scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-09-13 23:11:40 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 152

Merge remote-tracking branch 'origin/litellm_internal_staging' into litellm_post_call_pipeline_background_responses

Browse source
This commit is contained in:
mateo-berri 2026-09-08 16:34:42 -07:00
commit 804829049c
217 changed files with 12235 additions and 3503 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
.github/workflows/_test-unit-base.yml vendored
View file

@ -113,7 +113,7 @@ jobs:
if: steps.changes.outputs.decision != 'skip'
timeout-minutes: 8
run: |
.github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra google --extra proxy --extra semantic-router --extra saml --extra mongodb
.github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra google --extra proxy --extra semantic-router --extra saml
uv run --no-sync python -c 'import os, sys; print(sys.version); assert f"{sys.version_info.major}.{sys.version_info.minor}" == os.environ["UV_PYTHON"]'
- name: Cache Prisma binaries

4
.github/workflows/test-litellm-ui-unit.yml vendored
View file

@ -66,7 +66,7 @@ jobs:
BASE_SHA: ${{ github.event.pull_request.base.sha }}
HEAD_SHA: ${{ github.event.pull_request.head.sha }}
run: |
full_suite() { npm run test -- --run --pool forks --poolOptions.forks.maxForks=14; }
full_suite() { npm run test -- --run --pool forks --maxWorkers=14; }
if [ -z "$BASE_SHA" ]; then
echo "Push to $GITHUB_REF_NAME: running the full suite"
@ -95,4 +95,4 @@ jobs:
echo "Pull request: running tests related to ${#changed_files[@]} changed UI files"
npm run test -- related "${changed_files[@]}" --run --passWithNoTests \
--pool forks --poolOptions.forks.maxForks=14
--pool forks --maxWorkers=14

2
Dockerfile
View file

@ -67,7 +67,6 @@ RUN uv sync --frozen --no-install-project --no-install-workspace --no-default-gr
--extra semantic-router \
--extra saml \
--extra bedrock-realtime \
--extra mongodb \
--python python3.13
# Copy full source tree
@ -90,7 +89,6 @@ RUN uv sync --frozen --no-default-groups --no-editable \
--extra semantic-router \
--extra saml \
--extra bedrock-realtime \
--extra mongodb \
--python python3.13
RUN HOME=/opt/prisma XDG_CACHE_HOME=/opt/prisma/.cache PRISMA_BINARY_CACHE_DIR=/opt/prisma/binaries \

2
docker/Dockerfile.database
View file

@ -65,7 +65,6 @@ RUN uv sync --frozen --no-install-project --no-install-workspace --no-default-gr
--extra semantic-router \
--extra saml \
--extra bedrock-realtime \
--extra mongodb \
--python python3.13
# Copy full source tree
@ -88,7 +87,6 @@ RUN uv sync --frozen --no-default-groups --no-editable \
--extra semantic-router \
--extra saml \
--extra bedrock-realtime \
--extra mongodb \
--python python3.13
RUN HOME=/opt/prisma XDG_CACHE_HOME=/opt/prisma/.cache PRISMA_BINARY_CACHE_DIR=/opt/prisma/binaries \

3
docker/Dockerfile.non_root
View file

@ -71,7 +71,6 @@ RUN --mount=type=cache,target=/app/.cache/uv,id=litellm-uv-cache \
--extra semantic-router \
--extra saml \
--extra bedrock-realtime \
--extra mongodb \
--python python3.13
# Copy full source tree
@ -100,7 +99,6 @@ RUN --mount=type=cache,target=/app/.cache/uv,id=litellm-uv-cache \
--extra semantic-router \
--extra saml \
--extra bedrock-realtime \
--extra mongodb \
--python python3.13 \
--no-sources-package litellm-proxy-extras; \
else \
@ -111,7 +109,6 @@ RUN --mount=type=cache,target=/app/.cache/uv,id=litellm-uv-cache \
--extra semantic-router \
--extra saml \
--extra bedrock-realtime \
--extra mongodb \
--python python3.13; \
fi

6
docker/component_entrypoint.sh
View file

@ -1,5 +1,11 @@
#!/bin/sh
# stale samples from a previous container incarnation would be summed into the aggregate
if [ -n "$PROMETHEUS_MULTIPROC_DIR" ]; then
mkdir -p "$PROMETHEUS_MULTIPROC_DIR"
rm -f "$PROMETHEUS_MULTIPROC_DIR"/*.db
fi
case "$USE_DDTRACE" in
[Tt][Rr][Uu][Ee])
export DD_TRACE_OPENAI_ENABLED="False"

56
enterprise/litellm_enterprise/proxy/common_utils/check_batch_cost.py
View file

@ -142,6 +142,40 @@ class CheckBatchCost:
verbose_proxy_logger.error(f"CheckBatchCost: could not look up team alias for team {team_id}: {e}")
return None
async def _get_org_id(self, job: "LiteLLM_ManagedObjectTable", batch_id: str) -> str | None:
org_id = getattr(job, "org_id", None)
if org_id:
return org_id
api_key = getattr(job, "api_key", None)
team_id = getattr(job, "team_id", None)
if api_key:
try:
key_row: prisma_models.LiteLLM_VerificationToken | None = (
await self.prisma_client.db.litellm_verificationtoken.find_unique(
where={"token": api_key}
)
)
key_org_id = getattr(key_row, "organization_id", None) if key_row is not None else None
if key_org_id:
return key_org_id
except Exception as e:
verbose_proxy_logger.error(
f"CheckBatchCost: could not resolve the key's org for batch {batch_id}, "
f"still trying the team's: {e}"
)
if not team_id:
return None
try:
team_row: prisma_models.LiteLLM_TeamTable | None = (
await self.prisma_client.db.litellm_teamtable.find_unique(
where={"team_id": team_id}
)
)
return getattr(team_row, "organization_id", None) if team_row is not None else None
except Exception as e:
verbose_proxy_logger.error(f"CheckBatchCost: could not resolve the team's org for batch {batch_id}: {e}")
return None
async def _build_creator_attribution_metadata(
self, job: "LiteLLM_ManagedObjectTable", batch_id: str
) -> dict[str, object]:
@ -153,6 +187,10 @@ class CheckBatchCost:
user_api_key_alias; when it has no alias, or the key has since been rotated or
deleted, the field keeps the creating user's alias that _get_user_info filled in,
because a resolvable name is more useful on the spend row than a null.
user_api_key_org_id must be resolved here too: the spend update writer reads it
off this metadata to increment organization spend, so leaving it out silently
drops batch cost from org accounting for keys and teams that belong to one.
"""
api_key = getattr(job, "api_key", None)
team_id = getattr(job, "team_id", None)
@ -172,6 +210,9 @@ class CheckBatchCost:
team_alias = await self._get_team_alias(team_id)
if team_alias is not None:
metadata["user_api_key_team_alias"] = team_alias
org_id: Final = await self._get_org_id(job, batch_id)
if org_id is not None:
metadata["user_api_key_org_id"] = org_id
if isinstance(request_tags, list) and request_tags:
metadata["tags"] = [tag for tag in request_tags if isinstance(tag, str)]
@ -641,7 +682,7 @@ class CheckBatchCost:
from litellm.files.main import afile_content
from litellm.litellm_core_utils.get_llm_provider_logic import get_llm_provider
from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLogging
from litellm.litellm_core_utils.litellm_logging import deployment_pricing_model_info
from litellm.litellm_core_utils.litellm_logging import deployment_pricing_model_info, mask_api_base_credentials
from litellm.proxy.openai_files_endpoints.common_utils import (
_is_base64_encoded_unified_file_id,
)
@ -805,6 +846,7 @@ class CheckBatchCost:
function_id=str(uuid.uuid4()),
)
deployment_api_base: Final = deployment_info.litellm_params.api_base
logging_obj.update_environment_variables(
litellm_params={
# set the user-agent header so that S3 callback consumers can easily identify CheckBatchCost callbacks
@ -813,9 +855,17 @@ class CheckBatchCost:
"user-agent": CHECK_BATCH_COST_USER_AGENT,
}
},
"metadata": await self._build_creator_attribution_metadata(job, batch_id),
**({"api_base": mask_api_base_credentials(deployment_api_base)} if deployment_api_base else {}),
"metadata": {
**(await self._build_creator_attribution_metadata(job, batch_id)),
# spend logs read the deployment identity off these metadata keys, so
# without them the batch cost row carries no model_id or model_group
"model_info": {"id": model_id},
"model_group": deployment_info.model_name,
},
},
optional_params={},
custom_llm_provider=str(llm_provider) if llm_provider else None,
)
if not await self._claim_job_for_costing(job):
@ -833,6 +883,8 @@ class CheckBatchCost:
batch_models=batch_result.models,
batch_successful_requests=batch_result.successful_requests,
batch_failed_requests=batch_result.failed_requests,
batch_prompt_cost=batch_result.prompt_cost,
batch_completion_cost=batch_result.completion_cost,
)
except Exception:
await self._release_job_claim(job)

