inference/worker/utils.py - main - githubmirror/Open-Assistant

inference/worker/utils.py
300 строк10 KB
Oliver Stanley
Fixes for worker prompt truncation in ChatML case (#3673)
29 авг 2023, 20:38
29 авг 2023, 20:38adcf8dc
import collections
import random
import threading
import time
from typing import Iterable, Literal

import interface
import lorem
import pydantic
import requests
import sseclient
import transformers
import websocket
from loguru import logger
from oasst_shared.schemas import inference
from settings import settings

shared_tokenizer_lock = threading.Lock()


if settings.model_prompt_format == "chatml":
    special_tokens = {
        "prompter": "<|im_start|>user\n",
        "assistant": "<|im_start|>assistant\n",
        "system": "<|im_start|>system\n",
        "end": "<|im_end|>\n",
    }
else:
    special_tokens = {
        "prompter": "<|prompter|>",
        "assistant": "<|assistant|>",
        "system": "<|system|>",
        "end": "",
    }


class TokenBuffer:
    """
    A buffer for storing and managing tokens based on various conditions including stop sequences.

    The TokenBuffer class accumulates tokens while keeping track of the length and manages the tokens based on the stop
    sequences provided during initialization. Tokens can be added to the buffer and later on iterated upon finishing
    depending on the reason.
    """

    def __init__(self, stop_sequences: list[str]) -> None:
        self.stop_sequences = stop_sequences
        self.longest_stop_len = max((len(stop) for stop in stop_sequences), default=1)
        self.tokens = collections.deque()
        self.token_lens = collections.deque()
        self.total_len = 0

    def add(self, token: interface.Token):
        self.tokens.append(token)
        self.token_lens.append(len(token))
        self.total_len += len(token)
        while True:
            if not self.tokens:
                break
            head_len = self.token_lens[0]
            if self.total_len - head_len >= self.longest_stop_len:
                token = self.tokens.popleft()
                self.token_lens.popleft()
                self.total_len -= head_len
                yield token
            else:
                break

    def finish(self, reason: Literal["length", "eos_token", "stop_sequence"]) -> Iterable[interface.Token]:
        if reason == "stop_sequence":
            end_sequence = ""
            end_tokens = []
            while self.tokens:
                token = self.tokens.pop()
                end_tokens.append(token)
                end_sequence = token.text + end_sequence
                if end_sequence in self.stop_sequences:
                    break
            else:
                self.tokens.extend(reversed(end_tokens))
            yield from self.tokens
        elif reason == "eos_token":
            if self.tokens:
                self.tokens.pop()
            yield from self.tokens
        else:
            yield from self.tokens


def get_max_input_length(worker_config: inference.WorkerConfig, plugin_used: bool):
    """Get the maximum possible input length based on the worker config and whether a plugin is in use."""
    max_input_length = worker_config.model_config.max_input_length
    if plugin_used:
        max_input_length = max_input_length - 1
    return max_input_length


def get_tokens_until(tokens: list[int], target: list[int]) -> list[int]:
    if len(target) == 1:
        return tokens[: tokens.index(target[0])]

    for i in range(len(tokens) - len(target)):
        if tokens[i : i + len(target)] == target:
            break
    return tokens[:i]


def truncate_prompt(
    tokenizer: transformers.PreTrainedTokenizer,
    worker_config: inference.WorkerConfig,
    parameters: interface.GenerateStreamParameters,
    prompt: str,
    plugin_used: bool,
):
    """
    Truncate a prompt to ensure it does not exceed the maximum input length. Regardless of truncation, the system
    prompt is always retained if it is present. If truncation removes the final prompter prefix, a new one is added.

    The stream generation parameters are also updated with a maximum new tokens value which will not cause the total
    length to exceed the maximum specified in the worker's model config.
    """
    with shared_tokenizer_lock:
        ids = tokenizer.encode(prompt)
        # list of int IDs
        prompter_prefix_ids = tokenizer.encode(special_tokens["prompter"])

    system_prompt: str | None = None
    system_tokens: list[int] | None = None
    if prompt.startswith(special_tokens["system"]):
        system_prompt = prompt[: prompt.index(special_tokens["prompter"])]
        system_tokens = get_tokens_until(ids, prompter_prefix_ids)

    max_input_length = get_max_input_length(worker_config, plugin_used)

    if len(ids) > max_input_length:
        logger.debug(f"Prompt too long, left-truncating to {max_input_length} tokens")

        num_system_tokens = len(system_tokens) if system_tokens else 0
        # Maximum token allowed for the conversation, ex system prompt
        # We incorporate a buffer to allow for final inference tokenization differing from ours
        # This is a slightly hacky workaround and it would be better to find a cleaner solution
        max_conversation_length = max_input_length - num_system_tokens - int(0.01 * max_input_length)
        ids = ids[-(max_conversation_length - 1) :]

