Bash C C#C++CUDA Dart Go Java JavaScript Julia Kotlin Lua MATLAB PHP Python R Ruby Rust Scala SQL Swift TypeScript

Volume B · The TensorCore.si Library

Bash

The shell that automates everything.

Bash is the command shell on most Linux servers and the default scripting language for automation. Every AI engineer uses it to set up GPU machines, run training jobs, manage models and glue tools together.

Prologue

Why learn Bash?

You cannot avoid the terminal in AI and cloud work. Bash lets you automate repetitive tasks, control remote GPU servers and chain powerful tools together.

Everywhere: every Linux server, Docker container and Mac terminal has a shell.
Automation: turn any sequence of commands into a reusable script.
AI workflows: launch training, download models, check GPUs with nvidia-smi.

Prologue

Set up and run your first program

Bash is preinstalled on Linux. macOS uses zsh by default but includes bash; on Windows use WSL.

terminal

bash --version
chmod +x hello.sh
./hello.sh

hello.sh

#!/usr/bin/env bash
name="Ada"
echo "Hello, AI! Welcome to Bash, $name."

TipThe first line (shebang) tells the system to run the file with bash.

Chapter 1

Variables and quoting

Assign with no spaces around =. Read a variable with $name. Always wrap variables in double quotes to keep spaces intact. $(command) captures a command's output.

variables.sh

#!/usr/bin/env bash
model="llama3.1"
layers=32
today=$(date +%F)
echo "Model: $model, layers: $layers, date: $today"
gb=$(( 8 * 2 ))           # integer arithmetic
echo "FP16 size of 8B model: $gb GB"

Chapter 2

Conditions and loops

if uses [[ ... ]] tests. for loops over lists or files; while loops until a condition fails.

control.sh

#!/usr/bin/env bash
for gpu in H100 H200 B200; do
  echo "Benchmarking $gpu"
done
vram=24
if [[ $vram -ge 80 ]]; then
  echo "data center GPU"
elif [[ $vram -ge 24 ]]; then
  echo "can run a 30B model at 4-bit"
else
  echo "small models only"
fi
for f in *.txt; do [[ -e "$f" ]] && wc -l "$f"; done

Chapter 3

Functions and arguments

Scripts receive arguments as $1, $2 and so on; $# is the count and $@ is all of them. Functions work the same way.

functions.sh

#!/usr/bin/env bash
set -euo pipefail          # stop on errors and unset variables

params() {
  local layers=$1 width=$2
  echo $(( 12 * layers * width * width ))
}
echo "Args: $# ($*)"
echo "Params: $(params 32 4096)"

Chapter 4

Pipes and text processing

The pipe | sends one command's output into another. Combined with grep, sort, awk and jq, you can process logs and JSON quickly.

pipes.sh

#!/usr/bin/env bash
# Top 5 most common words in a file
tr -cs '[:alpha:]' '\n' < notes.txt | tr '[:upper:]' '[:lower:]' | sort | uniq -c | sort -rn | head -5

# GPU memory use as CSV
nvidia-smi --query-gpu=name,memory.used,memory.total --format=csv

# List local Ollama models with jq
curl -s http://localhost:11434/api/tags | jq -r '.models[].name'

Chapter 5

Files, exit codes and background jobs

Every command returns an exit code (0 means success). Redirect output with > and 2>, and run long jobs in the background with & or nohup.

jobs.sh

#!/usr/bin/env bash
mkdir -p logs
if python train.py > logs/train.log 2>&1; then
  echo "training finished"
else
  echo "training failed with code $?" >&2
fi
nohup python serve.py > logs/serve.log 2>&1 &
echo "server started with PID $!"

Part II · AI with Bash

AI with Bash

Bash does not train models, but it runs almost every AI workflow: setting up environments, downloading weights, launching distributed training with torchrun, scheduling jobs on clusters with Slurm and calling model APIs with curl.

The essential libraries

curl + jqCall LLM APIs and parse JSON responses.

nvidia-smiCheck GPU use, memory and temperature.

Ollama CLIPull and run local models from the terminal.

Slurm (sbatch)Submit training jobs to GPU clusters.

