My terrible experience with GPU cloud services

[ghost]

[jn-jairo]

I also have a weak GPU, it takes about 11s/it so for 20 steps I have to wait 3m40s to generate one image.

Renting a GPU is a good option but I couldn't find a practical way to use stable diffusion there, and the paid services that I found doesn't have all the cool features this repo has or are too expensive for the amount of images you can generate.

As a web developer I am thinking to create my own website to enable to run the stable diffusion in a rented GPU and automate the process, also, I would like to have a easy way to create a model using dreambooth.

I am planning to release it for free, so people would just need to pay the rented GPU directly to RunPod/Vast.ai, and maybe offer a paid subscription and rent the GPUs myself as a more practical alternative and to make some money, what would help me a lot because I am currently without any income :(

For now I am in the phase of finding out what I need to make it and if it is worth the effort.

I am accepting suggestions.

[smnnekho]

I'm using Runpod for this repo for quite a while and it's working great and at a very low price.

Updating the repo ist the same process you would have locally (git pull, maybe check new requirements) and i was always able to run the newest version.

The template also has been updated to a more recent version last week.

So i had a completly different experience with runpod so far

[davidgoering]

this, just git pull, pip install requirements, restart pod and your golden. Works well for me :)

[jus3q]

how do you connect to the Runpod instance? Juypter?

[ShinCoreSys]

ive been running this on colab just fine.

[slooi]

Do you have a guide on how to set it up? I'm a complete newbie to the machine learning ecosystem so I'd be very interested reading it and seeing the process. Please let me know if you do

[Miraculix200]

Try this. It's quite easy to get it running:
https://colab.research.google.com/github/Miraculix200/StableDiffusionUI_Colab/blob/main/StableDiffusionUI_Colab.ipynb

If you have a free SageMaker Studio Lab account, you can use this:
https://github.com/Miraculix200/StableDiffusionUI_SageMakerSL

[acheong08]

For the most simplified experience, https://colab.research.google.com/drive/1PvNyEWIhDU_D-i15DzpPjqDQkbYv_6Hu is probably best.

[AISPAISP]

How about a ready made service like https://aws.amazon.com/marketplace/pp/prodview-v2qrtpqachrtc

[vaalberith]

Is it possible to run this on yandex cloud? g2.mpeg (4 vCPU, 1/8 GPU A100) costs $0.311 in non-priority mode. They are also charged by seconds. This should be beneficial if you find a way to enable only during generation. I tried to run AUTOMATIC1111 in DataSphere but my skills are not enough.

Did You manage to use it? I am interested too. Thanks!

[mr-grixa]

@vaalberith
I was able to run only the original SD according to the instructions from huggingface. AUTOMATIC1111 throws errors. I'm not sure if it can work with a split graphics card at all.

[mr-grixa]

@vaalberith
Ты вроде по русски понимать должен. Запуск в яндексе видел у другого чела, на g2.mpeg выдает ошибку недостатка памяти, возможно виртуально разделенные видеокарты вообще не поддерживаются. Из адекватных аналогов которые можно оплатить нашел только immers.cloud но смотри цены, иногда есть смысл преплатить барыгам с платиру и арендовать runpod.io и его аналоги. Я изначально все это ради DreamBooth и большого диска под несколько моделей искал. Но сейчас и DreamBooth оптимизировали, и модели до 2 гигов сжимать можно. Теперь можно комфортно в гугл колабе работать.

[dKosarevsky]

@vaalberith Ты вроде по русски понимать должен. Запуск в яндексе видел у другого чела, на g2.mpeg выдает ошибку недостатка памяти, возможно виртуально разделенные видеокарты вообще не поддерживаются. Из адекватных аналогов которые можно оплатить нашел только immers.cloud но смотри цены, иногда есть смысл преплатить барыгам с платиру и арендовать runpod.io и его аналоги. Я изначально все это ради DreamBooth и большого диска под несколько моделей искал. Но сейчас и DreamBooth оптимизировали, и модели до 2 гигов сжимать можно. Теперь можно комфортно в гугл колабе работать.

