b8748

The open-source project llama.cpp, maintained by ggml-org, has patched a critical bug in its inference server that was blocking users from running models in routing mode. The issue, documented in GitHub commit b8748, surfaced when users attempted to launch the server with a custom model preset file (using `--models-preset`) while also specifying a model alias (via `--alias`). This combination triggered a conflict error where the server incorrectly interpreted the custom alias as clashing with an existing model name, causing initialization to fail entirely. The fix, contributed by Adrien Gallouët, is straightforward: the server now simply ignores the `--alias` argument whenever `--models-preset` is in use, eliminating the conflict and allowing the routing system to start correctly.

This seemingly minor fix has significant implications for developers and companies deploying local AI models at scale. The routing mode in llama.cpp is essential for creating multi-model endpoints, where a single server can load and manage several different AI models (like various quantized versions of Llama, Gemma, or others) and direct requests appropriately. The bug was preventing users from leveraging preset configuration files—a key feature for reproducible and manageable deployments—if they also used aliases for convenience. The commit ensures that complex, production-style setups using presets for models like 'Gemma 4 E4B UD Q8_K_XL' can now be launched reliably across all supported platforms, including macOS, Linux, Windows, and iOS.