tokflow 1.1.0

LLM utility of streaming token realtime replacement processing

TokFlow

日本語

Utility that outputs tokens generated by a large language model (LLM) with sequential replacement processing

How it works

The tokens are entered one after the other as small pieces as shown below.

["He","llo"," ","t","h","ere","!<","N","L>m","y ","nam","e"," ","is"," tokfl","ow.","<","N","L>N","ice"," to ","me","et you."]

The input tokens are output, with <NL> replaced by \n each time.

You can specify any string to be replaced. Moreover, you can specify multiple replacement targets.

What is this library for?

I developed this for the purpose of outputting special tokens with successive replacements in sequential sentence generation using a large-scale language model, which is a generative AI, but it may also be used for other string stream processing.

Install

pip install tokflow

Usage

import time
from tokflow import TokFlow

TOKEN_GENERATOR_MOCK = ["He", "llo", " ", "t", "h", "ere", "!<", "N", "L>m", "y ", "nam", "e", " ", "is", " tokfl", "ow.",
                  "<", "N", "L>N", "ice", " to ", "me", "et you."]

# replace "<NL>" with "\n". "<NL>" is called "search target string".
# Multiple replacement conditions can be specified.
tokf = TokFlow([("<NL>", "\n")])

for input_token in TOKEN_GENERATOR_MOCK:

    output_token = tokf.put(input_token)

    # Input sequential tokens.
    # If there is a possibility that the token is a "search target string",
    # it is buffered for a while, so output_token may be empty for a while.
    print(f"{output_token}", end="", flush=True)

    # Included wait to show the sequential generation operation.
    time.sleep(0.3)


# Remember to output the remaining buffer at the very end. Buffers may be empty characters.
print(f"{tokf.flush()}", end="", flush=True)

Generation Options

The put method can take an optional parameter opts like put(text,opts).

opts can specify the format of the input and output, like {"in_type":"spot","out_type:"spot" }.

It behaves as follows:

in_type	out_type	Description
spot	spot	A mode that incrementally sends tokens to the put method, and outputs generated segments each time.
spot	full	A mode that incrementally sends tokens to the put method, but outputs the full sentence.
full	spot	A mode that sends the full sentence to the put method at once, but outputs generated segments each time.
full	full	A mode that sends the full sentence to the put method at once, and outputs the full sentence.

Notes:

• All text strings need to be sent to the put method before calling the flush method. Especially in full mode, all input strings are sent at once.
• If the output type (out_type) is full, the flush method must be called to obtain the final result.
• It's important to appropriately combine the call pattern of the put method and the use of the flush method to maintain consistency in each mode.

Code Example

Specify rules like condition = {"in_type": "full", "out_type": "full"}, and use condition as an argument for put and flush.

    tokf = TokFlow([("<NL>", "\n")])

    condition = {"in_type": "full", "out_type": "full"}
    prev_len = 0
    for input_token_base in get_example_texts():
        output_sentence = tokf.put(input_token_base, condition)

        print(f"output_sentence:{output_sentence}")

        if prev_len > len(output_sentence):
            raise ValueError("Length error")

        if "<NL>" in output_sentence:
            raise Exception("Failure Must be converted str found.")

        prev_len = len(output_sentence)

    output_sentence = tokf.flush(condition)

Processing

About Internal processing

Tokens are sequentially read in real time. The token read is combined with the tokens read so far, referred to as the "token buffer". In this sequential process, when a pre-specified string (hereafter referred to as the "search target string") appears in the token buffer, this string is replaced with another string (hereafter referred to as the "replacement string"). Since tokens are read sequentially, in the intermediate stage, a string that is unrelated to the search target string or part of the search target string accumulates in the token buffer. If the token buffer is composed in an order that cannot be a search target string, the token buffer is returned as the method's return value the moment such a determination is made. On the other hand, if the token buffer is composed in an order that could be a search target string, the return value remains an empty string until either the search target string appears or it is determined that it cannot be a search target string. In this way, by buffering until the appearance of the search target string, most sequential tokens can be displayed as they are, while replacement is delayed when necessary, enabling stream processing.