Software Engineering Courses (SECourses)•3y ago

Hi, i'm having issues with setting up xformers with the A5000 GPU in SD Dreambooth Whenever i click

Hi, i'm having issues with setting up xformers with the A5000 GPU in SD Dreambooth

Whenever i click Generate Class Images i'm getting :

Exception generating concepts: No operator found for memory_efficient_attention_forward with inputs: query : shape=(1, 2, 1, 40) (torch.float32) key : shape=(1, 2, 1, 40) (torch.float32) value : shape=(1, 2, 1, 40) (torch.float32) attn_bias : p : 0.0 flshattF is not supported because: xFormers wasn't build with CUDA support dtype=torch.float32 (supported: {torch.bfloat16, torch.float16}) tritonflashattF is not supported because: xFormers wasn't build with CUDA support dtype=torch.float32 (supported: {torch.bfloat16, torch.float16}) triton is not available requires A100 GPU cutlassF is not supported because: xFormers wasn't build with CUDA support smallkF is not supported because: xFormers wasn't build with CUDA support max(query.shape[-1] != value.shape[-1]) > 32 unsupported embed per head: 40

I've set everything based on your youtube tutorial here https://www.youtube.com/watch?v=Bdl-jWR3Ukc&t=158s

Kkicedz Hi, i'm having issues with setting up xformers with the A5000 GPU in SD Dreamboo...

Furkan Gözükara SECourses•3/22/23, 12:58 PM

generate class images with txt2img

Furkan Gözükara SECourses•3/22/23, 12:58 PM

much faster better

FFurkan Gözükara SECourses ```set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.6,max_split_size_m...

Garreth210•3/22/23, 12:58 PM

With or without the --

GGarreth210 With or without the --

Furkan Gözükara SECourses•3/22/23, 12:58 PM

i think like this can you test?

Garreth210•3/22/23, 12:58 PM

does not work

Garreth210•3/22/23, 12:58 PM

with the --

Furkan Gözükara SECourses•3/22/23, 12:59 PM

there is no -- the command.

Garreth210•3/22/23, 12:59 PM

trying without the --. I thought that is how the args are being separated

Garreth210•3/22/23, 1:00 PM

nope

Garreth210•3/22/23, 1:00 PM

does not work

GGarreth210 trying without the --. I thought that is how the args are being separated

Furkan Gözükara SECourses•3/22/23, 1:00 PM

let me show you

Furkan Gözükara SECourses•3/22/23, 1:00 PM

you are writing incorrectly

Garreth210•3/22/23, 1:00 PM

I see now

Garreth210•3/22/23, 1:00 PM

ahahah

GGarreth210 oh

Furkan Gözükara SECourses•3/22/23, 1:00 PM

Garreth210•3/22/23, 1:00 PM

Yup

Garreth210•3/22/23, 1:01 PM

I was being stupid

Garreth210•3/22/23, 1:01 PM

fingers crossed this works

Garreth210•3/22/23, 1:01 PM

then I will try train again with drambooth, set some checkpoints and then XYZ select from the checkpoints

GGarreth210 then I will try train again with drambooth, set some checkpoints and then XYZ se...

Furkan Gözükara SECourses•3/22/23, 1:04 PM

i think dreambooth still wouldnt work on 8 gb :d

Furkan Gözükara SECourses•3/22/23, 1:04 PM

but worth to try let me know

Garreth210•3/22/23, 1:04 PM

will do

Garreth210•3/22/23, 1:04 PM

I am just finding that runpod is harder than running it locally on windows

Garreth210•3/22/23, 1:05 PM

but I don't get close to the image quality your SD is producing

FFurkan Gözükara SECourses generate class images with txt2img

kicedzOP•3/22/23, 1:06 PM

On text2img im still getting the same error
NotImplementedError: No operator found for memory_efficient_attention_forwardmemory_efficient_attention_forward with inputs: query : shape=(2, 4096, 8, 40) (torch.float16) key : shape=(2, 4096, 8, 40) (torch.float16) value : shape=(2, 4096, 8, 40) (torch.float16) attn_bias : <class 'NoneType'> p : 0.0 cutlassFcutlassF is not supported because: xFormers wasn't build with CUDA support flshattFflshattF is not supported because: xFormers wasn't build with CUDA support tritonflashattFtritonflashattF is not supported because: xFormers wasn't build with CUDA support triton is not available requires A100 GPU smallkFsmallkF is not supported because: xFormers wasn't build with CUDA support dtype=torch.float16 (supported: {torch.float32}) max(query.shape[-1] != value.shape[-1]) > 32 unsupported embed per head: 40
Time taken: 1.82sTorch active/reserved: 2230/2246 MiB, Sys VRAM: 3577/24257 MiB (14.75%)

What could be wrong? I'm using SD 1.5.ckpt, and probably the text2img has nothing to do with Dreambooth

kicedzOP•3/22/23, 1:07 PM

python: 3.10.6 • torch: 1.13.1+cu117 • xformers: 0.0.17+6d62b0e.d20230322 • gradio: 3.16.2 • commit: a9fed7c3 • checkpoint: 4c86efd062

I'm currently new to this, apologies if i ask pretty basic questions.

Garreth210•3/22/23, 1:09 PM

No luck with dreambooth on a 3070ti 8gig of vram

Zono50•3/22/23, 1:10 PM

quick question, i can download LORA files from the civit.ai and not have to use prompts to invoke them, how can I do that with my lora model i'm training?

Zono50•3/22/23, 1:22 PM

also my models eyes are super blue and red, not sure why that is, i took the models photos, did face restoration, then used the face restored photos for my dataset