fix(kv-router): migrate raw zmq paths to libzmq and refresh lockfiles (#7871)

Signed-off-by: PeaBrane <yanrpei@gmail.com>

lib/runtime/Cargo.toml

@@ -29,8 +29,7 @@ anyhow = { workspace = true }
 async-nats = { workspace = true }
 async-stream = { workspace = true }
 async-trait = { workspace = true }
-async_zmq = { workspace = true }
-zmq = { workspace = true }
+tmq = { workspace = true }
 lru = { version = "0.12" }
 axum = { workspace = true }
 blake3 = { workspace = true }

tests/router/common.py

@@ -871,14 +871,14 @@ def _test_router_indexers_sync(
             router_event_threads=router_event_threads,
         )
 
-        # If standalone indexer mode, launch mockers one-by-one and register.
+        # If standalone indexer mode, launch workers one-by-one and register.
         # We need to create a temporary endpoint just to discover worker IDs.
         if standalone_indexer_url:
             tmp_runtime = get_runtime(store_backend, request_plane)
             tmp_endpoint = tmp_runtime.endpoint(
                 f"{engine_workers.namespace}.{engine_workers.component_name}.generate"
             )
-            await engine_workers.launch_mockers_with_indexer(tmp_endpoint)
+            await engine_workers.launch_workers_with_indexer(tmp_endpoint)
 
         async def send_requests_to_router(router, num_requests, router_name, endpoint):
             # Now send the actual requests
@@ -1511,10 +1511,10 @@ def _test_router_decisions(
     # Create KvRouterConfig with lower snapshot threshold for testing
     # Use async to manage the test flow
     async def test_sync():
-        # If standalone indexer mode, launch mockers one-by-one and register.
+        # If standalone indexer mode, launch workers one-by-one and register.
         # Must happen before KvRouter creation since KvRouter blocks until workers appear.
         if standalone_indexer_url:
-            await engine_workers.launch_mockers_with_indexer(endpoint)
+            await engine_workers.launch_workers_with_indexer(endpoint)
 
         # Workers register one instance per process (not per dp_rank)
         expected_num_instances = engine_workers.num_workers

tests/router/e2e_harness.py

@@ -242,8 +242,10 @@ def run_indexers_sync_test(
     block_size: int,
     model_name: str,
     num_workers: int,
+    extra_process_kwargs: dict[str, Any] | None = None,
 ):
     nats_process, _etcd_process = runtime_services_dynamic_ports
+    process_kwargs = extra_process_kwargs or {}
 
     with engine_process_cls(
         request,
@@ -253,6 +255,7 @@ def run_indexers_sync_test(
         store_backend=store_backend,
         durable_kv_events=durable_kv_events,
         **{engine_args_name: engine_args},
+        **process_kwargs,
     ) as engine_workers:
         _test_router_indexers_sync(
             engine_workers=engine_workers,
@@ -264,4 +267,13 @@ def run_indexers_sync_test(
             test_nats_interruption=not durable_kv_events,
             nats_server=nats_process if not durable_kv_events else None,
             durable_kv_events=durable_kv_events,
+            standalone_indexer_url=getattr(
+                engine_workers, "standalone_indexer_url", None
+            ),
+            standalone_indexer_b_url=getattr(
+                engine_workers, "standalone_indexer_b_url", None
+            ),
+            test_zmq_replay=bool(
+                getattr(engine_workers, "standalone_indexer_url", None)
+            ),
         )