download and activate FLUX from huggingface is a real pain

[madias123]

I don't know where the problem is: Even with active token downloading and using FLUX models is a real pain and it works 1 out of 10. In worst case, I have to delete the user /.cache folder (on my root drive... I hate that) under Windows 11 and download everything again, it seems, that SD.next won't send the right token:
Next thing is, that downloading from huggingface stops without reason only a server restart restarts the download - often with ~5-10GB of loss data. (I have a perfect fiberglass internet connection, without interrupts)

18:52:20-576658 DEBUG    Cache clear
18:52:21-051643 INFO     HF login: token="hf_...FfMO" fn="C:\Users\test\.cache\huggingface\token"
18:52:21-178609 ERROR    Auth: repo="black-forest-labs/FLUX.1-schnell" login=True 401 Client Error. (Request ID:
                         Root=1-69331bd5-4a74d4c77bb5bb674b058767;22d9f21e-41d2-4cfe-82c9-ab4431663b4f)
                         Cannot access gated repo for url
                         https://huggingface.co/api/models/black-forest-labs/FLUX.1-schnell/auth-check.
                         Access to model black-forest-labs/FLUX.1-schnell is restricted. You must have access to it and
                         be authenticated to access it. Please log in.
18:52:21-181806 DEBUG    Load model: type=Flux repo="black-forest-labs/FLUX.1-schnell" cls=FluxPipeline offload=balanced
                         dtype=torch.bfloat16 args={'torch_dtype': torch.bfloat16}
18:52:21-183816 DEBUG    Load model: transformer="black-forest-labs/FLUX.1-schnell" cls=FluxTransformer2DModel
                         subfolder=transformer quant="None" loader=default args={'torch_dtype': torch.bfloat16}

Is there a better solution, instead of downloading directly in sdnext?
Why sdnext "forget" those tokens?

[madias123]

...again:

stucks at 97% and sdnext server is dead without log.

[madias123]

...and surprise, surprise: download is starting completely at 0%... like downloading in 1999.
here are my sys specs, I'm on the actual dev version:

[vladmandic]

i understand the frustration, but can we cut down on ironic tone - think of ppl that you're asking to help you, how does this motivate me?
anyhow:

18:52:21-051643 INFO HF login: token="hf_...FfMO" fn="C:\Users\test.cache\huggingface\token"

is this correct key?
you say sdnext forgets to use correct key, but did not specify what is correct.

I have to delete the user /.cache folder (on my root drive... I hate that)

why would you delete entire /.cache folder??
and if you dont want it to be on root drive, specify wherever you want.

that downloading from huggingface stops without reason

it works for 99% of users, so this is not a general issue.
but still, there are multiple download methods supported by sdnext and you can change them in settings.
since you only posted very small screenshot and not actual log, i cant really say which download method you're currently using,

...and surprise, surprise: download is starting completely at 0%... like downloading in 1999.

typically download continues where it stopped. again, if its restarting from 0, it indicates something else is wrong.

[madias123]

ok, sorry for being ironic, but it's hard to setup any SD environment on windows (red carpet) for me, but I needed it @work.
After cleaning up everything it seemed to work. It was really a problem with no space on my root device, where the .\cache folder is located. This is my big questions:
Why is everything relative to the install directory, eg. "K:\SDnext\sdnext" in my environment, but the .\cache folder (it's nothing related to windows folders/paths) is on "C:\Users\test.cache" so on a completely other drive? Which program set this to this path? Git or sdnext? This is totally strange, because it wont follow the rules of my system paths (e.g. is my tmp/temp folder on a different drive).
The better question would be: Where can I chance this? ;) Didn't find it here:

thank you for your big effort required to make this project possible.

Just one thing:
I don't know what equipment most people here are using, but I find that most "downloadable" models within SDnext (apart from SD and SDXL) aren't compatible with a 16GB card, especially Flux. With Flux, I absolutely have to use a variant like the "Flux deve fp8," which isn't suggested.

That's one of the few things I like about Amuse: it immediately checks compatibility, so users can quickly see if a model is even remotely suitable for their machine.

regards Matthias

[vladmandic]

re: cache - look for "Folder for Huggingface cache" setting
re: large models, yes - and we note in many places that quantization (not enabled by default) and offloading (enabled by default) are strongly recommended. but no, fp8 is never recommended. its a really bad numeric format. something like sdnq quant type using uint4 will be 2x better quality and 50% less memory than fp8.