v0.2.8
Release Notes (v0.2.8)
性能与显存优化
- 将 APOLLO+CAME 低秩更新路径融合到 CUDA 核心,大幅减少 Kernel 启动开销并优化临时张量生命周期,降低显存峰值。
Performance & Memory Optimization
- Fused the APOLLO+CAME low-rank update path into CUDA kernels, reducing launch overhead and optimizing temporary tensor lifecycles to lower peak memory usage.