New
srkrz23/vramkit — two-track VRAM compression for weights and the KV cache
srkrz23/vramkit is a Python repository implementing a lossless weight codec and KV cache compression that applies different quantization axes to keys and values.
Projects
srkrz23/vramkit
Lossless-verified VRAM compression for LLM inference: a weight codec (-27% mean / -33% best vs 4-bit on real Qwen weights, beats zstd, lossless) + drop-in PyTorch KV-cache compression (int8 ~2x near-lossless, outlier-aware int4 at cos-sim 0.999). ROCm/CUDA/MPS/CPU.
View on GitHub ↗- stars
- ★ 1
- Forks
- 0
- Language
- Python
- License
- —
- Latest release
- —
Posts
1No matching stories.