Skip to content

vllm.tracing

TRACE_HEADERS module-attribute 

TRACE_HEADERS = ['traceparent', 'tracestate']

_is_otel_imported module-attribute 

_is_otel_imported = True

logger module-attribute 

logger = init_logger(__name__)

otel_import_error_traceback module-attribute 

otel_import_error_traceback: str | None = None

BaseSpanAttributes

Source code in vllm/tracing.py 
41
42
class BaseSpanAttributes:  # type: ignore
    pass

SpanAttributes

Source code in vllm/tracing.py 
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
class SpanAttributes:
    # Attribute names copied from here to avoid version conflicts:
    # https://github.com/open-telemetry/semantic-conventions/blob/main/docs/gen-ai/gen-ai-spans.md
    GEN_AI_USAGE_COMPLETION_TOKENS = "gen_ai.usage.completion_tokens"
    GEN_AI_USAGE_PROMPT_TOKENS = "gen_ai.usage.prompt_tokens"
    GEN_AI_REQUEST_MAX_TOKENS = "gen_ai.request.max_tokens"
    GEN_AI_REQUEST_TOP_P = "gen_ai.request.top_p"
    GEN_AI_REQUEST_TEMPERATURE = "gen_ai.request.temperature"
    GEN_AI_RESPONSE_MODEL = "gen_ai.response.model"
    # Attribute names added until they are added to the semantic conventions:
    GEN_AI_REQUEST_ID = "gen_ai.request.id"
    GEN_AI_REQUEST_N = "gen_ai.request.n"
    GEN_AI_USAGE_NUM_SEQUENCES = "gen_ai.usage.num_sequences"
    GEN_AI_LATENCY_TIME_IN_QUEUE = "gen_ai.latency.time_in_queue"
    GEN_AI_LATENCY_TIME_TO_FIRST_TOKEN = "gen_ai.latency.time_to_first_token"
    GEN_AI_LATENCY_E2E = "gen_ai.latency.e2e"
    GEN_AI_LATENCY_TIME_IN_SCHEDULER = "gen_ai.latency.time_in_scheduler"
    # Time taken in the forward pass for this across all workers
    GEN_AI_LATENCY_TIME_IN_MODEL_FORWARD = "gen_ai.latency.time_in_model_forward"
    # Time taken in the model execute function. This will include model
    # forward, block/sync across workers, cpu-gpu sync time and sampling time.
    GEN_AI_LATENCY_TIME_IN_MODEL_EXECUTE = "gen_ai.latency.time_in_model_execute"
    GEN_AI_LATENCY_TIME_IN_MODEL_PREFILL = "gen_ai.latency.time_in_model_prefill"
    GEN_AI_LATENCY_TIME_IN_MODEL_DECODE = "gen_ai.latency.time_in_model_decode"
    GEN_AI_LATENCY_TIME_IN_MODEL_INFERENCE = "gen_ai.latency.time_in_model_inference"

GEN_AI_LATENCY_E2E class-attribute instance-attribute 

GEN_AI_LATENCY_E2E = 'gen_ai.latency.e2e'

GEN_AI_LATENCY_TIME_IN_MODEL_DECODE class-attribute instance-attribute 

GEN_AI_LATENCY_TIME_IN_MODEL_DECODE = (
    "gen_ai.latency.time_in_model_decode"
)

GEN_AI_LATENCY_TIME_IN_MODEL_EXECUTE class-attribute instance-attribute 

GEN_AI_LATENCY_TIME_IN_MODEL_EXECUTE = (
    "gen_ai.latency.time_in_model_execute"
)

GEN_AI_LATENCY_TIME_IN_MODEL_FORWARD class-attribute instance-attribute 

GEN_AI_LATENCY_TIME_IN_MODEL_FORWARD = (
    "gen_ai.latency.time_in_model_forward"
)

GEN_AI_LATENCY_TIME_IN_MODEL_INFERENCE class-attribute instance-attribute 

GEN_AI_LATENCY_TIME_IN_MODEL_INFERENCE = (
    "gen_ai.latency.time_in_model_inference"
)

GEN_AI_LATENCY_TIME_IN_MODEL_PREFILL class-attribute instance-attribute 

GEN_AI_LATENCY_TIME_IN_MODEL_PREFILL = (
    "gen_ai.latency.time_in_model_prefill"
)

GEN_AI_LATENCY_TIME_IN_QUEUE class-attribute instance-attribute 

GEN_AI_LATENCY_TIME_IN_QUEUE = (
    "gen_ai.latency.time_in_queue"
)

GEN_AI_LATENCY_TIME_IN_SCHEDULER class-attribute instance-attribute 

GEN_AI_LATENCY_TIME_IN_SCHEDULER = (
    "gen_ai.latency.time_in_scheduler"
)

