GPU Infrastructure for Open Source
CI at 7,000 runners · Dev servers from T4 to B300 · All on Kubernetes
Wouter Devriendt · Jean Schmidt
⚡
CI Infrastructure
Testing every PR across 6 GPU vendors.
7,367 concurrent runners at peak.
→ Jean Schmidt
🖥️
Dev Servers
SSH into GPUs in 15 seconds.
T4 to B300, multi-node ready.
→ Wouter Devriendt
⚡
CI Infrastructure
Testing every PR across 6 GPU vendors.
7,367 concurrent runners at peak.
→ Jean Schmidt
🖥️
Dev Servers
SSH into GPUs in 15 seconds.
T4 to B300, multi-node ready.
→ Wouter Devriendt
CI: Testing PyTorch
⚡
The Challenge
- ~100 PRs/day — every one needs GPU test suites
- 400K test cases per PR across CUDA, ROCm, XPU
- Peak: 7,367 concurrent runners on 2,100 nodes
- Was EC2-locked — couldn't accept donated GPU capacity
8,500
jobs / hour
1.5M
CI min / day
6
GPU vendors
⚡
CI Infrastructure
Testing every PR across 6 GPU vendors.
7,367 concurrent runners at peak.
→ Jean Schmidt
🖥️
Dev Servers
SSH into GPUs in 15 seconds.
T4 to B300, multi-node ready.
→ Wouter Devriendt
Dev: Building PyTorch
🖥️
The Challenge
- Contributors need SSH access to real GPUs
- Debug CUDA kernels, profile with ncu / nsys
- Multi-node distributed training — up to 64 GPUs
- No cloud account needed — just a GitHub username
~15s
cold start
T4→B300
GPU range
87M
pip installs/mo