[Tracking Issue] Pre-quantized RNN Operators Support (e.g., UnidirectionalSequenceLSTM from TFlite)

apache / tvm

Open deep learning compiler stack for cpu, gpu and specialized accelerators

https://tvm.apache.org/

Apache License 2.0

11.43k stars 3.41k forks source link

[Tracking Issue] Pre-quantized RNN Operators Support (e.g., UnidirectionalSequenceLSTM from TFlite) #13150

Open dftercero opened 1 year ago

dftercero commented 1 year ago

This issue is to track progress for Pre-quantized RNN Operators

It'd be great to expand support for recurrent quantized operators, such as the UnidirectionalSequenceLSTM from TensorFlow. Are there any timelines for this in the current roadmap?

Resources:

TFlite frontend: https://github.com/apache/tvm/blob/main/python/tvm/relay/frontend/tflite.py
Tensorflow fused RNN Operator: https://www.tensorflow.org/lite/models/convert/rnn

dftercero commented 1 year ago

@lhutton1 @leandron You guys have done a lot of work recently with quantized operators from TFLite based on #9187. Any thoughts on this? Thanks in advanced

lhutton1 commented 1 year ago

Hi @dftercero, I don't believe there is any ongoing work to support it yet. I don't know much about this operator, although, after a quick search of the codebase it looks like a bit more work would be needed to support the quantized variant of LSTM in TVM rather than simply making a conversion in the frontend

amroceva commented 1 month ago

Hi @hpanda-naut @lhutton1, is there any ongoing work to support the quantized version of it ? if no, can you give hints about the needed work to do so, I would like to contribute...