Skip to content

Same agent Smaller bill. The tools and infrastructure to get more work from your AI. Fewer tokens. Less compute. Lower cost.

[Feat]: Refactor vLLM OpenTelemetry Monitoring Instrumentation (#776)

Patcher0f3277a89a15
diff --git a/sdk/python/pyproject.toml b/sdk/python/pyproject.toml
index 850f1f90..e8fef7e2 100644
--- a/sdk/python/pyproject.toml
+++ b/sdk/python/pyproject.toml
@@ -1,6 +1,6 @@
[tool.poetry]
name = "openlit"
-version = "1.34.12"
+version = "1.34.13"
description = "OpenTelemetry-native Auto instrumentation library for monitoring LLM Applications and GPUs, facilitating the integration of observability into your GenAI-driven projects"
authors = ["OpenLIT"]
license = "Apache-2.0"
diff --git a/sdk/python/src/openlit/instrumentation/vllm/__init__.py b/sdk/python/src/openlit/instrumentation/vllm/__init__.py
index 9cbb8569..a72f3b1b 100644
--- a/sdk/python/src/openlit/instrumentation/vllm/__init__.py
+++ b/sdk/python/src/openlit/instrumentation/vllm/__init__.py
@@ -1,4 +1,3 @@
-# pylint: disable=useless-return, bad-staticmethod-argument, disable=duplicate-code
"""Initializer of Auto Instrumentation of vLLM Functions"""
from typing import Collection
@@ -14,15 +13,15 @@ _instruments = ("vllm >= 0.5.4",)
class VLLMInstrumentor(BaseInstrumentor):
"""
- An instrumentor for vLLM's client library.
+ An instrumentor for vLLM client library.
"""
def instrumentation_dependencies(self) -> Collection[str]:
return _instruments
def _instrument(self, **kwargs):
- application_name = kwargs.get("application_name", "default_application")
- environment = kwargs.get("environment", "default_environment")
+ application_name = kwargs.get("application_name", "default")
+ environment = kwargs.get("environment", "default")
tracer = kwargs.get("tracer")
metrics = kwargs.get("metrics_dict")
pricing_info = kwargs.get("pricing_info", {})
@@ -30,14 +29,13 @@ class VLLMInstrumentor(BaseInstrumentor):
disable_metrics = kwargs.get("disable_metrics")
version = importlib.metadata.version("vllm")
- # sync chat
+ # Chat completions
wrap_function_wrapper(
"vllm.entrypoints.llm",
"LLM.generate",
generate(version, environment, application_name,
- tracer, pricing_info, capture_message_content, metrics, disable_metrics),
+ tracer, pricing_info, capture_message_content, metrics, disable_metrics),
)
def _uninstrument(self, **kwargs):
- # Proper uninstrumentation logic to revert patched methods
pass
diff --git a/sdk/python/src/openlit/instrumentation/vllm/utils.py b/sdk/python/src/openlit/instrumentation/vllm/utils.py
index 6b5e99a9..7dcf318a 100644
--- a/sdk/python/src/openlit/instrumentation/vllm/utils.py
+++ b/sdk/python/src/openlit/instrumentation/vllm/utils.py
@@ -1,15 +1,15 @@
"""
-Utility functions for vLLM instrumentation.
+vLLM OpenTelemetry instrumentation utility functions
"""
-
import time
-from opentelemetry.sdk.resources import SERVICE_NAME, TELEMETRY_SDK_NAME, DEPLOYMENT_ENVIRONMENT
+
from opentelemetry.trace import Status, StatusCode
+
from openlit.__helpers import (
- calculate_tbt,
- get_chat_model_cost,
general_tokens,
- create_metrics_attributes,
+ get_chat_model_cost,
+ common_span_attributes,
+ record_completion_metrics,
)
from openlit.semcov import SemanticConvention
@@ -24,77 +24,81 @@ def get_inference_config(args, kwargs):
return args[1]
return None
+def format_content(prompts):
+ """
+ Process a list of prompts to extract content.
