Pinned Loading

1. turboquant-plus turboquant-plus Public

   TurboQuant+: 3-bit KV cache value quantization and group size optimization for long-context LLM inference

   Python 1

2. ChannelQuant ChannelQuant Public

   Near-lossless 4× KV-cache compression for GQA models at ~4.1 bits/value — per-channel-key INT4 + static outlier ROM, with a reproducible reference model and paper.

   Python

3. LonghornSilicon/kv-cache-engine LonghornSilicon/kv-cache-engine Public

   Hardware KV cache compression engine (SystemVerilog) using TurboQuant+ — keys at 4.25 bpv, values at ~3.0 bpv for 3–5× DRAM bandwidth reduction on LLM inference. Block 2 of the LonghornSilicon acce…

   SystemVerilog 1

4. LonghornSilicon/adaptive-precision-attention LonghornSilicon/adaptive-precision-attention Public

   Entropy-guided mixed-precision attention: evolutionary search discovers that entropy is the optimal discriminator for per-block quantization decisions

   Python

5. evoplace evoplace Public

   LLM-guided evolutionary VLSI placement: beating DreamPlace with evolved objective functions

   Python 1

6. covenant covenant Public

   Contract-driven GPU analytical placer (DREAMPlace fork overlay) with a noise-calibrated evaluation protocol, paired multi-seed CIs, liveness gates, calibration arms.

   Python