Mayank Jobanputra comments

Results 9 comments of


                                            Mayank Jobanputra

Run CI on Windows?

Can someone help me figure out how I should download a model to "tmp" folder on windows using `snapshot_download` so that I don't have to clean it up later? I...

FARMReader should convert local hf-transformers model to FARM if necessary

Closing this as it is being fixed by #3663 .

Feature: Training tracking support with multiple loggers (i.e. Tensorboard, wandb)

imo it is relevant, but I will discuss it with you offline!

Both demos broken

@wochinge The airgapped demo is supposed to throw the error: Because it's for education purposes and not for anyone just to use it directly. I agree this should be documented...

ImportError: cannot import name 'BaseDocumentStore' from partially initialized module 'haystack.document_stores.base' (most likely due to a circular import) (haystack/haystack/document_stores/base.py)

@bhawnapiryani can you share the command you used to install haystack? Also maybe the environment details (maybe which cloud you're using). I am unable to reproduce this locally on my...

ImportError: cannot import name 'BaseDocumentStore' from partially initialized module 'haystack.document_stores.base' (most likely due to a circular import) (haystack/haystack/document_stores/base.py)

@MozerWang can you share the environment details? like OS, python version, Haystack version?

ImportError: cannot import name 'BaseDocumentStore' from partially initialized module 'haystack.document_stores.base' (most likely due to a circular import) (haystack/haystack/document_stores/base.py)

@MozerWang I will check with the team on Monday if we have any guidelines on CentOS internally. For the cmd "pip install "farm-haystack[all]", the issue is with one package, we...

pretraining loss blows up with Triton Flash attention

I am facing the same issue, gonna try to switch to the torch implementation. Will report my findings here.

pretraining loss blows up with Triton Flash attention

@yzhang123 I was able to avoid the loss explosion by increasing the warm-up steps 3x than default and changing the learning rate scheduler to linear decay (with warm up). Maybe...