llama-cpp-python 0.3

Apply the Python 3 prompt first. These rules cover the llama-cpp-python 0.3 line and its python -m llama_cpp.server application, and were checked against 0.3.35. The package declares Python 3.8 or later; deploy on a maintained Python release and verify compatibility with the exact wheel, operating system, architecture, and accelerator backend.

Native artifacts and model acquisition

• Pin the wheel or source build, bundled llama.cpp revision, and accelerator libraries together; review native GGUF, tokenizer, and RPC advisories as well as Python dependencies before deployment.
• Provision models with hf_hub_download or snapshot_download using a full immutable commit in revision, then load verified local paths. In this version, Llama.from_pretrained and the server's hf_model_repo_id path do not forward a download revision; passing revision through constructor kwargs does not pin the download.
• Verify every GGUF shard, LoRA adapter, projector, and tokenizer file against approved digests from trusted publishers; review embedded chat templates. Loading the first GGUF shard also loads the remaining shards; checking only the first file is insufficient.
• Keep CONFIG_FILE, model, lora_path, clip_model_path, and hf_pretrained_model_name_or_path under operator control. Use read-only local artifact directories and provision Hugging Face tokenizers in advance so inference workers need no repository credentials or download access.
• Protect LLAMA_CPP_LIB_PATH, MTMD_CPP_LIB, CMAKE_ARGS, and accelerator library search directories from caller influence.