nx_gpu_broker.nx

nx_gpu_broker.nx -- THE RESOURCE-SHARING BRAIN (serving-census capstone). Admits/schedules a MIXED workload -- background image-gen (Z-Image DiT, big VRAM, low priority, preemptible) + interactive chat (small VRAM, high priority, latency-sensitive) -- onto ONE 16GB GPU, using BOTH levers together: (1) VRAM-AWARE admission (never oversubscribe the card; co-resident jobs whose VRAM sums <= budget run TOGETHER) and (2) QoS PREEMPTION (a high-priority arrival preempts enough low-priority resident jobs to fit, and the preempted job resumes later with its work intact). This is the operator's "resource sharing" highlight made a working decision-maker -- it composes the verified nx_vram_estimate (budgets) + nx_sched_qos (preemption) ideas into one broker. Pure logic, no GPU/model needed; VRAM never oversubscribed BY CONSTRUCTION (checked every op). license_tier: ORIGINAL expect_exit: 0