Conversation

Notices

1. Embed this notice
   VD-15 (vd-15@fediffusion.art)'s status on Wednesday, 25-Oct-2023 03:57:24 JST VD-15
   Oh yeah, this thing can train LoRAs alright. Shame I can't seem to get SD itself to run on multi GPU...
   • Embed this notice
     THOT POLICE (s8n@posting.lolicon.rocks)'s status on Wednesday, 25-Oct-2023 03:57:19 JST THOT POLICE

     in reply to
     @VD-15 you can get multiple instances running simultaneously. You should set your thermal throttling to kick in at 80c and ensure proper ventilation to your cards.
     
     I have been running 2 instances of a1111, if you provide the envvar CUDA_LAUNCH_BLOCKING=1, 2, etc. (zero indexed, zero is default) it will select the first available CUDA processor starting at the given device ID. So I just have a webui-user2.bat that has set CUDA_LAUNCH_BLOCKING=1
     
     this isn't a perfect solution because some files cannot be read by both instances simultaneously on windows or something, but if this were linux I could run multiple installs off an overlaid models directory with a bind mount or symlink
   • Embed this notice
      (mint@ryona.agency)'s status on Wednesday, 25-Oct-2023 03:57:19 JST 

     in reply to

     • THOT POLICE
     @s8n @VD-15 It *does* look like Linux from screenshot, but Windows had junctions (folder symlinks) since 2000 and file symlinks since XP, even though they added the mklink command only in 10.
     https://en.wikipedia.org/wiki/NTFS_links?useskin=vector
   • Embed this notice
     VD-15 (vd-15@fediffusion.art)'s status on Wednesday, 25-Oct-2023 03:57:20 JST VD-15

     in reply to

     • THOT POLICE
     @s8n This was just from training a LoHA across 3 cards.
     
     I'm getting about 6it/s at peak when I run SD proper, but I'm limited to a single card at the moment. Even if I could run it with multigpu, I don't want to because running all three of these at once is probably a fire hazard.
   • Embed this notice
     THOT POLICE (s8n@posting.lolicon.rocks)'s status on Wednesday, 25-Oct-2023 03:57:21 JST THOT POLICE

     in reply to
     @VD-15 what's your iterations per second looking like, are you using progressive upscaling?
   • Embed this notice
     VD-15 (vd-15@fediffusion.art)'s status on Wednesday, 25-Oct-2023 03:57:22 JST VD-15

     in reply to
     Yeah, these cards reach like 90C under load. it's enough to cause the 3D printed brackets I made to start deforming because of the heat. I need to figure out how to cool these things better
   • Embed this notice
     VD-15 (vd-15@fediffusion.art)'s status on Wednesday, 25-Oct-2023 03:57:23 JST VD-15

     in reply to
     Also holy shit, the airflow in my cupboard is terrible, who could've seen this coming?
   • Embed this notice
     THOT POLICE (s8n@posting.lolicon.rocks)'s status on Wednesday, 25-Oct-2023 04:10:29 JST THOT POLICE

     in reply to

     • 
     @mint @VD-15 idk if that cuda arg works in linux or if there's a similar-but-different one, haven't yet fucked with that. All I have right now is a pair of 980tis in a windows box, unsure if it would be worth moving them to the linux machine if I upgraded to a different card
      likes this.
   • Embed this notice
      (mint@ryona.agency)'s status on Wednesday, 25-Oct-2023 04:11:58 JST 

     in reply to

     • THOT POLICE
     @s8n @VD-15 I think SD needs at least 6 gigs to generate images without --lowvram, and at least 12 to train loras/embeddings/whatever.
   • Embed this notice
     THOT POLICE (s8n@posting.lolicon.rocks)'s status on Wednesday, 25-Oct-2023 04:14:59 JST THOT POLICE

     in reply to

     • 
     @mint @VD-15 I use --medvram, I think the 980ti is the only maxwell card with 6 gigs
      likes this.
   • Embed this notice
     THOT POLICE (s8n@posting.lolicon.rocks)'s status on Wednesday, 25-Oct-2023 04:15:51 JST THOT POLICE

     in reply to

     • 
     • THOT POLICE
     @mint @VD-15 you should have seen how epically space engineers choked on having 12 gigs of vram
      likes this.