llmlingua-2 / llmlingua /prompt_compressor.py
qianhuiwu's picture
Initial commit.
cdba444
# Copyright (c) 2023 Microsoft
# Licensed under The MIT License [see LICENSE for details]
import bisect
import re
from collections import defaultdict
from typing import List
import numpy as np
import torch
import nltk
import tiktoken
from transformers import (
AutoConfig,
AutoModelForCausalLM,
AutoModelForTokenClassification,
AutoTokenizer,
)
import torch.nn.functional as F
import string
import copy
from torch.utils.data import DataLoader
from .utils import TokenClfDataset, seed_everything, is_begin_of_new_word, replace_added_token, get_pure_token
class PromptCompressor:
"""
PromptCompressor is designed for compressing prompts based on a given language model.
This class initializes with the language model and its configuration, preparing it for prompt compression tasks.
The PromptCompressor class is versatile and can be adapted for various models and specific requirements in prompt processing.
Users can specify different model names and configurations as needed for their particular use case.The architecture is
based on the paper "LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models". Jiang, Huiqiang, Qianhui Wu,
Chin-Yew Lin, Yuqing Yang, and Lili Qiu. "Llmlingua: Compressing prompts for accelerated inference of large language models."
arXiv preprint arXiv:2310.05736 (2023).
Args:
model_name (str, optional): The name of the language model to be loaded. Default is "NousResearch/Llama-2-7b-hf".
device_map (str, optional): The device to load the model onto, e.g., "cuda" for GPU. Default is "cuda".
model_config (dict, optional): A dictionary containing the configuration parameters for the model. Default is an empty dictionary.
open_api_config (dict, optional): A dictionary containing configuration for openai APIs that may be used in conjunction with the model. Default is an empty dictionary.
use_llmlingua2 (bool, optional): Whether to use llmlingua-2 compressor based on the paper
"LLMLingua-2: Context-Aware Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compression".
Zhuoshi Pan, Qianhui Wu, Huiqiang Jiang, Menglin Xia, Xufang Luo, Jue Zhang, Qingwei Lin, Victor Ruhle, Yuqing Yang, Chin-Yew Lin, H. Vicky Zhao, Lili Qiu, Dongmei Zhang.
"LLMLingua-2: Context-Aware Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compression". arXiv preprint arXiv:,
Default is True.
llmlingua2_config (dict, optional): A dictionary containing the configuration parameters for llmlingua-2. Default is
{
"max_batch_size": 50,
"max_force_token": 100, # max number of the tokens which will be forcely preserved
}
Example:
>>> compress_method = PromptCompressor(model_name="xxx/llmlingua-2-xlm-roberta-large-meetingbank", use_llmlingua2=True, )
>>> context = ["This is the first context sentence.", "Here is another context sentence."]
>>> result = compress_method.compress_prompt(context, use_context_level_filter=True, target_token=5)
>>> print(result["compressed_prompt"])
# This will print the compressed version of the context.
Note:
The `PromptCompressor` class requires the Hugging Face Transformers library and an appropriate environment to load and run the models.
"""
def __init__(
self,
model_name: str = "NousResearch/Llama-2-7b-hf",
device_map: str = "cuda",
model_config: dict = {},
open_api_config: dict = {},
use_llmlingua2: bool = True,
llmlingua2_config: dict = {},
):
self.model_name = model_name
self.use_llmlingua2 = use_llmlingua2
self.retrieval_model = None
self.retrieval_model_name = None
self.open_api_config = open_api_config
self.cache_bos_num = 10
self.prefix_bos_num = 100
self.oai_tokenizer = tiktoken.encoding_for_model("gpt-3.5-turbo")
self.load_model(model_name, device_map, model_config)
if use_llmlingua2:
self.init_llmlingua2(**llmlingua2_config)
def init_llmlingua2(
self,
max_batch_size: int = 50,
max_force_token: int = 100,
):
seed_everything(42)
self.max_batch_size = max_batch_size
self.max_seq_len = 512
self.max_force_token = max_force_token
self.special_tokens = set(self.tokenizer.special_tokens_map.values())
self.added_tokens = [f"[NEW{i}]" for i in range(max_force_token)]
self.tokenizer.add_special_tokens(
{"additional_special_tokens": self.added_tokens}
)
self.model.resize_token_embeddings(len(self.tokenizer))
def load_model(
self, model_name: str, device_map: str = "cuda", model_config: dict = {}
):
trust_remote_code = model_config.get("trust_remote_code", True)
if "trust_remote_code" not in model_config:
model_config["trust_remote_code"] = trust_remote_code
config = AutoConfig.from_pretrained(model_name, **model_config)
tokenizer = AutoTokenizer.from_pretrained(model_name, **model_config)
if model_config.get("pad_to_left", True):
tokenizer.padding_side = "left"
tokenizer.pad_token_id = (
config.pad_token_id if config.pad_token_id else tokenizer.eos_token_id
)
MODEL_CLASS = (
AutoModelForTokenClassification
if any("ForTokenClassification" in ar for ar in config.architectures)
else AutoModelForCausalLM
)
self.device = (
device_map
if any(key in device_map for key in ["cuda", "cpu", "mps"])
else "cuda"
)
if "cuda" in device_map or "cpu" in device_map:
model = MODEL_CLASS.from_pretrained(
model_name,
torch_dtype=model_config.get(
"torch_dtype", "auto" if device_map == "cuda" else torch.float32
),
device_map=device_map,
config=config,
ignore_mismatched_sizes=True,
**model_config,
)
else:
model = MODEL_CLASS.from_pretrained(
model_name,
device_map=device_map,
torch_dtype=model_config.get("torch_dtype", "auto"),
pad_token_id=tokenizer.pad_token_id,
**model_config,
)
self.tokenizer = tokenizer
self.model = model
self.context_idxs = []
self.max_position_embeddings = config.max_position_embeddings
def get_ppl(
self,
text: str,
granularity: str = "sentence",
input_ids=None,
attention_mask=None,
past_key_values=None,
return_kv=False,
end=None,
condition_mode: str = "none",
condition_pos_id: int = 0,
):
if input_ids is None:
tokenized_text = self.tokenizer(text, return_tensors="pt")
input_ids = tokenized_text["input_ids"].to(self.device)
attention_mask = tokenized_text["attention_mask"].to(self.device)
if past_key_values is not None:
past_length = past_key_values[0][0].shape[2]
else:
past_length = 0
if end is None:
end = input_ids.shape[1]
end = min(end, past_length + self.max_position_embeddings)
with torch.no_grad():
response = self.model(
input_ids[:, past_length:end],
attention_mask=attention_mask[:, :end],
past_key_values=past_key_values,
use_cache=True,
)
past_key_values = response.past_key_values
shift_logits = response.logits[..., :-1, :].contiguous()
shift_labels = input_ids[..., past_length + 1 : end].contiguous()
# Flatten the tokens
active = (attention_mask[:, past_length:end] == 1)[..., :-1].view(-1)
active_logits = shift_logits.view(-1, shift_logits.size(-1))[active]
active_labels = shift_labels.view(-1)[active]
loss_fct = torch.nn.CrossEntropyLoss(reduction="none")
loss = loss_fct(active_logits, active_labels)
if condition_mode == "before":
loss = loss[:condition_pos_id]
elif condition_mode == "after":
loss = loss[condition_pos_id:]
res = loss.mean() if granularity == "sentence" else loss
return (res, past_key_values) if return_kv else res
def __call__(self, *args, **kwargs):
return self.compress_prompt(*args, **kwargs)
def structured_compress_prompt(
self,
context: List[str],
instruction: str = "",
question: str = "",
rate: float = 0.5,
target_token: float = -1,
iterative_size: int = 200,
force_context_ids: List[int] = None,
force_context_number: int = None,
use_sentence_level_filter: bool = False,
use_context_level_filter: bool = True,
use_token_level_filter: bool = True,
keep_split: bool = False,
keep_first_sentence: int = 0,
keep_last_sentence: int = 0,
keep_sentence_number: int = 0,
high_priority_bonus: int = 100,
context_budget: str = "+100",
token_budget_ratio: float = 1.4,
condition_in_question: str = "none",
reorder_context: str = "original",
dynamic_context_compression_ratio: float = 0.0,
condition_compare: bool = False,
add_instruction: bool = False,
rank_method: str = "llmlingua",
concate_question: bool = True,
):
"""
Compresses the given prompt context based on a specified structure.
Each element of context should be segmented using one or more non-nested '<llmlingua></llmlingua>' tags.
Each '<llmlingua>' tag can include optional parameters 'rate' and 'compress' (e.g., '<llmlingua, rate=0.3, compress=True>'),
indicating the compression rate for that segment. Default values are 'rate=rate' and 'compress=True'.
When 'compress' is set to False, it overrides the 'rate' parameter, resulting in no compression for that segment.
Args:
context (List[str]): List of context strings divided by '<llmlingua></llmlingua>' tags with optional compression settings.
instruction (str, optional): Additional instruction text to be included in the prompt. Default is an empty string.
question (str, optional): A specific question that the prompt is addressing. Default is an empty string.
rate (float, optional): The compression rate is defined the same as in paper "Language Modeling Is Compression".
Delétang, Grégoire, Anian Ruoss, Paul-Ambroise Duquenne, Elliot Catt, Tim Genewein, Christopher Mattern,
Jordi Grau-Moya et al. "Language modeling is compression." arXiv preprint arXiv:2309.10668 (2023):
.. math::\text{Compression Rate} = \frac{\text{Compressed Size}}{\text{Raw Size}}
Default is 0.5. The actual compression rate is generally lower than the specified target, but there can be
fluctuations due to differences in tokenizers. If specified, it should be a float less than or equal
to 1.0, representing the target compression rate. ``rate``, is applicable only within the context-level filter
and the sentence-level filter. In the token-level filter, the rate for each segment overrides the global rate.
However, for segments where no specific rate is defined, the global rate serves as the default value. The final
compression rate of the entire text is a composite result of multiple compression rates applied across different sections.
target_token (float, optional): The global maximum number of tokens to be achieved. Default is -1, indicating no
specific target. The actual number of tokens after compression should generally be less than the specified target_token,
but there can be fluctuations due to differences in tokenizers. If specified, compression will be based on the target_token as
the sole criterion, overriding the ``rate``. ``target_token``, is applicable only within the context-level
filter and the sentence-level filter. In the token-level filter, the rate for each segment overrides the global target token.
