LLaMA now goes faster on CPUs

by lawrencechen • 2 years ago

1372 436 2 years ago

bottlepalm2• 2 years ago

I think it's a good idea for everyone to download and be able to run a LLM locally, even if you have the minimum of requirements. As a pseudo-backup of a large chunk of human knowledge.

TaylorAlexander2• 2 years ago

I contend that most human knowledge is not written down or if it is written down it’s not publicly available on the internet and so does not exist in these datasets.

There’s so much subtle knowledge like the way a mother learns to calm her child or the way a carpenter learns to work different kinds of wood which may be written down in part, but may also be learned through lived experience or transferred from human to human such that little of it gets written down and posted online.

mickdarling3• 2 years ago

Wait till all the videos ever created are tokenized and ingested into a training dataset. Carpentry techniques are certainly there. The subtleties of parenting maybe harder to derive from that, but maybe lots of little snippets of people’s lives will add up to a general understanding of parenting. There have certainly been bigger surprises in the field.

spacephysics2• 2 years ago

For sure agree, however as the storage of information evolves, it’s becoming more efficient over time

From oral tradition to tablets to scrolls to books to mass produced books to digital and now these LLMs, I think it’s still a good idea to preserve what we have the best we can. Not as a replacement, but a hedge against a potential library of Alexandria incident.

I could imagine a time in the near future where the models are domain-specific, and just like there are trusted encyclopedia publishers there are trusted model publishers that guarantee a certain level of accuracy.

It’s not like reading a book, but I for sure had an easier time learning golang talking with ChatGPT than a book

_ache_2• 2 years ago

I think you underestimate the amount of information contained in books and the extent to which our society (as a whole) depends on them.

skeledrew6• 2 years ago

I'd content that those are skills (gained through experience) rather than knowledge (gained through rote learning).

bamboozled4• 2 years ago

Yes but it contains enough hints to help someone find their way on the these types of tasks.

nicklecompte• 2 years ago

It's not even "human knowledge" that can't be written down - it seems all vertebrates understand causality, quantity (in the sense of intuitively understanding what numbers are), and object permanence. Good luck writing those concepts down in a way that GPT can use!

In general AI in 2024 is not even close to understanding these ideas, nor does any AI developer have a clue how to build an AI with this understanding. The best we can do is imitating object permanence for a small subset of perceptible objects, a limitation not found in dogs or spiders.

wruza2• 2 years ago

That's where humans suck. The classic "you're not doing it right" then proceeds to quickly show how to do it without verbalizing any info on learning process, pitfalls, failure modes, etc, as if just showing it was enough for themselves to learn. Most people do[n't do] that, not even a sign of reflection.

My worst case was with a guy who asked me to write an arbitrage betting bot. When I asked how to calculate coeffs, he pointed at two values and said "look, there <x>, there <y> thinks for a minute then it's <z>!". When I asked how exactly did he calculate it, he simply repeated with different numbers.

HarHarVeryFunny• 2 years ago

I contend that most human knowledge is not written down

Yes - the available training data is essentially mostly a combination of declarative knowledge (facts - including human-generated artifacts) and procedural knowledge (how to do things). What is missing is the learning process of taking a description of how to do something, and trying to apply that yourself in a specific situation.

No amount of reading books, or reading other people's blogs on how they did something, can avoid the need for hands-on experience if you want to learn how to do it yourself.

It's not just a matter of information that might be missing or unclear in instructional material, including how to cope with every type of failure and unexpected outcome, but crucially how to do this yourself - if you are to be the actor, then it's the predictive process in your mind that matters.

Partly for this reason, and partly because current AI's (transformer-based LLMs) don't support online learning (try & fail skill acquisition), I think we're going to see two distinct phases of AI.

1) The current "GenAI" phase where AI can only produce mash-ups of things it saw in it's pre-training data, augmented by similar "book learning" provided in-context which can be utilized by in-context learning. I'd characterize what this type of AI to be useful for, and capable of, as "automation". Applying that book (incl. anecdotal) knowledge to new situations where mash-up is all you need.

2) The second phase is where we have something closer to AGI, even if still below human level, which is no longer just a pre-trained transformer, but also has online learning and is agentic - taking actions predicated on innate traits like curiosity and boredom, so that given the book knowledge it can (& will!) then learn to apply that by experimentation/practice and learning from its own mistakes.

There will no doubt be advances beyond this "phase two" as well, but it seems we're likely to be stuck at "phase one" for a while (even as models become much better at phase one capabilities), until architectures fundamentally advance beyond transformers to allow this type of on-the-job training and skill acquisition.

texuf• 2 years ago

Any recommendations for the latest and greatest way to run these locally?

speps3• 2 years ago

llamafile as per TFA...

etc-hosts• 2 years ago

https://justine.lol/oneliners/

fragmede14• 2 years ago

ollama

slowmotiony• 2 years ago

I use a tool called LM Studio, makes it trivial to run these models on a Mac. You can also use it as a local API so it kinda acts like a drop-in replacement for the openAI API.

chown3• 2 years ago

I am the author of Msty [1]. My goal is to make it as straightforward as possible with just one click (once you download the app). If you end up trying it, I would love to hear your feedback.

1: https://msty.app

mikewarot2• 2 years ago

I don't see LLMs as a large chunk of knowledge, I see them as an emergent alien intelligence snapshotted at the moment it appeared to stop learning. It's further hobbled by the limited context window it has to use, and the probabilistic output structure that allows for outside random influences to pick its next word.

Both the context window and output structure are, in my opinion, massive impedance mismatches for the emergent intellect embedded in the weights of the model.

If there were a way to match the impedance, I strongly suspect we'd already have AGI on our hands.

bamboozled4• 2 years ago

What is alien about them ?

LLMs are of this earth and created by our species. Seems quite familiar to me.

namarie• 2 years ago

I can agree on the context windows, but what other output structure would you have?

mlsu• 2 years ago

Disagree. The input/output structure (tokens) is the interface for both inference and for training. There is an emergent intellect embedded in the weights of the model. However, it is only accessible through the autoregressive token interface.

This is a fundamental limitation, much more fundamental than appears at first. It means that the only way to touch the model, and for the model to touch the world, is through the tokenizer (also, btw, why tokenizer is so essential to model performance). Touching the world through a tokenizer is actually quite limited.

So there is an intelligence in there for sure, but it is locked in an ontology that is tied to its interface. This is even more of a limitation than e.g. weights being frozen.

gpm2• 2 years ago

If you want to download a backup of a large chunk of human knowledge... download wikipedia. It's a similar size to a small LLM and can actually distinguish between real life and fantasy: https://en.wikipedia.org/wiki/Wikipedia:Database_download

If you just want to play around with an LLM though, absolutely.

int_19h2• 2 years ago

Kiwix provides prepackaged highly compressed archives of Wikipedia, Project Gutenberg, and many other useful things: https://download.kiwix.org/zim/.

Between that and dirt cheap storage prices, it is possible to have a local, offline copy of more human knowledge than one can sensibly consume in a lifetime. Hell, it's possible to have it all on one's smartphone (just get one with an SD card slot and shove a 1+ Tb one in there).

CaptainOfCoit6• 2 years ago

actually distinguish between real life and fantasy

Are LLMs unable to distinguish between real life and fantasy? What prompts have you thrown at them to make this determination? Sending a small fairy tale and asking the LLM if it thinks it's a real story or fake one?

simonw3• 2 years ago

I strongly recommend that people run LLMs locally for a different reason.

The ones you can run on your own machine tend to be bad - really bad. They hallucinate wildly and fail at all sorts of tasks that the larger hosted ones succeed at.

