Tetris with GPUs
Kueue log
Private repositories. Opening them needs Fleet GitHub access.
fleet-ai/kueue-ray-adapter
The RayJob adapter
fleet-ai/kueue
Our Kueue changes
RayJob type
Native RayJob · one Workload
Sharded RayJob · one per GPU workgroup
Job runtime at full size · quanta
1
2
3
5
10
20
40
Job priority
q0 · top rung
q1
q2 · bottom rung
GPU workgroups
1
2
GPU workgroups
A
B
+ B
×
Scaling
Elastic
Fixed, as a gang
Topology
Unconstrained
Same leaf switch, preferred
Same leaf switch, required
Target GPUs
4
8
16
24
32
GPUs per pod
1
2
4
8
Pods per worker
1
2
4
Scaling
Elastic
Fixed, as a gang
Topology
Unconstrained
Same leaf switch, preferred
Same leaf switch, required
Target GPUs
4
8
16
24
32
GPUs per pod
1
2
4
8
Pods per worker
1
2
4
Workgroup priority
q0 · top rung
q1
q2 · bottom rung
Workgroup priority
q0 · top rung
q1
q2 · bottom rung
Shape
hover for details
Quota · 64 GPUs
Submit job
Submit random job
Self-play
Restart
End job
Cancel job
Fail job
