Episode 14 · Planned
Packaging, accelerator preflight and memory containment
Verify reproducible containers, accelerator visibility, model loading and bounded memory behavior.
Build reproducible serving containers and verify the NVIDIA container runtime/CDI path, driver/library compatibility, model loading and streaming contract. Create controlled accelerator-not-visible and memory-pressure drills. Locate a repeatable KV-cache or OOM boundary and verify at least one admission, rejection or recovery control.
Evidence: immutable manifests and digests, compatibility matrix, effective launch flags, failure signatures, memory telemetry, client-visible behavior and verified recovery state.