We have an educational application that we’d like to deploy in WebGL but we’re having performance issues. We have a custom ASR model that uses an LSTM for most of the heavy lifting but LSTM isn’t supported in GPUPixelBackend. Are there plans to support any RNNs in WebGL by any mechanism?
Is CPU an option for you?
We should fallback LSTM on cpu, but unfortunately we don’t atm.
You could edit the model to manually force that layer on the cpu if that works for you.
Also is WebGPU an option?