TurboQuant+ KV Cache Compression

TurboQuant+ — KV Cache Compression for Local LLM Inference

> Accelerate local LLM inference on Apple Silicon with 3.8-6.4x KV cache compression via PolarQuant + Walsh-Hadamard rotation.

Trigger Keywords

量化, KV压缩, 本地推理, llama.cpp, turboquant, KV cache, compression, Apple Silicon, Metal, turbo2, turbo3, turbo4

Overview

TurboQuant+ implements TurboQuant (ICLR 2026) for llama.cpp with Metal GPU kernels. It compresses the transformer KV cache to squeeze larger models and longer contexts into limited Apple Silicon memory — with minimal quality loss.

Core Capabilities

turbo2 (2-bit, 6.4x compression) — Extreme memory savings, +6.48% PPL. Best for asymmetric V-only compression.

turbo3 (3-bit, 4.6-5.1x compression) — Maximum memory savings with acceptable quality. +1.06% PPL vs q8_0.

turbo4 (4-bit, 3.8x compression) — Best quality/compression tradeoff. +0.23% PPL vs q8_0, closer to q8_0 than q4_0.

Asymmetric K/V — Keep K at q8_0 for attention quality, compress V aggressively. Rescues quality on low-bit weight models.

Boundary V — Layer-aware V compression (first 2 + last 2 layers at q8_0, rest turbo2). Recovers 37-91% of quality gap.

Sparse V dequant — Skip low-weight V positions during decode. +22.8% decode speed at 32K context, no PPL impact.

Dependencies

None. Works with the llama.cpp TurboQuant fork.

Configuration Guide

Basic Usage (llama-server)

# Recommended default — turbo4 symmetric
llama-server -m model.gguf --cache-type-k turbo4 --cache-type-v turbo4 -fa 1
Maximum compression — turbo3 symmetric
llama-server -m model.gguf --cache-type-k turbo3 --cache-type-v turbo3 -fa 1
Extreme compression — turbo2 (best with asymmetric)
llama-server -m model.gguf --cache-type-k q8_0 --cache-type-v turbo2 -fa 1

Asymmetric K/V (for Q4_K_M models)

Some low-bit weight models degrade with symmetric turbo. Use asymmetric K/V:

# K stays at q8_0, V compressed with turbo
llama-server -m model-Q4_K_M.gguf --cache-type-k q8_0 --cache-type-v turbo4 -fa 1
Even more V compression
llama-server -m model-Q4_K_M.gguf --cache-type-k q8_0 --cache-type-v turbo3 -fa 1

> Note: Larger models (70B, 104B) handle symmetric turbo fine. Asymmetric mainly benefits smaller Q4_K_M models.

Long Context on Large Models

For 70B+ models at 32K+ context on 128GB Macs, raise the GPU memory cap:

# Set to 90% of 128GB
sudo sysctl iogpu.wired_limit_mb=117964
Then run with turbo3 for maximum context
llama-server -m Llama-70B-Q4_K_M.gguf --cache-type-k turbo3 --cache-type-v turbo3 -c 65536 -fa 1

Recommended Configs by Scenario

| Scenario | K cache | V cache | Compression | PPL impact | |----------|---------|---------|-------------|------------| | Best quality | turbo4 | turbo4 | 3.8x | +0.23% | | Balanced | turbo3 | turbo3 | 4.6-5.1x | +1.06% | | Max compression | turbo2 | turbo2 | 6.4x | +6.48% | | Q4_K_M safe | q8_0 | turbo4 | ~3.8x V | +1.0% | | Boundary V | q8_0 | turbo2 | ~6x V | 37-91% quality recovered |

Apple Silicon Benchmarks (M5 Max 128GB)

Quality (wikitext-2)

| Cache | Compression | PPL | vs q8_0 | |-------|-------------|-----|---------| | q8_0 | 1.9x | 6.111 | baseline | | turbo4 | 3.8x | 6.125 | +0.23% | | turbo3 | 4.6x | 6.176 | +1.06% | | turbo2 | 6.4x | 6.507 | +6.48% |

Large Model Results

| Model | Config | PPL | Context | NIAH | |-------|--------|-----|---------|------| | Llama-70B Q4_K_M | turbo4/turbo4 | 3.461 | 48K | 30/30 | | Command-R+ 104B Q4_K_M | turbo3/turbo3 | 6.415 | 128K | 10/10 |

Speed

Prefill: turbo3 matches or exceeds q8_0 speed (1.0-1.1x)

Decode: turbo4 at ~0.93x q8_0, turbo3 at ~0.78-0.90x q8_0

Sparse V: +22.8% decode at 32K context, no quality loss

M1 Max 64GB Results (Community)

| KV | Prefill t/s | Decode t/s | vs q8_0 | |----|------------|-----------|---------| | q8_0 | 399.0 | 12.4 | — | | turbo4 | 365.0 | 16.6 | +33.9% |

Key Research Findings

1. V compression is free — Compressing V to 2-bit has zero measurable effect when K precision is maintained. Validated on Metal, CUDA RTX 4090, RTX 3090. 2. All quality loss comes from K compression — This is why asymmetric configs rescue quality. 3. Boundary layers are sensitive — Protecting first 2 + last 2 layers recovers 37-91% of quality gap. 4. turbo4 beats q4_0 in quality — Lower KL divergence, higher top-p agreement, at similar compression.

References

Getting Started Guide

Configuration Recommendations

TurboQuant Paper (Google Research)

Asymmetric K/V Compression

M5 Max Stress Test