[metrics][EPLB]: Support selected count of physical experts on each GPU (#9825)

2025-10-28 20:56:19 -07:00
parent 0297773a2f
commit 5e36a0b455
3 changed files with 54 additions and 8 deletions
@@ -999,3 +999,16 @@ class StorageMetricsCollector:
            self._log_histogram(self.histogram_prefetch_bandwidth, v)
        for v in storage_metrics.backup_bandwidth:
            self._log_histogram(self.histogram_backup_bandwidth, v)
+
+
+class ExpertDispatchCollector:
+    def __init__(self, ep_size: int) -> None:
+        from prometheus_client import Histogram
+
+        ep_size_buckets = [i for i in range(ep_size)]
+        self.eplb_gpu_physical_count = Histogram(
+            name="sglang:eplb_gpu_physical_count",
+            documentation="The selected count of physical experts on each layer and GPU rank.",
+            labelnames={"layer"},
+            buckets=ep_size_buckets,
+        )