Tiny add num retracted tokens metric (#15653)

python/sglang/srt/managers/scheduler.py

@@ -2077,8 +2077,16 @@ class Scheduler(
             new_token_gained = new_available_tokens - old_available_tokens
 
             self.num_retracted_reqs = len(retracted_reqs)
-            if self.enable_metrics and (x := len(retracted_reqs)) > 0:
+            if self.enable_metrics and len(retracted_reqs) > 0:
-                self.metrics_collector.increment_num_retracted_reqs(x)
+                self.metrics_collector.increment_retracted_reqs(
+                    num_retracted_reqs=len(retracted_reqs),
+                    num_retracted_input_tokens=sum(
+                        len(r.origin_input_ids) for r in retracted_reqs
+                    ),
+                    num_retracted_output_tokens=sum(
+                        len(r.output_ids) for r in retracted_reqs
+                    ),
+                )
             self.new_token_ratio = new_token_ratio
             for req in reqs_to_abort:
                 abort_reason: FINISH_ABORT = req.to_finish

python/sglang/srt/metrics/collector.py

@@ -379,6 +379,16 @@ class SchedulerMetricsCollector:
             documentation="Total number of retracted requests.",
             labelnames=labels.keys(),
         )
+        self.num_retracted_input_tokens_total = Counter(
+            name="sglang:num_retracted_input_tokens_total",
+            documentation="Total number of retracted input tokens.",
+            labelnames=labels.keys(),
+        )
+        self.num_retracted_output_tokens_total = Counter(
+            name="sglang:num_retracted_output_tokens_total",
+            documentation="Total number of retracted output tokens.",
+            labelnames=labels.keys(),
+        )
         self.num_paused_reqs = Gauge(
             name="sglang:num_paused_reqs",
             documentation="The number of paused requests by async weight sync.",
@@ -731,8 +741,19 @@ class SchedulerMetricsCollector:
     def observe_queue_time(self, latency: float) -> None:
         self._log_histogram(self.queue_time, latency)
 
-    def increment_num_retracted_reqs(self, num: int) -> None:
+    def increment_retracted_reqs(
-        self.num_retracted_reqs_total.labels(**self.labels).inc(num)
+        self,
+        num_retracted_reqs: int,
+        num_retracted_input_tokens: int,
+        num_retracted_output_tokens: int,
+    ) -> None:
+        self.num_retracted_reqs_total.labels(**self.labels).inc(num_retracted_reqs)
+        self.num_retracted_input_tokens_total.labels(**self.labels).inc(
+            num_retracted_input_tokens
+        )
+        self.num_retracted_output_tokens_total.labels(**self.labels).inc(
+            num_retracted_output_tokens
+        )
 
     def increment_cuda_graph_pass(self, value: bool) -> None:
         # leave room for piecewise cuda graph, etc

test/srt/test_metrics.py

@@ -66,6 +66,7 @@ class TestEnableMetrics(CustomTestCase):
         with (
             envs.SGLANG_ENABLE_METRICS_DP_ATTENTION.override(True),
             envs.SGLANG_ENABLE_METRICS_DEVICE_TIMER.override(True),
+            envs.SGLANG_TEST_RETRACT.override(True),
         ):
             process = popen_launch_server(
                 _MODEL_NAME,
@@ -82,12 +83,13 @@ class TestEnableMetrics(CustomTestCase):
             response = requests.post(
                 f"{DEFAULT_URL_FOR_TEST}/generate",
                 json={
-                    "text": "The capital of France is",
+                    "text": ["The capital of France is"] * 20,
                     "sampling_params": {
                         "temperature": 0,
-                        "max_new_tokens": 32,
+                        "max_new_tokens": 50,
                     },
                     "stream": True,
+                    "ignore_eos": True,
                 },
                 stream=True,
             )