llama.cpp/llama_cpp/llama.py

import os
import uuid
import time
import multiprocessing
from typing import List, Optional, Union, Generator, Sequence, Iterator
from collections import deque

from . import llama_cpp
from .llama_types import *


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

    def __init__(
        self,
        model_path: str,
        # NOTE: These parameters are likely to change in the future.
        n_ctx: int = 512,
        n_parts: int = -1,
        seed: int = 1337,
        f16_kv: bool = False,
        logits_all: bool = False,
        vocab_only: bool = False,
        use_mlock: bool = False,
        embedding: bool = False,
        n_threads: Optional[int] = None,
        n_batch: int = 8,
        last_n_tokens_size: int = 64,
    ):
        """Load a llama.cpp model from `model_path`.

        Args:
            model_path: Path to the model.
            n_ctx: Maximum context size.
            n_parts: Number of parts to split the model into. If -1, the number of parts is automatically determined.
            seed: Random seed. 0 for random.
            f16_kv: Use half-precision for key/value cache.
            logits_all: Return logits for all tokens, not just the last token.
            vocab_only: Only load the vocabulary no weights.
            use_mlock: Force the system to keep the model in RAM.
            embedding: Embedding mode only.
            n_threads: Number of threads to use. If None, the number of threads is automatically determined.
            n_batch: Maximum number of prompt tokens to batch together when calling llama_eval.
            last_n_tokens_size: Maximum number of tokens to keep in the last_n_tokens deque.

        Raises:
            ValueError: If the model path does not exist.

        Returns:
            A Llama instance.
        """
        self.model_path = model_path

        self.params = llama_cpp.llama_context_default_params()
        self.params.n_ctx = n_ctx
        self.params.n_parts = n_parts
        self.params.seed = seed
        self.params.f16_kv = f16_kv
        self.params.logits_all = logits_all
        self.params.vocab_only = vocab_only
        self.params.use_mlock = use_mlock
        self.params.embedding = embedding

        self.last_n_tokens_size = last_n_tokens_size
        self.last_n_tokens_data = deque(
            [llama_cpp.llama_token(0)] * self.last_n_tokens_size,
            maxlen=self.last_n_tokens_size,
        )
        self.tokens_consumed = 0
        self.n_batch = n_batch

        self.n_threads = n_threads or multiprocessing.cpu_count()

        if not os.path.exists(model_path):
            raise ValueError(f"Model path does not exist: {model_path}")

        self.ctx = llama_cpp.llama_init_from_file(
            self.model_path.encode("utf-8"), self.params
        )

    def tokenize(self, text: bytes) -> List[llama_cpp.llama_token]:
        """Tokenize a string.

        Args:
            text: The utf-8 encoded string to tokenize.

        Raises:
            RuntimeError: If the tokenization failed.

        Returns:
            A list of tokens.
        """
        assert self.ctx is not None
        n_ctx = llama_cpp.llama_n_ctx(self.ctx)
        tokens = (llama_cpp.llama_token * int(n_ctx))()
        n_tokens = llama_cpp.llama_tokenize(
            self.ctx,
            text,
            tokens,
            n_ctx,
            llama_cpp.c_bool(True),
        )
        if int(n_tokens) < 0:
            raise RuntimeError(f'Failed to tokenize: text="{text}" n_tokens={n_tokens}')
        return list(tokens[:n_tokens])

    def detokenize(self, tokens: List[llama_cpp.llama_token]) -> bytes:
        """Detokenize a list of tokens.

        Args:
            tokens: The list of tokens to detokenize.

        Returns:
            The detokenized string.
        """
        assert self.ctx is not None
        output = b""
        for token in tokens:
            output += llama_cpp.llama_token_to_str(self.ctx, token)
        return output

    def reset(self):
        """Reset the model state."""
        self.last_n_tokens_data.extend(
            [llama_cpp.llama_token(0)] * self.last_n_tokens_size
        )
        self.tokens_consumed = 0

    def eval(self, tokens: Sequence[llama_cpp.llama_token]):
        """Evaluate a list of tokens.

