Make log statistics interval configurable

Signed-off-by: Sakalya Deshpande <[email protected]>
vllm-project · Feb 16, 2025 · 87df0fa · 87df0fa
1 parent d3d547e
commit 87df0fa
Show file tree

Hide file tree

Showing 5 changed files with 20 additions and 10 deletions.
diff --git a/vllm/config.py b/vllm/config.py
@@ -2612,6 +2612,9 @@ class ObservabilityConfig:
 
     # If set, collects the model execute time for the request.
     collect_model_execute_time: bool = False
+
+    # Interval for collecting log_stats. default value is 5.0
+    log_stats_interval: float = 5.0
 
     def compute_hash(self) -> str:
         """

diff --git a/vllm/engine/arg_utils.py b/vllm/engine/arg_utils.py
@@ -126,6 +126,7 @@ class EngineArgs:
     max_num_seqs: Optional[int] = None
     max_logprobs: int = 20  # Default value for OpenAI Chat Completions API
     disable_log_stats: bool = False
+    log_stats_interval: float = 5.0
     revision: Optional[str] = None
     code_revision: Optional[str] = None
     rope_scaling: Optional[Dict[str, Any]] = None
@@ -1025,6 +1026,11 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
             "Different platforms may support different configs. Make sure the "
             "configs are valid for the platform you are using. The input format"
             " is like '{\"config_key\":\"config_value\"}'")
+
+        parser.add_argument('--log-stats-interval',
+                            type=float,
+                            default=EngineArgs.log_stats_interval,
+                            help='Interval for collecting log statistics.')
         return parser
 
     @classmethod
@@ -1319,7 +1325,7 @@ def create_engine_config(self,
             or "all" in detailed_trace_modules,
             collect_model_execute_time="worker" in detailed_trace_modules
             or "all" in detailed_trace_modules,
-        )
+            log_stats_interval=self.log_stats_interval)
 
         config = VllmConfig(
             model_config=model_config,

diff --git a/vllm/engine/llm_engine.py b/vllm/engine/llm_engine.py
@@ -62,7 +62,6 @@
 from vllm.version import __version__ as VLLM_VERSION
 
 logger = init_logger(__name__)
-_LOCAL_LOGGING_INTERVAL_SEC = 5
 
 _G = TypeVar("_G", bound=BaseTokenizerGroup, default=BaseTokenizerGroup)
 _O = TypeVar("_O", RequestOutput, PoolingRequestOutput)
@@ -370,11 +369,13 @@ def get_tokenizer_for_seq(sequence: Sequence) -> AnyTokenizer:
                 self.stat_loggers = {
                     "logging":
                     LoggingStatLogger(
-                        local_interval=_LOCAL_LOGGING_INTERVAL_SEC,
+                        local_interval=self.vllm_config.observability_config.
+                        log_stats_interval,
                         vllm_config=vllm_config),
                     "prometheus":
                     PrometheusStatLogger(
-                        local_interval=_LOCAL_LOGGING_INTERVAL_SEC,
+                        local_interval=self.vllm_config.observability_config.
+                        log_stats_interval,
                         labels=dict(
                             model_name=self.model_config.served_model_name),
                         vllm_config=vllm_config),

diff --git a/vllm/v1/engine/async_llm.py b/vllm/v1/engine/async_llm.py
@@ -56,7 +56,7 @@ def __init__(
         self.stat_loggers: List[StatLoggerBase] = []
         if self.log_stats:
             self.stat_loggers.extend([
-                LoggingStatLogger(),
+                LoggingStatLogger(vllm_config),
                 PrometheusStatLogger(vllm_config),
             ])
 

diff --git a/vllm/v1/metrics/loggers.py b/vllm/v1/metrics/loggers.py
@@ -15,8 +15,6 @@
 
 logger = init_logger(__name__)
 
-_LOCAL_LOGGING_INTERVAL_SEC = 5.0
-
 
 class StatLoggerBase(ABC):
 
@@ -28,8 +26,9 @@ def log(self, scheduler_stats: SchedulerStats,
 
 class LoggingStatLogger(StatLoggerBase):
 
-    def __init__(self):
+    def __init__(self, vllm_config: VllmConfig):
         self._reset(time.monotonic())
+        self.vllm_config = vllm_config
 
     def _reset(self, now):
         self.last_log_time = now
@@ -42,9 +41,10 @@ def _reset(self, now):
         self.prefix_caching_metrics = PrefixCachingMetrics()
 
     def _local_interval_elapsed(self, now: float) -> bool:
-        # Log every _LOCAL_LOGGING_INTERVAL_SEC.
+        # Log every configured log_stats_interval level
         elapsed_time = now - self.last_log_time
-        return elapsed_time > _LOCAL_LOGGING_INTERVAL_SEC
+        interval = self.vllm_config.observability_config.log_stats_interval
+        return elapsed_time > interval
 
     def _track_iteration_stats(self, iteration_stats: IterationStats):
         # Save tracked stats for token counters.