A training job script

A typical script that checks the GPU, logs everything and launches multi-GPU training with PyTorch's torchrun.

train.sh

#!/usr/bin/env bash
set -euo pipefail
NGPU=$(nvidia-smi -L | wc -l)
RUN="run-$(date +%Y%m%d-%H%M)"
mkdir -p "runs/$RUN"
echo "Found $NGPU GPU(s); logging to runs/$RUN"
nvidia-smi --query-gpu=name,memory.total --format=csv | tee "runs/$RUN/gpus.csv"
torchrun --nproc_per_node="$NGPU" train.py --epochs 3 --out "runs/$RUN" 2>&1 | tee "runs/$RUN/train.log"

Part III · Build an AI agent

Build a simple AI agent in Bash

With curl and jq you can build a small tool-calling agent. It is a great way to see exactly what goes over the wire between your code and the model.

Your goal→Model plans→Tool call→Your code runs the tool→Result back to model→Final answer

agent.sh

#!/usr/bin/env bash
set -euo pipefail
URL=http://localhost:11434/v1/chat/completions
TOOLS='[{"type":"function","function":{"name":"get_weather","description":"Weather for a city","parameters":{"type":"object","properties":{"city":{"type":"string"}},"required":["city"]}}}]'
MSGS='[{"role":"user","content":"Should I take an umbrella in Lahore today?"}]'

for round in 1 2 3 4 5; do
  BODY=$(jq -n --argjson m "$MSGS" --argjson t "$TOOLS" '{model:"llama3.1",messages:$m,tools:$t}')
  MSG=$(curl -s "$URL" -H 'Content-Type: application/json' -d "$BODY" | jq '.choices[0].message')
  MSGS=$(jq --argjson x "$MSG" '. + [$x]' <<< "$MSGS")
  if [[ $(jq '.tool_calls | length // 0' <<< "$MSG") -eq 0 ]]; then
    jq -r '.content' <<< "$MSG"; break
  fi
  while read -r CALL; do
    ID=$(jq -r '.id' <<< "$CALL")
    CITY=$(jq -r '.function.arguments | fromjson | .city' <<< "$CALL")
    RESULT="It is 24°C and sunny in $CITY."
    MSGS=$(jq --arg id "$ID" --arg r "$RESULT" '. + [{role:"tool",tool_call_id:$id,content:$r}]' <<< "$MSGS")
  done < <(jq -c '.tool_calls[]' <<< "$MSG")
done

For anything bigger, move to Python or TypeScript, but keep Bash for the glue around it.

Run it locallyInstall Ollama, run ollama pull llama3.1, and keep it running. The code talks to its OpenAI-compatible API at http://localhost:11434/v1, so the same program also works with vLLM, llama.cpp server or a cloud API by changing the URL, model name and adding an API key.

Part IV · Your path

Projects to build

Backup scriptCompress a folder and keep the last 7 copies.

GPU monitorLog nvidia-smi output every minute to CSV.

Model managerA menu script to pull, run and remove Ollama models.

Reference

Bash cheat sheet

Task	Syntax
Variable	`name="x"; echo "$name"`
Command output	`now=$(date)`
If	`if [[ -f file ]]; then ... fi`
Loop	`for f in *.log; do ...; done`
Pipe	`cmd1 \| cmd2`
Redirect	`cmd > out.log 2>&1`

Reference

Common mistakes

Spaces around = break assignments: write x=1, not x = 1.
Unquoted variables split on spaces; quote them: "$file".
Scripts keep running after errors unless you use set -e.

Reference

Where to go next

Bash Guide — trusted guide
ShellCheck — find bugs in scripts
The Linux Command Line — free book

On TensorCore.si: Python volume, What is a GPU cluster?.

Keyboard: [ previous language · ] next language · / search the library

Bash

Why learn Bash?

Set up and run your first program

Variables and quoting

Conditions and loops

Functions and arguments

Pipes and text processing

Files, exit codes and background jobs

AI with Bash

The essential libraries

A training job script

Build a simple AI agent in Bash

Bash learning roadmap

Beginner

Intermediate

Advanced

Projects to build

Bash cheat sheet

Common mistakes

Where to go next