Repository navigation
- Preliminary ROCm support (ROCm 10 only, gfx1100+)
- Faster offloading for AVX2 CPUs
- Smarter CPU core allocation/pinning by default
- Reduced system memory overhead
- Fix vision towers for GLM4, Qwen3-VL
- Other bugfixes and optimizations
Full Changelog: v1.5.4...v1.6.0