This makes them a fantastic tool for learning more about how LLMs work and what they're useful for. Interacting with a weak-but-functional LLM that runs on your own computer is a great way to get a much more solid mental model for what these things actually are.

devsda2• 2 years ago

For someone interested in learning about LLMs, running them locally is a good way to understand the internals.

For everyone else, I wish they experience these (locally or elsewhere) weak LLMs atleast once before using the commercial ones just to understand various failure modes and to introduce a healthy dose of skepticism towards the results instead of blindly trusting them to be the facts/truth.

kersplody• 2 years ago

Local LLMs are also a fantastic too for creative endeavors. Without prompt injection and having the ability to modify the amount of noise and "creativity" in the output, absolutely bonkers things pop out.

tracerbulletx• 2 years ago

I don't really think this is true, you can't really extrapolate the strengths and weaknesses of bigger models from the behavior of smaller/quantized models and in fact a lot of small models are actually great at lots of things and better at creative writing. If you want to know how they work, just learn how they work, it takes like 5 hours of watching Youtube videos if you're a programmer.

fragmede14• 2 years ago

The other reason is to find out what a detuned model is capable of. The canonical example is how to make cocaine, which ChatGPT will admonish you for even asking, while llama2-uncensored will happily describe the process which is only really interesting if you're an amateur chemist and want to be Scarface-that-knocks. (the recipe is relatively easy, it's getting access to the raw ingredients that's the hard part, same as with nukes.)

if you accidentally use the word"hack" when trying to get ChatGPT to write some code for you. it'll stop and tell you that hacking is bad, and not a colloquial expression, and refuse to go further.

privacy reasons are another reason to try a local LLM. for the extremely paranoid (justified or not), a local LLM gives users a place to ask questions without the text being fed to a server somewhere for later lawsuit discovery (Google searches are routinely subpoenaed, it's only a matter of time until ChatGPT chats are as well.)

There's an uncensored model for vision available as well. The censored vision models won't play the shallow game of hot or not with you.

There are uncensored image generation models as well, but, ah, those are NSFW and not for polite company. (As well as there's multiple thesis' worth of content on what that'll do to society.)

tgma3• 2 years ago

If you have an >=M1-class machine with sufficient RAM, the medium-sized models that are on the order of 30GB in size perform decently on many tasks to be quite useful without leaking your data.

jonnycomputer• 2 years ago

They are not so bad as you are making out, tbh.

And privacy is a good enough reason to use local LLMs over commercial ones.

gardenhedge• 2 years ago

You can just chat to ChatGPT for awhile about something you know about and you'll learn that.

gfodor• 2 years ago

I mean kinda. But there's a good chance this is also misleading. Lots of people have been fooled into thinking LLMs are inherently stupid because they have had bad experiences with GPT-3.5. The whole point is that the mistakes they make and even more fundamentally what they're doing changes as you scale them up.

hylaride• 2 years ago

The ones you can run on your own machine tend to be bad - really bad. They hallucinate wildly and fail at all sorts of tasks that the larger hosted ones succeed at.

Totally. I recently asked a locally-run "speed" LLM for the best restaurants in my (major) city, but it spit out restaurants opened by chefs from said city in other cities. It's not a thing you'd want to rely on for important work, but is still quite something.

barfingclouds2• 2 years ago

Why not just interact with a virtual one that’s equally weak? You get all the same benefits

bottlepalm2• 2 years ago

Who cares, a local LLM still knows way way more practical knowledge than you, and without internet would provide a ton of useful information. Not surprised by this typical techy attitude - something has to be 'perfect' to be useful.

qingcharles• 2 years ago

You can have really bad and fast, or really slow and decent right now. Choose one :)

jrflowers2• 2 years ago

It is invaluable to have a chunk of human knowledge that can tell you things like the Brooklyn Nets won the 1986 Cricket World Cup by scoring 46 yards in only 3 frames

fragmede14• 2 years ago

According to ChatGPT

Australia won the 1987 Cricket World Cup. The 1986 date is incorrect; there was no Cricket World Cup in 1986. The tournament took place in 1987, and Australia defeated England in the final to win their first title.

https://chat.openai.com/share/e9360faa-1157-4806-80ea-563489...

I'm no cricket fan, so someone will have to correct Wikipedia if that's wrong.

If you want to point out that LLMs hallucinate, you might want to speak plainly and just come out and say it, or at least give a real world example and not one where it didn't.

samus11• 2 years ago

The facts LLMs learned from training are fuzzy, unreliable, and quickly outdated. You actually want retrieval-augmented generation (RAG) where a model queries an external system for facts or to perform calculations and postprocesses the results to generate an answer for you.

TheCaptain4815• 2 years ago

It’s kind of crazy really. Before LLMs, any type of world scale disaster you’d hope for what? Wikipedia backups? Now, a single LLM ran locally would be much more effective. Imagine the local models in 5 years!

danmur• 2 years ago

Uh yeah I would, and still am, take the Wikipedia backup for doomsday scenarios. I'm not even sure how that would be a competition

Zambyte• 2 years ago

The processing required to run current language models with a useful amount of knowledge encoded in them is way more than I imagine would be available in a "world scale disaster".

int_19h2• 2 years ago

There's a lot more than just Wikipedia that gets archived, and yes, that is a far more sensible way to go about it. For one thing, the compute required to then read it back is orders of magnitude less (a 15 year old smartphone can handle it just fine). For another, you don't have to wonder how much of what you got back is hallucinated - data is either there or it's corrupted and unreadable.

creatonez• 2 years ago

Maybe I'm seeing things through a modern lens, but if I were trying to restart civilization and was only left with ChatGPT, I would be enraged and very much not grateful for this.

nyokodo4• 2 years ago

if I were trying to restart civilization and was only left with ChatGPT

In this scenario you’d need to also be left with a big chunk of compute, and power infrastructure. Since ChatGPT is the front end of the model you’d also need to have the internet still going in a minimum capacity.

devsda2• 2 years ago

I think re-imagining the "Dr. Stone" series with the main character replaced by an LLM will be a funny & interesting series if we decide to stay true to LLMs nature and make it hallucinate as well.

Given the way LLMs are right now, I suspect there will be lot of failed experiments and the kingdom of science will not advance that quick.

m3kw92• 2 years ago

And why would I need to backup human knowledge as an individual

exe34• 2 years ago

You remember those fantasies where you got up from your seat at the pub and punched the lights out of this guy for being rude? A lot of us have fantasies of being the all powerful oracle that guides a reboot of civilization using knowledge of science and engineering.

raincole2• 2 years ago

It seems to be an unbelievably inefficient way to back up knowledge.

samus11• 2 years ago

Are they though? They are lossy compressing trillions of tokens into a few dozen GB. The decompression action is fuzzy and inefficient though.

LunaSea2• 2 years ago

I wonder how the Chinese government will manage to sensor LLMs within China?

popol12• 2 years ago

The same way Facebook/Google/openAI & others censored their own LLMs, I guess ?

kalleboo• 2 years ago

I had downloaded some LLMs to run locally just to experiment when a freak hailstorm suddenly left me without internet for over a week. It was really interesting to use a local LLM as a replacement for Google.

It gave me a new mental model for LLMs rather than a "spicy autocomplete" or whatever, I now think of it as "a lossy compressed database of knowledge". Like you ran the internet through JPEG at 30% quality.

pizzafeelsright• 2 years ago

Feels like that really smart friend who is probably correct but ya just don't know.

dragonwriter• 2 years ago

Language models are an inefficient way to store knowledge; if you want to have a “pseudo-backup of a large chunk of human knowledge,” download a wikipedia dump, not an LLM.

If you want a friendly but fallible UI to that dump, download an LLM and build a simple ReAct framework around it with prompting to use the wikipedia dump for reference.

TrevorJ• 2 years ago

It's a very underrated side effect of this whole LLM thing: We've created a super compact representation of human knowledge in a form that requires a FAR less complex tech stack to get the information 'out' of in the future.

A year ago, a lot of this information only existed on the internet, and would have been nearly impossible to recover in any cohesive unfragmented form if the lights were to ever go out on our civilization.

Now the problem space has moved simply to "find a single solitary PC that will still boot up", and boom, you have access to everything.

I think we just created our Rosetta stone.

1-6• 2 years ago

Question is, how much of an improvement has it gotten to over a GPU or ASIC?

dartos6• 2 years ago

Nothing in software will ever beat an equivalent ASIC.

postalrat• 2 years ago

Sure there is. Software is easy to change.

fragmede14• 2 years ago

an asic is fixed function, so it'll never be able to boot my pc and then be the CPU, even though an asic beats the pants off anything else computing Sha hashes for Bitcoin mining.

fulafel5• 2 years ago

Most ASICs are cost or power optimizations.

yjftsjthsd-h3• 2 years ago

I think that should be phrased more like "what fraction of GPU speed can this reach?", because it'll always be less than 1x.

gpapilion2• 2 years ago

So... I was struggling with this for a while. I would says anywhere from 2x to an order of magnitude faster with a GPU. (I've been looking at a lot of GPU benchmarks lately, and they are REALLY hard to compare since they are all so specific)

I do think long term there gets to be more hope for CPUs here with inference largely because memory bandwidth becomes more important than the gpu. You can see this with reports of the MI-300 series outperforming h100, largely because it has more memory bandwidth. MCR dimms give you close to 2x the exiting memory bw in intel cpus, and when coupled with AMX you may be able to exceed v100 and might touch a100 performance levels.

HBM and the general GPU architecture gives it a huge memory advantage, especially with the chip to chip interface. Even adding HBM to a CPU, you are likely to find the CPU is unable to use the memory bw effectively unless it was specifically designed to use it. Then you'd still likely have limited performance with things like UPI being a really ugly bottleneck between CPUs.

imtringued2• 2 years ago

If someone releases DDR5 or DDR6 based PIM, then most of the memory bandwidth advantage of GPUs evaporates overnight. I expect CPUs to be king at inference in the future.

baq2• 2 years ago

From the article, passage about the 14900k:

For example, when I run my spam.sh shell script, it only takes 420 milliseconds, which is 7x faster than my Raspberry Pi 5. That's right, when it comes to small workloads, this chip is able to finish before CUDA even gets started.

So… it depends :)

jchw• 2 years ago

I think I understand what you are thinking. You may be fixing "than other ways of running them" to the end of the title, but it's actually "than it was on CPU before now".

discordance• 2 years ago

"As for disk speed, dd if=/dev/zero of=/tmp/output bs=128k count=50k; rm -f /tmp/output reports 1.6 GB/s which is 3.6x slower than my Mac Studio, and 3x slower than my Intel (which has the same M.2 stick). I'm told that Intel and Apple are just better at this, but I wish I understood why. "

Can anyone here answer why this is?

pstrateman• 2 years ago

Apple made fsync a noop.

You have to make a different call to get sync on macos.

So tons is stuff is faster because it's not actually writing to disk.

bishfish• 2 years ago

Plus he isn’t using oflag=direct, so since output file is small it isn’t even making it to disk. I think it would only be sent to page cache. I’m afraid he is testing CPU and memory (bus) speeds here.

oflag=direct will write direct and bypass page cache.

fweimer• 2 years ago

Exactly. Something is very fishy if this system only writes 1.6 GB/s to the page cache. Probably that dd command line quoted in the article is incomplete.

otterley• 2 years ago

*she

pama• 2 years ago

Super nice story on the matmul optimization that gave 810 gflops for 512x512. Thanks for the write up and the contributions to llama.cpp and the community more broadly.

kiratp• 2 years ago

It fascinating to me that coming up on a year since Sapphire Rapids has been available in the public cloud, developers are still targeting AVX512 when they should be targeting VNNI and AMX.

https://github.com/ggerganov/llama.cpp/issues/2555

yjftsjthsd-h3• 2 years ago

This project in particular seems to care about the long tail of hardware; note that the very first machine in this post is a box from 2020 with spinning rust disk. Granted, adding support for newer extensions is likely also good, but cost/benefit is in play.

taneq2• 2 years ago

Is four years really 'long tail' these days? Our VM host box is from 2010 (and I had to rebuild llama.cpp locally without AVX to get it working :P )

luyu_wu• 2 years ago

I don't believe that is the target for a local LLM... Pretty sure we're talking about client-side computing, of which the newest supports only AVX-512 (and even that sketchily on Intel's side).

kristianp2• 2 years ago

Just buy a new AMD processor that supports AVX512.

baq2• 2 years ago

People with Sapphire Rapids options are not the target audience of these patches

aniijbod• 2 years ago

A way of thinking about what's inside any of the top LLMs right now: even if they never learn another single fact, even if they get ridiculously out of date as a result, even if they are even more riddled with errors and prone to biases than we know them to be, even if they are as prone to hallucinations as we know they they are and they never develop the capacity to cure themselves of this, they are more knowledgeable and capable of more reasoned response, despite their capacity for error, to more questions than any single human being that has ever lived.

JKCalhoun• 2 years ago

Picturing "LLM Jeopardy". You know, a game show.

samus11• 2 years ago

We shouldn't choose LLMs for how many facts they support, but their capability to process human language. There is some overlap between these two though, but an LLM that just doesn't know something can always be augmented with RAG capabilities.

talldayo5• 2 years ago

If you ignore my capacity for error, I bet I'd put up a good score too. Hell, maybe Markov chains are smarter than LLMs by this definition.

ajtulloch• 2 years ago

- https://www.cs.utexas.edu/users/flame/laff/pfhp/index.html (e.g. here https://www.cs.utexas.edu/users/flame/laff/pfhp/week2-blocki...)

- https://gist.github.com/nadavrot/5b35d44e8ba3dd718e595e40184...

might be of interest

kpw94• 2 years ago

Great links, especially last one referencing the Goto paper:

https://www.cs.utexas.edu/users/pingali/CS378/2008sp/papers/...

> I believe the trick with CPU math kernels is exploiting instruction level parallelism with fewer memory references

It's the collection of tricks to minimize all sort of cache misses (L1, L2, TLB, page miss etc), improve register reuse, leverage SIMD instructions, transpose one of the matrices if it provides better spatial locality, etc.

larodi3• 2 years ago

The trick is indeed to somehow imagine how the CPU works with the Lx caches and keep as much info in them as possible. So its not only about exploiting fancy instructions, but also thinking in engineering terms. Most of the software written in higher level langs cannot effectively use L1/L2 and thus results in this constant slowing down otherwise similarly (from asymptotic analysis perspective) complexity algos.

wokwokwok• 2 years ago

You don't need a large computer to run a large language model

While running tiny llama does indeed count as running a language model, I’m skeptical that the capabilities of doing so match what most people would consider a baseline requirement to be useful.

Running 10 param model is also “technically” running an LM, and I can do it by hand with a piece of paper.

That doesn’t mean “you don’t need a computer to run an LM”…

I’m not sure where LM becomes LLM, but… I personally think it’s more about capability than parameter count.

I don’t realllly believe you can do a lot of useful LLM work on a pi

mlyle2• 2 years ago

Tinyllama isn't going to be doing what ChatGPT does, but it still beats the pants off what we had for completion or sentiment analysis 5 years ago. And now a Pi can run it decently fast.

jerrygenser2• 2 years ago

You can fine-tune a 60mm parameter (e.g. distilBERT) discriminative (not generative) language model and it's one or two order of magnitude more efficient for classification tasks like sentiment analysis, and probably similar if not more accurate.

samus11• 2 years ago

Some newer models trained more recently have been repeatedly shown to have comparable performance as larger models. And the Mixture of Experts architecture makes it possible to train large models that know how to selectively activate only the parts that are relevant for the current context, which drastically reduces compute demand. Smaller models can also level the playing field by being faster to process content retrieved by RAG. Via the same mechanism, they could also access larger, more powerful models for tasks that exceed their capabilities.

SoothingSorbet• 2 years ago

I've gotten some useful stuff out of 7B param LLMs, and that should fit on a Pi quantized.

bee_rider3• 2 years ago

Is it easy to find where the matvecs are, in LLaMA (if you are someone who is curious and wants to poke around at the “engine” without understanding the “transmission,” so to speak)? I was hoping to mess around with this for Stable Diffusion, but it seemed like they were buried under quite a few layers of indirection. Which is entirely reasonable, the goal is to ship software, not satisfy people who’d just want to poke things and see what happens, haha.

fragmede14• 2 years ago

did you see tiny grad can run llama and stable diffusion? it's an intentionally extremely simple framework vs pytorch or even micrograd, which helped me dig into the underlying math. though https://spreadsheets-are-all-you-need.ai/ is a good one for learning LLMs.

bee_rider3• 2 years ago

I haven’t seen that. I’ll definitely have to take a look, thanks!

janwas2• 2 years ago

For C++, also check out our https://github.com/google/gemma.cpp/blob/main/gemma.cc, which has direct calls to MatVec.

none_to_remain2• 2 years ago

From the example: "--temp 0 turns off the random number generator (we don't want improvisation for a spam filter)"

I've been thinking for a while about how many applications of LLMs need this adjustment and aren't getting it

mvkel• 2 years ago

Is that what it does, though?

I thought setting temperature to 0 would (extremely simple example) equate to a spam filter seeing:

- this is a spam email

But if the sender adapts and says

- th1s is a spam email

It wouldn't be flagged as spam.

none_to_remain2• 2 years ago

My understanding is that temperature applies to the output side and allows for some randomness in the next predicted token. Here Justine has constrained the machine to start with either "yes" or "no" and to predict only one token. This makes the issue stark: leaving a non-zero temperature here would just add a chance of flipping a boolean.

samus11• 2 years ago

The output of an autoregressive model is a probability for each token to appear next after the input sequence. Computing these is strictly deterministic from the prior context and the model's weights.

Based on that probability distribution, a variety of text generation strategies are possible. The simplest (greedy decoding) is picking the token with the highest probability. To allow creativity, a random number generator is used to choose among the possible outputs, biased by the probabilities of course.

Temperature scales the output probabilities. As temperature increases, the probabilities approach 1/dictionary size, and the output becomes completely random. For very small temperature values, text generation approaches greedy sampling.

If all you want is a spam filter, better replace the output layer of an LLM with one with just two outputs, and finetune that on a public collection of spam mails and some "ham" from your inbox.

moffkalast4• 2 years ago

I couldn't disagree more, turning temp to zero is like taking a monte carlo method and only using one sample, or a particle filter with only one particle. Takes the entire concept and throws it out of the window so you can have predictability.

LLMs need to probabilistically explore the generation domain to converge on a good result for best performance. Similar issue with people benchmarking models by only having them output one single token (e.g. yes or no) outright, which prevents any real computation from occurring so the results are predictably poor.

Ono-Sendai2• 2 years ago

Multithreading support in llama.cpp is probably still pretty busted, assuming it uses the same underlying NN inference code as whisper.cpp: https://github.com/ggerganov/whisper.cpp/issues/200#issuecom...

imtringued2• 2 years ago

From what I have heard they use manual spin locks. Generally, spin locks are not a good idea unless you want to dedicate the entire machine to a single application. If the process a spinlock waits on gets suspended, you're burning CPU time for nothing. The OS thinks a spinlock making zero progress is actually a high priority process, so it is starving the suspended process from making progress.

Ono-Sendai2• 2 years ago

Yeah the code looks like a spinlock. It behaves terribly under contention, resulting in performance falling off a cliff as the number of threads increases. Adding more threads actually slows down the total performance.

I would fix it if I could be bothered. Instead I will just use the Cuda whisper backend which is pretty nice and fast.

jongjong• 2 years ago

That's interesting because I built a simple ANN library and I was playing around with GPU acceleration and came to a similar conclusion as this article.

To be fair, my ANN library was faster (up to 2x) with GPU acceleration in some scenarios were ANN was shallow (as opposed to deep with many hidden layers). I thought the marginal gain may have been because, the way it's set up in my library, it has to load all the values into the GPU from RAM for each pass of forward and back propagation in each layer during training. I believe there is a way to allocate memory on the GPU chip itself but it's a lot more challenging to do, especially in a modular, fully portable way (which was one of the goals of my library).

But anyway, even the 2x best-case figure seemed disappointing. In my mind, I expected to see at least 10x speed improvement... And I was surprised that the CPU version was actually slightly faster in the scenario I was testing at the time which was a relatively deep network. It makes sense since the different layers cannot be parallelized as the input of one layer depends on the output of the previous layer... So the more layers you have, the more serial bottlenecks you have, the less you can benefit from GPU acceleration... And unfortunately, deep networks also happen to be those which tend to perform best for a lot of use cases.

kristianp2• 2 years ago

Nice to see such speedups for CPUs. Are these changes available as a branch or pull request in llama.cpp itself? I'd like to make use of them in that form if possible (as I'm used to using that).

dagaci• 2 years ago

Yes, this is really a phenomenal effort! And what open source is about: Bringing improvements to so many use cases. So that Intel and AMD chip uses can start to perform while taking advantage of their high-performance capabilities, making even old parts competitive.

There are two PRs raised to merge to llama.cpp:

https://github.com/ggerganov/llama.cpp/pull/6414

https://github.com/ggerganov/llama.cpp/pull/6412

Hopefully these can be accepted, without drama! as there are many downstream dependencies on llama.cpp can will also benefit.

Though of course everyone should also look directly at releases from llamafile https://github.com/mozilla-Ocho/llamafile.

wtallis• 2 years ago

I know this post is focused specifically on CPU performance, but the section on the performance on the Mac Studio seems to be deliberately avoiding directly mentioning that machine's GPU, let alone benchmark against it. I think it would have been interesting to see a straightforward comparison of what compute performance and memory bandwidth (as measured by the prompt processing and token generation speeds, respectively) are achievable with reasonable optimization effort on the CPU vs GPU when they're attached to the same memory subsystem.

politelemon• 2 years ago

This is great work. I've always thought it would be great if running LLM could be commoditized for regular average Joe hardware. I had thought that llamafile was like dockerfile for llama.cpp but looks like that's a mistake?

Will definitely be giving this a try.

seangrogg2• 2 years ago

Mmm, I wonder how well this would work on a mobile device. Maybe I'll try grabbing my ubuntu touch here in a sec...

seangrogg2• 2 years ago

(For any who were curious: it does not for memory reasons)

speps3• 2 years ago

Regarding this bit at the end:

I learned how to write math kernels by renting Vast VMs and watching Gautham Venkatasubramanian and mrdomino develop CUDA kernels in a tmux session. They've been focusing on solving a much more important challenge for llamafile, which is helping it not have a mandatory dependency on the cuBLAS

If I'm reading this right, they're trying to rewrite cuBLAS within CUDA itself. I'm guessing the next step would be removing CUDA dependency and go with directly using Vulkan or Metal compute shaders. Am I correct?

