Merge pull request #64 from aish1331/prometheus-metrics-client

Added Prometheus client to get model server metrics

Author: k8s-ci-robot

config.yml

@@ -11,3 +11,8 @@ tokenizer:
   pretrained_model_name_or_path: HuggingFaceTB/SmolLM2-135M-Instruct
 data:
   type: shareGPT
+metrics_client:
+  type: prometheus
+  prometheus:
+    url: http://localhost:9090
+    scrape_interval: 15

inference_perf/client/base.py

@@ -12,9 +12,82 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from abc import ABC, abstractmethod
-from typing import Tuple
+from typing import List, Tuple, TypedDict
+
+from pydantic import BaseModel
 from inference_perf.datagen import InferenceData
-from inference_perf.reportgen import ReportGenerator
+
+
+class RequestMetric(BaseModel):
+    stage_id: int
+    prompt_tokens: int
+    output_tokens: int
+    time_per_request: float
+
+
+class ModelServerPrometheusMetric:
+    def __init__(self, name: str, op: str, type: str, filters: str) -> None:
+        self.name = name
+        self.op = op
+        self.type = type
+        self.filters = filters
+
+
+class ModelServerMetrics(BaseModel):
+    # Throughput
+    prompt_tokens_per_second: float = 0.0
+    output_tokens_per_second: float = 0.0
+    requests_per_second: float = 0.0
+
+    # Latency
+    avg_request_latency: float = 0.0
+    median_request_latency: float = 0.0
+    p90_request_latency: float = 0.0
+    p99_request_latency: float = 0.0
+    avg_time_to_first_token: float = 0.0
+    median_time_to_first_token: float = 0.0
+    p90_time_to_first_token: float = 0.0
+    p99_time_to_first_token: float = 0.0
+    avg_time_per_output_token: float = 0.0
+    median_time_per_output_token: float = 0.0
+    p90_time_per_output_token: float = 0.0
+    p99_time_per_output_token: float = 0.0
+
+    # Request
+    total_requests: int = 0
+    avg_prompt_tokens: int = 0
+    avg_output_tokens: int = 0
+    avg_queue_length: int = 0
+
+
+# PrometheusMetricMetadata stores the mapping of metrics to their model server names and types
+# and the filters to be applied to them.
+# This is used to generate Prometheus query for the metrics.
+class PrometheusMetricMetadata(TypedDict):
+    # Throughput
+    prompt_tokens_per_second: ModelServerPrometheusMetric
+    output_tokens_per_second: ModelServerPrometheusMetric
+    requests_per_second: ModelServerPrometheusMetric
+
+    # Latency
+    avg_request_latency: ModelServerPrometheusMetric
+    median_request_latency: ModelServerPrometheusMetric
+    p90_request_latency: ModelServerPrometheusMetric
+    p99_request_latency: ModelServerPrometheusMetric
+    avg_time_to_first_token: ModelServerPrometheusMetric
+    median_time_to_first_token: ModelServerPrometheusMetric
+    p90_time_to_first_token: ModelServerPrometheusMetric
+    p99_time_to_first_token: ModelServerPrometheusMetric
+    avg_time_per_output_token: ModelServerPrometheusMetric
+    median_time_per_output_token: ModelServerPrometheusMetric
+    p90_time_per_output_token: ModelServerPrometheusMetric
+    p99_time_per_output_token: ModelServerPrometheusMetric
+
+    # Request
+    total_requests: ModelServerPrometheusMetric
+    avg_prompt_tokens: ModelServerPrometheusMetric
+    avg_output_tokens: ModelServerPrometheusMetric
+    avg_queue_length: ModelServerPrometheusMetric
 
 
 class ModelServerClient(ABC):
@@ -23,9 +96,14 @@ def __init__(self, *args: Tuple[int, ...]) -> None:
         pass
 
     @abstractmethod
-    def set_report_generator(self, reportgen: ReportGenerator) -> None:
-        self.reportgen = reportgen
+    async def process_request(self, data: InferenceData, stage_id: int) -> None:
+        raise NotImplementedError
 
