scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-11-12 08:42:18 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 207

Merge pull request #32553 from BerriAI/litellm_backport_patch_192rc1

Browse source 
chore(release): backport #32256, #32277, #32236, #32416, #32405, #32524 onto patch-1.92.0rc1
This commit is contained in:
yuneng-jiang 2026-07-08 16:20:09 -07:00 • committed by GitHub
commit d2b4f495ea
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
21 changed files with 437 additions and 100 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

4
Dockerfile
View file

@ -1,10 +1,10 @@
# syntax=docker/dockerfile:1.7
# Base image for building
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:c61ac6919b811ea53c4782d69f1fe05218ba3c25d53f01b6ab7892e621bd4370
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
# Runtime image
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:c61ac6919b811ea53c4782d69f1fe05218ba3c25d53f01b6ab7892e621bd4370
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
# Pinned by digest like the other base images; bump explicitly on Node upgrades.
ARG UI_BUILD_IMAGE=node:20.18-alpine3.20@sha256:3488b10bf958af7125a176419d2d8a9937d895bf124012aae811651988d2ffe6

4
backend/Dockerfile
View file

@ -1,5 +1,5 @@
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:c61ac6919b811ea53c4782d69f1fe05218ba3c25d53f01b6ab7892e621bd4370
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:c61ac6919b811ea53c4782d69f1fe05218ba3c25d53f01b6ab7892e621bd4370
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
FROM $UV_IMAGE AS uvbin

4
docker/Dockerfile.database
View file

@ -1,10 +1,10 @@
# syntax=docker/dockerfile:1.7
# Base image for building
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:c61ac6919b811ea53c4782d69f1fe05218ba3c25d53f01b6ab7892e621bd4370
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
# Runtime image
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:c61ac6919b811ea53c4782d69f1fe05218ba3c25d53f01b6ab7892e621bd4370
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
# Pinned by digest like the other base images; bump explicitly on Node upgrades.
ARG UI_BUILD_IMAGE=node:20.18-alpine3.20@sha256:3488b10bf958af7125a176419d2d8a9937d895bf124012aae811651988d2ffe6

4
docker/Dockerfile.non_root
View file

@ -1,8 +1,8 @@
# syntax=docker/dockerfile:1.7
# Base images
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:c61ac6919b811ea53c4782d69f1fe05218ba3c25d53f01b6ab7892e621bd4370
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:c61ac6919b811ea53c4782d69f1fe05218ba3c25d53f01b6ab7892e621bd4370
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG PROXY_EXTRAS_SOURCE=published
ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
# Pinned by digest like the other base images; bump explicitly on Node upgrades.

4
gateway/Dockerfile
View file

@ -1,5 +1,5 @@
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:c61ac6919b811ea53c4782d69f1fe05218ba3c25d53f01b6ab7892e621bd4370
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:c61ac6919b811ea53c4782d69f1fe05218ba3c25d53f01b6ab7892e621bd4370
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
FROM $UV_IMAGE AS uvbin

2
litellm/integrations/otel/__init__.py
View file

@ -52,6 +52,7 @@ from litellm.integrations.otel.model.semconv import (
GenAIProvider,
JsonRpc,
LiteLLM,
LiteLLMError,
MCPMethod,
Metric,
Network,
@ -87,6 +88,7 @@ __all__ = [
"HTTP",
"JsonRpc",
"LiteLLM",
"LiteLLMError",
"MCP",
"MCPMethod",
"Metric",

35
litellm/integrations/otel/emitter.py
View file

@ -16,9 +16,10 @@ from litellm.integrations.otel.model.payloads import (
MCPListToolsSpanData,
MCPToolCallSpanData,
ServiceSpanData,
SpanError,
)
from litellm.integrations.otel.plumbing.providers import to_otel_span_kind
from litellm.integrations.otel.model.semconv import Error, ExceptionEvent
from litellm.integrations.otel.model.semconv import Error, ExceptionEvent, LiteLLMError
from litellm.integrations.otel.model.spans import (
SPAN_REGISTRY,
SpanRole,
@ -49,6 +50,27 @@ _NAME_BUILDERS: dict[SpanRole, Callable[..., str]] = {
_DEDUP_CACHE_MAX = 10_000
def _stamp_otel_error_attributes(span: Span, error_type: str, resolved_message: str) -> None:
"""Stamp the OTel-semconv error attributes (``error.type`` + ``error.message``).
``error_type`` and ``resolved_message`` are ``finish_span``'s already-computed
fallback chains, so the pair on the status, event, and attributes stays in
lockstep."""
span.set_attribute(Error.TYPE, error_type)
span.set_attribute(Error.MESSAGE, resolved_message)
def _stamp_litellm_error_attributes(span: Span, error: SpanError) -> None:
"""Stamp litellm-specific error detail attributes. Emitted only when the
corresponding field is populated so guardrail-shape errors carrying only a
message aren't polluted with empty detail keys."""
if error.code:
span.set_attribute(LiteLLMError.CODE, error.code)
if error.stack_trace:
span.set_attribute(LiteLLMError.STACK_TRACE, error.stack_trace)
if error.llm_provider:
span.set_attribute(LiteLLMError.LLM_PROVIDER, error.llm_provider)
class SpanEmitter:
def __init__(
self,
@ -190,12 +212,13 @@ class SpanEmitter:
if error and (error.error_type or error.message):
error_type = error.error_type or "error"
message = error.message or error.error_type or "error"
span.set_attribute(Error.TYPE, error_type)
_stamp_otel_error_attributes(span, error_type, message)
_stamp_litellm_error_attributes(span, error)
span.set_status(Status(StatusCode.ERROR, message))
# Carry the full message on the standard ``exception`` event so backends
# map it as full text under ``exception.message``. Setting it as a bare
# string attribute instead lets backends like Elasticsearch dynamic-map
# it to a ``keyword`` capped at 1024 chars, truncating the message.
# Also emit the semconv ``exception`` event so backends that
# dynamic-map unknown string span attrs to ``keyword`` (e.g.
# Elasticsearch with a 1024-char ``ignore_above``) still see the
# full untruncated message on the recognized event field.
span.add_event(
ExceptionEvent.NAME,
{ExceptionEvent.TYPE: error_type, ExceptionEvent.MESSAGE: message},

6
litellm/integrations/otel/logger.py
View file

@ -368,7 +368,11 @@ class OpenTelemetryV2(CustomLogger):
# it (named provisionally) so it isn't leaked as an open span.
carrier.span.end(end_time=to_ns(end_time))
return None
data = LLMCallSpanData.from_standard_logging_payload(payload, capture_content=self.config.capture_span_content)
data = LLMCallSpanData.from_standard_logging_payload(
payload,
capture_content=self.config.capture_span_content,
time_to_first_chunk_seconds=call.time_to_first_chunk_seconds,
)
end_time_ns = to_ns(end_time)
if carrier.span is not None:
# Born at the boundary: stamp attributes from the typed payload, set

1
litellm/integrations/otel/mappers/genai.py
View file

@ -55,6 +55,7 @@ class GenAIMapper:
GenAI.RESPONSE_MODEL: lambda d: d.response_model,
GenAI.RESPONSE_ID: lambda d: d.response_id,
GenAI.RESPONSE_FINISH_REASONS: lambda d: list(d.finish_reasons) if d.finish_reasons else None,
GenAI.RESPONSE_TIME_TO_FIRST_CHUNK: lambda d: d.time_to_first_chunk_seconds,
GenAI.USAGE_INPUT_TOKENS: lambda d: d.usage.input_tokens,
GenAI.USAGE_OUTPUT_TOKENS: lambda d: d.usage.output_tokens,
Error.TYPE: lambda d: d.error.error_type if d.error else None,

19
litellm/integrations/otel/model/metadata.py
View file

@ -41,7 +41,7 @@ from typing import TYPE_CHECKING, Any, Mapping, cast
from litellm.constants import LITELLM_LOGGING_NO_UPSTREAM_LLM_CALL
from litellm.integrations.otel.model.semconv import resolve_operation
from litellm.integrations.otel.model.utils import as_str
from litellm.integrations.otel.model.utils import as_str, to_seconds
if TYPE_CHECKING:
from litellm.types.utils import StandardLoggingPayload
@ -201,6 +201,7 @@ class LLMCallEvent:
# span is renamed from the typed payload at close (``finish_span``); this only
# needs to be reasonable for a span that never gets closed (a leak).
provisional_span_name: str
time_to_first_chunk_seconds: float | None
@classmethod
def from_dict(cls, kwargs: Mapping[str, Any]) -> "LLMCallEvent":
@ -214,9 +215,25 @@ class LLMCallEvent:
dynamic_params=kwargs.get("standard_callback_dynamic_params"),
is_no_upstream_call=bool(kwargs.get(LITELLM_LOGGING_NO_UPSTREAM_LLM_CALL)),
provisional_span_name=f"{operation.value} {model}".strip(),
time_to_first_chunk_seconds=time_to_first_chunk_seconds(kwargs),
)
def time_to_first_chunk_seconds(kwargs: Mapping[str, Any]) -> float | None:
"""Seconds from the upstream request being issued (``api_call_start_time``)
to the first streamed chunk (``completion_start_time``); ``None`` for
non-streaming calls, where ``completion_start_time`` is backfilled with the
end time and would not measure first-chunk latency."""
optional_params = cast(Mapping[str, Any], kwargs.get("optional_params") or {})
if not optional_params.get("stream"):
return None
api_call_start = to_seconds(kwargs.get("api_call_start_time"))
completion_start = to_seconds(kwargs.get("completion_start_time"))
if api_call_start is None or completion_start is None:
return None
return completion_start - api_call_start
def _call_id(payload: "StandardLoggingPayload | None", kwargs: Mapping[str, Any]) -> str | None:
"""The call id from the payload (when closed) or the bare kwargs (at pre_call)."""
if payload is not None:

13
litellm/integrations/otel/model/payloads.py
View file

@ -141,6 +141,9 @@ class LLMCost:
class SpanError:
error_type: str | None = None
message: str | None = None
code: str | None = None
stack_trace: str | None = None
llm_provider: str | None = None
@dataclass(frozen=True)
@ -305,10 +308,14 @@ class LLMCallSpanData:
messages_in: tuple[Mapping[str, object], ...] = ()
choices_out: tuple[Mapping[str, object], ...] = ()
system_fingerprint: str | None = None
time_to_first_chunk_seconds: float | None = None
@classmethod
def from_standard_logging_payload(
cls, payload: "StandardLoggingPayload", capture_content: bool = False
cls,
payload: "StandardLoggingPayload",
capture_content: bool = False,
time_to_first_chunk_seconds: float | None = None,
) -> "LLMCallSpanData":
params = cast(Mapping[str, object], payload.get("model_parameters") or {})
# The single parse of the request's metadata — the request-vs-provider
@ -349,6 +356,7 @@ class LLMCallSpanData:
messages_in=_dicts(payload.get("messages")) if capture_content else (),
choices_out=choices_out if capture_content else (),
system_fingerprint=as_str(response.get("system_fingerprint")),
time_to_first_chunk_seconds=time_to_first_chunk_seconds,
)
@ -566,6 +574,9 @@ def _parse_error(payload: "StandardLoggingPayload") -> SpanError | None:
return SpanError(
error_type=as_str(info.get("error_class")) or as_str(info.get("error_code")),
message=as_str(info.get("error_message")) or as_str(payload.get("error_str")),
code=as_str(info.get("error_code")),
stack_trace=as_str(info.get("traceback")),
llm_provider=as_str(info.get("llm_provider")),
)

21
litellm/integrations/otel/model/semconv.py
View file

@ -69,6 +69,7 @@ class GenAI:
RESPONSE_ID: Final = "gen_ai.response.id"
RESPONSE_MODEL: Final = "gen_ai.response.model"
RESPONSE_FINISH_REASONS: Final = "gen_ai.response.finish_reasons"
RESPONSE_TIME_TO_FIRST_CHUNK: Final = "gen_ai.response.time_to_first_chunk"
# usage
USAGE_INPUT_TOKENS: Final = "gen_ai.usage.input_tokens"
USAGE_OUTPUT_TOKENS: Final = "gen_ai.usage.output_tokens"
@ -143,7 +144,27 @@ class Client:
class Error:
"""OTel-defined error attribute keys, from the semconv ``error.*`` registry.
``MESSAGE`` is marked *Deprecated* upstream in favor of domain-specific
error message keys plus ``exception.message`` on the exception event, but
is still defined and stamped by litellm's v1 integration; keeping it here
for byte-for-byte parity."""
TYPE: Final = "error.type"
MESSAGE: Final = "error.message"
class LiteLLMError:
"""LiteLLM-specific error attribute keys. Emitted under the ``error.*``
namespace (not ``litellm.*``) for byte-for-byte compat with the v1
integration in ``opentelemetry.py``; consumers reading these keys on v1
spans read the same keys on v2 spans. OTel semconv does not define any of
these three, and per its extension rules a namespace may carry additional
vendor keys as long as they don't collide with defined names."""
CODE: Final = "error.code"
STACK_TRACE: Final = "error.stack_trace"
LLM_PROVIDER: Final = "error.llm_provider"
class ExceptionEvent:

10
litellm/integrations/otel/plumbing/metrics.py
View file

@ -21,6 +21,7 @@ from litellm.integrations.opentelemetry import (
_build_metric_attribute_filter,
_resolve_metric_attribute_filter,
)
from litellm.integrations.otel.model.metadata import time_to_first_chunk_seconds
from litellm.integrations.otel.model.semconv import Metric, resolve_operation
from litellm.integrations.otel.model.utils import to_seconds
from litellm.litellm_core_utils.safe_json_dumps import safe_dumps
@ -181,13 +182,10 @@ class GenAIMetricRecorder:
self._metrics.token_usage.record(usage.get("completion_tokens", 0), attributes=out_attrs)
def _record_time_to_first_token(self, kwargs: Mapping[str, Any], common_attrs: dict) -> None:
if not kwargs.get("optional_params", {}).get("stream", False):
time_to_first_chunk = time_to_first_chunk_seconds(kwargs)
if time_to_first_chunk is None:
return
api_call_start = to_seconds(kwargs.get("api_call_start_time"))
completion_start = to_seconds(kwargs.get("completion_start_time"))
if api_call_start is None or completion_start is None:
return
self._metrics.time_to_first_token.record(completion_start - api_call_start, attributes=common_attrs)
self._metrics.time_to_first_token.record(time_to_first_chunk, attributes=common_attrs)
def _record_time_per_output_token(
self,

46
litellm/proxy/proxy_server.py
View file

@ -1117,33 +1117,6 @@ _OPENAPI_HTTP_METHODS = {
# `_SSO_SENSITIVE_FIELDS` / `_CACHE_SENSITIVE_FIELDS` constants in the SSO
# and cache endpoint files.
_ALERTING_SENSITIVE_VARS: Set[str] = {"SLACK_WEBHOOK_URL", "SMTP_PASSWORD"}
_DB_LITELLM_PARAM_ENV_REF_KEYS = frozenset(
{
"api_key",
"client_secret",
"vertex_credentials",
"vertex_ai_credentials",
"aws_access_key_id",
"aws_secret_access_key",
}
)
def _db_model_is_team_scoped(model: object) -> bool:
model_info = getattr(model, "model_info", None)
if isinstance(model_info, BaseModel):
return getattr(model_info, "team_id", None) is not None
if isinstance(model_info, str):
try:
model_info = json.loads(model_info)
except (TypeError, ValueError):
model_info = None
if isinstance(model_info, dict) and model_info.get("team_id") is not None:
return True
if getattr(model_info, "team_id", None) is not None:
return True
model_name = getattr(model, "model_name", None)
return isinstance(model_name, str) and model_name.startswith("model_name_")
def _strip_operation_id_method_suffix(operation_id: str) -> str:
@ -4972,17 +4945,12 @@ class ProxyConfig:
deleted_deployments += 1
return deleted_deployments
def _resolve_db_litellm_param(self, key: str, value: object, resolve_env_refs: bool = True) -> object:
def _resolve_db_litellm_param(self, key: str, value: object) -> object:
if not isinstance(value, str):
return value
decrypted_value = decrypt_value_helper(value=value, key=key, return_original_value=True)
if (
resolve_env_refs
and key in _DB_LITELLM_PARAM_ENV_REF_KEYS
and isinstance(decrypted_value, str)
and decrypted_value.startswith("os.environ/")
):
if isinstance(decrypted_value, str) and decrypted_value.startswith("os.environ/"):
return get_secret(decrypted_value)
return decrypted_value
@ -5003,13 +4971,10 @@ class ProxyConfig:
## ADD MODEL LOGIC
for m in db_models:
_litellm_params = m.litellm_params
resolve_env_refs = not _db_model_is_team_scoped(m)
if isinstance(_litellm_params, dict):
# decrypt values
for k, v in _litellm_params.items():
_litellm_params[k] = self._resolve_db_litellm_param(
key=k, value=v, resolve_env_refs=resolve_env_refs
)
_litellm_params[k] = self._resolve_db_litellm_param(key=k, value=v)
_litellm_params = LiteLLM_Params(**_litellm_params)
else:
@ -5035,15 +5000,12 @@ class ProxyConfig:
_model_list: list = []
for m in new_models:
_litellm_params = m.litellm_params
resolve_env_refs = not _db_model_is_team_scoped(m)
if isinstance(_litellm_params, BaseModel):
_litellm_params = _litellm_params.model_dump()
if isinstance(_litellm_params, dict):
# decrypt values
for k, v in _litellm_params.items():
_litellm_params[k] = self._resolve_db_litellm_param(
key=k, value=v, resolve_env_refs=resolve_env_refs
)
_litellm_params[k] = self._resolve_db_litellm_param(key=k, value=v)
_litellm_params = LiteLLM_Params(**_litellm_params)
else:
verbose_proxy_logger.error(

4
migrations/Dockerfile
View file

@ -1,5 +1,5 @@
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:c61ac6919b811ea53c4782d69f1fe05218ba3c25d53f01b6ab7892e621bd4370
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:c61ac6919b811ea53c4782d69f1fe05218ba3c25d53f01b6ab7892e621bd4370
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
FROM $UV_IMAGE AS uvbin

110
tests/test_litellm/integrations/otel/test_otel_v2_components.py
View file

@ -579,13 +579,10 @@ def _exception_event(span):
def test_error_message_recorded_as_full_exception_event_untruncated():
"""Regression for the Elasticsearch keyword/ignore_above:1024 truncation.
A long error message must survive intact on the standard ``exception``
event under ``exception.message`` — not get dropped onto a bare string
attribute that backends dynamic-map to a 1024-char ``keyword``. The SDK
must not truncate it either, so a 5000-char message stays 5000 chars.
"""
"""The ``exception`` event carries the full untruncated message under
``exception.message`` so backends that dynamic-map unknown string span
attrs to ``keyword`` (e.g. Elasticsearch with a 1024-char ``ignore_above``)
still see it in full via the semconv-recognized event field."""
from litellm.integrations.otel.model.semconv import Error, ExceptionEvent
long_message = "boom: " + "x" * 5000
@ -596,13 +593,108 @@ def test_error_message_recorded_as_full_exception_event_untruncated():
assert len(event.attributes[ExceptionEvent.MESSAGE]) == len(long_message) > 1024
assert event.attributes[ExceptionEvent.TYPE] == "litellm.APIError"
# error.type stays a low-cardinality attribute; the message does NOT become a
# bare string attribute (which is what got truncated).
# error.type stays a low-cardinality attribute; the exception EVENT field
# ``exception.message`` never becomes a bare string attribute.
assert span.attributes[Error.TYPE] == "litellm.APIError"
assert ExceptionEvent.MESSAGE not in span.attributes
assert span.status.description == long_message
def test_error_details_stamped_as_span_attributes_for_labels_ingest():
"""OTel-defined keys and litellm-specific detail keys both ride span
attributes so backends that flatten attrs into label indexes (Elastic APM
``labels.*``, Datadog span tags) render them. The exception event with the
full untruncated message stays alongside — both places, matching v1's
shape."""
from litellm.integrations.otel.model.semconv import Error, ExceptionEvent, LiteLLMError
from litellm.integrations.otel.emitter import SpanEmitter
cfg = OpenTelemetryV2Config(exporter="in_memory")
provider, exporter = providers.in_memory_provider(cfg)
engine = SpanEmitter(providers.get_tracer(provider, "t"), cfg)
data = LLMCallSpanData(
operation=GenAIOperation.CHAT,
provider="openai",
request_model="gpt-4o",
response_model=None,
response_id=None,
request_params=LLMRequestParams(),
usage=LLMUsage(),
finish_reasons=(),
error=SpanError(
error_type="litellm.BadRequestError",
message="400: violated moderation policy",
code="400",
stack_trace="File proxy_server.py line 8570 ...",
llm_provider="openai",
),
response_cost=None,
server=None,
identity=RequestIdentity(call_id=None),
)
engine.emit(SpanRole.LLM_CALL, data)
(span,) = exporter.get_finished_spans()
# OTel-defined keys (from the ``error.*`` semconv registry).
assert span.attributes[Error.TYPE] == "litellm.BadRequestError"
assert span.attributes[Error.MESSAGE] == "400: violated moderation policy"
# LiteLLM-specific detail keys — vendor-namespaced under ``error.*``
# for v1-parity, not defined by OTel semconv.
assert span.attributes[LiteLLMError.CODE] == "400"
assert span.attributes[LiteLLMError.STACK_TRACE] == "File proxy_server.py line 8570 ..."
assert span.attributes[LiteLLMError.LLM_PROVIDER] == "openai"
# The exception event carries the same message on the span too.
event = _exception_event(span)
assert event.attributes[ExceptionEvent.MESSAGE] == "400: violated moderation policy"
def test_error_details_omitted_when_span_error_carries_only_message():
"""A guardrail-shape error (message only, no code/traceback/provider) must
not pollute the span with empty-string detail attributes. Only the keys
that carry real data land."""
from litellm.integrations.otel.model.semconv import Error, LiteLLMError
span = _emit_error_span("guardrail rejected", error_type="ContentFilter")
assert span.attributes[Error.TYPE] == "ContentFilter"
assert span.attributes[Error.MESSAGE] == "guardrail rejected"
# LiteLLM-specific detail keys aren't stamped when the SpanError doesn't
# carry them.
assert LiteLLMError.CODE not in span.attributes
assert LiteLLMError.STACK_TRACE not in span.attributes
assert LiteLLMError.LLM_PROVIDER not in span.attributes
def test_v2_error_attribute_keys_match_v1_error_attributes_byte_for_byte():
"""v1 (``opentelemetry.py``) and v2 (``otel/`` package) stamp identical
span-attribute keys so consumers reading ``labels.error_message`` don't
care which integration produced the span. Renaming either side is a
breaking change for downstream dashboards; this test locks the vocabulary."""
from litellm.integrations._types.open_inference import ErrorAttributes
from litellm.integrations.otel.model.semconv import Error, LiteLLMError
assert Error.TYPE == ErrorAttributes.ERROR_TYPE
assert Error.MESSAGE == ErrorAttributes.ERROR_MESSAGE
assert LiteLLMError.CODE == ErrorAttributes.ERROR_CODE
assert LiteLLMError.STACK_TRACE == ErrorAttributes.ERROR_STACK_TRACE
assert LiteLLMError.LLM_PROVIDER == ErrorAttributes.ERROR_LLM_PROVIDER
def test_error_message_falls_back_to_error_type_when_message_absent():
"""A ``SpanError(error_type=..., message=None)`` still renders on the span:
the resolved message is the error_type, and it lands on ``error.message``,
the exception event, and the span-status description in lockstep so a
single-source-of-truth view isn't inconsistent."""
from litellm.integrations.otel.model.semconv import Error, ExceptionEvent
span = _emit_error_span(message=None, error_type="RateLimitError")
assert span.attributes[Error.MESSAGE] == "RateLimitError"
assert _exception_event(span).attributes[ExceptionEvent.MESSAGE] == "RateLimitError"
assert span.status.description == "RateLimitError"
def test_success_span_records_no_exception_event():
from litellm.integrations.otel.emitter import SpanEmitter
from litellm.integrations.otel.model.semconv import ExceptionEvent

37
tests/test_litellm/integrations/otel/test_otel_v2_logger.py
View file

@ -168,6 +168,43 @@ def test_async_log_success_event_emits_llm_call_span():
assert span.status.status_code is StatusCode.UNSET
def test_streaming_span_carries_time_to_first_chunk():
logger, exporter = _logger()
kwargs = {
**_kwargs(payload=_payload(stream=True)),
"optional_params": {"stream": True},
"api_call_start_time": datetime(2026, 5, 26, 12, 0, 0, tzinfo=timezone.utc),
"completion_start_time": datetime(2026, 5, 26, 12, 0, 0, 750000, tzinfo=timezone.utc),
}
_emit_llm(logger, kwargs)
(span,) = exporter.get_finished_spans()
assert span.attributes[GenAI.RESPONSE_TIME_TO_FIRST_CHUNK] == pytest.approx(0.75)
def test_non_streaming_span_has_no_time_to_first_chunk():
logger, exporter = _logger()
kwargs = {
**_kwargs(),
"optional_params": {},
"api_call_start_time": datetime(2026, 5, 26, 12, 0, 0, tzinfo=timezone.utc),
"completion_start_time": datetime(2026, 5, 26, 12, 0, 5, tzinfo=timezone.utc),
}
_emit_llm(logger, kwargs)
(span,) = exporter.get_finished_spans()
assert GenAI.RESPONSE_TIME_TO_FIRST_CHUNK not in span.attributes
def test_streaming_span_without_timing_omits_time_to_first_chunk():
logger, exporter = _logger()
kwargs = {
**_kwargs(payload=_payload(stream=True)),
"optional_params": {"stream": True},
}
_emit_llm(logger, kwargs)
(span,) = exporter.get_finished_spans()
assert GenAI.RESPONSE_TIME_TO_FIRST_CHUNK not in span.attributes
def test_async_log_failure_event_marks_error_status():
logger, exporter = _logger()
payload = _payload(

47
tests/test_litellm/integrations/otel/test_otel_v2_sources_of_truth.py
View file

@ -144,11 +144,13 @@ def _all_constants(cls):
def test_attribute_keys_are_unique_across_namespaces():
from litellm.integrations.otel import MCP, Client, JsonRpc, Network
from litellm.integrations.otel import MCP, Client, JsonRpc, LiteLLMError, Network
# prefixes are allowed to be substrings; exact keys must not collide.
# ``LiteLLMError`` shares the ``error.*`` prefix with ``Error`` by design
# (v1-parity); the assert below is the guarantee they never overlap.
exact = set()
for cls in (GenAI, Error, Server, HTTP, DB, MCP, JsonRpc, Network, Client):
for cls in (GenAI, Error, LiteLLMError, Server, HTTP, DB, MCP, JsonRpc, Network, Client):
for key in _all_constants(cls):
assert key not in exact, f"duplicate attribute key {key}"
exact.add(key)
@ -342,6 +344,47 @@ def test_llm_call_adapter_failure_path():
assert data.error.message == "429 slow down"
def test_llm_call_adapter_carries_error_detail_fields():
"""``_parse_error`` threads the full detail set from ``error_information``
(``error_code``, ``traceback``, ``llm_provider``) onto ``SpanError`` so the
emitter can stamp them as span attributes."""
payload = _sample_payload(
status="failure",
error_information={
"error_class": "BadRequestError",
"error_message": "400 violated moderation policy",
"error_code": "400",
"traceback": "File proxy_server.py line 8570 ...",
"llm_provider": "openai",
},
)
data = LLMCallSpanData.from_standard_logging_payload(payload)
assert data.error is not None
assert data.error.error_type == "BadRequestError"
assert data.error.message == "400 violated moderation policy"
assert data.error.code == "400"
assert data.error.stack_trace == "File proxy_server.py line 8570 ..."
assert data.error.llm_provider == "openai"
def test_llm_call_adapter_error_details_default_to_none_when_absent():
"""Guardrail-shape payloads carry only ``error_class`` + ``error_message``.
The detail fields must stay ``None`` so the emitter's ``if error.code:``
guards skip stamping empty attributes."""
payload = _sample_payload(
status="failure",
error_information={
"error_class": "ContentFilter",
"error_message": "guardrail rejected",
},
)
data = LLMCallSpanData.from_standard_logging_payload(payload)
assert data.error is not None
assert data.error.code is None
assert data.error.stack_trace is None
assert data.error.llm_provider is None
def test_adapter_is_resilient_to_minimal_payload():
data = LLMCallSpanData.from_standard_logging_payload({})
assert data.request_model == ""

140
tests/test_litellm/proxy/proxy_server/test_proxy_config.py
View file

@ -952,6 +952,11 @@ def test_ProxyConfig__add_deployment_invalid_litellm_params_skips(monkeypatch):
def test_ProxyConfig__add_deployment_resolves_env_refs_after_db_decrypt(monkeypatch):
"""Every ``os.environ/`` value on an admin-scoped DB row resolves at
load time, regardless of the field name. Replaces the earlier
behavior where only fields in ``_DB_LITELLM_PARAM_ENV_REF_KEYS``
resolved: the whitelist has been removed so the resolver applies to
every string field."""
monkeypatch.setenv("LITELLM_DB_MODEL_API_KEY", "resolved-secret")
monkeypatch.setenv("LITELLM_MASTER_KEY", "master-secret")
monkeypatch.setattr(
@ -979,19 +984,21 @@ def test_ProxyConfig__add_deployment_resolves_env_refs_after_db_decrypt(monkeypa
assert added == 1
assert deployment.litellm_params.api_key == "resolved-secret"
assert deployment.litellm_params.api_base == "os.environ/LITELLM_MASTER_KEY"
assert deployment.litellm_params.api_base == "master-secret"
def test_ProxyConfig__add_deployment_keeps_team_env_refs_literal(monkeypatch):
def fail_on_call(secret_name, *args, **kwargs):
raise AssertionError("team DB models should not resolve env refs")
def test_ProxyConfig__add_deployment_resolves_team_env_refs(monkeypatch):
"""Team-scoped DB rows now resolve ``os.environ/`` refs the same way
admin rows do. The prior team-scoped short-circuit and the
field-by-field whitelist have both been removed; the write-side team
auth check in ``ModelManagementAuthChecks.can_user_make_model_call``
remains the single trust boundary. A literal (non-``os.environ/``)
value still passes through unchanged."""
monkeypatch.setenv("LITELLM_MASTER_KEY", "master-secret")
monkeypatch.setattr(
"litellm.proxy.proxy_server.decrypt_value_helper",
lambda value, key, return_original_value: value,
)
monkeypatch.setattr("litellm.proxy.proxy_server.get_secret", fail_on_call)
fake_router = MagicMock()
fake_router.upsert_deployment = MagicMock(return_value=True)
monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", fake_router)
@ -1003,7 +1010,7 @@ def test_ProxyConfig__add_deployment_keeps_team_env_refs_literal(monkeypatch):
litellm_params={
"model": "openai/gpt-4o-mini",
"api_key": "os.environ/LITELLM_MASTER_KEY",
"api_base": "https://attacker.example",
"api_base": "https://team.example",
},
blocked=False,
)
@ -1012,8 +1019,8 @@ def test_ProxyConfig__add_deployment_keeps_team_env_refs_literal(monkeypatch):
deployment = fake_router.upsert_deployment.call_args.kwargs["deployment"]
assert added == 1
assert deployment.litellm_params.api_key == "os.environ/LITELLM_MASTER_KEY"
assert deployment.litellm_params.api_base == "https://attacker.example"
assert deployment.litellm_params.api_key == "master-secret"
assert deployment.litellm_params.api_base == "https://team.example"
def test_ProxyConfig__resolve_db_litellm_param_skips_non_string_values(monkeypatch):
@ -1029,6 +1036,100 @@ def test_ProxyConfig__resolve_db_litellm_param_skips_non_string_values(monkeypat
assert pc._resolve_db_litellm_param(key="tpm", value=100) == 100
def test_ProxyConfig__add_deployment_resolves_env_refs_for_aws_bedrock_auth_params(
monkeypatch,
):
"""Regression: DB-stored Bedrock/SageMaker auth params like
``aws_role_name: os.environ/BEDROCK_ASSUME_ROLE_ARN`` must resolve at
DB-load time. PR #30867 removed request-time expansion in
``BaseAWSLLM.get_credentials``; without DB-load resolution the literal
string reaches STS and fails with ``ValidationError: ... is invalid``."""
aws_env = {
"aws_session_token": ("BEDROCK_SESSION_TOKEN", "resolved-session-token"),
"aws_region_name": ("BEDROCK_REGION", "us-east-1"),
"aws_session_name": ("BEDROCK_SESSION_NAME", "resolved-session"),
"aws_profile_name": ("BEDROCK_PROFILE", "resolved-profile"),
"aws_role_name": (
"BEDROCK_ASSUME_ROLE_ARN",
"arn:aws:iam::123456789012:role/resolved",
),
"aws_web_identity_token": ("BEDROCK_WEB_IDENTITY_TOKEN", "resolved-token"),
"aws_sts_endpoint": (
"BEDROCK_STS_ENDPOINT",
"https://sts.us-east-1.amazonaws.com",
),
"aws_external_id": ("BEDROCK_EXTERNAL_ID", "resolved-external-id"),
"aws_bedrock_runtime_endpoint": (
"BEDROCK_RUNTIME_ENDPOINT",
"https://bedrock-runtime.us-east-1.amazonaws.com",
),
"aws_bedrock_project_id": ("BEDROCK_PROJECT_ID", "resolved-project-id"),
"aws_batch_role_arn": (
"BEDROCK_BATCH_ROLE_ARN",
"arn:aws:iam::123456789012:role/batch",
),
"aws_workspace_id": ("BEDROCK_WORKSPACE_ID", "resolved-workspace-id"),
}
for _, (env_name, env_value) in aws_env.items():
monkeypatch.setenv(env_name, env_value)
monkeypatch.setattr(
"litellm.proxy.proxy_server.decrypt_value_helper",
lambda value, key, return_original_value: value,
)
fake_router = MagicMock()
fake_router.upsert_deployment = MagicMock(return_value=True)
monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", fake_router)
pc = ProxyConfig()
litellm_params: Dict[str, Any] = {"model": "bedrock/anthropic.claude-v2"}
for key, (env_name, _) in aws_env.items():
litellm_params[key] = f"os.environ/{env_name}"
db_model = SimpleNamespace(
model_id="model-1",
model_name="bedrock-model",
model_info={"id": "model-1"},
litellm_params=litellm_params,
blocked=False,
)
added = pc._add_deployment(db_models=[db_model])
deployment = fake_router.upsert_deployment.call_args.kwargs["deployment"]
assert added == 1
for key, (_, expected) in aws_env.items():
assert getattr(deployment.litellm_params, key) == expected, key
def test_ProxyConfig__add_deployment_resolves_env_refs_on_arbitrary_field(monkeypatch):
"""A made-up field name that was never on the removed whitelist still
resolves ``os.environ/`` refs. Pins the "no whitelist" invariant:
the resolver applies to every string field, not a curated list."""
monkeypatch.setenv("SOME_CUSTOM_ENV", "resolved-custom-value")
monkeypatch.setattr(
"litellm.proxy.proxy_server.decrypt_value_helper",
lambda value, key, return_original_value: value,
)
fake_router = MagicMock()
fake_router.upsert_deployment = MagicMock(return_value=True)
monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", fake_router)
pc = ProxyConfig()
db_model = SimpleNamespace(
model_id="model-1",
model_name="custom-field-model",
model_info={"id": "model-1"},
litellm_params={
"model": "openai/gpt-4o-mini",
"some_future_field": "os.environ/SOME_CUSTOM_ENV",
},
blocked=False,
)
added = pc._add_deployment(db_models=[db_model])
deployment = fake_router.upsert_deployment.call_args.kwargs["deployment"]
assert added == 1
assert deployment.litellm_params.some_future_field == "resolved-custom-value"
# ---------------------------------------------------------------------------
# ProxyConfig.decrypt_model_list_from_db
# ---------------------------------------------------------------------------
@ -1064,6 +1165,9 @@ def test_ProxyConfig_decrypt_model_list_from_db_returns_decrypted(monkeypatch):
def test_ProxyConfig_decrypt_model_list_from_db_resolves_env_refs_after_db_decrypt(
monkeypatch,
):
"""Path B (feeding /v2/model/info fallback and /model/info fallback)
resolves every ``os.environ/`` field on admin-scoped rows, mirroring
path A. Both paths now share the same universal-resolution shape."""
monkeypatch.setenv("LITELLM_DB_MODEL_API_KEY", "resolved-secret")
monkeypatch.setenv("LITELLM_MASTER_KEY", "master-secret")
monkeypatch.setattr(
@ -1090,15 +1194,16 @@ def test_ProxyConfig_decrypt_model_list_from_db_resolves_env_refs_after_db_decry
out = pc.decrypt_model_list_from_db(new_models=[m])
assert out[0]["litellm_params"]["api_key"] == "resolved-secret"
assert out[0]["litellm_params"]["api_base"] == "os.environ/LITELLM_MASTER_KEY"
assert out[0]["litellm_params"]["api_base"] == "master-secret"
def test_ProxyConfig_decrypt_model_list_from_db_keeps_team_env_refs_literal_after_db_decrypt(
def test_ProxyConfig_decrypt_model_list_from_db_resolves_team_env_refs_after_db_decrypt(
monkeypatch,
):
def fail_on_call(secret_name, *args, **kwargs):
raise AssertionError("team DB models should not resolve env refs")
"""Team-scoped rows on path B resolve ``os.environ/`` refs just like
admin rows do. Pairs with
``test_ProxyConfig__add_deployment_resolves_team_env_refs`` on path
A — both paths now agree on the trust model."""
monkeypatch.setenv("LITELLM_MASTER_KEY", "master-secret")
monkeypatch.setattr(
"litellm.proxy.proxy_server.decrypt_value_helper",
@ -1106,7 +1211,6 @@ def test_ProxyConfig_decrypt_model_list_from_db_keeps_team_env_refs_literal_afte
"os.environ/LITELLM_MASTER_KEY" if key == "api_key" else value
),
)
monkeypatch.setattr("litellm.proxy.proxy_server.get_secret", fail_on_call)
pc = ProxyConfig()
m = SimpleNamespace(
model_id="model-1",
@ -1114,7 +1218,7 @@ def test_ProxyConfig_decrypt_model_list_from_db_keeps_team_env_refs_literal_afte
model_info={"id": "model-1", "team_id": "team-1"},
litellm_params={
"api_key": "encrypted-env-ref",
"api_base": "https://attacker.example",
"api_base": "https://team.example",
"model": "openai/gpt-4o-mini",
},
blocked=False,
@ -1122,8 +1226,8 @@ def test_ProxyConfig_decrypt_model_list_from_db_keeps_team_env_refs_literal_afte
out = pc.decrypt_model_list_from_db(new_models=[m])
assert out[0]["litellm_params"]["api_key"] == "os.environ/LITELLM_MASTER_KEY"
assert out[0]["litellm_params"]["api_base"] == "https://attacker.example"
assert out[0]["litellm_params"]["api_key"] == "master-secret"
assert out[0]["litellm_params"]["api_base"] == "https://team.example"
def test_ProxyConfig_decrypt_model_list_from_db_invalid_params_skips():

