grads))
@@ -373,10 +382,9 @@ def body_fn(iteri: torch.Tensor, _: torch.Tensor,
     else:
       loss, *carried_tensors = result
     loss /= context.num_gradient_steps
-    gradients = torch.autograd.grad(loss, model_parameters)
-    acc_grads = [prev_grad + grad for prev_grad, grad in zip(grads, gradients)]
-    return (iteri, loss, *iterable_tensors, *carried_tensors, *params,
-            *acc_grads)
+    loss.backward()
+    grads = [param.grad for param in params]
+    return (iteri, loss, *iterable_tensors, *carried_tensors, *params, *grads)
 
   init_grads = []
   # Initialize the gradients to zero.