silvic-net

TcpCollective: the collectives distributed silvic training runs on (allreduce_sum_f64, allreduce_sum_i64, allreduce_max_f64, allgather_bytes), over a TCP ring. silvic.ray.train creates one per Ray actor.

Each rank binds a listener (TcpCollective::bind, or bind_advertised to listen on 0.0.0.0 and advertise the node IP), the addresses are exchanged out of band, and TcpCollective::connect joins the ring: every rank connects to the next and accepts from the previous, with a handshake carrying a job token, rank and world size that rejects stray or mismatched peers. Allreduces are a ring reduce-scatter and allgather (an allgather below 4 KiB); every chunk folds in a fixed rank order, so results are deterministic for a world size, and exact in any order for the integer-valued histograms quantized training reduces.

Sockets are non-blocking and polled, so full-duplex ring steps never deadlock on large buffers. Every frame carries the operation, a sequence number and the length: ranks that call different collectives fail with an "out of step" error rather than exchanging garbage. A socket error, a closed peer or a timeout fails the collective naming the peer rank, shuts the rank's sockets so the failure travels around the ring, and fails every later call.

Tests

tests/ring.rs runs world sizes 1, 2, 3 and 5 on threads over 127.0.0.1: sums, maxima and gathers against direct computation for lengths 0, 1, world - 1, both sides of the 4 KiB switch and 1,000,003; bit-identical repeats; and the failure paths (an early exit, a hung rank, a wrong token, stray connections, diverging calls).