The deepest structural test gap: DistributedGemma4Model (and every distributed client model) requires a DHT + live servers, so the client forward path is untestable in offline CI. Both Gemma 4 live-run bug rounds funneled through exactly that hole -- blocks were bit-exact in isolation while the client<->server composition was unexercised.
Proposal: a loopback transport stub that satisfies the RemoteSequenceManager / InferenceSession interfaces but executes real WrappedBlocks in-process (no hivemind, no network, no processes). CI could then run true end-to-end client-vs-stock greedy comparisons for every arch on the tiny fixtures -- including the per-layer-input and shared-KV side channels, span slicing, and (with a fault-injection hook) retry paths.
🤖 Generated with Claude Code
The deepest structural test gap:
DistributedGemma4Model(and every distributed client model) requires a DHT + live servers, so the client forward path is untestable in offline CI. Both Gemma 4 live-run bug rounds funneled through exactly that hole -- blocks were bit-exact in isolation while the client<->server composition was unexercised.Proposal: a loopback transport stub that satisfies the
RemoteSequenceManager/InferenceSessioninterfaces but executes realWrappedBlocks in-process (no hivemind, no network, no processes). CI could then run true end-to-end client-vs-stock greedy comparisons for every arch on the tiny fixtures -- including the per-layer-input and shared-KV side channels, span slicing, and (with a fault-injection hook) retry paths.🤖 Generated with Claude Code