[Bugfix] Fix speculative decode seeded test (#6743)

tests/spec_decode/e2e/conftest.py

@@ -191,7 +191,8 @@ def create_llm_generator(baseline_or_test, request, common_llm_kwargs,
                 and llm.llm_engine.log_stats):
             for sate_logger in llm.llm_engine.stat_loggers.values():
                 sate_logger.local_interval = 0
-        set_random_seed(seed)
+        if seed is not None:
+            set_random_seed(seed)
 
         yield llm
         del llm

tests/spec_decode/e2e/test_seed.py

@@ -21,7 +21,8 @@ from .conftest import run_equality_correctness_test
         "num_speculative_tokens": 3,
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
-@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
+@pytest.mark.parametrize("baseline_llm_kwargs", [{"seed": 1}])
+@pytest.mark.parametrize("test_llm_kwargs", [{"seed": 5}])
 @pytest.mark.parametrize("batch_size", [1, 8, 32])
 @pytest.mark.parametrize("temperature", [0.1, 1.0])
 @pytest.mark.parametrize(
@@ -30,15 +31,26 @@ from .conftest import run_equality_correctness_test
         # Use smaller output len for fast test.
         10,
     ])
-@pytest.mark.parametrize("seed", [1])
-def test_seeded_consistency(baseline_llm_generator, batch_size: int,
-                            temperature: float, output_len: int):
+@pytest.mark.parametrize("seed", [None])
+def test_seeded_consistency(baseline_llm_generator, test_llm_generator,
+                            batch_size: int, temperature: float,
+                            output_len: int):
     """Verify outputs are consistent across multiple runs with same seed
     """
     run_equality_correctness_test(baseline_llm_generator,
-                                  baseline_llm_generator,
+                                  test_llm_generator,
                                   batch_size,
                                   max_output_len=output_len,
                                   temperature=temperature,
                                   seeded=True,
                                   force_output_len=True)
+
+    # Ensure this same test does fail if we _don't_ include per-request seeds
+    with pytest.raises(AssertionError):
+        run_equality_correctness_test(baseline_llm_generator,
+                                      test_llm_generator,
+                                      batch_size,
+                                      max_output_len=output_len,
+                                      temperature=temperature,
+                                      seeded=False,
+                                      force_output_len=True)