Jack BAI
Jack BAI
@jdvin I think maybe you should @ one of the authors of this repo. I used your patch and it works nicely overall. One issue is that the output representations...
@jdvin yeah I think it's better to just close this PR as the system has modification restrictions. Very grateful to your contribution here, it helped many of us getting the...
What error did you encounter exactly? I'm also trying to add a new model
Yeah the `pip install numpy==1.21.5` solution does not work for me either. Has anyone solved this problem? ``` lm = ecco.from_pretrained('valhalla/t5-small-qa-qg-hl')#, gpu=False) output = lm.generate(text, generate=20, do_sample=True, attribution=['ig', 'grad_x_input']) ```...
Just fixed the issue. Downgrading your `transformers` to `4.13.0` solves the problem. This is due to an API update from `transformers` update. @SumitDasTR @Peter-Zhoutuanjie cc @jalammar
Note that this solution is a quick fix. If you want to use models like `OPT`, `LLAMA`, etc., you still need to fix the code from source to make it...
Hey @srikanthmalla , do you have any update on this? I can also help with it.
@LAKan233 @srikanthmalla : use this lib instead: https://github.com/inseq-team/inseq It's a more recent and maintained one.
Same issue! However, we're only seeing this issue on a single node of B200x8. It does not happen when in multi-node setting (2-node or 3-node). cc @JJJYmmm
From my understanding this is not possible unless you do algorithmic optimization (some math). Salience maps is doing backprop from output to embedding. This process is just chain rule, and...