srkrz23/vramkit

VRAM compression for LLM inference, measured on a MacBook: lossless weight codec (-33% vs 4-bit on real Qwen weights, beats zstd) + drop-in PyTorch KV-cache compression. Key finding: quantize KV keys per-CHANNEL (not per-token) -> int8 KV near-lossless, +0.1% real perplexity at 2x. ROCm/CUDA/MPS/CPU.

星标

★ 2

复刻

0

语言

Python

许可证

—

最新版本

—