Ollama v0.40.15-rc1 updates llama.cpp dependency to version b12111

Ollama has released version 0.40.15-rc1, a release candidate that focuses on updating its core inference engine dependency. The primary change involves a version bump of the integrated llama.cpp library to build b12111.

This update also includes the removal of a specific local patch, 002-metal-unary-cast-constexpr.patch, which was previously used to address Metal-related compilation issues. This patch is no longer required as the upstream llama.cpp repository has implemented a native fix for the issue.

Technical changes in Ollama v0.40.15-rc1

• Dependency Update: The release bumps the internal llama.cpp version to build b12111, ensuring the runtime utilizes the latest upstream inference optimizations.
• Metal Backend Cleanup: The project has dropped the 002-metal-unary-cast-constexpr.patch, reducing the maintenance overhead of local patches.
• Upstream Alignment: The removal of the Metal patch indicates that the upstream llama.cpp project has successfully integrated a fix for unary cast constexpr issues, allowing Ollama to stay closer to the main repository code.

Why it matters

By aligning more closely with upstream llama.cpp, Ollama reduces the technical debt associated with maintaining custom patches for Apple Metal support. This ensures that users running local LLMs on macOS benefit from official upstream fixes and performance improvements more reliably.