[Misc][Metrics] expose requests preemptions in logger

tests/v1/metrics/test_engine_logger_apis.py

@@ -5,7 +5,9 @@
 import pytest
 
 from vllm.v1.engine.async_llm import AsyncEngineArgs, AsyncLLM
+from vllm.v1.metrics.loggers import LoggingStatLogger
 from vllm.v1.metrics.ray_wrappers import RayPrometheusStatLogger
+from vllm.v1.metrics.stats import IterationStats
 
 
 class DummyStatLogger:
@@ -31,6 +33,16 @@ def log_engine_initialized(self):
         self.engine_initialized = True
 
 
+class DummyLoggingStatLogger(LoggingStatLogger):
+    """
+    A dummy logging stat logger for testing purposes.
+    Implemented the record and log APIs
+    """
+
+    def get_num_preempted_reqs(self) -> int:
+        return self.num_preempted_reqs
+
+
 @pytest.fixture
 def log_stats_enabled_engine_args():
     """
@@ -62,7 +74,7 @@ async def test_async_llm_replace_default_loggers(
 @pytest.mark.asyncio
 async def test_async_llm_add_to_default_loggers(log_stats_enabled_engine_args):
     """
-    It's still possible to use custom stat loggers exclusively by passing 
+    It's still possible to use custom stat loggers exclusively by passing
     disable_log_stats=True in addition to a list of custom stat loggers.
     """
     # Create engine_args with disable_log_stats=True for this test
@@ -81,3 +93,41 @@ async def test_async_llm_add_to_default_loggers(log_stats_enabled_engine_args):
     assert engine.log_stats
 
     engine.shutdown()
+
+
+@pytest.mark.asyncio
+async def test_logger_iteration_stats(log_stats_enabled_engine_args):
+    """
+    """
+    # Create engine_args with disable_log_stats=True for this test
+    disabled_log_engine_args = copy.deepcopy(log_stats_enabled_engine_args)
+    disabled_log_engine_args.disable_log_stats = True
+
+    # Disable default loggers; pass custom stat logger to the constructor
+    engine = AsyncLLM.from_engine_args(disabled_log_engine_args,
+                                       stat_loggers=[DummyLoggingStatLogger])
+
+    dummy_logger = engine.logger_manager.per_engine_logger_dict[0][0]
+
+    assert len(engine.logger_manager.per_engine_logger_dict[0]) == 1
+    assert isinstance(dummy_logger, DummyLoggingStatLogger)
+
+    stats_1 = IterationStats()
+    stats_1.num_preempted_reqs = 1
+    stats_1.num_generation_tokens = 10
+    stats_1.num_prompt_tokens = 100
+
+    stats_2 = IterationStats()
+    stats_2.num_preempted_reqs = 2
+    stats_2.num_generation_tokens = 20
+    stats_2.num_prompt_tokens = 200
+
+    # Expect the record will update the local iteration stats correctly
+    dummy_logger.record(scheduler_stats=None, iteration_stats=stats_1)
+    dummy_logger.record(scheduler_stats=None, iteration_stats=stats_2)
+
+    assert dummy_logger.num_preempted_reqs == 3
+    assert dummy_logger.num_generation_tokens == 30
+    assert dummy_logger.num_prompt_tokens == 300
+
+    engine.shutdown()

vllm/v1/metrics/loggers.py

@@ -72,11 +72,13 @@ def _reset(self, now):
         # Tracked stats over current local logging interval.
         self.num_prompt_tokens: int = 0
         self.num_generation_tokens: int = 0
+        self.num_preempted_reqs: int = 0
 
     def _track_iteration_stats(self, iteration_stats: IterationStats):
         # Save tracked stats for token counters.
         self.num_prompt_tokens += iteration_stats.num_prompt_tokens
         self.num_generation_tokens += iteration_stats.num_generation_tokens
+        self.num_preempted_reqs += iteration_stats.num_preempted_reqs
 
     def _get_throughput(self, tracked_stats: int, now: float) -> float:
         # Compute summary metrics for tracked stats