nx_fabric_collective.nx

nx_fabric_collective.nx -- R4 of the SkyHammer fabric exceed ladder: COLLECTIVE COMMUNICATION (the actual AI-training workload the fabric exists to accelerate). Ladder R4, the last MISSING axis F. SkyHammer is "AI-native ... for GPU-era COLLECTIVE workloads" (fab_collective.raw: All-Reduce/All-Gather/ Broadcast). All-reduce = element-wise sum of N processes' vectors, delivered to ALL. The naive way routes everything through a root -> the root moves O(N)*data = a bottleneck that grows with N. RING all-reduce (NCCL's bandwidth-optimal algorithm) moves a CONSTANT ~2*data per process regardless of N: a reduce-scatter phase (N-1 steps) then an all-gather phase (N-1 steps), each process touching only data/N per step. This rung performs the REAL reduction (integer, no-float) on N in-memory process vectors and PROVES the result is byte-exact the true sum on every process, while measuring the per-process communication (the bottleneck) -- the honest software expression of "collective acceleration at scale". No external organ to compose (pure algorithm). expect_exit: 0 license_tier: ORIGINAL