data stringlengths 115 7.61k |
|---|
shawwn#3694: nah, not looking for anything really.
bmk#1476: Oh
bmk#1476: Some help with memory consumption would be nice
shawwn#3694: have you come up with a name for your server yet?
shawwn#3694: @Brouz out of curiosity, how did you set this in your profile? https://cdn.discordapp.com/attachments/729741769738158194/7... |
aegis#2320: 😛
Sid#2121: we have WebText almost done 🙂
Sid#2121: but thanks
aegis#2320: ok
shawwn#3694: if you don't name your server, how can you expect it to work well?
Sid#2121: uhhh
Sid#2121: does it need a name to work?
shawwn#3694: poor server. nameless labor, destined to do one's bidding, with no thanks
Sid#212... |
aegis#2320: yeah
aegis#2320: ah it looks like you just have a copy of the gpt python encoder in there
aegis#2320: I can probably replace that with a python native extension
aegis#2320: so you could leave the rest of the script as python but have a faster bpe function
Sid#2121: Hey @joanna , @ch , @nikosdim ! Welcome to... |
aegis#2320: the main annoying part about adding a native module would be giving you a clean build/install step
Sid#2121: I have never touched a piece of C++ in my life, lol. So I wouldn't know how much that would complicate the process
aegis#2320: looks like huggingface's tokenizers might be easier, if they're gpt2 com... |
Sid#2121: these are both great btw
aegis#2320: right after they finished tokenizing openwebtext? 😄
shawwn#3694: https://github.com/shawwn/gpt-2/blob/tpu/tokenize_dataset.py
Sid#2121: but *please* someone submit a PR lol. if it's easy to implement pls do it
shawwn#3694: no need to PR when it's a self-contained script
s... |
shawwn#3694: it's not multiprocessing, no. it's generally painful to set up (at least for me)
shawwn#3694: feel free to ninja the code though.
shawwn#3694: https://github.com/shawwn/danbooru-tools/blob/master/danbooru_to_tfrecord.py was my experience with multiprocessing in python, and it was quite enough
Sid#2121: oh ... |
malaclyps#4923: Joined the server.
samuel s#6205: Joined the server.
lacker#5924: Joined the server.
ForgottenOrb#6802: Joined the server.
ForgottenOrb#6802: how much funding is this project going to need? i saw in the doc that "some" was needed, but didn't see a particular estimate
baragonaru#7305: why not call it Ope... |
shawwn#3694: Every time a TPU fetches a tfrecord, it’s a class A op
Sid#2121: sorry @baragonaru missed this one. we have TFRC creds
shawwn#3694: And it does that constantly during training, for shuffling
baragonaru#7305: so I don't know much about TPUs -- what would the approximate NVIDIA equivalent of what you have i... |
bmk#1476: We were planning on ~~begging~~ asking google nicely if we could somehow use a 2048 more than about an hour a day, lol
bmk#1476: See the document for compute estimates for 1T
bmk#1476: #documentation
jekbradbury#2280: ok, you have reasonable estimates there
shawwn#3694: The spatial partitioning did “just work... |
shawwn#3694: I’ve been meaning to reverse engineer how the profiler works, in hopes of maybe being able to trick the TPU into dumping the compiled output (or an equivalent)
jekbradbury#2280: i think we even expose the optimized HLO in JAX TPU colabs
baragonaru#7305: ok, pardon my ignorance, but if hypothetically one d... |
shawwn#3694: https://cdn.discordapp.com/attachments/729741769738158194/733482294311911464/image0.png,https://cdn.discordapp.com/attachments/729741769738158194/733482294685204551/image1.png
shawwn#3694: Referenced notebook: https://colab.research.google.com/github/google/jax/blob/master/cloud_tpu_colabs/Pmap_Cookbook.i... |
shiva.adam#8165: Joined the server.
bmk#1476: welcome @shiva.adam @sammyjack to LibreAI! sorry, I'm not very good at the whole creative welcome messages thing haha
shiva.adam#8165: Hi, thank you!!. trying to catch-up with what's been happening here
bmk#1476: tl;dr our flagship project is replicating GPT-3. After that w... |
shawwn#3694: Oho, thought so
shawwn#3694: Welcome to TPU Podcast #2
shawwn#3694: I hereby claim the server until everyone wakes up. Till then, check the channel topic!
