Fix metrics (#1963)

python/sglang/srt/metrics/metrics_collector.py

@@ -213,19 +213,67 @@ class Metrics:
             name="sglang:e2e_request_latency_seconds",
             documentation="Histogram of End-to-end request latency in seconds",
             labelnames=labelnames,
-            buckets=build_1_2_5_buckets(max_model_len),
+            buckets=[
+                0.3,
+                0.5,
+                0.8,
+                1.0,
+                1.5,
+                2.0,
+                2.5,
+                5.0,
+                10.0,
+                15.0,
+                20.0,
+                30.0,
+                40.0,
+                50.0,
+                60.0,
+            ],
         )
         self.histogram_time_waiting_requests = Histogram(
             name="sglang:waiting_request_latency_seconds",
             documentation="Histogram of request waiting time in seconds",
             labelnames=labelnames,
-            buckets=build_1_2_5_buckets(max_model_len),
+            buckets=[
+                0.3,
+                0.5,
+                0.8,
+                1.0,
+                1.5,
+                2.0,
+                2.5,
+                5.0,
+                10.0,
+                15.0,
+                20.0,
+                30.0,
+                40.0,
+                50.0,
+                60.0,
+            ],
         )
         self.histogram_time_decode_requests = Histogram(
             name="sglang:decode_request_latency_seconds",
             documentation="Histogram of request decoding time in seconds",
             labelnames=labelnames,
-            buckets=build_1_2_5_buckets(max_model_len),
+            buckets=[
+                0.3,
+                0.5,
+                0.8,
+                1.0,
+                1.5,
+                2.0,
+                2.5,
+                5.0,
+                10.0,
+                15.0,
+                20.0,
+                30.0,
+                40.0,
+                50.0,
+                60.0,
+            ],
         )
 
 

python/sglang/srt/metrics/metrics_types.py

@@ -34,15 +34,12 @@ class Stats:
     num_running_req: int = 0
     num_waiting_req: int = 0
     gen_throughput: float = 0.0
-    num_token: int = 0
-    token_usage: float = 0.0
     waiting_queue: int = 0
     time_e2e_requests: List[float] = field(default_factory=list)
     time_waiting_requests: List[float] = field(default_factory=list)
     time_decode_requests: List[float] = field(default_factory=list)
     # system stats
     token_usage: float = 0.0
-    is_mixed_chunk: bool = False
     new_seq: int = 0
     new_token: int = 0
     cached_token: int = 0

python/sglang/srt/server.py

@@ -446,6 +446,9 @@ def launch_server(
     2. Inter-process communication is done through ICP (each process uses a different port) via the ZMQ library.
     """
 
+    if server_args.enable_metrics:
+        _set_prometheus_env()
+
     launch_engine(server_args=server_args)
 
     # Add api key authorization
@@ -454,7 +457,6 @@ def launch_server(
 
     # add prometheus middleware
     if server_args.enable_metrics:
-        _set_prometheus_env()
         add_prometheus_middleware(app)
 
     # Send a warmup request