#7177 [V2] Add GPU support for sandbox via FLYTE_SANDBOX_GPU env var Pull request opened by pingsutw Summary • Add opt-in NVIDIA GPU support for the Flyte sandbox, targeting machines like DGX Spark • When

FLYTE_SANDBOX_GPU=true

is set, the sandbox container gets

--gpus all

and auto-configures K3s with the NVIDIA container runtime and device plugin • No impact on non-GPU machines — GPU support is disabled by default Changes •

docker/sandbox-bundled/Makefile

: Add

FLYTE_SANDBOX_GPU

env var that conditionally passes

--gpus all

to Docker •

docker/sandbox-bundled/bin/k3d-entrypoint-gpu.sh

(new): Entrypoint script that, when GPU is enabled: • Configures K3s containerd to use

nvidia-container-runtime

as default runtime • Deploys NVIDIA k8s-device-plugin DaemonSet for

<http://nvidia.com/gpu|nvidia.com/gpu>

resource advertisement Usage # On GPU machine (e.g., DGX Spark) FLYTE_SANDBOX_GPU=true make start # On non-GPU machine (default) make start Test plan • Verify

make start

still works without

FLYTE_SANDBOX_GPU

(no regression on non-GPU machines) • Verify

FLYTE_SANDBOX_GPU=true make start

on a machine with NVIDIA GPU + Container Toolkit • Verify

<http://nvidia.com/gpu|nvidia.com/gpu>

resources are advertised:

kubectl describe node | grep nvidia

• Verify a pod requesting GPU can be scheduled:

kubectl run gpu-test --image=nvidia/cuda:12.0-base --restart=Never -- nvidia-smi

•

main

• #6583 • [V2] Add GPU support for sandbox via FLYTE_SANDBOX_GPU env var 👈 flyteorg/flyte