22
enterprise/litellm_enterprise/proxy/hooks/managed_files.py
View file

@ -280,6 +280,27 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
)
verbose_logger.debug(f"LiteLLM Managed File object with id={file_id} stored in db: {result}")
async def _resolve_creator_org_id(self, user_api_key_dict: UserAPIKeyAuth) -> Optional[str]:
if user_api_key_dict.org_id:
return user_api_key_dict.org_id
if not user_api_key_dict.team_id:
return None
from litellm.proxy.auth.auth_checks import get_team_object
from litellm.proxy.proxy_server import proxy_logging_obj, user_api_key_cache
try:
team: Final = await get_team_object(
team_id=user_api_key_dict.team_id,
prisma_client=self.prisma_client,
user_api_key_cache=user_api_key_cache,
parent_otel_span=user_api_key_dict.parent_otel_span,
proxy_logging_obj=proxy_logging_obj,
)
return team.organization_id
except Exception as e:
verbose_logger.warning(f"could not resolve org for managed object attribution: {e}")
return None
async def store_unified_object_id(
self,
unified_object_id: str,
@ -352,6 +373,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
"file_purpose": file_purpose,
"created_by": resolve_resource_owner_id(user_api_key_dict),
"team_id": user_api_key_dict.team_id,
"org_id": await self._resolve_creator_org_id(user_api_key_dict),
"updated_by": user_api_key_dict.user_id,
"status": file_object.status,
**attribution_columns,

2
gateway/Dockerfile
View file

@ -47,7 +47,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
--extra extra_proxy \
--extra semantic-router \
--extra bedrock-realtime \
--extra mongodb \
--python python3.13
# Stage 2 — copy source and install the project + workspace members.
@ -60,7 +59,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
--extra extra_proxy \
--extra semantic-router \
--extra bedrock-realtime \
--extra mongodb \
--python python3.13
RUN HOME=/opt/prisma XDG_CACHE_HOME=/opt/prisma/.cache PRISMA_BINARY_CACHE_DIR=/opt/prisma/binaries \

12
helm/litellm-helm/templates/deployment.yaml
View file

@ -152,6 +152,13 @@ spec:
{{- if .Values.billingMetrics.enabled }}
{{- include "litellm.billingMetricsEnv" . | nindent 12 }}
{{- end }}
{{- if .Values.metricsServer.enabled }}
{{- if eq (int .Values.metricsServer.port) (int .Values.service.port) }}
{{- fail "metricsServer.port must differ from service.port" }}
{{- end }}
- name: PROMETHEUS_METRICS_PORT
value: {{ .Values.metricsServer.port | quote }}
{{- end }}
{{- if .Values.migrationJob.enabled }}
# Schema updates are owned by the dedicated migrations Job; skip
# the proxy's startup `prisma db push` so N replicas don't race
@ -189,6 +196,11 @@ spec:
- name: http
containerPort: {{ .Values.service.port }}
protocol: TCP
{{- if .Values.metricsServer.enabled }}
- name: metrics
containerPort: {{ .Values.metricsServer.port }}
protocol: TCP
{{- end }}
livenessProbe:
httpGet:
path: {{ .Values.livenessProbe.path | quote }}

17
helm/litellm-helm/templates/service-metrics.yaml Normal file
View file

@ -0,0 +1,17 @@
{{- if .Values.metricsServer.enabled }}
apiVersion: v1
kind: Service
metadata:
name: {{ include "litellm.fullname" . }}-metrics
labels:
{{- include "litellm.labels" . | nindent 4 }}
spec:
type: ClusterIP
ports:
- port: {{ .Values.metricsServer.port }}
targetPort: metrics
protocol: TCP
name: metrics
selector:
{{- include "litellm.selectorLabels" . | nindent 4 }}
{{- end }}

2
helm/litellm-helm/templates/servicemonitor.yaml
View file

@ -26,7 +26,7 @@ spec:
{{- toYaml .namespaceSelector.matchNames | nindent 4 }}
{{- end }}
endpoints:
- port: http
- port: {{ ternary "metrics" "http" $.Values.metricsServer.enabled }}
path: /metrics/
interval: {{ .interval }}
scrapeTimeout: {{ .scrapeTimeout }}

106
helm/litellm-helm/tests/metrics_server_tests.yaml Normal file
View file

@ -0,0 +1,106 @@
suite: separate metrics server
templates:
- configmap-litellm.yaml
- deployment.yaml
- service.yaml
- service-metrics.yaml
- servicemonitor.yaml
tests:
- it: should not expose a metrics port or PROMETHEUS_METRICS_PORT by default
asserts:
- notContains:
path: spec.template.spec.containers[0].ports
content:
name: metrics
any: true
template: deployment.yaml
- notContains:
path: spec.template.spec.containers[0].env
content:
name: PROMETHEUS_METRICS_PORT
any: true
template: deployment.yaml
- lengthEqual:
path: spec.ports
count: 1
template: service.yaml
- hasDocuments:
count: 0
template: service-metrics.yaml
- it: should scrape the proxy port when the metrics server is disabled
template: servicemonitor.yaml
set:
serviceMonitor.enabled: true
asserts:
- equal:
path: spec.endpoints[0].port
value: http
- it: should wire the separate metrics server through container, a ClusterIP metrics service and servicemonitor
set:
metricsServer.enabled: true
metricsServer.port: 4101
serviceMonitor.enabled: true
service.type: LoadBalancer
asserts:
- contains:
path: spec.template.spec.containers[0].env
content:
name: PROMETHEUS_METRICS_PORT
value: "4101"
template: deployment.yaml
- contains:
path: spec.template.spec.containers[0].ports
content:
name: metrics
containerPort: 4101
protocol: TCP
template: deployment.yaml
- lengthEqual:
path: spec.ports
count: 1
template: service.yaml
- equal:
path: spec.type
value: LoadBalancer
template: service.yaml
- equal:
path: metadata.name
value: RELEASE-NAME-litellm-metrics
template: service-metrics.yaml
- equal:
path: spec.type
value: ClusterIP
template: service-metrics.yaml
- equal:
path: spec.ports
value:
- port: 4101
targetPort: metrics
protocol: TCP
name: metrics
template: service-metrics.yaml
- equal:
path: spec.selector
value:
app.kubernetes.io/name: litellm
app.kubernetes.io/instance: RELEASE-NAME
template: service-metrics.yaml
- equal:
path: spec.endpoints[0].port
value: metrics
template: servicemonitor.yaml
- equal:
path: spec.endpoints[0].path
value: /metrics/
template: servicemonitor.yaml
- it: should reject a metrics port equal to the proxy port
template: deployment.yaml
set:
metricsServer.enabled: true
metricsServer.port: 4000
asserts:
- failedTemplate:
errorMessage: metricsServer.port must differ from service.port

10
helm/litellm-helm/values.yaml
View file

@ -180,6 +180,16 @@ proxy_config:
general_settings:
master_key: os.environ/PROXY_MASTER_KEY
# Serve Prometheus /metrics from a separate process (PROMETHEUS_METRICS_PORT)
# so a scrape never runs on an inference worker. Adds a `metrics` port to the
# container and a dedicated ClusterIP `<release>-metrics` Service, and the
# ServiceMonitor scrapes it instead of the proxy port. The separate port has
# no virtual-key auth: keep it off public ingress. Needs the proxy image
# v1.101.0 or newer.
metricsServer:
enabled: false
port: 4001
resources:
{}
# Unset by default so the chart installs on small clusters such as Minikube, and so an

2
helm/litellm/templates/_helpers.tpl
View file

@ -441,3 +441,5 @@ ImplementationSpecific
{{- .pathType -}}
{{- end -}}
{{- end -}}
{{- define "litellm.gateway.prometheusMultiprocDir" -}}/tmp/litellm_prometheus_multiproc{{- end -}}

53
helm/litellm/templates/gateway/deployment.yaml
View file

@ -64,14 +64,25 @@ spec:
{{- if .Values.billingMetrics.enabled }}
{{- include "litellm.billingMetricsEnv" . | nindent 12 }}
{{- end }}
{{- if .Values.gateway.metricsServer.enabled }}
{{- if eq (int .Values.gateway.metricsServer.port) 4000 }}
{{- fail "gateway.metricsServer.port must differ from the gateway port 4000" }}
{{- end }}
- name: PROMETHEUS_MULTIPROC_DIR
value: {{ include "litellm.gateway.prometheusMultiprocDir" . }}
{{- end }}
{{- include "litellm.envFrom" .Values.gateway | nindent 10 }}
{{- if or .Values.gateway.config.create .Values.gateway.volumeMounts .Values.billingMetrics.enabled }}
{{- if or .Values.gateway.config.create .Values.gateway.volumeMounts .Values.billingMetrics.enabled .Values.gateway.metricsServer.enabled }}
volumeMounts:
{{- if .Values.gateway.config.create }}
- name: gateway-config
mountPath: /app/config/config.yaml
subPath: config.yaml
{{- end }}
{{- if .Values.gateway.metricsServer.enabled }}
- name: prometheus-multiproc
mountPath: {{ include "litellm.gateway.prometheusMultiprocDir" . }}
{{- end }}
{{- if .Values.billingMetrics.enabled }}
{{- include "litellm.billingMetricsVolumeMounts" . | nindent 12 }}
{{- end }}
@ -97,16 +108,54 @@ spec:
{{- end }}
resources:
{{- toYaml .Values.gateway.resources | nindent 12 }}
{{- if .Values.gateway.metricsServer.enabled }}
- name: metrics
image: "{{ .Values.gateway.image.repository }}:{{ .Values.gateway.image.tag | default .Chart.AppVersion }}"
imagePullPolicy: {{ .Values.gateway.image.pullPolicy }}
{{- with .Values.gateway.securityContext }}
securityContext:
{{- toYaml . | nindent 12 }}
{{- end }}
command:
- python
- -m
- litellm.proxy.prometheus_metrics_server
- --port
- {{ .Values.gateway.metricsServer.port | quote }}
env:
- name: PROMETHEUS_MULTIPROC_DIR
value: {{ include "litellm.gateway.prometheusMultiprocDir" . }}
ports:
- name: metrics
containerPort: {{ .Values.gateway.metricsServer.port }}
protocol: TCP
volumeMounts:
- name: prometheus-multiproc
mountPath: {{ include "litellm.gateway.prometheusMultiprocDir" . }}
readinessProbe:
tcpSocket: { port: metrics }
periodSeconds: 10
livenessProbe:
tcpSocket: { port: metrics }
periodSeconds: 15
failureThreshold: 6
resources:
{{- toYaml .Values.gateway.metricsServer.resources | nindent 12 }}
{{- end }}
{{- with .Values.gateway.extraContainers }}
{{- tpl (toYaml .) $ | nindent 8 }}
{{- end }}
{{- if or .Values.gateway.config.create .Values.gateway.volumes .Values.billingMetrics.enabled }}
{{- if or .Values.gateway.config.create .Values.gateway.volumes .Values.billingMetrics.enabled .Values.gateway.metricsServer.enabled }}
volumes:
{{- if .Values.gateway.config.create }}
- name: gateway-config
configMap:
name: {{ include "litellm.gateway.fullname" . }}-config
{{- end }}
{{- if .Values.gateway.metricsServer.enabled }}
- name: prometheus-multiproc
emptyDir: {}
{{- end }}
{{- if .Values.billingMetrics.enabled }}
{{- include "litellm.billingMetricsVolumes" . | nindent 8 }}
{{- end }}

18
helm/litellm/templates/gateway/service-metrics.yaml Normal file
View file

@ -0,0 +1,18 @@
{{- if and .Values.gateway.enabled .Values.gateway.metricsServer.enabled }}
apiVersion: v1
kind: Service
metadata:
name: {{ include "litellm.gateway.fullname" . }}-metrics
labels:
{{- include "litellm.commonLabels" . | nindent 4 }}
app.kubernetes.io/component: gateway
spec:
type: ClusterIP
ports:
- port: {{ .Values.gateway.metricsServer.port }}
targetPort: metrics
protocol: TCP
name: metrics
selector:
{{- include "litellm.gateway.selectorLabels" . | nindent 4 }}
{{- end }}

148
helm/litellm/tests/metrics_server_tests.yaml Normal file
View file

@ -0,0 +1,148 @@
suite: test gateway metrics sidecar
templates:
- gateway/configmap.yaml
- gateway/deployment.yaml
- gateway/service.yaml
- gateway/service-metrics.yaml
values:
- ./values/required.yaml
tests:
- it: adds no sidecar, volume, env or service port when the metrics server is off
asserts:
- lengthEqual:
path: spec.template.spec.containers
count: 1
template: gateway/deployment.yaml
- notContains:
path: spec.template.spec.containers[0].env
content:
name: PROMETHEUS_MULTIPROC_DIR
any: true
template: gateway/deployment.yaml
- notContains:
path: spec.template.spec.volumes
content:
name: prometheus-multiproc
any: true
template: gateway/deployment.yaml
- lengthEqual:
path: spec.ports
count: 1
template: gateway/service.yaml
- hasDocuments:
count: 0
template: gateway/service-metrics.yaml
- it: runs the metrics server as a sidecar over a shared multiproc dir and exposes it on a ClusterIP metrics service
set:
gateway.metricsServer.enabled: true
gateway.metricsServer.port: 4101
gateway.service.type: LoadBalancer
gateway.image.tag: v1.101.0
asserts:
- contains:
path: spec.template.spec.containers[0].env
content:
name: PROMETHEUS_MULTIPROC_DIR
value: /tmp/litellm_prometheus_multiproc
template: gateway/deployment.yaml
- contains:
path: spec.template.spec.containers[0].volumeMounts
content:
name: prometheus-multiproc
mountPath: /tmp/litellm_prometheus_multiproc
template: gateway/deployment.yaml
- equal:
path: spec.template.spec.containers[1].name
value: metrics
template: gateway/deployment.yaml
- equal:
path: spec.template.spec.containers[1].image
value: ghcr.io/berriai/litellm-gateway:v1.101.0
template: gateway/deployment.yaml
- equal:
path: spec.template.spec.containers[1].command
value:
- python
- -m
- litellm.proxy.prometheus_metrics_server
- --port
- "4101"
template: gateway/deployment.yaml
- equal:
path: spec.template.spec.containers[1].env
value:
- name: PROMETHEUS_MULTIPROC_DIR
value: /tmp/litellm_prometheus_multiproc
template: gateway/deployment.yaml
- equal:
path: spec.template.spec.containers[1].ports
value:
- name: metrics
containerPort: 4101
protocol: TCP
template: gateway/deployment.yaml
- equal:
path: spec.template.spec.containers[1].volumeMounts
value:
- name: prometheus-multiproc
mountPath: /tmp/litellm_prometheus_multiproc
template: gateway/deployment.yaml
- equal:
path: spec.template.spec.containers[1].readinessProbe.tcpSocket.port
value: metrics
template: gateway/deployment.yaml
- equal:
path: spec.template.spec.containers[1].livenessProbe.tcpSocket.port
value: metrics
template: gateway/deployment.yaml
- equal:
path: spec.template.spec.containers[1].resources.requests.cpu
value: 50m
template: gateway/deployment.yaml
- contains:
path: spec.template.spec.volumes
content:
name: prometheus-multiproc
emptyDir: {}
template: gateway/deployment.yaml
- lengthEqual:
path: spec.ports
count: 1
template: gateway/service.yaml
- equal:
path: spec.type
value: LoadBalancer
template: gateway/service.yaml
- equal:
path: metadata.name
value: RELEASE-NAME-litellm-gateway-metrics
template: gateway/service-metrics.yaml
- equal:
path: spec.type
value: ClusterIP
template: gateway/service-metrics.yaml
- equal:
path: spec.ports
value:
- port: 4101
targetPort: metrics
protocol: TCP
name: metrics
template: gateway/service-metrics.yaml
- equal:
path: spec.selector
value:
app.kubernetes.io/name: litellm
app.kubernetes.io/instance: RELEASE-NAME
app.kubernetes.io/component: gateway
template: gateway/service-metrics.yaml
- it: rejects a metrics port equal to the gateway port
template: gateway/deployment.yaml
set:
gateway.metricsServer.enabled: true
gateway.metricsServer.port: 4000
asserts:
- failedTemplate:
errorMessage: gateway.metricsServer.port must differ from the gateway port 4000

16
helm/litellm/values.yaml
View file

@ -268,6 +268,22 @@ gateway:
config:
create: true
proxy_config: {}
# Serve Prometheus /metrics from a `metrics` sidecar container (same image,
# `python -m litellm.proxy.prometheus_metrics_server`) that aggregates the
# workers' PROMETHEUS_MULTIPROC_DIR samples over a shared emptyDir, so a
# scrape never runs on an inference worker. Adds a `metrics` port to the pod
# and a dedicated ClusterIP `<gateway>-metrics` Service; point your scrape
# config at it. The port has no virtual-key auth: keep it off public ingress.
# Needs the gateway image v1.101.0 or newer.
metricsServer:
enabled: false
port: 4001
resources:
requests:
cpu: 50m
memory: 128Mi
limits:
memory: 512Mi
image:
repository: ghcr.io/berriai/litellm-gateway
tag: "" # defaults to .Chart.AppVersion

4
litellm-proxy-extras/litellm_proxy_extras/migrations/20260903230000_add_org_id_to_managed_object_table/migration.sql Normal file
View file

@ -0,0 +1,4 @@
-- Add org_id column to LiteLLM_ManagedObjectTable
-- Snapshots the creating key's organization at submission time, like team_id,
-- so CheckBatchCost can bill organization spend hours later without re-resolving
ALTER TABLE "LiteLLM_ManagedObjectTable" ADD COLUMN IF NOT EXISTS "org_id" TEXT;

1
litellm-proxy-extras/litellm_proxy_extras/schema.prisma
View file

@ -1036,6 +1036,7 @@ model LiteLLM_ManagedObjectTable { // for batches or finetuning jobs which use t
created_at DateTime @default(now())
created_by String?
team_id String?
org_id String? // creating key's organization at submission time; CheckBatchCost bills org spend against it
api_key String?
request_tags Json? @default("[]")
updated_at DateTime @updatedAt

4
litellm-proxy-extras/pyproject.toml
View file

@ -1,6 +1,6 @@
[project]
name = "litellm-proxy-extras"
version = "0.4.94"
version = "0.4.95"
description = "Additional files for the LiteLLM Proxy. Reduces the size of the main litellm package."
readme = "README.md"
requires-python = ">=3.9"
@ -26,7 +26,7 @@ required-version = ">=0.10.9"
module-root = ""
[tool.commitizen]
version = "0.4.94"
version = "0.4.95"
version_files = [
"pyproject.toml:^version",
"../pyproject.toml:litellm-proxy-extras==",

6
litellm/__init__.py
View file

@ -47,6 +47,7 @@ from typing import (
)
from litellm.types.integrations.datadog import DatadogInitParams
from litellm.types.integrations.newrelic import NewRelicInitParams
from litellm.litellm_core_utils.core_helpers import drop_params_env_flag
from litellm._logging import (
set_verbose,
_turn_on_debug,
@ -238,7 +239,7 @@ token: Optional[str] = (
)
telemetry = True
max_tokens: int = DEFAULT_MAX_TOKENS # OpenAI Defaults
drop_params = bool(os.getenv("LITELLM_DROP_PARAMS", False))
drop_params = drop_params_env_flag(os.environ, verbose_logger)
modify_params = bool(os.getenv("LITELLM_MODIFY_PARAMS", False))
use_chat_completions_url_for_anthropic_messages: bool = bool(
os.getenv("LITELLM_USE_CHAT_COMPLETIONS_URL_FOR_ANTHROPIC_MESSAGES", False)
@ -325,6 +326,9 @@ ssl_certificate: Optional[str] = None
user_url_validation: bool = True
user_url_allowed_hosts: List[str] = []
provider_url_destination_allowed_hosts: List[str] = []
#: "override" (default) or "additive": whether a key or team destination replaces
#: the operator's exporter for that backend or exports alongside it.
otel_tenant_destination_mode: str | None = None
ssl_ecdh_curve: Optional[str] = None # Set to 'X25519' to disable PQC and improve performance
disable_streaming_logging: bool = False
disable_token_counter: bool = False

52
litellm/batches/batch_utils.py
View file

@ -10,7 +10,7 @@ from litellm._logging import verbose_logger
from litellm.litellm_core_utils.get_litellm_params import AWS_CREDENTIAL_KWARGS_KEYS
from litellm.litellm_core_utils.llm_cost_calc.utils import parse_prompt_tokens_details
from litellm.types.llms.openai import Batch
from litellm.types.utils import CallTypes, ModelInfo, Usage
from litellm.types.utils import ModelInfo, Usage
from litellm.utils import token_counter
@ -23,6 +23,8 @@ class BatchCostUsageResult:
models: list[str]
successful_requests: int
failed_requests: int
prompt_cost: float = 0.0
completion_cost: float = 0.0
_COMPLETED_BATCH_STATUSES: Final = frozenset({"completed", "complete"})
@ -151,7 +153,8 @@ class _LineOutcome(Enum):
@dataclass(frozen=True, slots=True)
class _BatchOutputLineStats:
cost: float
prompt_cost: float
completion_cost: float
prompt_tokens: int
completion_tokens: int
total_tokens: int
@ -214,15 +217,16 @@ def _compute_output_line_stats(
raw_model: Final = response_body.get("model")
response_model: Final = raw_model if isinstance(raw_model, str) and raw_model else None
completion_details: Final = usage.completion_tokens_details
line_prompt_cost, line_completion_cost = _output_line_cost(
usage=usage,
custom_llm_provider=custom_llm_provider,
model_name=model_name,
response_model=response_model,
model_info=model_info,
)
return _BatchOutputLineStats(
cost=_output_line_cost(
response_body=response_body,
usage=usage,
custom_llm_provider=custom_llm_provider,
model_name=model_name,
response_model=response_model,
model_info=model_info,
),
prompt_cost=line_prompt_cost,
completion_cost=line_completion_cost,
prompt_tokens=usage.prompt_tokens,
completion_tokens=usage.completion_tokens,
total_tokens=usage.total_tokens,
@ -234,31 +238,24 @@ def _compute_output_line_stats(
def _output_line_cost(
response_body: Mapping[str, object],
usage: Usage,
custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic", "bedrock"],
model_name: str | None,
response_model: str | None,
model_info: ModelInfo | None,
) -> float:
) -> tuple[float, float]:
"""(prompt_cost, completion_cost) for one output line, priced at batch rates."""
from litellm.cost_calculator import batch_cost_calculator
if model_info is None and custom_llm_provider not in ("anthropic", "bedrock"):
return litellm.completion_cost(
completion_response=response_body,
custom_llm_provider=custom_llm_provider,
call_type=CallTypes.aretrieve_batch.value,
)
cost_model: Final = (
model_name if custom_llm_provider == "bedrock" and model_name else response_model or model_name or ""
)
prompt_cost, completion_cost = batch_cost_calculator(
return batch_cost_calculator(
usage=usage,
model=cost_model,
custom_llm_provider=custom_llm_provider,
model_info=model_info,
)
return prompt_cost + completion_cost
def _aggregate_batch_cost_usage_models(
@ -291,7 +288,9 @@ def _aggregate_batch_cost_usage_models(
**cache_token_params,
)
batch_models: Final = [model_name] if model_name else [stats.model for stats in line_stats if stats.model]
total_cost: Final = sum((stats.cost for stats in line_stats), 0.0)
total_prompt_cost: Final = sum((stats.prompt_cost for stats in line_stats), 0.0)
total_completion_cost: Final = sum((stats.completion_cost for stats in line_stats), 0.0)
total_cost: Final = total_prompt_cost + total_completion_cost
verbose_logger.debug(
"batch output aggregate: cost=%s usage=%s models=%s successful=%d failed=%d",
total_cost,
@ -306,6 +305,8 @@ def _aggregate_batch_cost_usage_models(
models=batch_models,
successful_requests=successful_requests,
failed_requests=failed_requests,
prompt_cost=total_prompt_cost,
completion_cost=total_completion_cost,
)
@ -330,7 +331,8 @@ def calculate_vertex_ai_batch_cost_and_usage(
"""
from litellm.cost_calculator import batch_cost_calculator
total_cost = 0.0
total_prompt_cost = 0.0 # rebind-ok: loop accumulator, matches total_tokens below
total_completion_cost = 0.0 # rebind-ok: loop accumulator, matches total_tokens below
total_tokens = 0
prompt_tokens = 0
completion_tokens = 0
@ -362,7 +364,8 @@ def calculate_vertex_ai_batch_cost_and_usage(
model=actual_model_name,
custom_llm_provider="vertex_ai",
)
total_cost += p_cost + c_cost
total_prompt_cost += p_cost
total_completion_cost += c_cost
except Exception as e:
verbose_logger.debug("vertex_ai batch cost calculation error for line: %s", str(e))
@ -370,6 +373,7 @@ def calculate_vertex_ai_batch_cost_and_usage(
completion_tokens += _completion
total_tokens += _total
total_cost: Final = total_prompt_cost + total_completion_cost
verbose_logger.info(
"vertex_ai batch cost: cost=%s, prompt=%d, completion=%d, total=%d, successful=%d, failed=%d",
total_cost,
@ -390,6 +394,8 @@ def calculate_vertex_ai_batch_cost_and_usage(
models=[actual_model_name],
successful_requests=successful_requests,
failed_requests=failed_requests,
prompt_cost=total_prompt_cost,
completion_cost=total_completion_cost,
)

