本機推論與微調
Unsloth 0.1.807-beta Switches AMD APUs to Vulkan by Default, Boosting Strix Halo Prefill Performance by Up to 23%
Unsloth Desktop now uses Vulkan for Strix Halo, Strix Point, and AMD iGPUs without ROCm support, while also reducing the MLX KV cache footprint and installation package size. The performance figures come from official tests, but users have reported that the interface still displays ROCm after upgrading, so automatic backend selection still needs verification on real hardware.
Unsloth released version 0.1.807-beta on September 8, overhauling backend selection for local AMD inference, memory usage on Apple Silicon, and container delivery. The most technically significant change is that Strix Halo and Strix Point now use Vulkan by default. On Linux, AMD iGPUs without a working ROCm installation no longer fall back directly to the CPU. Official benchmarks claim up to a 23% improvement in prompt processing on Strix Halo and up to an 8% improvement in generation. However, Unsloth did not disclose the model, quantization format, context length, or power settings used, so the results cannot be generalized to every RDNA GPU.
For the Apple MLX path, Unsloth claims that quantized KV cache reduces prompt memory usage by up to 74%, while training gated-delta models is up to 25% faster. The release also adds DoRA, more DPO loss functions, and streaming batched generation with independent sampling. The practical value of these changes goes beyond peak speed: a smaller KV cache lets Macs with unified memory retain longer contexts or support larger batch sizes, while independent sampling addresses the risk of chats in the same batch sharing generation behavior.
The release also upgrades the default PyTorch version from 2.10 to 2.11, signs the Windows `llama-server.exe` binary, reduces the Python wheel size by 44%, and publishes amd64 and arm64 Docker images containing PyTorch 2.11, CUDA 12.8, JupyterLab, llama.cpp, and whisper.cpp. The container tags `latest` and `studio` are mutable; reproducible deployments should pin a dated or versioned tag as well as the image digest.
Engineering teams should still validate the release on each machine. On release day, a Strix Halo user reported that the settings screen continued to show automatic/ROCm selection, which could reflect an interface-labeling issue, a setting retained from the previous installation, or backend routing that failed to take effect. Before deployment, teams should record the backend that is actually loaded, time to first token, generation speed, and output correctness. This remains a beta release, and the official claim of “more than 200 fixes” also broadens the scope required for regression testing.