Skip to content

Add vGPU regional GCS download bypass to Ubuntu driver installer and align Autopilot NRI device injector - #625

Open
jkru3 wants to merge 1 commit into
GoogleCloudPlatform:masterfrom
jkru3:vgpu-ubuntu-option-a
Open

Add vGPU regional GCS download bypass to Ubuntu driver installer and align Autopilot NRI device injector#625
jkru3 wants to merge 1 commit into
GoogleCloudPlatform:masterfrom
jkru3:vgpu-ubuntu-option-a

Conversation

@jkru3

@jkru3 jkru3 commented Jul 17, 2026

Copy link
Copy Markdown
Collaborator

Ubuntu Driver Installer (nvidia-driver-installer/ubuntu/entrypoint.sh):

  • Add is_vgpu() to detect fractional g4-standard-* GCE machine types via metadata and /etc/nvidia/gpu_fraction_divisor.txt
  • Add configure_vgpu_download() to dynamically resolve the local GCP region (metadata.google.internal/computeMetadata/v1/instance/zone) and construct regional GCS mirror URL (gs://cos-nvidia-gpu-drivers-\${region}/NVIDIA-Linux-x86_64-580.126.09-grid-gcp.run)
  • Update download_nvidia_installer() to fetch the GCE instance metadata ADC access token and attach Authorization: Bearer \${token} for authenticated GCS runfile downloads
  • Update download_kernel_src() to detect host kernel headers directly at /root/usr/src/linux-headers-\${KERNEL_VERSION} (or /root/lib/modules/.../build), skipping legacy container apt-get calls when running on modern GKE Ubuntu 24.04 nodes
  • Update run_nvidia_installer() and update_host_ld_cache() to stop services and unload active kernel modules (rmmod -f ecc nvidia_uvm nvidia_drm nvidia_modeset nvidia), pass --no-questions and --kernel-source-path, write 8 to gpu_fraction_divisor.txt, and synchronize compiled libraries to \${NVIDIA_INSTALL_DIR_HOST}/lib64 (/home/kubernetes/bin/nvidia/lib64/)

NRI Device Injector (nri_device_injector/nri-device-injector-autopilot.yaml):

  • Add nvidia-rtx-pro-6000-vws to nodeSelectorTerms under matchExpressions: cloud.google.com/gke-accelerator, bringing Autopilot to parity with standard nri-device-injector.yaml so character devices (/dev/nvidia-caps/*) are injected into vGPU container namespaces across all cluster modes

Change-Id: Ic7dd70954322d0585ef2c1d5afc7c3613645fd03

…align Autopilot NRI device injector

Ubuntu Driver Installer (`nvidia-driver-installer/ubuntu/entrypoint.sh`):
- Add is_vgpu() to detect fractional `g4-standard-*` GCE machine types via metadata and `/etc/nvidia/gpu_fraction_divisor.txt`
- Add `configure_vgpu_download()` to dynamically resolve the local GCP region (`metadata.google.internal/computeMetadata/v1/instance/zone`) and construct regional GCS mirror URL (`gs://cos-nvidia-gpu-drivers-\${region}/NVIDIA-Linux-x86_64-580.126.09-grid-gcp.run`)
- Update `download_nvidia_installer()` to fetch the GCE instance metadata ADC access token and attach `Authorization: Bearer \${token}` for authenticated GCS runfile downloads
- Update `download_kernel_src()` to detect host kernel headers directly at `/root/usr/src/linux-headers-\${KERNEL_VERSION}` (or `/root/lib/modules/.../build`), skipping legacy container apt-get calls when running on modern GKE Ubuntu 24.04 nodes
- Update `run_nvidia_installer()` and `update_host_ld_cache()` to stop services and unload active kernel modules (`rmmod -f ecc nvidia_uvm nvidia_drm nvidia_modeset nvidia`), pass `--no-questions` and `--kernel-source-path`, write 8 to gpu_fraction_divisor.txt, and synchronize compiled libraries to `\${NVIDIA_INSTALL_DIR_HOST}/lib64` (`/home/kubernetes/bin/nvidia/lib64/`)

NRI Device Injector (`nri_device_injector/nri-device-injector-autopilot.yaml`):
   - Add `nvidia-rtx-pro-6000-vws` to `nodeSelectorTerms` under `matchExpressions: cloud.google.com/gke-accelerator`, bringing Autopilot to parity with standard `nri-device-injector.yaml` so character devices (`/dev/nvidia-caps/*`) are injected into vGPU container namespaces across all cluster modes

Change-Id: Ic7dd70954322d0585ef2c1d5afc7c3613645fd03
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant