Kye
10 months ago
parent
fec511c630
commit
0a48ca05cc
@ -1,40 +1,23 @@
|
||||
torch==2.1.1
|
||||
transformers
|
||||
pandas
|
||||
langchain==0.1.13
|
||||
mkdocs
|
||||
mkdocs-material
|
||||
mkdocs-glightbox
|
||||
torch>=2.1.1,<3.0
|
||||
transformers==4.39.0
|
||||
asyncio>=3.4.3,<4.0
|
||||
einops==0.7.0
|
||||
langchain-core==0.1.33
|
||||
langchain-community==0.0.29
|
||||
langsmith==0.1.17
|
||||
langchain-openai==0.0.5
|
||||
httpx==0.24.1
|
||||
Pillow==9.4.0
|
||||
datasets==2.14.5
|
||||
pydantic==2.6.4
|
||||
huggingface-hub
|
||||
requests_mock
|
||||
pypdf==4.0.1
|
||||
accelerate==0.22.0
|
||||
loguru==0.7.2
|
||||
optimum
|
||||
diffusers
|
||||
toml
|
||||
tiktoken==0.5.2
|
||||
colored
|
||||
addict
|
||||
langchain-experimental==0.0.55
|
||||
backoff==2.2.1
|
||||
toml
|
||||
pypdf==4.1.0
|
||||
httpx==0.24.1
|
||||
ratelimit==2.2.1
|
||||
loguru==0.7.2
|
||||
pydantic==2.6.4
|
||||
tenacity==8.2.3
|
||||
Pillow==10.2.0
|
||||
termcolor==2.2.0
|
||||
opencv-python==4.9.0.80
|
||||
timm
|
||||
torchvision==0.16.1
|
||||
rich==13.5.2
|
||||
mkdocs
|
||||
mkdocs-material
|
||||
anthropic==0.2.5
|
||||
mkdocs-glightbox
|
||||
pre-commit==3.6.2
|
||||
psutil
|
||||
black
|
||||
tenacity
|
||||
supervision
|
||||
sentry-sdk
|
||||
@ -0,0 +1,6 @@
|
||||
from swarms.schedulers.agent_process import (
|
||||
AgentProcess,
|
||||
AgentProcessQueue,
|
||||
)
|
||||
|
||||
__all__ = ["AgentProcess", "AgentProcessQueue"]
|
||||
@ -0,0 +1,103 @@
|
||||
from datetime import datetime
|
||||
|
||||
from pydantic import BaseModel
|
||||
|
||||
from swarms.structs.omni_agent_types import agents
|
||||
from swarms.utils.loguru_logger import logger
|
||||
|
||||
|
||||
class AgentProcess(BaseModel):
|
||||
agent_id: int
|
||||
agent_name: str
|
||||
prompt: str
|
||||
response: str = None
|
||||
time: callable = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
|
||||
priority: int = 0
|
||||
status: str = "Waiting"
|
||||
pid: int = None
|
||||
|
||||
def set_pid(self, pid: int):
|
||||
self.pid = pid
|
||||
|
||||
def get_pid(self):
|
||||
return self.pid
|
||||
|
||||
def set_time(self, time: callable):
|
||||
self.time = time
|
||||
|
||||
def get_time(self):
|
||||
return self.time
|
||||
|
||||
|
||||
class AgentProcessQueue:
|
||||
"""
|
||||
A class representing a queue of agent processes.
|
||||
|
||||
Attributes:
|
||||
MAX_PID (int): The maximum process ID.
|
||||
pid_pool (list): A list representing the availability of process IDs.
|
||||
agent_process_queue (list): A list representing the queue of agent processes.
|
||||
|
||||
Methods:
|
||||
add(agent_process): Adds an agent process to the queue.
|
||||
print(): Prints the details of all agent processes in the queue.
|
||||
|
||||
Private Methods:
|
||||
_get_available_pid(): Returns an available process ID from the pool.
|
||||
"""
|
||||
|
||||
def __init__(self, max_pid: int = 1024):
|
||||
self.MAX_PID = max_pid
|
||||
self.pid_pool = [False for i in range(self.MAX_PID)]
|
||||
self.agent_process_queue = (
|
||||
[]
|
||||
) # Currently use list to simulate queue
|
||||
|
||||
def add(self, agents: agents):
|
||||
"""
|
||||
Adds an agent process to the queue.
|
||||
|
||||
Args:
|
||||
agent_process (AgentProcess): The agent process to be added.
