Performance compared to TiktokenSharp, Tokenizer

[lofcz]

Performance compared to TiktokenSharp, Tokenizer

[vwilson]

I have a branch using RegexOptions.Compiled in several places. I think this will improve performance in long-lived instances of the GptEncoding class, but it actually runs the extant unit tests slightly slower than the main branch. I'm still playing with it and have several MB of public domain novels to crunch. Maybe a consumer-level option to use RegexOptions.Compiled in a couple places at least would be called for. I will let you know when I have finished.

[Anurag-RTS]

Ran a quick encode benchmark with 5500 words input:

BenchmarkDotNet=v0.13.5, OS=Windows 10 (10.0.19045.3086/22H2/2022Update)
AMD Ryzen 7 1700, 1 CPU, 16 logical and 8 physical cores
.NET SDK=7.0.203
  [Host]     : .NET 7.0.5 (7.0.523.17405), X64 RyuJIT AVX2
  DefaultJob : .NET 7.0.5 (7.0.523.17405), X64 RyuJIT AVX2

Method	Mean	Error	StdDev
TokenizerEncode	450.5 us	4.61 us	4.31 us
TikTokenEncode	504.1 us	3.82 us	2.99 us
SharpTokenEncode	2,704.0 us	21.60 us	20.20 us

[dmitry-brazhenko]

Here is an updated benchmarks that I calclulated. Will check what can be improved

Method	Categories	Data	Mean	Median	Ratio
SharpTokenV1_0_28_Encode	Encode	1. (...)57. [19866]	3,374,121.7 ns	3,227,934.8 ns	1.00
TiktokenSharpV1_0_5_Encode	Encode	1. (...)57. [19866]	3,415,080.4 ns	3,399,812.1 ns	1.11
TokenizerLibV1_3_2_Encode	Encode	1. (...)57. [19866]	2,060,565.1 ns	2,027,039.6 ns	0.63
Tiktoken_Encode	Encode	1. (...)57. [19866]	921,540.6 ns	895,316.2 ns	0.28

[dmitry-brazhenko]

There is a new version that has a significant improvements thanks to @r-Larch 086544d

[dmitry-brazhenko]

Hi @lofcz !

Thanks for reaching out!

I planned to implement some performance improvements and compare it with other libs. But not yet. If you have any ideas of improvements, please let me know and I will use them for improvement :)

[dmitry-brazhenko]

Hey @UX-Ayoma !

THanks a lot for sharing!
I will explore what can be improved in the lib. I would say that there are lot of stuff to be done.

I started with some easy and added some optimisation for Decoding (#9).
I will explore and implement fixes for Encoding as well.

[dmitry-brazhenko]

Fixed encoding here: PR #11.

Will update with benchmarks and more improvements later.