Show HN: TokenDagger – A tokenizer faster than OpenAI's Tiktoken

Show HN: TokenDagger – A tokenizer faster than OpenAI's Tiktoken(github.com)

281 points by matthewolfe 1 year ago | 73 comments

TokenDagger is a drop-in replacement for OpenAI’s Tiktoken (the tokenizer behind Llama 3, Mistral, GPT-3.*, etc.). It’s written in C++ 17 with thin Python bindings, keeps the exact same BPE vocab/special-token rules, and focuses on raw speed.

I’m teaching myself LLM internals by re-implementing the stack from first principles. Profiling TikToken’s Python/Rust implementation showed a lot of time was spent doing regex matching. Most of my perf gains come from a) using a faster jit-compiled regex engine; and b) simplifying the algorithm to forego regex matching special tokens at all.

Benchmarking code is included. Notable results show: - 4x faster code sample tokenization on a single thread. - 2-3x higher throughput when tested on a 1GB natural language text file.

npalli 1 year ago |

Kudos, I think (in the short term at least) there is a large amount of perf. optimization to be found by coding parts of the whole AI/ML infrastructure in C++ like this one, not as a rewrite (god no!) but drop in and fix key bottlenecks. Anytime I see someone (seems Chinese engineers are good at this) put something out in C++, good chance some solid engineering tradeoffs have been made and dramatic improvement will be seen.

matthewolfe 1 year ago | |

Agreed. A former mentor of mine told me a nice way of viewing software development:

1. Make it work. 2. Make it fast. 3. Make it pretty.

Transformers & LLMs have been developed to a point where they work quite well. I feel as though we're at a stage where most substantial progress is being made on the performance side.

diggan 1 year ago | | |

Heh, seems people I've been learning from been biased away from beauty, as I know that as "Make It Work, Make It Right, Make It Fast".

binarymax 1 year ago | | |

The Huggingface transformers lib is currently undergoing a refactor to get rid of cruft and make it more extensible, hopefully with some perf gains.

jotux 1 year ago | | |

A similar concept dates back to 30BC: https://en.wikipedia.org/wiki/De_architectura

Firmitas, utilitas, venustas - Strong, useful, and beautiful.

saretup 1 year ago | |

And while we’re at it, let’s move away from Python altogether. In the long run it doesn’t make sense just because it’s the language ML engineers are familiar with.

tbalsam 1 year ago | | |

No! This is not good.

Iteration speed trumps all in research, most of what Python does is launch GPU operations, if you're having slowdowns from Pythonland then you're doing something terribly wrong.

Python is an excellent (and yes, fast!) language for orchestrating and calling ML stuff. If C++ code is needed, call it as a module.

bigyabai 1 year ago | | |

It makes plenty of sense. Python handles strings well, has a great package ecosystem, and is easy to write/learn for non-programmers. It can be easily embedded into a notebook (which is huge for academics) and is technically a "write once run anywhere" platform in theory. It's great.

If you think Python is a bad language for AI integrations, try writing one in a compiled language.

janalsncm 1 year ago | | |

Most of that is already happening under the hood. A lot of performance-sensitive code is already written in C or cython. For example numpy, scikit learn, pandas. Lots of torch code is either C or CUDA.

ML researchers aren’t using python because they are dumb. They use it because what takes 8 lines in Java can be done with 2 or 3 (including import json) in python for example.

ipsum2 1 year ago | |

Sort of. The key bottlenecks are not in tokenization, but running the actual CUDA kernels. Python actually has very little overhead. (See VLLM, which is primarily in Python). So when people (like deepseek) 'rewrite in C++', they're usually just rewriting CUDA kernels to be more efficient.

notatallshaw 1 year ago | |

It looks like TikToken is written in Rust (https://github.com/openai/tiktoken/tree/main/src), are the gains here actually from porting to C++?

fhub 1 year ago | | |

From the post

Profiling TikToken’s Python/Rust implementation showed a lot of time was spent doing regex matching. Most of my perf gains come from a) using a faster jit-compiled regex engine; and b) simplifying the algorithm to forego regex matching special tokens at all.

chrismustcode 1 year ago |

There’s something beautiful about creating a drop in replacement for something that improves performance substantially.

ScyllaDB comes to mind

matthewolfe 1 year ago | |

Agreed. I figured nobody would use it otherwise.

parhamn 1 year ago | | |

Put it in there readme & description. It's a big selling point.

pvg 1 year ago | | |

To be fair, many people have token stabbing needs.

pama 1 year ago |

Cool. Would it be possible to eliminate that little vocab format conversion requirement for the vocab I see in the test against tiktoken? It would be nice to have a fully compatible drop in replacement without having to think about details. It also would be nice to have examples that work the other way around: initialize tiktoken as you normally would, including any specialized extension of standard tokenizers, and then use that initialized tokenizer to initialize a new tokendagger and test identity of results.

matthewolfe 1 year ago | |

Alright, 0.1.1 should now be a true drop-in replacement. I'll write up some examples soon.

matthewolfe 1 year ago | |

Ah good catch. Updating this right now.

superlopuh 1 year ago |

Can someone familiar with performance of LLMs please tell me how important this is to the overall perf? I'm interested in looking into optimizing tokenizers, and have not yet run the measurements. I would have assumed that the cost is generally dominated by matmuls but am encouraged by the reception of this post in the comments.

refibrillator 1 year ago | |

Tokenization is typically done on CPU and is rarely (if ever) a bottleneck for training or inference.

GPU kernels typically dominate in terms of wall clock time, the only exception might be very small models.

Thus the latency of tokenization can essentially be “hidden”, by having the CPU prepare the next batch while the GPU finishes the current batch.

serjester 1 year ago | |

Tokenizing text is ridiculously small part of the overall computation that goes into serving a request. With that said if you’re doing this on petabytes of data, never hurts to have something faster.

odyssey7 1 year ago | | |

A language that isn’t memory-safe can definitely hurt. AI needs more security, not less.

matthewolfe 1 year ago | |

To echo the other replies, the tokenizer is definitely not the bottleneck. It just happens to be the first step in inference, so it's what I did first.

benreesman 1 year ago | |

Tokenization performance is complicated, but your guidepost is that the institutions with the resources and talent to do so choose to write extremely fast tokenizers: sentencepiece and tiktoken both pay dearly in complexity (particularly complexity of deployment because now you've got another axis of architecture-specific build/bundle/dylib to manage in addition to whatever your accelerator burden always was: its now aarch64 cross x86_64 cross CUDA capability...)

Sometimes it can overlap with accelerator issue, but pros look at flame graphs: a CPU core running the AVX lanes hard isn't keeping the bus fed, million things. People pre-tokenize big runs all the time.

I don't know why this thread is full of "nothing to see here", this obliterates the SOTA from the money is no object status quo: I'd like to think better of the community than the obvious which is that C++ is threatening a modest mindshare comeback against a Rust narrative that's already under pressure from the explosion of interest in Zig. Maybe there's a better reason.

kevmo314 1 year ago |

Nice work! I tried something similar a while back ago: https://github.com/kevmo314/tokie

The takeaway I also found was that the running cost was really dominated by pretokenization (the regex). It's cool to see that you found a faster way to run the regex, but have you tried comparing the performance of just swapping out the regex engine and leaving the actual BPE to tiktoken? I wonder if that is upstreamable?

matthewolfe 1 year ago | |

Cool!

I've reached out to the guy who maintains Tiktoken to talk about this.

22c 1 year ago | |

There is at least some awareness already when it comes to the performance of the regex engine:

https://github.com/openai/tiktoken/blob/main/src/lib.rs#L95-...

Tiberium 1 year ago |

Can you also compare the performance with https://github.com/huggingface/tokenizers/? Would be helpful, since the benchmark in the tiktoken readme seems to be very outdated.

binarymax 1 year ago | |

Anecdotally I've always found tiktoken to be far slower than huggingface tokenizers. I'm not sure why, as I haven't dug into tiktoken, but I'm a heavy user of HF's rust tokenizers

p0 1 year ago |

How does this compare to the BPE crate [1]? Its main selling point is support for incrementally re-tokenising text, but it's also faster than tiktoken.

[1] https://crates.io/crates/bpe

matthewolfe 1 year ago | |

I'm working on incremental re-tokenizing next. Then I'll run some benchmarks against this crate too.

frabcus 1 year ago |

Is there any way we can get local tokenizers for other LLMs? e.g. Gemini only offer a remote API for their tokenizer. Is it proprietary? Could we infer the token mapping somehow efficiently by making lots of calls?

Deathmax 1 year ago | |

Gemini uses SentencePiece [1], and the proprietary Gemini models share the same tokenizer vocabulary as Gemma [2, 3, 4].

Out of the large proprietary western AI labs (OpenAI, Anthropic, Google), only Anthropic with Claude 3 and newer lack local tokenizers.

[1] https://github.com/google/sentencepiece

[2] https://github.com/googleapis/python-aiplatform/blob/main/ve...

[3] https://storage.googleapis.com/deepmind-media/gemma/gemma-2-...: "We inherit from the large Gemini vocabulary (256k entries)."

[4] https://storage.googleapis.com/deepmind-media/gemma/Gemma3Re...: "We use the same tokenizer as Gemini 2.0."

matthewolfe 1 year ago | |

A lot of model-specific tokenizers have reference implementations ([0], [1]). Underlying them is a core algorithm like SentencePiece or Byte-pair encoding (BPE). Tiktoken and TokenDagger are BPE implementations. The wrapping "tokenizer" mostly deals with the quirks of the vocabulary and handling special tokens.

