History

Patrick Esser 81d560ef1f edge with extra channel		2022-08-06 10:42:32 +00:00
..
resume_512	old uncommitted stuff	2022-07-23 09:42:18 +00:00
resume_512_improvedaesthetic	old uncommitted stuff	2022-07-23 09:42:18 +00:00
resume_768_hr	old uncommitted stuff	2022-07-23 09:42:18 +00:00
v1-upscaling-f16-pretraining-512-aesthetics	launchers	2022-07-28 00:12:03 +02:00
v1_edgeinpainting	edge with extra channel	2022-08-06 10:42:32 +00:00
v1_iahr_torch111	take restored gs into account for decay	2022-07-28 21:55:17 +00:00
v1_iahr_torch111_ucg	resume v1ucg	2022-07-31 23:24:33 +00:00
v1_improvedaesthetics	final v1 restart	2022-07-18 23:47:36 +00:00
v1_improvedaesthetics_torch111	test gpu scripts and other launchers and configs	2022-07-22 09:56:22 +00:00
v1_inpainting_aesthetics-larger-masks	launchers	2022-07-28 00:12:03 +02:00
v1_inpainting_aesthetics-larger-masks-ucg	inpainting with dropping configs and slurmys	2022-08-03 23:32:00 +02:00
v1_inpainting_improvedaesthetics_torch111	slurm launcher	2022-07-24 13:28:19 +02:00
v1_laionhr_torch111	test gpu scripts and other launchers and configs	2022-07-22 09:56:22 +00:00
v2_laionhr1024	v2 on laionhr 1024	2022-07-14 23:36:08 +00:00
v2_laionhr1024_2	v2 1024 scripts	2022-07-23 10:30:11 +00:00
v2_pretraining	print ema steps and resume v2 pretraining	2022-08-02 20:32:22 +00:00
v3_pretraining	resume	2022-07-27 14:26:51 +00:00
README.md	ready to slurm	2022-07-06 22:52:52 +00:00

Example

Resume f8 @ 512 on Laion-HR

sbatch scripts/slurm/resume_512/sbatch.sh

Reuse

To reuse this as a template, copy sbatch.sh and launcher.sh somewhere. In sbatch.sh, adjust the lines

#SBATCH --job-name=stable-diffusion-512cont
#SBATCH --nodes=24

and the path to your launcher.sh in the last line,

srun bash /fsx/stable-diffusion/stable-diffusion/scripts/slurm/resume_512/launcher.sh

In launcher.sh, adjust CONFIG and EXTRA. Maybe give it a test run with debug flags uncommented and a reduced number of nodes.