[CUDA graphs] hotfix for test_graph_ (#64339)

Michael Carilli · facebook-github-bot · commit 24e50b8453d8 · 2021-08-31T22:34:10.000-07:00
Summary: Graphed workloads that try to capture a full backward pass must do warmup on a non-default stream. If warmup happens on the default stream, AccumulateGrad functions might tag themselves to run on the default stream, and therefore won't be capturable. ngimel and I suspect some test_cuda.py tests run with the default stream as the ambient stream, which breaks `test_graph_grad_scaling` because `test_graph_grad_scaling` does warmup on the ambient stream _assuming_ the ambient stream is a non-default stream. This PR explicitly sets a side stream for the warmup in `test_graph_grad_scaling`, which is what I should have done all along because it's what the new documentation recommends. I pushed the PR branch straight to the main pytorch repo because we need to run ci-all on it, and I'm not sure what the requirements are these days. Pull Request resolved: #64339 Reviewed By: mruberry Differential Revision: D30690711 Pulled By: ngimel fbshipit-source-id: 91ad75f46a11f311e25bc468ea184e22acdcc25a
diff --git a/test/test_cuda.py b/test/test_cuda.py
@@ -3683,8 +3683,13 @@ def test_graph_grad_scaling(self):
         static_grad = torch.ones_like(weight)
 
         # warmup
-        loss = (weight.half() * static_input).sum()
-        scaler.scale(loss).backward()
+        s = torch.cuda.Stream()
+        s.wait_stream(torch.cuda.current_stream())
+        with torch.cuda.stream(s):
+            loss = (weight.half() * static_input).sum()
+            scaler.scale(loss).backward()
+        torch.cuda.current_stream().wait_stream(s)
+
         opt.zero_grad(set_to_none=True)
 
         # capture
