langchain_community.llms.llamacpp.LlamaCpp

class langchain_community.llms.llamacpp.LlamaCpp

Bases: LLM

llama.cpp model.

To use, you should have the llama-cpp-python library installed, and provide the path to the Llama model as a named parameter to the constructor. Check out: https://github.com/abetlen/llama-cpp-python

Example

from langchain_community.llms import LlamaCpp
llm = LlamaCpp(model_path="/path/to/llama/model")

Create a new model by parsing and validating input data from keyword arguments.

Raises ValidationError if the input data cannot be parsed to form a valid model.

param cache: Optional[bool] = None

param callback_manager: Optional[BaseCallbackManager] = None

param callbacks: Callbacks = None

param echo: Optional[bool] = False

Whether to echo the prompt.

param f16_kv: bool = True

Use half-precision for key/value cache.

param grammar: Optional[Union[str, LlamaGrammar]] = None

grammar: formal grammar for constraining model outputs. For instance, the grammar can be used to force the model to generate valid JSON or to speak exclusively in emojis. At most one of grammar_path and grammar should be passed in.

param grammar_path: Optional[Union[str, Path]] = None

grammar_path: Path to the .gbnf file that defines formal grammars for constraining model outputs. For instance, the grammar can be used to force the model to generate valid JSON or to speak exclusively in emojis. At most one of grammar_path and grammar should be passed in.

param last_n_tokens_size: Optional[int] = 64

The number of tokens to look back when applying the repeat_penalty.

param logits_all: bool = False

Return logits for all tokens, not just the last token.

param logprobs: Optional[int] = None

The number of logprobs to return. If None, no logprobs are returned.

param lora_base: Optional[str] = None

The path to the Llama LoRA base model.

param lora_path: Optional[str] = None

The path to the Llama LoRA. If None, no LoRa is loaded.

param max_tokens: Optional[int] = 256

The maximum number of tokens to generate.

param metadata: Optional[Dict[str, Any]] = None

Metadata to add to the run trace.

param model_kwargs: Dict[str, Any] [Optional]

Any additional parameters to pass to llama_cpp.Llama.

param model_path: str [Required]

The path to the Llama model file.

param n_batch: Optional[int] = 8

Number of tokens to process in parallel. Should be a number between 1 and n_ctx.

param n_ctx: int = 512

Token context window.

param n_gpu_layers: Optional[int] = None

Number of layers to be loaded into gpu memory. Default None.

param n_parts: int = -1

Number of parts to split the model into. If -1, the number of parts is automatically determined.

param n_threads: Optional[int] = None

Number of threads to use. If None, the number of threads is automatically determined.

param repeat_penalty: Optional[float] = 1.1

The penalty to apply to repeated tokens.

param rope_freq_base: float = 10000.0

Base frequency for rope sampling.

param rope_freq_scale: float = 1.0

Scale factor for rope sampling.

param seed: int = -1

Seed. If -1, a random seed is used.

param stop: Optional[List[str]] = []

A list of strings to stop generation when encountered.

param streaming: bool = True

Whether to stream the results, token by token.

param suffix: Optional[str] = None

A suffix to append to the generated text. If None, no suffix is appended.

param tags: Optional[List[str]] = None

Tags to add to the run trace.

param temperature: Optional[float] = 0.8

The temperature to use for sampling.

param top_k: Optional[int] = 40

The top-k value to use for sampling.

param top_p: Optional[float] = 0.95

The top-p value to use for sampling.

param use_mlock: bool = False

Force system to keep model in RAM.

param use_mmap: Optional[bool] = True

Whether to keep the model loaded in RAM

param verbose: bool = True

Print verbose output to stderr.

param vocab_only: bool = False

Only load the vocabulary, no weights.

__call__(prompt: str, stop: Optional[List[str]] = None, callbacks: Optional[Union[List[BaseCallbackHandler], BaseCallbackManager]] = None, *, tags: Optional[List[str]] = None, metadata: Optional[Dict[str, Any]] = None, **kwargs: Any) → str

Check Cache and run the LLM on the given prompt and input.

async abatch(inputs: List[Union[PromptValue, str, List[BaseMessage]]], config: Optional[Union[RunnableConfig, List[RunnableConfig]]] = None, *, return_exceptions: bool = False, **kwargs: Any) → List[str]

Default implementation runs ainvoke in parallel using asyncio.gather.

The default implementation of batch works well for IO bound runnables.

Subclasses should override this method if they can batch more efficiently; e.g., if the underlying runnable uses an API which supports a batch mode.

async agenerate(prompts: List[str], stop: Optional[List[str]] = None, callbacks: Union[List[BaseCallbackHandler], BaseCallbackManager, None, List[Optional[Union[List[BaseCallbackHandler], BaseCallbackManager]]]] = None, *, tags: Optional[Union[List[str], List[List[str]]]] = None, metadata: Optional[Union[Dict[str, Any], List[Dict[str, Any]]]] = None, run_name: Optional[Union[str, List[str]]] = None, **kwargs: Any) → LLMResult

Run the LLM on the given prompt and input.

async agenerate_prompt(prompts: List[PromptValue], stop: Optional[List[str]] = None, callbacks: Union[List[BaseCallbackHandler], BaseCallbackManager, None, List[Optional[Union[List[BaseCallbackHandler], BaseCallbackManager]]]] = None, **kwargs: Any) → LLMResult

Asynchronously pass a sequence of prompts and return model generations.

This method should make use of batched calls for models that expose a batched API.

Use this method when you want to:

1. take advantage of batched calls,

2. need more output from the model than just the top generated value,

3. are building chains that are agnostic to the underlying language model

   type (e.g., pure text completion models vs chat models).

Parameters

• prompts – List of PromptValues. A PromptValue is an object that can be converted to match the format of any language model (string for pure text generation models and BaseMessages for chat models).

• stop – Stop words to use when generating. Model output is cut off at the first occurrence of any of these substrings.

• callbacks – Callbacks to pass through. Used for executing additional functionality, such as logging or streaming, throughout generation.

• **kwargs – Arbitrary additional keyword arguments. These are usually passed to the model provider API call.

Returns

An LLMResult, which contains a list of candidate Generations for each input

prompt and additional model provider-specific output.

async ainvoke(input: Union[PromptValue, str, List[BaseMessage]], config: Optional[RunnableConfig] = None, *, stop: Optional[List[str]] = None, **kwargs: Any) → str

Default implementation of ainvoke, calls invoke from a thread.

The default implementation allows usage of async code even if the runnable did not implement a native async version of invoke.

Subclasses should override this method if they can run asynchronously.

async apredict(text: str, *, stop: Optional[Sequence[str]] = None, **kwargs: Any) → str

Asynchronously pass a string to the model and return a string prediction.

Use this method when calling pure text generation models and only the top

candidate generation is needed.

Parameters

• text – String input to pass to the model.

• stop – Stop words to use when generating. Model output is cut off at the first occurrence of any of these substrings.

• **kwargs – Arbitrary additional keyword arguments. These are usually passed to the model provider API call.

Returns

Top model prediction as a string.

async apredict_messages(messages: List[BaseMessage], *, stop: Optional[Sequence[str]] = None, **kwargs: Any) → BaseMessage

Asynchronously pass messages to the model and return a message prediction.

Use this method when calling chat models and only the top

candidate generation is needed.

Parameters

• messages – A sequence of chat messages corresponding to a single model input.

• stop – Stop words to use when generating. Model output is cut off at the first occurrence of any of these substrings.

• **kwargs – Arbitrary additional keyword arguments. These are usually passed to the model provider API call.

Returns

Top model prediction as a message.

async astream(input: Union[PromptValue, str, List[BaseMessage]], config: Optional[RunnableConfig] = None, *, stop: Optional[List[str]] = None, **kwargs: Any) → AsyncIterator[str]

Default implementation of astream, which calls ainvoke. Subclasses should override this method if they support streaming output.

async astream_log(input: Any, config: Optional[RunnableConfig] = None, *, diff: bool = True, with_streamed_output_list: bool = True, include_names: Optional[Sequence[str]] = None, include_types: Optional[Sequence[str]] = None, include_tags: Optional[Sequence[str]] = None, exclude_names: Optional[Sequence[str]] = None, exclude_types: Optional[Sequence[str]] = None, exclude_tags: Optional[Sequence[str]] = None, **kwargs: Optional[Any]) → Union[AsyncIterator[RunLogPatch], AsyncIterator[RunLog]]

Stream all output from a runnable, as reported to the callback system. This includes all inner runs of LLMs, Retrievers, Tools, etc.

Output is streamed as Log objects, which include a list of jsonpatch ops that describe how the state of the run has changed in each step, and the final state of the run.

The jsonpatch ops can be applied in order to construct state.

Parameters

• input – The input to the runnable.

• config – The config to use for the runnable.

• diff – Whether to yield diffs between each step, or the current state.

• with_streamed_output_list – Whether to yield the streamed_output list.

• include_names – Only include logs with these names.

• include_types – Only include logs with these types.

• include_tags – Only include logs with these tags.

• exclude_names – Exclude logs with these names.

• exclude_types – Exclude logs with these types.

• exclude_tags – Exclude logs with these tags.

async atransform(input: AsyncIterator[Input], config: Optional[RunnableConfig] = None, **kwargs: Optional[Any]) → AsyncIterator[Output]

Default implementation of atransform, which buffers input and calls astream. Subclasses should override this method if they can start producing output while input is still being generated.

batch(inputs: List[Union[PromptValue, str, List[BaseMessage]]], config: Optional[Union[RunnableConfig, List[RunnableConfig]]] = None, *, return_exceptions: bool = False, **kwargs: Any) → List[str]

Default implementation runs invoke in parallel using a thread pool executor.

The default implementation of batch works well for IO bound runnables.

Subclasses should override this method if they can batch more efficiently; e.g., if the underlying runnable uses an API which supports a batch mode.

bind(**kwargs: Any) → Runnable[Input, Output]

Bind arguments to a Runnable, returning a new Runnable.

config_schema(*, include: Optional[Sequence[str]] = None) → Type[BaseModel]

The type of config this runnable accepts specified as a pydantic model.

To mark a field as configurable, see the configurable_fields and configurable_alternatives methods.

Parameters

include – A list of fields to include in the config schema.

Returns

A pydantic model that can be used to validate config.

configurable_alternatives(which: ConfigurableField, *, default_key: str = 'default', prefix_keys: bool = False, **kwargs: Union[Runnable[Input, Output], Callable[[], Runnable[Input, Output]]]) → RunnableSerializable[Input, Output]

configurable_fields(**kwargs: Union[ConfigurableField, ConfigurableFieldSingleOption, ConfigurableFieldMultiOption]) → RunnableSerializable[Input, Output]

classmethod construct(_fields_set: Optional[SetStr] = None, **values: Any) → Model

Creates a new model setting __dict__ and __fields_set__ from trusted or pre-validated data. Default values are respected, but no other validation is performed. Behaves as if Config.extra = ‘allow’ was set since it adds all passed values

copy(*, include: Optional[Union[AbstractSetIntStr, MappingIntStrAny]] = None, exclude: Optional[Union[AbstractSetIntStr, MappingIntStrAny]] = None, update: Optional[DictStrAny] = None, deep: bool = False) → Model

Duplicate a model, optionally choose which fields to include, exclude and change.

Parameters

• include – fields to include in new model

• exclude – fields to exclude from new model, as with values this takes precedence over include

• update – values to change/add in the new model. Note: the data is not validated before creating the new model: you should trust this data

• deep – set to True to make a deep copy of the model

Returns

new model instance

dict(**kwargs: Any) → Dict

Return a dictionary of the LLM.

classmethod from_orm(obj: Any) → Model

generate(prompts: List[str], stop: Optional[List[str]] = None, callbacks: Union[List[BaseCallbackHandler], BaseCallbackManager, None, List[Optional[Union[List[BaseCallbackHandler], BaseCallbackManager]]]] = None, *, tags: Optional[Union[List[str], List[List[str]]]] = None, metadata: Optional[Union[Dict[str, Any], List[Dict[str, Any]]]] = None, run_name: Optional[Union[str, List[str]]] = None, **kwargs: Any) → LLMResult

Run the LLM on the given prompt and input.

generate_prompt(prompts: List[PromptValue], stop: Optional[List[str]] = None, callbacks: Union[List[BaseCallbackHandler], BaseCallbackManager, None, List[Optional[Union[List[BaseCallbackHandler], BaseCallbackManager]]]] = None, **kwargs: Any) → LLMResult

Pass a sequence of prompts to the model and return model generations.

This method should make use of batched calls for models that expose a batched API.

Use this method when you want to:

1. take advantage of batched calls,

2. need more output from the model than just the top generated value,

3. are building chains that are agnostic to the underlying language model

   type (e.g., pure text completion models vs chat models).

Parameters

• prompts – List of PromptValues. A PromptValue is an object that can be converted to match the format of any language model (string for pure text generation models and BaseMessages for chat models).

• stop – Stop words to use when generating. Model output is cut off at the first occurrence of any of these substrings.

• callbacks – Callbacks to pass through. Used for executing additional functionality, such as logging or streaming, throughout generation.

• **kwargs – Arbitrary additional keyword arguments. These are usually passed to the model provider API call.

Returns

An LLMResult, which contains a list of candidate Generations for each input

prompt and additional model provider-specific output.

get_input_schema(config: Optional[RunnableConfig] = None) → Type[BaseModel]

Get a pydantic model that can be used to validate input to the runnable.

Runnables that leverage the configurable_fields and configurable_alternatives methods will have a dynamic input schema that depends on which configuration the runnable is invoked with.

This method allows to get an input schema for a specific configuration.

Parameters

config – A config to use when generating the schema.

Returns

A pydantic model that can be used to validate input.

classmethod get_lc_namespace() → List[str]

Get the namespace of the langchain object.

For example, if the class is langchain.llms.openai.OpenAI, then the namespace is [“langchain”, “llms”, “openai”]

get_num_tokens(text: str) → int

Get the number of tokens present in the text.

Useful for checking if an input will fit in a model’s context window.

Parameters

text – The string input to tokenize.

Returns

The integer number of tokens in the text.

get_num_tokens_from_messages(messages: List[BaseMessage]) → int

Get the number of tokens in the messages.

Useful for checking if an input will fit in a model’s context window.

Parameters

messages – The message inputs to tokenize.

Returns

The sum of the number of tokens across the messages.

get_output_schema(config: Optional[RunnableConfig] = None) → Type[BaseModel]

Get a pydantic model that can be used to validate output to the runnable.

Runnables that leverage the configurable_fields and configurable_alternatives methods will have a dynamic output schema that depends on which configuration the runnable is invoked with.

This method allows to get an output schema for a specific configuration.

Parameters

config – A config to use when generating the schema.

Returns

A pydantic model that can be used to validate output.

get_token_ids(text: str) → List[int]

Return the ordered ids of the tokens in a text.

Parameters

text – The string input to tokenize.

Returns

A list of ids corresponding to the tokens in the text, in order they occur

in the text.

invoke(input: Union[PromptValue, str, List[BaseMessage]], config: Optional[RunnableConfig] = None, *, stop: Optional[List[str]] = None, **kwargs: Any) → str

Transform a single input into an output. Override to implement.

Parameters

• input – The input to the runnable.

• config – A config to use when invoking the runnable. The config supports standard keys like ‘tags’, ‘metadata’ for tracing purposes, ‘max_concurrency’ for controlling how much work to do in parallel, and other keys. Please refer to the RunnableConfig for more details.

Returns

The output of the runnable.

classmethod is_lc_serializable() → bool

Is this class serializable?

json(*, include: Optional[Union[AbstractSetIntStr, MappingIntStrAny]] = None, exclude: Optional[Union[AbstractSetIntStr, MappingIntStrAny]] = None, by_alias: bool = False, skip_defaults: Optional[bool] = None, exclude_unset: bool = False, exclude_defaults: bool = False, exclude_none: bool = False, encoder: Optional[Callable[[Any], Any]] = None, models_as_dict: bool = True, **dumps_kwargs: Any) → unicode

Generate a JSON representation of the model, include and exclude arguments as per dict().

encoder is an optional function to supply as default to json.dumps(), other arguments as per json.dumps().

classmethod lc_id() → List[str]

A unique identifier for this class for serialization purposes.

The unique identifier is a list of strings that describes the path to the object.

map() → Runnable[List[Input], List[Output]]

Return a new Runnable that maps a list of inputs to a list of outputs, by calling invoke() with each input.

classmethod parse_file(path: Union[str, Path], *, content_type: unicode = None, encoding: unicode = 'utf8', proto: Protocol = None, allow_pickle: bool = False) → Model

classmethod parse_obj(obj: Any) → Model

classmethod parse_raw(b: Union[str, bytes], *, content_type: unicode = None, encoding: unicode = 'utf8', proto: Protocol = None, allow_pickle: bool = False) → Model

predict(text: str, *, stop: Optional[Sequence[str]] = None, **kwargs: Any) → str

Pass a single string input to the model and return a string prediction.

Use this method when passing in raw text. If you want to pass in specific

types of chat messages, use predict_messages.

Parameters

• text – String input to pass to the model.

• stop – Stop words to use when generating. Model output is cut off at the first occurrence of any of these substrings.

• **kwargs – Arbitrary additional keyword arguments. These are usually passed to the model provider API call.

Returns

Top model prediction as a string.

predict_messages(messages: List[BaseMessage], *, stop: Optional[Sequence[str]] = None, **kwargs: Any) → BaseMessage

Pass a message sequence to the model and return a message prediction.

Use this method when passing in chat messages. If you want to pass in raw text,

use predict.

Parameters

• messages – A sequence of chat messages corresponding to a single model input.

• stop – Stop words to use when generating. Model output is cut off at the first occurrence of any of these substrings.

• **kwargs – Arbitrary additional keyword arguments. These are usually passed to the model provider API call.

Returns

Top model prediction as a message.

save(file_path: Union[Path, str]) → None

Save the LLM.

Parameters

file_path – Path to file to save the LLM to.

Example: .. code-block:: python

llm.save(file_path=”path/llm.yaml”)

classmethod schema(by_alias: bool = True, ref_template: unicode = '#/definitions/{model}') → DictStrAny

classmethod schema_json(*, by_alias: bool = True, ref_template: unicode = '#/definitions/{model}', **dumps_kwargs: Any) → unicode

stream(input: Union[PromptValue, str, List[BaseMessage]], config: Optional[RunnableConfig] = None, *, stop: Optional[List[str]] = None, **kwargs: Any) → Iterator[str]

Default implementation of stream, which calls invoke. Subclasses should override this method if they support streaming output.

to_json() → Union[SerializedConstructor, SerializedNotImplemented]

to_json_not_implemented() → SerializedNotImplemented

transform(input: Iterator[Input], config: Optional[RunnableConfig] = None, **kwargs: Optional[Any]) → Iterator[Output]

Default implementation of transform, which buffers input and then calls stream. Subclasses should override this method if they can start producing output while input is still being generated.

classmethod update_forward_refs(**localns: Any) → None

Try to update ForwardRefs on fields based on this Model, globalns and localns.

classmethod validate(value: Any) → Model

with_config(config: Optional[RunnableConfig] = None, **kwargs: Any) → Runnable[Input, Output]

Bind config to a Runnable, returning a new Runnable.

with_fallbacks(fallbacks: Sequence[Runnable[Input, Output]], *, exceptions_to_handle: Tuple[Type[BaseException], ...] = (<class 'Exception'>,)) → RunnableWithFallbacksT[Input, Output]

Add fallbacks to a runnable, returning a new Runnable.

Parameters

• fallbacks – A sequence of runnables to try if the original runnable fails.

• exceptions_to_handle – A tuple of exception types to handle.

Returns

A new Runnable that will try the original runnable, and then each fallback in order, upon failures.

with_listeners(*, on_start: Optional[Listener] = None, on_end: Optional[Listener] = None, on_error: Optional[Listener] = None) → Runnable[Input, Output]

Bind lifecycle listeners to a Runnable, returning a new Runnable.

on_start: Called before the runnable starts running, with the Run object. on_end: Called after the runnable finishes running, with the Run object. on_error: Called if the runnable throws an error, with the Run object.

The Run object contains information about the run, including its id, type, input, output, error, start_time, end_time, and any tags or metadata added to the run.

with_retry(*, retry_if_exception_type: ~typing.Tuple[~typing.Type[BaseException], ...] = (<class 'Exception'>,), wait_exponential_jitter: bool = True, stop_after_attempt: int = 3) → Runnable[Input, Output]

Create a new Runnable that retries the original runnable on exceptions.

Parameters

• retry_if_exception_type – A tuple of exception types to retry on

• wait_exponential_jitter – Whether to add jitter to the wait time between retries

• stop_after_attempt – The maximum number of attempts to make before giving up

Returns

A new Runnable that retries the original runnable on exceptions.

with_types(*, input_type: Optional[Type[Input]] = None, output_type: Optional[Type[Output]] = None) → Runnable[Input, Output]

Bind input and output types to a Runnable, returning a new Runnable.

property InputType: TypeAlias

Get the input type for this runnable.

property OutputType: Type[str]

Get the input type for this runnable.

property config_specs: List[langchain_core.runnables.utils.ConfigurableFieldSpec]

List configurable fields for this runnable.

property input_schema: Type[pydantic.main.BaseModel]

The type of input this runnable accepts specified as a pydantic model.

property lc_attributes: Dict

List of attribute names that should be included in the serialized kwargs.

These attributes must be accepted by the constructor.

property lc_secrets: Dict[str, str]

A map of constructor argument names to secret ids.

For example,

{“openai_api_key”: “OPENAI_API_KEY”}

property output_schema: Type[pydantic.main.BaseModel]

The type of output this runnable produces specified as a pydantic model.

Examples using LlamaCpp

• Llama.cpp

• Run LLMs locally

• Set env var OPENAI_API_KEY or load from a .env file

• Use local LLMs

• WebResearchRetriever