metrics: support customer buckets for prompt/generation_tokens_histogram (#9634)

python/sglang/srt/metrics/collector.py

@@ -18,6 +18,8 @@ from dataclasses import dataclass
 from enum import Enum
 from typing import Dict, List, Optional, Union
 
+from sglang.srt.metrics.utils import generate_buckets
+from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils import get_bool_env_var
 
 SGLANG_TEST_REQUEST_TIME_STATS = get_bool_env_var("SGLANG_TEST_REQUEST_TIME_STATS")
@@ -309,6 +311,7 @@ class SchedulerMetricsCollector:
 class TokenizerMetricsCollector:
     def __init__(
         self,
+        server_args: ServerArgs,
         labels: Dict[str, str],
         bucket_time_to_first_token: Optional[List[float]] = None,
         bucket_inter_token_latency: Optional[List[float]] = None,
@@ -334,7 +337,7 @@ class TokenizerMetricsCollector:
         )
 
         if collect_tokens_histogram:
-            bucket_prompt_tokens = [
+            default_bucket_prompt_tokens = [
                 100,
                 300,
                 500,
@@ -363,9 +366,11 @@ class TokenizerMetricsCollector:
                 name="sglang:prompt_tokens_histogram",
                 documentation="Histogram of prompt token length.",
                 labelnames=labels.keys(),
-                buckets=bucket_prompt_tokens,
+                buckets=generate_buckets(
+                    server_args.prompt_tokens_buckets, default_bucket_prompt_tokens
+                ),
             )
-            bucket_generation_tokens = [
+            default_bucket_generation_tokens = [
                 100,
                 300,
                 500,
@@ -390,7 +395,10 @@ class TokenizerMetricsCollector:
                 name="sglang:generation_tokens_histogram",
                 documentation="Histogram of generation token length.",
                 labelnames=labels.keys(),
-                buckets=bucket_generation_tokens,
+                buckets=generate_buckets(
+                    server_args.generation_tokens_buckets,
+                    default_bucket_generation_tokens,
+                ),
             )
 
         self.cached_tokens_total = Counter(

python/sglang/srt/metrics/utils.py

@@ -0,0 +1,48 @@
+# Copyright 2023-2025 SGLang Team
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Utilities for Prometheus Metrics."""
+import math
+from typing import List
+
+
+def two_sides_exponential_buckets(
+    middle: float, base: float, count: int
+) -> List[float]:
+    buckets = []
+    half_count = math.ceil(count / 2)
+    distance = 1
+    buckets.append(middle)
+    for i in range(half_count):
+        distance *= base
+        buckets.append(middle + distance)
+        buckets.append(max(0, middle - distance))
+    return sorted(set(buckets))
+
+
+def generate_buckets(
+    buckets_rule: List[str], default_buckets: List[float]
+) -> List[float]:
+    if not buckets_rule:
+        buckets_rule = ["default"]
+
+    assert len(buckets_rule) > 0
+    rule = buckets_rule[0]
+    if rule == "tse":
+        middle, base, count = buckets_rule[1:]
+        assert float(base) > 1.0, "Base must be greater than 1.0"
+        return two_sides_exponential_buckets(float(middle), float(base), int(count))
+    if rule == "default":
+        return sorted(set(default_buckets))
+    assert rule == "customer"
+    return sorted(set([float(x) for x in buckets_rule[1:]]))