 Josh Siviour
|
98fc01764b
|
feat(ecs): GPU passthrough to task containers (Epic C) (#512)
* feat(ecs): thread resourceRequirements GPU count through conv, ContainerDef, AssignContainer and RunSpec
* feat(ecs): add GPU placement dimension — InstanceRecord capacity, fits/reserve/release, STRINGSET resource reporting
* feat(ecs): agent GPU discovery via nvidia-smi, pinning ledger, device-UUID report-back and DescribeTasks gpuIds
* feat(ecs): inject assigned GPU devices into containerd OCI spec via CDI (oci.WithCDIDevices)
* test(ecs): add GPU task e2e (nvidia-smi container, gpuIds assertion) to gpu e2e suite
* test(gpu): decouple base-AMI requirement from GPU fixture so ECS e2e runs without the ubuntu EC2 image
* test(gpu): pick smallest advertised GPU instance type to fit a single host
* fix(ecs): preserve pinned GPU UUIDs across STOPPED state change so DescribeTasks reports gpuIds for one-shot tasks
|
2026-07-13 10:54:55 +10:00 |
|