The published 9.1.sh URL is the PVE 9.x / Debian Trixie compatibility lane. It has been observed selecting PVE 9.2.10 correctly. Run it from a privileged shell on the Proxmox host; no Git checkout or directory under /root is needed.
set -o pipefail
wget -qO- https://devs-guide.github.io/proxmox/9.1.sh | bashThe runner performs repository setup, safe upgrades, managed Python/Ansible bootstrap, packages, accounts, SSH, security services, sysctl, power, logging, and certificates. metal.sh is not the PVE entrypoint.
The LAN-oriented credential policy intentionally defaults app, agent, proxmox, and root to username:username and reapplies it on reruns. Change the static bootstrap_default_password_mode setting, or use PROXMOX_BOOTSTRAP_PASSWORD_MODE=preserve|locked for a streamed run, to select another policy. UFW requires an explicit management subnet; RDP defaults closed.
set -o pipefail
wget -qO- https://devs-guide.github.io/proxmox/9.1.sh |
PROXMOX_MANAGEMENT_SUBNETS=10.0.0.0/24 bashSee the ordered bootstrap and entrypoint guide.
Stream the PVE 9 AMD runner directly from GitHub Pages. It repairs the stock PVE kernel module and firmware, then reports native temperatures:
set -o pipefail
wget -qO- https://devs-guide.github.io/proxmox/setup.gpu.amd.sh | sudo bash
set -o pipefail
wget -qO- https://devs-guide.github.io/proxmox/setup.gpu.amd.sh |
sudo bash -s -- --dry-run
set -o pipefail
wget -qO- https://devs-guide.github.io/proxmox/setup.gpu.amd.sh |
sudo bash -s -- --checkIf an AMD GPU is currently owned by vfio-pci, the default action accepts
only intact schema-v4 host state created by this repository. It restores the
exact recorded host-file subset, archives that state, and requires an
operator-controlled reboot before rerunning the default setup action. It does
not rebind the live GPU, reboot the host, or change VM configuration; existing
VM passthrough may stop working. Use --dry-run to validate the rollback
evidence without changes.
--check remains strictly read-only; the default action may temporarily wake
a runtime-suspended AMDGPU solely to sample temperatures, then restores it.
After reviewing temperatures, globally blacklist AMD host drivers when every AMD GPU is intended for passthrough and out-of-band access is available:
set -o pipefail
wget -qO- https://devs-guide.github.io/proxmox/setup.gpu.amd.sh |
sudo bash -s -- --blacklist --yesThe runner never reboots automatically. Remove only its managed blacklist with
wget -qO- https://devs-guide.github.io/proxmox/setup.gpu.amd.sh | sudo bash -s -- --unblacklist --yes.
The AMDGPU runner supports streamed mutations; VFIO configuration remains in
setup/vm/gpu.sh.
Keep amdgpu loaded until every target reports readable temperatures and is confirmed not to be the host boot display:
set -o pipefail
wget -qO- https://devs-guide.github.io/proxmox/setup/temperature.sh |
sudo bash -s -- --component gpu --vendor amd --require-driver amdgpu \
--require-readable --reject-boot-vga --output json
set -o pipefail
wget -qO- https://devs-guide.github.io/proxmox/setup/vm/gpu.sh |
bash -s -- --action inventory --release 9.1 --output jsonRequired AMD temperature sampling may temporarily resume a runtime-suspended GPU and restore its prior power policy. GPU inventory remains read-only. Download and inspect the passthrough runner before prepare, attach, detach, or unprepare. AMD early preparation and attachment require reviewed temperature evidence and --confirm-temperature-reviewed.
The current stable release is 0.0.6, which adds reviewed general local ZFS
provisioning to the managed Proxmox baseline. Prerelease 0.0.8-rc.1 publishes
the whole-GPU passthrough candidate together with the PVE 9 AMDGPU restoration
and temperature gate. The published Proxmox VE 6.4/Buster and Proxmox VE
9.x/Trixie lanes remain supported by the candidate implementation.
