Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Meta AI announces Massive Multilingual Speech code, models for 1000+ languages

github.com

Post reply on HN
Print view View original

181–190 of 232 posts

Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages

#181

Post by chaxor »

sneak wrote: post #109
itake wrote: post #104

Earlier quoted context omitted.

Yep... you spend hours messing around with docker containers and debugging all the weird build errors. I am less familiar with storing data in a db (for ml hosting concerns), but I'd imagine it would add overhead (as opposed to accessing files on disk). You also have to deal with hosting a db and configuring the schema.

You "spend hours messing around" with everything you don't know or understand at first. One could say the same about writing the software itself. At its core Dockerfiles are just shell scripts with worse syntax, so it's not really that much more to learn. Once you get it done once, you don't have to screw around with it anymore, and you have it on any box you want in seconds. In either case you have to spend hours sc…

'shell scripts with worse syntax' lol I wish shell could emulate Alpine on a non-linux box. Shell script with worse syntax for config VM may be closer to a qemu cloud init file.

Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages

#182

Post by throwaway2037 »

pdntspa wrote: post #163
simonw wrote: post #126

Earlier quoted context omitted.

I'm having trouble imagining what data I would store in a database as opposed to a filesystem if my goal is to experiment with large models like Stable Diffusion.

I would take GP's kind of dogmatic jibber jabber with a grain of salt. There is an unspoken and timeless elegance to the simplicity of running a program from a folder with files as state

Isn't terminfo db famous for this filesystem-as-db approach? File Vs DB: I say do whatever works for you. There is certainly more overhead in the DB route.

Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages

#183

Post by hzay »

armatav wrote: post #54

Imagine if we used these types of models for like 500 years and it locked their vocabulary in time, disallowing any further language blending; then somehow the servers turned off and nobody could communicate across language barriers anymore. Someone should write that down in some sort of short-story involving a really tall structure.

Imagine if someone said this about dictionaries.

Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages

#184

Post by woodson »

throwme_123 wrote: post #160

fairseq is a fairly missed naming opportunity. C-3PO would have been better.

fairseq has existed for many years and is Fb’s (now Meta’s) AI research group’s sequence-to-sequence modeling toolkit. It’s not specific to this MMS model.

Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages

#185

Post by fnordpiglet »

itake wrote: post #144
fnordpiglet wrote: post #122

Earlier quoted context omitted.

aws s3 sync + image snapshot

all that costs time :-/ When you're finished working, you got to wait for the syncs to finish (or I guess run them all the time?). For home computer setups, you can simply walk away when you need a break.

No, you write a script that runs the sync and shut down the instance. When for instance is stopped you don’t pay for it. Resuming it is a simple api call. You don’t even really need to do the sync, it’s just to be certain if the instance volume is lost you have a backup.

The shutdown / stop on an instance is like closing the lid on your laptop. When you start it again it resumes where it was left off. In the mean time the instance doesn’t occupy a VM.

A caveat is you can’t really do this with spot instances. You would need to do a sync and rebuild on start. But, again, scriptable easily.

Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages

#186

Post by helsinkiandrew »

itake wrote: post #106
nomel wrote: post #91

Earlier quoted context omitted.

Besides some great tooling out there if you wanted to roll your own, you can literally rent windows/linux computers, with persistent disks. If you have good internet, you can even use it as a gaming PC, as I do.

Is there an easy way to off-board the persistent disk to cheaper machines when you don't need the gpus? Like imagine, setting up and installing everything with the gpu attached, but when you're not using the gpu or all the cpu cores, you can disconnect them. If you have docs on how to do this, please let me know.

With AWS (and probably most other cloud VPC services) the disk is remote from the hardware so you can halt the CPU and just pay for the storage until you restart.

AWS also provides accessible datasets of training data:

https://aws.amazon.com/marketplace/search/results?trk=868d87...

Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages

#187

Post by throwbackhere »

crazygringo wrote: post #56
m3kw9 wrote: post #36

The problem with all these model releases is they have no demos or even video of it working. It’s all just download it and run it, like it’s an app.

They're intended for researchers/professionals not consumers, and I'm not sure how a video is going to be helpful? And the issue with a live demo is that these are resource-intensive, they're not just webpages. It's an entire project to figure out how to host them, scale them to handle peaks, pay for them, implement rate-limiting, and so forth. For the intended audience, download-and-run-it doesn't seem like an issue…

Even researchers and professionals appreciate a sample or two.

Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages

#189

Post by akiselev »

itake wrote: post #76
nharada wrote: post #70

Earlier quoted context omitted.

Wait why is renting a GPU in the cloud not a solution? You can even try multiple options and see which ones are capable enough for your use case. Look into some barebones cloud GPU services, for example Lambda Labs which is significantly cheaper than AWS/GCP but offers basically nothing besides the machine with a GPU. You could even try something like Vast in which people rent out their personal GPU machines for chea…

My annoyance was managing state. I’d have to spend hours installing tools, downloading data, updating code, then when I want to go to bed I have to package it up and store as much as I can on s3 before shutting off the $$ server.

You can try skypilot: https://skypilot.readthedocs.io/en/latest/

It handles storage, setup, etc for machine learning work loads across several providers - which helps a lot if you need one of the instances that rarely have capacity like 8x A100 pods.

Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages

#190

Post by anonzzzies »

nharada wrote: post #70
qwertox wrote: post #68

I would like to use stuff like this as a side-project. Buy a Nvidia Geforce GPU and stick it into my 24/7 server and play around with it in my free time, to see what can be done. The issue with all these AI models is that there's no information on which GPU is enough for which task. I'm absolutely clueless if a single RTX 4000 SFF with its 20GB VRAM and only 70W of max power usage will be a waste of money, or really…

Wait why is renting a GPU in the cloud not a solution? You can even try multiple options and see which ones are capable enough for your use case. Look into some barebones cloud GPU services, for example Lambda Labs which is significantly cheaper than AWS/GCP but offers basically nothing besides the machine with a GPU. You could even try something like Vast in which people rent out their personal GPU machines for chea…

I have not seen many setups that wouldn’t pay itself back (including energy in my case) within a year (sometimes even 6 months) with buying vs renting. For something that pays itself back that fast, and that is without renting it out myself, just training with it, I cannot see how I would want to rent one.

Edit; on Lambda labs, the only exception seems to be the H100; it would be 1.5 years or so, but even 2 years would still fast enough. I have an A100 which paid itself back; thinking of getting another one.

Post reply on HN