onnxruntime_backend
onnxruntime_backend copied to clipboard
Improve autocomplete to make it more robust against partial model configuration
Is your feature request related to a problem? Please describe. Currently, the auto-complete function does nothing if model config provides even a single output and input. See here: https://github.com/triton-inference-server/onnxruntime_backend/blob/main/src/onnxruntime.cc#L652-L680. The specification of the input/output may be incomplete. It may be missing dimension or data-type fields. However, the auto-complete will still choose to skip completing these I/O configs.
Describe the solution you'd like Instead of skipping auto-complete logic entirely when there are inputs/outputs present, auto-complete should still ensure that all the I/O have complete configuration and not fail.
Describe alternatives you've considered Alternative is for user to ensure they provide complete or no config information. This is little restrictive on the user. For example, they can not only specify partial config for handling special cases and still use auto-complete to fill in the missing information for them.