User problem
TauGrid currently publishes Ray image variants optimized for GPU workloads, while the available CPU-only image may lag current Ray releases. Some users running local Kind clusters, CPU-only AKS node pools, CI, or lightweight development environments may prefer an image without CUDA and other GPU-focused dependencies.
It is not yet clear whether there is enough demand to maintain and publish CPU-only variants alongside the GPU images.
Desired outcome
Use this issue to track community and user interest in current CPU-only Ray images.
If a CPU-only image would improve your workflow, add a reaction or comment with the environment and workload you would use it for. That signal can help prioritize the work. The issue can remain open as a reference until there is sufficient demand or the maintainers choose to implement it.
Proposed approach
If demand justifies the additional build, testing, publication, and support cost, introduce CPU-only variants for supported Ray releases. They would retain the common TauGrid Ray functionality, such as the Azure Linux base, Ray default/data/serve extras, multi-architecture support, and KubeRay health-probe compatibility, while excluding dependencies needed only for GPU workloads.
The exact tag format, supported version policy, build structure, release cadence, and validation requirements should be decided when the work is prioritized rather than prescribed by this issue.
Alternatives considered
- Continue using the existing CUDA-capable image for both GPU and CPU-only environments.
- Continue using the older published CPU-only image where its Ray version is sufficient.
- Use upstream Ray CPU images when the TauGrid-specific image behavior is not required.
- Let users build application-specific CPU images until maintaining an official variant is justified.
Additional context
The image source and version matrix are maintained under images/ray in this repository, while publication is handled by a separate Azure DevOps pipeline.
During #294 validation, using the current CUDA image in Kind worked correctly. A CPU-only image would primarily reduce unnecessary transfer and storage for non-GPU environments rather than unblock the workflow.
Confirmation
User problem
TauGrid currently publishes Ray image variants optimized for GPU workloads, while the available CPU-only image may lag current Ray releases. Some users running local Kind clusters, CPU-only AKS node pools, CI, or lightweight development environments may prefer an image without CUDA and other GPU-focused dependencies.
It is not yet clear whether there is enough demand to maintain and publish CPU-only variants alongside the GPU images.
Desired outcome
Use this issue to track community and user interest in current CPU-only Ray images.
If a CPU-only image would improve your workflow, add a reaction or comment with the environment and workload you would use it for. That signal can help prioritize the work. The issue can remain open as a reference until there is sufficient demand or the maintainers choose to implement it.
Proposed approach
If demand justifies the additional build, testing, publication, and support cost, introduce CPU-only variants for supported Ray releases. They would retain the common TauGrid Ray functionality, such as the Azure Linux base, Ray
default/data/serveextras, multi-architecture support, and KubeRay health-probe compatibility, while excluding dependencies needed only for GPU workloads.The exact tag format, supported version policy, build structure, release cadence, and validation requirements should be decided when the work is prioritized rather than prescribed by this issue.
Alternatives considered
Additional context
The image source and version matrix are maintained under
images/rayin this repository, while publication is handled by a separate Azure DevOps pipeline.During #294 validation, using the current CUDA image in Kind worked correctly. A CPU-only image would primarily reduce unnecessary transfer and storage for non-GPU environments rather than unblock the workflow.
Confirmation