Skip to content

Request: Add open source kernel module (nvos) builds to available driver versions, including beta branch #6

@mricharz

Description

@mricharz

Problem

Blackwell GPUs (RTX Pro 6000, RTX 5090, etc.) require the open source kernel modules — NVIDIA dropped proprietary kernel module support for sm_120 entirely.

Currently the plugin UI lists only proprietary driver versions as radio buttons:

  • Production branch: 580.x (proprietary only)
  • Feature branch: 590.x (proprietary only)
  • Latest/Beta: 595.x (proprietary only)

Open source builds (nvos-590.*) exist in the GitHub releases but are not listed in the plugin's available versions. Blackwell users can only get open source drivers via the latest_nos option — with no control over which version they get.

Request

  1. Add open source driver versions to the available versions list in the plugin UI, so users can explicitly select them (e.g. nvos-590.48.01)
  2. Build open source packages for the beta branch (595.x) as well — Blackwell users currently cannot access the latest drivers at all

Why this matters

  • cuBLAS 13.3 (driver 595+) brings up to 20% BF16 GEMM speedup on RTX Pro 6000
  • cuDNN 9.20 and CUDA runtime 13.2 also require driver 595+
  • Blackwell users are stuck on 590 with no path to newer drivers
  • compile.sh already supports open source builds via --kernel-module-type=open for any driver >= 560
  • The 595.45.04 tag exists on NVIDIA/open-gpu-kernel-modules

Attempted self-build

We tried building nvos-595.45.04 ourselves using the unraid_kernel Docker container and compile.sh, but ran into multiple issues:

  • nvidia-installer detects the host's loaded NVIDIA kernel modules via /proc and aborts, even with --no-questions
  • makepkg is missing in the container (bzroot-extracted-*/sbin/makepkg: No such file or directory)
  • CPU_COUNT and UNAME environment variables are not set by the container, causing compile.sh to silently fail or produce broken flags (--j--kernel-module-type=open)

The CI/Jenkins build presumably works in a clean environment without loaded NVIDIA modules. For end users, self-building is currently not feasible.

Context

Blackwell adoption is growing. Without open source builds for newer driver branches, Blackwell users are effectively second-class citizens in the plugin ecosystem.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions