hckrnws

Hoping the author can answer, I'm still learning about how this all works. My understanding is that inference is "using the model" so to speak. How is this faster than established inference engines specifically on Mac? Are models generic enough that if you build e.g. an inference engine focused on AMD GPUs or even Intel GPUs, would they achieve reasonable performance? I always assumed because Nvidia is king of AI that you had to suck it up, or is it just that most inference engines being used are married to Nvidia?

I would love to understand how universal these models can become.

Basically “faster” means better performance e.g. tokens/s without loosing quality (benchmarks scores for models). So when we say faster we provide more tokens per second than llama cpp. That means we effectively utilize hardware API available (for example we wrote our own kernels) to perform better.

We also wrote our inference engine in rust for mixlayer, happy to answer any questions from those trying to do the same.

Looks like this uses ndarray and mpsgraph (which I did not know about!), we opted to use candle instead.

Have you added it to HomeBrew and other package managers yet?

Also any app deployed to PROD but developed on Mac need to be consistent i.e. work on Linux/in container.

Needs an "API key".

https://github.com/trymirai/uzu-swift?tab=readme-ov-file#qui...

How does this compare to https://github.com/Anemll/Anemll?

Can you explain the type of quantization you support?

would https://docs.unsloth.ai/basics/kimi-k2-how-to-run-locally be faster with mirai?

right now, we support AWQ but are currently working on various quantization methods in https://github.com/trymirai/lalamo

In practice, how often do the models use the ANE? It sounds like you are optimizing for speed which in my experience always favors GPU.

You're right, modern edge devices are powerful enough to run small models, so the real bottleneck for a forward pass is usually memory bandwidth, which defines the upper theoretical limit for inference speed. Right now, we've figured out how to run computations in a granular way on specific processing units, but we expect the real benefits to come later when we add support for VLMs and advanced speculative decoding, where you process more than one token at a time

VLMs = very large models?

Probably vision language models.

"trymirai", every time I hear the word Mirai I think of the large IOT DDoS botnet. Maybe it's just me though.

I think of the goofy Toyota fuel cell car. I think a grand total of about 6 have been sold (leased) in california

How does this bench compared to MLX?

I use MLX in lmstudio and it doesn't have whatever issues llama cpp is showing here.

Qwen3-0.6B at 5 t/s doesn't make any sense. Something is clearly wrong for that specific model.

I'm curious about why the performance gains mentioned were so substantial for Qwen vs Llama?

it looks like llama.cpp has some performance issues with bf16

What are the units on the benchmark results? I’m guessing higher is better?

yeah, tokens per second

Wow! Sounds super interesting

Amazing!

How was your experience using Rust on this project? I'm considering a project in an adjacent space and I'm trying to decide between Rust, C, and Zig. Rust seems a bit burdensome with its complexity compared to C and Zig. Reminds me of C++ in its complexity (although not as bad). I find it difficult to walk through and understand a complicated Rust repository. I don't have that problem with C and Zig for the most part.

But I'm wondering if I just need to invest more time in Rust. How was your learning curve with the language?

You are confusing familiarity with intrinsic complexity. I have 20 years experience with C/C++ before switching to rust a few years ago. After the initial hurdle, it is way easier and very simple to follow.

Somewhat faster on small models. Requires new format.

Not sure what the goal is for this project? Not seeing how this presents adequate benefits to get adopted by the community

It's utilizing Apple ANE and probably other optimization tools provided by Apple's framework. Not sure if llama.cpp uses them, but if they're not then the benchmark on GitHub says it all.

Written in Rust is a big one for me.

just curios, will it be supported on iOS, it would be great to build local llm app with this project.

already) https://github.com/trymirai/uzu-swift

>faster than llama cpp in all of the use cases

What's your deliberate, well-thought roadmap for achieving adoption similar to llama cpp?

Probably getting acquired by Apple :)

Ollama is the leader isn't it?

Brew stats (downloads last 30 days)

Ollama - 28,232 Lama.cpp - 7,826

that’s exactly we are looking for not to waste on apis. Wonder how significant trade offs are

I just spun up a AWS EC2 g6.xlarge instance to do some llm work. The GPU is NVIDIA L4 24GB and costs $0.8048/per hour. Starting to think about switching to an Apple mac2-m2.metal instance for $0.878/ per hour. Big question is the Mac instance only has 24GB of unified memory.

Unified memory doesn't compare to a Nvidia GPU, the latter is much better.

Just depends on what performance level you need.

nice

Wondering why use Rust other than C++

The recommendation from the security agencies is to prefer Rust over C++ as less risk of exploits.

Checked and Lama.cpp used C++ (obviously) and Llama uses Go.

Why use C++?

So C++ users don't need to learn something new.

Why use C++ for greenfield projects?

I wonder why they didn’t use Fortran.

...or D? or Go? or Java? C#? Zig? etc they chose what they were most comfortable with. Rust is fine, it's not for everyone clearly, but those who use it produce high quality software, I would argue similar with Go, without all the unnecessary mental overhead of C or C++