Add LoRA metrics for potential auto scaling (#15149)

python/sglang/srt/metrics/collector.py

@@ -241,6 +241,11 @@ class SchedulerStats:
     # CUDA graph
     is_cuda_graph: float = 0.0
 
+    # LoRA pool metrics
+    lora_pool_slots_used: int = 0
+    lora_pool_slots_total: int = 0
+    lora_pool_utilization: float = 0.0
+
 
 @dataclass
 class DPCooperationInfo:
@@ -265,11 +270,13 @@ class SchedulerMetricsCollector:
     def __init__(
         self,
         labels: Dict[str, str],
+        enable_lora: bool = False,
     ) -> None:
         # We need to import prometheus_client after setting the env variable `PROMETHEUS_MULTIPROC_DIR`
         from prometheus_client import Counter, Gauge, Histogram, Summary
 
         self.labels = labels
+        self.enable_lora = enable_lora
         self.last_log_time = time.perf_counter()
 
         self.num_running_reqs = Gauge(
@@ -692,6 +699,27 @@ class SchedulerMetricsCollector:
                 labelnames=list(labels.keys()) + ["forward_mode"],
             )
 
+        # LoRA pool metrics (only created when LoRA is enabled)
+        if self.enable_lora:
+            self.lora_pool_slots_used = Gauge(
+                name="sglang:lora_pool_slots_used",
+                documentation="Number of LoRA adapter slots currently occupied in GPU memory.",
+                labelnames=labels.keys(),
+                multiprocess_mode="mostrecent",
+            )
+            self.lora_pool_slots_total = Gauge(
+                name="sglang:lora_pool_slots_total",
+                documentation="Total number of LoRA adapter slots available (max_loras_per_batch).",
+                labelnames=labels.keys(),
+                multiprocess_mode="mostrecent",
+            )
+            self.lora_pool_utilization = Gauge(
+                name="sglang:lora_pool_utilization",
+                documentation="LoRA pool utilization ratio (used/total). 1.0 means pool is full.",
+                labelnames=labels.keys(),
+                multiprocess_mode="mostrecent",
+            )
+
         self.new_token_ratio = Gauge(
             name="sglang:new_token_ratio",
             documentation="The new token ratio.",
@@ -868,6 +896,12 @@ class SchedulerMetricsCollector:
         # CUDA graph
         self._log_gauge(self.is_cuda_graph, stats.is_cuda_graph)
 
+        # LoRA pool metrics (only logged if LoRA is enabled)
+        if self.enable_lora:
+            self._log_gauge(self.lora_pool_slots_used, stats.lora_pool_slots_used)
+            self._log_gauge(self.lora_pool_slots_total, stats.lora_pool_slots_total)
+            self._log_gauge(self.lora_pool_utilization, stats.lora_pool_utilization)
+
         self.last_log_time = time.perf_counter()
 
     def log_grammar_stats(self, grammar_stats) -> None: