This is not an LLM obviously
, it's just for generating random names. But interesting to think of the possibilities of truly tiny language models if there were connected together.
Yes, a quick back of the envelope math is 0.65 * (memory bandwidth of the card / (model weights in bytes + kv cache in bytes) ~ practical decode tps. Below context around 32k (depends upon the model but again can be used as a placeholder number) you can ignore the kv cache in bytes and the math becomes just about memory bandwidth and model weights in bytes.
You could think of it as a standard decoder only LLM (almost all modern ones we use everyday), with some layers (experts) having parallel networks and conditionally based on the input token (per token) - the token is routed through some of these layers. In the case of a non MoE (dense) - each token goes through all layers, so the inference engine has to read all the layers and do a matrix (layer) times vector (token) computation, while in the case of MoE the number of layers per token that has to do the compute is substantially lesser, so one can expect much higher tps than a dense model at the same number of parameters (size - 7B, 27B etc)
Honestly not sure this is impressive. I ported microgpt to zig as a learning exercise, then moved scalar engines to NEON/metal just to see what happened. Besides metal being slower (I probably did something wrong, but it could be due to the fixed costs of memory transfer into the GPU not being worth it due to the small model).
Anyways, it was also stupid fast, particularly compared to the python version. But I was pretty sure that's irrelevant to real production architectures!
And it's only using AVX-2 and not AVX-512, AMX or ACE. Or built-in GPUs and NPUs (the M series doesn't emphasize matrix multiplication on the CPU side because it already has matrix multiplication units on the GPU, which is always attached).
Before the M5, there was no dedicated matrix multiplication hardware on the Apple Silicon GPU. Their solution was generally using the NPU and AMX coprocessors for tensor and matrix workloads.
It's a trivial example. This won't be useful outside of a VERY specific domain without more parameters. Many people need to know about the bitter lesson.
The point here is that the library's overhead cost is very low. The fact that a tiny model can reach 10M tokens per second means that the overhead of token decode, memory allocation, calling the model, etc. is very low. The model doesn't actually need to be useful to prove that point.
It’s interesting and worthy of genuine applaud for being a good starting point for further work.
That said, I am more interested in what size model this could manage while producing “just enough” tokens per second to work at a conversational rate. What are models in that class capable of doing for me?
Thank you for your kind reply. I appreciate your point completely and while I tried to moderate sounding dismissive of what was being done here, I think I could have done better.
I love "trivial" examples and everything you've said is tue.
I think the bitter lesson only talks about task performance but not computational efficiency. Could tiny models improve efficiency? Maybe by just using a general architecture on specialized data, so the artichecture itself is not task specific?
What sense of the word "atomic" is meant here?
Anyways, it was also stupid fast, particularly compared to the python version. But I was pretty sure that's irrelevant to real production architectures!
Am I reading this right? Then I need to try this on Strix Halo
https://en.wikipedia.org/wiki/Bitter_lesson
Over time, I'm sure we'll be able to filter information better and get parameter counts down, but I wouldn't count on that within the next 6 months.
That said, I am more interested in what size model this could manage while producing “just enough” tokens per second to work at a conversational rate. What are models in that class capable of doing for me?
I love "trivial" examples and everything you've said is tue.