        with shared_tokenizer_lock:
            prompt = tokenizer.decode(ids)

            if special_tokens["prompter"] not in prompt:
                prompt = special_tokens["prompter"] + prompt
                ids = tokenizer.encode(special_tokens["prompter"]) + ids

            if system_tokens:
                prompt = system_prompt + prompt
                ids = system_tokens + ids

    max_total_tokens = worker_config.model_config.max_total_length
    input_length = len(ids)
    spare = max_total_tokens - input_length - 1

    if not parameters.max_new_tokens:
        parameters.max_new_tokens = spare
    elif parameters.max_new_tokens > spare:
        logger.debug(f"Max new tokens too high, reducing to {spare}")
        parameters.max_new_tokens = spare

    return prompt


def wait_for_inference_server(http: "HttpClient", timeout: int = 600):
    """Wait for the "health" endpoint of the inference server to return status 200."""
    time_limit = time.time() + timeout
    while True:
        try:
            response = http.get("/health")
            response.raise_for_status()
        except (requests.HTTPError, requests.ConnectionError):
            if time.time() > time_limit:
                raise
            sleep_duration = random.uniform(0, 10)
            logger.warning(f"Inference server not ready. Retrying in {sleep_duration:.2f} seconds")
            time.sleep(sleep_duration)
        else:
            logger.info("Inference server is ready")
            break


def text_to_events(
    text: str, seed: int | None = None, pause: float = 0.0
) -> Iterable[interface.GenerateStreamResponse]:
    """
    Iterate over stream generation "events" derived from the given text, where each word in the text is treated as a
    generated "token".
    """
    tokens = text.split()
    for token in tokens[:-1]:
        yield interface.GenerateStreamResponse(
            token=interface.Token(
                text=token + " ",
                logprob=0.1,
                id=0,
            ),
        )
        if pause > 0:
            time.sleep(pause)
    yield interface.GenerateStreamResponse(
        token=interface.Token(
            text=tokens[-1],
            logprob=0.1,
            id=0,
        ),
        generated_text=text,
        details=interface.StreamDetails(
            finish_reason="length",
            generated_tokens=len(tokens),
            seed=seed,
        ),
    )


def lorem_events(seed):
    sentence = lorem.paragraph()
    yield from text_to_events(sentence, seed=seed, pause=0.2)


ws_lock = threading.Lock()


def send_response(
    ws: websocket.WebSocket,
    response: inference.WorkerResponse | inference.WorkerInfo,
):
    msg = response.json()
    with ws_lock:
        ws.send(msg)


class HttpClient(pydantic.BaseModel):
    """Basic HTTP client built around `requests`. Supports simple authentication."""

    base_url: str
    basic_auth_username: str | None = None
    basic_auth_password: str | None = None
    bearer_token: str | None = None

    @property
    def auth(self):
        if self.basic_auth_username and self.basic_auth_password:
            return self.basic_auth_username, self.basic_auth_password
        else:
            return None

    def _maybe_add_bearer_token(self, headers: dict[str, str] | None):
        if self.bearer_token:
            if headers is None:
                headers = {}
            headers["Authorization"] = f"Bearer {self.bearer_token}"
        return headers

    def get(self, path: str, **kwargs):
        kwargs["headers"] = self._maybe_add_bearer_token(kwargs.get("headers"))
        return requests.get(self.base_url + path, auth=self.auth, **kwargs)

    def post(self, path: str, **kwargs):
        kwargs["headers"] = self._maybe_add_bearer_token(kwargs.get("headers"))
        return requests.post(self.base_url + path, auth=self.auth, **kwargs)


def get_inference_server_stream_events(
    request: interface.GenerateStreamRequest,
) -> Iterable[interface.GenerateStreamResponse]:
    """Query the model inference server specified in the worker settings and stream the generation events."""
    http = HttpClient(
        base_url=settings.inference_server_url,
        basic_auth_username=settings.basic_auth_username,
        basic_auth_password=settings.basic_auth_password,
        bearer_token=settings.bearer_token,
    )
    response = http.post(
        settings.inference_server_route,
        json=request.dict(),
        stream=True,
        headers={"Accept": "text/event-stream"},
    )
    try:
        response.raise_for_status()
    except requests.HTTPError:
        logger.exception("Failed to get response from inference server")
        logger.error(f"Response: {response.text}")
        raise

    client = sseclient.SSEClient(response)
    for event in client.events():
        if event.event == "error":
            logger.error(f"Error from inference server: {event.data}")
            yield interface.GenerateStreamResponse(error=event.data)
            raise RuntimeError(f"Error from inference server: {event.data}")
        if event.event == "ping":
            continue
        stream_response = interface.GenerateStreamResponse.parse_raw(event.data)
        yield stream_response