My teammate @plan9better is implementing an inference pipeline for our speech to text model based on wav2vec2 architecutre. Our inference pipeline has some bespose extra heads that we pipe hidden states thru. In order to implement it from C++ we are using some ops from ctranslate2 to define the layers. Are you open to including bindings to those ops in ctranslate2-rs? These are bindings to C++ functions that allow defining "custom models" from userspace. It's useful when you have some bespoke models that you wanna use with ctranslate2.
My teammate @plan9better is implementing an inference pipeline for our speech to text model based on wav2vec2 architecutre. Our inference pipeline has some bespose extra heads that we pipe hidden states thru. In order to implement it from C++ we are using some ops from ctranslate2 to define the layers. Are you open to including bindings to those ops in ctranslate2-rs? These are bindings to C++ functions that allow defining "custom models" from userspace. It's useful when you have some bespoke models that you wanna use with ctranslate2.