mratsim
gpt-oss-120b: (unknown prefill), ~175 tok/s generation. I don't remember the prefill speed but it certainly was below 10k
Nemotron-3-Super: 14070 tok/s prefill, ~194.5 tok/s generation. (Tested fresh after reload, no caching, I have a screenshot.)
Nemotron-3-Super using NVFP4 and speculative decoding via MTP 5 tokens at a time as mentioned in Nvidia cookbook: https://docs.nvidia.com/nemotron/nightly/usage-cookbook/Nemo...
Also checked and apparently Nvidia Cutlass now supports generic convolutions: https://github.com/NVIDIA/cutlass
And BLS signatures are standards as well: https://www.ietf.org/archive/id/draft-irtf-cfrg-bls-signatur...
Similarly KZG polynomial commitments are generic, the Ethereum specific thing is the trusted setup ceremony which needs to be specialized per protocol.
The finite field backends and the elliptic curves supported are all generic as well: https://github.com/mratsim/constantine/blob/master/constanti...
Then the library defines its own types SecretBool and SecretWord throughough library that cannot be mixed with regular Nim types and are based on uint32/uint64 so the compiler cannot assume 0/1 and do boolean logic.
Assembly is used to avoid compilers getting too smart and reintroducing branches but the no-assembly fallback should still compile without branches.
By convention, all variable-time cryptographic functions are suffixed _vartime.
There is a plan to add automated constant-time checks as well https://github.com/mratsim/constantine/issues/7.
People don't have to comply to GDPR but if they want to serve EU folks then they don't have a choice.
Nim supports variadic generic, it's an arbitrary limitation so that shape and stride small vectors that describe a tebsor can be stack-allocated and fit in a cacheline.
Edit: it can be configured via a compile-time flag to 8 or 10 or anything: https://github.com/mratsim/Arraymancer/blob/master/src%2Farr...
And lots of funding for then vaporware
I go over some examples here: https://github.com/mratsim/constantine/issues/358#issuecomme...
- https://github.com/mratsim/constantine/wiki/Constant-time-ar...
Passwords, credentials and other high-risk high value intangible data are not associated with high-risk high-value AND our memory is not designed for random junk.
Maybe an anthropologist can study this phenomenon further.
We already have hints of that: our memory has evolved for thousands year to easily memorize places and spatial navigation, in thr past for food, water and danger. Today you go in a foreign place, just one visit and you know where everything is. At the same time, people can usually hold only ~7 concepts/numbers/objects in their head.
In memory competitions people leverage our spatial prowess with a technique called the memory palace, that was already used in ancient Greece and Roman Empire to recall anything, from bard tales to the Iliad and Odysseus.
I've been quite disappointed by French LLMs on Huggingface when I tried a month ago.
You can predict what your month will look like based on your rent, utilities and how expensive are your groceries from the past months.
The unpredictable is accident and sickness but that's what insurances and emergency funds are for.
- Cuda: https://github.com/mratsim/weave/issues/133 - OpenCL: https://github.com/mratsim/weave/issues/134
If ChatGPT doesn't have product-market fit, what actually has?
The issue here is the complexity of complying and closing all loopholes that would allow the government to bring the full weight of the legal system against a library writer.
To fully comply with this you would need as a library provider to fully KYC your clients so that there is a firewall between their US and non-US entities, and that travelling people don't bring out an encryption library at the same time.
It would be a operational nightmare.