|
||||
|
||||
Returns:
|
||||
None
|
||||
"""
|
||||
for agent in agents:
|
||||
agent_process = AgentProcess(
|
||||
agent_id=agent.id,
|
||||
agent_name=agent.agent_name,
|
||||
prompt=agent.short_memory.return_history_as_string(),
|
||||
)
|
||||
pid = self._get_available_pid()
|
||||
if pid is None:
|
||||
logger.warning("No available PID")
|
||||
return
|
||||
agent_process.set_pid(pid)
|
||||
agent_process.set_status("Waiting")
|
||||
self.agent_process_queue.append(agent_process)
|
||||
|
||||
def print(self):
|
||||
"""
|
||||
Prints the details of all agent processes in the queue.
|
||||
|
||||
Returns:
|
||||
None
|
||||
"""
|
||||
for agent_process in self.agent_process_queue:
|
||||
logger.info(
|
||||
f"| Agent-process ID: {agent_process.get_pid()} |"
|
||||
f" Status: {agent_process.get_status()} |"
|
||||
)
|
||||
|
||||
def _get_available_pid(self):
|
||||
"""
|
||||
Returns an available process ID from the pool.
|
||||
|
||||
Returns:
|
||||
int or None: The available process ID, or None if no ID is available.
|
||||
"""
|
||||
for i, used in enumerate(self.pid_pool):
|
||||
if not used:
|
||||
return i
|
||||
return None
|
||||
@ -1,22 +0,0 @@
|
||||
# from swarms.tokenizers.anthropic_tokenizer import (
|
||||
# AnthropicTokenizer,
|
||||
# import_optional_dependency,
|
||||
# )
|
||||
from swarms.tokenizers.base_tokenizer import BaseTokenizer
|
||||
from swarms.tokenizers.openai_tokenizers import OpenAITokenizer
|
||||
from swarms.tokenizers.r_tokenizers import (
|
||||
HuggingFaceTokenizer,
|
||||
SentencePieceTokenizer,
|
||||
Tokenizer,
|
||||
)
|
||||
|
||||
|
||||
__all__ = [
|
||||
"SentencePieceTokenizer",
|
||||
"HuggingFaceTokenizer",
|
||||
"Tokenizer",
|
||||
"BaseTokenizer",
|
||||
"OpenAITokenizer",
|
||||
# "import_optional_dependency",
|
||||
# "AnthropicTokenizer",
|
||||
]
|
||||
@ -1,95 +0,0 @@
|
||||
from __future__ import annotations
|
||||
|
||||
from dataclasses import dataclass
|
||||
from importlib import import_module
|
||||
from types import ModuleType
|
||||
|
||||
from anthropic import Anthropic
|
||||
|
||||
from swarms.tokenizers.base_tokenizer import BaseTokenizer
|
||||
|
||||
INSTALL_MAPPING = {
|
||||
"huggingface_hub": "huggingface-hub",
|
||||
"pinecone": "pinecone-client",
|
||||
"opensearchpy": "opensearch-py",
|
||||
}
|
||||
|
||||
|
||||
def import_optional_dependency(name: str) -> ModuleType | None:
|
||||
"""Import an optional dependency.
|
||||
|
||||
If a dependency is missing, an ImportError with a nice message will be raised.
|
||||
|
||||
Args:
|
||||
name: The module name.
|
||||
Returns:
|
||||
The imported module, when found.
|
||||
None is returned when the package is not found and `errors` is False.
|
||||
"""
|
||||
|
||||
package_name = INSTALL_MAPPING.get(name)
|
||||
install_name = package_name if package_name is not None else name
|
||||
|
||||
msg = (
|
||||
f"Missing optional dependency: '{install_name}'. "
|
||||
f"Use poetry or pip to install '{install_name}'."
|
||||
)
|
||||
try:
|
||||
module = import_module(name)
|
||||
except ImportError:
|
||||
raise ImportError(msg)
|
||||
|
||||
return module
|
||||
|
||||
|
||||
@dataclass
|
||||
class AnthropicTokenizer(BaseTokenizer):
|
||||
"""
|
||||
Tokenizer class for Anthropic models.]
|
||||
"""
|
||||
|
||||
max_tokens: int = 500
|
||||
client: Anthropic = None
|
||||
model: str = "claude-2.1"
|
||||
|
||||
def __post_init__(self):
|
||||
self.DEFAULT_MODEL: str = "claude-2.1"
|
||||
self.MODEL_PREFIXES_TO_MAX_TOKENS: dict[str, int] = {
|
||||
"claude-2.1": 200000,
|
||||
"claude": 100000,
|
||||
}
|
||||
self.model = self.model # or self.DEFAULT_MODEL
|
||||
self.max_tokens = self.max_tokens or self.default_max_tokens()
|
||||
self.client = (
|
||||
self.client
|
||||
or import_optional_dependency("anthropic").Anthropic()
|
||||
)
|
||||
|
||||
def default_max_tokens(self) -> int:
|
||||
"""
|
||||
Returns the default maximum number of tokens based on the model prefix.
