Breaking change in llama-server?

A recent update to the popular llama.cpp inference server has sparked controversy by automatically and irreversibly migrating user model files. The change, introduced in commit b8498, forces a one-time migration of the local cache from the traditional `~/.cache/llama.cpp/` directory to HuggingFace's standard cache location (`~/.cache/huggingface/hub/`). This action, which occurs without user consent upon launching the latest `llama-server` build, moves all models previously downloaded using the `-hf` flag and converts them into HuggingFace's blob format. The result is immediate failure for any existing automation scripts that reference the old file paths, disrupting model management and deployment workflows.

The core issue is the breaking nature of the change. Users report their launch scripts failing with errors like "failed to load model" because the .gguf files are no longer at their expected locations. This is particularly problematic for developers who manage and distribute .gguf models across multiple machines, as their entire toolchain is now broken. The community backlash centers on the lack of a warning or an opt-out mechanism before making such a permanent alteration to a user's local file system. Many see this as a heavy-handed consequence of HuggingFace's increasing integration with the project, raising concerns about future stability and user control in the open-source AI tooling ecosystem.