Fix first-token metadata and reused attention-layer indexing (#39328)

Co-authored-by: Jinghui Zhang <jinghui@meta.com>
Co-authored-by: Lucia Fang <116399278+luccafong@users.noreply.github.com>

test/registered/unit/model_executor/model_runner_components/test_cuda_graph_setup.py

@@ -48,6 +48,25 @@ def test_pipeline_attention_metadata_is_indexed_by_global_layer_id():
     assert companions[24] is companion24
 
 
+def test_reuse_tables_pass_through_but_distinct_duplicates_raise():
+    looped = SimpleNamespace(layer_id=1)
+    companion = object()
+    attention_in = [SimpleNamespace(layer_id=0), looped, looped]
+    companions_in = [None, companion, companion]
+
+    attention, companions = index_attention_layers_by_global_id(
+        attention_in, companions_in
+    )
+
+    assert attention is attention_in
+    assert companions is companions_in
+
+    with pytest.raises(ValueError, match="duplicate attention layer_id: 2"):
+        index_attention_layers_by_global_id(
+            [SimpleNamespace(layer_id=2), SimpleNamespace(layer_id=2)], [None, None]
+        )
+
+
 def test_model_runner_can_override_decode_graph_runner(monkeypatch):
     from sglang.srt.runtime_context import get_context
 

test/registered/unit/observability/test_req_time_stats.py

@@ -46,6 +46,26 @@ class TestSetstatePreservesUnsetTimeSentinels(CustomTestCase):
         self.assertAlmostEqual(hop2.wait_queue_entry_time, 123.456 - 9.0)
 
 
+class TestOutputMetaInfo(CustomTestCase):
+    def test_first_token_latency_requires_valid_timestamps(self):
+        for created, first, expected in (
+            (1.0, 1.5, 0.5),
+            (0.0, 1.5, None),
+            (1.0, 0.0, None),
+            (1.0, 1.0, None),
+            (1.0, 0.5, None),
+        ):
+            with self.subTest(created=created, first=first):
+                stats = rts.APIServerReqTimeStats()
+                stats.created_time = created
+                stats.first_token_time = first
+                meta_info = stats.convert_to_output_meta_info()
+                if expected is None:
+                    self.assertNotIn("first_token_latency", meta_info)
+                else:
+                    self.assertAlmostEqual(meta_info["first_token_latency"], expected)
+
+
 class TestConvertToGenAiSpanAttrs(CustomTestCase):
     def _stats_after_first_token(self) -> rts.APIServerReqTimeStats:
         stats = rts.APIServerReqTimeStats()