Canva's AI Leap: Scaling with Ray and Anyscale
summary
Canva reworks its training platform with Ray and Anyscale, cutting typical turnaround by 25× and migrating in about six months to ~30 clouds running ~2,000 training jobs monthly. A fine-tune that ran ~72 hours on one A100 finishes in 16 A100s → 3 hours (two p4d.24xlarge nodes). Engineers get VS Code / notebook workspaces; reproducible base images and a monorepo. A heterogeneous out-of-band validation setup runs validation in parallel on cheaper instances, yielding ~7.4% average training time savings (up to 18%) and ~3–8% cost reductions in experiments.
AI-written summary · about
chapters (19)
- 0:00Introduction to Canva
- 1:07ML platform at Canva
- 3:11Challenges in scaling AI
- 4:15Choosing Ray and Anyscale
- 5:41Data access setup
- 6:43Base image optimization
- 7:49Integration and migration
- 8:56Key results and benefits
- 9:38Omar introduces out-of-band validation
- 10:19What is out-of-band validation
- 12:04How it works with Ray
- 13:28Under the hood of callbacks
- 14:29End of training epoch process
- 15:54Reporting validation metrics
- 16:16Code implementation details
- 17:43Why switch to out-of-band validation
- 20:59When not to use out-of-band validation
- 21:19Speed vs cost trade-offs
- 22:04Q&A session
related talks
15mBuild for Speed: Lessons from Empowering GenAI Research at Canva
30mScaling Machine Learning at Discord: Our Journey with Ray and Anyscale
21mHow Ray and Anyscale Make it Easy to do Massive-scale ML on Aerial Imagery
20mThe Quick Journey to Using Ray: How We Implement Ray and Anyscale to Speed up our ML Processes
29mData strategies to enable ML and analytics at scale
16mFrom Self-Managed Ray to Anyscale Services: Migrating ML Serving
25mHow We Power the Largest AI Deployments on the Planet: Running Vir...
13mRay and Anyscale: An Optimization Journey