[dynamo] graph break on setattr requires_grad (pytorch#113163)

Main: `RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn`
This PR: graph breaks and eager applies the mutation, new tensors are tracked

Fixes pytorch#109505 (the original bug does not occur, but a new bug where the mutation isn't applied - because AOTAutograd is not `requires_grad` mutation aware - is mitigated)

Pull Request resolved: pytorch#113163
Approved by: https://github.com/bdhirsh

Author: jon-chuang

test/dynamo/test_repros.py

@@ -3497,6 +3497,33 @@ def test_addr_alpha_beta_out(self):
         compiled_fn(inp, vec1, vec2, alpha=alpha, beta=beta, out=compile_out)
         self.assertTrue(same(out, compile_out))
 
+    def test_setattr_requires_grad_graph_breaks(self):
+        def fn(x):
+            z = x + 4
+            x.requires_grad = True
+            y = x * z
+            return y
+
+        for backend in ["count", "eager", "aot_eager"]:
+            if backend == "count":
+                backend = CompileCounter()
+            opt_fn = torch.compile(fn, backend=backend)
+
+            eager = torch.zeros(5)
+            compiled = eager.clone()
+
+            out_eager = fn(eager)
+            out_opt = opt_fn(compiled)
+
+            self.assertEqual(out_eager, out_opt)
+
+            out_eager.sum().backward()
+            out_opt.sum().backward()
+
+            self.assertEqual(eager, compiled)
+            if isinstance(backend, CompileCounter):
+                self.assertEqual(backend.frame_count, 2)  # graph breaks
+
     def test_inductor_no_recursionerror_on_for_loops(self):
         def forward(x):
             for _ in range(1000):

torch/_dynamo/variables/builtin.py

@@ -1199,7 +1199,13 @@ def call_setattr(
             tx.output.side_effects.is_attribute_mutation(obj)
             and name_var.is_python_constant()
         ):
-            tx.output.side_effects.store_attr(obj, name_var.as_python_constant(), val)
+            name = name_var.as_python_constant()
+            if name == "requires_grad" and isinstance(obj, variables.TensorVariable):
+                unimplemented(
+                    "mutating requires_grad can introduce a new leaf from non-leaf or vice versa in "
+                    "the middle of the graph, which aot_autograd does not currently know how to handle. "
+                )
+            tx.output.side_effects.store_attr(obj, name, val)
             return val
         elif isinstance(obj, variables.UserDefinedObjectVariable):
             unimplemented(