llama.cpp Release b11078 Adds Environment Variables for Server Sampling Options
llama.cpp release b11078 introduces environment variable support (via LLAMA_ARG_*) for configuring sampling parameters such as temperature, top-p, min-p, and penalties in llama-server. This allows server operators to fully manage inference settings using environment configuration files like systemd EnvironmentFile on Linux. This update simplifies infrastructure automation and modern container deployments (such as Docker, Kubernetes, and systemd services) for self-hosted LLM endpoints. Operators no longer need to pass lengthy, complex command-line arguments to customize model generation behavior. The newly supported flag mappings include --temp, --top-p, --min-p, --repeat-penalty, --presence-penalty, and --frequency-penalty. Updated binaries are provided across Linux, macOS, Windows, and Android, though the macOS Apple Silicon KleidiAI build remains disabled.
## BACKGROUND
llama.cpp is a popular open-source C/C++ inference framework that enables running Large Language Models efficiently on consumer hardware. Sampling parameters like temperature, top-p, and min-p control output randomness and text generation quality, while llama-server provides an HTTP interface compatible with OpenAI APIs.