Deployment discrimination against poor people
The difficulty of deploying a stable diffusion model.
After exploring model deployment options for my stick man in Stable diffusion for waitbutwhy, I encountered a significant problem: all available solutions are either expensive or overly complex.
Let me show how the world is so unfair to poor people to deploy a model.
In short, complex solutions require you to set up everything yourself, including downloading models, storing customized datasets, and potentially writing a new model class, before deploying it to the cloud. Expensive solutions, on the other hand, have everything already set up on their services.
Initially, I excluded Hugging Face due to its high cost and Bento + EC2 due to its complexity. Therefore, I narrowed down my options to Modal and Sagemaker. I generated a technical plan for each tool to compare and decide which one to use for the project.
At first, I was happy and confident to use Sagemaker, but then something went wrong.
BOOM! I didn’t realize that the “2 months free” offer only allows CPU usage, which means I cannot use a GPU. Unfortunately, the existing model in Sagemaker requires a GPU to train.
Today, I became so frustrated that I accidentally took the bus in the opposite direction and decided to attend a tea party instead of working. However, I’m still determined to deploy my model and get it working. So, I came up with the idea of finding a sugar daddy. Who would be the most likely daddy? Easy.
Tim Urban.
I haven’t thought through the strategy to persuade the prospective sugar daddy yet, but I’ll update you in the future. If you have any ideas, please let me know. Let’s strategize together.