We have been working with onnx-mlir for years now at Groq. As a result we have added some additional features, bug fixes, and have some general divergence.

Time permitting could we discuss a strategy for us upstreaming our changes in the onnx-mlir Tuesday meeting.

Below is hopefully a comprehensive list of what we have

• Adding splitv11, squeezev11, unsqueezev11 decompositions
  Add decompositions for v11 split, squeeze, and unsqueeze #1702
• Topological Sorting check and implementation for onnx nodes
  Topological sort implementation #1725
• MatMul shape inference reuse for QLinearMatMul, MatMulInt, and MatMul
  add MatMulInt shape inference #1761
• GatherND verifer fixes
  Fix for GatherND verifier #1754
• FP16 protobuf ingestion
  ingest float16 #1701
• Scatter verifier fix
  normalize axis in ScatterElements verify #1760
• Lots of Constant Propagation (our implementation differs from the buffer implementation upstream) We ended up moving our divergent const prop into our codebase. And don't need this merged to delete our fork
• Add pad support for Conv Transpose
  Add more functionality to conv transpose shape inference #1771
• Add Upsample shape inference
  Add shape inference for Upsample #1776
• make valueInfo call permissive
  Add value info fix #1779
• derive kernel shape for conv op when not provided
  Derive kernel shape when not provided #1780

We see big benefits in minimizing our divergence from upstream. I'm actively working to upstream these changes. I want to make sure that we do it in a way that doesn't overwhelm.