However, for segments where no specific rate is defined, the global rate calculated from global target token serves
as the default value. The final target token of the entire text is a composite result of multiple compression rates
applied across different sections.
iterative_size (int, optional): The number of tokens to consider in each iteration of compression. Default is 200.
force_context_ids (List[int], optional): List of specific context IDs to always include in the compressed result. Default is None.
force_context_number (int, optional): The number of context sections to forcibly include. Default is None.
use_sentence_level_filter (bool, optional): Whether to apply sentence-level filtering in compression. Default is False.
use_context_level_filter (bool, optional): Whether to apply context-level filtering in compression. Default is True.
use_token_level_filter (bool, optional): Whether to apply token-level filtering in compression. Default is True.
keep_split (bool, optional): Whether to preserve the original separators without compression. Default is False.
keep_first_sentence (int, optional): Number of sentences to forcibly preserve from the start of the context. Default is 0.
keep_last_sentence (int, optional): Number of sentences to forcibly preserve from the end of the context. Default is 0.
keep_sentence_number (int, optional): Total number of sentences to forcibly preserve in the compression. Default is 0.
high_priority_bonus (int, optional): Bonus score for high-priority sentences to influence their likelihood of being retained. Default is 100.
context_budget (str, optional): Token budget for the context-level filtering, expressed as a string to indicate flexibility. Default is "+100".
token_budget_ratio (float, optional): Ratio to adjust token budget during sentence-level filtering. Default is 1.4.
condition_in_question (str, optional): Specific condition to apply to question in the context. Default is "none".
reorder_context (str, optional): Strategy for reordering context in the compressed result. Default is "original".
dynamic_context_compression_ratio (float, optional): Ratio for dynamically adjusting context compression. Default is 0.0.
condition_compare (bool, optional): Whether to enable condition comparison during token-level compression. Default is False.
add_instruction (bool, optional): Whether to add the instruction to the prompt prefix. Default is False.
rank_method (str, optional): Method used for ranking elements during compression. Default is "llmlingua".
concate_question (bool, optional): Whether to concatenate the question to the compressed prompt. Default is True.
Returns:
dict: A dictionary containing:
- "compressed_prompt" (str): The resulting compressed prompt.
- "origin_tokens" (int): The original number of tokens in the input.
- "compressed_tokens" (int): The number of tokens in the compressed output.
- "ratio" (str): The compression ratio achieved, calculated as the original token number divided by the token number after compression.
- "rate" (str): The compression rate achieved, in a human-readable format.
- "saving" (str): Estimated savings in GPT-4 token usage.
"""
if not context:
context = [" "]
if isinstance(context, str):
context = [context]
context = [
self.tokenizer.decode(self.tokenizer(c, add_special_tokens=False).input_ids)
for c in context
]
context_tokens_length = [self.get_token_length(c) for c in context]
instruction_tokens_length, question_tokens_length = self.get_token_length(
instruction
), self.get_token_length(question)
if target_token == -1:
target_token = (
(
instruction_tokens_length
+ question_tokens_length
+ sum(context_tokens_length)
)
* rate
- instruction_tokens_length
- (question_tokens_length if concate_question else 0)
)
else:
rate = target_token / sum(context_tokens_length)
(
context,
context_segs,
context_segs_rate,
context_segs_compress,
) = self.segment_structured_context(context, rate)
return self.compress_prompt(
context,
instruction,
question,
rate,
target_token,
iterative_size,
force_context_ids,
force_context_number,
use_sentence_level_filter,
use_context_level_filter,
use_token_level_filter,
keep_split,
keep_first_sentence,
keep_last_sentence,
keep_sentence_number,
high_priority_bonus,
context_budget,
token_budget_ratio,
condition_in_question,
reorder_context,
dynamic_context_compression_ratio,
condition_compare,
add_instruction,
rank_method,
concate_question,
context_segs=context_segs,
context_segs_rate=context_segs_rate,
context_segs_compress=context_segs_compress,
)
def compress_prompt(
self,
context: List[str],
instruction: str = "",
question: str = "",
rate: float = 0.5,
target_token: float = -1,
iterative_size: int = 200,
force_context_ids: List[int] = None,
force_context_number: int = None,
use_sentence_level_filter: bool = False,
use_context_level_filter: bool = True,
use_token_level_filter: bool = True,
keep_split: bool = False,
keep_first_sentence: int = 0,
keep_last_sentence: int = 0,
keep_sentence_number: int = 0,
high_priority_bonus: int = 100,
context_budget: str = "+100",
token_budget_ratio: float = 1.4,
condition_in_question: str = "none",
reorder_context: str = "original",
dynamic_context_compression_ratio: float = 0.0,
condition_compare: bool = False,
add_instruction: bool = False,
rank_method: str = "llmlingua",
concate_question: bool = True,
context_segs: List[str] = None,
context_segs_rate: List[float] = None,
context_segs_compress: List[bool] = None,
target_context: int = -1,
context_level_rate: float = 1.0,
context_level_target_token: int = -1,
return_word_label: bool = False,
word_sep: str = "\t\t|\t\t",
label_sep: str = " ",
token_to_word: str = "mean",
force_tokens: List[str] = [],
force_reserve_digit: bool = False,
drop_consecutive: bool = False,
chunk_end_tokens: List[str] = [".", "\n"],
):
"""
Compresses the given context.
Args:
context (List[str]): List of context strings that form the basis of the prompt.
instruction (str, optional): Additional instruction text to be included in the prompt. Default is an empty string.
question (str, optional): A specific question that the prompt is addressing. Default is an empty string.
rate (float, optional): The maximum compression rate target to be achieved. The compression rate is defined
the same as in paper "Language Modeling Is Compression". Delétang, Grégoire, Anian Ruoss, Paul-Ambroise Duquenne,
Elliot Catt, Tim Genewein, Christopher Mattern, Jordi Grau-Moya et al. "Language modeling is compression."
arXiv preprint arXiv:2309.10668 (2023):
.. math::\text{Compression Rate} = \frac{\text{Compressed Size}}{\text{Raw Size}}
Default is 0.5. The actual compression rate is generally lower than the specified target, but there can be
fluctuations due to differences in tokenizers. If specified, it should be a float less than or equal
to 1.0, representing the target compression rate.
target_token (float, optional): The maximum number of tokens to be achieved. Default is -1, indicating no specific target.
The actual number of tokens after compression should generally be less than the specified target_token, but there can
be fluctuations due to differences in tokenizers. If specified, compression will be based on the target_token as
the sole criterion, overriding the ``rate``.
iterative_size (int, optional): The number of tokens to consider in each iteration of compression. Default is 200.
force_context_ids (List[int], optional): List of specific context IDs to always include in the compressed result. Default is None.
force_context_number (int, optional): The number of context sections to forcibly include. Default is None.
use_sentence_level_filter (bool, optional): Whether to apply sentence-level filtering in compression. Default is False.
use_context_level_filter (bool, optional): Whether to apply context-level filtering in compression. Default is True.
use_token_level_filter (bool, optional): Whether to apply token-level filtering in compression. Default is True.
keep_split (bool, optional): Whether to preserve the original separators without compression. Default is False.
keep_first_sentence (int, optional): Number of sentences to forcibly preserve from the start of the context. Default is 0.
keep_last_sentence (int, optional): Number of sentences to forcibly preserve from the end of the context. Default is 0.
keep_sentence_number (int, optional): Total number of sentences to forcibly preserve in the compression. Default is 0.
high_priority_bonus (int, optional): Bonus score for high-priority sentences to influence their likelihood of being retained. Default is 100.
context_budget (str, optional): Token budget for the context-level filtering, expressed as a string to indicate flexibility. Default is "+100".
token_budget_ratio (float, optional): Ratio to adjust token budget during sentence-level filtering. Default is 1.4.
condition_in_question (str, optional): Specific condition to apply to question in the context. Default is "none".
reorder_context (str, optional): Strategy for reordering context in the compressed result. Default is "original".
dynamic_context_compression_ratio (float, optional): Ratio for dynamically adjusting context compression. Default is 0.0.
condition_compare (bool, optional): Whether to enable condition comparison during token-level compression. Default is False.
add_instruction (bool, optional): Whether to add the instruction to the prompt prefix. Default is False.
rank_method (str, optional): Method used for ranking elements during compression. Default is "llmlingua".
concate_question (bool, optional): Whether to concatenate the question to the compressed prompt. Default is True.
target_context (int, optional): The maximum number of contexts to be achieved. Default is -1, indicating no specific target.
context_level_rate (float, optional): The minimum compression rate target to be achieved in context level. Default is 1.0.
context_level_target_token (float, optional): The maximum number of tokens to be achieved in context level compression.
Default is -1, indicating no specific target. Only used in the coarse-to-fine compression senario.
force_context_ids (List[int], optional): List of specific context IDs to always include in the compressed result. Default is None.
return_word_label (bool, optional): Whether to return word with corresponding label. Default is False.
word_sep (str, optional): The sep token used in fn_labeled_original_prompt to partition words. Default is "\t\t|\t\t".
label_sep (str, optional): The sep token used in fn_labeled_original_prompt to partition word and label. Default is " ".
token_to_word (str, optional): How to convert token probability to word probability. Default is "mean".
force_tokens (List[str], optional): List of specific tokens to always include in the compressed result. Default is [].
force_reserve_digit (bool, optional): Whether to forcibly reserve tokens that containing digit (0,...,9). Default is False.
drop_consecutive (bool, optinal): Whether to drop tokens which are in 'force_tokens' but appears consecutively in compressed prompt.
Default is False.
chunk_end_tokens (List[str], optinal): The early stop tokens for segmenting chunk. Default is [".", "\n"],
Returns:
dict: A dictionary containing:
- "compressed_prompt" (str): The resulting compressed prompt.
- "compressed_prompt_list" (List[str]): List of the resulting compressed prompt. Only used in llmlingua2.
- "fn_labeled_original_prompt" (str): original words along with their labels
indicating whether to reserve in compressed prompt, in the format (word label_sep label)
Only used in llmlingua2 when return_word_label = True.
- "origin_tokens" (int): The original number of tokens in the input.