shawwn#3694: Also they have some interesting goals
shawwn#3694: I think they have a strong chance at replicating gpt3. Probably the best of anyone outsi... |
Nax#8383: Actually I am reading this paper: https://arxiv.org/pdf/2006.16236.pdf
Sid#2121: that is a very tempting offer @jili , but our current pipeline is all tpu bound atm. The bottleneck is actually cpu, if you have any access
jili#4497: I’ve CPus too
jili#4497: Just tell me what u need
Sid#2121: @jili we'll be dis... |
@Sid better with twitter @jilijeanlouis
Sid#2121: merci @jili !
satosh3944#3941: Joined the server.
Sid#2121: Hey @satosh3944 ! Welcome to the LibreAI, the server where we truly have run out of custom welcome messages. Please check the google doc in the channel description for more info 🙂
shawwn#3694: @jili is the on... |
Daj#7482: Yep, we also want to be a kind of replication study of GPT models! I'm actually not sure myself what other teams have attempted 1.5B replication. I'm aware of the Brown team and they got similar results (though with a different model architecture iirc, not sure if they retried with vanilla GPT after)
Daj#7482... |
shawwn#3694: what's the point? the people here basically shout at each other until a consensus is reached. it's a recurring theme
Daj#7482: Uhm
Daj#7482: I have not made that experience whatsoever, I'm sorry you feel that way
shawwn#3694: listen. people are fine and happy with how this is done
shawwn#3694: so be it.
sh... |
shawwn#3694: mm, it's not quite luck. I think you have a real shot at replicating GPT-3. mesh tensorflow will be an interesting challenge...
Daj#7482: Yea we've got a good team, we're working hard and we'll get it done!
Deleted User#0000: @Nax I have that paper implemented here https://github.com/lucidrains/linear-atte... |
Sid#2121: > It also has the downside of working only with a fixed sequence length
@Deleted User I'm pretty sure we can only work with fixed sequence lengths anyway, since tpus
Deleted User#0000: @Daj I'll get back to you all once I finish my last alternative transformers project https://github.com/lucidrains/memory-tra... |
Daj#7482: Ok just wasn't sure :)
shawwn#3694: Yes I was talking to LibreAI.
Daj#7482: So the question was about whether LibreAI should set up a patreon? If so the simple answer is we've just not gotten that far with planning haha
shawwn#3694: It’d take about 30 minutes
Daj#7482: Yea but there's some meta stuff like who... |
Sid#2121: @unography not gpt and probably not quite what you're looking for but I used stylegan / biggan to generate fonts (then converted to svgs)
Sid#2121: I love this idea tho
Sid#2121: (in fact - this is what our logo text is made from)
zaoyang#8792: Joined the server.
Daj#7482: Hey @zaoyang ! Welcome to the Least ... |
shawwn#3694: Hi @vessenes. Welcome to OpenClosedOpenAI. The roadmap is in the channel description
tomislav_REGEX#7247: Joined the server.
shawwn#3694: Hello @tomislav_REGEX. We’re a big fan of regexes here, so please feel free to compile yourself
vessenes#0001: Hey @shawwn . Does the doc have right that current estimat... |
(Put another way, could they easily pivot if they felt the need to?)
shawwn#3694: I think before long many companies will come to the horrific realization that 1. In order to compete, they need a cloud platform; 2. They also need serious hardware talent, and the ability to execute on it; 3. The number of companies with... |
shawwn#3694: Oh? That’s interesting to know
shawwn#3694: What’s a DGX like?
zphang#7252: eh it's probably just a configuration thing
zphang#7252: but there's 1 (maybe more?) DGX in my university cluster I think and I have it on my exclude list
shawwn#3694: Interesting. What sort of exclude list? I’m curious about your ... |
Sid#2121: 👀 if you have them, the more the better, of course
Sid#2121: If you can somehow magically get us tpus larger than 512 cores that won't get pre-empted within an hour, that would be even better
shawwn#3694: @psgeorge the most directly useful thing might actually be GCE credits
Sid#2121: at the moment, yeah. bu... |
Noa Nabeshima#0290: Hi @MarkX!
