-
Notifications
You must be signed in to change notification settings - Fork 152
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
When CMD downloads large models, it always interrupts. #232
Comments
Tbh, it's never a good idea to download large files not with a download manager. It may seem convenient for smaller sized files, yes, but large models, like used for SD, text generation and now speech is just too much error-prone. Especially if your connection is not the fastest. |
I agree. |
This issue is linked with the other issue made by martjay in the suno repository - suno-ai/bark#505 |
@martjay I recommend creating an issue here as well - https://github.com/huggingface/huggingface_hub/issues since this is the "source" of the failing download. |
Did you all download the model without interruption? It's so frustrating. |
I did see many installation problems when I was elsewhere. But where I live gigabit connections are the standart. |
It's not a good choice to assume things based on yourself. You have always to consider the worst possible thing. And since gigabit connections are nowhere "standard" all around the world, there is always the problem of failing while downloading with slow connections. Therefore, the best possibility right now, is to provide the needed URL and the save location on the hard drive, in case the user has to download manually. The devs just have to implement this known information, for example, like this: |
HuggingFace creates the downloader software that is built in to the models that are then integrated into this project. |
A downloader software is/would not be needed if a model is downloadable from their website. And still, if it is that way, it makes what I wrote even more important. |
Just to reclarify - I'm not saying that the download process is amazing and
it doesn't screw people over. What I am saying is that it's not my code
that a) does the download b) chooses how things are downloaded.
I can only change very few downloads, and those are not the big ones like
bark etc.
Yes bark is open source and could be modified, and it could even be me who
modifies it, but that's the point - bark, tortoise, audio craft and more
projects would need to be changed.
This isn't stable diffusion where it's just a checkpoint. Bark and
tortoise, for example, are two entirely different projects on almost every
level. Meanwhile something like Realistic Vision and Anythingv5 are
basically the same thing, same model, same architecture everything, just a
different checkpoint (ckpt) file.
The only model that can be reasonably self downloaded is tortoise, probably
because there are different checkpoints and fine tuning. My repo does not
prevent this. You can choose your own checkpoint files.
That's why I'm saying - if HuggingFace download is so prevalent, they ought
to make it work. Obviously downloading big files shouldn't be an issue, we
have had open source torrent software for ages. If they can't make it work,
we need to start pressuring the projects with an alternative that does work.
…On Thu, Dec 21, 2023, 2:10 PM MusicHunter ***@***.***> wrote:
HuggingFace creates the downloader software that is built in to the models
that are then integrated into this project.
A downloader software is/would not be needed if a model is downloadable
from their website. And still, if it is that way, it makes what I wrote
even more important. There is no reason to not show the user WHAT is
downloaded and to WHERE it's saved.
[image: image]
<https://private-user-images.githubusercontent.com/6757283/291204691-309fc976-0426-46bd-bb9e-1e9988728e49.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.8XN0ZUG0FNuXIGBYgR1zsVK3LqonDbaDqg0-GmvKofs>
downloading
"4dda87e5dfafc1b59351131d9610002b06fbecc50793a0e3c8ab2e534176cd7f" to
C:\Users.cache\huggingface\hub\models--facebook--musicgen-melody\blobs
Sometimes it's not a matter of internet speed, you know? Some countries
are unable to connect to the huggingface website and require VPN access,
but I have also verified this issue and the answer is no. Because even if I
change several VPNs, the download will still be interrupted, but if I
download all the models from the huggingface website, it never interrupts,
and even if it interrupts, I can still continue. But when downloading CMD,
if I interrupt, everything starts over and it keeps looping. I don't
understand the mechanism of huggingface, maybe the download address will
automatically refresh? I'm not sure, but the model URL and save directory
are necessary, really.
—
Reply to this email directly, view it on GitHub
<#232 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/ABTRXI32VK2DT6KW7XTSBUDYKQRIZAVCNFSM6AAAAABATQLOMWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQNRWGE2DEMRSGY>
.
You are receiving this because you commented.Message ID:
***@***.***>
|
Are you also unable to determine the name of each project download model and the folder where it is saved? It's really frustrating. I need to find a solution to the download issue for all AI projects, but currently, that's all I have in mind. |
Maybe I can test it with a similar VPN at the end of next month. I know it's frustrating, with that VPN I had to download and install automatic1111 multiple times since several things kept failing. Because of how HuggingFace works, the model names are hashes, like my screenshot above. It might be possible to determine this. Maybe there's a better solution as well. There must be many people facing the same problem. |
Yes, I hope to solve this problem. It's really a big problem because almost all project models rely on huggingface downloads |
pytorch_model.bin: 12%|███████▎ | 157M/1.26G [01:00<07:02, 2.61MB/s] The above exception was the direct cause of the following exception: Traceback (most recent call last): During handling of the above exception, another exception occurred: Traceback (most recent call last): |
I found a solution today. edit "file_download.py" force_download=True, Can successfully complete the download now. |
That's awesome! |
When CMD downloads large models, it always interrupts. I don't know how to solve this problem, so can you tell me all the models link that need to be downloaded and the save path?
The text was updated successfully, but these errors were encountered: