Deprecation of --mmap, --mlock, and --direct-io in llama.cpp v0.4.1

The release of llama.cpp version 0.4.1 introduces several enhancements, including support for new models like Maple 20B-A1B and Tencent Hy 4, improved JSON schema handling, and updates to ggml v0.24.0. Among the notable changes, the deprecated command-line arguments --mmap, --mlock, and --direct-io have been officially removed and replaced with the unified --load-mode argument, according to the project's GitHub release notes.

Operators planning to upgrade should carefully review their existing configurations and scripts to ensure compatibility with the new --load-mode argument. Any workflows relying on the removed arguments will need to be updated to avoid runtime errors or unexpected behavior. This change reflects a broader trend in software projects to consolidate and simplify configuration options, but it also underscores the importance of validating operational dependencies before upgrading.