Sadly it didn't. It talked about "perplexities" and showed some floating point numbers.
I want to see examples like "here's a prompt against a model and the same prompt against a quantized version of that model, see how they differ."
I want to see examples like "here's a prompt against a model and the same prompt against a quantized version of that model, see how they differ."
The closer we get to assessing human like performance, the tougher it is, because it becomes more subjective and less deterministic by the nature of the task. I don't know the answer, but I know that for the metrics we have it's not so easy to translate them into any idea about the kind of performance on some specific thing you might want to do with the model.
I think this right here is the answer to measuring and comparing model performance.
Instead of trying to compare models holistically, we should be comparing them for specific problem sets and use cases... the same as we compare humans against one another.
Using people as an example, a hiring manager doesn't compare 2 people holistically, they compare 2 people based on how well they're expected to perform a certain task or set of tasks.
We should be measuring and comparing models discriminately rather than holistically.
Would setting the temperature to zero degrade the quality of response?
For an llm, a trivial change in wording could produce a big change in answer, same as running it again with a new random seed. "Prompt engineering" is basically overfitting if not approached methodically. For example, it would be interesting to try deliberate permutations of an input that don't change the meaning and see how the answer changes as part of an evaluation.
Look at this minimal implementation (Karpathy's) of LLaMA, the only randomness is in the "sample" function that comes in at non-zero temperature, otherwise its easy to see everything is deterministic: https://github.com/karpathy/llama2.c/blob/master/run.c
Otoh, with MoE like GPT-4 has, it can still vary at zero temperature.
Subjective ratings are different, but for compression things are quite well defined.
That seems simplistic, but its really simple as that. Naive 3 bit quantization will turn llama 7B into blubbering nonsense.
But llama.cpp quantization is good! I recommend checking out the graphs ikawrakow made for their K-quants implementation:
https://github.com/ggerganov/llama.cpp/pull/1684
Basically, the more you quantize with K-quant, the dumber the model gets. 2 bit llama 13B quant, for instance, is about as dumb as 7B F16, but the dropoff is not nearly as severe from 3-6 bits.
The dropdown at the top selects which comparison: Falcon compares GGML, Vicuna compares bits and bytes. I have some more comparisons planned, feel free to open an issue if you'd like to see something specific: https://github.com/the-crypt-keeper/can-ai-code
e.g. take a look at these two rows:
llama-65b.ggmlv3.q4_K_M.bin 4.90639 llama.cpp
llama-65b.ggmlv3.q3_K_M.bin 5.01299 llama.cpp
If you want more fine-grained explanations of how generation of certain types of texts get clobbered, you would probably need to prepare datasets comprised of that type of string, and measure the perplexity delta on that subset. i.e.
dperplexity/dquantization(typed_inputs).
dtype/dquantization(all_outputs).What is insane though is how far we've taken it. I remember when INT8 from NVIDIA seemed like a nigh-pipedream!
"But what we found with these neural networks is, if you use 32 bits, they're just fine. And then you use 16 bits, and they're just fine. And then with eight bits, you need to use a couple of tricks and then it's just fine.
And now we find if you can go to four bits, and for some networks, that's much easier. For some networks, it's much more difficult, but then you need a couple more tricks. And so it seems they're much more robust."
That will be really interesting for FPGAs, because the current ones are basically oceans of 4-bit computers.
Yes, you can gang together a pair of 4LUTs to make a 5LUT, and a pair of 5LUTs to make a 6LUT, but you halve your parallelism each time you do that. OTOH you can't turn a 4LUT into a pair of 3LUTs on any currently-manufactured FPGA. It's simply the "quantum unit" of currently-available hardware -- and it's been that way for at least 15 years (Altera had 3LUTs back in the 2000s). There's no fundamental reason for the number 4 -- but it is a very, very deep local minimum for the current (non-AI) customers of FPGA vendors.
I'm not sure what you mean by this. You've always been able to run GPT3 on a single server (your typical 8xA100).
And basically all servers will have 8xA100 (maybe 4xA100). Nobody bothers with a single A100 (of course in a VM you might have access to only one)
for those wondering: no this is not the norm. My lab at CMU doesn't own any A100s (we have A6000s).
It doesn’t really make sense to BTO. What you gain economically you lose in the science you can do.
But nobody could have anticipated this.
Please take an extra ten seconds to speak in proper human language!
You could save on the worlds carbon footprint by reducing the number of times humans have to search for “what is NVIDIA hgx?” or is it “what is AMD HGX” and then subsequently visiting the websites to see if that’s right or not.
However, there’s a difference between an acronym known to the broader public versus some single shot, context-specific one!
wishing, or guessing something, without actual experience, doesn't make it true.
DGX boxes do not grow on trees, especially these days
The interesting implication about running an LLM on a single board computer is that if it's a proof of concept for an LLM on a smartphone. If you have a model that can produce useful results on a Ras Pi, you have something that could potentially run on hundreds of millions of smartphones. I'm not sure what the use case is for running an LLM on your phone instead of the cloud, but it opens some interesting possibilities. It depends just how useful such a small LLM could be.
I would elaborate and say, anywhere that your computer is resource constrained ( ram, processing power ) but you still want to make up articles for your Amazon Affiliate blog
Is it: the weights all happen to be where float is sparse, so quantization ends up increasing fidelity? Or is it more of a “worse is better” dropout-type situation?
https://oobabooga.github.io/blog/posts/perplexities/
Essentially, you lose some accuracy and there might be some weird answers and probably more likely to go off the rail and hallucinate. But the quality loss is lower the more parameters you have. So for very large model sizes the differences might be negligible. Also, this is the cost of inference only. Training is a whole other beast and requires much more power.
Still, we are looking at GPT3 level of performance on one server rack. That says something when less than a year ago, such AI was literally magic and only run on a massive datacenter. Bandwidth and memory size are probably, in my ignorance mind, easier to increase than raw compute so maybe we will soon actually have "smart" devices.