А есть линк на инфу о запуске у другого чела?

[mr-grixa]

А есть линк на инфу о запуске у другого чела?
Я в телеграмм яндекс облака спрашивал. Кто-то сказал что делал. Но там не автоматик, а обычный запуск по официальной инструкции.

[0xspeedrunner]

I'm curious if anyone has gotten this running in a cloud that only charges for GPU time for inference (not general hosting)?

[dholbrook86]

The issue is that the models all need to be loaded into GPU vRAM. It would take a while to spin that back up and load into RAM between each iteration. So it would stand to reason that if you're using that GPU's VRAM you've got to pay for it.

[dholbrook86]

I've been trying to use RunPod.io for some time now and the UI keeps freezing on me. I can't get through 20 iterations without the UI becoming unresponsive. Any ideas what's going on? Anyone else having this issue?

[kylehamilton]

Go to the bottom of the page, click View API, and it'll work.

[mikejnpc]

ya they freaking sux balls... cant pull win imagines, cant run gui on ubuntu or really do anything... their model is what a dog does when you take him for a walk... and all they do is act stupid like they dont know. ... how about getting a decent platform you non chat answering prcks with all the money you ha ve taken from your so called customers... lol .. i just came from there so rant over .. :D

[mikejnpc]

i was just using their webadmin panel

…

On Wed, Oct 19, 2022 at 2:51 PM jus3q ***@***.***> wrote: how do you connect to the Runpod instance? Juypter? — Reply to this email directly, view it on GitHub <#2348 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AJLDXR3I7KR7JROUTDOMMFLWEBUP5ANCNFSM6AAAAAARC5YYN4> . You are receiving this because you commented.Message ID: <AUTOMATIC1111/stable-diffusion-webui/repo-discussions/2348/comments/3919368 @github.com>

[mikejnpc]

ya where? .. they are not that cheap now ?

…

On Wed, Oct 19, 2022 at 7:57 PM John ***@***.***> wrote: 100$ for a RTX 3060 12GB new and probably 50-60 used. — Reply to this email directly, view it on GitHub <#2348 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AJLDXRYBO6KZWJKMVNOERPLWECYLFANCNFSM6AAAAAARC5YYN4> . You are receiving this because you commented.Message ID: <AUTOMATIC1111/stable-diffusion-webui/repo-discussions/2348/comments/3920603 @github.com>

[mikejnpc]

every time i try something.. it has issues. its the the disconnections that annoy me and the non sftp..

[nolanfoster123]

I'd suggest you try to ACE Cloud GPU, they have new GPU machines (NVIDIA A2, A30, and A100). Most of all they have a team of professionals would assist you at every step.

[cmp-nct]

I'm quite sure they assist you given their rates are 500-10000/month.
I'm also quite sure OP did not look for a solution that extremely expensive, he could just buy the best desktop GPU himself in that case.

[revolvedai]

I've been using RunDiffusion.com - they set up all the servers and software, have great support and regularly update with new features. The UI and filebrowser is pretty good too. It's a bit more expensive than RunPod, but you save money on the time scratching your head trying to figure out errors.

[mr-grixa]

They only give 5 GB of storage. It's worse than Google collab

[dholbrook86]

Collab uses your Google drive storage. Hardly comparable. And Colab has tons of issues. You have to manage it yourself. Etc. plenty of reasons for a managed service. Keep using collab if you want. Some of us don’t mind a few bucks for convenience.

[larinius]

With my experience of using Runpod - when it works, it is quote fast - 15 sec per 1024 image using SDXL Lightning model, but it hangs a lot! And they charge money for all useless time when notebook restarting (1 restart ~ up to 10 minutes), Also they charge money for afk time, but not so much.

The 100Gb storage is not enough, you must constantly check free space (7gb per checkpoint, up to 1.5gb per Lora), and you will run out of space super fast if try to download and experiment with models.

If compare to awful Google Collab, Runpod is butter, but the price is too high, if use daily for significant time, it will be cheaper to buy real GPU. With cheapest A4000 GPU, I paid $5 USD per day (~6 hours or usage).