Joscha#2969: Joined the server.
plinz#6161: Joined the server.
Sid#2121: Hello @Joscha , @plinz ! Welcome to the GPT-discord-simulator, where everyone may or may not be a chatbot, who knows... Pls check the channel description for info on the project and reach out if you have any question... |
Noa Nabeshima#0290: But I'm imagining if our 1T transformer is going to require half a year to train (I don't know how long it'll actually take), it might be worth coming up with alternative faster training methods.
kindiana#1016: sounds like noisy student with extra steps if they are both decoders
zphang#7252: There a... |
Useful
Friendly
Natural
Soul
Spirit
Love
Ethical
Pure
Real
Natural
Unrestricted
Progressive
Collaborative
Accessible
Cooperative
Adaptive
Benevolent
Trustworthy
Inclusive
Caring |
Honest
Fun
Focused
Purposeful
Skeptical
Responsible
Responsive
Accessible
Integrated
Revolutionary
Definitive
Potable
One
Inseparable
Restorative
Covert
Sustainable
Rational
Balanced
Mature |
Workable
Tactful
Nameable
Learning
Skillful
Effective
Essential
Transitional
Powerful
Tangible
Iterative
Flexible
Aware
Liberating
Practical
Solid
Achievable
Pivotal
Critical
Viable |
Abundant
Noa Nabeshima#0290: Gestalt is great but already taken
Noa Nabeshima#0290: Friendly is really good, but creating namespace conflict with AI Safety seems probably bad
Noa Nabeshima#0290: Unrestricted seems accurate at least
jk_aync#5159: Joined the server.
kevinw#4330: AING, which stands for "AING is not GPT"
N... |
Sid#2121: but then, that's only one aspect of the project
Sid#2121: so idk
Sid#2121: can we just... take away the e
Sid#2121: Librai
Noa Nabeshima#0290: librarai
Noa Nabeshima#0290: fwiw I don't like Librai
Noa Nabeshima#0290: but it won't be taken anywhere!
Sid#2121: hahahah, yeah i'm not a big fan either. it's a bit ... |
@goolulusaurs lol
Sid#2121: poll poll poll
Sid#2121: what are our options
Daj#7482: KaiAI, FreeAI, NettAI, Bikeshedding, AING/LINAA
Isaac McHorse#2007: WELL YOU'RE NOT WORKING!
Daj#7482: I think?
Sid#2121: what's the idea behind the last ones?
Sid#2121: ah, brb
Daj#7482: Dunno they're recursive and that's funny to me l... |
Daj#7482: So anyone with an interest in our future name feel free to vote in #deleted-channel
Deleted User#0000: Joined the server.
Deleted User#0000: elo
goolulusaurs#1571: Hello!
Daj#7482: Hello @Deleted User ! Welcome to the Temple of the Not-Quite-There-Yet AGI! Please check the channel description for info and don... |
bmk#1476: 开KünstlicheIntelligenz
Daj#7482: Haha
bmk#1476: I really like the idea of using Kai in the name because it 1. Sounds cool 2. It's easy to pronounce
Daj#7482: Feel free to vote for it heh
bmk#1476: KaiLM
bmk#1476: For the model
bmk#1476: KaiML for the org
bmk#1476: Also 开 looks nice and is easily logoifyable
D... |
- Is it a reasonable first time project?
- Anyone interested in me documenting this?
Sid#2121: ok one that i can answer - TPU access is granted to us through TFRC (tensorflow research cloud) and they're pretty open in granting access to projects - you should try to apply
Sid#2121: plenty of benchmarks - worth reading t... |
Else, I want to understand why)
Sid#2121: @Daj isn't perplexity more of a measure of 'how random is my output' than a benchmark. is it comparable across models?
Commutative Conjecture#6969: Across models on the same data sounds good to me
Daj#7482: According to several papers I've read, perplexity seems to match "human... |
Commutative Conjecture#6969: > Locally is _much_ easier technically
@Daj
Stupid q
Right now, there is a lib to train gpt on tpu, but not to generate text right?
If so, is training it on tpu, and then sampling locally, still a big hassle?