# Proxmox VE 6.4 / Debian Buster
wget -qO- https://devs-guide.github.io/proxmox/6.4.sh | bash
# Proxmox VE 9.1 / Debian Trixie
wget -qO- https://devs-guide.github.io/proxmox/9.1.sh | bashThe PVE 9.1 compatibility lane uses the Proxmox no-subscription repository. Credential and firewall policy are documented above and in docs/readme.md.
See the 0.0.8-rc.1 release notes for the candidate scope,
safety boundary, and current acceptance evidence. ZFS-backed LXC/Samba delivery
remains planned for 0.0.7. Final 0.0.8 requires the complete PVE 6.4/PVE 9
passthrough and rollback acceptance matrix; Local Model Inventory remains
planned after that release.
Every existing and future GitHub release page follows the release publication policy and its review template. The exact title and body require human approval before publication or editing. Before a new release is published, older release pages must be audited and reconciled with the same structure without changing their tags, source commits, or assets.
The 0.0.6 feature discovers local disks and writes an editable
./zpool.config. It does not create a pool during discovery or configuration.
Its storage-oriented defaults are pool zfspool, dataset zfspool/archive,
and mountpoint /media/zfspool/archive; Samba and LXC remain later consumers
of that dataset rather than part of its name. Interactive disk review is
pretty-printed JSON by default.
mkdir -p ./zfs-setup
cd ./zfs-setup
wget -qO- https://devs-guide.github.io/proxmox/setup/storage/zfs.sh | \
bash -s -- --type sas --size 6TBDownload the runner before generating or applying a destructive plan. Review the design in the ZFS feature plan and follow the human acceptance procedure for review, planning, explicitly authorized creation, and verification.
- Proxmox-native feature runners keep their existing
setup/...layout such assetup/vlan.shand publish aliases such assetup.vlan.sh. - Non-Proxmox CLI application installers for Proxmox hosts should use the
cli.{app_name}naming family.
Current convention:
- source runner:
setup/cli.{app_name}.sh - published runner:
setup.cli.{app_name}.sh - Debian-side playbook when needed:
ansible/debian/cli.{app_name}.yml
Current example:
wget -qO- https://devs-guide.github.io/proxmox/setup.cli.codex.sh | bashEvery published Bash helper uses a .sh path and can bootstrap its shared
library when streamed directly into Bash.
Structured helper output uses --output json and requires jq.
Human operator documentation:
- End-to-end and manual VM restore runbook
- Whole-GPU passthrough test runbook
- PVE 6.4 / Buster GPU passthrough runbook
- PVE 9.1 / Trixie GPU passthrough runbook
- Manual GPU passthrough acceptance examples
- PVE 9 multi-AMD macOS acceptance runbook and evidence
- Whole-GPU passthrough implementation plan
- SSH setup and key rotation
- Archive inspection and rsync transfer
- Temporary restore storage
GPU platform and PCI identities are always discovered live. Stream only the read-only inventory action; download and inspect the runner before a dry-run or mutation:
wget -qO- https://devs-guide.github.io/proxmox/setup/vm/gpu.sh | \
bash -s -- --action inventory --output jsonwget -qO- https://devs-guide.github.io/proxmox/cli/ssh/sync.sh | \
bash -s -- --action setup --remote-host 10.0.0.11Rotate the dedicated key only after the fresh key is installed and verified:
wget -qO- https://devs-guide.github.io/proxmox/cli/ssh/sync.sh | \
bash -s -- --action setup --remote-host 10.0.0.11 \
--key-rotation --yeswget -qO- https://devs-guide.github.io/proxmox/cli/storage/temp.sh | \
bash -s -- --action status --vm 200wget -qO- https://devs-guide.github.io/proxmox/cli/rsync/fetch.sh | \
bash -s -- --action inspect --remote-host 10.0.0.11 \
--remote-path /backup/vzdump-qemu-200-date.vma.zst