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 Instrumentation (#759)

Patcherabc1e68db921
diff --git a/sdk/python/pyproject.toml b/sdk/python/pyproject.toml
index c5cea232..4f460328 100644
--- a/sdk/python/pyproject.toml
+++ b/sdk/python/pyproject.toml
@@ -1,6 +1,6 @@
[tool.poetry]
name = "openlit"
-version = "1.34.1"
+version = "1.34.2"
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/google_ai_studio/async_google_ai_studio.py b/sdk/python/src/openlit/instrumentation/google_ai_studio/async_google_ai_studio.py
index 9ae62fb3..9a574ba0 100644
--- a/sdk/python/src/openlit/instrumentation/google_ai_studio/async_google_ai_studio.py
+++ b/sdk/python/src/openlit/instrumentation/google_ai_studio/async_google_ai_studio.py
@@ -59,14 +59,12 @@ def async_generate(version, environment, application_name,
version=version,
)
- return response
-
except Exception as e:
handle_exception(span, e)
logger.error("Error in trace creation: %s", e)
- # Return original response
- return response
+ # Return original response
+ return response
return wrapper
diff --git a/sdk/python/src/openlit/instrumentation/google_ai_studio/google_ai_studio.py b/sdk/python/src/openlit/instrumentation/google_ai_studio/google_ai_studio.py
index a82413ec..2c6c8b15 100644
--- a/sdk/python/src/openlit/instrumentation/google_ai_studio/google_ai_studio.py
+++ b/sdk/python/src/openlit/instrumentation/google_ai_studio/google_ai_studio.py
@@ -59,14 +59,12 @@ def generate(version, environment, application_name,
version=version,
)
- return response
-
except Exception as e:
handle_exception(span, e)
logger.error("Error in trace creation: %s", e)
- # Return original response
- return response
+ # Return original response
+ return response
return wrapper
diff --git a/sdk/python/src/openlit/instrumentation/google_ai_studio/utils.py b/sdk/python/src/openlit/instrumentation/google_ai_studio/utils.py
index 4e31a2fe..cad85156 100644
--- a/sdk/python/src/openlit/instrumentation/google_ai_studio/utils.py
+++ b/sdk/python/src/openlit/instrumentation/google_ai_studio/utils.py
@@ -2,10 +2,8 @@
Google AI Studio 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_ttft,
response_as_dict,
@@ -117,6 +115,7 @@ def common_chat_logic(scope, pricing_info, environment, application_name, metric
scope._span.set_attribute(SemanticConvention.GEN_AI_SYSTEM, SemanticConvention.GEN_AI_SYSTEM_GEMINI)
scope._span.set_attribute(SemanticConvention.GEN_AI_REQUEST_MODEL, request_model)
scope._span.set_attribute(SemanticConvention.SERVER_PORT, scope._server_port)
+ scope._span.set_attribute(SemanticConvention.SERVER_ADDRESS, scope._server_address)
inference_config = scope._kwargs.get('config', {})
@@ -142,7 +141,6 @@ def common_chat_logic(scope, pricing_info, environment, application_name, metric
scope._span.set_attribute(SemanticConvention.GEN_AI_USAGE_INPUT_TOKENS, scope._input_tokens)
scope._span.set_attribute(SemanticConvention.GEN_AI_USAGE_OUTPUT_TOKENS, scope._output_tokens)
scope._span.set_attribute(SemanticConvention.GEN_AI_USAGE_REASONING_TOKENS, scope._reasoning_tokens)
- scope._span.set_attribute(SemanticConvention.SERVER_ADDRESS, scope._server_address)
scope._span.set_attribute(SemanticConvention.GEN_AI_OUTPUT_TYPE,
'text' if isinstance(scope._llmresponse, str) else 'json')
diff --git a/sdk/python/src/openlit/instrumentation/vllm/__init__.py b/sdk/python/src/openlit/instrumentation/vllm/__init__.py
index acd3b083..9cbb8569 100644
--- a/sdk/python/src/openlit/instrumentation/vllm/__init__.py
+++ b/sdk/python/src/openlit/instrumentation/vllm/__init__.py
@@ -32,7 +32,7 @@ class VLLMInstrumentor(BaseInstrumentor):
# sync chat
wrap_function_wrapper(
- "vllm",
+ "vllm.entrypoints.llm",
"LLM.generate",
generate(version, environment, application_name,
tracer, pricing_info, capture_message_content, metrics, disable_metrics),
diff --git a/sdk/python/src/openlit/instrumentation/vllm/utils.py b/sdk/python/src/openlit/instrumentation/vllm/utils.py
new file mode 100644
index 00000000..6b5e99a9
--- /dev/null
+++ b/sdk/python/src/openlit/instrumentation/vllm/utils.py
@@ -0,0 +1,161 @@
+"""
+Utility functions for vLLM instrumentation.
