You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Oh sorry! Yes this one is the model https://huggingface.co/togethercomputer/GPT-JT-6B-v1. Does this have similarities with other models that I could try to implement or even use directly? I would have no idea where to start but willing to help
Ah, nice! We support GPT 2, so maybe that can be used as a building block? Or at least you can compare GPT 2 Python's implementation with our GPT 2 and then to the same to implement your own. :)
The reference hf/transformers implementation of GPT-J is here. The implementation should be for the most part similar to any other text model we have, like GPT-2. By a brief look I think we may need to adjust/extend our attention implementation to support the rotary position embedding, but it's fine to modify the current code as necessary and we can find the best way to make it configurable later.
Activity
josevalim commentedon Jan 29, 2023
The above is an app, do you have a link to the model implementation and params from HF? :)
lorenzosinisi commentedon Jan 30, 2023
Oh sorry! Yes this one is the model https://huggingface.co/togethercomputer/GPT-JT-6B-v1. Does this have similarities with other models that I could try to implement or even use directly? I would have no idea where to start but willing to help
josevalim commentedon Jan 30, 2023
Ah, nice! We support GPT 2, so maybe that can be used as a building block? Or at least you can compare GPT 2 Python's implementation with our GPT 2 and then to the same to implement your own. :)
lorenzosinisi commentedon Jan 30, 2023
Thanks! I will give it a try :) thanks! Any MR I can look at for reference?
BTW this model already knows some Elixir
❤️
jonatanklosko commentedon Jan 30, 2023
The reference hf/transformers implementation of GPT-J is here. The implementation should be for the most part similar to any other text model we have, like GPT-2. By a brief look I think we may need to adjust/extend our attention implementation to support the rotary position embedding, but it's fine to modify the current code as necessary and we can find the best way to make it configurable later.
lorenzosinisi commentedon Feb 11, 2023
Working on this but it is gonna take a while 'cause I am new to transformers
[-]Support for togethercomputer/GPT-JT ?[/-][+]Support GPT-J[/+]