Skip to content

Latest commit

 

History

History
23 lines (19 loc) · 1.13 KB

File metadata and controls

23 lines (19 loc) · 1.13 KB

KV-Cache Comparison Sources

The README chart is a simple memory-scaling model. It compares linear KV-cache compression systems against Catalyst Brain's fixed public SDK state.

Sources used:

  • TurboQuant: Google Research / arXiv, "Online Vector Quantization for Efficient High-Throughput LLM Inference" (arXiv:2504.19874, https://arxiv.org/abs/2504.19874). The chart uses the reported 3.5 bits/channel quality-neutral point as the modeled memory ratio versus FP16.
  • KIVI: "A Tuning-Free Asymmetric 2bit Quantization for KV Cache" (arXiv:2402.02750, https://arxiv.org/abs/2402.02750). The chart models raw 2-bit KV storage versus FP16.
  • PyramidKV: "Dynamic KV Cache Compression based on Pyramidal Information Funneling" (arXiv:2406.02069, https://arxiv.org/abs/2406.02069). The chart uses the paper's 12% KV retention setting.
  • Catalyst Brain HKVC: fixed 4096-dim state measured through the public catalyst-brain SDK APIs in scripts/generate_scaling_chart.py.

This repository deliberately exposes only adapter code, public benchmark models, and public SDK calls. The closed-source Catalyst Brain SDK remains the engine.