[MISC] Add lora requests to metrics (#9477)

Co-authored-by: Kunjan Patel <kunjanp_google_com@vllm.us-central1-a.c.kunjanp-gke-dev-2.internal>

vllm/engine/llm_engine.py

@@ -1,4 +1,5 @@
 import time
+from collections import Counter as collectionsCounter
 from collections import deque
 from contextlib import contextmanager
 from dataclasses import dataclass
@@ -1617,6 +1618,25 @@ class LLMEngine:
         n_requests: List[int] = []
         finished_reason_requests: List[str] = []
 
+        # Lora requests
+        running_lora_adapters = dict(
+            collectionsCounter([
+                running_request.lora_request.lora_name
+                for scheduler in self.scheduler
+                for running_request in scheduler.running
+                if running_request.lora_request
+            ]))
+        waiting_lora_adapters = dict(
+            collectionsCounter([
+                waiting_request.lora_request.lora_name
+                for scheduler in self.scheduler
+                for waiting_request in scheduler.waiting
+                if waiting_request.lora_request
+            ]))
+        max_lora_stat = "0"
+        if self.lora_config:
+            max_lora_stat = str(self.lora_config.max_loras)
+
         # NOTE: This loop assumes prefill seq_groups are before
         # decode seq_groups in scheduled_seq_groups.
         if scheduler_outputs is not None:
@@ -1738,7 +1758,9 @@ class LLMEngine:
             num_generation_tokens_requests=num_generation_tokens_requests,
             n_requests=n_requests,
             finished_reason_requests=finished_reason_requests,
-        )
+            max_lora=str(max_lora_stat),
+            waiting_lora_adapters=list(waiting_lora_adapters.keys()),
+            running_lora_adapters=list(running_lora_adapters.keys()))
 
     def add_lora(self, lora_request: LoRARequest) -> bool:
         return self.model_executor.add_lora(lora_request)

vllm/engine/metrics.py

@@ -34,7 +34,11 @@ class Metrics:
     See https://prometheus.github.io/client_python/multiprocess/ for more
     details on limitations.
     """
+
     labelname_finish_reason = "finished_reason"
+    labelname_waiting_lora_adapters = "waiting_lora_adapters"
+    labelname_running_lora_adapters = "running_lora_adapters"
+    labelname_max_lora = "max_lora"
     _gauge_cls = prometheus_client.Gauge
     _counter_cls = prometheus_client.Counter
     _histogram_cls = prometheus_client.Histogram
@@ -55,6 +59,16 @@ class Metrics:
             documentation="Number of requests waiting to be processed.",
             labelnames=labelnames,
             multiprocess_mode="sum")
+        self.gauge_lora_info = self._gauge_cls(
+            name="vllm:lora_requests_info",
+            documentation="Running stats on lora requests.",
+            labelnames=[
+                self.labelname_running_lora_adapters,
+                self.labelname_max_lora,
+                self.labelname_waiting_lora_adapters,
+            ],
+            multiprocess_mode="livemostrecent",
+        )
         self.gauge_scheduler_swapped = self._gauge_cls(
             name="vllm:num_requests_swapped",
             documentation="Number of requests swapped to CPU.",
@@ -426,6 +440,9 @@ class PrometheusStatLogger(StatLoggerBase):
         for datum in data:
             histogram.labels(**self.labels).observe(datum)
 
+    def _log_gauge_string(self, gauge, data: Dict[str, str]) -> None:
+        gauge.labels(**data).set(1)
+
     def _log_prometheus(self, stats: Stats) -> None:
         # System state data
         self._log_gauge(self.metrics.gauge_scheduler_running,
@@ -442,7 +459,17 @@ class PrometheusStatLogger(StatLoggerBase):
                         stats.cpu_prefix_cache_hit_rate)
         self._log_gauge(self.metrics.gauge_gpu_prefix_cache_hit_rate,
                         stats.gpu_prefix_cache_hit_rate)
-
+        # Including max-lora in metric, in future this property of lora
+        # config maybe extended to be dynamic.
+        lora_info = {
+            self.metrics.labelname_running_lora_adapters:
+            ",".join(stats.running_lora_adapters),
+            self.metrics.labelname_waiting_lora_adapters:
+            ",".join(stats.waiting_lora_adapters),
+            self.metrics.labelname_max_lora:
+            stats.max_lora,
+        }
+        self._log_gauge_string(self.metrics.gauge_lora_info, lora_info)
         # Iteration level data
         self._log_counter(self.metrics.counter_num_preemption,
                           stats.num_preemption_iter)

vllm/engine/metrics_types.py

@@ -51,6 +51,9 @@ class Stats:
     num_generation_tokens_requests: List[int]
     n_requests: List[int]
     finished_reason_requests: List[str]
+    waiting_lora_adapters: List[str]
+    running_lora_adapters: List[str]
+    max_lora: str
 
     spec_decode_metrics: Optional["SpecDecodeWorkerMetrics"] = None