Revert "[Dynamo] Allow inlining into AO quantization modules (#152934)" (#161567)

This reverts commit 20e2ca3e29ce9eb33eef17db077696222c175764.

Fixes https://github.com/pytorch/pytorch/issues/157434

Pull Request resolved: https://github.com/pytorch/pytorch/pull/161567
Approved by: https://github.com/Lucaskabela

test/dynamo/test_repros.py

@@ -6486,21 +6486,6 @@ def forward(self, s77 : torch.SymInt, s27 : torch.SymInt, L_x_ : torch.Tensor):
         with torch.no_grad():
             model(x)
 
-    def test_ao_fake_quantize_tracing(self):
-        import torch.ao.quantization.fake_quantize
-
-        q = torch.ao.quantization.FusedMovingAvgObsFakeQuantize()
-
-        def fn(x):
-            return q(x)
-
-        x = torch.ones(2, 2)
-        opt_fn = torch.compile(fn, backend="eager", fullgraph=True)
-        res = opt_fn(x)
-        eager_res = fn(x)
-
-        self.assertEqual(res, eager_res)
-
     def test_typed_dict(self):
         class LlavaImagePixelInputs(TypedDict):
             type: Literal["pixel_values"]

torch/_dynamo/trace_rules.py

@@ -3420,7 +3420,6 @@ MOD_INLINELIST = [
     "torch._tensor",
     "torch.amp.autocast_mode",
     "torch.ao.nn",
-    "torch.ao.quantization.fake_quantize",
     "torch.autograd.function",
     "torch.backends.cuda",
     "torch.cuda.amp.autocast_mode",