Production Metrics#
vLLM exposes a number of metrics that can be used to monitor the health of the system. These metrics are exposed via the /metrics endpoint on the vLLM OpenAI compatible API server.
The following metrics are exposed:
class Metrics:
labelname_finish_reason = "finished_reason"
_base_library = prometheus_client
def __init__(self, labelnames: List[str], max_model_len: int):
# Unregister any existing vLLM collectors
self._unregister_vllm_metrics()
# Config Information
self.info_cache_config = prometheus_client.Info(
name='vllm:cache_config',
documentation='information of cache_config')
# System stats
# Scheduler State
self.gauge_scheduler_running = self._base_library.Gauge(
name="vllm:num_requests_running",
documentation="Number of requests currently running on GPU.",
labelnames=labelnames)
self.gauge_scheduler_waiting = self._base_library.Gauge(
name="vllm:num_requests_waiting",
documentation="Number of requests waiting to be processed.",
labelnames=labelnames)
self.gauge_scheduler_swapped = self._base_library.Gauge(
name="vllm:num_requests_swapped",
documentation="Number of requests swapped to CPU.",
labelnames=labelnames)
# KV Cache Usage in %
self.gauge_gpu_cache_usage = self._base_library.Gauge(
name="vllm:gpu_cache_usage_perc",
documentation="GPU KV-cache usage. 1 means 100 percent usage.",
labelnames=labelnames)
self.gauge_cpu_cache_usage = self._base_library.Gauge(
name="vllm:cpu_cache_usage_perc",
documentation="CPU KV-cache usage. 1 means 100 percent usage.",
labelnames=labelnames)
# Iteration stats
self.counter_num_preemption = self._base_library.Counter(
name="vllm:num_preemptions_total",
documentation="Cumulative number of preemption from the engine.",
labelnames=labelnames)
self.counter_prompt_tokens = self._base_library.Counter(
name="vllm:prompt_tokens_total",
documentation="Number of prefill tokens processed.",
labelnames=labelnames)
self.counter_generation_tokens = self._base_library.Counter(
name="vllm:generation_tokens_total",
documentation="Number of generation tokens processed.",
labelnames=labelnames)
self.histogram_time_to_first_token = self._base_library.Histogram(
name="vllm:time_to_first_token_seconds",
documentation="Histogram of time to first token in seconds.",
labelnames=labelnames,
buckets=[
0.001, 0.005, 0.01, 0.02, 0.04, 0.06, 0.08, 0.1, 0.25, 0.5,
0.75, 1.0, 2.5, 5.0, 7.5, 10.0
])
self.histogram_time_per_output_token = self._base_library.Histogram(
name="vllm:time_per_output_token_seconds",
documentation="Histogram of time per output token in seconds.",
labelnames=labelnames,
buckets=[
0.01, 0.025, 0.05, 0.075, 0.1, 0.15, 0.2, 0.3, 0.4, 0.5, 0.75,
1.0, 2.5
])
# Request stats
# Latency
self.histogram_e2e_time_request = self._base_library.Histogram(
name="vllm:e2e_request_latency_seconds",
documentation="Histogram of end to end request latency in seconds.",
labelnames=labelnames,
buckets=[1.0, 2.5, 5.0, 10.0, 15.0, 20.0, 30.0, 40.0, 50.0, 60.0])
# Metadata
self.histogram_num_prompt_tokens_request = self._base_library.Histogram(
name="vllm:request_prompt_tokens",
documentation="Number of prefill tokens processed.",
labelnames=labelnames,
buckets=build_1_2_5_buckets(max_model_len),
)
self.histogram_num_generation_tokens_request = \
self._base_library.Histogram(
name="vllm:request_generation_tokens",
documentation="Number of generation tokens processed.",
labelnames=labelnames,
buckets=build_1_2_5_buckets(max_model_len),
)
self.histogram_best_of_request = self._base_library.Histogram(
name="vllm:request_params_best_of",
documentation="Histogram of the best_of request parameter.",
labelnames=labelnames,
buckets=[1, 2, 5, 10, 20],
)
self.histogram_n_request = self._base_library.Histogram(
name="vllm:request_params_n",
documentation="Histogram of the n request parameter.",
labelnames=labelnames,
buckets=[1, 2, 5, 10, 20],
)
self.counter_request_success = self._base_library.Counter(
name="vllm:request_success_total",
documentation="Count of successfully processed requests.",
labelnames=labelnames + [Metrics.labelname_finish_reason])
# Deprecated in favor of vllm:prompt_tokens_total
self.gauge_avg_prompt_throughput = self._base_library.Gauge(
name="vllm:avg_prompt_throughput_toks_per_s",
documentation="Average prefill throughput in tokens/s.",
labelnames=labelnames,
)
# Deprecated in favor of vllm:generation_tokens_total
self.gauge_avg_generation_throughput = self._base_library.Gauge(
name="vllm:avg_generation_throughput_toks_per_s",
documentation="Average generation throughput in tokens/s.",
labelnames=labelnames,
)
def _unregister_vllm_metrics(self) -> None:
for collector in list(self._base_library.REGISTRY._collector_to_names):
if hasattr(collector, "_name") and "vllm" in collector._name:
self._base_library.REGISTRY.unregister(collector)
class RayMetrics(Metrics):
"""
RayMetrics is used by RayPrometheusStatLogger to log to Ray metrics.
Provides the same metrics as Metrics but uses Ray's util.metrics library.
"""
_base_library = ray_metrics
def __init__(self, labelnames: List[str], max_model_len: int):
if ray_metrics is None:
raise ImportError("RayMetrics requires Ray to be installed.")
super().__init__(labelnames, max_model_len)
def _unregister_vllm_metrics(self) -> None:
# No-op on purpose
pass