Training a 7B model on 4090 clusters: feasibility boundaries

VRAM walls, interconnect bottlenecks, and when to switch to A100 — lessons from a long run.