>>> print(model("The quick brown fox jumps ", stop=["."])["choices"][0]["text"])
the lazy dog
Loading a chat model
>>> import llama_cpp
>>> model = llama_cpp.Llama(
... model_path="path/to/model",
... chat_format="llama-2",
... )
>>> print(model.create_chat_completion(
... messages=[{
... "role": "user",
... "content": "what is the meaning of life?"
... }]
... ))
Args:
model_path: Path to the model.
n_gpu_layers: Number of layers to offload to GPU (-ngl). If -1, all layers are offloaded.
split_mode: How to split the model across GPUs. See llama_cpp.LLAMA_SPLIT_* for options.
main_gpu: main_gpu interpretation depends on split_mode: LLAMA_SPLIT_MODE_NONE: the GPU that is used for the entire model. LLAMA_SPLIT_MODE_ROW: the GPU that is used for small tensors and intermediate results. LLAMA_SPLIT_MODE_LAYER: ignored
tensor_split: How split tensors should be distributed across GPUs. If None, the model is not split.
rpc_servers: Comma separated list of RPC servers to use for offloading
vocab_only: Only load the vocabulary no weights.
use_mmap: Use mmap if possible.
use_mlock: Force the system to keep the model in RAM.
kv_overrides: Key-value overrides for the model.
seed: RNG seed, -1 for random
n_ctx: Text context, 0 = from model
n_batch: Prompt processing maximum batch size
n_ubatch: Physical batch size
n_threads: Number of threads to use for generation
n_threads_batch: Number of threads to use for batch processing
rope_scaling_type: RoPE scaling type, from `enum llama_rope_scaling_type`. ref: https://github.com/ggml-org/llama.cpp/pull/2054
pooling_type: Pooling type, from `enum llama_pooling_type`.
rope_freq_base: RoPE base frequency, 0 = from model
rope_freq_scale: RoPE frequency scaling factor, 0 = from model
yarn_ext_factor: YaRN extrapolation mix factor, negative = from model
yarn_attn_factor: YaRN magnitude scaling factor
yarn_beta_fast: YaRN low correction dim
yarn_beta_slow: YaRN high correction dim
yarn_orig_ctx: YaRN original context size
logits_all: Return logits for all tokens, not just the last token. Must be True for completion to return logprobs.
embedding: Embedding mode only.
offload_kqv: Offload K, Q, V to GPU.
flash_attn: Use flash attention.
no_perf: Measure performance timings.
last_n_tokens_size: Maximum number of tokens to keep in the last_n_tokens deque.
lora_base: Optional path to base model, useful if using a quantized base model and you want to apply LoRA to an f16 model.
lora_path: Path to a LoRA file to apply to the model.
numa: numa policy
chat_format: String specifying the chat format to use when calling create_chat_completion.
chat_handler: Optional chat handler to use when calling create_chat_completion.
draft_model: Optional draft model to use for speculative decoding.
tokenizer: Optional tokenizer to override the default tokenizer from llama.cpp.
verbose: Print verbose output to stderr.
type_k: KV cache data type for K (default: f16)
type_v: KV cache data type for V (default: f16)
spm_infill: Use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this.