nx_swarm_gpu.nx

nx_swarm_gpu.nx -- SWARM FABRIC GPU/VRAM inventory + SPLIT-MODEL SHARD PLANNER (F792). The keystone of split-model serve (exo/Petals-class, sovereign): given a model too big for ONE GPU, decide how to shard its layers across the fleet's LIVE free VRAM. Its OWN small organ (NEVER bolt onto nx_swarm_beat -- the CPU-beacon 13-token contract stays untouched); the pulse layer reports a GPU row here, the hub reads a fresh VRAM inventory + computes a shard plan. Composes the proven nx_swarm_beat store pattern (replace-by- node snap, flock'd log, fabric-clock freshness, path law). nx_swarm_gpu put "<GPU row>" -- validate 5-token row -> log append + snap replace nx_swarm_gpu show <window_sec> -- FRESH/STALE per node (the live VRAM inventory) nx_swarm_gpu plan <snap> <window_sec> <model_mb> [util%] -- greedy shard plan across fresh nodes' free VRAM; REFUSE if fleet usable < model (never oversubscribe) nx_swarm_gpu now -- fabric clock (stamp rows in the freshness units) nx_swarm_gpu [gate] -- self-gate incl the liar-killers GPU row (positional): GPU <node> <vram_total_mb> <vram_free_mb> <ts_us> util% = headroom margin (default 85): usable = free_vram * util%/100 -- leaves room for activations + KV-cache; a GPU is NEVER filled to 100% VRAM (an OOM mid-decode is the never-brick analog here). Greedy snap-order fill minimizes the shard count = fewer cross-node activation transfers (Amdahl-friendly); biggest-first = a v2 sort. PATH LAW mirrors nx_swarm_beat: CLI paths end .log/.snap, no "..", bare-basename-in-CWD or /tmp only. license_tier: ORIGINAL expect_exit:0