Gpu
Gpu
- k8s-gpu-scheduler-lab
partial
A controlled comparison of Kubernetes GPU schedulers on identical traces, on a real control plane with simulated GPU nodes. K0 and four degenerate baselines are built; Kueue, Volcano and KAI are planned.
- Epilog GPU Validator
built
A Slurm Epilog check of the GPUs a job just used: it drains a node on evidence of a persistent GPU fault, and never on a failed query without fault evidence or on a transient condition.
- GPU Reaper
built
Find wasted GPU allocations on a Slurm cluster and, only if enabled, drain and cancel them, with guardrails that refuse to act on missing, stale or unreadable telemetry.
- How KV-Cache Paging Works in vLLM — and Why It Matters for Production
A walkthrough of PagedAttention, the KV-cache memory manager from the vLLM paper (Kwon et al., SOSP 2023), and why paging matters for serving many concurrent LLM sessions.