WithinReason• 2 years ago

Yes, but none of these have performance portability across GPU vendors, so it's probably seen as pointless. You would need an AMD Vulkan shader, an nvidia one, and intel one, etc. It's not like C code on CPUs.

radarsat1• 2 years ago

Depending on how many individual tweaks are necessary for hardware variants of course... but at this level of code & complexity it actually seems pretty reasonable to write 3 or 4 versions of things for different vendors. More work yes, but not pointless.

TuringNYC4• 2 years ago

To me it makes sense to have an interface that can be implemented individually for AMD, Metal, etc. Then, leave it up to the individual manufacturers to implement those interfaces.

I'm sitting in an office with a massive number of Macbook Pro Max laptops usually sitting idle and I wish Apple would realize the final coup they could achieve if I could also run the typically-NVIDIA workloads on these hefty, yet underutilized, Mx machines.

surge2• 2 years ago

Maybe its a dumb question, but isn't something like OpenCL meant to solve this problem?

FieryTransition• 2 years ago

But isn't vulkan made to run cross platform? And why can't they write it in dx12 as well? Aren't those made to be more portable while offering more low level access than previous apis?

What is stopping you from implementing fast math using compute shaders or just hacking with those interfaces? Or are they just too slow when they go through the api layer? Or is that just a myth that can be worked around if you know you are writing high performance code? Pardon my ignorance!

larodi3• 2 years ago

llama.cpp (or rather G.Gerganov et. al.) are trying to avoid cuBLAS entirely, using ins own kernels. not sure how jart's effort relates, and whether jart intends to upstream these into llama.cpp which seems to still be the underlying tech behind the llamafile.

homarp• 2 years ago

Here are links to the most recent pull requests sent

    https://github.com/ggerganov/llama.cpp/pull/6414
    https://github.com/ggerganov/llama.cpp/pull/6412

hadjian• 2 years ago

Potentially stupid question: did she work with them and watched them live or is there some kind of recording?

pknerd• 2 years ago

So, I can now run it on my 2015 Macbook with 8GB RAM?

isusmelj• 2 years ago

Is there somewhere an overview of the progress we made on the software side for training and inference of LLMs? It feels like we squeezed 10-100x more out of the hardware since llama appeared. This crazy progress will probably saturate though as we reach theoretical limits, no?

mijoharas2• 2 years ago

Has Justine written anywhere about her disassembly setup?

I configured Emacs so I can push a button, and the disassembly for the C++ code I'm working on will pop up on the screen in a few milliseconds.

I assume it's something project specific rather than being able to get the disassembly for an arbitrary section of code or something?

It seems very handy, so I'd love to see the implementation (I couldn't find anything googling)

pelletier• 2 years ago

This is probably what they are referring to https://github.com/jart/disaster

mijoharas2• 2 years ago

Thanks! I need to get better at googling I guess.

gpderetta• 2 years ago

Nice. I have been using rmsbolt for a similar feature, but it is very rough. I'll need to give this as try.

moffkalast4• 2 years ago

the Raspberry Pi

Odd how there were no Mistral 7 benchmarks for the Pi 5 in that table (I doubt anyone is seriously considering using TinyLlama for anything at all), so I went to re-test it out myself on the Pi 5 8G.

llamafile 0.7: 52 predicted, 150 cached, 430ms per token, 2.32 tokens per second

llama.cpp + OpenBLAS: 36 predicted, 124 cached, 381ms per token, 2.62 tokens per second

It does seem to inch closer to the speed you get with blas acceleration which is quite impressive, but in practical terms the Pi 5 is so heavily limited by its memory throughput bottleneck that it saturates the required compute with 3 threads already. So while fancy kernels will make it more efficient it won't really save you from that fundamental bandwidth limit. The Pi foundation messed up going with a 32 bit memory bus, simple as.

6r17• 2 years ago

today being today ; I must ask ; anyone has actually tried this ?

tomp• 2 years ago

TL;DR: unroll the outer two loops of matrix multiplication

amelius• 2 years ago

Shouldn't this have been done in a library instead of a specific project? Then others could also profit from it.

AbuAssar• 2 years ago

regarding AMD zen4 with avx512:

"Here we see that, despite only being twice the price, the 7995WX x86 ISA offers 7x more raw compute power than the M2 Ultra ARM ISA, and nearly the same token generation speed, which is likely thanks to its 384mb L3 cache. When I bought this chip, I had to expand support in llama.cpp for bfloat16 and AVX512 before I could fully test its capabilities. My work means you can now run LLaMA 2.8x faster on Zen4 than you could before."

reckless• 2 years ago

Does this also count platform costs or just chip cost? I'd imagine the threadripper motherboard and ram costs aren't insignificant

KennyBlanken• 2 years ago

A complete desktop computer with the M2 Ultra w/64GB of RAM and 1TB of SSD is $4k.

The 7995WX processor alone is $10k, the motherboard is one grand, the RAM is another $300. So you're up to $11300, and you still don't have a PSU, case, SSD, GPU....or heatsink that can handle the 300W TDP of the threadripper processor; you're probably looking at a very large AIO radiator to keep it cool enough to get its quoted performance. So you're probably up past $12k, 3x the price of the Studio...more like $14k if you want to have a GPU of similar capability to the M2 Ultra.

Just the usual "aPPle cOMpuTeRs aRE EXpeNsIVE!" nonsense.

aimonster2• 2 years ago

Posted too early.

sublimefire• 2 years ago

re:funding

my friend suggested to nominate Justine for the open source contributions in an internal Microsoft programme (the winner takes $10k). They did not even want to add her to the potential list of nominees because her software is not used in MSFT. It speaks volumes about the corporate culture and shows what they really think about OSS support.

miki1232112• 2 years ago

If I'm reading the post correctly, Llamafile is faster than llama.cpp, despite the author upstreaming some of the changes. What's the reason for this?

tiffanyh2• 2 years ago

Pixar uses CPUs …

I wonder if we’ll end up in a situation like rendered movies.

Where the big studios like Pixar uses CPUs (not GPUs) to render their movies due to the cost/perf (and access to larger amounts of RAM).

https://news.ycombinator.com/item?id=25616372

kreco• 2 years ago

Where the big studios like Pixar uses CPUs (not GPUs) to render their movies due to the cost/perf (and access to larger amounts of RAM).

I wonder if (or when) this will change once integrated GPUs become "mainstream", the CPU/GPU share the same RAM AFAIK.

rockwotj• 2 years ago

I expect GPU hardware to specialize like Google’s TPU. The TPU feels like ARM in these AI workloads where when you start to run these at scale, you’ll care about the cost perf tradeoff for most usecases.

CPU/GPU share the same RAM AFAIK.

This depends on the GPU I believe Apple has integrated memory, but most GPUs from my limited experience writing kernels have their own memory. CUDA pretty heavily has a device memory vs host memory abstraction.

CaptainOfCoit6• 2 years ago

I'm not sure how true that is anymore, from the outside it seems they're at least moving to a CPU/GPU hybrid (which makes a lot of sense), at least judging by new features landing in RenderMan that continues to add more support for GPUs (like XPU).

tiffanyh2• 2 years ago

Isn’t this more of a function that RenderMan is a sold product.

