Deprecate --collect-tokens-histogram, auto-collect with --enable-metrics (#23595)

python/sglang/srt/managers/tokenizer_manager.py

@@ -470,7 +470,6 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
                 bucket_time_to_first_token=self.server_args.bucket_time_to_first_token,
                 bucket_e2e_request_latency=self.server_args.bucket_e2e_request_latency,
                 bucket_inter_token_latency=self.server_args.bucket_inter_token_latency,
-                collect_tokens_histogram=self.server_args.collect_tokens_histogram,
             )
 
             start_cpu_monitor_thread("tokenizer")

python/sglang/srt/observability/metrics_collector.py

@@ -1151,13 +1151,11 @@ class TokenizerMetricsCollector:
         bucket_time_to_first_token: Optional[List[float]] = None,
         bucket_inter_token_latency: Optional[List[float]] = None,
         bucket_e2e_request_latency: Optional[List[float]] = None,
-        collect_tokens_histogram: bool = False,
     ) -> None:
         # We need to import prometheus_client after setting the env variable `PROMETHEUS_MULTIPROC_DIR`
         from prometheus_client import Counter, Histogram
 
         self.labels = labels or {}
-        self.collect_tokens_histogram = collect_tokens_histogram
 
         self.prompt_tokens_total = Counter(
             name="sglang:prompt_tokens_total",
@@ -1171,7 +1169,6 @@ class TokenizerMetricsCollector:
             labelnames=labels.keys(),
         )
 
-        if collect_tokens_histogram:
         default_bucket_prompt_tokens = [
             100,
             300,
@@ -1413,7 +1410,6 @@ class TokenizerMetricsCollector:
         if has_grammar:
             self.num_so_requests_total.labels(**labels).inc(1)
         self.histogram_e2e_request_latency.labels(**labels).observe(float(e2e_latency))
-        if self.collect_tokens_histogram:
         self.prompt_tokens_histogram.labels(**labels).observe(float(prompt_tokens))
         self.generation_tokens_histogram.labels(**labels).observe(
             float(generation_tokens)

python/sglang/srt/server_args.py

@@ -420,7 +420,6 @@ class ServerArgs:
     bucket_time_to_first_token: Optional[List[float]] = None
     bucket_inter_token_latency: Optional[List[float]] = None
     bucket_e2e_request_latency: Optional[List[float]] = None
-    collect_tokens_histogram: bool = False
     prompt_tokens_buckets: Optional[List[str]] = None
     generation_tokens_buckets: Optional[List[str]] = None
     gc_warning_threshold_secs: float = 0.0
@@ -4803,9 +4802,8 @@ class ServerArgs:
         )
         parser.add_argument(
             "--collect-tokens-histogram",
-            action="store_true",
+            action=DeprecatedAction,
-            default=ServerArgs.collect_tokens_histogram,
+            help="Deprecated. Token histograms are now automatically collected when --enable-metrics is set.",
-            help="Collect prompt/generation tokens histogram.",
         )
         bucket_rule = (
             "Supports 3 rule types: 'default' uses predefined buckets; 'tse <middle> <base> <count>' "