Thanks for the report. I've also noticed there is some sort of memory leak. Let me know if your solution works!

from imaginairy.

Thanks for the report. I've also noticed there is some sort of memory leak. Let me know if your solution works!

I'm sorry that I wasn't able to solve the problem completely at this moment. I collecting the torch.cuda.memory_allocated() generated each time,I think the problem may be this pipelined image enhancement,The problem is fix-faces, to be exact. If I have the time so check it out tonight.

imagine with --fix-faces --upscale

repeat.1

Max VRAM used for this generation: 4.29G. Current VRAM utilization:2.51G

repeat.2

Max VRAM used for this generation: 4.29G. Current VRAM utilization:2.72G (leak 200MB)❌

repeat.3

Max VRAM used for this generation: 4.29G. Current VRAM utilization:2.92G (leak 200MB)❌

imagine without --fix-faces --upscale

repeat.1

Max VRAM used for this generation: 4.29G. Current VRAM utilization:2.17G

repeat.2

Max VRAM used for this generation: 4.29G. Current VRAM utilization:2.17G (leak 0MB)✅

repeat.3

Max VRAM used for this generation: 4.29G. Current VRAM utilization:2.17G (leak 0MB)✅

imagine only without --upscale

repeat.1

Max VRAM used for this generation: 4.29G. Current VRAM utilization:2.27G (leak 112MB)

repeat.2

Max VRAM used for this generation: 4.29G. Current VRAM utilization:2.38G (leak 112MB) ❌

repeat.3

Max VRAM used for this generation: 4.29G. Current VRAM utilization:2.49G (leak 112MB) ❌

imagine only without --fix-faces

repeat.1

Max VRAM used for this generation: 4.29G. Current VRAM utilization:2.29G (leak 0MB)

repeat.2

Max VRAM used for this generation: 4.29G. Current VRAM utilization:2.29G (leak 0MB)✅

repeat.3

Max VRAM used for this generation: 4.29G. Current VRAM utilization:2.29G (leak 0MB)✅

from imaginairy.

imagine "xxx" \
--repeats 3 \
        --steps 50 \
        --prompt-strength 7.5 \
        --sampler-type k_lms \
        --width 512 \
        --height 512 \
        --fix-faces

@brycedrennan I believe I've solved this problem by changing the file structure of Codeformer. I'll submit my changes later as PR:)

🤖🧠 imaginAIry received 1 prompt(s) and will repeat them 3 times to create 3 images.

repeat.1
>>   Max VRAM used for this generation: 4.29G. Current VRAM utilization:2.96G
repeat.2
>>   Max VRAM used for this generation: 4.29G. Current VRAM utilization:2.96G
repeat.3
>>   Max VRAM used for this generation: 4.29G. Current VRAM utilization:2.96G

from imaginairy.

awesome thank you. I'd love to know how you output those memory stats as well. we can add a leak detector in the future

from imaginairy.

awesome thank you. I'd love to know how you output those memory stats as well. we can add a leak detector in the future

I see you write "Noisy logs are gone (which was surprisingly hard to accomplish)" in the description of the front page, so no unauthorized addition to the output spoils the beauty, :) I do this locally 👇

                print(
                    f'>>   Max VRAM used for this generation:',
                    '%4.2fG.' % (torch.cuda.max_memory_allocated() / 1e9),
                    'Current VRAM utilization:'
                    '%4.2fG' % (torch.cuda.memory_allocated() / 1e9),
                )

from imaginairy.