+"""
+
+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,
+)
+from openlit.semcov import SemanticConvention
+
+def get_inference_config(args, kwargs):
+ """
+ Safely extract inference configuration from args or kwargs.
+ """
+
+ if 'sampling_params' in kwargs:
+ return kwargs['sampling_params']
+ if len(args) > 1:
+ return args[1]
+ return None
+
+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
+ 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_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
+
+ if capture_message_content:
+ 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)
+
+ # Add a single event for prompt
+ 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={
+ SemanticConvention.GEN_AI_CONTENT_COMPLETION: completion,
+ },
+ )
+
+ 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)
+
+def process_chat_response(instance, response, request_model, pricing_info, server_port, server_address,
+ environment, application_name, metrics, start_time, span, args, kwargs,
+ capture_message_content=False, disable_metrics=False, version="1.0.0"):
+ """
+ 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,
+ 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 71f9701b..0de2cef0 100644
--- a/sdk/python/src/openlit/instrumentation/vllm/vllm.py
+++ b/sdk/python/src/openlit/instrumentation/vllm/vllm.py
@@ -4,170 +4,62 @@ Module for monitoring vLLM API calls.
import logging
import time
-from opentelemetry.trace import SpanKind, Status, StatusCode
-from opentelemetry.sdk.resources import SERVICE_NAME, TELEMETRY_SDK_NAME, DEPLOYMENT_ENVIRONMENT
+from opentelemetry.trace import SpanKind
from openlit.__helpers import (
- get_chat_model_cost,
handle_exception,
- general_tokens,
- create_metrics_attributes,
set_server_address_and_port
)
+from openlit.instrumentation.vllm.utils import (
+ process_chat_response
+)
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):
+ tracer, pricing_info, capture_message_content, metrics, disable_metrics):
"""
- Generates a telemetry wrapper for generate to collect metrics.
-
- Args:
- version: Version of the monitoring package.
- environment: Deployment environment (e.g., production, staging).
- application_name: Name of the application using the vLLM API.
- tracer: OpenTelemetry tracer for creating spans.
- pricing_info: Information used for calculating the cost of vLLM usage.
- capture_message_content: Flag indicating whether to trace the actual content.
-
- Returns:
- A function that wraps the generate method to add telemetry.
+ Generates a telemetry wrapper for GenAI function call
"""
def wrapper(wrapped, instance, args, kwargs):
"""
- Wraps the 'generate' API call to add telemetry.
-
- This collects metrics such as execution time, cost, and token usage, and handles errors
- gracefully, adding details to the trace for observability.
-
- Args:
- wrapped: The original 'generate' method to be wrapped.
- instance: The instance of the class where the original method is defined.
- args: Positional arguments for the 'generate' method.
- kwargs: Keyword arguments for the 'generate' method.
-
- Returns:
- The response from the original 'generate' method.
+ Wraps the GenAI function call.
"""
- server_address, server_port = set_server_address_and_port(instance, "api.cohere.com", 443)
+ 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"
span_name = f"{SemanticConvention.GEN_AI_OPERATION_TYPE_CHAT} {request_model}"
- # pylint: disable=line-too-long
- with tracer.start_as_current_span(span_name, kind= SpanKind.CLIENT) as span:
+ with tracer.start_as_current_span(span_name, kind=SpanKind.CLIENT) as span:
start_time = time.time()
response = wrapped(*args, **kwargs)
- end_time = time.time()
try:
- # Set base span attribues
- span.set_attribute(TELEMETRY_SDK_NAME, "openlit")
- span.set_attribute(SemanticConvention.GEN_AI_SYSTEM,
- SemanticConvention.GEN_AI_SYSTEM_VLLM)
- span.set_attribute(SemanticConvention.GEN_AI_OPERATION,
- SemanticConvention.GEN_AI_OPERATION_TYPE_CHAT)
- span.set_attribute(SemanticConvention.SERVER_PORT,
- server_port)
- span.set_attribute(SemanticConvention.GEN_AI_REQUEST_MODEL,
- request_model)
- span.set_attribute(SemanticConvention.GEN_AI_RESPONSE_MODEL,
- request_model)
- span.set_attribute(SemanticConvention.SERVER_ADDRESS,
- server_address)
- span.set_attribute(SemanticConvention.GEN_AI_OUTPUT_TYPE,
- "text")
-
- # Set base span attribues (Extras)
- span.set_attribute(DEPLOYMENT_ENVIRONMENT,
- environment)
- span.set_attribute(SERVICE_NAME,
- application_name)
- span.set_attribute(SemanticConvention.GEN_AI_REQUEST_IS_STREAM,
- False)
- span.set_attribute(SemanticConvention.GEN_AI_SERVER_TTFT,
- end_time - start_time)
- span.set_attribute(SemanticConvention.GEN_AI_SDK_VERSION,
- version)
-
- input_tokens = 0
- output_tokens = 0
- cost = 0
-
- if capture_message_content:
- prompt_attributes = {}
- completion_attributes = {}
-
- for i, output in enumerate(response):
- prompt_attributes[f"{SemanticConvention.GEN_AI_CONTENT_PROMPT}.{i}"] = output.prompt
- completion_attributes[f"{SemanticConvention.GEN_AI_CONTENT_COMPLETION}.{i}"] = output.outputs[0].text
- input_tokens += general_tokens(output.prompt)
- output_tokens += general_tokens(output.outputs[0].text)
-
- # Add a single event for all prompts
- span.add_event(
- name=SemanticConvention.GEN_AI_CONTENT_PROMPT_EVENT,
- attributes=prompt_attributes,
- )
-
- # Add a single event for all completions
- span.add_event(
- name=SemanticConvention.GEN_AI_CONTENT_COMPLETION_EVENT,
- attributes=completion_attributes,
- )
-
- span.set_attribute(SemanticConvention.GEN_AI_USAGE_INPUT_TOKENS,
- input_tokens)
- span.set_attribute(SemanticConvention.GEN_AI_USAGE_OUTPUT_TOKENS,
- output_tokens)
- span.set_attribute(SemanticConvention.GEN_AI_USAGE_TOTAL_TOKENS,
- input_tokens + output_tokens)
-
- # Calculate cost of the operation
- cost = get_chat_model_cost(request_model, pricing_info,
- input_tokens, output_tokens)
- span.set_attribute(SemanticConvention.GEN_AI_USAGE_COST,
- cost)
-
- span.set_status(Status(StatusCode.OK))
-
- if disable_metrics is False:
- 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=request_model,
- server_address=server_address,
- server_port=server_port,
- response_model=request_model,
- )
-
- metrics["genai_client_usage_tokens"].record(
- input_tokens + output_tokens, attributes
- )
- metrics["genai_client_operation_duration"].record(
- end_time - start_time, attributes
- )
- metrics["genai_server_ttft"].record(
- end_time - start_time, attributes
- )
- metrics["genai_requests"].add(1, attributes)
- metrics["genai_completion_tokens"].add(output_tokens, attributes)
- metrics["genai_prompt_tokens"].add(input_tokens, attributes)
- metrics["genai_cost"].record(cost, attributes)
-
- # Return original response
- return response
-
+ response = process_chat_response(
+ instance=instance,
+ response=response,
+ request_model=request_model,
+ pricing_info=pricing_info,
+ server_port=server_port,
+ server_address=server_address,
+ environment=environment,
+ application_name=application_name,
+ metrics=metrics,
+ start_time=start_time,
+ span=span,
+ args=args,
+ kwargs=kwargs,
+ capture_message_content=capture_message_content,
+ disable_metrics=disable_metrics,
+ version=version,
+ )
except Exception as e:
handle_exception(span, e)
logger.error("Error in trace creation: %s", e)
- # Return original response
- return response
+ return response
return wrapper