Epic hydra_vortex#470. After PR #98 (RPC stale-graph fix), the engine hangs in common_params_fit_impl 'getting device memory data for initial parameters' when an RPC peer is registered — never reaches load_tensors. 2 independent trials with a healthy peer. Old engine (60bde13, pre-#98) completes fit in ~30s with the same peer. Suspect: the fit path enumerates peer device memory via the RPC layer and the new code deadlocks there. Reproduce: run llama-engine with --rpc-engine and observe fit hang.
Epic hydra_vortex#470. After PR #98 (RPC stale-graph fix), the engine hangs in common_params_fit_impl 'getting device memory data for initial parameters' when an RPC peer is registered — never reaches load_tensors. 2 independent trials with a healthy peer. Old engine (60bde13, pre-#98) completes fit in ~30s with the same peer. Suspect: the fit path enumerates peer device memory via the RPC layer and the new code deadlocks there. Reproduce: run llama-engine with --rpc-engine and observe fit hang.