Disaggregated Quantization Boosts 1-bit LLM Accuracy by 32+ Points and TTFT by 1.78x
ISTA-DASLab · hf · 2026-09-29
ISTA-DASLab proposes Disaggregated Quantization (DQ), which specializes compute formats, weights, and storage placement separately for prefill and decode phases. Highlights:
- On Qwen 3 and Gemma 3, removing activation quantization only on decode improves decode-heavy task accuracy at no extra inference cost
- Training compute-native prefill weights matches or beats weight-only inference at 2-3-bit decode
- Training an NVFP4 prefiller on top of released Qwen3.8-27B GGUF decoders lifts 1-bit accuracy by 32.5 points on MMLU-Pro and 35.3 on MMMU-Pro without touching the decode checkpoint
- Offloaded disaggregated prefill (ODP) streams weights from SSD, delivering a 1.78x time-to-first-token speedup at 8K prompts in llama.cpp
- Validated in vLLM disaggregated serving and via post-training quantization on models up to 2.8T parameters
More from Infra
- Sarvam AI Launches Token Factory: India-Hosted Inference for Open Models — itsOmSarraf_ · 2026-11-16
- BIS Issues Interim Final Rule Addressing Remote Access to Advanced AI Chips — apples_jimmy · 2026-11-16
- Jailbroken PS4 runs a 9B open-source model at up to 43 tok/s by talking directly to its 2013-era GPU — AIFlow_ML · 2026-11-16
- Mobile AI Factory: agents auto-build on-device SLMs with 19x less GPU time — pmttyji · 2026-11-16
- Stratechery's Fictional Obituary: Imagining Jensen Huang's 2026 Exit From Nvidia — Gregory_C_Allen · 2026-11-16
- vLLM serves DeepSeek V4.1-Flash at 954K tokens/s per GPU, $38.96/GPU-hour on GB300 NVL72 — bookwormengr · 2026-11-16