SLIM-ARC v1.0.0
🎉 SLIM-ARC v1.0.0 Release
核心成果
- 64.5× 累计 decode 加速 (0.08 → 5.16 t/s)
- 80B MoE 在 8GB RAM 运行 (+850% over baseline)
- FlashAttention +71.4% (零代码改动)
- GSM8K 精度验证 (75% Qwen3-4B, 50% 80B Q4_K_M)
优化技术链
Baseline → +MADV_RANDOM → +KV q4_0 → +IQ4_XS → +FlashAttention
文档
快速开始
git clone https://github.com/Nexa-Language/SLIM-ARC.git
cd SLIM-ARC
git clone https://github.com/ggml-org/llama.cpp src/llama-upstream
python3 scripts/apply-slim-arc.py
cd src/llama-upstream && cmake -B build -DGGML_CPU_REPACK=OFF && cmake --build build -j中山大学 · 依托Agent答辩OS团队 · 2026 OS设计赛 Proj 59