0.50.2
📦 bitsandbytesView on GitHub →
✨ 3 features🐛 6 fixes🔧 5 symbols
Summary
This release includes ROCm 10.0 builds for Linux and Windows, adds support for more AMD GPU targets, and improves performance for specific 4-bit GEMM operations on NVIDIA GB10 GPUs.
✨ New Features
- ROCm 10.0 build is now included for both Linux and Windows.
- Additional AMD GPU targets are included in prebuilt wheels.
- Performance improvement for certain 4bit GEMM shapes on NVIDIA GB10.
🐛 Bug Fixes
- Masked the absmax store in Triton 4-bit quantize kernels.
- Added sm121 (GB10) to the wide_n_simt 4-bit GEMM dispatch rule.
- Harden HIP version for native library naming and selection for ROCm.
- Restored Wave64 warp size for all gfx9 targets for ROCm.
- Expanded ROCm architecture coverage in CI.
- Added CI for ROCm 10.0.