diff --git a/nemo_rl/models/generation/dynamo/dynamo_generation.py b/nemo_rl/models/generation/dynamo/dynamo_generation.py
index 91bf3d25a32..c65380260c0 100644
--- a/nemo_rl/models/generation/dynamo/dynamo_generation.py
+++ b/nemo_rl/models/generation/dynamo/dynamo_generation.py
@@ -184,7 +184,7 @@ def __init__(
         )
         self._token_wrapper_server: Optional[DynamoTokenWrapperServer] = None
         self._dynamo_frontend_base_url = ""
-        self.dp_openai_server_base_urls: list[Optional[str]] = []
+        self.dp_openai_server_base_urls: list[str] = []
         self._refit_channel: DynamoRefitChannel | None = None
         self._metrics_sampler: DynamoMetricsSampler | None = None
         try:
@@ -217,7 +217,9 @@ def __init__(
                     flush=True,
                 )
             else:
-                self.dp_openai_server_base_urls = [None]
+                # No token wrapper means no OpenAI-compatible server to advertise;
+                # stay empty rather than holding a slot open with a placeholder.
+                self.dp_openai_server_base_urls = []
                 print(f"  [Dynamo] Forwarding rollouts to {url}", flush=True)
 
             if vllm_cfg.enable_vllm_metrics_logger:
diff --git a/nemo_rl/models/generation/megatron/megatron_generation.py b/nemo_rl/models/generation/megatron/megatron_generation.py
index 53294bce379..eb201df4be5 100644
--- a/nemo_rl/models/generation/megatron/megatron_generation.py
+++ b/nemo_rl/models/generation/megatron/megatron_generation.py
@@ -238,7 +238,7 @@ def __init__(
         self._policy_config = config
         self.cfg: MCoreGenerationConfig = config["generation"]
         # Populated after the first prepare_for_generation (which starts the HTTP server).
-        self.dp_openai_server_base_urls: list[Optional[str]] = []
+        self.dp_openai_server_base_urls: list[str] = []
         # Installed by setup via create_weight_synchronizer.
         self.weight_synchronizer: Optional["WeightSynchronizer"] = None
 
diff --git a/nemo_rl/models/generation/trtllm/trtllm_generation.py b/nemo_rl/models/generation/trtllm/trtllm_generation.py
index 422e2099856..2bf396d35f2 100644
--- a/nemo_rl/models/generation/trtllm/trtllm_generation.py
+++ b/nemo_rl/models/generation/trtllm/trtllm_generation.py
@@ -261,16 +261,23 @@ def _get_tied_worker_bundle_indices(
             )
         return tied_groups
 
-    def _report_dp_openai_server_base_urls(self) -> list[Optional[str]]:
+    def _report_dp_openai_server_base_urls(self) -> list[str]:
         """Collect HTTP server base URLs from each DP-rank-0 worker."""
         if not self.cfg["trtllm_cfg"].get("expose_http_server"):
-            urls = [cast(Optional[str], None)] * self.dp_size
-            return urls
+            # Empty rather than one None per DP rank: with no server running there is
+            # no per-rank slot to hold open, and the length said otherwise.
+            return []
         futures = self.worker_group.run_all_workers_single_data(
             "report_dp_openai_server_base_url",
             run_rank_0_only_axes=["tensor_parallel"],
         )
-        return ray.get(futures)
+        results = ray.get(futures)
+        # All-or-nothing, like the guarded branch above: expose_http_server is one
+        # config value for every worker. Collapsing the whole list rather than
+        # filtering keeps one entry per DP rank for anything that reads it by index.
+        if not any(results):
+            return []
+        return results
 
     def _report_device_id(self) -> list[list[str]]:
         futures = self.worker_group.run_all_workers_single_data(
diff --git a/nemo_rl/models/generation/vllm/vllm_generation.py b/nemo_rl/models/generation/vllm/vllm_generation.py
index 382ba1dded5..b3d16cc1bd8 100644
--- a/nemo_rl/models/generation/vllm/vllm_generation.py
+++ b/nemo_rl/models/generation/vllm/vllm_generation.py
@@ -524,10 +524,10 @@ def _report_device_id(self) -> list[list[str]]:
         results = ray.get(futures)
         return results
 
-    def _report_dp_openai_server_base_urls(self) -> list[Optional[str]]:
+    def _report_dp_openai_server_base_urls(self) -> list[str]:
         """Report the data parallel OpenAI server base URLs of vLLM workers, only populated if it is async vLLM engine and the HTTP server is active."""
         if not self.cfg["vllm_cfg"]["async_engine"]:
-            return [None]  # Not applicable since this is sync
+            return []  # Not applicable since this is sync
 
         # Use run_all_workers_single_data for methods that don't need data
         futures = self.worker_group.run_all_workers_single_data(
@@ -536,22 +536,34 @@ def _report_dp_openai_server_base_urls(self) -> list[Optional[str]]:
         )
         # Wait for all futures to complete
         results = ray.get(futures)
+        # A worker leaves base_url unset unless expose_http_server started a server, so
+        # an async engine without one answers with all Nones -- report that as no URLs,
+        # for the same reason the sync branch returns empty. Collapse the whole list
+        # rather than filtering entry by entry: expose_http_server is one config value
+        # for every worker, so the answer is all-or-nothing, and GenerationFleetHealth
+        # indexes this list by shard and length-checks it against shard_count.
+        if not any(results):
+            return []
         return results
 
-    def _collect_reserved_urls(self) -> list[Optional[str]]:
+    def _collect_reserved_urls(self) -> list[str]:
         """Collect reserved URLs from DP leaders before model loading.
 
         Only called when defer_model_load=True. Workers have bound ports
         during __init__ and can report their reserved URLs immediately.
         """
         if not self.cfg["vllm_cfg"]["async_engine"]:
-            return [None]
+            return []
 
         futures = self.worker_group.run_all_workers_single_data(
             "get_reserved_url",
             run_rank_0_only_axes=["tensor_parallel", "pipeline_parallel"],
         )
         results = ray.get(futures)
+        # Only expose_http_server reserves a socket, so without one every worker
+        # reports None here too. Collapsed as a whole, per the note above.
+        if not any(results):
+            return []
         return results
 
     def load_and_start(self) -> None:
diff --git a/tests/unit/models/generation/test_dynamo_generation.py b/tests/unit/models/generation/test_dynamo_generation.py
index 7a53e212f89..c64b3c97ab2 100644
--- a/tests/unit/models/generation/test_dynamo_generation.py
+++ b/tests/unit/models/generation/test_dynamo_generation.py
@@ -164,7 +164,8 @@ def test_runtime_start_world_size_sender_geometry_and_shutdown(monkeypatch) -> N
 
     assert calls[:2] == ["init", "start"]
     assert generation.frontend_url == "http://10.0.0.1:3000/v1"
-    assert generation.dp_openai_server_base_urls == [None]
+    # No token wrapper started, so there is no OpenAI-compatible server to name.
+    assert generation.dp_openai_server_base_urls == []
     assert generation.get_inference_world_size() == 4
     sender = generation.get_collective_sender_spec()
     assert sender.nccl_peer == "vllm"
diff --git a/tests/unit/models/generation/test_vllm_generation.py b/tests/unit/models/generation/test_vllm_generation.py
index 4429f26446e..a1fc6c3cce0 100644
--- a/tests/unit/models/generation/test_vllm_generation.py
+++ b/tests/unit/models/generation/test_vllm_generation.py
@@ -233,6 +233,66 @@ def test_vllm_generation_rejects_partial_refit_pause_and_resume(
         generation.resume_generation_after_refit()
 
 
+def _url_reporting_generation(*, async_engine: bool) -> VllmGeneration:
+    generation = VllmGeneration.__new__(VllmGeneration)
+    generation.cfg = {"vllm_cfg": {"async_engine": async_engine}}
+    generation.worker_group = MagicMock()
+    return generation
+
+
+def test_sync_engine_reports_no_openai_server_urls() -> None:
+    """A sync engine runs no HTTP server, so there is no URL to report."""
+    generation = _url_reporting_generation(async_engine=False)
+
+    assert generation._report_dp_openai_server_base_urls() == []
+    assert generation._collect_reserved_urls() == []
+
+
+def test_async_engine_without_http_server_reports_no_urls(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """Workers answer None unless expose_http_server started a server.
+
+    The list used to keep those Nones, which made a plain truth test read as
+    "we have URLs" when every entry stood for the absence of one.
+    """
+    generation = _url_reporting_generation(async_engine=True)
+    monkeypatch.setattr(ray, "get", MagicMock(return_value=[None, None]))
+
+    assert generation._report_dp_openai_server_base_urls() == []
+    assert generation._collect_reserved_urls() == []
+
+
+def test_async_engine_with_http_server_reports_served_urls(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    generation = _url_reporting_generation(async_engine=True)
+    served = ["http://10.0.0.1:8000/v1", "http://10.0.0.2:8000/v1"]
+    monkeypatch.setattr(ray, "get", MagicMock(return_value=list(served)))
+
+    assert generation._report_dp_openai_server_base_urls() == served
+    assert generation._collect_reserved_urls() == served
+
+
+def test_partially_reported_urls_keep_one_entry_per_rank(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """Report emptiness as a whole list, never by dropping individual entries.
+
+    ``expose_http_server`` is one config value for every worker, so a mixed answer
+    is not reachable today. Asserted anyway because ``GenerationFleetHealth``
+    indexes these URLs by shard and rejects a list whose length does not match
+    ``shard_count``: shortening one would turn a degraded shard into a startup
+    failure with a length error that names neither cause.
+    """
+    generation = _url_reporting_generation(async_engine=True)
+    partial = ["http://10.0.0.1:8000/v1", None]
+    monkeypatch.setattr(ray, "get", MagicMock(return_value=list(partial)))
+
+    assert generation._report_dp_openai_server_base_urls() == partial
+    assert generation._collect_reserved_urls() == partial
+
+
 def test_sampling_params_preserve_bad_words():
     worker = object.__new__(VllmGenerationWorkerImpl)
     worker.cfg = {
diff --git a/tests/unit/models/generation/trtllm/test_trtllm_generation.py b/tests/unit/models/generation/trtllm/test_trtllm_generation.py
index 1123ebae58a..a91ef3ce0b9 100644
--- a/tests/unit/models/generation/trtllm/test_trtllm_generation.py
+++ b/tests/unit/models/generation/trtllm/test_trtllm_generation.py
@@ -170,6 +170,27 @@ def test_cross_node_tp_replicas_use_unified_placement_group():
     ]
 
 
+def test_no_http_server_reports_an_empty_url_list():
+    """Without a server there is no per-rank slot, so the length said nothing.
+
+    This used to be one None per DP rank, which is truthy and carries a length
+    that reads like real URLs were collected.
+    """
+    generation = _bare_generation(dp_size=2)
+    generation.dp_size = 2
+
+    assert generation._report_dp_openai_server_base_urls() == []
+
+
+def test_exposed_http_server_reports_served_addresses(monkeypatch):
+    generation = _bare_generation(dp_size=2, expose_http_server=True)
+    generation.dp_size = 2
+    served = ["http://10.0.0.1:8000/v1", "http://10.0.0.2:8000/v1"]
+    monkeypatch.setattr(trtllm_generation.ray, "get", lambda futures: list(served))
+
+    assert generation._report_dp_openai_server_base_urls() == served
+
+
 @pytest.mark.asyncio
 async def test_generate_async_dispatches_round_robin_and_returns_leader_index():
     generation = _bare_generation(dp_size=2)