- "compressed_tokens" (int): The number of tokens in the compressed output.
- "ratio" (str): The compression ratio achieved, calculated as the original token number divided by the token number after compression.
- "rate" (str): The compression rate achieved, in a human-readable format.
- "saving" (str): Estimated savings in GPT-4 token usage.
"""
if self.use_llmlingua2:
return self.compress_prompt_llmlingua2(
context,
rate=rate,
target_token=target_token,
use_context_level_filter=use_context_level_filter,
use_token_level_filter=use_token_level_filter,
target_context=target_context,
context_level_rate=context_level_rate,
context_level_target_token=context_level_target_token,
force_context_ids=force_context_ids,
return_word_label=return_word_label,
word_sep=word_sep,
label_sep=label_sep,
token_to_word=token_to_word,
force_tokens=force_tokens,
force_reserve_digit=force_reserve_digit,
drop_consecutive=drop_consecutive,
chunk_end_tokens=chunk_end_tokens,
)
assert (
rate <= 1.0
), "Error: 'rate' must not exceed 1.0. The value of 'rate' indicates compression rate and must be within the range [0, 1]."
if not context:
context = [" "]
if isinstance(context, str):
context = [context]
assert not (
rank_method == "longllmlingua" and not question
), "In the LongLLMLingua, it is necessary to set a question."
if condition_compare and "_condition" not in condition_in_question:
condition_in_question += "_condition"
if rank_method == "longllmlingua":
if condition_in_question == "none":
condition_in_question = "after"
elif rank_method == "llmlingua":
condition_in_question = (
"none"
if "_condition" not in condition_in_question
else "none_condition"
)
origin_tokens = len(
self.oai_tokenizer.encode(
"\n\n".join([instruction] + context + [question]).strip()
)
)
context_tokens_length = [self.get_token_length(c) for c in context]
instruction_tokens_length, question_tokens_length = self.get_token_length(
instruction
), self.get_token_length(question)
if target_token == -1:
target_token = (
(
instruction_tokens_length
+ question_tokens_length
+ sum(context_tokens_length)
)
* rate
- instruction_tokens_length
- (question_tokens_length if concate_question else 0)
)
condition_flag = "_condition" in condition_in_question
condition_in_question = condition_in_question.replace("_condition", "")
if len(context) > 1 and use_context_level_filter:
context, dynamic_ratio, context_used = self.control_context_budget(
context,
context_tokens_length,
target_token,
force_context_ids,
force_context_number,
question,
condition_in_question,
reorder_context=reorder_context,
dynamic_context_compression_ratio=dynamic_context_compression_ratio,
rank_method=rank_method,
context_budget=context_budget,
context_segs=context_segs,
context_segs_rate=context_segs_rate,
context_segs_compress=context_segs_compress,
)
if context_segs is not None:
context_segs = [context_segs[idx] for idx in context_used]
context_segs_rate = [context_segs_rate[idx] for idx in context_used]
context_segs_compress = [
context_segs_compress[idx] for idx in context_used
]
else:
dynamic_ratio = [0.0] * len(context)
segments_info = []
if use_sentence_level_filter:
context, segments_info = self.control_sentence_budget(
context,
target_token,
keep_first_sentence=keep_first_sentence,
keep_last_sentence=keep_last_sentence,
keep_sentence_number=keep_sentence_number,
high_priority_bonus=high_priority_bonus,
token_budget_ratio=token_budget_ratio,
question=question,
condition_in_question=condition_in_question,
rank_method=rank_method,
context_segs=context_segs,
context_segs_rate=context_segs_rate,
context_segs_compress=context_segs_compress,
)
elif context_segs is not None:
for context_idx in range(len(context)):
segments_info.append(
[
(len(seg_text), seg_rate, seg_compress)
for seg_text, seg_rate, seg_compress in zip(
context_segs[context_idx],
context_segs_rate[context_idx],
context_segs_compress[context_idx],
)
]
)
segments_info = [
self.concate_segment_info(segment_info) for segment_info in segments_info
]
if condition_flag:
prefix = question + "\n\n" + instruction if add_instruction else question
if (
self.get_token_length(prefix + "\n\n") + iterative_size * 2
> self.max_position_embeddings
):
tokens = self.tokenizer(prefix, add_special_tokens=False).input_ids
prefix = self.tokenizer.decode(
tokens[: self.prefix_bos_num]
+ tokens[
len(tokens)
- self.max_position_embeddings
+ 2
+ self.prefix_bos_num
+ 2 * iterative_size :
]
)
start = self.get_prefix_length(prefix + "\n\n", context[0])
context = [prefix] + context
else:
start = 0
if use_token_level_filter:
context = self.iterative_compress_prompt(
context,
target_token,
iterative_size=iterative_size,
keep_split=keep_split,
start=start,
dynamic_ratio=dynamic_ratio,
condition_compare=condition_compare,
segments_info=segments_info,
)
compressed_prompt = (
self.tokenizer.batch_decode(context[0])[0]
.replace("<s> ", "")
.replace("<s>", "")
)
else:
if condition_flag:
context = context[1:]
compressed_prompt = "\n\n".join(context)
res = []
if instruction:
res.append(instruction)
if compressed_prompt.strip():
res.append(compressed_prompt)
if question and concate_question:
res.append(question)
compressed_prompt = "\n\n".join(res)
compressed_tokens = len(self.oai_tokenizer.encode(compressed_prompt))
saving = (origin_tokens - compressed_tokens) * 0.06 / 1000
ratio = 1 if compressed_tokens == 0 else origin_tokens / compressed_tokens
rate = 1 / ratio
return {
"compressed_prompt": compressed_prompt,
"origin_tokens": origin_tokens,
"compressed_tokens": compressed_tokens,
"ratio": f"{ratio:.1f}x",
"rate": f"{rate * 100:.1f}%",
"saving": f", Saving ${saving:.1f} in GPT-4.",
}
def compress_prompt_llmlingua2(
self,
context: List[str],
rate: float = 0.5,
target_token: int = -1,
use_context_level_filter: bool = False,
use_token_level_filter: bool = True,
target_context: int = -1,
context_level_rate: float = 1.0,
context_level_target_token: int = -1,
force_context_ids: List[int] = [],
return_word_label: bool = False,
word_sep: str = "\t\t|\t\t",
label_sep: str = " ",
token_to_word: str = "mean",
force_tokens: List[str] = [],
force_reserve_digit: bool = False,
drop_consecutive: bool = False,
chunk_end_tokens: List[str] = [".", "\n"],
):
"""
Compresses the given context, instruction and question.
Args:
context (List[str]): List of context strings that form the basis of the prompt.
rate (float, optional): The minimum compression rate target to be achieved. Default is 0.5. The actual compression rate
generally exceeds the specified target, but there can be fluctuations due to differences in tokenizers. If specified,
it should be a float greater than or equal to 1.0, representing the target compression rate.
target_token (int, optional): The maximum number of tokens to be achieved. Default is -1, indicating no specific target.
The actual number of tokens after compression should generally be less than the specified target_token, but there can
be fluctuations due to differences in tokenizers. If specified, compression will be based on the target_token as
the sole criterion, overriding the rate.
target_context (int, optional): The maximum number of contexts to be achieved. Default is -1, indicating no specific target.
Only used in the coarse-to-fine compression.
context_level_rate (float, optional): The minimum compression rate target to be achieved in context level. Default is 1.0.
Only used in the coarse-to-fine compression.
context_level_target_token (float, optional): The maximum number of tokens to be achieved in context level compression.
Default is -1, indicating no specific target. Only used in the coarse-to-fine compression senario.
force_context_ids (List[int], optional): List of specific context IDs to always include in the compressed result. Default is None.
return_word_label (bool, optional): Whether to return word with corresponding label. Default is False.
word_sep (str, optional): The sep token used in fn_labeled_original_prompt to partition words. Default is "\t\t|\t\t".
label_sep (str, optional): The sep token used in fn_labeled_original_prompt to partition word and label. Default is " ".
token_to_word (str, optional): How to convert token probability to word probability. Default is "mean".
force_tokens (List[str], optional): List of specific tokens to always include in the compressed result. Default is [].
force_reserve_digit (bool, optional): Whether to forcibly reserve tokens that containing digit (0,...,9). Default is False.
drop_consecutive (bool, optinal): Whether to drop tokens which are in 'force_tokens' but appears consecutively in compressed prompt.
Default is False.
chunk_end_tokens (List[str], optional): The early stop tokens for segmenting chunk. Default is [".", "\n"].
Returns:
dict: A dictionary containing:
- "compressed_prompt" (str): The resulting compressed prompt.
- "compressed_prompt_list" (List[str]): List of the resulting compressed prompt.
- "fn_labeled_original_prompt" (str): original words along with their labels
indicating whether to reserve in compressed prompt, in the format (word label_sep label)
- "origin_tokens" (int): The original number of tokens in the input.
- "compressed_tokens" (int): The number of tokens in the compressed output.
- "ratio" (str): The compression ratio achieved, in a human-readable format.
- "rate" (str): The compression rate achieved, in a human-readable format.
- "saving" (str): Estimated savings in GPT-4 token usage.
"""
assert len(force_tokens) <= self.max_force_token
token_map = {}
for i, t in enumerate(force_tokens):
if len(self.tokenizer.tokenize(t)) != 1:
token_map[t] = self.added_tokens[i]
chunk_end_tokens = copy.deepcopy(chunk_end_tokens)
for c in chunk_end_tokens:
if c in token_map:
chunk_end_tokens.append(token_map[c])
chunk_end_tokens = set(chunk_end_tokens)
if type(context) == str:
context = [context]
context = copy.deepcopy(context)
if len(context) == 1 and use_context_level_filter:
use_context_level_filter = False
n_original_token = 0
context_chunked = []
for i in range(len(context)):
n_original_token += self.get_token_length(context[i], use_oai_tokenizer=True)
for ori_token, new_token in token_map.items():
context[i] = context[i].replace(ori_token, new_token)
context_chunked.append(self.__chunk_context(context[i], chunk_end_tokens=chunk_end_tokens))
if use_context_level_filter:
# want use_context_level_filter but do not specify any parameters in context level?
# we will set context_level_rate = (rate + 1.0) / 2 if specify rate or target_token * 2 if specify target_token
if (
target_context <= 0
and context_level_rate >= 1.0
and context_level_target_token <= 0
):
if target_token < 0 and rate < 1.0:
context_level_rate = (
(rate + 1.0) / 2 if use_token_level_filter else rate
)
print(
f"set context level compression rate to {context_level_rate}."
