[RFC] Features you want to see

[ma7dev]

In this issue, you are welcome to discuss features that you want to see included in the library.

[ma7dev]

The currently supported operations for Dynamic Quantization based on 1 are:

• [x] nn.Linear
• [ ] nn.LSTM
• [ ] nn.GRU
• [x] nn.RNNCell
• [x] nn.GRUCell
• [x] nn.LSTMCell
• [x] nn.EmbeddingBad

We might need to look into Static Quantization as they provide coverage for nn.Conv1d/2d/3d, nn.Embedding, and Activations 1.

Another idea is that FX Graph mode will be the go-to for future development but it is under construction, so we might be able to contribute to PyTorch and utilize it in our library 2.

• Introduction to Quantization on PyTorch
• (PROTOTYPE) FX GRAPH MODE QUANTIZATION USER GUIDE
• (PROTOTYPE) FX GRAPH MODE POST TRAINING STATIC QUANTIZATION
• (PROTOTYPE) FX GRAPH MODE POST TRAINING DYNAMIC QUANTIZATION
• DYNAMIC QUANTIZATION - testing model
  • Model size (when saving the model in the disk, what its size)
  • Latency (speed to run forward pass)
  • Accuracy
• Quantization Triage - hunting ground