生产指标#
vLLM 公开了一些指标，可用于监控系统的健康状况。这些指标通过 vLLM OpenAI 兼容 API 服务器上的 /metrics 端点公开。
您可以使用 Python 或 Docker 启动服务器
vllm serve unsloth/Llama-3.2-1B-Instruct
然后查询端点以获取服务器的最新指标
$ curl http://0.0.0.0:8000/metrics
# HELP vllm:iteration_tokens_total Histogram of number of tokens per engine_step.
# TYPE vllm:iteration_tokens_total histogram
vllm:iteration_tokens_total_sum{model_name="unsloth/Llama-3.2-1B-Instruct"} 0.0
vllm:iteration_tokens_total_bucket{le="1.0",model_name="unsloth/Llama-3.2-1B-Instruct"} 3.0
vllm:iteration_tokens_total_bucket{le="8.0",model_name="unsloth/Llama-3.2-1B-Instruct"} 3.0
vllm:iteration_tokens_total_bucket{le="16.0",model_name="unsloth/Llama-3.2-1B-Instruct"} 3.0
vllm:iteration_tokens_total_bucket{le="32.0",model_name="unsloth/Llama-3.2-1B-Instruct"} 3.0
vllm:iteration_tokens_total_bucket{le="64.0",model_name="unsloth/Llama-3.2-1B-Instruct"} 3.0
vllm:iteration_tokens_total_bucket{le="128.0",model_name="unsloth/Llama-3.2-1B-Instruct"} 3.0
vllm:iteration_tokens_total_bucket{le="256.0",model_name="unsloth/Llama-3.2-1B-Instruct"} 3.0
vllm:iteration_tokens_total_bucket{le="512.0",model_name="unsloth/Llama-3.2-1B-Instruct"} 3.0
...
以下是公开的指标
class Metrics:
"""
vLLM uses a multiprocessing-based frontend for the OpenAI server.
This means that we need to run prometheus_client in multiprocessing mode
See https://prometheus.github.io/client_python/multiprocess/ for more
details on limitations.
"""
labelname_finish_reason = "finished_reason"
labelname_waiting_lora_adapters = "waiting_lora_adapters"
labelname_running_lora_adapters = "running_lora_adapters"
labelname_max_lora = "max_lora"
_gauge_cls = prometheus_client.Gauge
_counter_cls = prometheus_client.Counter
_histogram_cls = prometheus_client.Histogram
def __init__(self, labelnames: List[str], vllm_config: VllmConfig):
# Unregister any existing vLLM collectors (for CI/CD)
self._unregister_vllm_metrics()
max_model_len = vllm_config.model_config.max_model_len
# Use this flag to hide metrics that were deprecated in
# a previous release and which will be removed future
self.show_hidden_metrics = \
vllm_config.observability_config.show_hidden_metrics
# System stats
# Scheduler State
self.gauge_scheduler_running = self._gauge_cls(
name="vllm:num_requests_running",
documentation="Number of requests currently running on GPU.",
labelnames=labelnames,
multiprocess_mode="sum")
self.gauge_scheduler_waiting = self._gauge_cls(
name="vllm:num_requests_waiting",
documentation="Number of requests waiting to be processed.",
labelnames=labelnames,
multiprocess_mode="sum")
self.gauge_lora_info = self._gauge_cls(
name="vllm:lora_requests_info",
documentation="Running stats on lora requests.",
labelnames=[
self.labelname_running_lora_adapters,
self.labelname_max_lora,
self.labelname_waiting_lora_adapters,
],
multiprocess_mode="livemostrecent",
)
# Deprecated in 0.8 - KV cache offloading is not used in V1
# Hidden in 0.9, due to be removed in 0.10
if self.show_hidden_metrics:
self.gauge_scheduler_swapped = self._gauge_cls(
name="vllm:num_requests_swapped",
documentation=(
"Number of requests swapped to CPU. "
"DEPRECATED: KV cache offloading is not used in V1"),
labelnames=labelnames,
multiprocess_mode="sum")
# KV Cache Usage in %
self.gauge_gpu_cache_usage = self._gauge_cls(
name="vllm:gpu_cache_usage_perc",
documentation="GPU KV-cache usage. 1 means 100 percent usage.",
labelnames=labelnames,
multiprocess_mode="sum")
# Deprecated in 0.8 - KV cache offloading is not used in V1
# Hidden in 0.9, due to be removed in 0.10
if self.show_hidden_metrics:
self.gauge_cpu_cache_usage = self._gauge_cls(
name="vllm:cpu_cache_usage_perc",
documentation=(
"CPU KV-cache usage. 1 means 100 percent usage. "
"DEPRECATED: KV cache offloading is not used in V1"),
labelnames=labelnames,
multiprocess_mode="sum")
self.gauge_cpu_prefix_cache_hit_rate = self._gauge_cls(
name="vllm:cpu_prefix_cache_hit_rate",
documentation=(
"CPU prefix cache block hit rate. "
"DEPRECATED: KV cache offloading is not used in V1"),
labelnames=labelnames,
multiprocess_mode="sum")
# Deprecated in 0.8 - replaced by queries+hits counters in V1
# Hidden in 0.9, due to be removed in 0.10
if self.show_hidden_metrics:
self.gauge_gpu_prefix_cache_hit_rate = self._gauge_cls(
name="vllm:gpu_prefix_cache_hit_rate",
documentation=("GPU prefix cache block hit rate. "
"DEPRECATED: use vllm:gpu_prefix_cache_queries "
"and vllm:gpu_prefix_cache_queries in V1"),
labelnames=labelnames,
multiprocess_mode="sum")
# Iteration stats
self.counter_num_preemption = self._counter_cls(
name="vllm:num_preemptions_total",
documentation="Cumulative number of preemption from the engine.",
labelnames=labelnames)
self.counter_prompt_tokens = self._counter_cls(
name="vllm:prompt_tokens_total",
documentation="Number of prefill tokens processed.",
labelnames=labelnames)
self.counter_generation_tokens = self._counter_cls(
name="vllm:generation_tokens_total",
documentation="Number of generation tokens processed.",
labelnames=labelnames)
buckets = [1, 8, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8096]
if not vllm_config.model_config.enforce_eager:
buckets = vllm_config.compilation_config.\
cudagraph_capture_sizes.copy()
buckets.sort()
self.histogram_iteration_tokens = self._histogram_cls(
name="vllm:iteration_tokens_total",
documentation="Histogram of number of tokens per engine_step.",
labelnames=labelnames,
buckets=buckets)
self.histogram_time_to_first_token = self._histogram_cls(
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, 20.0, 40.0, 80.0, 160.0, 640.0,
2560.0
])
self.histogram_time_per_output_token = self._histogram_cls(
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, 5.0, 7.5, 10.0, 20.0, 40.0, 80.0
])
# Request stats
# Latency
request_latency_buckets = [
0.3, 0.5, 0.8, 1.0, 1.5, 2.0, 2.5, 5.0, 10.0, 15.0, 20.0, 30.0,
40.0, 50.0, 60.0, 120.0, 240.0, 480.0, 960.0, 1920.0, 7680.0
]
self.histogram_e2e_time_request = self._histogram_cls(
name="vllm:e2e_request_latency_seconds",
documentation="Histogram of end to end request latency in seconds.",
labelnames=labelnames,
buckets=request_latency_buckets)
self.histogram_queue_time_request = self._histogram_cls(
name="vllm:request_queue_time_seconds",
documentation=
"Histogram of time spent in WAITING phase for request.",
labelnames=labelnames,
buckets=request_latency_buckets)
self.histogram_inference_time_request = self._histogram_cls(
name="vllm:request_inference_time_seconds",
documentation=
"Histogram of time spent in RUNNING phase for request.",
labelnames=labelnames,
buckets=request_latency_buckets)
self.histogram_prefill_time_request = self._histogram_cls(
name="vllm:request_prefill_time_seconds",
documentation=
"Histogram of time spent in PREFILL phase for request.",
labelnames=labelnames,
buckets=request_latency_buckets)
self.histogram_decode_time_request = self._histogram_cls(
name="vllm:request_decode_time_seconds",
documentation=
"Histogram of time spent in DECODE phase for request.",
labelnames=labelnames,
