** Description changed: [Availability] - The package dgx-desktop-defaults is already in Ubuntu universe. - The package dgx-desktop-defaults builds for the architectures it is designed to work on. - It currently builds and works for architectures: all - Link to package [link](https://launchpad.net/ubuntu/+source/dgx-desktop-defaults) [Rationale] - The package dgx-desktop-defaults is required in Ubuntu main because Nvidia DGX Spark would benefit to have it all the time. - The package dgx-desktop-defaults will generally be useful for a large part of our user base. - Currently, the users of these machines get the packages only if Subiquity runs online, thanks to ubuntu-drivers. The binary packages within dgx-desktop-defaults needs to be in main because it needs to be present on the live installer ISO so that offline installations work too. - There is no other/better way to solve this that is already in main or should go universe->main instead of this. - This is the first time package will be in main - All binary packages built by dgx-desktop-defaults need to be in main. - It would be great and useful to community/processes to have the package in Ubuntu main by the next point release of 26.04, to be in the built ISO. [Security] - No CVEs/security issues in this software in the past - no executables in `/sbin` and `/usr/sbin` - Package does install services, timers or recurring jobs - cuda-gpu-select.service tune the file `/etc/cuda-gpu-select/env` on boot - dgx-desktop-pro-activation.service loads pro tokens from UEFI on startup - systemd generator dgx-desktop-docker-gpus.sh tunes containerd override configuration during boot - systemd generator dgx-desktop-load-realtek-driver.sh loads one realtek driver or the other, depending on the kernel - no `suid` or `sgid` binaries - Packages does not open privileged ports (ports < 1024). - Package does not expose any external endpoints - Packages does not contain extensions to security-sensitive software (filters, scanners, plugins, UI skins, ...) [Quality assurance - function/usage] - The package works well right after install [Quality assurance - maintenance] - The package is maintained well in Ubuntu and does not have too many, long-term & critical, open bugs - Ubuntu [https://bugs.launchpad.net/ubuntu/+source/dgx-desktop-defaults/+bug](https://bugs.launchpad.net/ubuntu/+source/dgx-desktop-defaults/+bug) - The package does deal with exotic hardware, such hardware is available to the team for debugging, test, verification and development via: - testflinger under the following queue: [nvidia-n1x-spark-prod](https://testflinger.canonical.com/queues/nvidia-n1x-spark-prod) - testflinger under the following queue: [gb300-galaxy](https://testflinger.canonical.com/queues/gb300-galaxy) - the nvidia global partnership. Also exotic is opinionated, DGX Sparks are popular customer products. [Quality assurance - testing] - The package does not run a test at build time because it runs its test suite as autopkgtest. - We have access to such hardware in the team to test package changes. + - Test plan + - The metas (`hwe-dgx-gb10-meta`, `hwe-dgx-gb300ws-meta`) should match the right hardware, respectively DGX Sparks and DGX Stations. + - `dgx-desktop-pro-activation`: a freshly provisioned DGX Spark should automatically try to attach to a Pro subscription. If the token is compatible with the distro, the attach should work. + - `dgx-desktop-sbsa-gwdt-options`, `dgx-desktop-disable-init-on-alloc`, `dgx-desktop-disable-numa-balancing`, `dgx-desktop-crashkernel-configuration`, `dgx-desktop-cppc-cpufreq-options`, `dgx-desktop-enable-power-meter-cap` : the relevant kernel argument should be applied on boot. The device should boot correctly, and not get unattended reboot after 30min. The `power_cap` sysfs nodes should be exposed. + - `dgx-desktop-limits` is quite trivial, any change done here would be entirely revisiting the behavior so there is no regression to look for really + - `dgx-desktop-nvidia-fs-loader`: the `nvidia-fs` module should load during the boot. + - `dgx-desktop-enable-persistenced`: The GPU should be snappy. `sudo docker run --rm --runtime=nvidia --gpus all ubuntu nvidia-smi` should run instantly several time in a row, and not take a few seconds each time. + - `dgx-desktop-nvme-interrupt-coalescing`: interrupt coalescing should be enable with machines with the relevant nvme vendor (Samsung, Kioxia, Micron). That can be verified with + ``` + $ sudo nvme get-feature /dev/nvme0 -f 8 + get-feature:0x08 (Interrupt Coalescing), Current value:0x00000107 + ``` + A disabled state would be `0x00000000`. + - `dgx-desktop-docker-gpus`: `sudo docker run --rm --runtime=nvidia --gpus all ubuntu nvidia-smi` should work. The nvidia docker runtime should only be able to use compute GPUs + That can be verified with the following line, where the dGPU is expected NOT to appear. + ``` + $ sudo docker run --rm --gpus all ubuntu nvidia-smi -L + GPU 0: NVIDIA GB300 (UUID: GPU-4205f1fd-9c05-ef26-2a33-34578900205c) + ``` + - `dgx-desktop-nvidia-cuda-environment`: on a DGX GB300 workstation, after a reboot, CUDA should only run on the GB300 GPU and not on a potential dGPU. That can be verified with + ``` + echo "$CUDA_VISIBLE_DEVICES" + nvidia-smi -i "$CUDA_VISIBLE_DEVICES" --query-gpu=name --format=csv,noheader + ``` + - `dgx-desktop-arp-configuration` can be tested through these commands: + ``` + # On the prober — ask "who has 10.0.0.11" + arping -c 3 10.102.182.27 + # Baseline: you may see replies from multiple MAC addresses (flux) + # then install the package + # On the prober + arping -c 3 10.102.182.27 # should be answered ONLY by eth1's MAC + arping -c 3 10.102.182.28 # should be answered ONLY by eth0's MAC + ``` + - `dgx-desktop-app-profiles-mixed-coherency`: when running an opengl workload like glxgears, the dGPU should always be picked. That can be verified with `nvidia-smi`. [Quality assurance - packaging] - debian/watch is not present because it is a native package - debian/control defines a correct Maintainer field - Lintian overrides are not present - This package does not rely on obsolete or about to be demoted packages. - The package will be installed by default, but does not ask debconf questions higher than medium - Packaging and build is easy, link to debian/rules [UI standards] - Application is not end-user facing (does not need translation) [Dependencies] - Used check-mir from ubuntu-dev-tools to validate all dependencies or recommends are in main. [Standards compliance] - This package correctly follows FHS and Debian Policy [Maintenance/Owner] - The owning team will be Partner Engineering and I have their acknowledgment for that commitment - The future owning team is already subscribed to the package - This does not use static builds - This does not use vendored code - This package is not rust based - The package has been built within the last 3 months in the archive - This change will not impact other teams [Background information] - The Package description explains the package well - No Upstream Name - it's a native package
-- You received this bug notification because you are a member of Ubuntu Bugs, which is subscribed to Ubuntu. https://bugs.launchpad.net/bugs/2154816 Title: [MIR] dgx-desktop-defaults To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/dgx-desktop-defaults/+bug/2154816/+subscriptions -- ubuntu-bugs mailing list [email protected] https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
