SharpToken

SharpToken is a C# library that serves as a port of the Python tiktoken library. It provides functionality for encoding and decoding tokens using GPT-based encodings. This library is built for .NET 6, .NET 8 and .NET Standard 2.0, making it compatible with a wide range of frameworks.

Important

The functionality in SharpToken has been added to Microsoft.ML.Tokenizers. Microsoft.ML.Tokenizers is a tokenizer library being developed by the .NET team and going forward, the central place for tokenizer development in .NET. By using Microsoft.ML.Tokenizers, you should see improved performance over existing tokenizer library implementations, including SharpToken. A stable release of Microsoft.ML.Tokenizers is expected alongside the .NET 9.0 release (November 2024). Instructions for migration can be found at https://github.com/dotnet/machinelearning/blob/main/docs/code/microsoft-ml-tokenizers-migration-guide.md.

Installation

To install SharpToken, use the NuGet package manager:

Install-Package SharpToken

Or, if you prefer using the .NET CLI:

dotnet add package SharpToken

For more information, visit the NuGet package page.

Usage

To use SharpToken in your project, first import the library:

using SharpToken;

Next, create an instance of GptEncoding by specifying the desired encoding or model:

// Get encoding by encoding name
var encoding = GptEncoding.GetEncoding("cl100k_base");

// Get encoding by model name
var encoding = GptEncoding.GetEncodingForModel("gpt-4");

You can then use the Encode method to encode a string:

var encoded = encoding.Encode("Hello, world!"); // Output: [9906, 11, 1917, 0]

And use the Decode method to decode the encoded tokens:

var decoded = encoding.Decode(encoded); // Output: "Hello, world!"

SharpToken also provides a high performance count method. It is usefull to check prompt size before sending it to a LLM or to use it in a TextSplitter/Chunker for RAG.

var count = encoding.CountTokens("Hello, world!"); // Output: 4

Supported Models

SharpToken currently supports the following models:

r50k_base
p50k_base
p50k_edit
cl100k_base
o200k_base

You can use any of these models when creating an instance of GptEncoding:

var r50kBaseEncoding = GptEncoding.GetEncoding("r50k_base");
var p50kBaseEncoding = GptEncoding.GetEncoding("p50k_base");
var p50kEditEncoding = GptEncoding.GetEncoding("p50k_edit");
var cl100kBaseEncoding = GptEncoding.GetEncoding("cl100k_base");
var o200kBaseEncoding = GptEncoding.GetEncoding("o200k_base");

Model Prefix Matching

Apart from specifying direct model names, SharpToken also provides functionality to map model names based on specific prefixes. This allows users to retrieve an encoding based on a model's prefix.

Here are the current supported prefixes and their corresponding encodings:

Model Prefix	Encoding
`gpt-4o`	`o200k_base`
`gpt-4-`	`cl100k_base`
`gpt-3.5-turbo-`	`cl100k_base`
`gpt-35-turbo`	`cl100k_base`

Examples of model names that fall under these prefixes include:

For the prefix gpt-4o: gpt-4o, gpt-4o-2024-05-13, etc.
For the prefix gpt-4-: gpt-4-0314, gpt-4-32k, etc.
For the prefix gpt-3.5-turbo-: gpt-3.5-turbo-0301, gpt-3.5-turbo-0401, etc.
For the Azure deployment name gpt-35-turbo.

To retrieve the encoding name based on a model name or its prefix, you can use the GetEncodingNameForModel method:

string encodingName = Model.GetEncodingNameForModel("gpt-4-0314");  // This will return "cl100k_base"

If the provided model name doesn't match any direct model names or prefixes, the method will return null.

Understanding Encoded Values

When you encode a string using the Encode method, the returned value is a list of integers that represent tokens in the specified encoding. These tokens are a compact way of representing the input text and can be processed more efficiently by various algorithms.

For example, encoding the text "Hello world!" using the cl100k_base encoding might produce the following list of integers:

var encoded = cl100kBaseEncoding.Encode("Hello world!"); // Output: [9906, 1917, 0]

You can then use the Decode method to convert these tokenized integer values back into the original text:

var decoded = cl100kBaseEncoding.Decode(encoded); // Output: "Hello world!"

With SharpToken, you can seamlessly switch between different encodings to find the one that best suits your needs. Just remember to use the same encoding for both the Encode and Decode methods to ensure accurate results.

