transformer-optimization

Author shares experience hitting diminishing returns with FP16 + ONNX + pruning on 162 MB transformer, seeks advice on next best steps among quantization, distillation, low-rank factorization, or hardware-specific tricks.