)
if target_token >= 0:
context_level_target_token = (
target_token * 2 if use_token_level_filter else target_token
)
print(
f"set context level target token to {context_level_target_token}."
)
if target_context >= 0:
context_level_rate = min(target_context / len(context), 1.0)
# print(f'override context level compression rate to {context_level_rate} because you specified target_context = {target_context}.')
if context_level_target_token >= 0:
context_level_rate = min(
context_level_target_token / n_original_token, 1.0
)
# print(f'override context level compression rate to {context_level_rate} because you specified context_level_target_token = {context_level_target_token}.')
context_probs, context_words = self.__get_context_prob(
context_chunked,
token_to_word=token_to_word,
force_tokens=force_tokens,
token_map=token_map,
force_reserve_digit=force_reserve_digit,
)
threshold = np.percentile(
context_probs, int(100 * (1 - context_level_rate))
)
reserved_context = []
context_label = [False] * len(context_probs)
for i, p in enumerate(context_probs):
if p >= threshold or (
force_context_ids is not None and i in force_context_ids
):
reserved_context.append(context_chunked[i])
context_label[i] = True
n_reserved_token = 0
for chunks in reserved_context:
for c in chunks:
n_reserved_token += self.get_token_length(c, use_oai_tokenizer=True)
if target_token >= 0:
rate = min(target_token / n_reserved_token, 1.0)
print(
f"override compression rate to {rate} because you specified target_token = {target_token}."
)
if use_token_level_filter:
compressed_context, word_list, word_label_list = self.__compress(
reserved_context,
reduce_rate=max(0, 1 - rate),
token_to_word=token_to_word,
force_tokens=force_tokens,
token_map=token_map,
force_reserve_digit=force_reserve_digit,
drop_consecutive=drop_consecutive,
)
else:
compressed_context, word_list, word_label_list = self.__compress(
reserved_context,
reduce_rate=0,
token_to_word=token_to_word,
force_tokens=force_tokens,
token_map=token_map,
force_reserve_digit=force_reserve_digit,
drop_consecutive=drop_consecutive,
)
print(
"return the original text because you specify use_token_level_filter=False"
)
n_compressed_token = 0
for c in compressed_context:
n_compressed_token += self.get_token_length(c, use_oai_tokenizer=True)
saving = (n_original_token - n_compressed_token) * 0.06 / 1000
ratio = (
1 if n_compressed_token == 0 else n_original_token / n_compressed_token
)
res = {
"compressed_prompt": "\n\n".join(compressed_context),
"compressed_prompt_list": compressed_context,
"origin_tokens": n_original_token,
"compressed_tokens": n_compressed_token,
"ratio": f"{ratio:.1f}x",
"rate": f"{1 / ratio * 100:.1f}%",
"saving": f", Saving ${saving:.1f} in GPT-4.",
}
if return_word_label:
words = []
labels = []
j = 0
for i in range(len(context)):
if context_label[i]:
words.extend(word_list[j])
labels.extend(word_label_list[j])
j += 1
else:
words.extend(context_words[i])
labels.extend([0] * len(context_words[i]))
word_label_lines = word_sep.join(
[f"{word}{label_sep}{label}" for word, label in zip(words, labels)]
)
res["fn_labeled_original_prompt"] = word_label_lines
return res
if target_token > 0:
rate = min(target_token / n_original_token, 1.0)
print(
f"override compression rate to {rate} \
because you specified target_token = {target_token}."
)
if use_token_level_filter:
compressed_context, word_list, word_label_list = self.__compress(
context_chunked,
reduce_rate=max(0, 1 - rate),
token_to_word=token_to_word,
force_tokens=force_tokens,
token_map=token_map,
force_reserve_digit=force_reserve_digit,
drop_consecutive=drop_consecutive,
)
else:
compressed_context, word_list, word_label_list = self.__compress(
context_chunked,
reduce_rate=0,
token_to_word=token_to_word,
force_tokens=force_tokens,
token_map=token_map,
force_reserve_digit=force_reserve_digit,
drop_consecutive=drop_consecutive,
)
print(
"return the original text because you specify use_token_level_filter=False"
)
n_compressed_token = 0
for c in compressed_context:
n_compressed_token += self.get_token_length(c, use_oai_tokenizer=True)
saving = (n_original_token - n_compressed_token) * 0.06 / 1000
ratio = 1 if n_compressed_token == 0 else n_original_token / n_compressed_token
res = {
"compressed_prompt": "\n\n".join(compressed_context),
"compressed_prompt_list": compressed_context,
"origin_tokens": n_original_token,
"compressed_tokens": n_compressed_token,
"ratio": f"{ratio:.1f}x",
"rate": f"{1 / ratio * 100:.1f}%",
"saving": f", Saving ${saving:.1f} in GPT-4.",
}
if return_word_label:
words = []
labels = []
for w_list, l_list in zip(word_list, word_label_list):
words.extend(w_list)
labels.extend(l_list)
# new_words = []
# new_labels = []
# for i in range(len(words)):
# word, label = words[i], labels[i]
# if word in string.punctuation:
# if labels[i-1] == 1 and label == 1 and i > 0:
# new_words[-1] += word
# else:
# new_words.append(word)
# new_labels.append(label)
# word_label_lines = word_sep.join([f'{word}{label_sep}{label}' for word, label in zip(new_words, new_labels)])
word_label_lines = word_sep.join(
[f"{word}{label_sep}{label}" for word, label in zip(words, labels)]
)
res["fn_labeled_original_prompt"] = word_label_lines
return res
def get_token_length(self, text: str, add_special_tokens: bool = True, use_oai_tokenizer: bool = False):
if use_oai_tokenizer:
return len(self.oai_tokenizer.encode(text))
else:
return len(
self.tokenizer(text, add_special_tokens=add_special_tokens).input_ids
)
def get_prefix_length(self, prefix: str, text: str):
possible_prefix_token = max(self.get_token_length(prefix, False) - 3, 1)
full_input_ids = self.tokenizer(
prefix + text[:100], add_special_tokens=False
).input_ids
for i in range(possible_prefix_token, len(full_input_ids)):
cur_prefix = self.tokenizer.decode(full_input_ids[:i])
if cur_prefix == prefix:
break
assert self.tokenizer.decode(full_input_ids[i:]) == text[:100]
return i
def get_condition_ppl(
self,
text: str,
question: str,
condition_in_question: str = "none",
granularity: str = "sentence",
):
if condition_in_question == "none":
return self.get_ppl(text, granularity=granularity)
elif condition_in_question == "before":
return self.get_ppl(
question + text,
granularity=granularity,
condition_mode="after",
condition_pos_id=self.get_token_length(question) - 1,
)
elif condition_in_question == "after":
return self.get_ppl(
text + question,
granularity=granularity,
condition_mode="after",
condition_pos_id=self.get_token_length(text) - 1,
)
def get_dynamic_compression_ratio(
self,
context: list,
target_token: float,
iterative_size: int,
dynamic_ratio: list,
start: int,
seg_info: List[List[tuple]] = None,
):
def get_ratio(base: float, delta: float):
return max(min(1, base + delta), 0)
context_length = [self.get_token_length(ii, False) + 2 for ii in context]
if start:
context_length = context_length[1:]
tau = target_token / (sum(context_length) + 1)
res, idx, last, last_target = [], 0, 1, []
while idx < len(context_length):
if last + context_length[idx] >= iterative_size:
last_target.append(
(iterative_size - last, get_ratio(tau, dynamic_ratio[idx]))
)
res.append(last_target)
last = last + context_length[idx] - iterative_size
if last > iterative_size:
k = last // iterative_size
res.extend(
[[(iterative_size, get_ratio(tau, dynamic_ratio[idx]))]] * k
)
last -= k * iterative_size
last_target = (
[(last, get_ratio(tau, dynamic_ratio[idx]))] if last else []
)
else:
last += context_length[idx]
last_target.append(
(context_length[idx], get_ratio(tau, dynamic_ratio[idx]))
)
idx += 1
if last_target:
res.append(last_target)
return res
def get_structured_dynamic_compression_ratio(
self,
context: list,
iterative_size: int,
dynamic_ratio: list,
start: int,
seg_info: List[List[tuple]] = None,
):
if start:
pure_context = context[1:]
else:
pure_context = context
global_dynamic_rate, global_dynamic_compress, segments = [], [], []
for context_idx, text in enumerate(pure_context):
text_seen = 0
for seg_idx, (seg_len, seg_rate, seg_compress) in enumerate(
seg_info[context_idx]
):
seg_text = text[text_seen : text_seen + seg_len]
if (
seg_idx == len(seg_info[context_idx]) - 1
and context_idx != len(pure_context) - 1
):
seg_text += "\n\n"
segments.append(seg_text)
if seg_compress:
global_dynamic_rate.append(seg_rate)
else:
global_dynamic_rate.append(1.0)
global_dynamic_compress.append(seg_compress)
text_seen += seg_len
origin_text = "\n\n".join(pure_context)
assert len("".join(segments)) == len(origin_text)
assert len(segments) == len(global_dynamic_rate) == len(global_dynamic_compress)
text_input_ids = self.tokenizer(
"\n\n".join(context), add_special_tokens=False
).input_ids[start:]
assert self.tokenizer.decode(text_input_ids) == origin_text
dynamic_compression_ratio = self.token_segment(
text_input_ids,
iterative_size,
segments,
global_dynamic_rate,
global_dynamic_compress,
)
return dynamic_compression_ratio
def token_segment(
self,
text_input_ids: List[int],
iterative_size: int,
segments: List[str],
global_dynamic_rate: List[float],
global_dynamic_compress: List[bool],
):
decode_window = 3
seg_idx, seg_seen, token_seen_num, last_rate = 0, 0, 0, -1
dynamic_compression_rate, local_compresssion_rate = [], []
