this is before and after the crash. The small peak is the model loading
this is before and after the crash. The small peak is the model loading
Fso it dies immediately
Fset virtual ram = 100 gb
F
Fthis is mine
Fi have 96 GB
Jinstant
Fso 173 - 96 GB is my virtual ram
Fyou have 0 virtual ram
JI'll try
Jwith 100gb of virtual ram is trying to do something
Fyep
Ftrying to load models :d
Fthis is a 14b model not a small thing
Fit is like twice of FLUX DEV
JIt's processing now, but it's ultra slow. +1 hour to generate a video with 50 frames
J
Fmodels are slow
Fbut working 
Fhopefully i will bring teacache i expect 40% speed up
Fyou can also do lower resolution, or 20 steps or lower frames
Ji use for testing 50steps, but a low res like 480x480 or 512x512, this is ready in aprox 9mins, only if i see a good result, i can then start a highres with the same seed. i do batch rendering the whole night, so i wake up and have tons of clips ready 
Fnice strategy
Fi got to be afk now need to sleep see you tomorrow hopefully
FPatreon
Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon
Fnow afk 
Vwhat do you think about this doc..
nvidia refuses to give us more vram
modders increased 24gb vram to 96gb.. problem solved
xDD
you should have bought this..i think it will be better than 5090 ...
nvidia refuses to give us more vram
modders increased 24gb vram to 96gb.. problem solved
xDD
you should have bought this..i think it will be better than 5090 ...
DWhich factory is this?
Vsomewhere in shenzhen..they say 48gb model is stable..96gb version is in testing phase and has some stability issue..
�There are a few people on reit in the LocalLama subreddit who have bought these cards and share their experience. While you get more vram their is no warranty of these cards can die any time.
�This is a great model and gives almost flux quality images at 6b and also comes with apache 2.0 - https://huggingface.co/THUDM/CogView4-6B
J@Dr. Furkan Gözükara with the new unified downloader, should we set the CLIP_G model also? Or only CLIP_L like before?
Fclip g used in SD 3.5
Fauto downloaded
Fit is amazing
Fi would recommend buy if possible
JFor the upscaling model, do you recommend the .safetensors or .pth file?
Nand how to properly monitor video card resources between neural networks on 1 computer?
Fboth works i think
Fcan you eloborate your question?
NI have several neural networks running on 1 system
I want to track how GPU resources are allocated
I want to track how GPU resources are allocated
Fnvitop shows limited info sadly
Fstill gives some info here
F
Fpython exe app py. app py is wan 2.1
RI tested it briefly and was able to generate really good images. I don't think there is a tutorial or guide to fine-tune it, though.
Fi can make installer hopefully but i presume it will come to swarmui soon
RI'm actually positively surprised. It generates images very similar to Flux Dev (I'm using a few prompts I use with Flux).
�Yes it's performance is very close to flux with just half the parameters.
