The llama.cpp project released version b9923, featuring a significant refactor of the server-stream component to address issues with the Server-Sent Events (SSE) replay buffer. This update introduces a pimpl pattern and scopes public stream functions under the `server_stream_` prefix to improve code organization.

  • Session and manager state are now guarded by a mutex, with condvar predicates set under the lock while keeping cancelled atomic for lock-free polling.
  • Comments have been trimmed to focus on concurrency, lifetime, and ordering rationale, removing those that merely restate the code.
  • Stream traces were moved to the debug level to reduce default log noise during bring-up diagnostics.
  • The `stream_read_status` enum was moved into the .cpp file to remain behind the pimpl boundary.

The release includes binaries for macOS (Apple Silicon and Intel), iOS, Linux (CPU, Vulkan, ROCm, OpenVINO, SYCL), Android, Windows (CPU, CUDA 12/13, Vulkan, OpenCL, OpenVINO, SYCL, HIP), and openEuler.