MLX Header Vendoring

This release introduces a vendoring approach for MLX-C headers, allowing Go to build the project without requiring CMake as a build dependency. Previously, CMake needed to run first to generate the necessary headers. Now the headers are included directly in the repository.

Build Simplification

By removing the Go build tag dependency, MLX-based functionality is enabled by default across platforms. The vendored headers are automatically refreshed whenever CMake runs, making it easy to keep them synchronized when MLX versions are updated.

Platform Support

The changes have been verified to work on both Windows and Linux. Additional improvements in this release include preference for ROCm v6 on Windows to avoid compatibility issues with v7, and CI hardening to handle flaky Chocolatey repository servers.

What Developers Need to Know

• No more CMake prerequisite: Building Ollama no longer requires CMake to be installed first
• MLX support enabled by default: MLX-based code is now part of the standard build
• Header synchronization: Headers stay in sync with MLX versions automatically during CMake runs