Optimized multi-layer LSTM for GPU

Hello everyone,

I am looking for a TVM implementation of an optimized multi-layer LSTM for GPU. I tried this one : https://github.com/dmlc/tvm/blob/master/topi/recipe/rnn/lstm.py but it only supports one layer and I don’t know if it is an optimal implementation.

Do you know where I can find this implementation ?

Thank you.

Hmmm we didn’t implement a multi layer one yet, but you might be interested in this tutorial

Thank you for your response.

Is this implementation https://github.com/dmlc/tvm/blob/master/topi/recipe/rnn/lstm.py an optimal one ? So I can modify it to support multi-layer lstm using the tutorial that you cited ?

This is a good one, but defining “optimal” is non-trivial