llama.cpp/llama_cpp/server/app.py

import json
import multiprocessing
from threading import Lock
from functools import partial
from typing import Iterator, List, Optional, Union, Dict
from typing_extensions import TypedDict, Literal

import llama_cpp

import anyio
from anyio.streams.memory import MemoryObjectSendStream
from starlette.concurrency import run_in_threadpool, iterate_in_threadpool
from fastapi import Depends, FastAPI, APIRouter, Request
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel, Field
from pydantic_settings import BaseSettings
from sse_starlette.sse import EventSourceResponse


class Settings(BaseSettings):
    model: str = Field(
        description="The path to the model to use for generating completions."
    )
    model_alias: Optional[str] = Field(
        default=None,
        description="The alias of the model to use for generating completions.",
    )
    n_ctx: int = Field(default=2048, ge=1, description="The context size.")
    n_gpu_layers: int = Field(
        default=0,
        ge=0,
        description="The number of layers to put on the GPU. The rest will be on the CPU.",
    )
    tensor_split: Optional[List[float]] = Field(
        default=None,
        description="Split layers across multiple GPUs in proportion.",
    )
    rope_freq_base: float = Field(default=10000, ge=1, description="RoPE base frequency")
    rope_freq_scale: float = Field(default=1.0, description="RoPE frequency scaling factor")
    seed: int = Field(
        default=1337, description="Random seed. -1 for random."
    )
    n_batch: int = Field(
        default=512, ge=1, description="The batch size to use per eval."
    )
    n_threads: int = Field(
        default=max(multiprocessing.cpu_count() // 2, 1),
        ge=1,
        description="The number of threads to use.",
    )
    f16_kv: bool = Field(default=True, description="Whether to use f16 key/value.")
    use_mlock: bool = Field(
        default=llama_cpp.llama_mlock_supported(),
        description="Use mlock.",
    )
    use_mmap: bool = Field(
        default=llama_cpp.llama_mmap_supported(),
        description="Use mmap.",
    )
    embedding: bool = Field(default=True, description="Whether to use embeddings.")
    low_vram: bool = Field(
        default=False,
        description="Whether to use less VRAM. This will reduce performance.",
    )
    last_n_tokens_size: int = Field(
        default=64,
        ge=0,
        description="Last n tokens to keep for repeat penalty calculation.",
    )
    logits_all: bool = Field(default=True, description="Whether to return logits.")
    cache: bool = Field(
        default=False,
        description="Use a cache to reduce processing times for evaluated prompts.",
    )
    cache_type: Literal["ram", "disk"] = Field(
        default="ram",
        description="The type of cache to use. Only used if cache is True.",
    )
    cache_size: int = Field(
        default=2 << 30,
        description="The size of the cache in bytes. Only used if cache is True.",
    )
    vocab_only: bool = Field(
        default=False, description="Whether to only return the vocabulary."
    )
    verbose: bool = Field(
        default=True, description="Whether to print debug information."
    )
    host: str = Field(default="localhost", description="Listen address")
    port: int = Field(default=8000, description="Listen port")
    interrupt_requests: bool = Field(
        default=True,
        description="Whether to interrupt requests when a new request is received.",
    )


router = APIRouter()

settings: Optional[Settings] = None
llama: Optional[llama_cpp.Llama] = None


def create_app(settings: Optional[Settings] = None):
    if settings is None:
        settings = Settings()
    app = FastAPI(
        title="🦙 llama.cpp Python API",
        version="0.0.1",
    )
    app.add_middleware(
        CORSMiddleware,
        allow_origins=["*"],
        allow_credentials=True,
        allow_methods=["*"],
        allow_headers=["*"],
    )
    app.include_router(router)
    global llama
    llama = llama_cpp.Llama(
        model_path=settings.model,
        n_gpu_layers=settings.n_gpu_layers,
        tensor_split=settings.tensor_split,
        rope_freq_base=settings.rope_freq_base,
        rope_freq_scale=settings.rope_freq_scale,
        seed=settings.seed,
        f16_kv=settings.f16_kv,
        use_mlock=settings.use_mlock,
        use_mmap=settings.use_mmap,
        embedding=settings.embedding,
        logits_all=settings.logits_all,
        n_threads=settings.n_threads,
        n_batch=settings.n_batch,
        n_ctx=settings.n_ctx,
        last_n_tokens_size=settings.last_n_tokens_size,
        vocab_only=settings.vocab_only,
        verbose=settings.verbose,
    )
    if settings.cache:
        if settings.cache_type == "disk":
            if settings.verbose:
                print(f"Using disk cache with size {settings.cache_size}")
            cache = llama_cpp.LlamaDiskCache(capacity_bytes=settings.cache_size)
        else:
            if settings.verbose:
                print(f"Using ram cache with size {settings.cache_size}")
            cache = llama_cpp.LlamaRAMCache(capacity_bytes=settings.cache_size)