|
||||
"""
|
||||
tokens = next(
|
||||
v
|
||||
for k, v in self.MODEL_PREFIXES_TO_MAX_TOKENS.items()
|
||||
if self.model.startswith(k)
|
||||
)
|
||||
return tokens
|
||||
|
||||
def count_tokens(self, text: str | list) -> int:
|
||||
"""
|
||||
Counts the number of tokens in the given text.
|
||||
|
||||
Args:
|
||||
text: The input text.
|
||||
|
||||
Returns:
|
||||
The number of tokens in the text.
|
||||
|
||||
Raises:
|
||||
ValueError: If the input text is not a string.
|
||||
"""
|
||||
if isinstance(text, str):
|
||||
return self.client.count_tokens(text)
|
||||
else:
|
||||
raise ValueError("Text must be a string.")
|
||||
@ -1,55 +0,0 @@
|
||||
from __future__ import annotations
|
||||
|
||||
from abc import ABC, abstractmethod
|
||||
from dataclasses import dataclass, field
|
||||
|
||||
|
||||
@dataclass
|
||||
class BaseTokenizer(ABC):
|
||||
"""
|
||||
Base class for tokenizers.
|
||||
|
||||
Attributes:
|
||||
stop_sequences (List[str]): List of stop sequences.
|
||||
max_tokens (int): Maximum number of tokens.
|
||||
stop_token (str): Stop token.
|
||||
"""
|
||||
|
||||
max_tokens: int
|
||||
stop_token: str = "<|Response|>"
|
||||
|
||||
def __post_init__(self):
|
||||
self.stop_sequences: list[str] = field(
|
||||
default_factory=lambda: ["<|Response|>"],
|
||||
init=False,
|
||||
)
|
||||
|
||||
def count_tokens_left(self, text: str | list[dict]) -> int:
|
||||
"""
|
||||
Counts the number of tokens left based on the given text.
|
||||
|
||||
Args:
|
||||
text (Union[str, List[dict]]): The text to count tokens from.
|
||||
|
||||
Returns:
|
||||
int: The number of tokens left.
|
||||
"""
|
||||
diff = self.max_tokens - self.count_tokens(text)
|
||||
|
||||
if diff > 0:
|
||||
return diff
|
||||
else:
|
||||
return 0
|
||||
|
||||
@abstractmethod
|
||||
def count_tokens(self, text: str | list[dict]) -> int:
|
||||
"""
|
||||
Counts the number of tokens in the given text.
|
||||
|
||||
Args:
|
||||
text (Union[str, List[dict]]): The text to count tokens from.
|
||||
|
||||
Returns:
|
||||
int: The number of tokens.
|
||||
"""
|
||||
...
|
||||
@ -1,181 +0,0 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import logging
|
||||
from dataclasses import dataclass, field
|
||||
|
||||
import tiktoken
|
||||
from tiktoken import Encoding
|
||||
|
||||
from swarms.tokenizers.base_tokenizer import BaseTokenizer
|
||||
|
||||
|
||||
@dataclass
|
||||
class OpenAITokenizer(BaseTokenizer):
|
||||
"""
|
||||
A class representing an OpenAI tokenizer.
|
||||
|
||||
Attributes:
|
||||
- DEFAULT_OPENAI_GPT_3_COMPLETION_MODEL (str): The default OpenAI GPT-3 completion model.
|
||||
- DEFAULT_OPENAI_GPT_3_CHAT_MODEL (str): The default OpenAI GPT-3 chat model.
|
||||
- DEFAULT_OPENAI_GPT_4_MODEL (str): The default OpenAI GPT-4 model.
|
||||
- DEFAULT_ENCODING (str): The default encoding.
|
||||
- DEFAULT_MAX_TOKENS (int): The default maximum number of tokens.
|
||||
- TOKEN_OFFSET (int): The token offset.
|
||||
- MODEL_PREFIXES_TO_MAX_TOKENS (dict): A dictionary mapping model prefixes to maximum tokens.
|
||||
- EMBEDDING_MODELS (list): A list of embedding models.
|
||||
- model (str): The model name.
|
||||
|
||||
Methods:
|
||||
- __post_init__(): Initializes the OpenAITokenizer object.
|
||||
- encoding(): Returns the encoding for the model.
|
||||
- default_max_tokens(): Returns the default maximum number of tokens.
|
||||
- count_tokens(text, model): Counts the number of tokens in the given text.
|
||||
- len(text, model): Returns the length of the text in tokens.
|
||||
"""
|
||||
|
||||
model: str = "gpt-2"
|
||||
|
||||
def __post_init__(self):
|
||||
"""
|
||||
Initializes the OpenAITokenizer object.
|
||||
Sets the default maximum number of tokens.
|
||||
"""
|
||||
self.max_tokens: int = field(
|
||||
default_factory=self.default_max_tokens
|
||||
)
|
||||
|
||||
self.DEFAULT_OPENAI_GPT_3_COMPLETION_MODEL = (
|
||||
"text-davinci-003"
|
||||
)
|
||||
self.DEFAULT_OPENAI_GPT_3_CHAT_MODEL = "gpt-3.5-turbo"
|
||||
self.DEFAULT_OPENAI_GPT_4_MODEL = "gpt-4"
|
||||
self.DEFAULT_ENCODING = "cl100k_base"
|
||||
self.EFAULT_MAX_TOKENS = 2049
|
||||
self.TOKEN_OFFSET = 8
|
||||
|
||||
self.MODEL_PREFIXES_TO_MAX_TOKENS = {
|
||||
"gpt-4-1106": 128000,
|
||||
"gpt-4-32k": 32768,
|
||||
"gpt-4": 8192,
|
||||
"gpt-3.5-turbo-16k": 16384,
|
||||
"gpt-3.5-turbo": 4096,
|
||||
"gpt-35-turbo-16k": 16384,
|
||||
"gpt-35-turbo": 4096,
|
||||
"text-davinci-003": 4097,
|
||||
"text-davinci-002": 4097,
|
||||
"code-davinci-002": 8001,
|
||||
"text-embedding-ada-002": 8191,
|
||||
"text-embedding-ada-001": 2046,
|
||||
}
|
||||
|
||||
self.EMBEDDING_MODELS = [
|
||||
"text-embedding-ada-002",
|
||||
"text-embedding-ada-001",
|
||||
]
|
||||
|
||||
@property
|
||||
def encoding(self) -> Encoding:
|
||||
"""
|
||||
Returns the encoding for the model.
|
||||
If the model is not found, returns the default encoding.
|
||||
"""
|
||||
try:
|
||||
return tiktoken.encoding_for_model(self.model)
|
||||
except KeyError:
|
||||
return tiktoken.get_encoding(self.DEFAULT_ENCODING)
|
||||
|
||||
def default_max_tokens(self) -> int:
|
||||
"""
|
||||
Returns the default maximum number of tokens based on the model.
|
||||
"""
|
||||
tokens = next(
|
||||
v
|
||||
for k, v in self.MODEL_PREFIXES_TO_MAX_TOKENS.items()
|
||||
if self.model.startswith(k)
|
||||
)
|
||||
offset = (
|
||||
0
|
||||
if self.model in self.EMBEDDING_MODELS
|
||||
else self.TOKEN_OFFSET
|
||||
)
|
||||
|
||||
return (
|
||||
tokens if tokens else self.DEFAULT_MAX_TOKENS
|
||||
) - offset
|
||||
|
||||
def count_tokens(
|
||||
self, text: str | list[dict], model: str | None = None
|
||||
) -> int:
|
||||
"""
|
||||
Counts the number of tokens in the given text.
|
||||
If the text is a list of messages, counts the tokens for each message.
|
||||
If a model is provided, uses that model for encoding.
|
||||
"""
|
||||
if isinstance(text, list):
|
||||
model = model if model else self.model
|
||||
|
||||
try:
|
||||
encoding = tiktoken.encoding_for_model(model)
|
||||
except KeyError:
|
||||
logging.warning(
|
||||
"model not found. Using cl100k_base encoding."
|
||||
)
|
||||
encoding = tiktoken.get_encoding("cl100k_base")
|
||||
|
||||
if model in {
|
||||
"gpt-3.5-turbo-0613",
|
||||
"gpt-3.5-turbo-16k-0613",
|
||||
"gpt-4-0314",
|
||||
"gpt-4-32k-0314",
|
||||
"gpt-4-0613",
|
||||
"gpt-4-32k-0613",
|
||||
}:
|
||||
tokens_per_message = 3
|
||||
tokens_per_name = 1
|
||||
elif model == "gpt-3.5-turbo-0301":
|
||||
tokens_per_message = 4
|
||||
tokens_per_name = -1
|
||||
elif "gpt-3.5-turbo" in model or "gpt-35-turbo" in model:
|
||||
logging.info(
|
||||
"gpt-3.5-turbo may update over time. Returning"
|
||||
" num tokens assuming gpt-3.5-turbo-0613."
|
||||
)
|
||||
return self.count_tokens(
|
||||
text, model="gpt-3.5-turbo-0613"
|
||||
)
|
||||
elif "gpt-4" in model:
|
||||
logging.info(
|
||||
"gpt-4 may update over time. Returning num tokens"
|
||||
" assuming gpt-4-0613."
|
||||
)
|
||||
return self.count_tokens(text, model="gpt-4-0613")
|
||||
else:
|
||||
raise NotImplementedError(
|
||||
"token_count() is not implemented for model"
|
||||
f" {model}. See"
|
||||
" https://github.com/openai/openai-python/blob/main/chatml.md"
|
||||
" for information on how messages are converted"
|
||||
" to tokens."
|
||||
)
|
||||
|
||||
num_tokens = 0
|
||||
|
||||
for message in text:
|
||||
num_tokens += tokens_per_message
|
||||
for key, value in message.items():
|
||||
num_tokens += len(encoding.encode(value))
|
||||
if key == "name":
|
||||
num_tokens += tokens_per_name
|
||||
|
||||
num_tokens += 3
|
||||
|
||||
return num_tokens
|
||||
else:
|
||||
return len(self.encoding.encode(text))
|
||||
|
||||
def len(self, text: str | list[dict], model: str | None):
|
||||
"""
|
||||
Returns the length of the text in tokens.
|
||||
If a model is provided, uses that model for encoding.
|
||||
"""
|
||||
return self.count_tokens(text, model)
|
||||
@ -1,422 +0,0 @@
|
||||
# Copyright (c) OpenMMLab. All rights reserved.
|
||||
import json
|
||||
import os
|
||||
import os.path as osp
|
||||
from collections import deque
|
||||
from typing import List, Optional, Sequence, Union
|
||||
|
||||
import torch
|
||||
|
||||
from swarms.utils.get_logger import get_logger
|
||||
|
||||
|
||||
class SentencePieceTokenizer:
|
||||
"""Tokenizer of sentencepiece.
|
||||
|
||||
Args:
|
||||
model_file (str): the path of the tokenizer model
|
||||
"""
|
||||
|
||||
def __init__(self, model_file: str):
|
||||
from sentencepiece import SentencePieceProcessor
|
||||
|
||||
self.model = SentencePieceProcessor(model_file=model_file)
|
||||
self._prefix_space_tokens = None
|
||||
# for stop words
|
||||
self._maybe_decode_bytes: bool = None
|
||||
# TODO maybe lack a constant.py
|
||||
self._indexes_tokens_deque = deque(maxlen=10)
|
||||
self.max_indexes_num = 5
|
||||
self.logger = get_logger("lmdeploy")
|
||||
|
||||
@property
|
||||
def vocab_size(self):
|
||||
"""vocabulary size."""
|
||||
return self.model.vocab_size()
|
||||
|
||||
@property
|
||||
def bos_token_id(self):
|
||||
"""begine of the sentence token id."""
|
||||
return self.model.bos_id()
|
||||
|
||||
@property
|
||||
def eos_token_id(self):
|
||||
"""end of the sentence token id."""
|
||||
return self.model.eos_id()
|
||||
|
||||
@property
|
||||
def prefix_space_tokens(self):
|
||||
"""tokens without prefix space."""
|
||||
if self._prefix_space_tokens is None:
|
||||
vocab = self.model.IdToPiece(list(range(self.vocab_size)))
|
||||
self._prefix_space_tokens = {
|
||||
i
|
||||
for i, tok in enumerate(vocab)
|
||||
if tok.startswith("▁")
|
||||
}
|
||||
return self._prefix_space_tokens
|
||||
|
||||
def _maybe_add_prefix_space(self, tokens, decoded):
|
||||
"""maybe add prefix space for incremental decoding."""
|
||||
if (
|
||||
tokens
|
||||
and not decoded.startswith(" ")
|
||||
and tokens[0] in self.prefix_space_tokens
|
||||
):
|
||||
return " " + decoded
|
||||
else:
|
||||
return decoded
|
||||
|
||||
def indexes_containing_token(self, token: str):
|
||||
"""Return all the possible indexes, whose decoding output may contain
|
||||
the input token."""
