Unsloth released version 0.1.808-beta, a performance and reliability update covering diffusion, AMD, and MLX workloads1. Diffusion inference is 1.2x to 1.7x faster across INT8 and FP8 pathways, AMD Strix devices gain a reported 20% speed boost over ROCm by defaulting to Vulkan, and gated-delta models train up to 25% faster on Apple Silicon. The release includes over 250 bug fixes, 60% smaller binaries, quantized MLX KV caches using up to 74% less prompt memory, and an updated PyTorch 2.11 base.
Unsloth v0.1.808 Delivers 1.7x Faster Diffusion, 250-Plus Bug Fixes
Unsloth's v0.1.808-beta release accelerates diffusion inference up to 1.7x, boosts AMD performance 20% via Vulkan, and ships over 250 bug fixes with 60%…