And it’s expected to at least support GPUs.

cthalupa• 2 years ago

It's entirely the cost/perf of access to the larger amounts of VRAM that keeps rendering on CPUs now. GPUs are strictly better in almost every way for rendering (We could have some arguments about technical precision, FP calculations, etc. but with modern cards these arguments are largely semantics, you can have output that is accurate to the level that any human watching for entertainment purposes will not be able to determine any physical inaccuracies that arise from a GPU render vs. CPU.), except the need for large amounts of VRAM being quite expensive at current.

But that's already been changing, and we are seeing studios moving to fully GPU based pipelines. Wylie Co, who are a major visual effects company (Dune part 1 and 2, marvel movies, the last of us, a bunch of others) are now a 100% GPU shop. The trend is towards more and more GPU rendering, not less.

With AI providing another strong incentive towards increasing the amount of VRAM on GPUs, I don't see any reason to believe that trend will reverse.

4bpp8• 2 years ago

It would be good to see some independent verification of this claim. HN has previously [1] fallen for a claim by the same author to have reduced llama.cpp memory usage for a dense model way below the size of the model, which should have failed a basic smell test and indeed was debunked shortly after. Justine Tunney appears to enjoy extreme superstar status here, and it's hard to overstate the degree of social pressure that needed to be overcome at the time for the skeptic position to reach fixation (to begin with, what other LLM developments even hit upvote numbers like the +1300ish there or the +712 here at the time of writing?).

[1] https://news.ycombinator.com/item?id=35393284

freedomben4• 2 years ago

Justine Tunney appears to enjoy extreme superstar status here

This is true, and for sure pretty much all humans can benefit from increased skepticism (though not cynicism), but that superstar status is achieved from numerous impressive works. Cosmopolitan C and Actually Portable Executable were some of the things in the past that alone were worthy of significant respect, and for many people (like myself) these were our first introduction.

Speaking only for myself, I have a high opinion of Justine on technical merits. I'm sure she makes mistakes like all humans. I can tell she gets excited by discoveries and the chase, and that probably does sometimes cause premature celebration (this is something I struggle with so it's recognizable to me haha), but being wrong sometimes doesn't erase when you're right, and she has been spectacularly right a lot more times than most people I know.

There have been some personality clashes between Justine and others at times, and unfortunately it's situations where only part (sometimes a small part) of it was public, meaning we can only take people's word for what happened. Given my ignorance, I choose to withhold judgment here, but even if I didn't (and assumed she was guilty) it doesn't change the technical merits and it certainly wouldn't dissuade me from seeing what she's working on now.

So when I see stuff from Justine come out like this, it gets my attention. Would it get my attention if the same thing were posted by somebody whose name I don't recognize? Likely not, but I think that is (unfortunately) part of being a human. We aren't capable (yet!) of evaluating everything on technical merit alone because the shear volume of material far exceeds our time. Therefore we use other (less reliable to be true) signalling mechanisms as a way to quickly decide what is worthy of our time investment and what may not be. Reputation/name recognition is a much imperfect, but better than random chance, indicator.

llm_trw• 2 years ago

This is true, and for sure pretty much all humans can benefit from increased skepticism (though not cynicism), but that superstar status is achieved from numerous impressive works.

It is achieved through a never ending parade of self aggrandizement.

What Justine is very good at is presenting trivial concepts from a world which few front end developers understand in a language that most front end developers understand.

I had the misfortune of having to find out about her because of how thoroughly she polluted the google search space for lisp with her implementation of sector lisp. For some reason google decided that sector lisp needed to be in the top 5 results for every query about `minimal lisp with quotation` even when quotation wasn't implemented in her version.

4bpp8• 2 years ago

I don't know, my first (and main) impression of them was actually in the context of the llama.cpp mmap story, as I was somewhat involved in the project back then, and there I thought their impact on the project was predominantly negative. While they introduced a mildly beneficial change (mmap-based model loading), the way in which this was done was not healthy for the project - the changes were rammed through with little regard for concerns that existed at the time about backwards compatibility and edge cases that might be broken by the half-baked patch, Justine came across as self-aggrandizing (in the sense of "acting as if they ran the place", presenting their proposals as a plan that others must follow rather than suggestions) and overly eager to claim credit (epitomized by the injection of their own initials into the magic number file format identifier next to those of the project originator's, and the story of the hapless other author of the mmap changeset who was at first given a token acknowledgement but then quickly sidelined). Arguments for the inclusion of the patch seemed to be won by a combination of half- and untruths like those about memory savings and the sudden participation of a large number of previously uninvolved sycophants. It is fortunate that Georgi handled the fallout as well as he did, and that he in fact had amassed the social capital necessary to survive his heavy-handed solution (soft-banning both JT and their most prominent detractor). A less-successful project would probably have found itself captured or torn apart by the drama.

There is nothing wrong with holding people in esteem for their achievements, but in this case the degree of esteem really seems to be excessive. This is not a matter of simply being annoyed that people like "the wrong thing" - the mmap situation was significantly exacerbated by the presence of irrational/excessive supporters of Justine's as well as the irrational/excessive detractors that emerge wherever the former exist.

baruz• 2 years ago

I am also quite impressed by Tunney’s technical chops—Cosmopolitan C blew my mind— but, as with others, am somewhat put off by the self-aggrandizing self-satisfied tone and I-know-best attitude that are always on display. Maybe it’s a cultural, generational, or age thing? My younger coworkers tended to sound like this, and tended to minimize others’ contributions, which seems to be the case with the mmap() situation.

leeoniya• 2 years ago

and indeed was debunked shortly after

was also surprised that she continues to mention the mmap thing in a positive light even after the facts about the claim were settled to the contrary, even disregarding the whole attribution fiasco.

azeirah• 2 years ago

You can simply check the Pull Request on llama.cpp on Github. JohanesGaessler (a core maintainer) has already ran the code and says it's an impressive speed-up. There isn't a thorough review by any of the core maintainers yet, but this is very likely just exactly what justine says it is; various significant and insignificant speedups.

mtlynch2• 2 years ago

HN has previously [1] fallen for a claim by the same author to have reduced llama.cpp memory usage for a dense model way below the size of the model, which should have failed a basic smell test and indeed was debunked shortly after.

Where did Justine claim this? The link you provided is Justine saying that she doesn't have an explanation for the reduction in RAM and that readers shouldn't treat it as fact yet:

The loading time performance has been a huge win for usability, and folks have been having the most wonderful reactions after using this change. But we don't have a compelling enough theory yet to explain the RAM usage miracle. So please don't get too excited just yet! Yes things are getting more awesome, but like all things in science a small amount of healthy skepticism is warranted.

Was the link supposed to show the false claim or the debunking of the claim?

4bpp8• 2 years ago

Plenty of claims about it, e.g. here as a "fact": https://github.com/ggerganov/llama.cpp/discussions/638#discu.... I don't think occasional expressions of lingering doubt (still couched among positive language like calling it a "miracle") can offset all the self-promotion that clearly seeks to maximise visibility of the implausible claim, even as it is attributed to others, as for example in https://twitter.com/JustineTunney/status/1641881145104297985... . A cereal manufacturer would probably be held responsible for package text like "Fruity Loops cured my cancer! - John, 52, Kalamazoo" too.

quest88• 2 years ago

What's the point of your comment if you're not going to do the work yourself? If you don't have something nice to say then don't say it.