For this project, I think there is value in building some of these model-specific quirks into the library. Could see some minor performance gains and generally make it easier to integrate with. It's probably not too much work to keep up with newer models. Tokenizers change much less frequently.

[0] https://github.com/meta-llama/llama-models/blob/01dc8ce46fec...

[1] https://github.com/mistralai/mistral-common/tree/main/src/mi...

weberer 1 year ago | |

I thought Gemini used SentencePiece

https://github.com/google/sentencepiece

fkyoureadthedoc 1 year ago |

Would be cool to see WASM bindings for this here https://github.com/dqbd/tiktoken

Or maybe even your speedups from "b" in the pure js implementation

justinhj 1 year ago |

I've been playing with tokenization too. Starting from Kaparthy's Python minbpe I set myself the task of training a tokenizer on wikitext (500mb) in a reasonable time. I got the C++ version down to about 50 minutes compared to the original Python code (estimated) several months.

Haven't really spent much time looking at encode and decode but I plan to incorporate these regex modifications when I do!

https://github.com/justinhj/minbpe-cc

konsalexee 1 year ago |

> simplifying the algorithm to forego regex matching special tokens at all

Does that mean there could be cases with less quality in terms of tokenization?

matthewolfe 1 year ago | |

The output should be identical, assuming no bugs.

The Tiktoken implementation takes a collection of all special tokens upon initialization and compiles them into a regex by joining them with `|` [0]. Then the actual encoding process checks for matches on this expression.

Models like Llama 4 define a list of 1,135 special tokens. Notably, 1,115 of those are "reserved" special tokens! So this yields a huge regexp of special tokens that shouldn't be considered at all.

TokenDagger does not do this. Instead, simple string matching is used. This works because we don't need to consider the entire special vocabulary every time. The caller of `encode` must explicitly define which special tokens should be considered [1]. So it's faster to check against the much smaller list we _know_ is being used.

[0] https://github.com/openai/tiktoken/blob/main/src/lib.rs#L476

[1] https://github.com/openai/tiktoken/blob/main/tiktoken/core.p...

anonymoushn 1 year ago | | |

Isn't this incorrect? If the user doesn't specify what to do with almost all of the special tokens, you still must detect them so you can raise an error.

pamelafox 1 year ago |

Just curious whether it's possible to push any of your performance improvements to tiktoken itself?

matthewolfe 1 year ago | |

I probably will. Was hesitant initially, because adding PCRE2 as a dependency might cause issues to existing projects. I believe this was discussed briefly in a closed PR with other performance improvements.

polynomial 1 year ago |

Just to note that Tiktoken is still the tokenizer behind the GPT-4x series, it just uses a different token model. (Post only says GPT-3, implying they were using something else for subsequent iterations.)

manishsharan 1 year ago |

Is there a tokenizer someone can recommend for code ? I have tried CodeBert but maybe I am using it wrong as my results with it were pretty bad.

isjustintime 1 year ago |

Very cool. We use Tiktoken and I'd love to see the performance impact. Pretty great decision to make it drop-in compatible.

matrix2596 1 year ago |

is is possible for your tokenizer to give different tokenization ever then openai tokenizer? i am asking because there are multiple ways to tokenize the same string?? sry if i am mistaken

matthewolfe 1 year ago | |

Should be the same. Both use Byte-Pair Encoding (BPE) as underlying algo.

semiinfinitely 1 year ago |

I'm relieved to see that its not written in rust

matthewolfe 1 year ago | |

haha, I thought about it.

singularity2001 1 year ago |

this is still the outdated architecture without special tokens for numbers like out-of-vocab tokens like NUM_FLOAT(3.1415) right?

EGreg 1 year ago |

What about pairing this with BigBird and Mamba?

sheerun 1 year ago |

Now that byte-patch-level embeddings are discovered?

silentsea90 1 year ago |

"I’m teaching myself LLM internals by re-implementing the stack from first principles." - curious what resources you're using? Any books or courses, or just building it straight up? Great work!

matthewolfe 1 year ago | |

Modal's GPU glossary is a good overview about how GPUs work [0]. Karpathy's LLM overview is a good high level overview on LLMs [1]. 3b1b's video (and subsequent videos) on transformers was excellent at helping me understand the math at a high level [2]. This matrix multiplication optimization worklog helped me understand writing better CUDA (not for beginner intro though) [3].

During this process I also asked ChatGPT a lot of questions.

I'm definitely open to suggestions about "how to learn" with all the new tools we have. I felt this has not been straightforward to figure out.

[0] https://modal.com/gpu-glossary

[1] https://www.youtube.com/watch?v=7xTGNNLPyMI

[2] https://www.youtube.com/watch?v=wjZofJX0v4M

[3] https://siboehm.com/articles/22/CUDA-MMM

janwilmake 1 year ago |

You know what's also faster to roughly get the amount of tokens? string.length/5

_flux 1 year ago | |

It is not helpful in actual tokenization, though.