{"id":2117,"date":"2026-01-28T02:45:41","date_gmt":"2026-01-28T10:45:41","guid":{"rendered":"https:\/\/wp.c9h.org\/cj\/?p=2117"},"modified":"2026-01-29T00:57:37","modified_gmt":"2026-01-29T08:57:37","slug":"part-2-taming-the-beast-deep-dive-into-the-proxy-aware-gpu-initialization-action","status":"publish","type":"post","link":"https:\/\/wp.c9h.org\/cj\/?p=2117","title":{"rendered":"Part 2: Taming the Beast – Deep Dive into the Proxy-Aware GPU Initialization Action"},"content":{"rendered":"Part
\n2: Taming the Beast – Deep Dive into the Proxy-Aware GPU Initialization
\nAction<\/h1>\n

In Part 1 of this series, we laid the network foundation for running
\nsecure Dataproc clusters. Now, let\u2019s zoom in on the core component
\nresponsible for installing and configuring NVIDIA GPU drivers and the
\nassociated ML stack in this restricted environment: the
\ninstall_gpu_driver.sh<\/code> script from the LLC-Technologies-Collier\/initialization-actions<\/a>
\nrepository (branch gpu-202601<\/code>).<\/p>\n

This isn\u2019t just any installation script; it has been significantly
\nenhanced to handle the nuances of Secure Boot and to operate seamlessly
\nbehind an HTTP\/S proxy.<\/p>\nThe
\nChallenge: Installing GPU Drivers Without Direct Internet<\/h2>\n

Our goal was to create a Dataproc custom image with NVIDIA GPU
\ndrivers, sign the kernel modules for Secure Boot, and ensure the entire
\nprocess works seamlessly when the build VM and the eventual cluster
\nnodes have no direct internet access, relying solely on an HTTP\/S proxy.
\nThis involved:<\/p>\n

    \n
  1. Proxy-Aware Build:<\/strong> Ensuring all build steps within
    \nthe custom image creation process (package downloads, driver downloads,
    \nGPG keys, etc.) correctly use the customer\u2019s proxy.<\/li>\n
  2. Secure Boot Signing:<\/strong> Integrating kernel module
    \nsigning using keys managed in GCP Secret Manager, especially when
    \ndrivers are built from source.<\/li>\n
  3. Conda Environment:<\/strong> Reliably and speedily installing
    \na complex Conda environment with PyTorch, TensorFlow, Rapids, and other
    \nGPU-accelerated libraries through the proxy.<\/li>\n
  4. Dataproc Integration:<\/strong> Making sure the custom image
    \nworks correctly with Dataproc\u2019s own startup, agent processes, and
    \ncluster-specific configurations like YARN.<\/li>\n<\/ol>\nThe
    \nDevelopment Journey: Key Enhancements in
    \ninstall_gpu_driver.sh<\/code><\/h2>\n

    To address these challenges, the script incorporates several key
    \nfeatures:<\/p>\n