chore: Expose new kv_cache metrics from trtllm backend (#6469)

components/src/dynamo/trtllm/publisher.py

@@ -304,6 +304,7 @@ class Publisher:
         component_gauges: LLMBackendMetrics,
         zmq_endpoint: Optional[str] = None,
         enable_local_indexer: bool = False,
+        metrics_collector=None,
     ):
         self.endpoint = endpoint
         self.engine = engine
@@ -313,6 +314,7 @@ class Publisher:
         self.metrics_labels = metrics_labels
         self.component_gauges = component_gauges
         self.enable_local_indexer = enable_local_indexer
+        self.metrics_collector = metrics_collector
         self.attention_dp_size = engine.get_attention_dp_size()
 
         # The first few kv events from the model engine are always "created" type events.
@@ -482,6 +484,16 @@ class Publisher:
             )
             self.component_gauges.set_gpu_cache_usage("0", gpu_cache_usage)
 
+            # Log iteration stats to TRT-LLM MetricsCollector (PR #11243)
+            # This populates trtllm_kv_cache_hit_rate and trtllm_kv_cache_utilization gauges
+            if self.metrics_collector and hasattr(
+                self.metrics_collector, "log_iteration_stats"
+            ):
+                try:
+                    self.metrics_collector.log_iteration_stats(stat)
+                except Exception as e:
+                    logging.warning(f"Failed to log iteration stats: {e}")
+
         await self._polling_loop(
             lambda: self.engine.llm.get_stats_async(timeout=_STATS_TIMEOUT_SEC),
             handle_stat,
@@ -766,6 +778,7 @@ async def get_publisher(
     component_gauges: LLMBackendMetrics,
     zmq_endpoint: Optional[str] = None,
     enable_local_indexer: bool = False,
+    metrics_collector=None,
 ):
     publisher = Publisher(
         endpoint,
@@ -776,6 +789,7 @@ async def get_publisher(
         component_gauges=component_gauges,
         zmq_endpoint=zmq_endpoint,
         enable_local_indexer=enable_local_indexer,
+        metrics_collector=metrics_collector,
     )
     try:
         publisher.initialize()

components/src/dynamo/trtllm/request_handlers/handler_base.py

@@ -802,13 +802,24 @@ class HandlerBase(BaseGenerativeHandler):
                         )
 
                     # Log metrics to TensorRT-LLM MetricsCollector when request finishes
+                    # NOTE: TRT-LLM 1.3.0rc5 (PR #11243) renamed log_metrics_dict → log_request_metrics_dict
                     if (
                         res.finished
                         and self.metrics_collector
                         and hasattr(res, "metrics_dict")
                     ):
                         try:
-                            self.metrics_collector.log_metrics_dict(res.metrics_dict)
+                            if hasattr(
+                                self.metrics_collector,
+                                "log_request_metrics_dict",
+                            ):
+                                self.metrics_collector.log_request_metrics_dict(
+                                    res.metrics_dict
+                                )
+                            else:
+                                self.metrics_collector.log_metrics_dict(
+                                    res.metrics_dict
+                                )
                         except Exception as e:
                             logging.warning(f"Failed to log TensorRT-LLM metrics: {e}")
 

components/src/dynamo/trtllm/workers/llm_worker.py

@@ -181,6 +181,10 @@ async def init_llm_worker(
         "max_beam_width": config.max_beam_width,
         "max_batch_size": config.max_batch_size,
         "return_perf_metrics": config.publish_events_and_metrics,
+        # enable_iter_perf_stats is required for PyTorch backend to compute iteration-level
+        # stats (KV cache utilization, hit rate). TensorRT backend always has this enabled.
+        # See TRT-LLM PR #11243: MetricsCollector.log_iteration_stats() needs these stats.
+        "enable_iter_perf_stats": config.publish_events_and_metrics,
         "kv_connector_config": kv_connector_config,
     }
 
@@ -493,6 +497,7 @@ async def init_llm_worker(
                 component_gauges=component_gauges,
                 zmq_endpoint=trtllm_zmq_bind_endpoint,
                 enable_local_indexer=config.enable_local_indexer,
+                metrics_collector=metrics_collector,
             ) as publisher:
                 handler_config.publisher = publisher
                 handler = RequestHandlerFactory().get_request_handler(handler_config)