ci(mm-router): parallel worker boot + pre_merge gater for MM routing

examples/backends/vllm/launch/agg_multimodal_router.sh

@@ -131,6 +131,7 @@ COMMON_ENV=(
 
 GPU_MEM_ARGS=$(build_vllm_gpu_mem_args)
 
+# Phase 1: launch all workers in parallel.
 for i in $(seq 1 "${NUM_WORKERS}"); do
     WORKER_PORT=$((VLLM_SYSTEM_PORT_BASE + (i - 1) * 2))
     KV_EVENTS_PORT=$((KV_EVENTS_PORT_BASE + (i - 1)))
@@ -150,6 +151,11 @@ for i in $(seq 1 "${NUM_WORKERS}"); do
         --max-model-len "${MAX_MODEL_LEN}" \
         --kv-events-config "${KV_EVENTS_CONFIG}" \
         ${GPU_MEM_ARGS} ${VLLM_EXTRA_ARGS} "${PASSTHRU_ARGS[@]}" &
+done
+
+# Phase 2: wait for all workers to be ready.
+for i in $(seq 1 "${NUM_WORKERS}"); do
+    WORKER_PORT=$((VLLM_SYSTEM_PORT_BASE + (i - 1) * 2))
     wait_ready "http://127.0.0.1:${WORKER_PORT}/health" "vLLM backend $i"
 done
 

tests/serve/multimodal_profiles/vllm.py

@@ -100,21 +100,11 @@
                 requested_vllm_kv_cache_bytes=1_719_075_000,
                 tests=[MmCase(payload=make_video_payload(["red", "static", "still"]))],
             ),
-            # `agg_router` exercises agg_multimodal_router.sh: Rust frontend
-            # with the `lightseek-mm` feature, MM-aware KV routing, multi-worker.
-            # Smoke-level on post_merge so regressions to the script's plumbing
-            # (worker boot order, ZMQ KV events, MM-routing build) surface in
-            # CI. The fine-grained routing-correctness assertions live in
-            # tests/mm_router/test_router_rust_mm_router_e2e.py.
-            #
-            # The payload sends two identical MM requests and asserts the
-            # second sees cached_tokens > 0 — proves the warm worker reused
-            # its KV cache, which only happens if the router routed both
-            # requests to the same worker. If routing silently regressed to
-            # text-prefix only, both requests would still succeed but the
-            # second's cached_tokens would be 0 and this case would fail.
+            # Pre_merge gater for the lightseek MM-routing path. Fine-grained
+            # assertions live in tests/mm_router/test_router_rust_mm_router_e2e.py
+            # (post_merge).
             "agg_router": TopologyConfig(
-                marks=[pytest.mark.post_merge],
+                marks=[pytest.mark.pre_merge],
                 timeout_s=400,
                 profiled_vram_gib=18.7,
                 requested_vllm_kv_cache_bytes=1_719_075_000,

tests/utils/multimodal.py

@@ -96,7 +96,7 @@ def make_image_payload_cached_tokens(
                     ],
                 }
             ],
-            "max_tokens": 100,
+            "max_tokens": 50,
             "temperature": 0.0,
             "stream": False,
         },