GEN_AI_LATENCY_TIME_TO_FIRST_TOKEN class-attribute instance-attribute 

GEN_AI_LATENCY_TIME_TO_FIRST_TOKEN = (
    "gen_ai.latency.time_to_first_token"
)

GEN_AI_REQUEST_ID class-attribute instance-attribute 

GEN_AI_REQUEST_ID = 'gen_ai.request.id'

GEN_AI_REQUEST_MAX_TOKENS class-attribute instance-attribute 

GEN_AI_REQUEST_MAX_TOKENS = 'gen_ai.request.max_tokens'

GEN_AI_REQUEST_N class-attribute instance-attribute 

GEN_AI_REQUEST_N = 'gen_ai.request.n'

GEN_AI_REQUEST_TEMPERATURE class-attribute instance-attribute 

GEN_AI_REQUEST_TEMPERATURE = 'gen_ai.request.temperature'

GEN_AI_REQUEST_TOP_P class-attribute instance-attribute 

GEN_AI_REQUEST_TOP_P = 'gen_ai.request.top_p'

GEN_AI_RESPONSE_MODEL class-attribute instance-attribute 

GEN_AI_RESPONSE_MODEL = 'gen_ai.response.model'

GEN_AI_USAGE_COMPLETION_TOKENS class-attribute instance-attribute 

GEN_AI_USAGE_COMPLETION_TOKENS = (
    "gen_ai.usage.completion_tokens"
)

GEN_AI_USAGE_NUM_SEQUENCES class-attribute instance-attribute 

GEN_AI_USAGE_NUM_SEQUENCES = 'gen_ai.usage.num_sequences'

GEN_AI_USAGE_PROMPT_TOKENS class-attribute instance-attribute 

GEN_AI_USAGE_PROMPT_TOKENS = 'gen_ai.usage.prompt_tokens'

contains_trace_headers 

contains_trace_headers(headers: Mapping[str, str]) -> bool
Source code in vllm/tracing.py 
129
130
def contains_trace_headers(headers: Mapping[str, str]) -> bool:
    return any(h in headers for h in TRACE_HEADERS)

extract_trace_context 

extract_trace_context(
    headers: Mapping[str, str] | None,
) -> Context | None
Source code in vllm/tracing.py 
90
91
92
93
94
95
def extract_trace_context(headers: Mapping[str, str] | None) -> Context | None:
    if is_otel_available():
        headers = headers or {}
        return TraceContextTextMapPropagator().extract(headers)
    else:
        return None

extract_trace_headers 

extract_trace_headers(
    headers: Mapping[str, str],
) -> Mapping[str, str]
Source code in vllm/tracing.py 
98
99
def extract_trace_headers(headers: Mapping[str, str]) -> Mapping[str, str]:
    return {h: headers[h] for h in TRACE_HEADERS if h in headers}

get_span_exporter 

get_span_exporter(endpoint)
Source code in vllm/tracing.py 
74
75
76
77
78
79
80
81
82
83
84
85
86
87
def get_span_exporter(endpoint):
    protocol = os.environ.get(OTEL_EXPORTER_OTLP_TRACES_PROTOCOL, "grpc")
    if protocol == "grpc":
        from opentelemetry.exporter.otlp.proto.grpc.trace_exporter import (
            OTLPSpanExporter,
        )
    elif protocol == "http/protobuf":
        from opentelemetry.exporter.otlp.proto.http.trace_exporter import (
            OTLPSpanExporter,  # type: ignore
        )
    else:
        raise ValueError(f"Unsupported OTLP protocol '{protocol}' is configured")

    return OTLPSpanExporter(endpoint=endpoint)

init_tracer 

init_tracer(
    instrumenting_module_name: str,
    otlp_traces_endpoint: str,
) -> Tracer | None
Source code in vllm/tracing.py 
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
def init_tracer(
    instrumenting_module_name: str, otlp_traces_endpoint: str
) -> Tracer | None:
    if not is_otel_available():
        raise ValueError(
            "OpenTelemetry is not available. Unable to initialize "
            "a tracer. Ensure OpenTelemetry packages are installed. "
            f"Original error:\n{otel_import_error_traceback}"
        )
    trace_provider = TracerProvider()

    span_exporter = get_span_exporter(otlp_traces_endpoint)
    trace_provider.add_span_processor(BatchSpanProcessor(span_exporter))
    set_tracer_provider(trace_provider)

    tracer = trace_provider.get_tracer(instrumenting_module_name)
    return tracer

is_otel_available 

is_otel_available() -> bool
Source code in vllm/tracing.py 
51
52
def is_otel_available() -> bool:
    return _is_otel_imported

log_tracing_disabled_warning 

log_tracing_disabled_warning() -> None
Source code in vllm/tracing.py 
133
134
135
@run_once
def log_tracing_disabled_warning() -> None:
    logger.warning("Received a request with trace context but tracing is disabled")