pretrain_opt6.7b not working

[oaefou]

How to fix it ?

1.mp4

[goblin776655]

Same here!
Plz fix!

[Tps-F]

Can't you generate captions for other images?
and please show me config

[oaefou]

On any images pretrain_opt6.7b doesn't work. 2.7b works well.

[goblin776655]

Can't you generate captions for other images? and please show me config

All default, just change the model from 2.7b to pretrain_opt6.7b

[Tps-F]

I will check the operation again. I will try to fix it by tomorrow.

[Tps-F]

By the way, can you tell me what GPU you are using?

[oaefou]

5900x / 32gb ram / rtx 3090

By the way, can you tell me what GPU you are using?

[Tps-F]

Can I get screenshot or something of what is in %UserProfile%\.cache\huggingface\hub ?

[oaefou]

Can I get screenshot or something of what is in %UserProfile%\.cache\huggingface\hub ?

[oaefou]

.cache\torch\hub\checkpoints

[Tps-F]

Thank you
The model is downloading fine and the extension itself is working fine.
It might be a problem on the BLIP2 side, I'll try some things.

[LexCybermac]

Same issue here, 3900x/32gb ram/3090. Fingers crossed that a fix isn't too much work to find.

[Tps-F]

It still works in my environment.
Can you verify once with CPU? but I know it would take an awful lot of time.

change

sd-webui-blip2/scripts/main.py

Lines 23 to 29 in 9f6ce7b

	def get_device():
	if torch.cuda.is_available():
	return "cuda"
	elif torch.backends.mps.is_available():
	return "mps"
	else:
	return "cpu"

to

def get_device():
    return "cpu"

[oaefou]

It still works in my environment. Can you verify once with CPU? but I know it would take an awful lot of time.

change

sd-webui-blip2/scripts/main.py

Lines 23 to 29 in 9f6ce7b

	def get_device():
	if torch.cuda.is_available():
	return "cuda"
	elif torch.backends.mps.is_available():
	return "mps"
	else:
	return "cpu"

to

def get_device():
    return "cpu"

Does not work.

[Tps-F]

Thank you for testing

By the way, is the time until the end of the program shorter than others?
on cuda

[R3xpook]

For me it works fine...tho i got cuda OOM i don't know if this repo have some code borrowed from https://github.com/pharmapsychotic/clip-interrogator but i was able to run 6.7b (it has a low vram option)

[Tps-F]

The code in that repository is not in use, but I will implement half option #13 soon!

[jcrsantiago]

Any updates on this issue?

[Tps-F]

Sorry, I've been busy and haven't gotten around to much.
I will do it as soon as possible.