Clément Dumas

icon indicating a website link butanium.github.io icon indicating an email [email protected]

icon company ENS Paris-Saclay icon location Paris-Saclay, France icon location I'm Clément, student at ENS Paris Saclay, France.

Results 87 comments of Clément Dumas

[Feature Request] Make token indexing work for batched input and UnifiedTransformer

Oh wait, seems like token indexing is supposed to work only with `tracer.invoke` calls. It would be nice if it also works on directly batched input, not sure how easy...

[Feature Request] Make token indexing work for batched input and UnifiedTransformer

Ok but with UnifiedTransformer, token indexing doesn't work as padding side is right by default : ```py l = ["ab dfez zd", "a", "b"] from nnsight import LanguageModel model =...

[Feature Request] Make token indexing work for batched input and UnifiedTransformer

@JadenFiotto-Kaufman I guess we can close this as UnifiedTransformer is no longer supported?

Make tensor creation functions like torch.zeros traceable via nnsight

Isn't that already supported ? I've run remote code with `torch.arange` and it worked iirc

Error when using stopping_criteria in .generate if remote=True

Hi, sorry I posted the wrong trace, I got a recursion error too 😅 I'll look into it

Error when using stopping_criteria in .generate if remote=True

So if I understand correctly, as every stopping criteria is a different class inheriting StoppingCriteria it might not be possible for nnsight to support this argument on remote execution ?

Error when using stopping_criteria in .generate if remote=True

Stopping criteria is an abstract class meant to be inherited. If I understand correctly you'd need to manually add a set of class inheriting StoppingCriteria right ?

Does not check if authors' names are bolded

Almost missed this. Would be good if it were flagged indeed

IndexError when flat inputs are concatenated during trace

Not sure why you're using NNsight rather than LanguageModel here. Here is a MWE of the issue. If NNsight is used, both example fail, if LanguageModel is used both example...

Support flash attention `flash-attn --no-build-isolation` with `uv sync`

Just to make sure: there is no magic uv flags to make this work on Windows without going through custom wheels / full build from scratch (according to https://github.com/Dao-AILab/flash-attention/issues/1469 this...