fd0f4a0365
6d224861 came from a --no-checkout worktree whose empty index made the
commit drop every file except two. This restores 05b50d2b's full tree and
correctly adds stacks/stem95su/gdrive-sync.tf + the service-catalog stem95su
entry. Forward-only (parent=6d224861, no force-push); [ci skip] since the
live infra was never applied from the broken commit.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
80 lines
3 KiB
YAML
80 lines
3 KiB
YAML
driver:
|
||
enabled: true
|
||
# repository: nvcr.io/nvidia/driver
|
||
# choose a driver version compatible with your GPU + CUDA 12.x (example)
|
||
# NVIDIA GPU driver - https://docs.nvidia.com/datacenter/cloud-native/gpu-operator/latest/platform-support.html#known-issue
|
||
# https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/
|
||
# 13.x >= 580
|
||
# 12.x >= 525, <580
|
||
# 11.x >= 450, <525
|
||
#
|
||
# Delete the cluster policy before each change
|
||
# version: "575.57.08" # CUDA 12.9
|
||
#
|
||
# 2026-05-17: tried bumping to 580.x with kernelModuleType=open but
|
||
# NVIDIA has NOT published any nvcr.io/nvidia/driver:*-ubuntu26.04
|
||
# images yet (skopeo list-tags shows 0 ubuntu26.04 tags vs 779 for
|
||
# ubuntu22.04 and 206 for ubuntu24.04). Chart v26.3.1's operator
|
||
# auto-detects the host OS (k8s-node1 was upgraded to Ubuntu 26.04
|
||
# with kernel 7.0.0-15-generic) and picks `<version>-ubuntu26.04` —
|
||
# which then 404s on pull. Rolled back to chart v25.10.1 + this
|
||
# 570.195.03 pin, which uses the ubuntu24.04 image suffix. That
|
||
# image still can't compile against kernel 7.0.0 (apt sources are
|
||
# 24.04 noble, which doesn't ship linux-headers-7.0.0-15-generic),
|
||
# so the host kernel needs to be rolled back to 6.8.0-117-generic
|
||
# (still installed in /boot) before the driver can come up.
|
||
# See post-mortem 2026-05-17-gpu-driver-ubuntu2604-mismatch.md.
|
||
version: "570.195.03" # CUDA 12.8 — pinned until NVIDIA ships ubuntu26.04 images
|
||
upgradePolicy:
|
||
autoUpgrade: false
|
||
|
||
# 2026-05-17: bumped from the namespace LimitRange default of 128Mi.
|
||
# The driver-installer's `apt-get install linux-headers-<kernel>` step
|
||
# exceeded 128Mi and OOMKilled (exit 137) before producing any visible
|
||
# output beyond "Installing Linux kernel headers...". 2Gi limit gives
|
||
# the apt + module-compile phase enough headroom (peak observed ~1.4Gi
|
||
# while DKMS builds the kernel module).
|
||
resources:
|
||
requests:
|
||
cpu: "50m"
|
||
memory: "822Mi"
|
||
limits:
|
||
memory: "2Gi"
|
||
|
||
# 2026-05-25: extended startup probe from 120 to 300 failures.
|
||
# On k8s-node1 (6 vCPUs, 16Gi RAM, Ubuntu 24.04 + 6.8.0-117-generic),
|
||
# the full driver install sequence — apt install linux-headers (~2min) +
|
||
# gcc make -j16 kernel module compilation (~12min) + nvidia-installer
|
||
# file copy (~7min) = ~21min total, which exactly exhausted the default
|
||
# 120×10s=20min window (exit 137 = SIGKILL from startup probe).
|
||
# 300×10s = 50min gives 2.5× headroom on this hardware.
|
||
startupProbe:
|
||
failureThreshold: 300
|
||
|
||
devicePlugin:
|
||
config:
|
||
name: time-slicing-config
|
||
|
||
# DCGM Exporter - reduced to 768Mi (actual usage ~489Mi, 1.5x margin)
|
||
dcgmExporter:
|
||
resources:
|
||
requests:
|
||
memory: "768Mi"
|
||
limits:
|
||
memory: "768Mi"
|
||
|
||
# CUDA Validator - reduced from 1024Mi to 256Mi (one-shot job)
|
||
validator:
|
||
resources:
|
||
requests:
|
||
memory: "256Mi"
|
||
limits:
|
||
memory: "256Mi"
|
||
|
||
# Tolerate GPU node taint for all GPU operator components
|
||
daemonsets:
|
||
tolerations:
|
||
- key: "nvidia.com/gpu"
|
||
operator: "Equal"
|
||
value: "true"
|
||
effect: "NoSchedule"
|