Patching LoRAs for KModel slow with Automatic (fp16 LoRA).

[ZeroCool22]

Patching LoRA by precomputing model weights.
Patching LoRAs for KModel:  94%|█████████████████████████████████████████████████▊   | 286/304 [01:16<00:02,  6.48it/s]ERROR lora diffusion_model.single_blocks.32.linear1.weight Allocation on device
Patching LoRA weights failed. Retrying by offloading models.
Patching LoRAs for KModel: 100%|█████████████████████████████████████████████████████| 304/304 [01:58<00:00,  2.56it/s]
LoRA patching has taken 118.75 seconds
Moving model(s) has taken 130.87 seconds

[kalle07]

i have also 16GB for me works this checkpoine fine https://civitai.com/models/638187?modelVersionId=721627

maybe try to choose the "Diffusion in Low Bits" on top not automatic but 16bit !

[ZeroCool22]

maybe try to choose the "Diffusion in Low Bits" on top not automatic but 16bit !

It's already selected, check the image again.

[kalle07]

i said NOT automatic try all 3 options in 16bit and may one work with the checkpoint above ... but anyway FLUX is still not that stable ...

[ZeroCool22]

@lllyasviel

Why the patching is so damn slow?

Isn't supposed if we use Automatic (fp16 LoRA) the patching must be almost instantly?

[kalle07]

2 weeks ago for me worked if i reduce in top line the "GPU weight" -> 7000 MB GGUF and LORA isnt yet very well programmed .... BE PATIENT ! anyway flux isnt that good at the moment, lets say only in some small parts

[ZeroCool22]

2 weeks ago for me worked if i reduce in top line the "GPU weight" -> 7000 MB GGUF and LORA isnt yet very well programmed .... BE PATIENT ! anyway flux isnt that good at the moment, lets say only in some small parts

I understand, but i think FLUX is amazing, the Inpainting give the most perfect results i have seen so far and the best part, is you are not obligated to do a min. res of 1024x1024 (neither on normal txt2img) like SDXL needs. It's amazing that you can get perfect results at any resolution. That make things so much faster.

[ZeroCool22]

reduce in top line the "GPU weight" -> 7000 MB GGUF

But doing that, you are not taking the benefit of all your VRAM (dunno what GPU you have).

[kalle07]

16GB RTX and if you read the CMD lines it tryes to free/clean the vram ... for what ever ... and if that fails it lora-patch was slow ... but this was 2 weeks ago ... now with the checkpoint i mentioned is all ok

[ZeroCool22]

16GB RTX and if you read the CMD lines it tryes to free/clean the vram ... for what ever ... and if that fails it lora-patch was slow ... but this was 2 weeks ago ... now with the checkpoint i mentioned is all ok

It shouldn't matter what checkpoint you use.

Anyway, now seems to work fine again....