A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 18:15 UTC
[CI/Build] Pin OpenTelemetry versions and make errors clearer (#7266)
[CI/Build] Pin OpenTelemetry versions and make a availability errors clearer (#7266)
Ronen Schaffer committed
Aug 20, 2024 at 20:02 UTC
2aa00d59ada908cbf58e241fb83b1a9b5c02f71b
4 files changed
+31
-19
.buildkite/test-pipeline.yaml
+4
-4
@@ -111,10 +111,10 @@ steps:
111
commands:
112
- pytest -v -s metrics
113
- "pip install \
114
- opentelemetry-sdk \
115
- opentelemetry-api \
116
- opentelemetry-exporter-otlp \
117
- opentelemetry-semantic-conventions-ai"
114
+ 'opentelemetry-sdk>=1.26.0,<1.27.0' \
115
+ 'opentelemetry-api>=1.26.0,<1.27.0' \
116
+ 'opentelemetry-exporter-otlp>=1.26.0,<1.27.0' \
117
+ 'opentelemetry-semantic-conventions-ai>=0.4.1,<0.5.0'"
118
- pytest -v -s tracing
119
120
##### fast check tests #####
examples/production_monitoring/Otel.md
+4
-4
@@ -3,10 +3,10 @@
3
1. Install OpenTelemetry packages:
4
```
5
pip install \
6
- opentelemetry-sdk \
7
- opentelemetry-api \
8
- opentelemetry-exporter-otlp \
9
- opentelemetry-semantic-conventions-ai
6
+ 'opentelemetry-sdk>=1.26.0,<1.27.0' \
7
+ 'opentelemetry-api>=1.26.0,<1.27.0' \
8
+ 'opentelemetry-exporter-otlp>=1.26.0,<1.27.0' \
9
+ 'opentelemetry-semantic-conventions-ai>=0.4.1,<0.5.0'
10
```
11
12
1. Start Jaeger in a docker container:
vllm/config.py
+6
-4
@@ -12,7 +12,7 @@ from vllm.logger import init_logger
12
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
13
from vllm.model_executor.models import ModelRegistry
14
from vllm.platforms import current_platform
15
-from vllm.tracing import is_otel_installed
15
+from vllm.tracing import is_otel_available, otel_import_error_traceback
16
from vllm.transformers_utils.config import get_config, get_hf_text_config
17
from vllm.utils import (STR_NOT_IMPL_ENC_DEC_CUDAGRAPH, GiB_bytes,
18
cuda_device_count_stateless, get_cpu_memory, is_cpu,
@@ -1721,9 +1721,11 @@ class ObservabilityConfig:
1721
collect_model_execute_time: bool = False
1722
1723
def __post_init__(self):
1724
- if not is_otel_installed() and self.otlp_traces_endpoint is not None:
1725
- raise ValueError("OpenTelemetry packages must be installed before "
1726
- "configuring 'otlp_traces_endpoint'")
1724
+ if not is_otel_available() and self.otlp_traces_endpoint is not None:
1725
+ raise ValueError(
1726
+ "OpenTelemetry is not available. Unable to configure "
1727
+ "'otlp_traces_endpoint'. Ensure OpenTelemetry packages are "
1728
+ f"installed. Original error:\n{otel_import_error_traceback}")
1729
1730
if ((self.collect_model_forward_time
1731
or self.collect_model_execute_time)
vllm/tracing.py
+17
-7
@@ -8,7 +8,8 @@ TRACE_HEADERS = ["traceparent", "tracestate"]
8
9
logger = init_logger(__name__)
10
11
-_is_otel_installed = False
11
+_is_otel_imported = False
12
+otel_import_error_traceback: Optional[str] = None
13
try:
14
from opentelemetry.context.context import Context
15
from opentelemetry.sdk.environment_variables import (
@@ -19,8 +20,14 @@ try:
20
from opentelemetry.trace import SpanKind, Tracer, set_tracer_provider
21
from opentelemetry.trace.propagation.tracecontext import (
22
TraceContextTextMapPropagator)
22
- _is_otel_installed = True
23
+ _is_otel_imported = True
24
except ImportError:
25
+ # Capture and format traceback to provide detailed context for the import
26
+ # error. Only the string representation of the error is retained to avoid
27
+ # memory leaks.
28
+ # See https://github.com/vllm-project/vllm/pull/7266#discussion_r1707395458
29
+ import traceback
30
+ otel_import_error_traceback = traceback.format_exc()
31
32
class Context: # type: ignore
33
pass
@@ -35,14 +42,17 @@ except ImportError:
42
pass
43
44
38
-def is_otel_installed() -> bool:
39
- return _is_otel_installed
45
+def is_otel_available() -> bool:
46
+ return _is_otel_imported
47
48
49
def init_tracer(instrumenting_module_name: str,
50
otlp_traces_endpoint: str) -> Optional[Tracer]:
44
- assert is_otel_installed(), ("OpenTelemetry packages must be installed "
45
- "prior to initializing a tracer")
51
+ if not is_otel_available():
52
+ raise ValueError(
53
+ "OpenTelemetry is not available. Unable to initialize "
54
+ "a tracer. Ensure OpenTelemetry packages are installed. "
55
+ f"Original error:\n{otel_import_error_traceback}")
56
trace_provider = TracerProvider()
57
58
span_exporter = get_span_exporter(otlp_traces_endpoint)
@@ -70,7 +80,7 @@ def get_span_exporter(endpoint):
80
81
def extract_trace_context(
82
headers: Optional[Mapping[str, str]]) -> Optional[Context]:
73
- if is_otel_installed():
83
+ if is_otel_available():
84
headers = headers or {}
85
return TraceContextTextMapPropagator().extract(headers)
86
else: