quic/aimet

Supports both post-training quantization (PTQ) and quantization-aware training (QAT) with specialized techniques like AdaRound, SeqMSE, and Data-Free Quantization that minimize accuracy loss during INT8 conversion. Integrates directly with PyTorch and ONNX model pipelines, offering automated layer-wise optimization and compression techniques including spatial SVD and channel pruning. Provides visualization tools for weight ranges and per-layer compression sensitivity to guide optimization decisions.

2,566 stars. Actively maintained with 87 commits in the last 30 days.