Task accelerators in a HealthOmics workflow definition

In the workflow definition, you can optionally specify the GPU accelerator-spec for a task. HealthOmics supports the following accelerator-spec values, along with the supported instance types:

Accelerator spec	Healthomics instance types
nvidia-tesla-t4	G4
nvidia-tesla-t4-a10g	G4 and G5
nvidia-tesla-a10g	G5
nvidia-t4-a10g-l4	G4, G5, and G6
nvidia-l4-a10g	G5 and G6
nvidia-l4	G6
nvidia-l40s	G6e

If you specify an accelerator type that supports multiple instance types, HealthOmics selects the instance type based on available capacity. If both instance types are available, HealthOmics gives preference to the lower cost instance. The exception is for the nvidia-t4-a10g-l4 task accelerator which gives preference to the latest generation instance available.

For details about the instance types, see Accelerated-computing instances.

In the following example, the workflow definition specifies nvidia-l4 as the accelerator:

WDL

task my_task {
 runtime {
    ...
    acceleratorCount: 1
    acceleratorType: "nvidia-l4"
 }
 ...
}          

NextFlow

process my_task {
 ...
 accelerator 1, type: "nvidia-l4"
 ...
}          

CWL

cwlVersion: v1.2
class: CommandLineTool
requirements:
  ...
  cwltool:CUDARequirement:
      cudaDeviceCountMin: 1
      cudaComputeCapability: "nvidia-l4"
      cudaVersionMin: "1.0"
          

Advanced resource configuration

You can also configure a task to reserve multiple accelerators in your preferred priority order using the omicsResourceFallbackOrder directive. This directive is currently available only for WDL. HealthOmics tries each accelerator profile in order you define until one succeeds, including an optional final CPU profile.

This is useful in the following scenarios:

• GPU to GPU fallback – List GPU types in priority order (for example, try nvidia-l40s first, then fall back to nvidia-l4).

• GPU to CPU fallback – Add a final CPU-only profile so the task completes even when no GPU capacity is available in the Region.

For full details about omicsResourceFallbackOrder, including per-profile field reference, validation rules, and a complete WDL example, see Advanced resource configuration.