buckets=request_latency_buckets)
# Deprecated in 0.8 - duplicates vllm:request_queue_time_seconds:
# Hidden in 0.9, due to be removed in 0.10
if self.show_hidden_metrics:
self.histogram_time_in_queue_request = self._histogram_cls(
name="vllm:time_in_queue_requests",
documentation=
("Histogram of time the request spent in the queue in seconds. "
"DEPRECATED: use vllm:request_queue_time_seconds instead."),
labelnames=labelnames,
buckets=request_latency_buckets)
# Deprecated in 0.8 - use prefill/decode/inference time metrics
# Hidden in 0.9, due to be removed in 0.10
if self.show_hidden_metrics:
self.histogram_model_forward_time_request = self._histogram_cls(
name="vllm:model_forward_time_milliseconds",
documentation=
("Histogram of time spent in the model forward pass in ms. "
"DEPRECATED: use prefill/decode/inference time metrics instead"
),
labelnames=labelnames,
buckets=build_1_2_3_5_8_buckets(3000))
self.histogram_model_execute_time_request = self._histogram_cls(
name="vllm:model_execute_time_milliseconds",
documentation=
("Histogram of time spent in the model execute function in ms."
"DEPRECATED: use prefill/decode/inference time metrics instead"
),
labelnames=labelnames,
buckets=build_1_2_3_5_8_buckets(3000))
# Metadata
self.histogram_num_prompt_tokens_request = self._histogram_cls(
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._histogram_cls(
name="vllm:request_generation_tokens",
documentation="Number of generation tokens processed.",
labelnames=labelnames,
buckets=build_1_2_5_buckets(max_model_len),
)
self.histogram_max_num_generation_tokens_request = self._histogram_cls(
name="vllm:request_max_num_generation_tokens",
documentation=
"Histogram of maximum number of requested generation tokens.",
labelnames=labelnames,
buckets=build_1_2_5_buckets(max_model_len))
self.histogram_n_request = self._histogram_cls(
name="vllm:request_params_n",
documentation="Histogram of the n request parameter.",
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.",
labelnames=labelnames + [Metrics.labelname_finish_reason])
# Speculative decoding stats
self.gauge_spec_decode_draft_acceptance_rate = self._gauge_cls(
name="vllm:spec_decode_draft_acceptance_rate",
documentation="Speulative token acceptance rate.",
labelnames=labelnames,
multiprocess_mode="sum")
self.gauge_spec_decode_efficiency = self._gauge_cls(
name="vllm:spec_decode_efficiency",
documentation="Speculative decoding system efficiency.",
labelnames=labelnames,
multiprocess_mode="sum")
self.counter_spec_decode_num_accepted_tokens = (self._counter_cls(
name="vllm:spec_decode_num_accepted_tokens_total",
documentation="Number of accepted tokens.",
labelnames=labelnames))
self.counter_spec_decode_num_draft_tokens = self._counter_cls(
name="vllm:spec_decode_num_draft_tokens_total",
documentation="Number of draft tokens.",
labelnames=labelnames)
self.counter_spec_decode_num_emitted_tokens = (self._counter_cls(
name="vllm:spec_decode_num_emitted_tokens_total",
documentation="Number of emitted tokens.",
labelnames=labelnames))
以下指标已弃用，并将在未来版本中移除
vllm:num_requests_swapped、vllm:cpu_cache_usage_perc和vllm:cpu_prefix_cache_hit_rate，因为 V1 中未使用 KV 缓存卸载。vllm:gpu_prefix_cache_hit_rate已被 V1 中的 queries+hits 计数器取代。vllm:time_in_queue_requests，因为它与vllm:request_queue_time_seconds重复。vllm:model_forward_time_milliseconds和vllm:model_execute_time_milliseconds，因为应使用 prefill/decode/inference 时间指标来代替。
注意：当指标在 X.Y 版本中被弃用时，它们会在 X.Y+1 版本中被隐藏，但可以使用 --show-hidden-metrics-for-version=X.Y 逃生舱口重新启用，然后在 X.Y+2 版本中被移除。