feat: add GPU support to metaflow-dev minikube setup

[cnaples79]

Summary

• Add optional GPU support for minikube in metaflow-dev with intelligent auto-detection
• Provide manual control via MINIKUBE_ENABLE_GPU environment variable
• Enable GPU workloads like @resources(gpu=1) in local development environments

Changes Made

• Auto-detection logic: Detects NVIDIA (nvidia-smi) and AMD (rocm-smi) GPUs automatically
• Environment variable control: MINIKUBE_ENABLE_GPU=auto|true|false (default: auto)
• User feedback: Informative messages about GPU detection status during startup
• Help documentation: Updated help text with environment variable usage
• Conditional flag addition: Adds --gpus all to minikube start only when appropriate

Modes of Operation

1. auto (default): Automatically detects GPU availability and enables if found
2. true: Force enables GPU support regardless of detection
3. false: Explicitly disables GPU support

Test Plan

• ✅ Verified Makefile syntax with make help
• ✅ Tested dry-run with make -n setup-minikube (shows no GPU detected message)
• ✅ Tested forced enable with MINIKUBE_ENABLE_GPU=true (correctly adds --gpus all flag)
• ✅ Confirmed help text displays new environment variable documentation

Example Usage

# Auto-detect GPU (default behavior)
make setup-minikube

# Force enable GPU support
MINIKUBE_ENABLE_GPU=true make setup-minikube

# Explicitly disable GPU support  
MINIKUBE_ENABLE_GPU=false make setup-minikube

Fixes #2606

[feltech]

Depending on how Docker is configured, --gpus all might not work, and instead you need to use --devices nvidia.com/gpu=all or similar.

(I had this problem with Docker on NixOS)

[cnaples79]

Thanks for the feedback, I'll update.

[feltech]

Apologies, I may have misled you here, the problem I referred to was using the docker command-line, rather than the minikube command-line. They both accept a --gpus argument, but I think minikube is a bit more clever about it. Indeed, --devices doesn't seem to be valid for minikube:

$ minikube start --devices nvidia.com/gpu=all
Error: unknown flag: --devices

My bad, got my wires totally crossed.

[cnaples79]

Ahhh I see. No worries, I'll update the PR with the correct fix.

[cnaples79]

Thanks for the feedback @feltech! I've updated the implementation to address the Docker compatibility concerns:

Changes Made

🔧 Improved Docker Compatibility:

• Default to --devices nvidia.com/gpu=all for NVIDIA GPUs (more compatible with different Docker configurations)
• Keep --gpus all for AMD/other GPUs
• This addresses the NixOS Docker issue you mentioned

⚙️ Enhanced Control Options:
Added MINIKUBE_GPU_FLAG environment variable for explicit control:

• auto (default): Smart selection based on GPU type
• gpus: Force --gpus all format
• devices: Force --devices nvidia.com/gpu=all format
• Custom value: User-provided flag (e.g., --devices nvidia.com/gpu=2)

Example Usage

# Auto-detect best GPU flag (default)
make setup-minikube

# Force devices format (good for Docker compatibility issues)
MINIKUBE_GPU_FLAG=devices make setup-minikube

# Force legacy gpus format
MINIKUBE_GPU_FLAG=gpus make setup-minikube

# Custom GPU specification
MINIKUBE_GPU_FLAG="--devices nvidia.com/gpu=2" make setup-minikube

This should resolve the Docker configuration compatibility issues while maintaining flexibility for different setups. Let me know if this addresses your concerns!

[cnaples79]

Thanks for the clarification, and you're absolutely right — minikube doesn't support --devices. I've updated the PR to remove the --devices path and always pass --gpus all to minikube start when GPU is detected or forced via MINIKUBE_ENABLE_GPU=true.

Summary of changes:

• Remove MINIKUBE_GPU_FLAG and the --devices nvidia.com/gpu=all path
• Keep simple/valid --gpus all for minikube
• Preserve auto‑detection and MINIKUBE_ENABLE_GPU env var controls

If you want me to also document the separate Docker CLI considerations (for folks not using minikube), I can add a short note in the devtools help.