        Args:
            tokens: The list of tokens to evaluate.
        """
        assert self.ctx is not None
        n_ctx = int(llama_cpp.llama_n_ctx(self.ctx))
        for i in range(0, len(tokens), self.n_batch):
            batch = tokens[i : min(len(tokens), i + self.n_batch)]
            n_past = min(n_ctx - len(batch), self.tokens_consumed)
            return_code = llama_cpp.llama_eval(
                ctx=self.ctx,
                tokens=(llama_cpp.llama_token * len(batch))(*batch),
                n_tokens=llama_cpp.c_int(len(batch)),
                n_past=llama_cpp.c_int(n_past),
                n_threads=llama_cpp.c_int(self.n_threads),
            )
            if int(return_code) != 0:
                raise RuntimeError(f"llama_eval returned {return_code}")
            self.last_n_tokens_data.extend(batch)
            self.tokens_consumed += len(batch)

    def sample(
        self,
        top_k: int,
        top_p: float,
        temp: float,
        repeat_penalty: float,
    ):
        """Sample a token from the model.

        Args:
            top_k: The top-k sampling parameter.
            top_p: The top-p sampling parameter.
            temp: The temperature parameter.
            repeat_penalty: The repeat penalty parameter.

        Returns:
            The sampled token.
        """
        assert self.ctx is not None
        # Temporary workaround for https://github.com/ggerganov/llama.cpp/issues/684
        if temp == 0.0:
            temp = 1.0
            top_p = 0.0
            top_k = 1
        return llama_cpp.llama_sample_top_p_top_k(
            ctx=self.ctx,
            last_n_tokens_data=(llama_cpp.llama_token * self.last_n_tokens_size)(
                *self.last_n_tokens_data
            ),
            last_n_tokens_size=llama_cpp.c_int(self.last_n_tokens_size),
            top_k=llama_cpp.c_int(top_k),
            top_p=llama_cpp.c_float(top_p),
            temp=llama_cpp.c_float(temp),
            repeat_penalty=llama_cpp.c_float(repeat_penalty),
        )

    def generate(
        self,
        tokens: Sequence[llama_cpp.llama_token],
        top_k: int,
        top_p: float,
        temp: float,
        repeat_penalty: float,
    ) -> Generator[
        llama_cpp.llama_token, Optional[Sequence[llama_cpp.llama_token]], None
    ]:
        """Create a generator of tokens from a prompt.

        Examples:
            >>> llama = Llama("models/ggml-7b.bin")
            >>> tokens = llama.tokenize(b"Hello, world!")
            >>> for token in llama.generate(tokens, top_k=40, top_p=0.95, temp=1.0, repeat_penalty=1.1):
            ...     print(llama.detokenize([token]))

        Args:
            tokens: The prompt tokens.
            top_k: The top-k sampling parameter.
            top_p: The top-p sampling parameter.
            temp: The temperature parameter.
            repeat_penalty: The repeat penalty parameter.

        Yields:
            The generated tokens.
        """
        assert self.ctx is not None
        self.reset()
        while True:
            self.eval(tokens)
            token = self.sample(
                top_k=top_k,
                top_p=top_p,
                temp=temp,
                repeat_penalty=repeat_penalty,
            )
            tokens_or_none = yield token
            tokens = [token]
            if tokens_or_none is not None:
                tokens.extend(tokens_or_none)

    def create_embedding(self, input: str) -> Embedding:
        """Embed a string.

        Args:
            input: The utf-8 encoded string to embed.

        Returns:
            An embedding object.
        """
        assert self.ctx is not None
        tokens = self.tokenize(input.encode("utf-8"))
        self.reset()
        self.eval(tokens)
        n_tokens = len(tokens)
        embedding = llama_cpp.llama_get_embeddings(self.ctx)[
            : llama_cpp.llama_n_embd(self.ctx)
        ]
        return {
            "object": "list",
            "data": [
                {
                    "object": "embedding",
                    "embedding": embedding,
                    "index": 0,
                }
            ],
            "model": self.model_path,
            "usage": {
                "prompt_tokens": n_tokens,
                "total_tokens": n_tokens,
            },
        }

    def embed(self, input: str) -> List[float]:
        """Embed a string.

        Args:
            input: The utf-8 encoded string to embed.

        Returns:
            A list of embeddings
        """
        return list(map(float, self.create_embedding(input)["data"][0]["embedding"]))

    def _create_completion(
        self,
        prompt: str,
        suffix: Optional[str] = None,
        max_tokens: int = 16,
        temperature: float = 0.8,
        top_p: float = 0.95,
        logprobs: Optional[int] = None,
        echo: bool = False,
        stop: List[str] = [],
        repeat_penalty: float = 1.1,
        top_k: int = 40,
        stream: bool = False,
    ) -> Union[Iterator[Completion], Iterator[CompletionChunk],]:
        assert self.ctx is not None
        completion_id = f"cmpl-{str(uuid.uuid4())}"
        created = int(time.time())
        completion_tokens: List[llama_cpp.llama_token] = []
        # Add blank space to start of prompt to match OG llama tokenizer
        prompt_tokens = self.tokenize(b" " + prompt.encode("utf-8"))
        text = b""
        returned_characters = 0

        if len(prompt_tokens) + max_tokens > int(llama_cpp.llama_n_ctx(self.ctx)):
            raise ValueError(
                f"Requested tokens exceed context window of {llama_cpp.llama_n_ctx(self.ctx)}"
            )

        if stop != []:
            stop_sequences = [s.encode("utf-8") for s in stop]
        else:
            stop_sequences = []

        finish_reason = None
        for token in self.generate(
            prompt_tokens,
            top_k=top_k,
            top_p=top_p,
            temp=temperature,
            repeat_penalty=repeat_penalty,
        ):
            if token == llama_cpp.llama_token_eos():
                text = self.detokenize(completion_tokens)
                finish_reason = "stop"
                break
            completion_tokens.append(token)

            all_text = self.detokenize(completion_tokens)
            any_stop = [s for s in stop_sequences if s in all_text]
            if len(any_stop) > 0:
                first_stop = any_stop[0]
                text = all_text[: all_text.index(first_stop)]
                finish_reason = "stop"
                break

            if stream:
                start = returned_characters
                longest = 0
                # We want to avoid yielding any characters from
                # the generated text if they are part of a stop
                # sequence.
                for s in stop_sequences:
                    for i in range(len(s), 0, -1):
                        if all_text.endswith(s[:i]):
                            if i > longest:
                                longest = i
                            break
                text = all_text[: len(all_text) - longest]
                returned_characters += len(text[start:])
                yield {
                    "id": completion_id,
                    "object": "text_completion",
                    "created": created,
                    "model": self.model_path,
                    "choices": [
                        {
                            "text": text[start:].decode("utf-8"),
                            "index": 0,
                            "logprobs": None,
                            "finish_reason": None,
                        }
                    ],
                }
            if len(completion_tokens) >= max_tokens:
                text = self.detokenize(completion_tokens)
                finish_reason = "length"
                break

        if finish_reason is None:
            finish_reason = "length"

        if stream:
            yield {
                "id": completion_id,
                "object": "text_completion",
                "created": created,
                "model": self.model_path,
                "choices": [
                    {
                        "text": text[returned_characters:].decode("utf-8"),
                        "index": 0,
                        "logprobs": None,
                        "finish_reason": finish_reason,
                    }
                ],
            }
            return

        text = text.decode("utf-8")

        if echo:
            text = prompt + text

        if suffix is not None:
            text = text + suffix

        if logprobs is not None:
            raise NotImplementedError("logprobs not implemented")

        yield {
            "id": completion_id,
            "object": "text_completion",
            "created": created,
            "model": self.model_path,
            "choices": [
                {
                    "text": text,
                    "index": 0,
                    "logprobs": None,
                    "finish_reason": finish_reason,
                }
            ],
            "usage": {
                "prompt_tokens": len(prompt_tokens),
                "completion_tokens": len(completion_tokens),
                "total_tokens": len(prompt_tokens) + len(completion_tokens),
            },
        }

    def create_completion(
        self,
        prompt: str,
        suffix: Optional[str] = None,
        max_tokens: int = 128,
        temperature: float = 0.8,
        top_p: float = 0.95,
        logprobs: Optional[int] = None,
        echo: bool = False,
        stop: List[str] = [],
        repeat_penalty: float = 1.1,
        top_k: int = 40,
        stream: bool = False,
    ) -> Union[Completion, Iterator[CompletionChunk]]:
        """Generate text from a prompt.

        Args:
            prompt: The prompt to generate text from.
            suffix: A suffix to append to the generated text. If None, no suffix is appended.
            max_tokens: The maximum number of tokens to generate.
            temperature: The temperature to use for sampling.
            top_p: The top-p value to use for sampling.
            logprobs: The number of logprobs to return. If None, no logprobs are returned.
            echo: Whether to echo the prompt.
            stop: A list of strings to stop generation when encountered.
            repeat_penalty: The penalty to apply to repeated tokens.
            top_k: The top-k value to use for sampling.
            stream: Whether to stream the results.

        Raises:
            ValueError: If the requested tokens exceed the context window.
            RuntimeError: If the prompt fails to tokenize or the model fails to evaluate the prompt.

        Returns:
            Response object containing the generated text.
        """
        completion_or_chunks = self._create_completion(
            prompt=prompt,
            suffix=suffix,
            max_tokens=max_tokens,
            temperature=temperature,
            top_p=top_p,
            logprobs=logprobs,
            echo=echo,
            stop=stop,
            repeat_penalty=repeat_penalty,
            top_k=top_k,
            stream=stream,
        )
        if stream:
            chunks: Iterator[CompletionChunk] = completion_or_chunks
            return chunks
        completion: Completion = next(completion_or_chunks)  # type: ignore
        return completion

    def __call__(
        self,
        prompt: str,
        suffix: Optional[str] = None,
        max_tokens: int = 128,
        temperature: float = 0.8,
        top_p: float = 0.95,
        logprobs: Optional[int] = None,
        echo: bool = False,
        stop: List[str] = [],
        repeat_penalty: float = 1.1,
        top_k: int = 40,
        stream: bool = False,
    ) -> Union[Completion, Iterator[CompletionChunk]]:
        """Generate text from a prompt.

        Args:
            prompt: The prompt to generate text from.
            suffix: A suffix to append to the generated text. If None, no suffix is appended.
            max_tokens: The maximum number of tokens to generate.
            temperature: The temperature to use for sampling.
            top_p: The top-p value to use for sampling.
            logprobs: The number of logprobs to return. If None, no logprobs are returned.
            echo: Whether to echo the prompt.
            stop: A list of strings to stop generation when encountered.
            repeat_penalty: The penalty to apply to repeated tokens.
            top_k: The top-k value to use for sampling.
            stream: Whether to stream the results.

        Raises:
            ValueError: If the requested tokens exceed the context window.
            RuntimeError: If the prompt fails to tokenize or the model fails to evaluate the prompt.

        Returns:
            Response object containing the generated text.
        """
        return self.create_completion(
            prompt=prompt,
            suffix=suffix,
            max_tokens=max_tokens,
            temperature=temperature,
            top_p=top_p,
            logprobs=logprobs,
            echo=echo,
            stop=stop,
            repeat_penalty=repeat_penalty,
            top_k=top_k,
            stream=stream,
        )

    def _convert_text_completion_to_chat(
        self, completion: Completion
    ) -> ChatCompletion:
        return {
            "id": "chat" + completion["id"],
            "object": "chat.completion",
            "created": completion["created"],
            "model": completion["model"],
            "choices": [
                {
                    "index": 0,
                    "message": {
                        "role": "assistant",
                        "content": completion["choices"][0]["text"],
                    },
                    "finish_reason": completion["choices"][0]["finish_reason"],
                }
            ],
            "usage": completion["usage"],
        }

    def _convert_text_completion_chunks_to_chat(
        self,
        chunks: Iterator[CompletionChunk],
    ) -> Iterator[ChatCompletionChunk]:
        for i, chunk in enumerate(chunks):
            if i == 0:
                yield {
                    "id": "chat" + chunk["id"],
                    "model": chunk["model"],
                    "created": chunk["created"],
                    "object": "chat.completion.chunk",
                    "choices": [
                        {
                            "index": 0,
                            "delta": {
                                "role": "assistant",
                            },
                            "finish_reason": None,
                        }
                    ],
                }
            yield {
                "id": "chat" + chunk["id"],
                "model": chunk["model"],
                "created": chunk["created"],
                "object": "chat.completion.chunk",
                "choices": [
                    {
                        "index": 0,
                        "delta": {
                            "content": chunk["choices"][0]["text"],
                        },
                        "finish_reason": chunk["choices"][0]["finish_reason"],
                    }
                ],
            }

    def create_chat_completion(
        self,
        messages: List[ChatCompletionMessage],
        temperature: float = 0.8,
        top_p: float = 0.95,
        top_k: int = 40,
        stream: bool = False,
        stop: List[str] = [],
        max_tokens: int = 128,
        repeat_penalty: float = 1.1,
    ) -> Union[ChatCompletion, Iterator[ChatCompletionChunk]]:
        """Generate a chat completion from a list of messages.

        Args:
            messages: A list of messages to generate a response for.
            temperature: The temperature to use for sampling.
            top_p: The top-p value to use for sampling.
            top_k: The top-k value to use for sampling.
            stream: Whether to stream the results.
            stop: A list of strings to stop generation when encountered.
            max_tokens: The maximum number of tokens to generate.
            repeat_penalty: The penalty to apply to repeated tokens.

        Returns:
            Generated chat completion or a stream of chat completion chunks.
        """
        instructions = """Complete the following chat conversation between the user and the assistant. System messages should be strictly followed as additional instructions."""
        chat_history = "\n".join(
            f'{message["role"]} {message.get("user", "")}: {message["content"]}'
            for message in messages
        )
        PROMPT = f" \n\n### Instructions:{instructions}\n\n### Inputs:{chat_history}\n\n### Response:\nassistant: "
        PROMPT_STOP = ["###", "\nuser: ", "\nassistant: ", "\nsystem: "]
        completion_or_chunks = self(
            prompt=PROMPT,
            stop=PROMPT_STOP + stop,
            temperature=temperature,
            top_p=top_p,
            top_k=top_k,
            stream=stream,
            max_tokens=max_tokens,
            repeat_penalty=repeat_penalty,
        )
        if stream:
            chunks: Iterator[CompletionChunk] = completion_or_chunks  # type: ignore
            return self._convert_text_completion_chunks_to_chat(chunks)
        else:
            completion: Completion = completion_or_chunks  # type: ignore
            return self._convert_text_completion_to_chat(completion)

    def __del__(self):
        if self.ctx is not None:
            llama_cpp.llama_free(self.ctx)
            self.ctx = None

    @staticmethod
    def token_eos() -> llama_cpp.llama_token:
        """Return the end-of-sequence token."""
        return llama_cpp.llama_token_eos()

    @staticmethod
    def token_bos() -> llama_cpp.llama_token:
        """Return the beginning-of-sequence token."""
        return llama_cpp.llama_token_bos()