Optimizing Transformer model size & inference beyond FP16 + ONNX (pruning/graph opt didn't help much)

Hi everyone, I’ve been working on optimizing a transformer-based neural network for both inference speed and model size, but I feel like I’ve hit a plateau and would appreciate some guidance. So far I’ve converted weights to FP16 (about 2× size reduction), exported and optimized with ONNX Runtime for inference speed, and tried both unstructured and structured pruning as well as ONNX graph optimizations, but none of these gave significant additional gains, and I’m still around ~162 MB per model.