FP4 KV cache with no degradation sparks debate on aggressive inference compression
stochasticchasm · x · 2026-09-11
Engineers are discussing an aggressive inference optimization: FP4 KV cache with no noticeable degradation, combined with compressing KVs via CSA, leaves models barely using any KV cache.
The author adds nuance: within a single request you still need the KV cache in HBM/RAM, but between requests re-prefilling is negligible compute at 500K context. Kernel reduction order and similar details may matter a lot for performance.
More from Infra
- NVIDIA ships NVFP4-quantized Qwen3.8-27B, trending on Hugging Face — nvidia · 2026-09-11
- Qwen 3.8 125B-A6B runs 15.3% faster on Mac via speculative decoding on mlx.fast — TheMoonMidas · 2026-09-11
- Olam Labs CEO: only compute and data remain as bottlenecks to AGI — garrytan · 2026-09-11
- Open-sourced inference acceleration for structure-based models ships benchmarked and documented — AllThingsApx · 2026-09-11
- NVIDIA's $89B quarterly data center revenue means compute is no longer an IT expense — PeterDiamandis · 2026-09-11
- Pure-CPU DeepSeek V4.1 Inference Project Hits 6 TPS on a Xeon for Overnight Agent Jobs — Qwen30bEnjoyer · 2026-09-11