b8517

The open-source project llama.cpp, maintained by ggml-org, has released a significant update with commit b8517. This commit introduces a new, lower-level API function called `llama_model_load_from_fd`. This function allows developers to load AI models directly using a file descriptor (an integer handle to an open file) instead of relying on a traditional file path string. The change represents a shift towards more efficient and secure model loading mechanisms, particularly beneficial for applications running in sandboxed or containerized environments where direct file system access can be restricted or slow.

The technical implementation involved refactoring the internal loading logic to consistently use `FILE` pointers across the codebase, addressing feedback from code reviews. The update also includes a fix for the `llama-model-saver` tool and improvements to roundtrip tests to ensure model integrity. This foundational change paves the way for more advanced features, such as loading models directly from memory or over networks, by abstracting the source of the model data. The commit is part of the continuous optimization of llama.cpp, which is crucial for deploying models from Meta's Llama family and others on everything from servers to edge devices.