b8571

The open-source project llama.cpp, maintained by ggml-org, has released a critical fix in commit b8571 (SHA e397d38). This patch resolves a severe crash bug (SIGSEGV) in the library's JSON schema pattern converter. The converter, which transforms regex patterns into formal grammars for AI parsing, would fail when encountering non-capturing groups—a common regex syntax like (?:...). The root cause was a parser that, upon seeing '(' followed by '?', would push a warning but fail to advance past the '?:' characters. This led to a subsequent recursive function call incorrectly interpreting a '?' as a quantifier and calling `.back()` on an empty vector, triggering undefined behavior and a crash.

This bug is particularly relevant for developers serving OpenAI-compatible tool/function calls, as clients often send complex regex patterns for input validation within their JSON schemas. A prime example cited is intricate date validation patterns (e.g., `^(?:(?:\d\d[2468][048]|...)-02-29|...)$`). The fix involves safely skipping the '?:' syntax to treat non-capturing groups as regular groups, and for other unsupported syntax like lookaheads (?=, ?!), it safely skips to the matching parenthesis. The update also improves parenthesis depth counting to handle escaped characters and adjusts error checks. The fix is currently implemented in the C++ core, with test cases added, while the JavaScript and Python bindings do not yet support this syntax. The release includes pre-built binaries for macOS (Apple Silicon/Intel), Linux (CPU, Vulkan, ROCm), Windows (CPU, CUDA, Vulkan), and openEuler.