-
Notifications
You must be signed in to change notification settings - Fork 84
7. Limitations and Known Issues
This page outlines current limitations and known issues in our project:
We currently only support the following dialects for your MLIR file:
- TensorFlow (tf)
- TensorFlow Lite (tfl)
- StableHLO (stablehlo)
The list will expand in the future. You can also use our Adapter Extension API to build your own custom adapters. The example can be found here.
We only support PyTorch models in ExportedProgram format. You can obtain the ExportedProgram through torch.export.export.
The recommended way to visualize ExportedProgram is to use our Python API as it avoids the potential package version conflicts and you don't have to serialize your exported program into .pt2 file.
However, you are also welcome to save it and upload it to UI by using
torch.export.save(exported_program, 'exported_program.pt2')
Note:
- ExportedProgram format is not stable and doesn't have API/schema stability.
- The PyTorch version used for exporting the model must match with the version of torch installed locally to ensure compatibility.
- The current adapter implementation supports
torch >= 2.2.0. No compatibility guarantee for versions before2.2.0.
Visualization of the original JAX model is not supported yet.
You need to export the JAX model either to a TF SavedModel or a StableHLO MLIR. Here is a tutorial guiding you to set up with both approaches.
Some ops that generate unnamed nested regions might not contain all the op information, eg. stablehlo.reduce_scatter, stablehlo.map, stablehlo.sort, etc. They will be displayed in Model Explorer, but their nested regions might be missing for now.
Windows is supported exclusively through the Windows Subsystem for Linux (WSL). Native Windows installations are not currently supported.
If you encounter issues related to these limitations, please search our issue tracker for potential solutions or workarounds. If your issue is not already documented, please create a new issue with detailed information to help us investigate and address the problem.