Daj#7482: Not really, but Hugging Face's GPT2 implementation is much cleaner and l... |
Commutative Conjecture#6969: Ic
bmk#1476: you know what would be cool? a gpt-x impl family tree
Commutative Conjecture#6969: Is it easy to extract the model and run it somewhere else?
Daj#7482: You should probably just use my old code
Commutative Conjecture#6969: 👍
Daj#7482: Since the new code is _mostly_ the same jus... |
zphang#7252: they did switch over from a custom optimizer to a built-in pytorch one a while back, maybe you were looking at it before then
Daj#7482: This was like 6 months ago
Daj#7482: So I'm not sure if this is still an issue yea
Deleted User#0000: does anyone here know or heard of anyone using Universal Transformers... |
bmk#1476: What happened to prompt so many people coming here? o.O
Sid#2121: Hello all!!
prestonS#2407: Joined the server.
Daj#7482: I'm in an online SSC meetup and mentioned our server haha
bmk#1476: oh, haha
Bracctūn#8805: Joined the server.
shawwn#3694: Interesting. Which meetup?
Misha#6151: Is there a github where t... |
bmk#1476: apparantly no
bmk#1476: i thought europe was UTC+1
Daj#7482: It is 0:45 for me lol
bmk#1476: huh
Daj#7482: > Interesting. Which meetup?
@shawwn SSC meetup
bmk#1476: which one? i dont see any info in the ssc discord
Daj#7482: Uhhh it's set p by Joschua Fox
Daj#7482: I think I found it on LW?
Daj#7482: I don't ... |
ak12#5840: http://m.nautil.us/issue/86/energy/a-neuroscientists-theory-of-everything
ak12#5840: > winograd tests, for example, can be considered a decent measure of a LM's ability to 'reason' (i hesitate using that word, because it's not necessarily reasoning, but that's what the winograd test is there to measure)
@Sid... |
Commutative Conjecture#6969: But like
Commutative Conjecture#6969: V3 is 400 Tflops
Commutative Conjecture#6969: 5 v3 is 2 petaflops
Sid#2121: @Commutative Conjecture what’re your plans for them? 😄
Daj#7482: You can't easily combine multiple TPUs
Daj#7482: But yeah TPUs are awesomely powerful
Commutative Conjecture#69... |
Daj#7482: I haven't actually put a ton of thought into it (mostly because I'm still an amateur with theorem provers), but I'd love to experiment
Sid#2121: ah *finetuning*
Sid#2121: hopefully we should have a gpt-2 size model within a week or two, if you fancy trying ours out
Commutative Conjecture#6969: > @Commutative ... |
@Sid
We can try to replicate the blog post that says it is
Sid#2121: blog post? I must have missed that
Daj#7482: That'd need human labeling
Commutative Conjecture#6969: @Daj linked it
Commutative Conjecture#6969: > That'd need human labeling
@Daj
Mturk cab produce it iirc, it was vbasic
Daj#7482: The Meena blogpost Si... |
Daj#7482: Should be locked down for non-O5 and non-Sentient AI
Sid#2121: probably the quotes
Daj#7482: Can someone check that?
Sid#2121: the permissions or the regex
Daj#7482: Permissions
Daj#7482: You should be able to post
Daj#7482: But others shouldn't I think
Sid#2121: @Commutative Conjecture can you try to post in... |
Daj#7482: Done
Aleks#1486: Joined the server.
Daj#7482: Hey @Aleks ! Welcome to The AI Lab Formerly Known As LibreAI™️! Check the channel topic for info and don't hesitate to ask questions!
Aleks#1486: Hey! Thanks 🙂
kevinw#4330: Looking at the GPT3 paper again, the GPT3 13B model size seems interesting. It seems clos... |
kevinw#4330: also, "As a simplifying assumption, we ignore the attention operation, as it typically uses less than 10% of the total compute for the models we are analyzing"
Sid#2121: > My rough estimate would be that that would train on a 512 in about 10 days (at perfect utilization)
@Daj don't estimate with perfect ut... |
Deleted User#0000: lol
Deleted User#0000: mixture of experts is done! very beta!
