Tools
Unsloth 0.1.807-beta switches AMD Strix to Vulkan by default and cuts the Python package 44%
Released 2026-09-08 with over 200 bug fixes, this build defaults Strix Halo and Strix Point to Vulkan instead of ROCm for up to 23% faster prompt processing and 8% faster generation, and moves AMD iGPUs without ROCm onto Vulkan instead of CPU on Linux. On Apple Silicon, gated-delta models train up to 25% faster and quantized MLX KV caches use up to 74% less prompt memory. PyTorch moves 2.10 to 2.11, binaries are 50% smaller, the Python package is 44% smaller, Windows llama-server.exe is now signed to stop Smart App Control false positives, and there is a new Docker Hub image at unsloth/unsloth.
Source
↳ Follow the thread