for i in range(len(text_input_ids)):
if i < decode_window:
id_pre, id_cur = text_input_ids[:i], text_input_ids[: i + 1]
else:
id_pre, id_cur = (
text_input_ids[i - decode_window + 1 : i],
text_input_ids[i - decode_window + 1 : i + 1],
)
cur_word = self.tokenizer.decode(id_cur)[
len(self.tokenizer.decode(id_pre)) :
]
cur_word_len = len(cur_word)
if cur_word_len and cur_word_len >= len(segments[seg_idx]) - seg_seen:
possible_rate, possible_compress = [], []
while (
cur_word_len and cur_word_len >= len(segments[seg_idx]) - seg_seen
):
possible_rate.append(global_dynamic_rate[seg_idx])
possible_compress.append(global_dynamic_compress[seg_idx])
cur_word_len -= len(segments[seg_idx]) - seg_seen
seg_idx += 1
seg_seen = 0
if cur_word_len:
possible_rate.append(global_dynamic_rate[seg_idx])
possible_compress.append(global_dynamic_compress[seg_idx])
new_rate = 1.0 if False in possible_compress else min(possible_rate)
else:
new_rate = global_dynamic_rate[seg_idx]
if new_rate != last_rate and i - token_seen_num:
local_compresssion_rate.append((i - token_seen_num, last_rate))
token_seen_num = i
last_rate = new_rate
seg_seen += cur_word_len
if (i + 1) % iterative_size == 0:
if token_seen_num != i + 1:
local_compresssion_rate.append((i + 1 - token_seen_num, last_rate))
token_seen_num = i + 1
dynamic_compression_rate.append(local_compresssion_rate[:])
local_compresssion_rate = []
if token_seen_num != len(text_input_ids):
local_compresssion_rate.append(
(len(text_input_ids) - token_seen_num, last_rate)
)
if local_compresssion_rate != []:
dynamic_compression_rate.append(local_compresssion_rate[:])
return dynamic_compression_rate
def control_context_budget(
self,
context: List[str],
context_tokens_length: List[int],
target_token: float,
force_context_ids: List[int] = None,
force_context_number: int = None,
question: str = "",
condition_in_question: str = "none",
reorder_context: str = "original",
dynamic_context_compression_ratio: float = 0.0,
rank_method: str = "longllmlingua",
context_budget: str = "+100",
context_segs: List[List[str]] = None,
context_segs_rate: List[List[float]] = None,
context_segs_compress: List[List[bool]] = None,
):
demostrations_sort = self.get_rank_results(
context,
question,
rank_method,
condition_in_question,
context_tokens_length,
)
if target_token < 0:
target_token = 100
target_token = eval("target_token" + context_budget)
res = []
used = force_context_ids if force_context_ids is not None else []
if context_segs is not None:
for idx, _ in enumerate(context):
if False in context_segs_compress[idx]:
used.append(idx)
self.context_idxs.append([x for idx, (x, _) in enumerate(demostrations_sort)])
for idx, _ in demostrations_sort:
if idx >= len(context_tokens_length):
continue
target_token -= context_tokens_length[idx]
if idx not in used:
used.append(idx)
if target_token < 0 or (
force_context_number is not None and len(res) >= force_context_number
):
break
original_used = used
if reorder_context == "original":
used = sorted(used)
elif reorder_context == "two_stage":
l, r = [_ for idx, _ in enumerate(used) if idx % 2 == 0], [
_ for idx, _ in enumerate(used) if idx % 2 == 1
]
used = l + r[::-1]
if dynamic_context_compression_ratio > 0:
N = len(used)
dynamic_ratio = [
i * (abs(dynamic_context_compression_ratio) / (N - 1)) if N > 1 else 0
for i in range(-(N - 1), N, 2)
][::-1]
dynamic_ratio_map = {i: j for i, j in zip(original_used, dynamic_ratio)}
dynamic_ratio = [dynamic_ratio_map[i] for i in used]
else:
dynamic_ratio = [0.0] * len(used)
res = [context[idx] for idx in used if idx < len(context)]
return res, dynamic_ratio, used
def control_sentence_budget(
self,
context: List[str],
target_token: float,
keep_first_sentence: int = 0,
keep_last_sentence: int = 0,
keep_sentence_number: int = 0,
high_priority_bonus: int = 100,
token_budget_ratio: float = 1.4,
question: str = "",
condition_in_question: str = "none",
rank_method: str = "longllmlingua",
context_segs: List[List[str]] = None,
context_segs_rate: List[List[float]] = None,
context_segs_compress: List[List[bool]] = None,
):
def keep_sentence(dem_idx: int, sent_keep: int):
idxs = sorted(dem_g[dem_idx], key=lambda x: sentence_ppl[x])[:sent_keep]
for idx in idxs:
sentence_ppl[idx] += high_priority_bonus
def sync_sentence(segments, text):
seg_num = len(segments)
new_segments = []
text_seen = 0
seg_idx, cur_seg_seen = 0, 0
for i, s in enumerate(text):
while seg_idx < seg_num and s != segments[seg_idx][cur_seg_seen]:
if cur_seg_seen < len(segments[seg_idx]) - 1:
cur_seg_seen += 1
continue
new_segments.append(text[text_seen:i])
text_seen = i
seg_idx += 1
cur_seg_seen = 0
cur_seg_seen += 1
if seg_idx == seg_num:
break
if cur_seg_seen == len(segments[seg_idx]):
new_segments.append(text[text_seen : i + 1])
text_seen = i + 1
seg_idx += 1
cur_seg_seen = 0
if text_seen < len(text):
new_segments.append(text[text_seen:])
assert len("".join(new_segments)) == len(text)
return new_segments
sentences = [nltk.sent_tokenize(c) for c in context]
dem_g, s2de, idx = defaultdict(set), defaultdict(int), 0
for idx_d, s in enumerate(sentences):
for _ in s:
dem_g[idx_d].add(idx)
s2de[idx] = idx_d
idx += 1
if context_segs is not None:
context_segs = [
sync_sentence(s, "".join(c)) for s, c in zip(context_segs, sentences)
]
sen2seg_ratio = {}
idx = 0
for idx_d, sentences_each_context in enumerate(sentences):
segments_length = [len(s) for s in context_segs[idx_d]]
seg_idx, cur_seg_seen = 0, 0
for sentence in sentences_each_context:
sentence_seg_ratio = []
remain = len(sentence)
while remain:
if segments_length[seg_idx] - cur_seg_seen <= remain:
new_seg_len = segments_length[seg_idx] - cur_seg_seen
sentence_seg_ratio.append(
(
new_seg_len,
context_segs_rate[idx_d][seg_idx],
context_segs_compress[idx_d][seg_idx],
)
)
seg_idx += 1
cur_seg_seen = 0
remain -= new_seg_len
else:
sentence_seg_ratio.append(
(
remain,
context_segs_rate[idx_d][seg_idx],
context_segs_compress[idx_d][seg_idx],
)
)
cur_seg_seen += remain
remain = 0
sen2seg_ratio[idx] = sentence_seg_ratio
idx += 1
context_sentences = [s for ii in sentences for s in ii]
sentence_tokens_length = [
self.get_token_length(sentence) for sentence in context_sentences
]
N = len(context_sentences)
flags = list(range(len(context_sentences)))
if len(sentence_tokens_length) == 1:
return context
if rank_method == "longllmlingua":
sentence_ppl = [
self.get_condition_ppl(sentence, question, condition_in_question)
.cpu()
.numpy()
.item()
for sentence in context_sentences
]
if keep_first_sentence:
sentence_ppl[:keep_first_sentence] = [
ii + high_priority_bonus
for ii in sentence_ppl[:keep_first_sentence]
]
if keep_last_sentence:
sentence_ppl[-keep_last_sentence:] = [
ii + high_priority_bonus
for ii in sentence_ppl[-keep_last_sentence:]
]
if keep_sentence_number:
for dem_idx in range(len(sentences)):
keep_sentence(dem_idx, keep_sentence_number)
sort_direct = -1 if condition_in_question == "none" else 1
sent_sort = sorted(
enumerate(sentence_ppl), key=lambda x: sort_direct * x[1]
)
else:
sent_sort = self.get_rank_results(
context_sentences,
question,
rank_method,
condition_in_question,
[0] * len(context_sentences),
)
sentence_flags = [False] * N
if target_token < 0:
target_token = 100
target_token *= token_budget_ratio
res = []
for idx, _ in sent_sort:
idx = flags[idx]
target_token -= sentence_tokens_length[idx]
sentence_flags[idx] = True
if target_token < 0:
break
if context_segs is not None:
for idx in range(N):
preserved = [sen_seg_info[2] for sen_seg_info in sen2seg_ratio[idx]]
if False in preserved:
sentence_flags[idx] = True
idx = 0
res = []
new_segments_info = []
for s in sentences:
tmp = [jj for ii, jj in enumerate(s) if sentence_flags[idx + ii]]
res.append("".join(tmp))
if context_segs is not None:
segment_ratio = []
for ii in range(len(s)):
if sentence_flags[idx + ii]:
segment_ratio.extend(sen2seg_ratio[idx + ii])
new_segments_info.append(segment_ratio)
idx += len(s)
if context_segs is not None:
new_segments_info = [
self.concate_segment_info(segment_info)
for segment_info in new_segments_info
]
return res, new_segments_info
def get_compressed_input(
self,
loss,
input_ids,
attention_mask,
end=200,
iterative_size=200,
threshold=0.5,
keep_flag=None,
split_token_id: int = 13,
start: int = 0,
self_loss=None,
self_input_ids=None,
self_attention_mask=None,
):
if self_loss is not None:
need_idx = torch.concat(
[
loss[:start] > 0,
self_loss[: loss[start:].shape[0]] - loss[start:] > threshold,
loss[:1] > 0,
]
)
else:
need_idx = torch.concat([loss > threshold, loss[:1] > 0])
need_idx[end:] = 1
need_idx[: end - iterative_size] = 1
loss = loss[need_idx[:-1]]
if self_loss is not None:
if need_idx.shape[0] < self_loss.shape[0] + start + 1:
need_idx = torch.cat(
[
need_idx,
torch.ones(
self_loss.shape[0] - need_idx.shape[0] + start + 1,
dtype=torch.bool,
).to(need_idx.device),
]
)
self_loss = self_loss[need_idx[start:-1]]
if need_idx.shape[0] < input_ids.shape[1]:
need_idx = torch.cat(
[
need_idx,
torch.ones(
input_ids.shape[1] - need_idx.shape[0], dtype=torch.bool
).to(need_idx.device),
]
)
elif need_idx.shape[0] > input_ids.shape[1]:
need_idx = need_idx[: input_ids.shape[1]]
if keep_flag is not None:
