meta-nvidia: bump NVIDIA stack to 580.95.05 (Blackwell CC support)

[kvinwang]

Summary

Adds NVIDIA driver 580.95.05 recipes for nvidia, nvidia-fabricmanager, and libnvidia-nscq. Required for RTX PRO 6000 Blackwell Server Edition (10de:2bb5) confidential compute under Intel TDX — the existing 570.172.08 driver predates Pro 6000 SE and bails out with GPU confidential compute capability is not enabled when loaded inside a TDX guest.

Yocto picks the highest available PV by default, so adding these recipes is enough to make builds use 580.95.05; the 570 recipes are kept for now.

What's in each recipe

• nvidia_580.95.05.bb — same structure as 570 recipe, only SRC_URI[md5sum] / SRC_URI[sha256sum] updated.
• nvidia-fabricmanager_580.95.05.bb — same as 570. The 580 archive ships extra gb300_* topology files; current do_install lists files explicitly so the new ones aren't picked up. Harmless for non-GB300 hosts; could be turned into a glob in a follow-up.
• libnvidia-nscq_580.95.05.bb — 580 archive no longer ships bin/nscq-cli, so do_install and FILES:${PN} were trimmed accordingly.

Verification

Built dstack-nvidia-dev-rootfs with the new stack and deployed to a TDX guest with one RTX PRO 6000 Blackwell Server Edition passthrough on Ubuntu intel-tdx kernel 6.14.0-1008-intel + tdx-patched QEMU. Inside the guest:

NVRM version: NVIDIA UNIX Open Kernel Module 580.95.05
GPU 0: NVIDIA RTX PRO 6000 Blackwell Server Edition

==============NVSMI CONF-COMPUTE LOG==============
    CC State                   : ON
    Multi-GPU Mode             : None
    CPU CC Capabilities        : INTEL TDX
    GPU CC Capabilities        : CC Capable
    CC GPUs Ready State        : Ready
    Protected memory size      : 99461312 KiB
    Unprotected memory size    : 0 KiB
    Environment                : PRODUCTION

Fabric Manager version is : 580.95.05
/usr/lib/libnvidia-nscq.so.580.95.05

Real workload: PyTorch 2.11+cu128 sees sm_120, runs cuBLAS GEMM (~78 TFLOP/s FP32, ~115 TFLOP/s BF16 on a single Pro 6000) entirely inside the TDX guest with CC ON.

Test plan

• bitbake nvidia succeeds
• bitbake nvidia-fabricmanager libnvidia-nscq succeeds
• bitbake dstack-nvidia-dev-rootfs succeeds; rootfs contains nvidia.ko@580.95.05, nv-fabricmanager reporting 580.95.05, libnvidia-nscq.so.580.95.05
• TDX guest boots, NVIDIA module loads on RTX PRO 6000 Blackwell SE
• nvidia-smi conf-compute -q shows CC State: ON
• PyTorch CUDA workload runs end-to-end

[Copilot]

Pull request overview

Adds Yocto recipes for the NVIDIA 580.95.05 driver stack to enable confidential compute support on newer Blackwell-era GPUs (e.g., RTX PRO 6000 Blackwell Server Edition) while keeping existing 570.x recipes in place.

Changes:

• Add nvidia_580.95.05.bb with updated 580.95.05 installer checksums/URL.
• Add nvidia-fabricmanager_580.95.05.bb to package Fabric Manager 580.95.05.
• Add libnvidia-nscq_580.95.05.bb, updated for the 580 archive layout (no nscq-cli).

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
meta-nvidia/recipes-graphics/nvidia/nvidia_580.95.05.bb	New 580.95.05 driver recipe (runfile extraction + includes).
meta-nvidia/recipes-graphics/nvidia/nvidia-fabricmanager_580.95.05.bb	New 580.95.05 Fabric Manager recipe and install manifest.
meta-nvidia/recipes-graphics/nvidia/libnvidia-nscq_580.95.05.bb	New 580.95.05 NSCQ library recipe; install/package list adjusted to archive contents.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[kvinwang]

Addressed @copilot-pull-request-reviewer feedback in 85e7e17 + added version pinning:

1. nvidia-fabricmanager_580.95.05.bb — replaced the explicit topology file list with a for f in ${S}/share/nvidia/nvswitch/*; do install …; done loop. Verified build picks up the new gb300_nvl72r{1,2}_c2g4_topology files that the old list was missing (29 topology files installed total).

2. Version pin (meta-dstack/conf/distro/dstack.conf):

   NVIDIA_VERSION = "580.95.05"
   PREFERRED_VERSION_nvidia                = "${NVIDIA_VERSION}"
   PREFERRED_VERSION_nvidia-fabricmanager  = "${NVIDIA_VERSION}"
   PREFERRED_VERSION_libnvidia-nscq        = "${NVIDIA_VERSION}"
   

   So the three packages always bump together — kernel module ABI and userspace libs are paired. bitbake-getvar --value confirms all four resolve to 580.95.05. Future bumps only need to change NVIDIA_VERSION.

[Copilot]

Addressed @copilot-pull-request-reviewer feedback in 85e7e17 + added version pinning:

1. nvidia-fabricmanager_580.95.05.bb — replaced the explicit topology file list with a `for f in ${S}/sha...

Both changes are already in place at commit 85e7e17:

• nvidia-fabricmanager_580.95.05.bb uses the glob loop (for f in ${S}/share/nvidia/nvswitch/*) instead of the explicit file list.
• meta-dstack/conf/distro/dstack.conf has NVIDIA_VERSION = "580.95.05" with the three PREFERRED_VERSION_* variables pinning all packages together.

No further changes needed.