        cache = llama_cpp.LlamaCache(capacity_bytes=settings.cache_size)
        llama.set_cache(cache)

    def set_settings(_settings: Settings):
        global settings
        settings = _settings

    set_settings(settings)
    return app


llama_outer_lock = Lock()
llama_inner_lock = Lock()


def get_llama():
    # NOTE: This double lock allows the currently streaming llama model to
    # check if any other requests are pending in the same thread and cancel
    # the stream if so.
    llama_outer_lock.acquire()
    release_outer_lock = True
    try:
        llama_inner_lock.acquire()
        try:
            llama_outer_lock.release()
            release_outer_lock = False
            yield llama
        finally:
            llama_inner_lock.release()
    finally:
        if release_outer_lock:
            llama_outer_lock.release()


def get_settings():
    yield settings


model_field = Field(description="The model to use for generating completions.", default=None)

max_tokens_field = Field(
    default=16, ge=1, le=2048, description="The maximum number of tokens to generate."
)

temperature_field = Field(
    default=0.8,
    ge=0.0,
    le=2.0,
    description="Adjust the randomness of the generated text.\n\n"
    + "Temperature is a hyperparameter that controls the randomness of the generated text. It affects the probability distribution of the model's output tokens. A higher temperature (e.g., 1.5) makes the output more random and creative, while a lower temperature (e.g., 0.5) makes the output more focused, deterministic, and conservative. The default value is 0.8, which provides a balance between randomness and determinism. At the extreme, a temperature of 0 will always pick the most likely next token, leading to identical outputs in each run.",
)

top_p_field = Field(
    default=0.95,
    ge=0.0,
    le=1.0,
    description="Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P.\n\n"
    + "Top-p sampling, also known as nucleus sampling, is another text generation method that selects the next token from a subset of tokens that together have a cumulative probability of at least p. This method provides a balance between diversity and quality by considering both the probabilities of tokens and the number of tokens to sample from. A higher value for top_p (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text.",
)

stop_field = Field(
    default=None,
    description="A list of tokens at which to stop generation. If None, no stop tokens are used.",
)

stream_field = Field(
    default=False,
    description="Whether to stream the results as they are generated. Useful for chatbots.",
)

top_k_field = Field(
    default=40,
    ge=0,
    description="Limit the next token selection to the K most probable tokens.\n\n"
    + "Top-k sampling is a text generation method that selects the next token only from the top k most likely tokens predicted by the model. It helps reduce the risk of generating low-probability or nonsensical tokens, but it may also limit the diversity of the output. A higher value for top_k (e.g., 100) will consider more tokens and lead to more diverse text, while a lower value (e.g., 10) will focus on the most probable tokens and generate more conservative text.",
)

repeat_penalty_field = Field(
    default=1.1,
    ge=0.0,
    description="A penalty applied to each token that is already generated. This helps prevent the model from repeating itself.\n\n"
    + "Repeat penalty is a hyperparameter used to penalize the repetition of token sequences during text generation. It helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient.",
)

presence_penalty_field = Field(
    default=0.0,
    ge=-2.0,
    le=2.0,
    description="Positive values penalize new tokens based on whether they appear in the text so far, increasing the model's likelihood to talk about new topics.",
)

frequency_penalty_field = Field(
    default=0.0,
    ge=-2.0,
    le=2.0,
    description="Positive values penalize new tokens based on their existing frequency in the text so far, decreasing the model's likelihood to repeat the same line verbatim.",
)

mirostat_mode_field = Field(
    default=0,
    ge=0,
    le=2,
    description="Enable Mirostat constant-perplexity algorithm of the specified version (1 or 2; 0 = disabled)",
)

mirostat_tau_field = Field(
    default=5.0,
    ge=0.0,
    le=10.0,
    description="Mirostat target entropy, i.e. the target perplexity - lower values produce focused and coherent text, larger values produce more diverse and less coherent text",
)

mirostat_eta_field = Field(
    default=0.1, ge=0.001, le=1.0, description="Mirostat learning rate"
)


class CreateCompletionRequest(BaseModel):
    prompt: Union[str, List[str]] = Field(
        default="", description="The prompt to generate completions for."
    )
    suffix: Optional[str] = Field(
        default=None,
        description="A suffix to append to the generated text. If None, no suffix is appended. Useful for chatbots.",
    )
    max_tokens: int = max_tokens_field
    temperature: float = temperature_field
    top_p: float = top_p_field
    mirostat_mode: int = mirostat_mode_field
    mirostat_tau: float = mirostat_tau_field
    mirostat_eta: float = mirostat_eta_field
    echo: bool = Field(
        default=False,
        description="Whether to echo the prompt in the generated text. Useful for chatbots.",
    )
    stop: Optional[Union[str, List[str]]] = stop_field
    stream: bool = stream_field
    logprobs: Optional[int] = Field(
        default=None,
        ge=0,
        description="The number of logprobs to generate. If None, no logprobs are generated.",
    )
    presence_penalty: Optional[float] = presence_penalty_field
    frequency_penalty: Optional[float] = frequency_penalty_field
    logit_bias: Optional[Dict[str, float]] = Field(None)
    logprobs: Optional[int] = Field(None)

    # ignored or currently unsupported
    model: Optional[str] = model_field
    n: Optional[int] = 1
    best_of: Optional[int] = 1
    user: Optional[str] = Field(default=None)

    # llama.cpp specific parameters
    top_k: int = top_k_field
    repeat_penalty: float = repeat_penalty_field
    logit_bias_type: Optional[Literal["input_ids", "tokens"]] = Field(None)

    model_config = {
        "json_schema_extra": {
            "examples": [
                {
                    "prompt": "\n\n### Instructions:\nWhat is the capital of France?\n\n### Response:\n",
                    "stop": ["\n", "###"],
                }
            ]
        }
    }


def make_logit_bias_processor(
    llama: llama_cpp.Llama,
    logit_bias: Dict[str, float],
    logit_bias_type: Optional[Literal["input_ids", "tokens"]],
):
    if logit_bias_type is None:
        logit_bias_type = "input_ids"

    to_bias: Dict[int, float] = {}
    if logit_bias_type == "input_ids":
        for input_id, score in logit_bias.items():
            input_id = int(input_id)
            to_bias[input_id] = score

    elif logit_bias_type == "tokens":
        for token, score in logit_bias.items():
            token = token.encode("utf-8")
            for input_id in llama.tokenize(token, add_bos=False):
                to_bias[input_id] = score

    def logit_bias_processor(
        input_ids: List[int],
        scores: List[float],
    ) -> List[float]:
        new_scores = [None] * len(scores)
        for input_id, score in enumerate(scores):
            new_scores[input_id] = score + to_bias.get(input_id, 0.0)

        return new_scores

    return logit_bias_processor


@router.post(
    "/v1/completions",
)
async def create_completion(
    request: Request,
    body: CreateCompletionRequest,
    llama: llama_cpp.Llama = Depends(get_llama),
) -> llama_cpp.Completion:
    if isinstance(body.prompt, list):
        assert len(body.prompt) <= 1
        body.prompt = body.prompt[0] if len(body.prompt) > 0 else ""

    exclude = {
        "n",
        "best_of",
        "logit_bias",
        "logit_bias_type",
        "user",
    }
    kwargs = body.model_dump(exclude=exclude)

    if body.logit_bias is not None:
        kwargs['logits_processor'] = llama_cpp.LogitsProcessorList([
            make_logit_bias_processor(llama, body.logit_bias, body.logit_bias_type),
        ])

    if body.stream:
        send_chan, recv_chan = anyio.create_memory_object_stream(10)

        async def event_publisher(inner_send_chan: MemoryObjectSendStream):
            async with inner_send_chan:
                try:
                    iterator: Iterator[llama_cpp.CompletionChunk] = await run_in_threadpool(llama, **kwargs)  # type: ignore
                    async for chunk in iterate_in_threadpool(iterator):
                        await inner_send_chan.send(dict(data=json.dumps(chunk)))
                        if await request.is_disconnected():
                            raise anyio.get_cancelled_exc_class()()
                        if settings.interrupt_requests and llama_outer_lock.locked():
                            await inner_send_chan.send(dict(data="[DONE]"))
                            raise anyio.get_cancelled_exc_class()()
                    await inner_send_chan.send(dict(data="[DONE]"))
                except anyio.get_cancelled_exc_class() as e:
                    print("disconnected")
                    with anyio.move_on_after(1, shield=True):
                        print(
                            f"Disconnected from client (via refresh/close) {request.client}"
                        )
                        raise e

        return EventSourceResponse(
            recv_chan, data_sender_callable=partial(event_publisher, send_chan)
        ) # type: ignore
    else:
        completion: llama_cpp.Completion = await run_in_threadpool(llama, **kwargs)  # type: ignore
        return completion


class CreateEmbeddingRequest(BaseModel):
    model: Optional[str] = model_field
    input: Union[str, List[str]] = Field(description="The input to embed.")
    user: Optional[str] = Field(default=None)

    model_config = {
        "json_schema_extra": {
            "examples": [
                {
                    "input": "The food was delicious and the waiter...",
                }
            ]
        }
    }


@router.post(
    "/v1/embeddings",
)
async def create_embedding(
    request: CreateEmbeddingRequest, llama: llama_cpp.Llama = Depends(get_llama)
):
    return await run_in_threadpool(
        llama.create_embedding, **request.model_dump(exclude={"user"})
    )


class ChatCompletionRequestMessage(BaseModel):
    role: Literal["system", "user", "assistant"] = Field(
        default="user", description="The role of the message."
    )
    content: str = Field(default="", description="The content of the message.")


class CreateChatCompletionRequest(BaseModel):
    messages: List[ChatCompletionRequestMessage] = Field(
        default=[], description="A list of messages to generate completions for."
    )
    max_tokens: int = max_tokens_field
    temperature: float = temperature_field
    top_p: float = top_p_field
    mirostat_mode: int = mirostat_mode_field
    mirostat_tau: float = mirostat_tau_field
    mirostat_eta: float = mirostat_eta_field
    stop: Optional[List[str]] = stop_field
    stream: bool = stream_field
    presence_penalty: Optional[float] = presence_penalty_field
    frequency_penalty: Optional[float] = frequency_penalty_field
    logit_bias: Optional[Dict[str, float]] = Field(None)

    # ignored or currently unsupported
    model: Optional[str] = model_field
    n: Optional[int] = 1
    user: Optional[str] = Field(None)

    # llama.cpp specific parameters
    top_k: int = top_k_field
    repeat_penalty: float = repeat_penalty_field
    logit_bias_type: Optional[Literal["input_ids", "tokens"]] = Field(None)

    model_config = {
        "json_schema_extra": {
            "examples": [
                {
                    "messages": [
                        ChatCompletionRequestMessage(
                            role="system", content="You are a helpful assistant."
                        ).model_dump(),
                        ChatCompletionRequestMessage(
                            role="user", content="What is the capital of France?"
                        ).model_dump(),
                    ]
                }
            ]
        }
    }


@router.post(
    "/v1/chat/completions",
)
async def create_chat_completion(
    request: Request,
    body: CreateChatCompletionRequest,
    llama: llama_cpp.Llama = Depends(get_llama),
    settings: Settings = Depends(get_settings),
) -> llama_cpp.ChatCompletion:
    exclude = {
        "n",
        "logit_bias",
        "logit_bias_type",
        "user",
    }
    kwargs = body.model_dump(exclude=exclude)

    if body.logit_bias is not None:
        kwargs['logits_processor'] = llama_cpp.LogitsProcessorList([
            make_logit_bias_processor(llama, body.logit_bias, body.logit_bias_type),
        ])

    if body.stream:
        send_chan, recv_chan = anyio.create_memory_object_stream(10)

        async def event_publisher(inner_send_chan: MemoryObjectSendStream):
            async with inner_send_chan:
                try:
                    iterator: Iterator[llama_cpp.ChatCompletionChunk] = await run_in_threadpool(llama.create_chat_completion, **kwargs)  # type: ignore
                    async for chat_chunk in iterate_in_threadpool(iterator):
                        await inner_send_chan.send(dict(data=json.dumps(chat_chunk)))
                        if await request.is_disconnected():
                            raise anyio.get_cancelled_exc_class()()
                        if settings.interrupt_requests and llama_outer_lock.locked():
                            await inner_send_chan.send(dict(data="[DONE]"))
                            raise anyio.get_cancelled_exc_class()()
                    await inner_send_chan.send(dict(data="[DONE]"))
                except anyio.get_cancelled_exc_class() as e:
                    print("disconnected")
                    with anyio.move_on_after(1, shield=True):
                        print(
                            f"Disconnected from client (via refresh/close) {request.client}"
                        )
                        raise e

        return EventSourceResponse(
            recv_chan,
            data_sender_callable=partial(event_publisher, send_chan),
        ) # type: ignore
    else:
        completion: llama_cpp.ChatCompletion = await run_in_threadpool(
            llama.create_chat_completion, **kwargs  # type: ignore
        )
        return completion


class ModelData(TypedDict):
    id: str
    object: Literal["model"]
    owned_by: str
    permissions: List[str]


class ModelList(TypedDict):
    object: Literal["list"]
    data: List[ModelData]


@router.get("/v1/models")
async def get_models(
    settings: Settings = Depends(get_settings),
) -> ModelList:
    assert llama is not None
    return {
        "object": "list",
        "data": [
            {
                "id": settings.model_alias
                if settings.model_alias is not None
                else llama.model_path,
                "object": "model",
                "owned_by": "me",
                "permissions": [],
            }
        ],
    }