Customizing task resources

When defining a task function, you can specify resource requirements for the pod that runs the task. Flyte will take this into account to ensure that the task pod is scheduled to run on a Kubernetes node that meets the specified resource profile.

Resources are specified in the @fl.task decorator. Here is an example:

from flytekit.extras.accelerators import A100

@fl.task(
    requests=Resources(mem="120Gi", cpu="44", ephemeral_storage="100Gi"),
    limits=Resources(mem="200Gi", cpu="100", gpu="12", ephemeral_storage="200Gi"),
    accelerator=GPUAccelerator("nvidia-tesla-a100")
)
def my_task()
    ...

There are three separate resource-related settings:

  • requests
  • limits
  • accelerator

The requests and limits settings

The requests and limits settings each takes a Resource object, which itself has five possible attributes:

  • cpu: Number of CPU cores (in whole numbers or millicores (m)).
  • gpu: Number of GPU cores (in whole numbers or millicores (m)).
  • mem: Main memory (in Mi, Gi, etc.).
  • ephemeral_storage: Ephemeral storage (in Mi, Gi etc.).

Note that CPU and GPU allocations can be specified either as whole numbers or in millicores (m). For example, cpu="2500m" means two and a half CPU cores and gpu="3000m", meaning three GPU cores.

The requests setting tells the system that the task requires at least the resources specified and therefore the pod running this task should be scheduled only on a node that meets or exceeds the resource profile specified.

The limits setting serves as a hard upper bound on the resource profile of nodes to be scheduled to run the task. The task will not be scheduled on a node that exceeds the resource profile specified (in any of the specified attributes).

GPUs take only limits

GPUs should only be specified in the limits section of the task decorator:

  • You should specify GPU requirements only in limits, not in requests, because Kubernetes will use the limits value as the requests value anyway.
  • You can specify GPU in both limits and requests but the two values must be equal.
  • You cannot specify GPU requests without specifying limits.

The accelerator setting

The accelerator setting further specifies the type of specialized hardware required for the task. This can be a GPU, a specific variation of a GPU, a fractional GPU, or a different hardware device, such as a TPU.

See Accelerators for more information.

The with_overrides method

When requests, limits, or accelerator are specified in the @fl.task decorator, they apply every time that a task is invoked from a workflow. In some cases, you may wish to change the resources specified from one invocation to another. To do that, use the with_overrides method of the task function.

For example:

@fl.task
def my_task(ff: FlyteFile):
    ...

@fl.workflow
def my_workflow():
    my_task(ff=smallFile)
    my_task(ff=bigFile).with_overrides(requests=Resources(mem="120Gi", cpu="10"))