Allocate device memory to container

Allocate a certain size of device memory by specifying resources such as nvidia.com/gpumem. Optional, each unit of nvidia.com/gpumem equals 1 MiB.

resources:
  limits:
    nvidia.com/gpu: 1 # requesting 1 GPU
    nvidia.com/gpumem: 3000 # Each GPU contains 3000 MiB device memory

warning

Set nvidia.com/gpumem as a plain integer (a count of MiB), not a Kubernetes quantity. A suffixed value such as 16Gi is parsed as its byte count (about 17 billion), which overflows HAMi's 32-bit memory field and is silently truncated, often to 0. A zero memory request passes the scheduler's memory check unconditionally, so the pod can be placed on a GPU that is already full and then fails with out-of-memory errors at runtime. Use the integer form shown above (for example 3000).

Allocate a percentage of device memory by specifying resource nvidia.com/gpumem-percentage. Optional, each unit of nvidia.com/gpumem-percentage equals 1% of device memory.

resources:
  limits:
    nvidia.com/gpu: 1 # requesting 1 GPU
    nvidia.com/gpumem-percentage: 50 # Each GPU contains 50% device memory

note

nvidia.com/gpumem and nvidia.com/gpumem-percentage cannot be assigned together