-
Notifications
You must be signed in to change notification settings - Fork 3.8k
Closed
Labels
frontend:onnxpython/tvm/relay/frontend/onnx.pypython/tvm/relay/frontend/onnx.pytype:rfc-trackingRFC progress tracking. Ref: https://github.com/apache/tvm-rfcsRFC progress tracking. Ref: https://github.com/apache/tvm-rfcs
Description
Right now ONNX converter to Relay works by using operator name to match up with proper converter.
The issue is we increasingly have to deal with operators from different onnx modules/domains. Onnx nodes also have this sort of namespace https://github.com/onnx/onnx/blob/main/docs/IR.md#nodes too. In the future, we might have name collisions between like names in different namespaces. We should handle this properly.
Not a big deal until it happens.
List of operators not under default onnx operator domain (please add):
- Gelu (https://com.microsoft/)
- EmbedLayerNormalization (https://com.microsoft/)
- SkipLayerNormalization (https://com.microsoft/)
- Attention (https://com.microsoft/)
Metadata
Metadata
Assignees
Labels
frontend:onnxpython/tvm/relay/frontend/onnx.pypython/tvm/relay/frontend/onnx.pytype:rfc-trackingRFC progress tracking. Ref: https://github.com/apache/tvm-rfcsRFC progress tracking. Ref: https://github.com/apache/tvm-rfcs