My read was that it meant "small", since they already say "dependency-free" in the second half of the sentence. But it's also a strange way to say "small" so I'm still not sure.
This is pretty cool. Implemented something similar myself (a really small language model with ~10M params) just to teach myself the ML behind the LLMs. Did not implement it in C obviously, just use PyTorch, but it's interesting to go through the c file and see how he has implemented stuff I took for granted in Python in C.
Anyway, I just tested this out myself on my AMD Ryzen 9 9800x3d. I got 7647173 tok/sec using karpathy's Shakespeare dataset https://raw.githubusercontent.com/karpathy/char-rnn/master/d.... Going to play around with it and see if I can get a CUDA kernal built to see what it could do on a 5090. Claude estimates with napkin math that we could get around 2B tok/s
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.
I think that is a sort of a reverse scaling fallacy. Given the right resources and environments, many small models can function together in an emergent way. I’ve been on the lookout for an SLM version of Conway’s game of life. SLM always reminds me of slime molds, which demonstrate a form of intelligence which is remarkable.
> Given the right resources and environments, many small models can function together in an emergent way.
What is this based on? Every researcher I've heard talk about this says it's exactly not true, as an uncontested rule, because the larger models will more effectively contain the smaller models, and use them together in ways that the connections between the smaller models can't. Remember, even MOE is to save compute/memory, not to help performance/parameter.
That is my understanding as well. Thousands of monkeys do not equal or surpass a man, intellectually, even if working together. There is some intrinsic super linear scaling in intelligence.
You're comparing completely different training data, harness overhead, and cost function. You're also comparing brains, which aren't really related to this discussion at all.
But, it depends on what you're measuring. By spatial/navigational memory, yes, elephants are far far better. Reasoning, no. It would be interesting to see what an elephant or whale eugenics program could result in, since humans have that pesky (or maybe instrumental?) birth canal problem.
This is not a good comparison, because the brain doesn't only do "being smart" - it has to do things like innervate muscle and other tissue through the body, elephants will require more neurons given their larger size, just to be able to *walk*
LLMs are an imprecise, more of a marketing term, to define Transformer models based on the self-attention mechanism, trained with massives amounts of data.
And this implements a transformer. Actually it is a very cool didactic example.
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.
Parameter size and total number of parameters so ultimately the total size of the model in memory.
This leads to some interesting optimizations. You can quantize all the parameters (or certain layers) of a model and halve or quarter the memory requirement but maintain most of the model's intelligence. This increases the token rate inversely with the size reduction.
Popular quantizations for local models are 8-bit and 4-bit parameter sizes. The Blackwell series of nVidia chips now even support native FP4 math making 4-bit quantizations even faster.
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).
No, but if you are guaranteed to have a GPU or NPU packaged together it becomes like the SPUs in the Cell processor - you can’t offload instructions (unless you use a trap mechanism to a subroutine) but you can have code that hides the setup, the different ISA, and the result retrieval, behind an API call.
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?
Is it that impressive? It is a model generating short strings from scratch, so I do not think there is significant token parsing or memory allocation going on.
In compute performance, 10M tok/s * 4096 parameter/tok * 4 byte/param = ~160 GiB/sec implied memory bandwidth, vs nominal ~300 GB/sec for the M5 Pro's RAM. That seems even less notable once we consider that 4096 parameter * 4 byte/param = 16 KiB of parameters fit easily within L1 data cache (the M5's efficiency cores each have 64 KiB of L1 data cache). 40 GFLOPs means ~10 FLOP/cycle, about 2.5 NEON instructions per cycle given that the core instructions are fused multiply-adds (FMAs). A random web page I found says the M5 family cores have a 4-wide SIMD block, meaning this gets about 63% utilization: respectable but not super high.
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?
reply