finally got it to run on Windows

[leomaxwell973]

I finally got it to run on Windows :D

thanks to the models posted here by this user
#77 (comment)

Requirements adjusted:
easydict==1.10
tokenizers==0.12.1
numpy>=1.19.2
ftfy==6.1.1
transformers==4.18.0
imageio==2.15.0
fairscale==0.4.6
ipdb
open-clip-torch==2.0.2
xformers==0.0.16
chardet==5.1.0
torchdiffeq==0.2.3
opencv-python>=4.4.0.46
opencv-python-headless==4.7.0.68
torchsde==0.2.6
simplejson==3.18.4
scikit-learn
scikit-image
rotary-embedding-torch==0.2.1
pynvml==11.5.0
pytorch-lightning==1.4.2
torchmetrics==0.6.0
gradio==3.39.0
imageio-ffmpeg
piq
torch==1.13.1+cu117
torchvision==0.14.1+cu117
torchaudio
flash-attn==0.2.1

Windows Triton install (to stop log overflow and compatibility... and optimization? questionable, but log looks clean with it at the very least, sadly, my windows NCCL doesn't work without a windows NCCL torch, can't find the "prototype" mentioned on the torch website :( )
https://github.com/PrashantSaikia/Triton-for-Windows

Running on Python 3.10 with miniconda3 environment

switch all NCCL backends on .yamls (the entrance.py s not the training ones, name is listed at top of training yaml for whichever you're using). Switch to GLOO

and then various tinker's here and there, and knowing what can't be touched, I.E. on DreamMachine it seems it must be left on chunk size 1, and you must name the placeholder mid-save string for examples. Too many little fixes to name at this point and each yaml/VidGen is different so just follow the log and use best guesses for the misc. bugs.

but yep. Can finally confirm works on windows :P