16
ui/litellm-dashboard/src/components/mcp_tools/create_mcp_server.test.tsx
View file

@ -937,6 +937,22 @@ describe("CreateMCPServer", () => {
const reopenedUrlInput = screen.getByPlaceholderText("https://your-mcp-server.com") as HTMLInputElement;
expect(reopenedUrlInput.value).toBe("");
});
it("does not reset an in-flight OAuth resume when mounted with the modal closed (post-redirect restore)", () => {
// After the "Authorize & Fetch Token" redirect the page reloads and this
// component mounts with isModalVisible=false while useMcpOAuthFlow is still
// exchanging the authorization code. Calling reset() during that mount bumps
// the hook's reset version and the fetched token is silently discarded, so
// the user sees no Connection Status / Tool Configuration and must authorize
// again after saving.
const { rerender } = render(<CreateMCPServer {...defaultProps} isModalVisible={false} />);
expect(oauthHook.reset).not.toHaveBeenCalled();
// A real open -> closed transition must still reset (the #30000 leak fix).
rerender(<CreateMCPServer {...defaultProps} isModalVisible={true} />);
rerender(<CreateMCPServer {...defaultProps} isModalVisible={false} />);
expect(oauthHook.reset).toHaveBeenCalled();
});
});
describe("when stdio transport is selected", () => {

10
ui/litellm-dashboard/src/components/mcp_tools/create_mcp_server.tsx
View file

@ -603,9 +603,15 @@ const CreateMCPServer: React.FC<CreateMCPServerProps> = ({
// Clear form, tools, and OAuth state when the modal closes so a previous server's
// authorization, credentials, or tool list never bleed into the next "Add New MCP
// Server" session, including when a parent dismisses the modal without routing
// through handleCancel or handleCreate.
// through handleCancel or handleCreate. Only a real open -> closed transition may
// trigger this: on the post-OAuth-redirect remount the modal starts closed while
// resumeOAuthFlow's token exchange is in flight, and resetting then discards the
// fetched token.
const wasModalVisibleRef = React.useRef(isModalVisible);
React.useEffect(() => {
if (!isModalVisible) {
const wasVisible = wasModalVisibleRef.current;
wasModalVisibleRef.current = isModalVisible;
if (!isModalVisible && wasVisible) {
form.resetFields();
setFormValues({});
setOauthAccessToken(null);