     @abstractmethod
-    async def process_request(self, data: InferenceData, stage_id: int) -> None:
+    def get_request_metrics(self) -> List[RequestMetric]:
+        raise NotImplementedError
+
+    @abstractmethod
+    def get_prometheus_metric_metadata(self) -> PrometheusMetricMetadata:
+        # assumption: all metrics clients have metrics exported in Prometheus format
         raise NotImplementedError

inference_perf/client/mock_client.py

@@ -11,27 +11,27 @@
 # 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.
-from inference_perf.datagen import InferenceData
-from inference_perf.reportgen import ReportGenerator, RequestMetric
-from .base import ModelServerClient
 import asyncio
+from typing import List
+from inference_perf.datagen import InferenceData
+from .base import ModelServerClient, RequestMetric
 
 
 class MockModelServerClient(ModelServerClient):
     def __init__(self) -> None:
-        pass
-
-    def set_report_generator(self, reportgen: ReportGenerator) -> None:
-        self.reportgen = reportgen
+        self.request_metrics: List[RequestMetric] = list()
 
     async def process_request(self, payload: InferenceData, stage_id: int) -> None:
         print("Processing request - " + str(payload.data) + " for stage - " + str(stage_id))
         await asyncio.sleep(3)
-        self.reportgen.collect_request_metrics(
+        self.request_metrics.append(
             RequestMetric(
                 stage_id=stage_id,
                 prompt_tokens=0,
                 output_tokens=0,
                 time_per_request=3,
             )
         )
+
+    def get_request_metrics(self) -> List[RequestMetric]:
+        return self.request_metrics

inference_perf/client/vllm_client.py

@@ -12,11 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from inference_perf.datagen import InferenceData
-from inference_perf.reportgen import ReportGenerator, RequestMetric
 from inference_perf.config import APIType, CustomTokenizerConfig
 from inference_perf.utils import CustomTokenizer
-from .base import ModelServerClient
-from typing import Any, Optional
+from .base import ModelServerClient, ModelServerPrometheusMetric, PrometheusMetricMetadata, RequestMetric
+from typing import Any, Optional, List
 import aiohttp
 import json
 import time
@@ -42,9 +41,67 @@ def __init__(self, uri: str, model_name: str, tokenizer: Optional[CustomTokenize
                 print("Falling back to usage metrics.")
         else:
             print("Tokenizer path is empty. Falling back to usage metrics.")
+        self.request_metrics: List[RequestMetric] = list()
 
-    def set_report_generator(self, reportgen: ReportGenerator) -> None:
-        self.reportgen = reportgen
+        self.prometheus_metric_metadata: PrometheusMetricMetadata = {
+            "avg_queue_length": ModelServerPrometheusMetric(
+                "vllm:num_requests_waiting", "mean", "gauge", "model_name='%s'" % self.model_name
+            ),
+            "avg_time_to_first_token": ModelServerPrometheusMetric(
+                "vllm:time_to_first_token_seconds", "mean", "histogram", "model_name='%s'" % self.model_name
+            ),
+            "median_time_to_first_token": ModelServerPrometheusMetric(
+                "vllm:time_to_first_token_seconds", "median", "histogram", "model_name='%s'" % self.model_name
+            ),
+            "p90_time_to_first_token": ModelServerPrometheusMetric(
+                "vllm:time_to_first_token_seconds", "p90", "histogram", "model_name='%s'" % self.model_name
+            ),
+            "p99_time_to_first_token": ModelServerPrometheusMetric(
+                "vllm:time_to_first_token_seconds", "p99", "histogram", "model_name='%s'" % self.model_name
+            ),
+            "avg_time_per_output_token": ModelServerPrometheusMetric(
+                "vllm:time_per_output_token_seconds", "mean", "histogram", "model_name='%s'" % self.model_name
+            ),
+            "median_time_per_output_token": ModelServerPrometheusMetric(
+                "vllm:time_per_output_token_seconds", "median", "histogram", "model_name='%s'" % self.model_name
+            ),
+            "p90_time_per_output_token": ModelServerPrometheusMetric(
+                "vllm:time_per_output_token_seconds", "p90", "histogram", "model_name='%s'" % self.model_name
+            ),
+            "p99_time_per_output_token": ModelServerPrometheusMetric(
+                "vllm:time_per_output_token_seconds", "p99", "histogram", "model_name='%s'" % self.model_name
+            ),
+            "avg_prompt_tokens": ModelServerPrometheusMetric(
+                "vllm:prompt_tokens_total", "mean", "counter", "model_name='%s'" % self.model_name
+            ),
+            "prompt_tokens_per_second": ModelServerPrometheusMetric(
+                "vllm:prompt_tokens_total", "rate", "counter", "model_name='%s'" % self.model_name
+            ),
+            "avg_output_tokens": ModelServerPrometheusMetric(
+                "vllm:generation_tokens_total", "mean", "counter", "model_name='%s'" % self.model_name
+            ),
+            "output_tokens_per_second": ModelServerPrometheusMetric(
+                "vllm:generation_tokens_total", "rate", "counter", "model_name='%s'" % self.model_name
+            ),
+            "total_requests": ModelServerPrometheusMetric(
+                "vllm:e2e_request_latency_seconds_count", "increase", "counter", "model_name='%s'" % self.model_name
+            ),
+            "requests_per_second": ModelServerPrometheusMetric(
+                "vllm:e2e_request_latency_seconds_count", "rate", "counter", "model_name='%s'" % self.model_name
+            ),
+            "avg_request_latency": ModelServerPrometheusMetric(
+                "vllm:e2e_request_latency_seconds", "mean", "histogram", "model_name='%s'" % self.model_name
+            ),
+            "median_request_latency": ModelServerPrometheusMetric(
+                "vllm:e2e_request_latency_seconds", "median", "histogram", "model_name='%s'" % self.model_name
+            ),
+            "p90_request_latency": ModelServerPrometheusMetric(
+                "vllm:e2e_request_latency_seconds", "p90", "histogram", "model_name='%s'" % self.model_name
+            ),
+            "p99_request_latency": ModelServerPrometheusMetric(
+                "vllm:e2e_request_latency_seconds", "p99", "histogram", "model_name='%s'" % self.model_name
+            ),
+        }
 
     def _create_payload(self, payload: InferenceData) -> dict[str, Any]:
         if payload.type == APIType.Completion:
@@ -93,7 +150,7 @@ async def process_request(self, data: InferenceData, stage_id: int) -> None:
                             prompt_tokens = usage.get("prompt_tokens", 0)
                             output_tokens = usage.get("completion_tokens", 0)
 
-                        self.reportgen.collect_request_metrics(
+                        self.request_metrics.append(
                             RequestMetric(
                                 stage_id=stage_id,
                                 prompt_tokens=prompt_tokens,
@@ -105,3 +162,9 @@ async def process_request(self, data: InferenceData, stage_id: int) -> None:
                         print(await response.text())
             except aiohttp.ClientConnectorError as e:
                 print("vLLM Server connection error:\n", str(e))
+
+    def get_request_metrics(self) -> List[RequestMetric]:
+        return self.request_metrics
+
+    def get_prometheus_metric_metadata(self) -> PrometheusMetricMetadata:
+        return self.prometheus_metric_metadata

inference_perf/config.py

@@ -38,6 +38,11 @@ class LoadType(Enum):
     POISSON = "poisson"
 
 
+class MetricsClientType(Enum):
+    PROMETHEUS = "prometheus"
+    DEFAULT = "default"
+
+
 class LoadStage(BaseModel):
     rate: int = 1
     duration: int = 1
@@ -66,8 +71,14 @@ class ReportConfig(BaseModel):
     pass
 
 
-class MetricsConfig(BaseModel):
-    url: str
+class PrometheusClientConfig(BaseModel):
+    scrape_interval: Optional[int] = 15
+    url: str = "http://localhost:9090"
+
+
+class MetricsClientConfig(BaseModel):
+    type: MetricsClientType
+    prometheus: Optional[PrometheusClientConfig] = None
 
 
 class VLLMConfig(BaseModel):
@@ -86,7 +97,7 @@ class Config(BaseModel):
     data: Optional[DataConfig] = DataConfig()
     load: Optional[LoadConfig] = LoadConfig(stages=[LoadStage()])
     report: Optional[ReportConfig] = ReportConfig()
-    metrics: Optional[MetricsConfig] = MetricsConfig(url="")
+    metrics_client: Optional[MetricsClientConfig] = None
     storage: Optional[StorageConfig] = StorageConfig()
     vllm: Optional[VLLMConfig] = None
     tokenizer: Optional[CustomTokenizerConfig] = None

inference_perf/main.py

@@ -11,14 +11,16 @@
 # 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.
+import time
 from typing import List
 from inference_perf.loadgen import LoadGenerator
-from inference_perf.config import DataGenType
+from inference_perf.config import DataGenType, MetricsClientType
 from inference_perf.datagen import DataGenerator, MockDataGenerator, HFShareGPTDataGenerator
 from inference_perf.client import ModelServerClient, vLLMModelServerClient
+from inference_perf.metrics.base import MetricsClient, PerfRuntimeParameters
+from inference_perf.metrics.prometheus_client import PrometheusMetricsClient
 from inference_perf.client.storage import StorageClient, GoogleCloudStorageClient
 from inference_perf.reportgen import ReportGenerator, ReportFile
-from inference_perf.metrics import MockMetricsClient
 from inference_perf.config import read_config
 import asyncio
 
@@ -35,13 +37,12 @@ def __init__(
         self.loadgen = loadgen
         self.reportgen = reportgen
         self.storage_clients = storage_clients
-        self.client.set_report_generator(self.reportgen)
 
     def run(self) -> None:
         asyncio.run(self.loadgen.run(self.client))
 
-    def generate_reports(self) -> List[ReportFile]:
-        return asyncio.run(self.reportgen.generate_reports())
+    def generate_reports(self, runtime_parameters: PerfRuntimeParameters) -> List[ReportFile]:
+        return asyncio.run(self.reportgen.generate_reports(runtime_parameters=runtime_parameters))
 
     def save_reports(self, reports: List[ReportFile]) -> None:
         for storage_client in self.storage_clients:
@@ -53,7 +54,7 @@ def main_cli() -> None:
 
     # Define Model Server Client
     if config.vllm:
-        client = vLLMModelServerClient(
+        model_server_client = vLLMModelServerClient(
             uri=config.vllm.url, model_name=config.vllm.model_name, tokenizer=config.tokenizer, api_type=config.vllm.api
         )
     else:
@@ -76,13 +77,13 @@ def main_cli() -> None:
         raise Exception("load config missing")
 
     # Define Metrics Client
-    if config.metrics:
-        metricsclient = MockMetricsClient(uri=config.metrics.url)
-    else:
-        raise Exception("metrics config missing")
+    metrics_client: MetricsClient | None = None
+    if config.metrics_client:
+        if config.metrics_client.type == MetricsClientType.PROMETHEUS and config.metrics_client.prometheus:
+            metrics_client = PrometheusMetricsClient(config=config.metrics_client.prometheus)
 
     # Define Report Generator
-    reportgen = ReportGenerator(metricsclient)
+    reportgen = ReportGenerator(metrics_client)
 
     # Define Storage Clients
     storage_clients: List[StorageClient] = []
@@ -91,13 +92,18 @@ def main_cli() -> None:
             storage_clients.append(GoogleCloudStorageClient(config=config.storage.google_cloud_storage))
 
     # Setup Perf Test Runner
-    perfrunner = InferencePerfRunner(client, loadgen, reportgen, storage_clients)
+    perfrunner = InferencePerfRunner(model_server_client, loadgen, reportgen, storage_clients)
+
+    start_time = time.time()
 
     # Run Perf Test
     perfrunner.run()
 
-    # Generate Reports
-    reports = perfrunner.generate_reports()
+    end_time = time.time()
+    duration = end_time - start_time  # Calculate the duration of the test
+
+    # Generate Report after the tests
+    reports = perfrunner.generate_reports(PerfRuntimeParameters(start_time, duration, model_server_client))
 
     # Save Reports
     perfrunner.save_reports(reports=reports)

inference_perf/metrics/README.md

@@ -5,10 +5,9 @@ This repository provides clients to query performance metrics from various monit
 ## Supported Monitoring Platforms
 
 **Available now**:
-- None
+- Self Deployed Prometheus
 
 **Todo**:
 - Google Cloud Monitoring
 - AWS CloudWatch
-- Azure Monitor
-- Self Deployed Prometheus
+- Azure Monitor

inference_perf/metrics/__init__.py

@@ -11,7 +11,7 @@
 # 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.
-from .base import MetricsClient, MetricsSummary
+from .base import MetricsClient
 from .mock_client import MockMetricsClient
 
-__all__ = ["MetricsClient", "MetricsSummary", "MockMetricsClient"]
+__all__ = ["MetricsClient", "MockMetricsClient"]