|
||||
# traversing vocab is time consuming, can not be accelerated with
|
||||
# multi threads (computation) or multi process (can't pickle tokenizer)
|
||||
# so, we maintain latest 10 stop words and return directly if matched
|
||||
for _token, _indexes in self._indexes_tokens_deque:
|
||||
if token == _token:
|
||||
return _indexes
|
||||
if token == " ": # ' ' is special
|
||||
token = "▁"
|
||||
vocab = self.model.IdToPiece(list(range(self.vocab_size)))
|
||||
indexes = [i for i, voc in enumerate(vocab) if token in voc]
|
||||
if len(indexes) > self.max_indexes_num:
|
||||
indexes = self.encode(token, add_bos=False)[-1:]
|
||||
self.logger.warning(
|
||||
f"There are too many(>{self.max_indexes_num})"
|
||||
f" possible indexes may decoding {token}, we will use"
|
||||
f" {indexes} only"
|
||||
)
|
||||
self._indexes_tokens_deque.append((token, indexes))
|
||||
return indexes
|
||||
|
||||
def encode(self, s: str, add_bos: bool = True, **kwargs):
|
||||
"""Tokenize a prompt.
|
||||
|
||||
Args:
|
||||
s (str): a prompt
|
||||
Returns:
|
||||
list[int]: token ids
|
||||
"""
|
||||
return self.model.Encode(s, add_bos=add_bos, **kwargs)
|
||||
|
||||
def decode(self, t: Sequence[int], offset: Optional[int] = None):
|
||||
"""De-tokenize.
|
||||
|
||||
Args:
|
||||
t (List[int]): a list of token ids
|
||||
offset (int): for incrementally decoding. Default to None, which
|
||||
means not applied.
|
||||
Returns:
|
||||
str: text of decoding tokens
|
||||
"""
|
||||
if isinstance(t, torch.Tensor):
|
||||
t = t.tolist()
|
||||
t = t[offset:]
|
||||
out_string = self.model.Decode(t)
|
||||
if offset:
|
||||
out_string = self._maybe_add_prefix_space(t, out_string)
|
||||
return out_string
|
||||
|
||||
def __call__(self, s: Union[str, Sequence[str]]):
|
||||
"""Tokenize prompts.
|
||||
|
||||
Args:
|
||||
s (str): prompts
|
||||
Returns:
|
||||
list[int]: token ids
|
||||
"""
|
||||
import addict
|
||||
|
||||
add_bos = False
|
||||
add_eos = False
|
||||
|
||||
input_ids = self.model.Encode(
|
||||
s, add_bos=add_bos, add_eos=add_eos
|
||||
)
|
||||
return addict.Addict(input_ids=input_ids)
|
||||
|
||||
|
||||
class HuggingFaceTokenizer:
|
||||
"""Tokenizer of sentencepiece.
|
||||
|
||||
Args:
|
||||
model_dir (str): the directory of the tokenizer model
|
||||
"""
|
||||
|
||||
def __init__(self, model_dir: str):
|
||||
from transformers import AutoTokenizer
|
||||
|
||||
model_file = osp.join(model_dir, "tokenizer.model")
|
||||
backend_tokenizer_file = osp.join(model_dir, "tokenizer.json")
|
||||
model_file_exists = osp.exists(model_file)
|
||||
self.logger = get_logger("lmdeploy")
|
||||
if (
|
||||
not osp.exists(backend_tokenizer_file)
|
||||
and model_file_exists
|
||||
):
|
||||
self.logger.warning(
|
||||
"Can not find tokenizer.json. "
|
||||
"It may take long time to initialize the tokenizer."
|
||||
)
|
||||
self.model = AutoTokenizer.from_pretrained(
|
||||
model_dir, trust_remote_code=True
|
||||
)
|
||||
self._prefix_space_tokens = None
|
||||
# save tokenizer.json to reuse
|
||||
if (
|
||||
not osp.exists(backend_tokenizer_file)
|
||||
and model_file_exists
|
||||
):
|
||||
if hasattr(self.model, "backend_tokenizer"):
|
||||
if os.access(model_dir, os.W_OK):
|
||||
self.model.backend_tokenizer.save(
|
||||
backend_tokenizer_file
|
||||
)
|
||||
|
||||
if self.model.eos_token_id is None:
|
||||
generation_config_file = osp.join(
|
||||
model_dir, "generation_config.json"
|
||||
)
|
||||
if osp.exists(generation_config_file):
|
||||
with open(generation_config_file) as f:
|
||||
cfg = json.load(f)
|
||||
self.model.eos_token_id = cfg["eos_token_id"]
|
||||
elif hasattr(self.model, "eod_id"): # Qwen remote
|
||||
self.model.eos_token_id = self.model.eod_id
|
||||
|
||||
# for stop words
|
||||
self._maybe_decode_bytes: bool = None
|
||||
# TODO maybe lack a constant.py
|
||||
self._indexes_tokens_deque = deque(maxlen=10)
|
||||
self.max_indexes_num = 5
|
||||
self.token2id = {}
|
||||
|
||||
@property
|
||||
def vocab_size(self):
|
||||
"""vocabulary size."""
|
||||
return self.model.vocab_size
|
||||
|
||||
@property
|
||||
def bos_token_id(self):
|
||||
"""begine of the sentence token id."""
|
||||
return self.model.bos_token_id
|
||||
|
||||
@property
|
||||
def eos_token_id(self):
|
||||
"""end of the sentence token id."""
|
||||
return self.model.eos_token_id
|
||||
|
||||
@property
|
||||
def prefix_space_tokens(self):
|
||||
"""tokens without prefix space."""
|
||||
if self._prefix_space_tokens is None:
|
||||
vocab = self.model.convert_ids_to_tokens(
|
||||
list(range(self.vocab_size))
|
||||
)
|
||||
self._prefix_space_tokens = {
|
||||
i
|
||||
for i, tok in enumerate(vocab)
|
||||
if tok.startswith(
|
||||
"▁" if isinstance(tok, str) else b" "
|
||||
)
|
||||
}
|
||||
return self._prefix_space_tokens
|
||||
|
||||
def _maybe_add_prefix_space(
|
||||
self, tokens: List[int], decoded: str
|
||||
):
|
||||
"""maybe add prefix space for incremental decoding."""
|
||||
if (
|
||||
tokens
|
||||
and not decoded.startswith(" ")
|
||||
and tokens[0] in self.prefix_space_tokens
|
||||
):
|
||||
return " " + decoded
|
||||
else:
|
||||
return decoded
|
||||
|
||||
@property
|
||||
def maybe_decode_bytes(self):
|
||||
"""Check if self.model.convert_ids_to_tokens return not a str value."""
|
||||
if self._maybe_decode_bytes is None:
|
||||
self._maybe_decode_bytes = False
|
||||
vocab = self.model.convert_ids_to_tokens(
|
||||
list(range(self.vocab_size))
|
||||
)
|
||||
for tok in vocab:
|
||||
if not isinstance(tok, str):
|
||||
self._maybe_decode_bytes = True
|
||||
break
|
||||
return self._maybe_decode_bytes
|
||||
|
||||
def indexes_containing_token(self, token: str):
|
||||
"""Return all the possible indexes, whose decoding output may contain
|
||||
the input token."""
|
||||
# traversing vocab is time consuming, can not be accelerated with
|
||||
# multi threads (computation) or multi process (can't pickle tokenizer)
|
||||
# so, we maintain latest 10 stop words and return directly if matched
|
||||
for _token, _indexes in self._indexes_tokens_deque:
|
||||
if token == _token:
|
||||
return _indexes
|
||||
|
||||
if self.token2id == {}:
|
||||
# decode is slower than convert_ids_to_tokens
|
||||
if self.maybe_decode_bytes:
|
||||
self.token2id = {
|
||||
self.model.decode(i): i
|
||||
for i in range(self.vocab_size)
|
||||
}
|
||||
else:
|
||||
self.token2id = {
|
||||
self.model.convert_ids_to_tokens(i): i
|
||||
for i in range(self.vocab_size)
|
||||
}
|
||||
if token == " ": # ' ' is special
|
||||
token = "▁"
|
||||
indexes = [
|
||||
i
|
||||
for _token, i in self.token2id.items()
|
||||
if token in _token
|
||||
]
|
||||
if len(indexes) > self.max_indexes_num:
|
||||
indexes = self.encode(token, add_bos=False)[-1:]
|
||||
self.logger.warning(
|
||||
f"There are too many(>{self.max_indexes_num})"
|
||||
f" possible indexes may decoding {token}, we will use"
|
||||
f" {indexes} only"
|
||||
)
|
||||
self._indexes_tokens_deque.append((token, indexes))
|
||||
return indexes
|
||||
|
||||
def encode(self, s: str, add_bos: bool = True, **kwargs):
|
||||
"""Tokenize a prompt.