need_idx[keep_flag == 1] = 1
last = -1
if keep_flag is not None:
for ii in range(max(0, end - iterative_size), end):
if need_idx[ii] != 1:
continue
now = input_ids[0][ii].detach().cpu().item()
if (
now == split_token_id
and last == split_token_id
and keep_flag[ii].detach().cpu().item() == 0
):
need_idx[ii] = 0
else:
last = now
compressed_input_ids = input_ids[attention_mask == 1][need_idx].unsqueeze(0)
compressed_attention_mask = attention_mask[attention_mask == 1][
need_idx
].unsqueeze(0)
if self_loss is not None:
self_compressed_input_ids = self_input_ids[self_attention_mask == 1][
need_idx[start:]
].unsqueeze(0)
self_compressed_attention_mask = self_attention_mask[
self_attention_mask == 1
][need_idx[start:]].unsqueeze(0)
else:
self_compressed_input_ids, self_compressed_attention_mask = None, None
if keep_flag is not None:
if len(keep_flag) > len(need_idx):
keep_flag = torch.cat(
[
keep_flag[:start],
keep_flag[start : len(need_idx) + start][need_idx],
keep_flag[start + len(need_idx) :],
]
)
else:
keep_flag = keep_flag[need_idx]
end -= (need_idx[:end] == 0).sum()
return (
compressed_input_ids,
compressed_attention_mask,
keep_flag,
end,
loss,
self_loss,
self_compressed_input_ids,
self_compressed_attention_mask,
)
def get_estimate_threshold_base_distribution(
self, ppl, ratio: float, condition_flag: bool = False
):
if ratio == 1.0:
return float("-inf")
ppl = ppl[ppl != 10000]
target_token = max(0, min(len(ppl) - 1, int(len(ppl) * ratio) - 1))
return (
ppl.sort(descending=not condition_flag)
.values[target_token]
.detach()
.cpu()
.item()
)
def iterative_compress_prompt(
self,
context: List[str],
target_token: float,
iterative_size: int = 200,
keep_split: bool = False,
split_token_id: int = 13,
start: int = 0,
dynamic_ratio: list = None,
condition_compare: bool = False,
segments_info: List[List[tuple]] = None,
):
if segments_info is None or segments_info == []:
iterative_ratios = self.get_dynamic_compression_ratio(
context, target_token, iterative_size, dynamic_ratio, start
)
else:
iterative_ratios = self.get_structured_dynamic_compression_ratio(
context, iterative_size, dynamic_ratio, start, segments_info
)
context = "\n\n".join(context)
tokenized_text = self.tokenizer(
context, return_tensors="pt", add_special_tokens=False
)
input_ids = tokenized_text["input_ids"].to(self.device)
attention_mask = tokenized_text["attention_mask"].to(self.device)
N = (attention_mask == 1).sum()
compressed_input_ids, compressed_attention_mask = input_ids, attention_mask
if condition_compare:
self_input_ids, self_attention_mask = (
input_ids[:, start:],
attention_mask[:, start:],
)
self_compressed_input_ids, self_compressed_attention_mask = (
self_input_ids,
self_attention_mask,
)
end = min(iterative_size + start, compressed_input_ids.shape[1])
threshold, keep_flag = None, None
if keep_split:
input_ids_numpy = input_ids.cpu().detach().numpy()[0]
N = len(input_ids_numpy)
keep_flag = [
int(
(
ii > 0
and input_ids_numpy[ii] == split_token_id
and input_ids_numpy[ii - 1] == split_token_id
)
or (
ii < N - 1
and input_ids_numpy[ii] == split_token_id
and input_ids_numpy[ii + 1] == split_token_id
)
)
for ii in range(N)
]
keep_flag = torch.tensor(keep_flag).to(self.device)
past_key_values, past_loss, ready_end = None, None, 0
self_past_key_values, self_past_loss, self_ready_end = None, None, 0
pop_compressed_input_ids, pop_self_compressed_input_ids = None, None
idx = 0
while end <= compressed_input_ids.shape[1]:
if end > self.max_position_embeddings and past_key_values is not None:
# KV-Cache Compression
e, s = end - self.max_position_embeddings, min(
self.cache_bos_num + start, self.max_position_embeddings
)
if pop_compressed_input_ids is None:
pop_compressed_input_ids = compressed_input_ids[:, :e]
else:
pop_compressed_input_ids = torch.cat(
[pop_compressed_input_ids, compressed_input_ids[:, :e]], dim=-1
)
compressed_input_ids = compressed_input_ids[:, e:]
compressed_attention_mask = compressed_attention_mask[:, e:]
past_key_values = [
[
torch.cat([k[..., :s, :], k[..., s + e :, :]], dim=-2),
torch.cat([v[..., :s, :], v[..., s + e :, :]], dim=-2),
]
for k, v in past_key_values
]
if keep_flag is not None:
keep_flag = keep_flag[e:]
end, ready_end = end - e, ready_end - e
if condition_compare:
s = min(s, self_past_key_values[0][0].shape[2] - e)
self_ready_end -= e
if pop_self_compressed_input_ids is None:
pop_self_compressed_input_ids = self_compressed_input_ids[:, :e]
else:
pop_self_compressed_input_ids = torch.cat(
[
pop_self_compressed_input_ids,
self_compressed_input_ids[:, :e],
],
dim=-1,
)
self_compressed_input_ids = self_compressed_input_ids[:, e:]
self_compressed_attention_mask = self_compressed_attention_mask[
:, e:
]
self_past_key_values = [
[
torch.cat([k[..., :s, :], k[..., s + e :, :]], dim=-2),
torch.cat([v[..., :s, :], v[..., s + e :, :]], dim=-2),
]
for k, v in self_past_key_values
]
loss, past_key_values = self.get_ppl(
"",
"token",
compressed_input_ids,
compressed_attention_mask,
past_key_values=past_key_values,
return_kv=True,
end=end if idx else None,
)
if loss.shape[0] == 0:
break
if past_loss is not None:
if end - 1 > len(past_loss):
past_loss = torch.cat(
[past_loss, torch.zeros_like(loss)[: end - 1 - len(past_loss)]]
)
past_loss[ready_end : end - 1] = loss
loss = past_loss
else:
past_loss = loss
if idx:
past_key_values = [
[k[:, :, : end - iterative_size], v[:, :, : end - iterative_size]]
for k, v in past_key_values
]
else:
past_key_values = None
if condition_compare:
self_loss, self_past_key_values = self.get_ppl(
"",
"token",
self_compressed_input_ids,
self_compressed_attention_mask,
past_key_values=self_past_key_values,
return_kv=True,
end=end - start if idx else None,
)
if self_past_loss is not None:
if end - start - 1 > len(self_past_loss):
self_past_loss = torch.cat(
[
self_past_loss,
torch.zeros_like(self_loss)[
: end - 1 - start - len(self_past_loss)
],
]
)
self_past_loss[self_ready_end : end - start - 1] = self_loss
self_loss = self_past_loss
else:
self_past_loss = self_loss
if idx:
self_past_key_values = [
[
k[:, :, : end - iterative_size - start],
v[:, :, : end - iterative_size - start],
]
for k, v in self_past_key_values
]
else:
self_past_key_values = None
self_ready_end = (
end - start - iterative_size if not (start and idx == 0) else 0
)
ready_end = end - iterative_size if not (start and idx == 0) else 0
for delta_end, ratio in iterative_ratios[idx]:
loss = past_loss
if condition_compare:
self_loss = self_past_loss
threshold = self.get_estimate_threshold_base_distribution(
self_loss[: loss[start:].shape[0]] - loss[start:], ratio, False
)
else:
threshold = self.get_estimate_threshold_base_distribution(
loss, ratio, False
)
(
compressed_input_ids,
compressed_attention_mask,
keep_flag,
end,
past_loss,
self_past_loss,
self_compressed_input_ids,
self_compressed_attention_mask,
) = self.get_compressed_input(
loss,
compressed_input_ids,
compressed_attention_mask,
end - iterative_size + delta_end,
iterative_size=delta_end,
threshold=threshold,
keep_flag=keep_flag,
split_token_id=split_token_id,
start=start,
self_loss=self_loss if condition_compare else None,
self_input_ids=(
self_compressed_input_ids if condition_compare else None
),
self_attention_mask=(
self_compressed_attention_mask if condition_compare else None
),
)
end += iterative_size
idx += 1
if pop_compressed_input_ids is not None:
compressed_input_ids = torch.cat(
[pop_compressed_input_ids, compressed_input_ids], dim=-1
)
return compressed_input_ids[:, start:], compressed_attention_mask[:, start:]
def recover(
self,
original_prompt: str,
compressed_prompt: str,
response: str,
):
def match_from_compressed(response_word):
response_input_ids = self.tokenizer(
response_word, add_special_tokens=False
)["input_ids"]
response_set, response_c = set(response_input_ids), defaultdict(list)
for idx in range(M):
if original_input_ids[idx] in response_set:
response_c[original_input_ids[idx]].append(idx)
res, res_min, res_c = None, float("inf"), 1
n = len(response_input_ids)
for l in response_c[response_input_ids[0]]:
x, y, c = 0, l, 1
for x in range(1, n):
idx = bisect.bisect_right(response_c[response_input_ids[x]], y)
if (
idx >= len(response_c[response_input_ids[x]])
or response_c[response_input_ids[x]][idx] - y > 10
):
continue
c += 1
y = response_c[response_input_ids[x]][idx]
if c > res_c:
res_c = c
res_min = y - l + 1
res = (l, y + 1)
elif c == res_c and y - l + 1 < res_min:
res_min = y - l + 1
res = (l, y + 1)
if res is None:
return response_word
# while l > 0 and not self.tokenizer.convert_ids_to_tokens(original_input_ids[l]).startswith("_"):
# l -= 1
# while r < M - 1 and not self.tokenizer.convert_ids_to_tokens(original_input_ids[l]).startswith("_"):
# l -= 1
return self.tokenizer.decode(original_input_ids[res[0] : res[1]])
response_words = response.split(" ")
original_input_ids = self.tokenizer(original_prompt, add_special_tokens=False)[
"input_ids"
]
N, M = len(response_words), len(original_input_ids)
recovered_response_words = []
l = 0
while l < N:
if response_words[l] not in compressed_prompt:
recovered_response_words.append(response_words[l])
l += 1
continue
r = l
while (
r + 1 < N and " ".join(response_words[l : r + 2]) in compressed_prompt
