consolidated vs model safetensors - what's the difference?
Just looking to download this and noticed there are two sets of safetensors files? I assume it's the "model" ones that need to be downloaded?
why are there two copies of the model here. this sucks. what a waste of bandwidth and time...
hugginface-cli download ..... --exclude="*consolidated*" to skip over the consolidated files. I don't think the companies realize not everyone work/live with 10G-100G uplinks. =)
You have to first know there are multiple copies of the model in the repo
By the time you know you've already finished downloading the wrong one and started downloading the right one
Just wish the big guys would look around a little bit with a curiousity to fit in and learn our ways. We have good reasons for our conventions
It seems that the consolidated*-safetentors are the correct ones, according to the Usage in Model card.
Hey all, I opened the initial PR to offer the transformers' implementation out of the box with the repo.
To me it's quite similar to other repos where you might have multiple quants; you don't download all the quants, same as here you wouldn't download both implementations.
How do you get the files locally? Cloning the repo is a way but really isn't the best way to do so IMO: both mistral-inference and transformers will only download the files that they need, and in case you want to use a CLI you can use what
@Qubitium
mentions with huggingface-cli download and the --include and --excluse flags. Using this in conjunction with HF_HUB_ENABLE_HF_TRANSFER will be much much faster than using git clone to obtain the files.
Maybe for future repos we could include download options in the model card for both implementations, so that it's easier to see how to get about to downloading each of the two rather than both (which no-one wants to do, agree with all of you).
Just wish the big guys would look around a little bit with a curiousity to fit in and learn our ways. We have good reasons for our conventions
I understand the issue, but in this case what do you believe would be the best approach? As lysandre mentions this is to be able to offer the transformers implementation. Is there a recommended way you believe it would be better to go with?
Use two repos. (Just like everyone else does)
@ehartford why are you complaining? (again)
You want them to create 2 seperate repos just because you won't take 1 second out of your day to check what files you're downloading?
By the time you know you've already finished downloading
If your internet is that fast, why do you care?
fit in and learn our ways
lol
You made finetunes (which aren't that great) and now you act like you own the place. Get over yourself.
I fully support Mistral on this one. Who cares about the environment and HF bandwidth? You rock Mistral!!!
While we all appreciate what mistral does, cluttering things up like that is silly.
Just learned a magic trick: You can download Mistral Large twice by forgetting to add --exclude="*consolidated*". It's a neat way to waste bandwidth, time, money, and disk space!
Save your breath. They don't care.
Just learned a magic trick: You can download Mistral Large twice by forgetting to add
--exclude="*consolidated*". It's a neat way to waste bandwidth, time, money, and disk space!
lol