|
||||
|
||||
Args:
|
||||
s (str): a prompt
|
||||
Returns:
|
||||
list[int]: token ids
|
||||
"""
|
||||
encoded = self.model.encode(s, **kwargs)
|
||||
if not add_bos:
|
||||
# in the middle of a session
|
||||
if encoded and encoded[0] == self.bos_token_id:
|
||||
encoded = encoded[1:]
|
||||
return encoded
|
||||
|
||||
def decode(self, t: Sequence[int], offset: Optional[int] = None):
|
||||
"""De-tokenize.
|
||||
|
||||
Args:
|
||||
t (List[int]): a list of token ids
|
||||
offset (int): for incrementally decoding. Default to None, which
|
||||
means not applied.
|
||||
Returns:
|
||||
str: text of decoding tokens
|
||||
"""
|
||||
skip_special_tokens = True
|
||||
t = t[offset:]
|
||||
out_string = self.model.decode(
|
||||
t, skip_special_tokens=skip_special_tokens
|
||||
)
|
||||
if offset:
|
||||
out_string = self._maybe_add_prefix_space(t, out_string)
|
||||
return out_string
|
||||
|
||||
def __call__(self, s: Union[str, Sequence[str]]):
|
||||
"""Tokenize prompts.
|
||||
|
||||
Args:
|
||||
s (str): prompts
|
||||
Returns:
|
||||
list[int]: token ids
|
||||
"""
|
||||
add_special_tokens = False
|
||||
return self.model(s, add_special_tokens=add_special_tokens)
|
||||
|
||||
|
||||
class Tokenizer:
|
||||
"""Tokenize prompts or de-tokenize tokens into texts.
|
||||
|
||||
Args:
|
||||
model_file (str): the path of the tokenizer model
|
||||
"""
|
||||
|
||||
def __init__(self, model_file: str):
|
||||
if model_file.endswith(".model"):
|
||||
model_folder = osp.split(model_file)[0]
|
||||
else:
|
||||
model_folder = model_file
|
||||
model_file = osp.join(model_folder, "tokenizer.model")
|
||||
tokenizer_config_file = osp.join(
|
||||
model_folder, "tokenizer_config.json"
|
||||
)
|
||||
|
||||
model_file_exists = osp.exists(model_file)
|
||||
config_exists = osp.exists(tokenizer_config_file)
|
||||
use_hf_model = config_exists or not model_file_exists
|
||||
self.logger = get_logger("lmdeploy")
|
||||
if not use_hf_model:
|
||||
self.model = SentencePieceTokenizer(model_file)
|
||||
else:
|
||||
self.model = HuggingFaceTokenizer(model_folder)
|
||||
|
||||
@property
|
||||
def vocab_size(self):
|
||||
"""vocabulary size."""
|
||||
return self.model.vocab_size
|
||||
|
||||
@property
|
||||
def bos_token_id(self):
|
||||
"""begine of the sentence token id."""
|
||||
return self.model.bos_token_id
|
||||
|
||||
@property
|
||||
def eos_token_id(self):
|
||||
"""end of the sentence token id."""
|
||||
return self.model.eos_token_id
|
||||
|
||||
def encode(self, s: str, add_bos: bool = True, **kwargs):
|
||||
"""Tokenize a prompt.
|
||||
|
||||
Args:
|
||||
s (str): a prompt
|
||||
Returns:
|
||||
list[int]: token ids
|
||||
"""
|
||||
return self.model.encode(s, add_bos, **kwargs)
|
||||
|
||||
def decode(self, t: Sequence[int], offset: Optional[int] = None):
|
||||
"""De-tokenize.
|
||||
|
||||
Args:
|
||||
t (List[int]): a list of token ids
|
||||
offset (int): for incrementally decoding. Default to None, which
|
||||
means not applied.
|
||||
Returns:
|
||||
str: text of decoding tokens
|
||||
"""
|
||||
return self.model.decode(t, offset)
|
||||
|
||||
def __call__(self, s: Union[str, Sequence[str]]):
|
||||
"""Tokenize prompts.
|
||||
|
||||
Args:
|
||||
s (str): prompts
|
||||
Returns:
|
||||
list[int]: token ids
|
||||
"""
|
||||
return self.model(s)
|
||||
|
||||
def indexes_containing_token(self, token):
|
||||
"""Return all the possible indexes, whose decoding output may contain
|
||||
the input token."""
|
||||
encoded = self.encode(token, add_bos=False)
|
||||
if len(encoded) > 1:
|
||||
self.logger.warning(
|
||||
f"The token {token}, its length of indexes"
|
||||
f" {encoded} is over than 1. Currently, it can not be"
|
||||
" used as stop words"
|
||||
)
|
||||
return []
|
||||
return self.model.indexes_containing_token(token)
|
||||
Loading…
Reference in new issue