jacobtomlinson

GPU Python for the Real World: Practical GPU-Accelerated Python with RAPIDS

EuroSciPy 2025 | Share Link

Agenda

• RAPIDS/CUDA Python Overview (40 mins)
  • Explore some RAPIDS notebooks from the RAPIDS Community Tutorial
    • cuDF (Pandas/Dataframes)
    • cuML (scikit-learn/ML)
  • and/or explore some CUDA Python notebooks from the Accelerated Computing Hub Tutorial
• Numba (CUDA development in Python)

jacobtomlinson

GPU Utilization CLI demo

A fun little CLI tool to gather and disiplay GPU utilization for nodes in a cluster.

Inspired by this LinkedIn post.

Usage

jacobtomlinson

jacobtomlinson

channels:
  - rapidsai
  - conda-forge
  - nvidia
dependencies:
  - cudf=24.10
  - python=3.12
  - cuda-version>=12.0,<=12.5
  - s3fs
  - dask

jacobtomlinson

awk 'NR==FNR {map[$1]=$2; next} {print ($1 in map ? map[$1] : $1)}' <(alias | sed -E "s/^([^=]*)='?([^ ]*).*/\1 \2/") <(echo "SHELL HISTORY UNWRAPPED" `date +%Y` && history | gawk '{gsub(/^\s*[0-9]+\*?(\s*[0-9/T:]+)\s+/, "", $0); print $0}' | gawk '{gsub(/ \| /, "\n", $0); print $0}' | gawk ' { i=2; while ($1 ~ /^[A-Z0-9_]+=/) { $1=$i; i++ }; print $1 }') | sort | uniq -c | sort -n | tail -n 10

jacobtomlinson

cudf.pandas Coiled Demo

The RAPIDS cudf.pandas accelerator allows you to leverage the power of NVIDIA GPU acceleration in your pandas workflows.

Scripts that use pandas can be run via the cudf.pandas module to accelerate your code with zero-code change.

python my_code.py  # Uses the CPU
python -m cudf.pandas my_code.py  # Same pandas code uses the GPU

jacobtomlinson

### Apple Specific ###

# ignore OS X hidden meta files
.DS_Store
.AppleDouble
.LSOverride

# Icon must end with two \r
Icon

jacobtomlinson

import contextlib
import codecs
import subprocess
import pandas as pd

# Load list of global nameservers and country code information
print("Loading data sources...")
nameservers = pd.read_csv("https://public-dns.info/nameservers.csv")
countries = pd.read_csv("https://github.com/lukes/ISO-3166-Countries-with-Regional-Codes/raw/refs/heads/master/all/all.csv")

jacobtomlinson

When using LocalCUDACluster on a single node it is possible to scale your work out on a SLURM based HPC with a few small tweaks.

First install the Dask Runners package. (Note: this is a prototype and will be merged into dask-jobqueue in the future)

pip install git+https://github.com/jacobtomlinson/dask-hpc-runner.git

Then replace LocalCUDACluster with the SLURMRunner class.

jacobtomlinson