+ """
+
+ if isinstance(prompts, str):
+ return prompts
+ elif isinstance(prompts, list):
+ return "\n".join(str(prompt) for prompt in prompts)
+ else:
+ return str(prompts)
+
def common_chat_logic(scope, pricing_info, environment, application_name, metrics,
capture_message_content, disable_metrics, version, is_stream):
"""
Process chat request and generate Telemetry
"""
- scope._end_time = time.time()
- if len(scope._timestamps) > 1:
- scope._tbt = calculate_tbt(scope._timestamps)
-
- # Set base span attributes
- scope._span.set_attribute(TELEMETRY_SDK_NAME, "openlit")
- scope._span.set_attribute(SemanticConvention.GEN_AI_OPERATION, SemanticConvention.GEN_AI_OPERATION_TYPE_CHAT)
- scope._span.set_attribute(SemanticConvention.GEN_AI_SYSTEM, SemanticConvention.GEN_AI_SYSTEM_VLLM)
- scope._span.set_attribute(SemanticConvention.GEN_AI_REQUEST_MODEL, scope._request_model)
- scope._span.set_attribute(SemanticConvention.SERVER_PORT, scope._server_port)
- scope._span.set_attribute(SemanticConvention.SERVER_ADDRESS, scope._server_address)
-
- # Handle inference configuration
+ request_model = scope._request_model
+
+ # Extract prompts and completions from vLLM response
+ input_tokens = 0
+ output_tokens = 0
+ prompt = ""
+ completion = ""
+
+ for output in scope._response:
+ prompt += output.prompt + "\n"
+ if output.outputs and len(output.outputs) > 0:
+ completion += output.outputs[0].text + "\n"
+ input_tokens += general_tokens(output.prompt)
+ output_tokens += general_tokens(output.outputs[0].text)
+
+ cost = get_chat_model_cost(request_model, pricing_info, input_tokens, output_tokens)
+
+ # Common Span Attributes
+ common_span_attributes(scope,
+ SemanticConvention.GEN_AI_OPERATION_TYPE_CHAT, SemanticConvention.GEN_AI_SYSTEM_VLLM,
+ scope._server_address, scope._server_port, request_model, request_model,
+ environment, application_name, is_stream, scope._tbt, scope._ttft, version)
+
+ # Span Attributes for Request parameters
inference_config = get_inference_config(scope._args, scope._kwargs)
if inference_config:
- attributes = [
- (SemanticConvention.GEN_AI_REQUEST_FREQUENCY_PENALTY, 'frequency_penalty'),
- (SemanticConvention.GEN_AI_REQUEST_MAX_TOKENS, 'max_tokens'),
- (SemanticConvention.GEN_AI_REQUEST_PRESENCE_PENALTY, 'presence_penalty'),
- (SemanticConvention.GEN_AI_REQUEST_STOP_SEQUENCES, 'stop_sequences'),
- (SemanticConvention.GEN_AI_REQUEST_TEMPERATURE, 'temperature'),
- (SemanticConvention.GEN_AI_REQUEST_TOP_P, 'top_p'),
- (SemanticConvention.GEN_AI_REQUEST_TOP_K, 'top_k'),
- ]
-
- for attribute, key in attributes:
- value = getattr(inference_config, key, None)
- if value is not None:
- scope._span.set_attribute(attribute, value)
-
- scope._span.set_attribute(SemanticConvention.GEN_AI_RESPONSE_MODEL, scope._request_model)
+ scope._span.set_attribute(SemanticConvention.GEN_AI_REQUEST_MAX_TOKENS, getattr(inference_config, 'max_tokens', -1))
+ scope._span.set_attribute(SemanticConvention.GEN_AI_REQUEST_STOP_SEQUENCES, getattr(inference_config, 'stop_sequences', []))
+ scope._span.set_attribute(SemanticConvention.GEN_AI_REQUEST_TEMPERATURE, getattr(inference_config, 'temperature', 1.0))
+ scope._span.set_attribute(SemanticConvention.GEN_AI_REQUEST_TOP_P, getattr(inference_config, 'top_p', 1.0))
+ scope._span.set_attribute(SemanticConvention.GEN_AI_REQUEST_TOP_K, getattr(inference_config, 'top_k', -1))
+ scope._span.set_attribute(SemanticConvention.GEN_AI_REQUEST_PRESENCE_PENALTY,
+ getattr(inference_config, 'presence_penalty', 0.0))
+ scope._span.set_attribute(SemanticConvention.GEN_AI_REQUEST_FREQUENCY_PENALTY,
+ getattr(inference_config, 'frequency_penalty', 0.0))
+
+ # Span Attributes for Response parameters
scope._span.set_attribute(SemanticConvention.GEN_AI_OUTPUT_TYPE, "text")
- # Set base span attributes (Extras)
- scope._span.set_attribute(DEPLOYMENT_ENVIRONMENT, environment)
- scope._span.set_attribute(SERVICE_NAME, application_name)
- scope._span.set_attribute(SemanticConvention.GEN_AI_REQUEST_IS_STREAM, is_stream)
- scope._span.set_attribute(SemanticConvention.GEN_AI_SERVER_TBT, scope._tbt)
- scope._span.set_attribute(SemanticConvention.GEN_AI_SERVER_TTFT, scope._ttft)
- scope._span.set_attribute(SemanticConvention.GEN_AI_SDK_VERSION, version)
-
- input_tokens = 0
- output_tokens = 0
- cost = 0
+ # Span Attributes for Cost and Tokens
+ scope._span.set_attribute(SemanticConvention.GEN_AI_USAGE_INPUT_TOKENS, input_tokens)
+ scope._span.set_attribute(SemanticConvention.GEN_AI_USAGE_OUTPUT_TOKENS, output_tokens)
+ scope._span.set_attribute(SemanticConvention.GEN_AI_CLIENT_TOKEN_USAGE, input_tokens + output_tokens)
+ scope._span.set_attribute(SemanticConvention.GEN_AI_USAGE_COST, cost)
+ # Span Attributes for Content
if capture_message_content:
- prompt = ""
- completion = ""
+ scope._span.set_attribute(SemanticConvention.GEN_AI_CONTENT_PROMPT, prompt)
+ scope._span.set_attribute(SemanticConvention.GEN_AI_CONTENT_COMPLETION, completion)
- for output in scope._response:
- prompt += output.prompt + "\n"
- if output.outputs and len(output.outputs) > 0:
- completion += output.outputs[0].text + "\n"
- input_tokens += general_tokens(output.prompt)
- output_tokens += general_tokens(output.outputs[0].text)
-
- # Add a single event for prompt
+ # To be removed once the change to span_attributes (from span events) is complete
scope._span.add_event(
name=SemanticConvention.GEN_AI_CONTENT_PROMPT_EVENT,
attributes={
SemanticConvention.GEN_AI_CONTENT_PROMPT: prompt,
},
)
-
- # Add a single event for completion
scope._span.add_event(
name=SemanticConvention.GEN_AI_CONTENT_COMPLETION_EVENT,
attributes={
@@ -102,39 +106,14 @@ def common_chat_logic(scope, pricing_info, environment, application_name, metric
},
)
- scope._span.set_attribute(SemanticConvention.GEN_AI_USAGE_INPUT_TOKENS,
- input_tokens)
- scope._span.set_attribute(SemanticConvention.GEN_AI_USAGE_OUTPUT_TOKENS,
- output_tokens)
- scope._span.set_attribute(SemanticConvention.GEN_AI_CLIENT_TOKEN_USAGE,
- input_tokens + output_tokens)
-
- # Calculate cost of the operation
- cost = get_chat_model_cost(scope._request_model, pricing_info, input_tokens, output_tokens)
- scope._span.set_attribute(SemanticConvention.GEN_AI_USAGE_COST, cost)
-
scope._span.set_status(Status(StatusCode.OK))
- if disable_metrics is False:
- metrics_attributes = create_metrics_attributes(
- service_name=application_name,
- deployment_environment=environment,
- operation=SemanticConvention.GEN_AI_OPERATION_TYPE_CHAT,
- system=SemanticConvention.GEN_AI_SYSTEM_VLLM,
- request_model=scope._request_model,
- server_address=scope._server_address,
- server_port=scope._server_port,
- response_model=scope._request_model,
- )
- metrics['genai_client_operation_duration'].record(scope._end_time - scope._start_time, metrics_attributes)
- metrics['genai_server_tbt'].record(scope._tbt, metrics_attributes)
- metrics['genai_server_ttft'].record(scope._ttft, metrics_attributes)
- metrics['genai_requests'].add(1, metrics_attributes)
- metrics['genai_completion_tokens'].add(output_tokens, metrics_attributes)
- metrics['genai_prompt_tokens'].add(input_tokens, metrics_attributes)
- metrics['genai_cost'].record(cost, metrics_attributes)
- metrics['genai_client_usage_tokens'].record(
- input_tokens + output_tokens, metrics_attributes)
+ # Metrics
+ if not disable_metrics:
+ record_completion_metrics(metrics, SemanticConvention.GEN_AI_OPERATION_TYPE_CHAT, SemanticConvention.GEN_AI_SYSTEM_VLLM,
+ scope._server_address, scope._server_port, request_model, request_model, environment,
+ application_name, scope._start_time, scope._end_time, input_tokens, output_tokens,
+ cost, scope._tbt, scope._ttft)
def process_chat_response(instance, response, request_model, pricing_info, server_port, server_address,
environment, application_name, metrics, start_time, span, args, kwargs,
@@ -142,20 +121,23 @@ def process_chat_response(instance, response, request_model, pricing_info, serve
"""
Process chat request and generate Telemetry
"""
- self = type('GenericScope', (), {})()
- self._response = response
- self._start_time = start_time
- self._end_time = time.time()
- self._span = span
- self._ttft, self._tbt = self._end_time - self._start_time, 0
- self._server_address = server_address
- self._server_port = server_port
- self._request_model = request_model
- self._timestamps = []
- self._args = args
- self._kwargs = kwargs
-
- common_chat_logic(self, pricing_info, environment, application_name, metrics,
+
+ # Create scope object
+ scope = type("GenericScope", (), {})()
+
+ scope._response = response
+ scope._start_time = start_time
+ scope._end_time = time.time()
+ scope._span = span
+ scope._ttft, scope._tbt = scope._end_time - scope._start_time, 0
+ scope._server_address = server_address
+ scope._server_port = server_port
+ scope._request_model = request_model
+ scope._timestamps = []
+ scope._args = args
+ scope._kwargs = kwargs
+
+ common_chat_logic(scope, pricing_info, environment, application_name, metrics,
capture_message_content, disable_metrics, version, is_stream=False)
return response
diff --git a/sdk/python/src/openlit/instrumentation/vllm/vllm.py b/sdk/python/src/openlit/instrumentation/vllm/vllm.py
index 0de2cef0..16be3ae8 100644
--- a/sdk/python/src/openlit/instrumentation/vllm/vllm.py
+++ b/sdk/python/src/openlit/instrumentation/vllm/vllm.py
@@ -2,7 +2,6 @@
Module for monitoring vLLM API calls.
"""
-import logging
import time
from opentelemetry.trace import SpanKind
from openlit.__helpers import (
@@ -14,11 +13,8 @@ from openlit.instrumentation.vllm.utils import (
)
from openlit.semcov import SemanticConvention
-# Initialize logger for logging potential issues and operations
-logger = logging.getLogger(__name__)
-
-def generate(version, environment, application_name,
- tracer, pricing_info, capture_message_content, metrics, disable_metrics):
+def generate(version, environment, application_name, tracer, pricing_info,
+ capture_message_content, metrics, disable_metrics):
"""
Generates a telemetry wrapper for GenAI function call
"""
@@ -27,7 +23,6 @@ def generate(version, environment, application_name,
"""
Wraps the GenAI function call.
"""
-
server_address, server_port = set_server_address_and_port(instance, "http://127.0.0.1", 443)
request_model = instance.llm_engine.model_config.model or "facebook/opt-125m"
@@ -56,9 +51,9 @@ def generate(version, environment, application_name,
disable_metrics=disable_metrics,
version=version,
)
+
except Exception as e:
handle_exception(span, e)
- logger.error("Error in trace creation: %s", e)
return response