[PT FE] Report about FakeTensor during conversion (#27203)

**Details:** Report about FakeTensor during conversion **Ticket:** 149989 --------- Signed-off-by: Kazantsev, Roman <[email protected]> Co-authored-by: Maxim Vafin <[email protected]>
openvinotoolkit · Oct 23, 2024 · 6458855 · 6458855
1 parent ca926d6
commit 6458855
Showing 1 changed file with 9 additions and 0 deletions.
diff --git a/src/bindings/python/src/openvino/frontend/pytorch/utils.py b/src/bindings/python/src/openvino/frontend/pytorch/utils.py
@@ -46,6 +46,15 @@ def get_type_from_py_type(value):
 
 
 def torch_tensor_to_ov_const(torch_t: torch.Tensor, shared_memory=True):
+    is_fake_tensor = False
+    try:
+        from torch._prims import FakeTensor
+        is_fake_tensor = isinstance(torch_t, FakeTensor)
+    except:
+        pass
+    assert not is_fake_tensor, '`FakeTensor` is found in the graph during conversion. ' \
+                               'In order to avoid `FakeTensor` in the traced model, ' \
+                               'try to infer the model before exporting.'
     torch_t = torch_t.contiguous()
     if torch_t.dtype == torch.bfloat16:
         # reinterpret bfloat16 data as float16 to allow conversion to numpy