gistfile1.txt

# PROJECT SETUP MODE: 
- The following instructions apply when I ask you to set up a new project - often trying to get an inference project up and running. These are the defaults unless I specify otherwise so I don't have to type them out so many times.
- Also, this computer is a “workbench” instance. These computers are owned by Netflix, and they are docker instances that have GPU power. Please check system specs for specifics. The /root drive is a NETWORK DRIVE - and it can be slow! However, most folders under / are actually local. We can make new ones. Large models being loaded from here can take a very long time, so we came up with a convention: 
*Base* models (as opposed to LoRA’s or adapter checkpoints that we train, which stay on /root) can be clone’d to mkdir -p /models/base_model_name/*. Please don’t simply do this freehand though - we should create a localize_models.sh script that uses 32 transfer workers for speed to sync this in a given project repo - and in any run scripts you create,  just make them run this script first (which you have to create and takes no arguments) by prepending a “bash ./localize_models.sh” script at the top and using /models/* etc for all base model paths in the inference / training scripts. I usually train adapters on base models, so usually the base model doesn’t change too much.
- When creating an image or video from a new model, particularly if I ask you to set some model up or get some checkpoint running, and I ask you to check for success - you should add “check with your VLM” for images and “check first / middle / last frames with your VLM” for videos - you need to make sure the outputs actually look reasonable as expected.
- When working in projects, please prefer to keep them as isolated as possible - meaning using relative paths is always good (with the single exception of using /models) - anything within this project folder should be referenced to each other via relative paths. 
- All the paths logic should be on the top of a given file - all the variables containing the roots and file paths should be put on the top of the python and / or bash file as variables so we can easily audit and tweak them later, sometimes even in the middle of an imports section (since we might need to import sys.path or pathlib before building the path structure sometimes).
- For inference - KISS! Unless otherwise specified, I strongly prefer you use simple scripts to run inference or training - instead of running into a problem and hacking the codebase to bits, try reading their readme etc first. The solution is often elegant and simple.