):
r += 1
match_words = match_from_compressed(" ".join(response_words[l : r + 1]))
recovered_response_words.append(match_words)
l = r + 1
return " ".join(recovered_response_words)
def get_rank_results(
self,
context: list,
question: str,
rank_method: str,
condition_in_question: str,
context_tokens_length: list,
):
def get_distance_bm25(corpus, query):
from rank_bm25 import BM25Okapi
tokenized_corpus = [doc.split(" ") for doc in corpus]
bm25 = BM25Okapi(tokenized_corpus)
tokenized_query = query.split(" ")
doc_scores = bm25.get_scores(tokenized_query)
idx = [(ii, 0) for ii in (-doc_scores).argsort()]
return idx
def get_distance_gzip(corpus, query):
def get_score(x, y):
cx, cy = len(gzip.compress(x.encode())), len(gzip.compress(y.encode()))
cxy = len(gzip.compress(f"{x} {y}".encode()))
return (cxy - min(cx, cy)) / max(cx, cy)
import gzip
doc_scores = [get_score(doc, query) for doc in corpus]
idx = [(ii, 0) for ii in np.argsort(doc_scores)]
return idx
def get_distance_sentbert(corpus, query):
from sentence_transformers import SentenceTransformer, util
if self.retrieval_model is None or self.retrieval_model_name != rank_method:
self.retrieval_model = SentenceTransformer("multi-qa-mpnet-base-dot-v1")
self.retrieval_model_name = rank_method
doc_embeds = self.retrieval_model.encode(corpus)
query = self.retrieval_model.encode(query)
doc_scores = -util.dot_score(doc_embeds, query).cpu().numpy().reshape(-1)
idx = [(ii, 0) for ii in np.argsort(doc_scores)]
return idx
def get_distance_openai(corpus, query):
import openai
from sentence_transformers import util
openai.api_key = self.open_api_config.get("api_key", "")
openai.api_base = self.open_api_config.get(
"api_base", "https://api.openai.com/v1"
)
openai.api_type = self.open_api_config.get("api_type", "open_ai")
openai.api_version = self.open_api_config.get("api_version", "2023-05-15")
engine = self.open_api_config.get("engine", "text-embedding-ada-002")
def get_embed(text):
return openai.Embedding.create(
input=[text.replace("\n", " ")], engine=engine
)["data"][0]["embedding"]
doc_embeds = [get_embed(i) for i in corpus]
query = get_embed(query)
doc_scores = -util.dot_score(doc_embeds, query).cpu().numpy().reshape(-1)
idx = [(ii, 0) for ii in np.argsort(doc_scores)]
return idx
def get_distance_sentbert_bge(corpus, query):
from sentence_transformers import SentenceTransformer, util
if self.retrieval_model is None or self.retrieval_model_name != rank_method:
self.retrieval_model = SentenceTransformer("BAAI/bge-large-en-v1.5")
self.retrieval_model_name = rank_method
doc_embeds = self.retrieval_model.encode(
[i for i in corpus], normalize_embeddings=True
)
query = self.retrieval_model.encode(query, normalize_embeddings=True)
doc_scores = -util.dot_score(doc_embeds, query).cpu().numpy().reshape(-1)
idx = [(ii, 0) for ii in np.argsort(doc_scores)]
return idx
def get_distance_bge_ranker(corpus, query):
from transformers import AutoModelForSequenceClassification, AutoTokenizer
pairs = [[i, query] for i in corpus]
if self.retrieval_model is None or self.retrieval_model_name != rank_method:
tokenizer = AutoTokenizer.from_pretrained("BAAI/bge-reranker-large")
model = (
AutoModelForSequenceClassification.from_pretrained(
"BAAI/bge-reranker-large"
)
.eval()
.to(self.device)
)
self.retrieval_model = [tokenizer, model]
self.retrieval_model_name = rank_method
with torch.no_grad():
inputs = self.retrieval_model[0](
pairs,
padding=True,
truncation=True,
return_tensors="pt",
max_length=512,
).to(self.device)
scores = (
self.retrieval_model[1](**inputs, return_dict=True)
.logits.view(
-1,
)
.float()
)
idx = [(ii, 0) for ii in np.argsort(-scores.cpu())]
return idx
def get_distance_bge_llmembedder(corpus, query):
from transformers import AutoModel, AutoTokenizer
if self.retrieval_model is None or self.retrieval_model_name != rank_method:
tokenizer = AutoTokenizer.from_pretrained("BAAI/llm-embedder")
model = (
AutoModel.from_pretrained("BAAI/llm-embedder")
.eval()
.to(self.device)
)
self.retrieval_model = [tokenizer, model]
self.retrieval_model_name = rank_method
instruction_qa_query = (
"Represent this query for retrieving relevant documents: "
)
instruction_qa_key = "Represent this document for retrieval: "
queries = [instruction_qa_query + query for _ in corpus]
keys = [instruction_qa_key + key for key in corpus]
with torch.no_grad():
query_inputs = self.retrieval_model[0](
queries,
padding=True,
truncation=True,
return_tensors="pt",
max_length=512,
).to(self.device)
key_inputs = self.retrieval_model[0](
keys,
padding=True,
truncation=True,
return_tensors="pt",
max_length=512,
).to(self.device)
query_outputs = self.retrieval_model[1](**query_inputs)
key_outputs = self.retrieval_model[1](**key_inputs)
# CLS pooling
query_embeddings = query_outputs.last_hidden_state[:, 0]
key_embeddings = key_outputs.last_hidden_state[:, 0]
# Normalize
query_embeddings = torch.nn.functional.normalize(
query_embeddings, p=2, dim=1
)
key_embeddings = torch.nn.functional.normalize(
key_embeddings, p=2, dim=1
)
similarity = query_embeddings @ key_embeddings.T
idx = [(ii, 0) for ii in np.argsort(-similarity[0].cpu())]
return idx
def get_distance_jinza(corpus, query):
from numpy.linalg import norm
from transformers import AutoModel
def cos_sim(a, b):
return (a @ b.T) / (norm(a) * norm(b))
if self.retrieval_model is None or self.retrieval_model_name != rank_method:
model = (
AutoModel.from_pretrained(
"jinaai/jina-embeddings-v2-base-en", trust_remote_code=True
)
.eval()
.to(self.device)
)
self.retrieval_model = model
self.retrieval_model_name = rank_method
doc_embeds = self.retrieval_model.encode(corpus)
query = self.retrieval_model.encode(query)
doc_scores = cos_sim(doc_embeds, query)
idx = [(ii, 0) for ii in np.argsort(-doc_scores)]
return idx
def get_distance_voyageai(corpus, query):
import voyageai
from sentence_transformers import util
voyageai.api_key = self.open_api_config.get("voyageai_api_key", "")
def get_embed(text):
return voyageai.get_embedding(text, model="voyage-01")
doc_embeds = [get_embed(i) for i in corpus]
query = get_embed(query)
doc_scores = -util.dot_score(doc_embeds, query).cpu().numpy().reshape(-1)
idx = [(ii, 0) for ii in np.argsort(doc_scores)]
return idx
def get_distance_cohere(corpus, query):
import cohere
api_key = self.open_api_config.get("cohere_api_key", "")
co = cohere.Client(api_key)
results = co.rerank(
model="rerank-english-v2.0", query=query, documents=corpus, top_n=20
)
c_map = {jj: ii for ii, jj in enumerate(corpus)}
doc_rank = [c_map[ii.document["text"]] for ii in results]
idx = [(ii, 0) for ii in doc_rank]
return idx
def get_distance_longllmlingua(corpus, query):
context_ppl = [
self.get_condition_ppl(
d,
query
+ " We can get the answer to this question in the given documents.",
condition_in_question,
)
- dl * 2 / 250 * 0
for d, dl in zip(corpus, context_tokens_length)
]
sort_direct = -1 if condition_in_question == "none" else 1
ys = sorted(enumerate(context_ppl), key=lambda x: sort_direct * x[1])
return ys
method = None
if rank_method == "bm25":
method = get_distance_bm25
elif rank_method == "gzip":
method = get_distance_gzip
elif rank_method == "sentbert":
method = get_distance_sentbert
elif rank_method == "openai":
method = get_distance_openai
elif rank_method in ["longllmlingua", "llmlingua"]:
method = get_distance_longllmlingua
elif rank_method == "bge":
method = get_distance_sentbert_bge
elif rank_method == "bge_reranker":
method = get_distance_bge_ranker
elif rank_method == "bge_llmembedder":
method = get_distance_bge_llmembedder
elif rank_method == "jinza":
method = get_distance_jinza
elif rank_method == "voyageai":
method = get_distance_voyageai
elif rank_method == "cohere":
method = get_distance_cohere
return method(context, question)
def segment_structured_context(
self,
context: List[str],
global_rate: float,
):
new_context, context_segs, context_segs_rate, context_segs_compress = (
[],
[],
[],
[],
)
for text in context:
if not text.startswith("<llmlingua"):
text = "<llmlingua>" + text
if not text.endswith("</llmlingua>"):
text = text + "</llmlingua>"
# Regular expression to match <llmlingua, rate=x, compress=y>content</llmlingua>, allowing rate and compress in any order
pattern = r"<llmlingua\s*(?:,\s*rate\s*=\s*([\d\.]+))?\s*(?:,\s*compress\s*=\s*(True|False))?\s*(?:,\s*rate\s*=\s*([\d\.]+))?\s*(?:,\s*compress\s*=\s*(True|False))?\s*>([^<]+)</llmlingua>"
matches = re.findall(pattern, text)
# Extracting segment contents
segments = [match[4] for match in matches]
# Extracting rate and compress, considering their possible positions
segs_rate = [
float(match[0]) if match[0] else (float(match[2]) if match[2] else None)
for match in matches
]
segs_compress = [
(
match[1] == "True"
if match[1]
else (match[3] == "True" if match[3] else None)
)
for match in matches
]
segs_compress = [
compress if compress is not None else True for compress in segs_compress
]
segs_rate = [
rate if rate else (global_rate if compress else 1.0)
for rate, compress in zip(segs_rate, segs_compress)
]
assert (
len(segments) == len(segs_rate) == len(segs_compress)
), "The number of segments, rates, and compress flags should be the same."
