engine: report COMBINED-static layer-split capability truthfully

[ddvnguyen]

Overview

Follow-up to the spike-ggml-org#381 reconciliation on #31 (the "PDL crash" was a misdiagnosis; PR #32 closed as inert). This is the first, safe step toward driving dense-model COMBINE through the layer-split path (ggml-org#383) instead of the runtime expert path.

The unified engine's model-loaded path hard-coded:

ctx_server.set_hydra_capabilities(true, flags.rpc_engine_peer, false, "", "solo");

even when launched with --rpc-engine PEER --tensor-split R0/R1 — i.e. the peer was registered as an RPC device before load_model (llama-engine.cpp:311-322) and llama.cpp's stock allocator placed whole layers across CUDA0 + RPC0 (COMBINED-static layer-split). So /health + INFO reported the engine as SOLO with an unreachable peer while it was actually layer-splitting onto that peer.

Change

Report the real startup state:

• peer_reachable now reflects whether the peer actually registered (was always false — a reporting bug).
• split_mode="layer" when a peer registered before load with a tensor-split, so Hydra Core can route to the layer-split engine instead of the runtime expert path (the add-RPC-after-load path behind the 35-graph-split regression + recurrent-state corruption on dense/SSM models).
• combined_pattern kept empty on purpose: it is the expert regex consumed by SET_EXPERT_MODE("combined") → llama_hydra_rebind_combined_experts (server-context.cpp:3651). Putting the tensor-split ratio there would trigger the after-load expert re-partition that layer-split exists to avoid.

Observability/routing only — no compute-path change.

Additional information

• Context and evidence: fork: resolve #376 — make COMBINE mode RPC work (gate-then-degrade) #31 (comment reconciling with hydra_vortex#381) and cuda: guard PDL can-use check on pre-Hopper devices (fix sm_86 crash) #32 (closed).
• Deferred to a build-capable session (needs the 2-GPU repro, flagged high-blast-radius by the author): the graph_compute degrade-to-solo fix (ggml_backend_rpc_graph_compute → GGML_STATUS_FAILED instead of GGML_ABORT, peer serve-loop break not return). It interacts with the fire-and-forget GRAPH_COMPUTE protocol and a client-only change is incomplete (the next RPC_STATUS_ASSERT still aborts), so it should be done + hardware-verified together. Tracked as hydra_vortex#395 Step 4.
• Companion (separate PR, hydra_vortex repo): route the dense profile to launch the engine with --rpc-engine PEER --tensor-split 21/44 (spike Doesn't compile due to missing headers for memcpy and assert ggml-org/llama.cpp#381's MTP-headroom winner) and stop sending expert-mode COMBINED requests for it.
• ⚠️ Not compiled in this environment (no CUDA / 2-GPU host). Needs cmake --build build_sm86_sm120 --target llama-engine, then confirm /health reports split_mode:"layer" and peer_reachable:true under the spike-Doesn't compile due to missing headers for memcpy and assert ggml-org/llama.cpp#381 config.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES — change drafted by an AI agent (Claude Code) from the fork: resolve #376 — make COMBINE mode RPC work (gate-then-degrade) #31 / spike-Doesn't compile due to missing headers for memcpy and assert ggml-org/llama.cpp#381 analysis; the author is responsible for reviewing, building, and hardware-verifying before merge. This project restricts AI-generated content — see AGENTS.md and CONTRIBUTING.md.

---

Generated by Claude Code