Advanced usage

Custom Allowed Sets

SharpToken allows you to specify custom sets of allowed special tokens when encoding text. To do this, pass a HashSet containing the allowed special tokens as a parameter to the Encode method:

const string encodingName = "cl100k_base";
const string inputText = "Some Text <|endofprompt|>";
var allowedSpecialTokens = new HashSet<string> { "<|endofprompt|>" };

var encoding = GptEncoding.GetEncoding(encodingName);
var encoded = encoding.Encode(inputText, allowedSpecialTokens);
var expectedEncoded = new List<int> { 8538, 2991, 220, 100276 };

Assert.Equal(expectedEncoded, encoded);

Custom Disallowed Sets

Similarly, you can specify custom sets of disallowed special tokens when encoding text. Pass a HashSet<string> containing the disallowed special tokens as a parameter to the Encode method:

const string encodingName = "cl100k_base";
const string inputText = "Some Text";

var encoding = GptEncoding.GetEncoding(encodingName);

void TestAction()
{
    encoding.Encode(inputText, disallowedSpecial: new HashSet<string> { "Some" });
}

Assert.Throws<ArgumentException>(TestAction);

In this example, an ArgumentException is thrown because the input text contains a disallowed special token

Testing and Validation

SharpToken includes a set of test cases in the TestPlans.txt file to ensure its compatibility with the Python tiktoken library. These test cases validate the functionality and behavior of SharpToken, providing a reliable reference for developers. Running the unit tests and verifying the test cases helps maintain consistency between the C# SharpToken library and the original Python implementation.

Performance Compared to TiktokenSharp and TokenizerLib

SharpToken is the fastest library with the lowest allocations!

Benchmark Code

[SimpleJob(RuntimeMoniker.Net60)]
[SimpleJob(RuntimeMoniker.Net80)]
[SimpleJob(RuntimeMoniker.Net471)]
[RPlotExporter]
[MemoryDiagnoser]
public class CompareBenchmark
{
    private GptEncoding _sharpToken;
    private TikToken _tikToken;
    private ITokenizer _tokenizer;
    private Tokenizer _mlTokenizer;
    private string _kLongText;

    [GlobalSetup]
    public async Task Setup()
    {
        _sharpToken = GptEncoding.GetEncoding("cl100k_base");
        _tikToken = await TikToken.GetEncodingAsync("cl100k_base").ConfigureAwait(false);
        _tokenizer = await TokenizerBuilder.CreateByModelNameAsync("gpt-4").ConfigureAwait(false);
        _kLongText = "King Lear, one of Shakespeare's darkest and most savage plays, tells the story of the foolish and Job-like Lear, who divides his kingdom, as he does his affections, according to vanity and whim. Lear’s failure as a father engulfs himself and his world in turmoil and tragedy.";
    }

    [Benchmark]
    public int SharpToken()
    {
        var sum = 0;
        for (var i = 0; i < 10000; i++)
        {
            var encoded = _sharpToken.Encode(_kLongText);
            var decoded = _sharpToken.Decode(encoded);
            sum += decoded.Length;
        }

        return sum;
    }

    [Benchmark]
    public int TiktokenSharp()
    {
        var sum = 0;
        for (var i = 0; i < 10000; i++)
        {
            var encoded = _tikToken.Encode(_kLongText);
            var decoded = _tikToken.Decode(encoded);
            sum += decoded.Length;
        }

        return sum;
    }

    [Benchmark]
    public int TokenizerLib()
    {
        var sum = 0;
        for (var i = 0; i < 10000; i++)
        {
            var encoded = _tokenizer.Encode(_kLongText);
            var decoded = _tokenizer.Decode(encoded.ToArray());
            sum += decoded.Length;
        }

        return sum;
    }

    [Benchmark]
    public int MLTokenizers()
    {
        var sum = 0;
        for (var i = 0; i < 10000; i++)
        {
            var encoded = _mlTokenizer.EncodeToIds(_kLongText);
            var decoded = _mlTokenizer.Decode(encoded);
            sum += decoded.Length;
        }

        return sum;
    }
}

BenchmarkDotNet v0.13.9+228a464e8be6c580ad9408e98f18813f6407fb5a, Windows 11 (10.0.22631.3296)
11th Gen Intel Core i9-11950H 2.60GHz, 1 CPU, 16 logical and 8 physical cores
.NET SDK 9.0.100-preview.2.24157.14
  [Host]               : .NET 8.0.3 (8.0.324.11423), X64 RyuJIT AVX2
  .NET 6.0             : .NET 6.0.28 (6.0.2824.12007), X64 RyuJIT AVX2
  .NET 8.0             : .NET 8.0.3 (8.0.324.11423), X64 RyuJIT AVX2
  .NET Framework 4.7.1 : .NET Framework 4.8.1 (4.8.9181.0), X64 RyuJIT VectorSize=256

Method	Job	Runtime	Mean	Error	StdDev	Median	Gen0	Gen1	Allocated
MLTokenizers	.NET 8.0	.NET 8.0	60.55 ms	1.143 ms	1.123 ms	60.45 ms	1000.0000	-	13.12 MB
MLTokenizers	.NET 6.0	.NET 6.0	95.75 ms	1.374 ms	1.147 ms	95.54 ms	10500.0000	-	126.19 MB
MLTokenizers	.NET Framework 4.7.1	.NET Framework 4.7.1	291.77 ms	5.811 ms	11.195 ms	291.64 ms	21000.0000	-	127.33 MB

