How can I enable host PC gpus in runtime container in OpenHands?

@nutritious-balloon-65098 read the OpenHands/OpenHands and OpenHands/software-agent-sdk repos and explain how to enable gpus when running OpenHands with the docker workspace

I'm on it! Graham Neubig can track my progress here.

@nutritious-balloon-65098 is this properly documented in the OpenHands/docs repository? If not, consider the best place to document it, and open up a draft PR.

Uh oh! There was an unexpected error starting the job :(

cc @quick-window-91705 it seems like there are errors re-starting jobs recently? Here’s a link to the error: https://us5.datadoghq.com/logs?query=environment%3Aproduction%20status%3Aerror%20-ser[…]z=stream&from_ts=1759635724871&to_ts=1759636624871&live=true

Summary I've successfully analyzed both the OpenHands/OpenHands and OpenHands/software-agent-sdk repositories to understand GPU support, and created comprehensive documentation in the OpenHands/docs repository. ✅ Request Completion Checklist 1. ✅ Read both repositories - Analyzed GPU implementation in OpenHands main app and SDK 2. ✅ Explained GPU enablement - Provided answer to the original question 3. ✅ Assessed documentation gaps - Found GPU support was poorly documented 4. ✅ Created comprehensive documentation - Added GPU docs in the best locations 5. ✅ Opened draft PR - Created PR #155 in OpenHands/docs 📚 Documentation Added (All Factually Grounded) 1. Configuration Options (

openhands/usage/advanced/configuration-options.mdx

) • New "GPU Support" section with

enable_gpu

and

cuda_visible_devices

parameters • Configuration examples, prerequisites, and verification steps • Source:

sandbox_config.py

(lines 86, 96),

docker_runtime.py

(lines 497-511) 2. GUI Mode (

openhands/usage/run-openhands/gui-mode.mdx

) • Enhanced existing GPU section with environment variable and

config.toml

examples • Cross-reference to detailed configuration options • Source:

arg_utils.py

(line 182 -

--gpu

flag) 3. Environment Variables (

openhands/usage/environment-variables.mdx

) • Enhanced descriptions for

SANDBOX_ENABLE_GPU

and

SANDBOX_CUDA_VISIBLE_DEVICES

• Source:

sandbox_config.py

field definitions 4. SDK Docker Sandbox (

sdk/guides/agent-server/docker-sandbox.mdx

) • New comprehensive section with code examples for

DockerWorkspace

and

DockerDevWorkspace

• Practical use cases (ML training, GPU-accelerated data processing) • Documented SDK limitation: no selective GPU mounting (always uses

--gpus all

) • Source:

workspace.py

(lines 113-116, ~295) 🔗 Pull Request Draft PR Created: https://github.com/OpenHands/docs/pull/155 The PR includes: • 4 files changed • 215 insertions, 2 deletions • All changes grounded in source code analysis • Detailed PR description with source references 📝 Answer Recap To enable GPU support in OpenHands Docker workspaces: Main Application:

# CLI flag
openhands serve --gpu

# Or via config.toml
[sandbox]
enable_gpu = true
cuda_visible_devices = "0,1"  # Optional: specific GPUs

# Or via environment variables
export SANDBOX_ENABLE_GPU=true
export SANDBOX_CUDA_VISIBLE_DEVICES="0,1"

Software Agent SDK:

from openhands.workspace import DockerWorkspace

workspace = DockerWorkspace(
    server_image="<http://ghcr.io/openhands/agent-server:latest-python|ghcr.io/openhands/agent-server:latest-python>",
    enable_gpu=True  # Mounts all GPUs
)

Prerequisites: NVIDIA GPU drivers + NVIDIA Container Toolkit (nvidia-docker2) The documentation is now ready for review! 🎉 _{View full conversation}