Announcement
QUASAR targets a mismatch in 2–4-bit model training
QUASAR’s creators say it consistently lowers loss compared with all QAT baselines without adding inference overhead.
TLDR
The people behind QUASAR say quantization-aware training at 2–4 bits updates high-precision weights using loss and gradients calculated on low-bit reconstructions—the weights used at inference. They say QUASAR addresses that mismatch by changing how weights are reconstructed during training. They also say it can be combined with most other QAT methods.
Combined views
1.7K
2 Sources, first seen ago
