b8821

The ggml-org team behind the widely-used llama.cpp project has released a significant server-side update with commit b8821. This commit primarily refines how the server handles media markers, which are identifiers for media content within prompts. The most notable change is the introduction of the LLAMA_MEDIA_MARKER environment variable, allowing developers to pin a specific marker at server startup. This replaces the previous method where tests had to dynamically fetch a random marker via an `/apply-template` endpoint, simplifying test fixtures and improving consistency. The hardcoded prompts in tests can now rely on a predetermined marker, making test behavior more reliable and reproducible.

A second major improvement is the implementation of thread-safe initialization for the `get_media_marker()` function. The developers addressed code review feedback by replacing a potentially unsafe global static variable with a C++11 static local variable initialized via a lambda. This 'magic static' pattern guarantees that the initialization happens exactly once, even in multi-threaded environments, without the need for explicit locking mechanisms. This change enhances the robustness of the llama.cpp server when handling concurrent requests, a critical feature for production deployments. The commit also includes various 'nits'—minor code cleanups and fixes—and the release was packaged with pre-built binaries for a wide range of platforms including macOS (Apple Silicon/Intel), Linux (CPU/Vulkan/ROCm), Windows (CPU/CUDA/Vulkan), and openEuler.