Metal079
Metal079
Excuse my ignorance but how do I run it? I seem to get this error trying to run either the lora or SD1.5 version `accelerate launch train_add_distill_lora_sd_wds.py --pretrained_teacher_model C:\Users\Pablo\Documents\mobians_api\sonicDiffusionV4 --train_shards_path_or_url...
Got it running, ran into bug saving though. Validation images also looked like random noise also. ``` Steps: 40%|▍| 400/1000 [35:56
Do this to fix it self.metric = torchmetrics.Accuracy(task="multiclass", num_classes=13) > sir, this is still not working.....can u pls correct the issue....because i am doing one project with this. > […](#)...
damn dead link
I get a similar issue using nexfort in wsl
install the cuda 11.8 version of pytorch not 12.1
> @Froggiehere > > We are now working on new projects: > > https://github.com/chengzeyi/Comfy-WaveSpeed Any plans on diffusers equivalents? I saw parattention but that seemed focused on flux and HunyuanVideo....
Yep still happens
Yep can confirm the same issue as well
> `--keep_vae_loaded` in TRAINER_EXTRA_ARGS in config.env should work around it without meaningfully changing the likelihood of VRAM OOM on 24GB unless you are really pushing the LoRA rank as far...