Pytorch and Huggingface Experiments

I'm running these on WSL2 on Windows 11 with a NVidia GEForce card and drivers.

WSL2 Ubuntu has a special CUDA Toolkit install available here. This installer avoids overwriting the /usr/lib/wsl/lib/libcuda.so file setup by the NVidia GeForce driver installer.

Scripts

• train.py and inference.py

  • straight from the PyTorch getting started. Shows how to save and load a trained model.

• transformerpipeline.py

  • demos the pipeline "helper" swiss army knife.

• tensors.py

  • basic tensor operations and cuda

• Microsoft phi 1.5 SML (Small Language Model) capable of non-trivial code/language one-shot results that can be run on a laptop/phone. Example run on RTX3050Ti Laptop GPU takes 8 seconds.

• Microsoft phi 2 SML (Small Language Model) capable of non-trivial code/language one-shot results that can be run on a laptop/phone. Example run on RTX3050Ti Laptop GPU takes 88 seconds.

• Langgraph

  • graphchat.py basic chat example
  • graphconditional.py conditional tool execution
  • visgraph.py visualization of graph functions