Class Llama

llama_cpp/llama.py:55–2447

High-level Python wrapper for a llama.cpp model.

class Llama:
    """High-level Python wrapper for a llama.cpp model."""

    __backend_initialized = False

    def __init__(
        self,
        model_path: str,
        *,
        # Model Params
        n_gpu_layers: int = 0,
        split_mode: int = llama_cpp.LLAMA_SPLIT_MODE_LAYER,
        main_gpu: int = 0,
        tensor_split: Optional[List[float]] = None,
        vocab_only: bool = False,
        use_mmap: bool = True,
        use_mlock: bool = False,
        kv_overrides: Optional[Dict[str, Union[bool, int, float, str]]] = None,
        # Context Params
        seed: int = llama_cpp.LLAMA_DEFAULT_SEED,
        n_ctx: int = 512,
        n_batch: int = 512,
        n_ubatch: int = 512,
        n_threads: Optional[int] = None,
        n_threads_batch: Optional[int] = None,
        rope_scaling_type: Optional[
            int
        ] = llama_cpp.LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED,
        pooling_type: int = llama_cpp.LLAMA_POOLING_TYPE_UNSPECIFIED,
        attention_type: int = llama_cpp.LLAMA_ATTENTION_TYPE_UNSPECIFIED,
        rope_freq_base: float = 0.0,
        rope_freq_scale: float = 0.0,
        yarn_ext_factor: float = -1.0,
        yarn_attn_factor: float = 1.0,
        yarn_beta_fast: float = 32.0,
        yarn_beta_slow: float = 1.0,
        yarn_orig_ctx: int = 0,
        logits_all: bool = False,
        embedding: bool = False,
        offload_kqv: bool = True,
        flash_attn: bool = False,
        op_offload: Optional[bool] = None,
        swa_full: Optional[bool] = None,
        # Sampling Params
        no_perf: bool = False,
        last_n_tokens_size: int = 64,
        # LoRA Params
        lora_base: Optional[str] = None,
        lora_scale: float = 1.0,
        lora_path: Optional[str] = None,
        # Backend Params
        numa: Union[bool, int] = False,
        # Chat Format Params
        chat_format: Optional[str] = None,
        chat_handler: Optional[llama_chat_format.LlamaChatCompletionHandler] = None,
        # Speculative Decoding
        draft_model: Optional[LlamaDraftModel] = None,
        # Tokenizer Override