You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Currently, the TorchScript path routinely tests model compatibility for ResNet18, MobileNetV2, EfficientNetB0, ResNet18 with Half precision, and HuggingFace BERT base uncased. I have added a PR which tests the FX path (via specifying ir="fx" directly in torch_tensorrt.compile(...)) for all of these models except BERT, which currently fails to compile with this method and also does not compile using torch_tensorrt.fx.compile(...) with either is_aten=True or is_aten=False, via the current main branch (deda87b). The model also fails to compile with torch._dynamo.optimize(...). An issue will be raised regarding compilation of HuggingFace BERT in FX. This (current) investigative issue will be closed and the PR will remain as a draft until BERT model compilation in FX is supported.
No description provided.
The text was updated successfully, but these errors were encountered: