I'd like to start off with a thank you for developing this library - it's awesome, and so convenient! :)
Cutting to the chase: your implementation of fastembed uses ort to handle the underlying ONNX inference (i.e. the actual "embed" in "fastembed"). While this allows it to be performant on supported platforms, my understanding is that it leads to a more complex build process and makes targeting *-musl (e.g.x86_64-unknown-linux-musl, my go-to for portable Linux programs) platforms a chore: to do so, you must build the onnxruntime library for musl yourself, and then configure the ort build to use your local onnxruntime with ORT_LIB_LOCATION=/path/to/musl/onnxruntime/build during compilation time.
Are there any plans to switch the ONNX inference library to something other than ort, like sonos/tract, or would such a change be better suited to a fork?
I'd like to start off with a thank you for developing this library - it's awesome, and so convenient! :)
Cutting to the chase: your implementation of
fastembed
usesort
to handle the underlying ONNX inference (i.e. the actual "embed" in "fastembed"). While this allows it to be performant on supported platforms, my understanding is that it leads to a more complex build process and makes targeting*-musl
(e.g.x86_64-unknown-linux-musl
, my go-to for portable Linux programs) platforms a chore: to do so, you must build theonnxruntime
library for musl yourself, and then configure theort
build to use your localonnxruntime
withORT_LIB_LOCATION=/path/to/musl/onnxruntime/build
during compilation time.Are there any plans to switch the ONNX inference library to something other than
ort
, like sonos/tract, or would such a change be better suited to a fork?