When we asked the community what hurts about GPUs on Cozystack, the answer was not a broken feature but a missing shape: there are no described scenarios and no configuration examples showing how a GPU workload actually gets deployed, and no explanation of why the platform does it this way. Customers ask for containers, and today the docs read as if virtual machines are the only story. https://t.me/cozystack/16793
The pieces exist and are individually documented. A tenant Kubernetes node group takes gpus and gets the device passed through, the gpuOperator addon installs the NVIDIA stack inside the tenant cluster, and the optional hami system package slices one physical GPU across pods. What is missing is the line connecting them: a reader who wants "my container gets a GPU" has to assemble that path from three separate pages and infer that the VM is an implementation detail of the node, not of their workload.
Concretely, the docs should carry:
- The end-to-end path for a containerised GPU workload, from node group with a GPU through gpu-operator to a pod requesting
nvidia.com/gpu, with the manifests spelled out.
- Why worker nodes are VMs and what that costs, so the VM-versus-container question is answered once instead of in every conversation. HAMi's sharing model, and its glibc constraint, belongs here too.
This is a documentation gap rather than a feature request, but it is the one that decides whether someone believes the platform can run their AI workloads at all.
When we asked the community what hurts about GPUs on Cozystack, the answer was not a broken feature but a missing shape: there are no described scenarios and no configuration examples showing how a GPU workload actually gets deployed, and no explanation of why the platform does it this way. Customers ask for containers, and today the docs read as if virtual machines are the only story. https://t.me/cozystack/16793
The pieces exist and are individually documented. A tenant Kubernetes node group takes
gpusand gets the device passed through, thegpuOperatoraddon installs the NVIDIA stack inside the tenant cluster, and the optionalhamisystem package slices one physical GPU across pods. What is missing is the line connecting them: a reader who wants "my container gets a GPU" has to assemble that path from three separate pages and infer that the VM is an implementation detail of the node, not of their workload.Concretely, the docs should carry:
nvidia.com/gpu, with the manifests spelled out.This is a documentation gap rather than a feature request, but it is the one that decides whether someone believes the platform can run their AI workloads at all.