Roxxik

I mean this is what MoE is the first step of. Slice your intelligence to make it more efficient without losing much capability.

CUA seems to disaggregate. Train specialist models and make delegation explicit. But I think the future for capable but efficient models involves doing it all in one package. I have no idea how that architecture would look like or how to train it.

Like a human would, I can walk and talk, without thinking hard about walking and still not stumbling. Specialized sub-circuits that interact. Not via a text protocol, but by directly interacting.

Did you see the Astra plays Portal 2 videos? You can see how separate viewing a scene, understanding the scene and acting in the scene still are. And in this model of interaction, a very capable model still makes stupid mistakes.

It is remarkable, that it works at all. But it is neither effective nor efficient. And in my opinion the real revolution comes not by capability, but by integration.

So to tie the loop: I think CUA makes an interesting first step towards specialization, but I am not sure if that is the logical conclusion of how AI should work.

I'd say recreating an existing website as is, is the easier task. Doing a look alike, even if the models do take some shortcuts on the way is doable right now. You have a template frontend to compare to. Now if it would scale to the number of users is a different ask and harder to validate for the models.

Building something new is where they struggle most.

My recommendation if you are willing to push it further is to keep KV and always needed weights pinned in memory, so they do not get evicted. This is likely already the case, as they are touched on each token. mmap is slow on evicted pages, as it does not load the whole tensor, but only the touched pages. And it does this through a page fault, thus blocking your code. So it loads a page hands control back to you and the code goes on to touch another evicted page, repeating the loop. Now on an HDD that is not a big problem (yes reads can be coalesced, but a HDD is fundamentally serial in reading) while an SSD can overlay reads better and it is good to keep a few reads in flight at all times to keep its queue fed.

One option is locking the pages. But for that size you need extra privileges.

I experimented with some options. For example: one problem with io_uring is that it still reads to page cache so your reads gets copied in memory after they landed. Now if you pass O_DIRECT that does not happen, but it has its own can of worms.

For full transparency: I had opus write the io_uring layer into llama.cpp for me. And it yielded something slightly short of a 2x tok/s speedup vs simple mmaping. Also I noticed that disabling the warmup and initial test dramatically increases startup time.

The medusa paper looks interesting. My work was a few months ago, multi-token decoding was not a thing then.

Not sure if mmapping is the right way. In my own tests I noticed that simple mmapping will produce many small reads and not keep the SSD queue saturated. So if RAM is large enough to cache most experts, that is a rounding error. But if the base weigths without experts fill more than half or RAM and you basically need to load in a few experts for each layer of each token, the latency gets important and mmap sadly blocks until the data is loaded. You can't do concurrent requests for multiple experts with mmap (but you know all the ones you need right after the router ran). And even going one step further, depending on the arch / the tensors you need, you could eagerly load some, start computing with them and load the rest of the expert tensors in the background (extra thread or async io) parallel to the compute. This is not really possible with mmap, even with madvise.

One further step is predicting which experts will be needed next token / next layer. LRU does this okish. But a learned projection from the hidden state can do better. Or even a simple correlation from past activated experts. Expert usage is heavily skewed.

I have prototyped something similar with ollama some months ago.

Do you mmap or issue reads on demand? Also do you use io_uring to interleave compute with io or do you spawn extra threads?

I also tried predicting which experts get reused and I managed to beat a simple LRU very slightly.

EDIT: That was on Kimi 2.5 but even worse quant than 4bit. IIRC it was 2.6 or so

IO is very bursty in these setups. When the router results are in you can start loading experts from SSD. In this brief moment the SSD is saturated.

Outside of that the SSD is idling.

Table 3 shows for K=4 experts an IO of 943 MB/Tok at 3.15 Tok/s giving an average IO of 2970 MB/s far below what the SSD could do.

I'm not sure, but not all expert weights are used immediately. Maybe they could do async reads for the down tensors parallelizing compute with IO.

Not sure if this works on Mac, I only tested my larger than RAM setup on Linux with io_uring O_DIRECT reads and I saw that about 20% of total reads do finish while my fused upgate matmul is already running.

Edit: Typos

Does an SSD meaningfully degrade by read only workloads?

Not only you understanding the how, but you not understanding the goal.

I often use AI successfully, but in a few cases I had, it was bad. That was when I didn't even know the end goal and regularly switched the fundamental assumptions that the LLM tried to build up.

One case was a simulation where I wanted to see some specific property in the convergence behavior, but I had no idea how it would get there in the dynamics of the simulation or how it should behave when perturbed.

So the LLM tried many fundamentally different approaches and when I had something that specifically did not work it immediately switched approaches.

Next time I get to work on this (toy) problem I will let it implement some of them, fully parametrize them and let me have a go with it. There is a concrete goal and I can play around myself to see if my specific convergence criterium is even possible.

They are explicitly not assuming anything about the content of the auxiliary space (full hard drive).

So the data might be incompressible and thus compressing it and restoring it afterwards would not work.

Edit: From the paper:

> One natural approach is to compress the data on the hard disk as much as possible, use the freed-up space for your computation and finally uncompress the data, restoring it to its original setting. But suppose that the data is not compressible. In other words, your scheme has to always work no matter the contents of the hard drive. Can you still make good use of this additional space?

So the trick is to do the computation forwards, but take care to only use reversible operations, store the result outside of the auxiliary "full" memory and then run the computation backwards, reversing all instructions and thus undoing their effect on the auxiliary space.

Which is called catalytic, because it wouldn't be able to do the computation in the amount of clean space it has, but can do it by temporarily mutating auxiliary space and then restoring it.

What I haven't yet figured out is how to do reversible instructions on auxiliary space. You can mutate a value depending on your input, but how do you use that value, since you can't assume anything about the contents of the auxiliary space and just overwriting with a constant (e.g. 0) is not reversible.

Maybe there is some xor like trick, where you can store two values in the same space and you can restore them, as long as you know one of the values.

Edit: After delving into the paper linked in another comment, which is rather mathy (or computer sciency in the original meaning of the phrase), I'd like to have a simple example of a program that can not run in it's amount of free space and actually needs to utilize the auxiliary space.

I'm not sure that the analogy stretches so far.

What even is an artificial neuron in an Artificial Neural Network executed on "normal" (non-neuromorphic) hardware? It is a set of weights and an activation function.

And you evaluate all neurons of a layer at the same time by multiplying their weights in a matrix by the incoming activations in a vector. Then you apply the activation function to get the outgoing activations.

Viewing this from a hardware perspective, there are no individual neurons, just matrix multiplications followed by activation functions.

I'm going out of my area of expertise here, I just started studying bioinformatics, but neurological neurons can't simply hold an activation because they communicate by depolarising their membrane. So they have to be spiking by their very nature of being a cell.

This depolarization costs a lot of energy, so they are incentived to do more with less activations.

Computer hardware doesn't have a membrane and thus can hold activations, it doesn't need spiking and these activations cost very little on their own.

So I'm not sure what we stand to gain from more complicated artificial neurons.

On the other hand, artificial neutral networks do need a lot of memory bandwidth to load in these weights. So an approach that better integrates storage and execution might help. If that is memristor tech or something else.

I think it is a little unusual, but nothing unheard of.

I started as a junior IT Consultant in Hamburg, Germany with a 3 month notice, but could only quit every half year. So for quitting at the end of june I had to give notice in march, and then the next opportunity was september for quitting at the end of december.

But since I'm going back to university now, we made a "Aufhebungsvertrag" so that I quit at the end of september. You have no right to it, but I have a good relation to my employer so they accepted.

EDITED for clarity

This looks like german to me, but i haven't ever seen these words used together. I thought it might be some old german and googled it, but searching "bewunderungeifersucht" just yields your comment and "bewunderungseifersucht", which i deemed more likely, yields nothing.

Disclaimer: I'm from germany