perf: vLLM metrics optimization (#66)

Author: yinggeh

src/model.py

@@ -161,6 +161,7 @@ def init_engine(self):
         self.llm_engine = AsyncLLMEngine.from_engine_args(aync_engine_args)
 
         # Create vLLM custom metrics
+        self.vllm_metrics = None
         if (
             "REPORT_CUSTOM_METRICS" in self.model_config["parameters"]
             and self.model_config["parameters"]["REPORT_CUSTOM_METRICS"]["string_value"]
@@ -174,9 +175,10 @@ def init_engine(self):
                 }
                 # Add vLLM custom metrics
                 engine_config = self.llm_engine.engine.model_config
-                self.llm_engine.add_logger(
-                    "triton", VllmStatLogger(labels, engine_config.max_model_len)
+                self.vllm_metrics = VllmStatLogger(
+                    labels, engine_config.max_model_len, self.logger
                 )
+                self.llm_engine.add_logger("triton", self.vllm_metrics)
             except pb_utils.TritonModelException as e:
                 if "metrics not supported" in str(e):
                     # Metrics are disabled at the server
@@ -572,6 +574,10 @@ def finalize(self):
             self._response_thread.join()
             self._response_thread = None
 
+        # Shutdown the logger thread.
+        if self.vllm_metrics is not None:
+            self.vllm_metrics.finalize()
+
         # When using parallel tensors, the stub process may not shutdown due to
         # unreleased references, so manually run the garbage collector once.
         self.logger.log_info("[vllm] Running Garbage Collector on finalize...")

src/utils/metrics.py

@@ -24,6 +24,8 @@
 # (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
+import queue
+import threading
 from typing import Dict, List, Union
 
 import triton_python_backend_utils as pb_utils
@@ -170,11 +172,18 @@ def __init__(self, labels: List[str], max_model_len: int):
 class VllmStatLogger(VllmStatLoggerBase):
     """StatLogger is used as an adapter between vLLM stats collector and Triton metrics provider."""
 
-    # local_interval not used here. It's for vLLM logs to stdout.
-    def __init__(self, labels: Dict, max_model_len: int) -> None:
+    def __init__(self, labels: Dict, max_model_len: int, log_logger) -> None:
         # Tracked stats over current local logging interval.
+        # local_interval not used here. It's for vLLM logs to stdout.
         super().__init__(local_interval=0)
         self.metrics = TritonMetrics(labels, max_model_len)
+        self.log_logger = log_logger
+
+        # Starting the metrics thread. It allows vLLM to keep making progress
+        # while reporting metrics to triton metrics service.
+        self._logger_queue = queue.Queue()
+        self._logger_thread = threading.Thread(target=self.logger_loop)
+        self._logger_thread.start()
 
     def info(self, type: str, obj: SupportsMetricsInfo) -> None:
         pass
@@ -190,7 +199,7 @@ def _log_counter(self, counter, data: Union[int, float]) -> None:
             None
         """
         if data != 0:
-            counter.increment(data)
+            self._logger_queue.put_nowait((counter, "increment", data))
 
     def _log_histogram(self, histogram, data: Union[List[int], List[float]]) -> None:
         """Convenience function for logging list to histogram.
@@ -203,7 +212,7 @@ def _log_histogram(self, histogram, data: Union[List[int], List[float]]) -> None
             None
         """
         for datum in data:
-            histogram.observe(datum)
+            self._logger_queue.put_nowait((histogram, "observe", datum))
 
     def log(self, stats: VllmStats) -> None:
         """Report stats to Triton metrics server.
@@ -246,3 +255,24 @@ def log(self, stats: VllmStats) -> None:
             self._log_counter(metric, data)
         for metric, data in histogram_metrics:
             self._log_histogram(metric, data)
+
+    def logger_loop(self):
+        while True:
+            item = self._logger_queue.get()
+            # To signal shutdown a None item will be added to the queue.
+            if item is None:
+                break
+            metric, command, data = item
+            if command == "increment":
+                metric.increment(data)
+            elif command == "observe":
+                metric.observe(data)
+            else:
+                self.log_logger.log_error(f"Undefined command name: {command}")
+
+    def finalize(self):
+        # Shutdown the logger thread.
+        self._logger_queue.put(None)
+        if self._logger_thread is not None:
+            self._logger_thread.join()
+            self._logger_thread = None