diff --git a/components/src/dynamo/trtllm/tests/test_trtllm_autodeploy.py b/components/src/dynamo/trtllm/tests/test_trtllm_autodeploy.py
index 3394a912e2d3..fa4785a86d7b 100644
--- a/components/src/dynamo/trtllm/tests/test_trtllm_autodeploy.py
+++ b/components/src/dynamo/trtllm/tests/test_trtllm_autodeploy.py
@@ -28,8 +28,11 @@
     # the `gpu_0` marker.
     pytest.mark.gpu_1,
     pytest.mark.pre_merge,
-    pytest.mark.profiled_vram_gib(0),
 ]
+
+# Intentionally unprofiled: these import-heavy, zero-VRAM tests run in the
+# sequential GPU stage so TensorRT-LLM initialization is shared.
+
 _PYTORCH_LLM_CLS_NAME = "dynamo.trtllm.engine.LLM"
 _AUTODEPLOY_LLM_CLS_NAME = "tensorrt_llm._torch.auto_deploy.LLM"
 
diff --git a/components/src/dynamo/trtllm/tests/test_trtllm_image_token_resolver.py b/components/src/dynamo/trtllm/tests/test_trtllm_image_token_resolver.py
index fca91ea24b8b..0fe58f45aeee 100644
--- a/components/src/dynamo/trtllm/tests/test_trtllm_image_token_resolver.py
+++ b/components/src/dynamo/trtllm/tests/test_trtllm_image_token_resolver.py
@@ -27,9 +27,11 @@
     pytest.mark.multimodal,
     pytest.mark.unit,
     pytest.mark.gpu_1,
-    pytest.mark.profiled_vram_gib(0),
 ]
 
+# Intentionally unprofiled: these import-heavy, zero-VRAM tests run in the
+# sequential GPU stage so TensorRT-LLM initialization is shared.
+
 
 def _config(model: str = "org/model") -> SimpleNamespace:
     return SimpleNamespace(model=model, revision=None)
diff --git a/components/src/dynamo/trtllm/tests/test_trtllm_logits_runtime.py b/components/src/dynamo/trtllm/tests/test_trtllm_logits_runtime.py
index c87b008539e4..2c423860da46 100644
--- a/components/src/dynamo/trtllm/tests/test_trtllm_logits_runtime.py
+++ b/components/src/dynamo/trtllm/tests/test_trtllm_logits_runtime.py
@@ -38,9 +38,11 @@
     pytest.mark.trtllm,
     pytest.mark.pre_merge,
     pytest.mark.gpu_1,
-    pytest.mark.profiled_vram_gib(0),
 ]
 
+# Intentionally unprofiled: the zero-VRAM tests run in the sequential GPU stage
+# so TensorRT-LLM initialization is shared. The CUDA test overrides this below.
+
 
 # ---------------------------------------------------------------------------
 # TRT-LLM attach_logits_processors contract
@@ -157,10 +159,9 @@ def test_adapter_invokes_or_logs_on_bad_shape(shape, expect_invoke, caplog):
         )
 
 
-# Unlike the rest of this module (CPU-only mocks, module-level
-# profiled_vram_gib(0)), this test initializes a real CUDA context, so the
-# GPU-parallel scheduler must reserve VRAM for it instead of packing it onto
-# an already-full GPU as a zero-VRAM filler.
+# Unlike the rest of this module (unprofiled CPU-only mocks), this test
+# initializes a real CUDA context, so the GPU-parallel scheduler must reserve
+# VRAM for it.
 @pytest.mark.profiled_vram_gib(2.0)
 @pytest.mark.requested_trtllm_vram_gib(2.0)
 def test_adapter_enters_engine_cuda_stream():
diff --git a/components/src/dynamo/trtllm/tests/test_trtllm_main_init.py b/components/src/dynamo/trtllm/tests/test_trtllm_main_init.py
index aba992af1f5c..4e7961610461 100644
--- a/components/src/dynamo/trtllm/tests/test_trtllm_main_init.py
+++ b/components/src/dynamo/trtllm/tests/test_trtllm_main_init.py
@@ -15,9 +15,11 @@
     pytest.mark.trtllm,
     pytest.mark.gpu_1,
     pytest.mark.pre_merge,
-    pytest.mark.profiled_vram_gib(0),
 ]
 
+# Intentionally unprofiled: these import-heavy, zero-VRAM tests run in the
+# sequential GPU stage so TensorRT-LLM initialization is shared.
+
 
 def test_tensorrt_llm_metrics_collector_import():
     """Test that we can import MetricsCollector from TensorRT-LLM."""
diff --git a/components/src/dynamo/trtllm/tests/test_trtllm_multimodal_processor.py b/components/src/dynamo/trtllm/tests/test_trtllm_multimodal_processor.py
index 450009d62e5a..88580251eabd 100644
--- a/components/src/dynamo/trtllm/tests/test_trtllm_multimodal_processor.py
+++ b/components/src/dynamo/trtllm/tests/test_trtllm_multimodal_processor.py
@@ -29,9 +29,11 @@
     pytest.mark.multimodal,
     pytest.mark.pre_merge,
     pytest.mark.gpu_1,
-    pytest.mark.profiled_vram_gib(0),
 ]
 
+# Intentionally unprofiled: these import-heavy, zero-VRAM tests run in the
+# sequential GPU stage so TensorRT-LLM initialization is shared.
+
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize(
diff --git a/components/src/dynamo/trtllm/tests/test_trtllm_unit.py b/components/src/dynamo/trtllm/tests/test_trtllm_unit.py
index bbfbb79bbc57..02891813d6b0 100644
--- a/components/src/dynamo/trtllm/tests/test_trtllm_unit.py
+++ b/components/src/dynamo/trtllm/tests/test_trtllm_unit.py
@@ -42,9 +42,11 @@
     pytest.mark.trtllm,
     pytest.mark.gpu_1,
     pytest.mark.pre_merge,
-    pytest.mark.profiled_vram_gib(0),
 ]
 
+# Intentionally unprofiled: these import-heavy, zero-VRAM tests run in the
+# sequential GPU stage so TensorRT-LLM initialization is shared.
+
 
 # Create TRTLLM-specific CLI args fixture
 # This will use monkeypatch to write to argv
diff --git a/examples/backends/trtllm/launch/disagg_multimodal.sh b/examples/backends/trtllm/launch/disagg_multimodal.sh
index f112c0b511aa..bf5dfb9f63f1 100755
--- a/examples/backends/trtllm/launch/disagg_multimodal.sh
+++ b/examples/backends/trtllm/launch/disagg_multimodal.sh
@@ -28,11 +28,16 @@ fi
 HTTP_PORT="${DYN_HTTP_PORT:-8000}"
 print_launch_banner --multimodal "Launching Disaggregated Multimodal Serving (2 GPUs)" "$MODEL_PATH" "$HTTP_PORT"
 
+# Multi-worker launchers use numbered status ports; do not pass the single-worker
+# alias to the frontend.
+unset DYN_SYSTEM_PORT
+
 # run frontend
 # dynamo.frontend accepts either --http-port flag or DYN_HTTP_PORT env var (defaults to 8000)
 python3 -m dynamo.frontend &
 
 # run prefill worker
+DYN_SYSTEM_PORT=${DYN_SYSTEM_PORT1:-8081} \
 CUDA_VISIBLE_DEVICES=$PREFILL_CUDA_VISIBLE_DEVICES python3 -m dynamo.trtllm \
   --model-path "$MODEL_PATH" \
   --served-model-name "$SERVED_MODEL_NAME" \
@@ -42,6 +47,7 @@ CUDA_VISIBLE_DEVICES=$PREFILL_CUDA_VISIBLE_DEVICES python3 -m dynamo.trtllm \
   --disaggregation-mode prefill &
 
 # run decode worker
+DYN_SYSTEM_PORT=${DYN_SYSTEM_PORT2:-8082} \
 CUDA_VISIBLE_DEVICES=$DECODE_CUDA_VISIBLE_DEVICES python3 -m dynamo.trtllm \
   --model-path "$MODEL_PATH" \
   --served-model-name "$SERVED_MODEL_NAME" \
diff --git a/tests/kvbm_integration/test_consolidator_config_unit.py b/tests/kvbm_integration/test_consolidator_config_unit.py
index 17fd4507f944..27911a04c041 100644
--- a/tests/kvbm_integration/test_consolidator_config_unit.py
+++ b/tests/kvbm_integration/test_consolidator_config_unit.py
@@ -49,12 +49,13 @@ def test_env_var_disables(self):
             assert should_enable_consolidator(arg_map) is False
 
 
+# Intentionally unprofiled: these import-heavy, zero-VRAM tests run in the
+# sequential GPU stage so TensorRT-LLM initialization is shared.
 @pytest.mark.unit
 @pytest.mark.pre_merge
 @pytest.mark.kvbm
 @pytest.mark.trtllm
 @pytest.mark.gpu_1
-@pytest.mark.profiled_vram_gib(0)
 class TestShouldEnableConsolidatorTyped:
     """Tests that need trtllm (requires GPU for import)."""
 
diff --git a/tests/serve/test_trtllm.py b/tests/serve/test_trtllm.py
index 808e17b9db2a..4f8c2aa0fb1a 100644
--- a/tests/serve/test_trtllm.py
+++ b/tests/serve/test_trtllm.py
@@ -352,11 +352,12 @@ class TRTLLMConfig(EngineConfig):
             pytest.mark.pre_merge,
             pytest.mark.profiled_vram_gib(15.0),
             pytest.mark.requested_trtllm_kv_tokens(1056),
+            pytest.mark.timeout(360),  # 3x measured 118s CI runtime
         ],
         model="Qwen/Qwen3-VL-2B-Instruct",
         frontend_port=DefaultPort.FRONTEND.value,
-        timeout=900,
-        delayed_start=120,
+        timeout=300,
+        health_check_workers=True,
         request_payloads=[
             multimodal_payload_default(
                 text="Describe what you see in this image.",
@@ -366,6 +367,9 @@ class TRTLLMConfig(EngineConfig):
         env={
             "PREFILL_CUDA_VISIBLE_DEVICES": "0",
             "DECODE_CUDA_VISIBLE_DEVICES": "0",
+            # Make worker /health readiness depend on a successful one-token
+            # engine canary instead of the system-status server alone.
+            "DYN_HEALTH_CHECK_ENABLED": "true",
         },
     ),
     "e_pd_multimodal": TRTLLMConfig(
diff --git a/tests/serve/test_trtllm_mm_hashes_protocol.py b/tests/serve/test_trtllm_mm_hashes_protocol.py
index ffb70b8c3a86..2779eb08ba21 100644
--- a/tests/serve/test_trtllm_mm_hashes_protocol.py
+++ b/tests/serve/test_trtllm_mm_hashes_protocol.py
@@ -26,9 +26,11 @@
     pytest.mark.multimodal,
     pytest.mark.unit,
     pytest.mark.gpu_1,
-    pytest.mark.profiled_vram_gib(0),
 ]
 
+# Intentionally unprofiled: these import-heavy, zero-VRAM tests run in the
+# sequential GPU stage so TensorRT-LLM initialization is shared.
+
 
 def test_trtllm_qwen2vl_uses_in_vocab_image_markers() -> None:
     from tensorrt_llm._torch.models.modeling_qwen2vl import Qwen2VLInputProcessorBase