Add cache_config_info metric. (#17273)

python/sglang/srt/managers/scheduler.py

@@ -607,6 +607,11 @@ class Scheduler(
                 f"{'available_cpu_mem' if self.device == 'cpu' else 'available_gpu_mem'}={avail_mem:.2f} GB"
             )
 
+        if self.enable_metrics and hasattr(self, "metrics_collector"):
+            self.metrics_collector.emit_cache_config_info(
+                self.page_size, self.max_total_num_tokens // self.page_size
+            )
+
     def init_cache_with_memory_pool(self):
         server_args = self.server_args
 

python/sglang/srt/metrics/collector.py

@@ -849,6 +849,16 @@ class SchedulerMetricsCollector:
             ],
         )
 
+        # This is a work-around Info metric since Info metrics are not supported in Prometheus.
+        # Similar to vLLM, https://github.com/vllm-project/vllm/blob/main/vllm/v1/metrics/loggers.py
+        # If more Info metrics are needed, we can create a common _log_info function.
+        self.cache_config_info = Gauge(
+            name="sglang:cache_config_info",
+            documentation="Cache configuration information.",
+            labelnames=["page_size", "num_pages"],
+            multiprocess_mode="mostrecent",
+        )
+
     def _log_gauge(self, gauge, data: Union[int, float]) -> None:
         # Convenience function for logging to gauge.
         gauge.labels(**self.labels).set(data)
@@ -1066,6 +1076,9 @@ class SchedulerMetricsCollector:
             )
         self.num_grammar_total.labels(**self.labels).inc(1)
 
+    def emit_cache_config_info(self, page_size: int, num_pages: int) -> None:
+        self.cache_config_info.labels(page_size=page_size, num_pages=num_pages).set(1)
+
 
 class TokenizerMetricsCollector:
     def __init__(