Implementation:Kubeflow Pipelines GPU Resource Request Sample

Knowledge Sources	Kubeflow_Pipelines
Domains	Pipeline_Sample, Resource_Management, GPU
Last Updated	2026-02-13 14:00 GMT

Overview

Sample pipeline demonstrating GPU resource requests and accelerator type constraints for pipeline tasks using the KFP SDK.

Description

This sample (53 lines) shows how to request GPU resources. The `training_job` component uses a PyTorch CUDA base image and verifies GPU availability. The `generate_resource_constraints_request` component returns GPU count and accelerator type. GPU count can be set dynamically via `set_accelerator_limit()`, but accelerator type must be hardcoded due to SDK limitations.

Usage

Reference this sample when building GPU-enabled pipelines targeting specific hardware like NVIDIA Tesla K80.

Code Reference

Source Location

• Repository: Kubeflow_Pipelines
• File: samples/core/resource_spec/runtime_resource_request_gpu.py
• Lines: 1-53

Signature

@dsl.component(base_image='pytorch/pytorch:1.7.1-cuda11.0-cudnn8-runtime')
def training_job():
    """Checks torch.cuda.is_available(), raises if no GPU."""

@dsl.component
def generate_resource_constraints_request() -> NamedTuple('Outputs', [('nbr_gpus', str), ('accelerator', str)]):
    """Returns nbr_gpus='1', accelerator='NVIDIA_TESLA_K80'."""

@dsl.pipeline
def resource_constraint_request_pipeline():
    """Pipeline: generate GPU constraints -> training job with GPU."""

Import

from kfp import dsl
from typing import NamedTuple

I/O Contract

Inputs

Name	Type	Required	Description
(none)	—	—	Pipeline takes no inputs

Outputs

Name	Type	Description
Compiled YAML	file	Pipeline IR YAML for submission to KFP

Usage Examples

GPU Pipeline Pattern

from kfp import dsl

@dsl.pipeline
def gpu_pipeline():
    constraints = generate_resource_constraints_request()
    task = training_job()
    task.set_accelerator_limit(constraints.outputs['nbr_gpus'])
    # Accelerator type must be hardcoded (SDK limitation)
    task.set_accelerator_type('NVIDIA_TESLA_K80')

Related Pages

• Environment:Kubeflow_Pipelines_Python_SDK