PACE: Post-Hoc 4-Bit Quantization of Linear-Attention Recurrent States Without Fine-Tuning

huberlin · hf · 2026-10-22

Humboldt-Universität zu Berlin introduces PACE, a post-hoc 4-bit quantizer for linear-attention recurrent states. Instead of rounding every token, PACE rounds once every few tokens with a small buffer and uses column-wise scales exploiting how large gated-delta state entries cluster in columns. This cuts error accumulation, keeping accuracy at 4 bits without fine-tuning and letting more sequences share an accelerator.

Original post →

More from Infra

Infra channel →