Gianluca Detommaso

icon company Helsing icon location Berlin icon location AI Research at Helsing. Before Senior Applied Scientist at AWS. PhD in Statistical Applied Mathematics.

Results 3 comments of Gianluca Detommaso

Mesh in restore checkpoint when target is None

Thanks a lot for the explanation, it is clearer now! Meanwhile, let me signal [this](https://github.com/google/orbax/issues/381) issue I raised in the Orbax repo. I found out that `restore_args_from_target` gives problems when...

`no_repeat_ngram_size` has no effect for Flax model

@mayankagarwals @sanchit-gandhi It seems also `num_return_sequences` does not work for the Flax model, and indeed it seems missing from [transformers/src/transformers/generation/flax_utils.py](https://github.com/huggingface/transformers/blob/151425ddb29d4ad1a121e8cce62000a2ac52d3ba/src/transformers/generation/flax_utils.py#L488). Worth doing a more general comparison of functionalities, I guess....

[WIP] Add Llama Flax Implementation

Hi @vvvm23 and @sanchit-gandhi, do you guys have a timeline for this effort? Asking because I would love to import FlaxLlama from Hugging Face, but if it is going to...