Why we write our own C and C++ inference engines
localai.io
localai.io
1. The comparison of vLLM to vLLM.cpp never shows more improvement than a handful of tokens per second. That's less than 0.05x improvement on every run, and the gap doesn't grow as concurrency increases. The comparison doesn't show any real net improvement, let alone justify the project.
2. The depth anything comparison isn't apples to apples. Of course a q8_0 quant is faster than f32. It's 4x less data to chew on.
3. This point is silly, it again fights against the argument that writing your own c++ engines are worth it. It's a bug, just fix it:
> The reason it is faster has nothing to do with writing better matmul kernels than PyTorch. Two positional embeddings, the DPT head’s UV embedding and the backbone’s bicubic position embedding, were being recomputed on every forward pass with single-threaded scalar sin, cos and bicubic loops, even though they depend only on the input geometry and are identical every call.
They argue against their own point again just after:
> For a biometric pipeline, matching the reference exactly matters more than being faster than it.
Okay, then don't rewrite it! It's not faster anyway!
4. If the argument is that the venv is large, then rewriting it in C++ doesn't seem like the answer, it seems like a lot of work and maintenance to avoid having to cull unreachable files in your venv. In a past life I maintained a simple denylist for files in node_modules. The low hanging fruit is plentiful and generally very safe.
Having said that if you know of some magic bullet to fix the problems with the Python ecosystem (that vLLM and Pytorch are apparently unaware of) then I will be happy to try it because our CI is constantly blocked by Python related issues.
> Deploying a Python inference stack means resolving a dependency tree at install time, on the target machine, against whatever CUDA and glibc that machine has. Deploying a ggml port means copying a shared library and a GGUF file.
For open-source code to get adoption and support, the theory of its construction needs to be understood by humans because at the end of the day, some human will be responsible for its operation. I would call this the Terrence Tao principle: Output of AI is only as useful as how well it can be digested by humans.
vLLM.cpp's runtime is pure C++, but AOT Triton compilation is done at build time in Python.
While the page looks vibe-coded[1], the content itself does not have any AI tells. What are the tells you are seeing?
[1] Too many sites I find on HN frontpage these days slow my PC to a crawl. I assume they are all using the same autogenerated HTML, Javascrip and CSS to make animated backgrounds :-( On this specific site scrolling is laggy.
"depth estimation that beats PyTorch on CPU in half the memory" — "…beats X in Y…"
"Most LocalAI backends wrap somebody else’s engine, and that is the right default." — "…and that is the right"
"MLX and the rest are maintained by people who are better at those models than we are" — "better at those models than we are" — it's this thing that LLMs do where they are kind of weirdly confident but overly deferential
"This post is about what those ports buy" — "…buy" used in this context
"Same model, 1.31x the speed" — "Same X, something Y" — it's this overconfident yet deferential writing style
The further I read, the more tells there are. I find it incredibly tiring to read LLM generated prose and I'm not sure why. Is it because I'm aware it's not human written and have an unconscious bias? Or is it because the style is just full-on, "Not X but Y. Those performance gains are bought, not earned. This stops, that starts. Read on, or don't, that's the follow-up"
Its like listening to an insufferable clever dick, who is not as bright as they think they are. You would also find it incredibly irritating if a human talked like that
It always reads like a 14 year old arguing.
I think it's because of the laggy scrolling that I didn't read the whole thing anyway, just the first few screens.
Stop. Seek help.
> We are ahead at all six points, and five of those six are ties. Our run-to-run noise band is 0.5%, and concurrency 2 through 32 land between 0.7% and 1.7%, so the honest reading is that only the single-stream case (4.5%) is clearly outside noise. Output is token-for-token identical to vLLM at every point on that curve. Peak host memory is 24.88 GiB against 28.18 GiB.
I get the motives but it's tiring. I myself sometimes check my own (unpublished) writing or have friends preview it and the same feedback comes out, it's all hand written.
People on HN have accused me of sounding like an AI one time or another, English is my second language and sometimes my ADD goofs my writing into something that sounds like gibberish even though I fully understand what I wrote others might not, so yeah.
"blland", "llame",... ?
if you've done any amount of optimizaation or hill-climbing work (kernel optimization, autoresearch type shit), you would know that these models LOVE the terms "buy" and "cost" in this sense. they use them nearly constantly (along with "budget", and even "credit" (even when working with RL/credit assignment!))
> The method, the measurements, and what it costs us.
> That is the general shape of these wins.
> Parity is the gate, speed is the follow-up
I could go on and on, but you probably get the point. If you don't find anything funny with the above, you might have not been enough-exposed to slop.
And okay, your second argument is that I just don't know slop because I am not exposed to it? But you don't know anything about me or what I am exposed.
You're just making random claims and stating they are correct without any evidence or arguments.
This post is incredibly obviously AI generated, to the extent that I doubt a human author edited it at all. Not "written with AI assistance" but full on "give Claude some bullets and hit publish." It contains tons of tropes that show up in all AI writing and which people are highlighting here.
What would convince you of that?
A big problem with the existing engines like llama or sd is that they don’t support optimal graph compilation. Usually this means about a real 2 or 3x multiplier loss relative to optimal. Cuda graphs do okay but they still leave a lot on the floor
It’s usually worth it to optimize in that context if you are willing to peer into the mechanics.
Of course that’s expensive. You need to know how to appropriately pipeline and merge your kernels.
Wasm size from 30Mb to 300kb and 1.5x speedup. It's definitely worth it for performance or distribution size.
But the cpp port of vllm looks great, that'd be great if you'll maintain that. I hit the same limitations with vllm.
Well, it's clearly not 1-shot / low effort.
(can we agree there's a difference between low effort 1-shotting a prompt and those who use LLMs interactively?)
The one reason why I can't take a mega hardline stance against AI for writing: Not everyone speaks English as a first language.
People with awful English speaking and writing skills are using LLMs to contribute content in English. This is a good thing. They used to try their best and suffix every post with "sorry for my english".
The author's name is Ettore Di Giacinto - I put 2 and 2 together on this one.
I give a pass when non-native English speakers are just using it to write better English and keep up with the rest of us. It wasn't super obvious anyway, it wasn't interruptive.
The content is beyond that, and I more than welcome stuff like this here.