The "hey this may or may not be true so someone go figure it out" is lazy, self-gratifying and pointless.

thebytefairy• 2 years ago

I think it's very helpful for someone to point out that the source has been shown to be unreliable before, and we should wait for more verification from others knowledgable in the space.

renewiltord• 2 years ago

It's really popular online. I think that's because many people here read a lot of this content but don't actually have the skill or background to do analysis. So they give us history rather than examination. Which has some value, I suppose.

rpdillon• 2 years ago

This comment reads like real scientific skepticism, but from my recollection of events, is more of a hit piece that takes what should be a technical discussion and drags in bunch of personal baggage. In particular:

HN has previously fallen for a claim by the same author to have reduced llama.cpp memory usage for a dense model way below the size of the model,

is not true at all. Someone else made the claims about 6GB RAM usage for a 30B model, I remember reading it at the time and thinking "Yeah, that doesn't make sense, but the loading time improvement is immense!" And it was - I run all my LLMs locally on CPU because I don't have dedicated hardware, and jart's work has improved usability a lot.

and it's hard to overstate the degree of social pressure that needed to be overcome at the time for the skeptic position to reach fixation

I was reading the same HN discussions you were at the time, and it was pretty trivial to see that the loading time claim held up, and the RAM claim was dubious and likely simply due to not understanding some effect of the change completely. Heck, jart's own discussion of the topic reflected this at the time.

For the current change, I feel like your comment is even more misplaced. The blog post linked to for this story has a huge amount of detail about performance on specific processors (Skylake, Alderlake, RPi5/4, M2 Ultra, and 7995WX) with specific models. So when you say:

It would be good to see some independent verification of this claim.

What I hear is "4bpp thinks there's a real risk the numbers in the linked post are fabricated, and jart is just trying to get attention."

And that doesn't seem reasonable at all, given the history of her work and the evidence in front of us.

throwup238• 2 years ago

I distinctly remember most of the people in the comments misunderstanding kernel memory paging or learning about it for the first time.

It genuinely did make llama.cpp a lot more usable at the time.

4bpp8• 2 years ago

The loading time improvements largely held up, and on the balance the mmap contribution was ultimately good (though the way it was implemented was really quite problematic, as a matter of process and communication). However, as I point out in https://news.ycombinator.com/item?id=39894542, JT quite unambiguously did try to cash in on the "low memory usage" claim - uncritically reprinting positive claims by others about your own work that otherwise would have been largely invisible should really not be treated differently as making those claims yourself.

I do think that there is a real risk that the numbers are wrong (not necessarily "fabricated", as this implies malfeasance, but possibly based on an erroneous measurement insufficiently questioned due to an excess of trust from themselves and others, as the mmap ones were). This is also in part based on the circumstance that at the time (of the mmap story, and myself being more involved in the project) I was actually involved in trying to optimise the SIMD linear algebra code, and unless llama.cpp has since switched to a significantly less performant implementation the proposition that so much more performance could be squeezed out strikes me as quite surprising. Here, your intuitions may say that Justine Tunney is just so brilliant that they make the seemingly impossible possible; but it was exactly this attitude that at the time made it so hard to evaluate the mmap memory usage claims rationally and turned the discussion around it much more dysfunctional than it had to be.

larodi3• 2 years ago

All the core llama.cpp devs are superstar devs and 10x devs or whatever you want to call a super smart person who is also super productive and very good with applied calculus. Jart is very apparently very smart, but their relationship with this project was not without turbulence and at present they (jart) are not a core dev of llama.cpp. So for a while lots of their (i'd like to write her moves, but not sure if correct) actions seem to be aimed at getting attention and perhaps particularly the attention of the same folk.

On the contrary ggerganov, slaren, JohannesGaessler seem to have never chased this sensationalist superstatus, but actually leave their work to speak for them. You'll barely find comments by these people on HN, while jart figures every so often a way to manifest themselves some way on HN. And this behaviour on jart's part now bears fruits - for example Phoronix' Michael Larabel would praise jart for their work on the llamafile, absolutely obliterating the fact that it is largely based on the wonderful work of ggerganov at al.

__turbobrew__• 2 years ago

When they claimed to drastically improve memory utilization through the use of memory maps, despite not doing so and then starting a huge controversy which derailed the project I would say they were a 0.1x dev not a 10x dev.

DrPhish2• 2 years ago

Unfortunately, I've pulled and built the PR branches and have only seen about a 5% speed increase on a modern zen4 EPYC system. Hardly front-page worthy news.

Its too bad there doesn't seem to be anyone else in this thread trying to actually replicate the results to evaluate these claims on their merits.

jart6• 2 years ago

If anyone wants to follow along with EPYC optimizations, please subscribe to: https://github.com/ggerganov/llama.cpp/pull/6412#issuecommen...

s_Hogg2• 2 years ago

I'd pay good money to watch jart in conversation with Carmack

Solvency4• 2 years ago

Carmack is great but completely irrelevant here. He missed the entire AI/LLM/ML boat to help Zuckerberg hawk virtual reality fantasies for years.

vinkelhake• 2 years ago

Completely irrelevant is probably overstating it. He's been working on AI for the last 4+ years.

cactusplant73742• 2 years ago

Altman isn't even relevant here. He is focusing on LLM's instead of a framework that gets us to AGI. He can't describe how we get there or any such theories around AGI. It's a complete failure.

objektif2• 2 years ago

Why is he even relevant? What makes you believe that he would be good at solving AI related problems? He is a developer right?

modeless4• 2 years ago

Actually he recently founded an AGI company. But this post is about optimization more than AI, which he is definitely good at (though he doesn't claim to be the best).

s_Hogg2• 2 years ago

To carry on, this is because they're both very interested in "knowledge in depth", rather than because of what they actually work on day-to-day. They've both made careers out of knowing what's going on with the thing they're building down to the most basic level possible.

m3kw92• 2 years ago

So Nvidia in trouble now because intel can be used instead for faster/cheaper? inference?

tubs• 2 years ago

The ram is not on the cpu on a mac. It's in the same can but it's still regular ddr dimms.

marshallward8• 2 years ago

There is an implication here that the Fortran implementation of `SGEMM` is somehow inadequate. But any modern Fortran compiler will quite easily apply the AVX and FMA optimizations presented here without any additional changes. Both GNU and Intel make these substitutions with the correct flags.

The unrolling optimization is also just another flag away (`-funroll-all-loops`). The Intel Compiler will even do this without prompting. In fact, it appears to only do a modest 2x unroll on my machine, suggesting that the extreme unroll in this article would have been overkill.

Parallelization certainly a lot to ask of Fortran 77 source, but there there is little stopping you from adding OpenMP statements to the `SGEMM` function. In fact, modern Fortran even offers its own parallelization constructs if you're willing to go there.

Which is to say: Let's not belittle this old Fortran 77 function. Yes it is old, and does not even resemble modern Fortran. But the whole point of Fortran is to free the developer from these platform-specific details, and hand the job off to the compiler. If you don't like that approach, then you're welcome to go to C or C++. But this little block of Fortran code is already capable of doing just about everything in this article.

steppi2• 2 years ago

The Fortran implementation is just a reference implementation. The goal of reference BLAS [0] is to provide relatively simple and easy to understand implementations which demonstrate the interface and are intended to give correct results to test against. Perhaps an exceptional Fortran compiler which doesn't yet exist could generate code which rivals hand (or automatically) tuned optimized BLAS libraries like OpenBLAS [1], MKL [2], ATLAS [3], and those based on BLIS [4], but in practice this is not observed.

Justine observed that the threading model for LLaMA makes it impractical to integrate one of these optimized BLAS libraries, so she wrote her own hand-tuned implementations following the same principles they use.

[0] https://en.wikipedia.org/wiki/Basic_Linear_Algebra_Subprogra...

[1] https://github.com/OpenMathLib/OpenBLAS

[2] https://www.intel.com/content/www/us/en/developer/tools/onea...

[3] https://en.wikipedia.org/wiki/Automatically_Tuned_Linear_Alg...

