Skip to content

ECS  >  Structures  >  NvidiaMpsAllocation

NvidiaMpsAllocation

Structure Class

NvidiaMpsAllocation dataclass

Bases: _SerdeShape

The GPU memory and compute limits for a container that shares a GPU by using NVIDIA Multi-Process Service (MPS). Amazon ECS assigns the container a single GPU that has enough memory available. Other containers that use MPS can share the same GPU.

Attributes

max_compute_percent class-attribute instance-attribute
max_compute_percent: int | None = ...

The maximum percentage of the GPU's compute capacity that the container can use. Amazon ECS sets the CUDA_MPS_ACTIVE_THREAD_PERCENTAGE environment variable in the container to this value. This value is a limit, not a reservation.

If you don't specify this parameter, Amazon ECS doesn't set CUDA_MPS_ACTIVE_THREAD_PERCENTAGE, and the active thread percentage of the MPS server on the container instance applies. By default, this is 100 percent.

memory instance-attribute
memory: int

The amount of GPU memory, in MiB, to allocate to the container. Amazon ECS assigns the container a GPU that has at least the specified amount of memory available, and sets the CUDA_MPS_PINNED_DEVICE_MEM_LIMIT environment variable in the container to limit its memory to this amount. The memory must be available on a single GPU. If more than one container in a task uses nvidiaMps, Amazon ECS assigns them the same GPU, which must have enough memory available for all of them.