Is it possible to run this on something other than google's cloud platform?
GPT-J-6B: 6B JAX-Based Transformer
81–84 of 84 posts
Re: GPT-J-6B: 6B JAX-Based Transformer
#82Is it possible to run this on something other than google's cloud platform?
I'm running it comfortably on my 3090, although it's a really snug fit for the VRAM, and that's with a number of fixes to significantly reduce its memory use from https://github.com/AeroScripts/mesh-transformer-jax .
Re: GPT-J-6B: 6B JAX-Based Transformer
#83Second I think we may need to try a series of different targeting options. We've tried fitness people, vitamins people, and coffee people. The fitness people were closest, but the other two didn't care. I think we could get them to care if we use different messaging. For example, we really haven't said anything about why this tablet should go in your coffee and not your water. We need to talk more about the flavoring aspect of this to the coffee people and tell them that it's so much cheaper than creamer and tastes better - plus it makes them healthy whereas creamer will give them cancer.
For vitamin people...well I think we've been misreading who these are. People who care about vitamins are already too healthy to use a standard multi-vitamin. They have their own supplement stack they've developed over years. We should be targeting the couch potatoes of the world. People with an interest in fast food, or specific chains like Wendy's, BK etc. Our messaging can basically be, "are you lazy af but want to get 1% healthier? Drop this in your coffee and go." It feels like that's the multi-vitamin demographic.
If we run out of other stuff, I think we could try those two streams.
Re: GPT-J-6B: 6B JAX-Based Transformer
#84Earlier quoted context omitted.
I'm running it comfortably on my 3090, although it's a really snug fit for the VRAM, and that's with a number of fixes to significantly reduce its memory use from https://github.com/AeroScripts/mesh-transformer-jax .
Out of curiosity, how fast are your inferences with this setup?
I'm not sure yet what settings I'll end up with if I decide to play with this more. per_replica_batch=3, seq=1024, gen_len=64 would give an experience roughly similar to the AI Dungeon that I'm used to, though less clever than the Dragon model, and a bit slower at about 10 seconds per batch.