9
litellm/completion_extras/litellm_responses_transformation/transformation.py
View file

@ -370,7 +370,12 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
and isinstance(tool_call.get("custom"), dict)
)
for msg in messages:
leading_system_count: Final = next(
(index for index, msg in enumerate(messages) if msg.get("role") != "system"),
len(messages),
)
for index, msg in enumerate(messages):
role = msg.get("role")
content = msg.get("content", "")
tool_calls = msg.get("tool_calls")
@ -378,7 +383,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
if role == "system":
# Extract system message as instructions
if isinstance(content, str):
if isinstance(content, str) and index < leading_system_count:
if instructions:
# Concatenate multiple system prompts with a space
instructions = f"{instructions} {content}"

3
litellm/constants.py
View file

@ -1461,7 +1461,10 @@ LITELLM_METADATA_FIELD: Final = "litellm_metadata"
OLD_LITELLM_METADATA_FIELD: Final = "metadata"
RETURN_RAW_MODEL_NAME_METADATA_KEY: Final = "_complexity_router_return_raw_model_name"
SESSION_DEPLOYMENT_AFFINITY_TTL_METADATA_KEY: Final = "_session_deployment_affinity_ttl"
OUTPUT_TOKEN_CEILING_PARAMS: Final = frozenset({"max_tokens", "max_completion_tokens", "max_output_tokens"})
CLIENT_OUTPUT_CEILING_METADATA_KEY: Final = "_client_output_ceiling"
CONSUMED_REQUEST_TAGS_METADATA_KEY: Final = "_consumed_request_tags"
ROUTING_REQUEST_TAGS_METADATA_KEY: Final = "_routing_request_tags"
INTERNAL_CALL_ORIGIN_METADATA_KEY: Final = "internal_call_origin"
SESSION_ID_GENERATED_METADATA_KEY: Final = "litellm_session_id_generated"
SESSION_ID_OMITTED_METADATA_KEY: Final = "litellm_session_id_omitted"

40
litellm/cost_calculator.py
View file

@ -2414,6 +2414,46 @@ class RealtimeAPITokenUsageProcessor(BaseTokenUsageProcessor):
)
_RESPONSES_WS_BILLABLE_EVENT_TYPES: Final = frozenset({"response.completed", "response.incomplete"})
class _ResponsesWsEventResponse(BaseModel):
usage: Mapping[str, object] | None = None
class _ResponsesWsEvent(BaseModel):
type: str = ""
response: _ResponsesWsEventResponse | None = None
class ResponsesWebSocketTokenUsageProcessor(BaseTokenUsageProcessor):
@staticmethod
def collect_usage_from_responses_ws_results(
results: Sequence[Mapping[str, object]],
) -> tuple[Usage, ...]:
events: Final = tuple(_ResponsesWsEvent.model_validate(result) for result in results)
return tuple(
ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage( # pyright: ignore[reportPrivateUsage] # same shared transform the realtime processor uses
event.response.usage
)
for event in events
if event.type in _RESPONSES_WS_BILLABLE_EVENT_TYPES
and event.response is not None
and event.response.usage is not None
)
@staticmethod
def collect_and_combine_usage_from_responses_ws_results(
results: Sequence[Mapping[str, object]],
) -> Usage:
collected_usage_objects: Final = ResponsesWebSocketTokenUsageProcessor.collect_usage_from_responses_ws_results(
results
)
return ResponsesWebSocketTokenUsageProcessor.combine_usage_objects(
list(collected_usage_objects) # mutable-ok: combine_usage_objects requires a list parameter
)
_TRANSCRIPTION_COMPLETED_EVENT_TYPE: Final = "conversation.item.input_audio_transcription.completed"

13
litellm/integrations/callback_configs.json
View file

@ -356,11 +356,24 @@
"description": "OpenTelemetry collector endpoint URL",
"required": true
},
"otel_traces_endpoint": {
"type": "text",
"ui_name": "Traces Endpoint URL",
"description": "Complete trace export URL used verbatim when the collector does not serve /v1/traces (OTel v2 only)",
"required": false
},
"otel_headers": {
"type": "text",
"ui_name": "Headers",
"description": "Headers for OTEL exporter (e.g., x-honeycomb-team=YOUR_API_KEY)",
"required": false
},
"otel_exporter_otlp_protocol": {
"type": "select",
"ui_name": "Export Protocol",
"description": "OTLP wire format for trace exports. Use http/json for collectors that cannot decode protobuf",
"options": ["http/protobuf", "http/json"],
"required": false
}
},
"description": "OpenTelemetry Logging Integration"

28
litellm/integrations/mlflow.py
View file

@ -133,17 +133,17 @@ class MlflowLogger(CustomLogger):
if final_response:
end_time_ns: Final = int(end_time.timestamp() * 1e9)
self._extract_and_set_chat_attributes(span, kwargs, final_response)
self._end_span_or_trace(
span=span,
outputs=final_response,
status=SpanStatusCode.OK,
end_time_ns=end_time_ns,
)
# Remove the stream_id from the map
with self._lock:
self._stream_id_to_span.pop(litellm_call_id)
try:
self._extract_and_set_chat_attributes(span, kwargs, final_response)
self._end_span_or_trace(
span=span,
outputs=final_response,
status=SpanStatusCode.OK,
end_time_ns=end_time_ns,
)
finally:
with self._lock:
self._stream_id_to_span.pop(litellm_call_id, None)
def _add_chunk_events(self, span, response_obj):
from mlflow.entities import SpanEvent
@ -282,15 +282,15 @@ class MlflowLogger(CustomLogger):
"""End an MLflow span or a trace."""
if span.parent_id is None:
self._client.end_trace(
trace_id=span.request_id,
span.request_id,
outputs=outputs,
status=status,
end_time_ns=end_time_ns,
)
else:
self._client.end_span(
trace_id=span.request_id,
span_id=span.span_id,
span.request_id,
span.span_id,
outputs=outputs,
status=status,
end_time_ns=end_time_ns,

