# Conda

- Modules come later. Now just load the module and mention what it does.
- Mostly going through the [python-conda](https://scicomp.aalto.fi/triton/apps/python-conda) section
- Should we demonstrate the first time setup for environments? I guess so.
- [first-time-setup](https://scicomp.aalto.fi/triton/apps/python-conda/#first-time-setup)
- Start with installing R and demonstrating the version
- I don't want people to only associate Conda with Python, so start with something else
- Big example: Conda env for LLM
- Note: use pytorch from
- Recommend using an environment file and recreating when changed.

Plan (40 minutes)

- Introduce conda (10min)
- General dependency manager
- Language agnostic
- Lot's of binary packages

- Example: create environment with R (10min)
- Explain what each line does while typing

- Why you should track environments (5min)
- Makes your code reproducible
- Helps debugging
- Helps set-up for other team members
- Allows transitioning to another system

- Example: Python LLM environment (10 min)
- Note that you probably want the Triton LLM module

- Wrap up

### Introduce Conda (10 min)

- General dependency manager
- Language agnostic
- Lot's of binary packages

- First time setup
- Channels

### create environment with R (10min)

- Create environment file, go through line by line

name: tidyverse
channels:
  - conda-forge
dependencies:
  - r-tidyverse

- Show that it works

source activate tidyverse
vim hello_tidy.R
Rscript hello_tidy.R

- hello_tidy.R

library(tidyverse)

if ("dplyr" %in% .packages()) {
  cat("tidyverse (via dplyr) loaded successfully.\n")
  cat("dplyr version:", as.character(packageVersion("dplyr")), "\n")
} else {
  cat("Error: tidyverse (or dplyr) could not be loaded.\n")
}

### Why you should track environments (10min)

- Points not in the docs:
- Makes your code reproducible
- Allows version controlling your environments
- You can have the environment file with the code and the output
- Helps debugging
- Helps set-up for other team members
- Allows transitioning to another system

- Points in the docs: Solves most Python dependency issues

### Python LLM environment (10 min)

- Environment with pytorch on Triton
- First run without, then add `export CONDA_OVERRIDE_CUDA=12.6`. The first time there will be an error.

name: pytorch
channels:
  - nvidia
  - conda-forge
dependencies:
  - python==3.12
  - pytorch-gpu>=2.6,<2.7
  - torchvision
  - torchaudio
  - transformers