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.
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.
(different circumstances have different nuances about what "better" means, it isn't always performance optimization; some do substitute "faster" for "better" here, but I think it loses generality then).
I always heard the "Make it Right" as "Make it Beautiful", where Right and Beautiful would mean "non-hacky, easily maintainable, easily extendable, well tested, and well documented"
My mentor used say it is the difference between a screw and glue.
You can glue some things together and prove that it works, but eventually you learn that anytime you had to break something to fix it, you should've used a screw.
It is trade off in coupling - the glue binds tightly over the entire surface but a screw concentrates the loads, so needs maintenance to stay tight.
You only really know which is "right" it if you test it to destruction.
All of that advice is probably sounding date now, even in material science the glue might be winning (see the Tesla bumper or Lotus Elise bonding videos - every screw is extra grams).
Making it work can be a hacky, tech debt laden implementation. Making it right involves refactoring/rewriting with an eye towards maintainability, testability, etc etc
In ML, often it does work to a degree even if it's not 100% correct. So getting it working at all is all about hacking b/c most ideas are bad and don't work. Then you'll find wins by incrementally correcting issues with the math / data / floating point precision / etc.
Depends on your definition of "right" and "work". It could be a big ball of mud that always returns exactly the required response (so it 'works'), but be hellish hard change and very picky about dependencies and environment (so it's not 'right').
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.
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.
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.
So great there are 8 of them. 800% better than all the rest!
> If you think Python is a bad language for AI integrations, try writing one in a compiled language.
I'll take this challenge, all day, every day, so long as I and the hypothetical 'move fast and break things' have equal "must run in prod" and "must be understandable by some other human" qualifiers
What type is `array`? Don't worry your pretty head about it, feed it whatever type you want and let Sentry's TypeError sort it out <https://github.com/openai/whisper/blob/v20250625/whisper/aud...> Oh, sorry, and you wanted to know what `pad_or_trim` returns? Well that's just, like, your opinion man
Tracks with me, I don't like using Python for real programming. Try explaining any of your "Python sucks" catechisms to a second-year statistics student though. If you'd rather teach them C++, be my guest. If you want to make them indebted to proprietary infra like Mojo or CUDA, knock yourself out.
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.
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.
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.
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.
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.
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.
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.
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.
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?
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.
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
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.
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?
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.
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!
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.
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.
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.)
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
"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!
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.
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.
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.
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".
I've usually heard/said it as
(different circumstances have different nuances about what "better" means, it isn't always performance optimization; some do substitute "faster" for "better" here, but I think it loses generality then).i like this version best
I always heard the "Make it Right" as "Make it Beautiful", where Right and Beautiful would mean "non-hacky, easily maintainable, easily extendable, well tested, and well documented"
I've always heard it (and said it) as:
What's the difference between make it work and make it right? Aren't they the same thing?
> make it work and make it right?
My mentor used say it is the difference between a screw and glue.
You can glue some things together and prove that it works, but eventually you learn that anytime you had to break something to fix it, you should've used a screw.
It is trade off in coupling - the glue binds tightly over the entire surface but a screw concentrates the loads, so needs maintenance to stay tight.
You only really know which is "right" it if you test it to destruction.
All of that advice is probably sounding date now, even in material science the glue might be winning (see the Tesla bumper or Lotus Elise bonding videos - every screw is extra grams).
Making it work can be a hacky, tech debt laden implementation. Making it right involves refactoring/rewriting with an eye towards maintainability, testability, etc etc
Yeah, if it's not right, it doesn't work.
In ML, often it does work to a degree even if it's not 100% correct. So getting it working at all is all about hacking b/c most ideas are bad and don't work. Then you'll find wins by incrementally correcting issues with the math / data / floating point precision / etc.
Depends on your definition of "right" and "work". It could be a big ball of mud that always returns exactly the required response (so it 'works'), but be hellish hard change and very picky about dependencies and environment (so it's not 'right').
Nope, it's right, but it's not pretty.
Not true. Things can work with hacks. Your standards might consider it unacceptable to have hacks. So you can have a “make it right” stage.
Fair chance I'm remembering it wrong :D
The Huggingface transformers lib is currently undergoing a refactor to get rid of cruft and make it more extensible, hopefully with some perf gains.
A similar concept dates back to 30BC: https://en.wikipedia.org/wiki/De_architectura
Firmitas, utilitas, venustas - Strong, useful, and beautiful.
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.
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.
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.
> has a great package ecosystem
So great there are 8 of them. 800% better than all the rest!
> If you think Python is a bad language for AI integrations, try writing one in a compiled language.
I'll take this challenge, all day, every day, so long as I and the hypothetical 'move fast and break things' have equal "must run in prod" and "must be understandable by some other human" qualifiers
What type is `array`? Don't worry your pretty head about it, feed it whatever type you want and let Sentry's TypeError sort it out <https://github.com/openai/whisper/blob/v20250625/whisper/aud...> Oh, sorry, and you wanted to know what `pad_or_trim` returns? Well that's just, like, your opinion man
Tracks with me, I don't like using Python for real programming. Try explaining any of your "Python sucks" catechisms to a second-year statistics student though. If you'd rather teach them C++, be my guest. If you want to make them indebted to proprietary infra like Mojo or CUDA, knock yourself out.
I'm still teaching them Python.
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.
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.
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++?
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.
There’s something beautiful about creating a drop in replacement for something that improves performance substantially.
ScyllaDB comes to mind
Agreed. I figured nobody would use it otherwise.
Put it in there readme & description. It's a big selling point.
Thanks, I clarified it.
To be fair, many people have token stabbing needs.
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.
Alright, 0.1.1 should now be a true drop-in replacement. I'll write up some examples soon.
Ah good catch. Updating this right now.
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.
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.
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.
A language that isn’t memory-safe can definitely hurt. AI needs more security, not less.
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.
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.
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?
Cool!
I've reached out to the guy who maintains Tiktoken to talk about this.
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-...
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.
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
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
I'm working on incremental re-tokenizing next. Then I'll run some benchmarks against this crate too.
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?
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."
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...
I thought Gemini used SentencePiece
https://github.com/google/sentencepiece
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
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
> 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?
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...
Just curious whether it's possible to push any of your performance improvements to tiktoken itself?
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.
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.)
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.
Very cool. We use Tiktoken and I'd love to see the performance impact. Pretty great decision to make it drop-in compatible.
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
Should be the same. Both use Byte-Pair Encoding (BPE) as underlying algo.
this is still the outdated architecture without special tokens for numbers like out-of-vocab tokens like NUM_FLOAT(3.1415) right?
I'm relieved to see that its not written in rust
haha, I thought about it.
What about pairing this with BigBird and Mamba?
Now that byte-patch-level embeddings are discovered?
"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!
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
[dead]
[flagged]
You know what's also faster to roughly get the amount of tokens? string.length/5
It is not helpful in actual tokenization, though.