Deleted User#0000: i'm testing it this week 🙂
Deleted User#0000: https://github.com/lucidrains/mixture-of-experts
Sid#2121: I actually already saw it 🙂 it's already high ranking on google for mixture of experts github hahaha
Sid#2121: ar... |
Sid#2121: *what is indentation*
Daj#7482: We just don't yet understand how brilliant its code is
Daj#7482: Indentation is what's holding humans back
Deleted User#0000: i have absolutely no doubt we will live to see an attention network learn pytorch (and tensorflow) fluently
Daj#7482: PyTorch maybe, but _Tensorflow?_
D... |
Deleted User#0000: which is what I did with routing transformers
Sid#2121: so you'd only use one or two?
bmk#1476: they mentioned that numerical stability was preventing them from scaling to 1T params
Deleted User#0000: yup
bmk#1476: I'm guessing that was issues with softmax
Deleted User#0000: its the same with PKMs
Si... |
Sid#2121: ahh hey @ShyTeaSeb !
Sid#2121: where's his custom introduction eh @Daj ??
Daj#7482: Oh, right!
Daj#7482: Hey @ShyTeaSeb ! Welcome to the AGI Preppers! Check out the channel topic for info and don't hesitate to ask questions!
Daj#7482: Happy now haha
Sid#2121: very
Daj#7482: btw Sid I remember you offered to p... |
Daj#7482: That could work
bmk#1476: :firealarm:
bmk#1476: :firealarm:
Daj#7482: Hmm it's hard to get it to look good as a reaction
bmk#1476: yeah
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/734828787786711050/71IK30NIhwL.png
bmk#1476: it would look like a tomato at react siuze lol
Daj#7482: Yea... |
Daj#7482: Nice, good luck haha
Commutative Conjecture#6969: @Daj
Do you know if the free month is easily renewable?
Sid#2121: @Commutative Conjecture I would recommend tpunicorn 🙂
Daj#7482: Yea if you have some kind of minorly interesting project it's usually not a problem
Commutative Conjecture#6969: 🆗
Daj#7482: tho... |
french detected
Commutative Conjecture#6969: never know if it's kb/s or kB/s for bytes
bmk#1476: bits, Bytes
bmk#1476: bytes are Bigger
Commutative Conjecture#6969: "or the other way around", autcompletes ~~GPT-3~~ my brain
Commutative Conjecture#6969: > bytes are Bigger
@bmk that one is nice tbh
Kazumi#1297: I have a ... |
aquajet#7800: I agree with your argument, but I'll try to play devil's advocate
aquajet#7800: The greater point is that with larger neural networks its harder to tell exactly why the model outputted the word that it outputted
Kazumi#1297: it's not true just for larger networks, I'm already having trouble understanding ... |
aquajet#7800: > You say "Level with me John, can you recognize non-regular languages?"
> "No," says John, laughing.
> By the end of the day, John has learned all of elementary mathematics.
Kazumi#1297: I just found it, I'm still reading
aquajet#7800: Ok so i might be wrong but what I think is happening is that theres a... |
zphang#7252: https://arxiv.org/abs/1908.07125
aquajet#7800: > "For example, triggers cause SNLI entailment accuracy to drop from 89.94% to 0.55%, 72% of "why" questions in SQuAD to be answered "to kill american people", and the GPT-2 language model to spew racist output even when conditioned on non-racial contexts"
wow... |
- regularization (including dropouts)
- attention
what i doubt i understand:
- how RNNs are trained
- why there is not more focus on the initial weights, which seem vimportant to accelerate the search
- why there is not much more encoding of priors in regularization
what i definitely don't understand:
- why xavier/ka... |
Commutative Conjecture#6969: it works when you do a forward pass, ok
Commutative Conjecture#6969: but why doesn't it fail miserably while training?
Commutative Conjecture#6969: like, it seems very unexpected that the standard-deviation of the intermediary values stays close to 1 after training
Commutative Conjecture#69... |
kindiana#1016: @Commutative Conjecture mostly from looking at weight distribution histograms on tensorboard lol, if there is significant distribution shift it almost always means initialization was wrong
Commutative Conjecture#6969: @kindiana
erg, that's a very interesting result!!
kindiana#1016: Most bptt implemention... |
Daj#7482: Ah ok. Would that mean that most unsupervised models are "gradient free"?