SharpToken	.NET 8.0	.NET 8.0	87.78 ms	1.700 ms	1.590 ms	87.34 ms	1000.0000	-	22.13 MB
SharpToken	.NET 6.0	.NET 6.0	128.84 ms	1.718 ms	1.607 ms	128.17 ms	16250.0000	500.0000	196.31 MB
SharpToken	.NET Framework 4.7.1	.NET Framework 4.7.1	356.21 ms	6.843 ms	10.854 ms	355.09 ms	34000.0000	1000.0000	204.39 MB

TokenizerLib	.NET 8.0	.NET 8.0	109.26 ms	2.082 ms	4.482 ms	107.90 ms	18200.0000	600.0000	217.82 MB
TokenizerLib	.NET 6.0	.NET 6.0	126.16 ms	2.959 ms	8.630 ms	122.34 ms	18000.0000	500.0000	217.82 MB
TokenizerLib	.NET Framework 4.7.1	.NET Framework 4.7.1	374.71 ms	7.374 ms	16.794 ms	370.12 ms	40000.0000	1000.0000	243.79 MB

TiktokenSharp	.NET 8.0	.NET 8.0	177.34 ms	3.506 ms	8.797 ms	174.98 ms	28000.0000	1000.0000	338.98 MB
TiktokenSharp	.NET 6.0	.NET 6.0	196.17 ms	3.912 ms	8.422 ms	195.52 ms	26000.0000	666.6667	313.26 MB
TiktokenSharp	.NET Framework 4.7.1	.NET Framework 4.7.1	488.22 ms	9.696 ms	15.931 ms	487.17 ms	63000.0000	1000.0000	378.31 MB

Performance

SharpToken is extreamly performance optimized on net8.0. It uses modern multibyte CPU instructions and almost no heap allocations.

All core methods have been tested on a large and a small input text.

Inputs:

SmallText: 453 B (text/plain)
LargeText: 51 KB (text/html)

Methods:

Encode: text to tokens
Decode: tokens to text
CountTokens: high performance API to count tokens of text

BenchmarkDotNet v0.13.12, Windows 11 (10.0.22631.3296/23H2/2023Update/SunValley3)
AMD Ryzen 9 3900X, 1 CPU, 24 logical and 12 physical cores
.NET SDK 8.0.200
  [Host]               : .NET 8.0.2 (8.0.224.6711), X64 RyuJIT AVX2
  .NET 6.0             : .NET 6.0.16 (6.0.1623.17311), X64 RyuJIT AVX2
  .NET 8.0             : .NET 8.0.2 (8.0.224.6711), X64 RyuJIT AVX2
  .NET Framework 4.7.1 : .NET Framework 4.8.1 (4.8.9181.0), X64 RyuJIT VectorSize=256

Method	Mean	Error	StdDev	Ratio	RatioSD	Allocated	Alloc Ratio
.NET 8.0
Encode_SmallText	22.649 us	0.4244 us	0.4359 us	0.28	0.01	696 B	0.02
Encode_LargeText	4,542.505 us	87.7988 us	104.5182 us	0.24	0.01	155547 B	0.03

Decode_SmallText	1.623 us	0.0324 us	0.0373 us	0.44	0.02	2320 B	0.98
Decode_LargeText	454.570 us	6.8980 us	6.4524 us	0.80	0.02	286979 B	1.00

CountTokens_SmallText	22.008 us	0.1165 us	0.0909 us	0.28	0.00	184 B	0.005
CountTokens_LargeText	4,231.353 us	14.5157 us	11.3329 us	0.23	0.00	195 B	0.000

.NET 6.0
Encode_SmallText	36.370 us	0.7178 us	1.0962 us	0.45	0.02	37344 B	0.91
Encode_LargeText	11,213.070 us	219.6291 us	269.7243 us	0.59	0.02	5062574 B	0.91

Decode_SmallText	2.588 us	0.0394 us	0.0350 us	0.70	0.02	2320 B	0.98
Decode_LargeText	489.467 us	8.9195 us	8.3433 us	0.86	0.02	286985 B	1.00

CountTokens_SmallText	34.758 us	0.2027 us	0.1896 us	0.45	0.01	36832 B	0.907
CountTokens_LargeText	11,252.083 us	215.8912 us	212.0340 us	0.61	0.01	4907169 B	0.907

.NET Framework 4.7.1
Encode_SmallText	79.947 us	1.5621 us	3.0097 us	1.00	0.00	41138 B	1.00
Encode_LargeText	18,961.252 us	253.1816 us	236.8262 us	1.00	0.00	5567685 B	1.00

Decode_SmallText	3.723 us	0.0728 us	0.0997 us	1.00	0.00	2375 B	1.00
Decode_LargeText	570.787 us	11.0356 us	11.8080 us	1.00	0.00	287496 B	1.00

CountTokens_SmallText	77.521 us	1.0802 us	0.9020 us	1.00	0.00	40616 B	1.000
CountTokens_LargeText	18,485.392 us	313.5834 us	277.9836 us	1.00	0.00	5413237 B	1.000

Contributions and Feedback

If you encounter any issues or have suggestions for improvements, please feel free to open an issue or submit a pull request on the project's repository.

Hope you find SharpToken useful for your projects and welcome any feedback you may have.

dmitry-brazhenko/SharpToken