KVA projectors boost local LLM prefill up to 1.85x with tunable perplexity tradeoff
Public_Umpire_1099 · reddit · 2026-09-24
- A Redditor implemented KVA projectors (inspired by DeepSeek V4.1 Flash and HySparse2/MiMo-V3) for Qwen3.8 Flash Next, running locally on 2x R9700 with 128GB DDR5.
- Tunable speed/quality tradeoff depending on start layer: 1.85x prefill speedup (1700→3150 t/s) at +8% perplexity from layer 12; 1.7x at +5% from layer 16; 1.45x at +2.6% from layer 24.
- Unlike other implementations: one full per-layer map using Tikhonov/ridge regression (1.5GB VRAM), predicting only later-layer inputs while real weights compute keys/values, with a selectable start layer instead of a fixed layer 24. The author found extra training nearly worthless — it's architectural levers that matter.
- A second multi-layer version achieves 1.55x speedup at only +2% perplexity, with a projected path to 1.5x under 1% ppl increase. V1 is stable and usable in R9V now; V2 is behind a --ced quality flag. Both the IQ4XS model and projector repos are on Hugging Face.
More from Infra
- Anthropic paid $5B for short-term compute in Q3 as spot prices run ~2x long-term deals — tengyanAI · 2026-11-11
- UK Sovereign AI invests in Fractile to build next-gen frontier AI chips — charlieharris01 · 2026-11-11
- What happened when I moved vector retrieval into Snowflake SQL instead of orchestrating RAG outside — dima806_dima · 2026-11-11
- AI chip startup Fractile raises $720M Series C for DRAM-on-processor design — Tom_Westgarth15 · 2026-11-11
- Mobile AI Factory: agentic pipeline builds on-device SLMs that beat 5-shot frontier models — kalyan_kpl · 2026-11-11
- Vivo's Starlink direct-to-cell service hits 10M users in two weeks across Brazil — XFreeze · 2026-11-11