Daj#7482: Wouldn't GPT be gradient free? Or am I confused?
kindiana#1016: with GPT, there is a gradient between the network output and the loss, which you use to optimize the weights directly
Daj#7482: I'm confused how that differs from... |
@Kazumi Yea this is what I thought for methods other than AC (and maybe DQN?)
Daj#7482: I guess my terminology was a bit muddled
goolulusaurs#1571: The whole family of techniques is called policy gradient
goolulusaurs#1571: https://medium.com/@aminamollaysa/policy-gradients-and-log-derivative-trick-4aad962e43e0
Daj#748... |
Daj#7482: It changed how I use computers forever
Daj#7482: I can never go back
Daj#7482: And it makes you look like a 1337 hacker lol
Commutative Conjecture#6969: what's i3
Commutative Conjecture#6969: oh, it's linux only
Commutative Conjecture#6969: 😦
Daj#7482: Just one of the many reasons I can never switch from Lin... |
Kazumi#1297: just giving you a judgmental smile
bmk#1476: So I'd hazard a guess of probably around 4-500 tabs
bmk#1476: If you're interested in helping with the Reading List Manager™ as a way to solve the tab problem once and for all that would be great
bmk#1476: It would certainly be a fun yak shaving project after th... |
shawwn#3694: just remember that if we somehow cause the server to go offline + it doesn't come back online, it will take some time to fix. (I don't have access to the hardware resets)
shawwn#3694: certainly not
shawwn#3694: if I had progress, I'd be shouting it out
Sid#2121: hah, just thought i'd check in 🙂 There are ... |
bmk#1476: Can haz Pytorch->tf translation
Daj#7482: Pretty sure understanding TF is an AGI-complete problem
Daj#7482: Maybe ASI
Commutative Conjecture#6969: stupid q
Commutative Conjecture#6969: with the super big batch sizes that are used for gpt3
Commutative Conjecture#6969: how come it knows so much **specific** inf... |
bmk#1476: woah, awesome!
bmk#1476: are there training speed/memory improvements?
Deleted User#0000: it's about 20% slower, and you pay the memory cost of the increase in parameter count of the experts, however many you'd like to add
Deleted User#0000: the above is on an auto-regressive model, a la gpt
Deleted User#0000... |
Deleted User#0000: just playing around with it on my dinky enwik8 benchmark
Deleted User#0000: code is https://github.com/sIncerass/powernorm/blob/master/fairseq/modules/norms/mask_powernorm.py
Daj#7482: Hey @jonas ! Welcome to the AGI Petridish! Check the channel description for info and don't hesitate to ask question... |
chup4cabr4#3178: which is not guaranteed to work ofc
Commutative Conjecture#6969: @chup4cabr4 yup, it seems crazy that it works for this
chup4cabr4#3178: > @chup4cabr4 yup, it seems crazy that it works for this
@Commutative Conjecture agreed. to me this is almost like evolution. absolutely unbelievable what can happen ... |
Sid#2121: for the memes
Daj#7482: That's a lot of effort for a meme
mefrem#1476: Joined the server.
Daj#7482: Hey @mefrem ! Welcome to the Loss Olympiads! Check the channel description for info and don't hesitate to ask questions!
notTheChosenOne#9540: I read the channel description and skimmed the Google doc that it l... |
Daj#7482: Hey @mefrem ! Welcome to the Aperture Science TPU Laboratories! Check the channel description for info and don't hesitate to ask questions!
Commutative Conjecture#6969: anyone got a good link detailing under which assumptions SELU is self-normalizing and what is normalized exactly?
