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.
More from Infra
- Amazon is stockpiling $5B of Nvidia GB300 racks it can't power up yet — AIFlow_ML · 2026-10-22
- Ornith-1.5-122B MoE runs on a DGX Spark: 52 tok/s, 256K context recipe — lmoroney · 2026-10-22
- eMudhra intermediate CA private key compromise flagged as P1 bug at Mozilla — basedjensen · 2026-10-22
- 10 B300 nodes available now in the US at $7.5/hr with 1-year commitment — NirantK · 2026-10-22
- Vibe-Coded Android LLM App Targets Adreno GPU and Hexagon NPU, Seeks Testers — Puzzleheaded_Acadia1 · 2026-10-22
- Mac local inference survey: dense 27B remains the go-to for coding — DotaMate · 2026-10-22