1 related articles
A deep dive into how quantization-aware training (QAT) makes models robust to rounding error, plus BitNet's 1-bit and 1.58-bit ternary weight schemes, including STE, latent weights, and absmean quantization.