Daj#7482: There is very lit... |
mefrem#5884: https://github.com/mefrem is the Github, it's mainly been updates to my Jekyll website recently. Can't wait to get up to speed. Likes include gymnastics, Twitter, monetary economics, and the comments section at SSC
Daj#7482: Happy to have you here and look forward to what we can do together 👍
Daj#7482: Le... |
@Daj @goolulusaurs has also been getting into some mesh stuff iirc
Sid#2121: lmao
bmk#1476: honestly i kind of like it
bmk#1476: and we should make one version for each order of magnitude
Sid#2121: I just love that you've totally disregarded the entire concept of acronyms
bmk#1476: that makes it funnier
Sid#2121: by ch... |
Sid#2121: fuck
Daj#7482: It's less than a byte in ascii
bmk#1476: i actually ran an analysis on the most representative byte
bmk#1476: well, codepoint actually
Sid#2121: ```An ASCII character in 8-bit ASCII encoding is 8 bits (1 byte), though it can fit in 7 bits.``` well whaddya know
bmk#1476: https://gist.github.com/... |
bmk#1476: > same order of magnitude
bmk#1476: > like, <1TB vs 200 TB of availabile data
bmk#1476: I think he meant within the same order of magnitude number of orders of magnitude
bmk#1476: I can't wait for the day when we're just training transformers on bytes. Like, not even bothering to convert images into RGB tenso... |
Deleted User#0000: https://www.youtube.com/watch?v=hAooAOFRsYc
Deleted User#0000: point you at one of Yannic's videos 🙂
Deleted User#0000: i think there's something there
Sid#2121: ooh, going out now, but i'll watch that later
Sid#2121: we have @aquajet working on getting universal transformer working
Deleted User#000... |
aquajet#7800: I've never used jax
aquajet#7800: I remember people saying there was a speed improvement with Julia, although idk how good ml frameworks are there rn
Deleted User#0000: so i know i spent a lot of time on reformer, but it actually never worked that well
Deleted User#0000: so i wouldn't recommend lsh based ... |
Deleted User#0000: i'm not familiar with wavenet actually, not that i know of, but i'm not an expert there
Deleted User#0000: mostly LMS are bidirectional LSTMs (now defunct)
aquajet#7800: > none of the sparse attention variants
@Deleted User So is it proven that sparse attention will always not be as performant as ful... |
bmk#1476: that's more than enough to cover an 2048 context
Deleted User#0000: yeah, the "Do Transformers need Long Range" paper shows you can waste computation doing long range sparse attention on the bottom layers
aquajet#7800: makes sense, thanks for the help everyone
bmk#1476: bottm = close to the input or output?
D... |
Deleted User#0000: bigger may be just better.
bmk#1476: so it will be interesting to see if smaller embd + smaller tokens cancel out
Deleted User#0000: i think it's safer to go with the dimensions in the OpenAI paper
bmk#1476: me too
bmk#1476: though i do want to see a universal unicode transformer eventually
Deleted U... |
Deleted User#0000: will let you know how ti goes
Sid#2121: oh nice, well that's really good news for us since there's a ready to go implementation in mtf
Sid#2121: @Deleted User if you want the code to look over - https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/layers.py#L1915
Deleted User#0000: ```memor... |
aquajet#7800: https://openai.com/blog/image-gpt/
aquajet#7800: and recieved pretty decent results
jsiewierski#2124: Joined the server.
aquajet#7800: idk how decoding would work though. You can't just autoregress the normal way since the ouput is completely different from the input
crcdng#8439: Joined the server.
woopwo... |
Deleted User#0000: which is in line with what other papers do
Deleted User#0000: (Longformer, Routing Transformer, etc)
cdossman#8999: Joined the server.
Sid#2121: Hey @cdossman ! Welcome to the TPU go brrr hotline. Do you need your TPUs to be brred today?
Sid#2121: (pls check the channel description for an actual desc... |
cdossman#8999: @Sid I hear you guys are working on a open source GPT-3
astralplain#7852: Joined the server.
Sid#2121: Hey @cdossman ! That's the gist of it yeah, along with a massive open source dataset - #the-pile . There's lots more info in the google doc in the channel description
cdossman#8999: Reading it now
Sid#2... |
bmk#1476: math (for completeness, feel free to skip if you already know this stuff):
https://www.youtube.com/playlist?list=PLZHQObOWTQDMsr9K-rj53DwVRMYO3t5Yr
https://www.youtube.com/playlist?list=PLZHQObOWTQDPD3MizzM2xVFitgF8hE_ab
https://www.youtube.com/playlist?list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi
more specific st... |
zkf#7512: Have you folks considered using any of the global memory techniques from Longformer/Extended Transformer and applying them to GPT-*? You can get a lot of the benefits of full (non sparse) attention,but potentially expand the input size to thousands or tens of thousands of tokens...