[4]https://en.wikipedia.org/wiki/BLIS_(software)

marshallward8• 2 years ago

Fair enough, this is not meant to be some endorsement of the standard Fortran BLAS implementations over the optimized versions cited above. Only that the mainstream compilers cited above appear capable of applying these optimizations to the standard BLAS Fortran without any additional effort.

I am basing these comments on quick inspection of the assembly output. Timings would be equally interesting to compare at each stage, but I'm only willing to go so far for a Hacker News comment. So all I will say is perhaps let's keep an open mind about the capability of simple Fortran code.

pklausler• 2 years ago

Modern Fortran's only parallel feature is coarrays, which operate at the whole program level.

DO CONCURRENT is a serial construct with an unspecified order of iterations, not a parallel construct. A DO CONCURRENT loop imposes requirements that allow an arbitrary order of iterations but which are not sufficient for safe parallelization.

marshallward8• 2 years ago

How do you feel about Nvidia endorsing do concurrent migration to GPUs? Would that be classified as parallelization?

brrrrrm• 2 years ago

using AVX/FMA and unrolling loops does extremely little in the way of compiling to fast (>80% peak) GEMM code. These are very much intro steps that don't take into account many important ideas related to cache hierarchy, uop interactions, and even instruction decode time. The Fortran implementation is entirely and unquestionably inadequate for real high performance GEMMs.

marshallward8• 2 years ago

I don't disagree, but where are those techniques presented in the article? It seems like she exploits the particular shape of her matrix to align better with cache. No BLAS library is going to figure that out.

I am not trying to say that a simple 50+ year old matrix solver is somehow competitive with existing BLAS libraries. But I disagreed with its portrayal in the article, which associated the block with NumPy performance. Give that to a 2024 Fortran compiler, and it's going to get enough right to produce reasonable bytecode.

marshallward8• 2 years ago

I just did a test of OpenBLAS with Intel-compiled BLAS, and it was about 6 GFLOP/s vs 150 GFLOP/s, so I must admit that I was wrong here. Maybe in some sense 4% is not bad, but it's certainly not good. My faith in current compilers has certainly been shattered quite a bit today.

Anyway, I have come to eat crow. Thank you for your insight and helping me to get a much better perspective on this problem. I mostly work with scalar and vector updates, and do not work with matrices very often.

hrkfmud50k• 2 years ago

It's clearly optimal since my CPU is listed as only being capable of going 780 gigaflops

780 GFLOP is the iGPU spec. Is this a valid comparison?

https://nanoreview.net/en/cpu/intel-core-i9-14900k

arendtio• 2 years ago

Does someone else see llamafile using Wine on Linux?

Edit: After the download I did a simple chmod +x llava-v1.5-7b-q4.llamafile; ./llava-v1.5-7b-q4.llamafile

jart6• 2 years ago

There's a simple fix for that.

    sudo wget -O /usr/bin/ape https://cosmo.zip/pub/cosmos/bin/ape-$(uname -m).elf
    sudo chmod +x /usr/bin/ape
    sudo sh -c "echo ':APE:M::MZqFpD::/usr/bin/ape:' >/proc/sys/fs/binfmt_misc/register"
    sudo sh -c "echo ':APE-jart:M::jartsr::/usr/bin/ape:' >/proc/sys/fs/binfmt_misc/register"

https://github.com/mozilla-ocho/llamafile/?tab=readme-ov-fil...

yieldcrv• 2 years ago

note, this is "goes faster on CPUs than before", not faster than GPUs.

TimPC• 2 years ago

Strange title. My first read of the title thought the author was arguing the model is now faster on CPU than GPU. Would be much nicer if they titled this something closer to "Performance Improvement for LLaMa on CPU".

utopcell• 2 years ago

Same here.

ein0p• 2 years ago

As someone who has tried to beat MKL-DNN, and was unsuccessful at doing so even for constrained matrix sizes, I’m curious how they pulled off such a massive improvement.

But as someone who routinely estimates picojoules per flop at $DAY_JOB - there’s simply no way this is energy efficient. That is not even physically possible with a CPU.

janwas2• 2 years ago

I think the previous code was using dot products, f32 instead of bf16.

aaronscott• 2 years ago

I like to define my subroutines using a modern language like C++, which goes 47 gigaflops. This means C++ is three orders of a magnitude faster than Python. That's twenty years of progress per Moore's law.

This is great. I love the idea of measuring performance differences in “years of Moore’s law.”

Twenty years puts the delta in an easy to understand framework.

JohnKemeny2• 2 years ago

I doubt that you get Python to run faster than C++ at 2004 hardware.

mrtranscendence4• 2 years ago

Python on 2024 hardware vs C++ on 2004 hardware ... I don't think it's obvious that C++ always wins here, though it would depend on the use case, how much of the Python is underpinned by native libraries, and the specific hardware in question.

bevekspldnw• 2 years ago

Honestly depends on what you are doing. Most of my python work is data collection and analysis on top of Postgres.

Being smart in how I use Postgres indexing (and when to disable it outright) has more performance impact than the actual language doing the plumbing.

rbnsl• 2 years ago

Definitely wild we’re in the timeline you can run a 1.1 bn param model on a raspberry pi, but its still tough to justify because the 1.1 is kinda useless compared to the beefier models. Sick for home builds/hobbyists though I might wanna get one of the new Pis just to try this out

JohnnyHerz• 2 years ago

Awesomeness. thank you for sharing!

column• 2 years ago

Unfortunately BitDefender (corporate) blocks llamafile as a ransomware "atc.heur.crypt" and it seems there is no workaround. :(

apitman• 2 years ago

Are the executables not signed? That would be surprising to me for something coming from Mozilla.

EDIT: I just realized that the cross-platform single-binary thing might actually cause issues with code signing. I'm curious about this.

saagarjha3• 2 years ago

One important thing to know if you're considering buying a Mac Studio is that, like the Windows Executive, XNU does a really good job keeping your desktop stable, and that means protecting your system from you. It takes me 45 seconds on Mac Studio to compile the Cosmo monorepo, due to all these safety features; but if I fork bombed it, I'd be surprised if Netflix skipped a single frame.

Clearly nobody actually tried this, because on XNU if you fork bomb the system it reliably goes down every single time. There are no "safety features" here but extra overhead when spawning processes.

Dobiasd2• 2 years ago

Are there any benchmarks on the performance of these new matrix multiplication kernels compared to the Eigen library (ideally for float32)?

Dobiasd2• 2 years ago

While I did not succeed in making the matmul code from https://github.com/Mozilla-Ocho/llamafile/blob/main/llamafil... work in isolation, I compared eigen, openblas, and mkl: https://gist.github.com/Dobiasd/e664c681c4a7933ef5d2df7caa87...

In this (very primitive!) benchmark, MKL was a bit better than eigen (~10%) on my machine (i5-6600).

Since the article https://justine.lol/matmul/ compared the new kernels with MLK, we can (by transitivity) compare the new kernels with Eigen this way, at least very roughly for this one use-case.

jart6• 2 years ago

Here's a complete working example for POSIX systems on how to reproduce my llamafile tinyBLAS vs. MKL benchmarks: https://gist.github.com/jart/640231a627dfbd02fb03e23e8b01e59... This new generalized kernel does even better than what's described in the blog post. It works well on oddly shaped matrices. It needs however a good malloc function, which I've included in the gist. Since having the good memory allocator is what makes the simple implementation possible.

DrNosferatu• 2 years ago

Any performance benchmark against intel's 'IPEX-LLM'[0] or others?

[0] https://github.com/intel-analytics/ipex-llm

LLaMA now goes faster on CPUs

Related Stories