---

name: setup-and-backends description: "Helps install, repair, and verify the CUDA-backed MASt3R-SLAM environment, third-party editable packages, checkpoints, and backend smoke checks." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

setup-and-backends

Use this sub-skill when the task is about installing MASt3R-SLAM, fixing the CUDA build/toolchain, verifying the editable third-party packages, or staging checkpoints before any real SLAM run.

Triggers

• "install MASt3R-SLAM"
• "torch import fails"
• "CUDA extension build failed"
• "mast3r_slam_backends missing"
• "nvcc not found"
• "where do the checkpoints go"
• "RealSense/OpenGL import or backend issues"

What this route covers

• Python 3.11 environment creation and the verified CUDA 12.4 install shape.
• Submodule preparation for Eigen and pyimgui.
• Editable installs for thirdparty/mast3r, thirdparty/in3d, and the root package.
• CUDA extension build prerequisites (nvcc, cuda_runtime.h, ninja).
• Checkpoint presence, naming, and download planning.
• Import checks for mast3r_slam, mast3r_slam_backends, mast3r, in3d, lietorch, pyrealsense2, and evo.

What this route excludes

• One-off runtime commands for a video, RGB folder, or dataset sequence. Route those to run-slam.
• Dataset download/evaluation loops and evo_ape metrics. Route those to evaluation.
• Upstream MASt3R or Dust3R training/demos that do not belong to MASt3R-SLAM.

First reads

• references/installation.md
• references/backends-and-assets.md
• references/troubleshooting.md
• scripts/check_install.py
• scripts/checkpoint_manifest.py

Practical workflow

1. Confirm the environment uses Python 3.11 and a CUDA-capable torch build.
2. Run the install sequence in references/installation.md or adapt it to the current machine manager.
3. Run scripts/check_install.py --check-cuda.
4. If checkpoints are missing, use scripts/checkpoint_manifest.py to print the filenames and download commands.
5. If imports fail, use references/troubleshooting.md to map the symptom to the right repair step before touching runtime workflows.

Acceptance bar

Do not hand off this sub-skill as ready until all of these are true:

• python -m pip check passes in the target environment.
• mast3r_slam_backends imports.
• torch.cuda.is_available() is true on the intended GPU host.
• The root package imports from outside the checkout.
• The checkpoint manifest is understood and any missing assets are explicitly noted rather than silently skipped.

When the environment is already good, keep the answer short: point to the checkpoint manifest or the backend diagnostic script instead of repeating the whole install recipe.