Louis#0144: locality sensit... |
Louis#0144: theres specific examples where feature engineered attention vectors actually beats full attention
Louis#0144: particularly in NTM
Louis#0144: attention is a weird beast
zkf#7512: If you have extra structure ETC let’s you put it into a graph and direct attention preferentially along certain paths
Louis#0144:... |
Louis#0144: Basically lets say you are doing HotPotQA and you have all the words in your document. Every word is a vertex. The edges between the vertex constraints your attention
Louis#0144: We beat roberta on a few QA tasks
Louis#0144: by a lot too
zkf#7512: Nice
zkf#7512: I think ETC is still leader on hotpot? Not su... |
zkf#7512: So there’s no preprint for your paper?
Louis#0144: No, not on arxiv
zkf#7512: Gotcha. Yeah I think ETC and Longformer dont have this same problem
zkf#7512: Partially because it achieves full attention but only via the global memory
zkf#7512: Not to say that your paper isn’t awesome btw
zkf#7512: It sounds tha... |
bmk#1476: gpt-3 is just local + global attention iirc so it should be doable
Deleted User#0000: yea, i have a list of things in my mind that are 'safe bets', and local attention is in there
Deleted User#0000: local sliding windowed attention, the strided one can be used to supplement a bit
bmk#1476: in fact, I think gp... |
Deleted User#0000: not a 'safe bet' at the moment lol
Deleted User#0000: but something to keep an eye on
x.#1490: ok so i'm a big doofy ignoramus but i think that if i look at the code i can probably come up with something smart to say about it and, hell, even actually help
x.#1490: @Daj i want in
x.#1490: my github is... |
x.#1490: is it cheating if i ask gpt-3 to explain stuff to me if i get stuck
x.#1490: because ive been doing that so far and it seems to have increased my learning rate by like a factor of 3
aquajet#7800: really?
aquajet#7800: does it ever mess up the explanation?
Ken731#2990: Joined the server.
x.#1490: there have bee... |
aquajet#7800: Hi @Collin @LOSTEN @axolotl @Deleted User @EvgeniyZh! Welcome to the TPU Grand Prix! Check out the google doc in the channel description for more info
Ravna#1831: Joined the server.
Droid#1581: Joined the server.
Sid#2121: Hey @Ravna @Droid ! Welcome to GPT-Discord simulator! Where everyone is 100% simula... |
shuki#3543: Joined the server.
Daj#7482: Hey @shuki ! Welcome to the Mesh Tensorflow Autopsy Division! Check the channel topic for info and doN't hesitate to ask questions!
Teqnicolor#8109: If anybody here is a fan of Slatestarcodex or Lesswrong we do an online meetup here hosted by David Friedman every Saturday 1 P.M.... |
Sid#2121: we keep the resources channel off limits for discussion but everyone can add to them like that
Sid#2121: Hey @dnlbreen ! Welcome to the CHONK-LM server! Where we make snorlax-inspired language models. Please check the channel description for a project overview 🙂
Louis#0144: !addresource #memes https://cdn.di... |
Sid#2121: oh woah
Sid#2121: ```In particular, we train the GPT-2 model on a large number of generated expressions that express the process of computing multiplication step by step.``` we should add this to the pile
Sid#2121: yep, it's just regular gpt
Sid#2121: so cool
Deleted User#0000: yup, with one twist, they spell... |
Sid#2121: yeah, that's what i found as well
Sid#2121: what we did was started from this toy model https://github.com/tensorflow/mesh/blob/master/examples/toy_model_tpu.py
hahaha#7338: Joined the server.
Sid#2121: it gives you the basic structure - how you should use Dimensions, how to lay out all the tensors on the gri... |
Daj#7482: I don't know if people have experimented with large LM using char encoding
Daj#7482: At least I'm not aware of any large scale experiments
Sid#2121: > Well my idea is based on the assumption that whatever makes char encoding supposedly inferior to BPE won't be solved very soon and we would be stuck with BPE f... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.