assert all(
seg_rate <= 1.0 for seg_rate in segs_rate
), "Error: 'rate' must not exceed 1.0. The value of 'rate' indicates compression rate and must be within the range [0, 1]."
new_context.append("".join(segments))
context_segs.append(segments)
context_segs_rate.append(segs_rate)
context_segs_compress.append(segs_compress)
return new_context, context_segs, context_segs_rate, context_segs_compress
def concate_segment_info(
self,
segment_info: List[List[tuple]],
):
new_segment_info = []
for i, (seg_len, seg_ratio, seg_compress) in enumerate(segment_info):
if (
new_segment_info
and new_segment_info[-1][1] == seg_ratio
and new_segment_info[-1][2] == seg_compress
):
new_segment_info[-1] = (
new_segment_info[-1][0] + seg_len,
seg_ratio,
seg_compress,
)
else:
new_segment_info.append((seg_len, seg_ratio, seg_compress))
return new_segment_info
def __get_context_prob(
self,
context_list: list,
token_to_word="mean",
force_tokens: List[str]=[],
token_map: dict={},
force_reserve_digit: bool=False,
):
chunk_list = []
for chunks in context_list:
for c in chunks:
chunk_list.append(c)
dataset = TokenClfDataset(
chunk_list, tokenizer=self.tokenizer, max_len=self.max_seq_len
)
dataloader = DataLoader(
dataset, batch_size=self.max_batch_size, shuffle=False, drop_last=False
)
chunk_probs = []
chunk_words = []
with torch.no_grad():
for batch in dataloader:
ids = batch["ids"].to(self.device, dtype=torch.long)
mask = batch["mask"].to(self.device, dtype=torch.long) == 1
outputs = self.model(input_ids=ids, attention_mask=mask)
loss, logits = outputs.loss, outputs.logits
probs = F.softmax(logits, dim=-1)
for j in range(ids.shape[0]):
_probs = probs[j, :, 1]
_ids = ids[j]
_mask = mask[j]
active_probs = torch.masked_select(_probs, _mask)
active_ids = torch.masked_select(_ids, _mask)
tokens = self.tokenizer.convert_ids_to_tokens(
active_ids.squeeze().tolist()
)
token_probs = [prob for prob in active_probs.cpu().numpy()]
(
words,
valid_token_probs,
valid_token_probs_no_force,
) = self.__merge_token_to_word(
tokens,
token_probs,
force_tokens=force_tokens,
token_map=token_map,
force_reserve_digit=force_reserve_digit,
)
word_probs_no_force = self.__token_prob_to_word_prob(
valid_token_probs_no_force, convert_mode=token_to_word
)
if "xlm-roberta-large" in self.model_name:
for i in range(len(words)):
words[i] = words[i].lstrip("▁")
chunk_words.append(words)
chunk_probs.append(word_probs_no_force)
prev_idx = 0
context_probs = []
context_words = []
for chunk_list in context_list:
n_chunk = len(chunk_list)
context_probs.append([])
context_words.append([])
for i in range(n_chunk):
context_probs[-1].extend(chunk_probs[prev_idx + i])
context_words[-1].extend(chunk_words[prev_idx + i])
prev_idx = prev_idx + n_chunk
context_probs = [sum(probs) / len(probs) for probs in context_probs]
return context_probs, context_words
def __chunk_context(self, origin_text, chunk_end_tokens):
origin_list = []
origin_tokens = self.tokenizer.tokenize(origin_text)
n = len(origin_tokens)
st = 0
while st < n:
if st + self.max_seq_len > n - 1:
chunk = self.tokenizer.convert_tokens_to_string(origin_tokens[st:n])
origin_list.append(chunk)
break
else:
ed = st + self.max_seq_len
for j in range(0, ed - st):
if origin_tokens[ed - j] in chunk_end_tokens:
ed = ed - j
break
chunk = self.tokenizer.convert_tokens_to_string(
origin_tokens[st : ed + 1]
)
origin_list.append(chunk)
st = ed + 1
return origin_list
def __merge_token_to_word(self, tokens, token_probs, force_tokens, token_map, force_reserve_digit):
words = []
word_probs = []
word_probs_no_force = []
for token, prob in zip(tokens, token_probs):
if token in self.special_tokens:
continue
# add a new word
elif is_begin_of_new_word(token, self.model_name, force_tokens, token_map):
pure_token = get_pure_token(token, self.model_name)
prob_no_force = prob
if pure_token in force_tokens or pure_token in set(token_map.values()):
prob=1.0
token = replace_added_token(token, token_map)
words.append(token)
word_probs.append(
[
1.0
if force_reserve_digit
and bool(re.search(r"\d", token))
else prob
]
)
word_probs_no_force.append([prob_no_force])
# concatenate with previous token
else:
pure_token = get_pure_token(token, self.model_name)
words[-1] += pure_token
word_probs[-1].append(
1.0
if force_reserve_digit
and bool(re.search(r"\d", token))
else prob
)
word_probs_no_force[-1].append(prob_no_force)
return words, word_probs, word_probs_no_force
def __token_prob_to_word_prob(self, token_probs, convert_mode="mean"):
if convert_mode == "mean":
word_probs = [sum(p) / len(p) for p in token_probs]
elif convert_mode == "first":
word_probs = [p[0] for p in token_probs]
else:
raise NotImplementedError()
return word_probs
def __compress(
self,
context_list: list,
reduce_rate: float=0.5,
token_to_word: str="mean",
force_tokens: List[str]=[],
token_map: dict={},
force_reserve_digit: bool=False,
drop_consecutive: bool=False,
):
def split_string_to_words(input_string):
pattern = r'\b\w+\b|[<>=/!@#$%^&*()?":{}|\\`~;_+-]'
result = re.findall(pattern, input_string)
return result
# print(force_tokens, token_map, force_reserve_digit, drop_consecutive)
if reduce_rate <= 0:
words, word_labels = [], []
for i in range(len(context_list)):
chunk_list = context_list[i]
chunk_words = []
chunk_word_labels = []
for j in range(len(chunk_list)):
# replace to original token
for ori_token, new_token in token_map.items():
chunk_list[j] = chunk_list[j].replace(new_token, ori_token)
ws = split_string_to_words(chunk_list[j])
chunk_words.extend(ws)
chunk_word_labels.extend([1 for _ in range(len(ws))])
context_list[i] = "".join(chunk_list)
words.append(chunk_words)
word_labels.append(chunk_word_labels)
return context_list, words, word_labels
chunk_list = []
for chunks in context_list:
for c in chunks:
chunk_list.append(c)
dataset = TokenClfDataset(
chunk_list, tokenizer=self.tokenizer, max_len=self.max_seq_len
)
dataloader = DataLoader(
dataset, batch_size=self.max_batch_size, shuffle=False, drop_last=False
)
compressed_chunk_list = []
word_list = []
word_label_list = []
with torch.no_grad():
for batch in dataloader:
ids = batch["ids"].to(self.device, dtype=torch.long)
mask = batch["mask"].to(self.device, dtype=torch.long) == 1
outputs = self.model(input_ids=ids, attention_mask=mask)
loss, logits = outputs.loss, outputs.logits
probs = F.softmax(logits, dim=-1)
for j in range(ids.shape[0]):
chunk_probs = probs[j, :, 1]
chunk_ids = ids[j]
chunk_mask = mask[j]
active_probs = torch.masked_select(chunk_probs, chunk_mask)
active_ids = torch.masked_select(chunk_ids, chunk_mask)
tokens = self.tokenizer.convert_ids_to_tokens(
active_ids.squeeze().tolist()
)
token_probs = [prob for prob in active_probs.cpu().numpy()]
words, valid_token_probs, _ = self.__merge_token_to_word(
tokens=tokens,
token_probs=token_probs,
force_tokens=force_tokens,
token_map=token_map,
force_reserve_digit=force_reserve_digit,
)
word_probs = self.__token_prob_to_word_prob(
valid_token_probs, convert_mode=token_to_word
)
if drop_consecutive:
threshold = np.percentile(word_probs, int(100 * reduce_rate))
is_token_between = False
prev = None
for i, (word, word_prob) in enumerate(zip(words, word_probs)):
if word in force_tokens:
if is_token_between:
is_token_between = False
elif not is_token_between and word == prev:
word_probs[i] = 0.0
prev = word
else:
is_token_between |= word_prob > threshold
# calculate compression ratio w.r.t. gpt-4 tokenizer
new_token_probs = []
for word, word_prob in zip(words, word_probs):
num_token = len(self.oai_tokenizer.encode(word))
new_token_probs.extend([word_prob for _ in range(num_token)])
threshold = np.percentile(
new_token_probs, int(100 * reduce_rate + 1)
)
keep_words = []
word_labels = []
assert len(words) == len(word_probs)
for word, word_porb in zip(words, word_probs):
if word_porb > threshold:
if (
drop_consecutive
and word in force_tokens
and len(keep_words) > 0
and keep_words[-1] == word
):
word_labels.append(0)
else:
keep_words.append(word)
word_labels.append(1)
else:
word_labels.append(0)
keep_str = self.tokenizer.convert_tokens_to_string(keep_words)
if "xlm-roberta-large" in self.model_name:
for i in range(len(words)):
words[i] = words[i].lstrip("▁")
compressed_chunk_list.append(keep_str)
word_list.append(words[:])
word_label_list.append(word_labels[:])
compressed_context_list = []
original_word_list = []
original_word_label_list = []
prev_idx = 0
for chunk_list in context_list:
n_chunk = len(chunk_list)
compressed_context_list.append(
"".join(compressed_chunk_list[prev_idx : prev_idx + n_chunk])
)
original_word_list.append([])
original_word_label_list.append([])
for i in range(n_chunk):
original_word_list[-1].extend(word_list[prev_idx + i])
original_word_label_list[-1].extend(word_label_list[prev_idx + i])
prev_idx = prev_idx + n_chunk
return compressed_context_list, original_word_list, original_word_label_list