Repository navigation
Update nvidia-ctk --config-source flag from 'command' to 'file' - #76
Merged
cdesiniotis merged 1 commit intoJun 9, 2026
Merged
Conversation
Using command as a config source can be problematic with containerd. The 'containerd config dump' command does not reliably inform us of the config file version currently set in the top-level /etc/containerd/config.toml file. For newer versions of containerd that support config versions 3 (or even 4), we fail to generate a valid drop-in config file if the top-level config is using version 2 of the containerd config schema. In kind worker nodes (e.g. kindest/node:$k8s-ver), the top-level /etc/containerd/config.toml is populated with enough information for nvidia-ctk to create a functional drop-in file at /etc/containerd/conf.d/99-nvidia.toml. So it should be safe, at least in kind environments, to use 'file' as our config source. Signed-off-by: Christopher Desiniotis <cdesiniotis@nvidia.com>
tariq1890
approved these changes
Jun 9, 2026
dims
approved these changes
Jun 9, 2026
harche
added a commit
to harche/k8s-dra-driver-gpu
that referenced
this pull request
Oct 5, 2026
Implement WatchHealthStatus so the kubelet can surface the health of allocated GPUs in pod.status.containerStatuses[].allocatedResourcesStatus. Health is derived from the device taints the NVML health monitor already publishes in the ResourceSlice, so the pod status and the scheduler see the same classification: fatal XIDs and GPU loss are Unhealthy, unmonitored devices are Unknown, non-fatal XIDs stay Healthy with a message. Reports are sent on subscribe, on every taint change and every 10 seconds, and never wait behind a Prepare or Unprepare holding the device state lock. The DRAResourceHealth service is only advertised when the NVML health monitor runs (NVMLDeviceHealthCheck). Move the GCP nvkind e2e harness to Kubernetes 1.37, where ResourceHealthStatus is on by default: use kind v0.33.0 (kubeadm v1beta4) and rebuild nvkind at main (NVIDIA/nvkind#76, nvidia-ctk --config-source=file) against it. Also keep macOS tar from adding AppleDouble files that Helm fails to parse as CRDs.
harche
added a commit
to harche/k8s-dra-driver-gpu
that referenced
this pull request
Oct 5, 2026
Implement WatchHealthStatus so the kubelet can surface the health of allocated GPUs in pod.status.containerStatuses[].allocatedResourcesStatus. Health is derived from the device taints the NVML health monitor already publishes in the ResourceSlice, so the pod status and the scheduler see the same classification: fatal XIDs and GPU loss are Unhealthy, unmonitored devices are Unknown, non-fatal XIDs stay Healthy with a message. Reports are sent on subscribe, on every taint change and every 10 seconds, and never wait behind a Prepare or Unprepare holding the device state lock. The DRAResourceHealth service is only advertised when the NVML health monitor runs (NVMLDeviceHealthCheck). Move the GCP nvkind e2e harness to Kubernetes 1.37, where ResourceHealthStatus is on by default: use kind v0.33.0 (kubeadm v1beta4) and rebuild nvkind at main (NVIDIA/nvkind#76, nvidia-ctk --config-source=file) against it. Also keep macOS tar from adding AppleDouble files that Helm fails to parse as CRDs. Bump the mock NVML (NVIDIA/k8s-test-infra) to 57ef0165, which delivers injected XID events to NVML health monitors (NVIDIA/k8s-test-infra#735), and emulate 4 GPUs to match its 4-GPU gb200 profile; GPUs beyond the profile get a random UUID per process, so nvidia-smi cannot find them.
harche
added a commit
to harche/k8s-dra-driver-gpu
that referenced
this pull request
Oct 5, 2026
Implement WatchHealthStatus so the kubelet can surface the health of allocated GPUs in pod.status.containerStatuses[].allocatedResourcesStatus. Health is derived from the device taints the NVML health monitor already publishes in the ResourceSlice, so the pod status and the scheduler see the same classification: fatal XIDs and GPU loss are Unhealthy, unmonitored devices are Unknown, non-fatal XIDs stay Healthy with a message. Reports are sent on subscribe, on every taint change and every 10 seconds, and never wait behind a Prepare or Unprepare holding the device state lock. The DRAResourceHealth service is only advertised when the NVML health monitor runs (NVMLDeviceHealthCheck). Move the GCP nvkind e2e harness to Kubernetes 1.37, where ResourceHealthStatus is on by default: use kind v0.33.0 (kubeadm v1beta4) and rebuild nvkind at main (NVIDIA/nvkind#76, nvidia-ctk --config-source=file) against it. Also keep macOS tar from adding AppleDouble files that Helm fails to parse as CRDs. Bump the mock NVML (NVIDIA/k8s-test-infra) to 57ef0165, which delivers injected XID events to NVML health monitors (NVIDIA/k8s-test-infra#735), and emulate 4 GPUs to match its 4-GPU gb200 profile; GPUs beyond the profile get a random UUID per process, so nvidia-smi cannot find them.
harche
added a commit
to harche/k8s-dra-driver-gpu
that referenced
this pull request
Oct 5, 2026
Implement WatchHealthStatus so the kubelet can surface the health of allocated GPUs in pod.status.containerStatuses[].allocatedResourcesStatus. Health is derived from the device taints the NVML health monitor already publishes in the ResourceSlice, so the pod status and the scheduler see the same classification: fatal XIDs and GPU loss are Unhealthy, unmonitored devices are Unknown, non-fatal XIDs stay Healthy with a message. Reports are sent on subscribe, on every taint change and every 10 seconds, and never wait behind a Prepare or Unprepare holding the device state lock. The DRAResourceHealth service is only advertised when the NVML health monitor runs (NVMLDeviceHealthCheck). Move the GCP nvkind e2e harness to Kubernetes 1.37, where ResourceHealthStatus is on by default: use kind v0.33.0 (kubeadm v1beta4) and rebuild nvkind at main (NVIDIA/nvkind#76, nvidia-ctk --config-source=file) against it. Also keep macOS tar from adding AppleDouble files that Helm fails to parse as CRDs. Bump the mock NVML (NVIDIA/k8s-test-infra) to 57ef0165, which delivers injected XID events to NVML health monitors (NVIDIA/k8s-test-infra#735), and emulate 4 GPUs to match its 4-GPU gb200 profile; GPUs beyond the profile get a random UUID per process, so nvidia-smi cannot find them.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Using command as a config source can be problematic with containerd. The 'containerd config dump' command does not reliably inform us of the config file version currently set in the top-level /etc/containerd/config.toml file. For newer versions of containerd that support config versions 3 (or even 4), we fail to generate a valid drop-in config file if the top-level config is using version 2 of the containerd config schema.
In kind worker nodes (e.g. kindest/node:$k8s-ver), the top-level /etc/containerd/config.toml is populated with enough information for nvidia-ctk to create a functional drop-in file at /etc/containerd/conf.d/99-nvidia.toml. So it should be safe, at least in kind environments, to use 'file' as our config source.