Specialize symfloats when getting fake value involves complex args (pytorch#140832)

Fixed `PYTORCH_TEST_WITH_DYNAMO=1 tlp python test/test_sparse_csr.py TestSparseCSRCPU.test_sampled_addmm_cpu_complex64` when `specialize_float=False`

Pull Request resolved: pytorch#140832
Approved by: https://github.com/ezyang
ghstack dependencies: pytorch#140830

Author: bobrenjc93

torch/_dynamo/utils.py

@@ -2217,7 +2217,9 @@ def get_fake_value(node, tx, allow_non_graph_fake=False):
         # no matter it's lazy module or not, we should copy to fake mode.
         nnmodule = deepcopy_to_fake_tensor(nnmodule, tx.fake_mode)
 
-    if node.name in ["interpolate", "is_integer", "wrapped_gradient"]:
+    if node.name in ["interpolate", "is_integer", "wrapped_gradient"] or any(
+        isinstance(a, complex) for a in args
+    ):
         # We need to specialize symfloats for now. Eventually we should do a tensorify pass in dynamo.
         args = tuple(
             float(arg)