Add guard for non-default stream in DDP's autograd engine callback (#…

…40115)

Summary:
Pull Request resolved: #40115

Closes #37790
Closes #37944

A user may wish to run DDP's forward + backwards step under a non-default CUDA stream such as those created by `with torch.cuda.Stream(stream)`. In this case, the user should be responsible for synchronizing events on this stream with other streams used in the program (per the documentation at https://pytorch.org/docs/stable/notes/cuda.html#cuda-semantics), but currently DDP has a bug which causes DDP under non-default streams to fail.

If a user does the following:
```
model = DDP(...)
loss = model(inptut).sum()
loss.backward()
grad = model.module.weight.grad()
average = dist.all_reduce(grad)
```

There is a chance that `average` and `grad` will not be equal. This is because the CUDA kernels corresponding to the  `all_reduce` call may run before `loss.backward()`'s kernels are finished. Specifically, in DDP we copy the allreduced gradients back to the model parameter gradients in an autograd engine callback, but this callback runs on the default stream. Note that this can also be fixed by the application synchronizing on the current stream, although this should not be expected, since the application is not using the current stream at all.

This PR fixes the issue by passing the current stream into DDP's callback.

Tested by adding a UT `test_DistributedDataParallel_non_default_stream` that fails without this PR
ghstack-source-id: 106481208

Differential Revision: D22073353

fbshipit-source-id: 70da9b44e5f546ff8b6d8c42022ecc846dff033e

test/distributed/test_distributed.py

@@ -1906,6 +1906,45 @@ def test_DistributedDataParallel_requires_grad(self):
         # a module without gradients shouldn't be accepted
         self.assertRaises(AssertionError, lambda: nn.parallel.DistributedDataParallel(nn.Module()))
 
+    @unittest.skipIf(
+        BACKEND != "nccl" and BACKEND != "gloo",
+        "Only NCCL and GLOO backend support DistributedDataParallel",
+    )
+    @skip_if_lt_x_gpu(2)
+    @skip_if_rocm
+    def test_DistributedDataParallel_non_default_stream(self):
+        stream = torch.cuda.Stream()
+        rank = self.rank
+        with torch.cuda.stream(stream):
+            net = torch.nn.parallel.DistributedDataParallel(
+                torch.nn.Linear(1, 1, bias=False).cuda(rank), device_ids=[rank]
+            )
+            for i in range(1000):
+                # Clear gradients manually
+                grad = net.module.weight.grad
+                if grad is not None:
+                    grad.detach_()
+                    grad.zero_()
+                # Forward + BW
+                batch = torch.tensor([rank]).float().cuda(rank)
+                loss = net(batch).sum()
+                loss.backward()
+                # For each worker, the gradient on the weight should be worker_rank.
+                grad = net.module.weight.grad
+                avg = grad.clone()
+                # All-reducing the gradient averages should give us the gradient
+                # average. If not, then one of the workers has not correctly
+                # written back the averaged gradient before this all-reduce call.
+                dist.all_reduce(avg)
+                world_size = int(os.environ["WORLD_SIZE"])
+                avg.div_(world_size)
+                expected_grad = sum(i for i in range(world_size)) / world_size
+                self.assertEqual(
+                    avg[0, 0],
+                    expected_grad,
+                    msg=f"Expected gradient of {expected_grad} but got {avg} on rank {self.rank}",
+                )
+
     @unittest.skipIf(BACKEND != 'nccl' and BACKEND != 'gloo',
                      "Only Nccl & Gloo backend support DistributedDataParallel")
     @skip_if_no_gpu

torch/csrc/distributed/c10d/reducer.cpp

@@ -3,6 +3,7 @@
 #include <functional>
 
 #include <c10/core/DeviceGuard.h>
+#include <c10/core/StreamGuard.h>
 #include <c10/util/Exception.h>
 #include <torch/csrc/autograd/engine.h>
 #include <torch/csrc/autograd/function_hook.h>
@@ -483,8 +484,17 @@ void Reducer::mark_variable_ready(VariableIndex index) {
     }
     local_used_work_ = process_group_->allreduce(local_used_maps_dev_);
 
+    // The autograd engine uses the default stream when running callbacks, so we
+    // pass in the current CUDA stream in case it is not the default.
+    c10::DeviceType deviceType = replica.contents.device().type();
+    const c10::impl::VirtualGuardImpl guard =
+        c10::impl::VirtualGuardImpl{deviceType};
+    const c10::Stream currentStream =
+        guard.getStream(replica.contents.device());
     torch::autograd::Engine::get_default_engine().queue_callback([=] {
       std::unique_lock<std::mutex> lock(this->mutex_);
+      // Run callback with the current stream
+      c10::OptionalStreamGuard currentStreamGuard{currentStream};
       this->finalize_backward();
       // Rebuild bucket if this is the first time to rebuild
       if (!rebuilt_params_.empty()) {