diff --git a/python/sglang/srt/managers/scheduler_metrics_mixin.py b/python/sglang/srt/managers/scheduler_metrics_mixin.py
index bc990f70a..da2e220a1 100644
--- a/python/sglang/srt/managers/scheduler_metrics_mixin.py
+++ b/python/sglang/srt/managers/scheduler_metrics_mixin.py
@@ -172,6 +172,8 @@ class SchedulerMetricsMixin:
             self.stats.num_grammar_queue_reqs = len(self.grammar_queue)
             self.stats.cache_hit_rate = cache_hit_rate
 
+            self.stats.max_total_num_tokens = self.max_total_num_tokens
+
             # Retract
             self.stats.num_retracted_reqs = self.num_retracted_reqs
             self.stats.num_paused_reqs = self.num_paused_reqs
@@ -320,6 +322,8 @@ class SchedulerMetricsMixin:
             self.stats.num_grammar_queue_reqs = len(self.grammar_queue)
             self.stats.cache_hit_rate = cache_hit_rate
 
+            self.stats.max_total_num_tokens = self.max_total_num_tokens
+
             # Speculative decoding
             self.stats.spec_accept_rate = spec_accept_rate
             self.stats.spec_accept_length = spec_accept_length
diff --git a/python/sglang/srt/metrics/collector.py b/python/sglang/srt/metrics/collector.py
index b4af2288f..ef0d73cc9 100644
--- a/python/sglang/srt/metrics/collector.py
+++ b/python/sglang/srt/metrics/collector.py
@@ -184,6 +184,8 @@ class SchedulerStats:
     num_running_reqs_offline_batch: int = 0
     cache_hit_rate: float = 0.0
 
+    max_total_num_tokens: int = 0
+
     # Speculative decoding
     spec_accept_length: float = 0.0
     spec_accept_rate: float = 0.0
@@ -294,6 +296,13 @@ class SchedulerMetricsCollector:
             multiprocess_mode="mostrecent",
         )
 
+        self.max_total_num_tokens = Gauge(
+            name="sglang:max_total_num_tokens",
+            documentation="Maximum total number of tokens in the KV cache pool.",
+            labelnames=labels.keys(),
+            multiprocess_mode="mostrecent",
+        )
+
         # Speculative decoding
         self.spec_accept_length = Gauge(
             name="sglang:spec_accept_length",
@@ -640,6 +649,8 @@ class SchedulerMetricsCollector:
         )
         self._log_gauge(self.cache_hit_rate, stats.cache_hit_rate)
 
+        self._log_gauge(self.max_total_num_tokens, stats.max_total_num_tokens)
+
         # Speculative decoding
         self._log_gauge(self.spec_accept_length, stats.spec_accept_length)
         self._log_gauge(self.spec_accept_rate, stats.spec_accept_rate)