nx_batch_infer.nx

nx_batch_infer.nx -- WORKER MESH BACKEND: a sovereign FUSED-BATCH inference kernel that gives the real throughput win sd.cpp cannot (it runs n>1 sequentially). The core of any forward pass is C = A x W (a batch of A-rows through a shared weight matrix W). The whole point of batching is WEIGHT REUSE: load each W element ONCE and apply it to the WHOLE batch. Per-request serving reloads W for every request; a fused batch streams W once. This proves it MECHANICALLY + MEASURED: * fused (weight-stationary: for k,for j: w=W[k][j]; for b: C[b][j]+=A[b][k]*w) -> W streamed 1x * seq (per-request: for b: for j: for k: C[b][j]+=A[b][k]*W[k][j]) -> W streamed B times Both compute the IDENTICAL result (integer, bit-exact) -- the ONLY difference is memory traffic on W. With W larger than cache the fused schedule is dramatically faster: THAT is the batching-capable-backend win, on our own stack, deterministic. Timed with the monotonic clock; the gate refuses to pass unless results are bit-exact AND fused is at least as fast. license_tier: ORIGINAL