[Frontend] Add max_tokens prometheus metric (#9881)

Signed-off-by: Tomer Asida <tomera@ai21.com>

tests/entrypoints/openai/test_metrics.py

@@ -70,10 +70,14 @@ EXPECTED_VALUES = {
     [("_sum", _NUM_REQUESTS * _NUM_GENERATION_TOKENS_PER_REQUEST),
      ("_count", _NUM_REQUESTS)],
     "vllm:request_params_n": [("_count", _NUM_REQUESTS)],
+    "vllm:request_params_max_tokens":
+    [("_sum", _NUM_REQUESTS * _NUM_GENERATION_TOKENS_PER_REQUEST),
+     ("_count", _NUM_REQUESTS)],
     "vllm:prompt_tokens": [("_total",
                             _NUM_REQUESTS * _NUM_PROMPT_TOKENS_PER_REQUEST)],
-    "vllm:generation_tokens":
-    [("_total", _NUM_REQUESTS * _NUM_PROMPT_TOKENS_PER_REQUEST)],
+    "vllm:generation_tokens": [
+        ("_total", _NUM_REQUESTS * _NUM_PROMPT_TOKENS_PER_REQUEST)
+    ],
     "vllm:request_success": [("_total", _NUM_REQUESTS)],
 }
 
@@ -149,6 +153,9 @@ EXPECTED_METRICS = [
     "vllm:request_params_n_sum",
     "vllm:request_params_n_bucket",
     "vllm:request_params_n_count",
+    "vllm:request_params_max_tokens_sum",
+    "vllm:request_params_max_tokens_bucket",
+    "vllm:request_params_max_tokens_count",
     "vllm:num_preemptions_total",
     "vllm:prompt_tokens_total",
     "vllm:generation_tokens_total",

tests/metrics/test_metrics.py

@@ -365,6 +365,7 @@ def assert_metrics(engine: LLMEngine, disable_log_stats: bool,
             "vllm:request_prompt_tokens",
             "vllm:request_generation_tokens",
             "vllm:request_params_n",
+            "vllm:request_params_max_tokens",
         ]
         for metric_name in request_histogram_metrics:
             metric_value = REGISTRY.get_sample_value(f"{metric_name}_count",

vllm/engine/llm_engine.py

@@ -1685,6 +1685,7 @@ class LLMEngine:
         num_prompt_tokens_requests: List[int] = []
         num_generation_tokens_requests: List[int] = []
         n_requests: List[int] = []
+        max_tokens_requests: List[int] = []
         finished_reason_requests: List[str] = []
 
         # Lora requests
@@ -1792,6 +1793,8 @@ class LLMEngine:
                     ])
                     if seq_group.sampling_params is not None:
                         n_requests.append(seq_group.sampling_params.n)
+                        max_tokens_requests.append(
+                            seq_group.sampling_params.max_tokens)
                     finished_reason_requests.extend([
                         SequenceStatus.get_finished_reason(seq.status)
                         for seq in seq_group.get_finished_seqs()
@@ -1847,6 +1850,7 @@ class LLMEngine:
             num_prompt_tokens_requests=num_prompt_tokens_requests,
             num_generation_tokens_requests=num_generation_tokens_requests,
             n_requests=n_requests,
+            max_tokens_requests=max_tokens_requests,
             finished_reason_requests=finished_reason_requests,
             max_lora=str(max_lora_stat),
             waiting_lora_adapters=list(waiting_lora_adapters.keys()),

vllm/engine/metrics.py

@@ -179,6 +179,12 @@ class Metrics:
             labelnames=labelnames,
             buckets=[1, 2, 5, 10, 20],
         )
+        self.histogram_max_tokens_request = self._histogram_cls(
+            name="vllm:request_params_max_tokens",
+            documentation="Histogram of the max_tokens request parameter.",
+            labelnames=labelnames,
+            buckets=build_1_2_5_buckets(max_model_len),
+        )
         self.counter_request_success = self._counter_cls(
             name="vllm:request_success_total",
             documentation="Count of successfully processed requests.",
@@ -547,6 +553,8 @@ class PrometheusStatLogger(StatLoggerBase):
             self.metrics.histogram_num_generation_tokens_request,
             stats.num_generation_tokens_requests)
         self._log_histogram(self.metrics.histogram_n_request, stats.n_requests)
+        self._log_histogram(self.metrics.histogram_max_tokens_request,
+                            stats.max_tokens_requests)
 
     def _log_prometheus_interval(self, prompt_throughput: float,
                                  generation_throughput: float) -> None:

vllm/engine/metrics_types.py

@@ -53,6 +53,7 @@ class Stats:
     num_prompt_tokens_requests: List[int]
     num_generation_tokens_requests: List[int]
     n_requests: List[int]
+    max_tokens_requests: List[int]
     finished_reason_requests: List[str]
     waiting_lora_adapters: List[str]
     running_lora_adapters: List[str]