vLLM dropped CUDA 12.9: an 11-step migration checklist for self-hosted LLMs on Kubernetes GPU nodes

vLLM v0.33.0 replaces CUDA 12.9 artifacts with CUDA 13.0-only wheels and images, making upgrades for Kubernetes-hosted inference fleets a node-level migration rather than a container tag update. CUDA 13 requires NVIDIA Linux driver 580.65.06 or later and no longer supports Maxwell, Pascal, or Volta GPUs, meaning V100-based fleets must remain pinned to vLLM v0.32.0’s CUDA 12.9 build or replace their hardware, while Turing and newer GPUs can migrate after driver updates. The recommended approach is a blue/green deployment: inventory hardware and software versions, preserve and mirror the old image by digest for rollback, create a separate GPU node pool with updated drivers and GPU stack components, rebuild CUDA-dependent extensions, smoke-test and canary the new release, benchmark performance and memory behavior, then carefully drain the old pool with extended termination grace periods for streaming traffic. The discussion emphasizes validating latency, throughput, GPU memory, KV-cache pressure, correctness, and multi-GPU NCCL behavior instead of relying on pod health alone, and advises defining measurable rollback thresholds before migration. It also notes differences among EKS, GKE, and AKS driver defaults, capacity and quota constraints during dual-pool operation, the limited usefulness of CUDA forward-compatibility packages, and the tradeoff between temporarily pinning older software, managing infrastructure through a platform such as Qovery, or using a managed inference API.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.