55
litellm/integrations/otel/logger.py
View file

@ -16,9 +16,11 @@ from opentelemetry.trace import (
Span,
Tracer,
get_current_span,
get_tracer_provider,
set_span_in_context,
use_span,
)
from opentelemetry.trace import TracerProvider as ApiTracerProvider
import litellm
from litellm._logging import verbose_logger
@ -63,6 +65,7 @@ from litellm.integrations.otel.plumbing.metrics import (
create_genai_metrics,
)
from litellm.integrations.otel.plumbing.providers import (
attach_tenant_fan_out,
build_tracer_provider,
get_event_logger,
get_meter,
@ -85,6 +88,7 @@ if TYPE_CHECKING:
)
LITELLM_TRACER_NAME: Final = "litellm"
_published_v2_provider: ApiTracerProvider | None = None
def _span_error_from_exception(
@ -180,7 +184,9 @@ class OpenTelemetryV2(CustomLogger):
self.config: OpenTelemetryV2Config = config or OpenTelemetryV2Config(**kwargs)
self.callback_name = callback_name
self._tracer_provider: TracerProvider = (
tracer_provider if tracer_provider is not None else build_tracer_provider(self.config)
tracer_provider
if tracer_provider is not None
else build_tracer_provider(self.config, tenant_overrides=True)
)
self.tracer: Tracer = get_tracer(self._tracer_provider, LITELLM_TRACER_NAME)
self._metrics_recorder = self._init_metrics(meter_provider)
@ -195,6 +201,11 @@ class OpenTelemetryV2(CustomLogger):
self._open_llm_calls: OrderedDict[str, _LLMCallSpan] = OrderedDict()
self._init_otel_logger_on_litellm_proxy()
@property
def tracer_provider(self) -> TracerProvider:
"""The provider this logger emits through, read-only to its callers."""
return self._tracer_provider
def _init_metrics(self, meter_provider: "MeterProvider | None") -> "GenAIMetricRecorder | None":
"""Create the six GenAI histograms when metrics are enabled, else ``None``.
@ -863,12 +874,33 @@ def publish_global_otel_v2_provider(
``opentelemetry.trace.set_tracer_provider``) are injected so the publish step is
unit-testable without reading or mutating real global OTel state. Returns the
logger whose provider was published.
The published provider is also the one that fans spans out to key/team
destinations, because it is the only provider the whole request tree passes
through; see :func:`attach_tenant_fan_out`. It is remembered for
:func:`fan_out_provider` because neither the OTel global (``set_tracer_provider``
keeps the first provider it was ever handed) nor
``proxy_server.open_telemetry_logger`` (a legacy v1 logger can hold that slot)
reliably leads back to it.
"""
global _published_v2_provider
logger: Final = select_global_otel_v2_logger(in_memory_loggers, registered=registered)
set_global_provider(logger._tracer_provider)
attach_tenant_fan_out(logger.tracer_provider, *_v2_configs(in_memory_loggers, logger))
set_global_provider(logger.tracer_provider)
_published_v2_provider = logger.tracer_provider # rebind-ok: startup records the one provider carrying the fan-out
return logger
def _v2_configs(in_memory_loggers: Sequence[object], logger: "OpenTelemetryV2") -> tuple[OpenTelemetryV2Config, ...]:
"""Every v2 logger's config, the published logger's first.
Each preset keeps its own provider and exporters, so the accounts the operator
writes to are spread over all of them, not held by the published logger alone.
"""
others: Final = tuple(cb.config for cb in in_memory_loggers if isinstance(cb, OpenTelemetryV2) and cb is not logger)
return (logger.config, *others)
def _registered_v2_logger() -> "OpenTelemetryV2 | None":
try:
from litellm.proxy import proxy_server
@ -904,6 +936,25 @@ def seed_request_identity(user_api_key_dict: object, model: str | None = None) -
logger.seed_request_identity(user_api_key_dict, model=model)
def fan_out_provider() -> ApiTracerProvider:
"""The provider :func:`publish_global_otel_v2_provider` gave the tenant fan-out.
Read off the publish itself, not the OTel global and not the registered logger:
the global keeps whichever provider claimed it first (auto-instrumentation, a
legacy logger), and the registered slot can hold a v1 logger while the publish
picked a v2 one from ``_in_memory_loggers``. Either detour lands on a provider
with no fan-out and drops every destination at auth.
"""
published: Final = _published_v2_provider
if published is not None:
return published
logger: Final = _registered_v2_logger()
if logger is not None:
attach_tenant_fan_out(logger.tracer_provider, logger.config)
return logger.tracer_provider
return get_tracer_provider()
@contextmanager
def phase_span(name: str) -> "Iterator[Span | None]":
logger: Final = _registered_v2_logger()

3
litellm/integrations/otel/mappers/legacy.py
View file

@ -23,6 +23,7 @@ from litellm.integrations.otel.model.payloads import (
ServiceSpanData,
ToolDefinition,
)
from litellm.integrations.otel.model.semconv import Error
# Attribute keys in the semconv-ai / Traceloop vocabulary.
_LEGACY_SYSTEM: Final = "gen_ai.system"
@ -36,7 +37,7 @@ _LEGACY_PRESENCE_PENALTY: Final = "llm.presence_penalty"
_LEGACY_STOP_SEQUENCES: Final = "llm.chat.stop_sequences"
_LEGACY_SERVICE: Final = "service"
_LEGACY_CALL_TYPE: Final = "call_type"
_LEGACY_ERROR: Final = "error"
_LEGACY_ERROR: Final = Error.MESSAGE_LEGACY
class LegacyMapper:

27
litellm/integrations/otel/model/config.py
View file

@ -69,9 +69,17 @@ class ExporterSpec(BaseModel):
kind: str = Field(
default="console",
description="console | in_memory | otlp_http | otlp_grpc | <factory kind>",
description="console | in_memory | otlp_http | http/json | otlp_grpc | <factory kind>",
)
endpoint: str | None = None
traces_endpoint: str | None = Field(
default=None,
description=(
"Complete OTLP/HTTP trace URL, used verbatim. Set this when the "
"collector serves traces on a path other than ``/v1/traces``; "
"``endpoint`` is a base URL the signal path is appended to."
),
)
headers: str | None = None
owner: ExporterOwner | None = Field(
default=None,
@ -127,6 +135,14 @@ class OpenTelemetryV2Config(BaseSettings):
default=None,
validation_alias=AliasChoices("OTEL_ENDPOINT", "OTEL_EXPORTER_OTLP_ENDPOINT"),
)
traces_endpoint: str | None = Field(
default=None,
validation_alias=AliasChoices("OTEL_TRACES_ENDPOINT", "OTEL_EXPORTER_OTLP_TRACES_ENDPOINT"),
description=(
"Complete OTLP/HTTP trace URL for the single-destination shorthand, "
"used verbatim instead of ``endpoint`` + ``/v1/traces``."
),
)
headers: str | None = Field(
default=None,
validation_alias=AliasChoices("OTEL_HEADERS", "OTEL_EXPORTER_OTLP_HEADERS"),
@ -250,17 +266,22 @@ class OpenTelemetryV2Config(BaseSettings):
@model_validator(mode="after")
def _normalize(self) -> "OpenTelemetryV2Config":
# An endpoint with the default exporter kind implies OTLP/HTTP.
if self.endpoint and self.exporter == "console":
if (self.endpoint or self.traces_endpoint) and self.exporter == "console":
self.exporter = "otlp_http"
# When no explicit destinations are given, fold the single-destination
# shorthand into one spec so the provider always has a destination.
# shorthand into one spec so the provider always has a destination. A spec
# with no fields set is how the presets tell "nothing configured" from an
# operator who asked for the console by name.
if not self.exporters:
self.exporters = [
ExporterSpec(
kind=self.exporter,
endpoint=self.endpoint,
traces_endpoint=self.traces_endpoint,
headers=self.headers,
)
if not self.model_fields_set.isdisjoint(("exporter", "endpoint", "headers"))
else ExporterSpec()
]
# Ensure ``genai`` is always present and first.
names = list(self.mapper_names)

49
litellm/integrations/otel/model/destination.py Normal file
View file

@ -0,0 +1,49 @@
"""The resolved OTLP destination a request's traces export to.
Backend-agnostic on purpose: every OTEL backend reduces to an endpoint plus auth
headers. The per-backend field mapping lives in ``presets.destinations``.
"""
from collections.abc import Mapping
from typing import Final
from urllib.parse import quote
from pydantic import BaseModel, ConfigDict, Field
class OtelDestination(BaseModel):
model_config = ConfigDict(frozen=True)
endpoint: str
headers: Mapping[str, str] = Field(default_factory=dict)
resource_attributes: Mapping[str, str] = Field(default_factory=dict)
callback_name: str | None = None
protocol: str | None = Field(
default=None,
description=(
"OTLP transport, defaulting to the backend's own. Not derivable from the "
"scheme: Arize's ``https://otlp.arize.com/v1`` is gRPC."
),
)
def header_string(self) -> str:
"""Render headers as the ``k=v,k2=v2`` form an ``ExporterSpec`` expects.
Values are percent-encoded because ``providers.parse_headers`` decodes them
with the SDK's W3C-Baggage parser: a value carrying a ``,`` or ``=`` (a
Langfuse project name, a base64 Authorization payload ending in ``==``)
would otherwise be split into bogus pairs on the way back out.
"""
return ",".join(f"{key}={quote(value, safe='')}" for key, value in self.headers.items())
def cache_key(self) -> tuple[str, tuple[tuple[str, str], ...], tuple[tuple[str, str], ...], str | None]:
"""Identity for processor reuse, so one destination means one exporter."""
return (
self.endpoint,
tuple(sorted(self.headers.items())),
tuple(sorted(self.resource_attributes.items())),
self.protocol,
)
NO_DESTINATIONS: Final[tuple[OtelDestination, ...]] = ()

3
litellm/integrations/otel/model/semconv.py
View file

@ -204,6 +204,9 @@ class Error:
TYPE: Final = "error.type"
MESSAGE: Final = "error.message"
# The same text under the bare key the semconv-ai / Traceloop vocabulary uses
# (see ``LegacyMapper``), so anything reading or redacting error text covers both.
MESSAGE_LEGACY: Final = "error"
class LiteLLMError:

68
litellm/integrations/otel/plumbing/context.py
View file

@ -1,8 +1,9 @@
"""Trace-context + Baggage helpers."""
import os
from collections.abc import Mapping
from contextvars import ContextVar, Token
from typing import Final
from typing import TYPE_CHECKING, Final
from opentelemetry import baggage
from opentelemetry.context import Context, get_current
@ -21,6 +22,9 @@ from opentelemetry.trace.propagation.tracecontext import (
from litellm.integrations.otel.model.semconv import HTTP
if TYPE_CHECKING:
from litellm.integrations.otel.model.destination import OtelDestination
_PROPAGATOR: Final = TraceContextTextMapPropagator()
# The request's root span — the FastAPI-owned SERVER span — captured ONCE when the
@ -304,3 +308,65 @@ def extract_traceparent(headers: Mapping[str, str]) -> Context | None:
return None
carrier: Final = {str(key).lower(): value for key, value in headers.items()}
return _PROPAGATOR.extract(carrier)
# The OTLP destinations this request's key or team pointed its traces at, resolved
# once during auth. A ``ContextVar`` for the same reason the root span above is one:
# it rides the request task's context into the ``asyncio.create_task`` children that
# close the LLM span, and it is visible to every ``SpanProcessor.on_end`` that fires
# on the request task. Stateful MCP handlers set and reset it per message; the
# request-task value otherwise dies with that task.
_request_destinations: Final['ContextVar[tuple["OtelDestination", ...]]'] = ContextVar(
"litellm_otel_request_destinations", default=()
)
def set_request_destinations(destinations: 'tuple["OtelDestination", ...]') -> "Token[tuple[OtelDestination, ...]]":
"""Anchor the destinations this request exports to and return a reset token."""
return _request_destinations.set(destinations)
def reset_request_destinations(token: "Token[tuple[OtelDestination, ...]]") -> None:
_request_destinations.reset(token)
def request_destinations() -> 'tuple["OtelDestination", ...]':
"""The destinations resolved for this request, empty outside a proxy request."""
return _request_destinations.get()
#: ``litellm_settings: otel_tenant_destination_mode`` and its env equivalent.
ADDITIVE_DESTINATION_MODE: Final = "additive"
OTEL_TENANT_DESTINATION_MODE_ENV: Final = "LITELLM_OTEL_TENANT_DESTINATION_MODE"
def tenant_destinations_are_additive() -> bool:
"""Whether a tenant destination exports alongside the operator's own exporter.
Override is the default: the tenant's traffic reaches the tenant's account and
nowhere else. Operators running one org-wide backend across every team set this
to ``additive`` so the same trace lands in both places.
"""
import litellm
configured: Final = litellm.otel_tenant_destination_mode or os.environ.get(OTEL_TENANT_DESTINATION_MODE_ENV)
return isinstance(configured, str) and configured.strip().lower() == ADDITIVE_DESTINATION_MODE
def destination_backends() -> frozenset[str]:
"""Backends this request resolved a tenant destination for.
The fan-out already carries the whole trace to those destinations, so the
per-request tracer route must never send a second copy, in either mode.
"""
return frozenset(d.callback_name for d in _request_destinations.get() if d.callback_name)
def suppressed_backends() -> frozenset[str]:
"""Backends whose operator-level exporters this request must NOT reach.
Empty under ``additive``, where the operator keeps its copy of every span.
"""
if tenant_destinations_are_additive():
return frozenset()
return destination_backends()

70
litellm/integrations/otel/plumbing/otlp_json.py Normal file
View file

@ -0,0 +1,70 @@
"""OTLP/HTTP span exporter that sends the OTLP/JSON encoding instead of protobuf.
The SDK only ships a protobuf OTLP/HTTP exporter; this reuses its transport and
retry loop and swaps the payload for OTLP/JSON (enums as integers, ids as hex).
"""
import base64
import json
from collections.abc import Mapping, Sequence
from types import MappingProxyType
from typing import Final, TypeAlias
from google.protobuf.json_format import MessageToDict
from opentelemetry.exporter.otlp.proto.common.trace_encoder import encode_spans
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk.trace import ReadableSpan
JSON_CONTENT_TYPE: Final = "application/json"
_HEX_ID_KEYS: Final = frozenset({"traceId", "spanId", "parentSpanId"})
_JsonValue: TypeAlias = "Mapping[str, _JsonValue] | Sequence[_JsonValue] | str | int | float | bool | None"
_JsonObject: TypeAlias = Mapping[str, "_JsonValue"]
def _objects(node: _JsonObject, key: str) -> tuple[_JsonObject, ...]:
items: Final = node.get(key)
if isinstance(items, str) or not isinstance(items, Sequence):
return ()
return tuple(item for item in items if isinstance(item, Mapping))
def _hex_ids(node: _JsonObject) -> _JsonObject:
return MappingProxyType(
{
key: base64.b64decode(item).hex() if key in _HEX_ID_KEYS and isinstance(item, str) else item
for key, item in node.items()
}
)
def _hex_span(span: _JsonObject) -> _JsonObject:
links: Final = _objects(span, "links")
if not links:
return _hex_ids(span)
return MappingProxyType({**_hex_ids(span), "links": tuple(_hex_ids(link) for link in links)})
def _hex_scope_spans(scope: _JsonObject) -> _JsonObject:
return MappingProxyType({**scope, "spans": tuple(_hex_span(span) for span in _objects(scope, "spans"))})
def _hex_resource_spans(resource: _JsonObject) -> _JsonObject:
scope_spans: Final = tuple(_hex_scope_spans(scope) for scope in _objects(resource, "scopeSpans"))
return MappingProxyType({**resource, "scopeSpans": scope_spans})
def encode_spans_json(spans: Sequence[ReadableSpan]) -> bytes:
payload: Final[_JsonObject] = MessageToDict(encode_spans(spans), use_integers_for_enums=True)
resource_spans: Final = tuple(_hex_resource_spans(resource) for resource in _objects(payload, "resourceSpans"))
hexed: Final[_JsonObject] = MappingProxyType({**payload, "resourceSpans": resource_spans})
return json.dumps(hexed, default=dict, separators=(",", ":")).encode()
class OTLPJsonSpanExporter(OTLPSpanExporter):
def __init__(self, endpoint: str | None, headers: dict[str, str]) -> None: # mutable-ok: SDK __init__ takes Dict
super().__init__(endpoint=endpoint, headers=headers)
self._session.headers["Content-Type"] = JSON_CONTENT_TYPE
def _serialize_spans(self, spans: Sequence[ReadableSpan]) -> bytes:
return encode_spans_json(spans)

711
litellm/integrations/otel/plumbing/providers.py
View file

@ -1,9 +1,14 @@
"""Provider / exporter factory + the Baggage span processor."""
from collections.abc import Callable, Iterable
import queue
import threading
import time
from collections import OrderedDict
from collections.abc import Callable, Iterable, Mapping, Sequence
from types import MappingProxyType
from typing import TYPE_CHECKING, Any, Final, Literal
from opentelemetry import _logs, baggage, metrics
from opentelemetry import _logs, baggage, metrics, trace
from opentelemetry._events import EventLogger
from opentelemetry._logs import LoggerProvider, NoOpLoggerProvider
from opentelemetry.context import Context
@ -19,7 +24,8 @@ from opentelemetry.sdk._logs.export import (
)
from opentelemetry.sdk.metrics import MeterProvider as SDKMeterProvider
from opentelemetry.sdk.resources import Resource
from opentelemetry.sdk.trace import ReadableSpan, SpanProcessor, TracerProvider
from opentelemetry.sdk.trace import Event, ReadableSpan, SpanProcessor, TracerProvider
from opentelemetry.sdk.trace import Span as SDKSpan
from opentelemetry.sdk.trace.export import (
BatchSpanProcessor,
ConsoleSpanExporter,
@ -29,18 +35,35 @@ from opentelemetry.sdk.trace.export import (
from opentelemetry.sdk.trace.export.in_memory_span_exporter import (
InMemorySpanExporter,
)
from opentelemetry.trace import Span, SpanKind, Tracer
from opentelemetry.trace import Span, SpanKind, Status, Tracer
from opentelemetry.util.re import parse_env_headers
from opentelemetry.util.types import Attributes, AttributeValue
from litellm._logging import verbose_logger
from litellm._version import version as litellm_version
from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
from litellm.integrations.otel.model.semconv import LiteLLM
from litellm.integrations.otel.model.semconv import (
DB,
MCP,
Error,
ExceptionEvent,
GenAI,
LiteLLM,
LiteLLMError,
Server,
)
from litellm.integrations.otel.model.spans import LiteLLMSpanKind
from litellm.integrations.otel.plumbing.context import (
request_destinations,
suppressed_backends,
)
if TYPE_CHECKING:
from opentelemetry.metrics import Meter
from opentelemetry.sdk.metrics.export import MetricReader
from litellm.integrations.otel.model.destination import OtelDestination
_SPAN_KIND_BY_ROLE_KIND: Final[dict[LiteLLMSpanKind, SpanKind]] = {
LiteLLMSpanKind.SERVER: SpanKind.SERVER,
LiteLLMSpanKind.CLIENT: SpanKind.CLIENT,
@ -136,7 +159,8 @@ def parse_headers(raw: str | None) -> dict[str, str]:
_IN_MEMORY_KINDS: Final = ("in_memory", "inmemory", "memory")
_OTLP_HTTP_KINDS: Final = ("otlp_http", "http", "http/protobuf", "http/json")
_OTLP_HTTP_JSON_KINDS: Final = ("http/json",)
_OTLP_HTTP_KINDS: Final = ("otlp_http", "http", "http/protobuf", *_OTLP_HTTP_JSON_KINDS)
_OTLP_GRPC_KINDS: Final = ("otlp_grpc", "grpc")
@ -164,13 +188,20 @@ def _exporter_from_spec(spec: ExporterSpec) -> SpanExporter:
return factory(spec)
if kind in _IN_MEMORY_KINDS:
return InMemorySpanExporter()
if kind in _OTLP_HTTP_JSON_KINDS:
from litellm.integrations.otel.plumbing.otlp_json import OTLPJsonSpanExporter
return OTLPJsonSpanExporter(
endpoint=spec.traces_endpoint or _otlp_traces_endpoint(spec.endpoint),
headers=parse_headers(spec.headers),
)
if kind in _OTLP_HTTP_KINDS:
from opentelemetry.exporter.otlp.proto.http.trace_exporter import (
OTLPSpanExporter as HTTPExporter,
)
return HTTPExporter(
endpoint=_otlp_traces_endpoint(spec.endpoint),
endpoint=spec.traces_endpoint or _otlp_traces_endpoint(spec.endpoint),
headers=parse_headers(spec.headers),
)
if kind in _OTLP_GRPC_KINDS:
@ -194,6 +225,555 @@ def _processor_for(exporter: SpanExporter, use_simple: bool | None) -> SpanProce
return SimpleSpanProcessor(exporter) if use_simple else BatchSpanProcessor(exporter)
#: Distinct tenant destinations whose exporters stay alive. Each holds a connection
#: pool and a batch thread, so the cache is bounded and evicts least-recently-used.
_MAX_CACHED_DESTINATION_PROCESSORS: Final = 32
#: Workers closing shed destination processors, bounding the threads a tenant can
#: create by cycling its destination config.
_DRAIN_WORKERS: Final = 2
#: Shed processors waiting to be closed before the fan-out stops building new ones.
#: Each still owns a batch thread until its close returns, and a collector that never
#: answers makes every close take the exporter's full timeout, so past this many the
#: operator's exporter keeps the span instead (see ``deliverable``).
_MAX_PENDING_DRAINS: Final = 64
#: How long ``shutdown`` waits for spans already being forwarded, so teardown closes
#: no processor under one. Bounded: an exporter that never returns must not hold the
#: proxy open.
_SHUTDOWN_DRAIN_SECONDS: Final = 5.0
#: An exporter's account: its normalized endpoint and the credentials it presents.
_SinkKey = tuple[str, tuple[tuple[str, str], ...]]
#: Header names that spell one credential two ways. Arize's operator exporter sends
#: ``space_id`` where a tenant destination sends ``arize-space-id``.
_CREDENTIAL_ALIASES: Final = MappingProxyType({"arize_space_id": "space_id"})
class _DrainPool:
"""Closes shed destination processors off the span-export path.
``shutdown`` flushes over the network and is reached from ``on_end``, so closing
one inline would let a single unreachable tenant collector stall every other
tenant's spans behind it. A fixed set of workers rather than a thread per
processor means a tenant cycling its destination config cannot spawn threads as
fast as it can send requests; slow shutdowns queue behind each other.
The workers are daemons and belong to the fan-out that sheds the processors, so
neither an unreachable collector nor a lazily built process-wide singleton can
hold the proxy open on the way down.
"""
def __init__(
self,
workers: int = _DRAIN_WORKERS,
pending: "queue.Queue[SpanProcessor | None] | None" = None,
capacity: int = _MAX_PENDING_DRAINS,
) -> None:
self._workers: Final = workers
self._capacity: Final = capacity
self._lock: Final = threading.Lock()
self._closed = False
self._backlog = 0 # guarded by ``_lock``: submitted processors whose close has not returned
self._pending: Final[queue.Queue[SpanProcessor | None]] = pending if pending is not None else queue.Queue()
self._threads: Final = tuple(
threading.Thread(target=self._drain_until_closed, daemon=True, name="litellm-otel-destination-drain")
for _ in range(workers)
)
for worker in self._threads:
worker.start()
def submit(self, processor: SpanProcessor) -> None:
"""Queue ``processor`` for closing, or hand it off once the pool is retired.
The check and the put share one lock. Reading a closed flag on its own leaves
room for :meth:`close` to run in between, and the processor would land behind
the sentinels every worker has already exited on.
Past close there is no worker left to take it, and the caller is whichever
thread just ended a span, so closing it inline would park that thread on a
network flush the shutdown deadline has already stopped waiting for. The extra
thread is bounded by the same close: the fan-out stops handing processors out
at that point, so only the ones already exporting when it happened arrive here.
"""
with self._lock:
if not self._closed:
self._backlog += 1
self._pending.put(processor)
return
threading.Thread(
target=_shutdown_quietly,
args=(processor,),
daemon=True,
name="litellm-otel-destination-drain-straggler",
).start()
def saturated(self) -> bool:
"""Whether enough closes are outstanding that building another processor must wait.
The workers close in order and each close blocks for as long as its exporter
does, so a collector that stopped answering would otherwise turn every new
destination into one more batch thread parked behind them, for as long as the
tenants keep rotating. Holding the count here rather than reading the queue
keeps the two processors a worker is mid-close on in the total.
"""
with self._lock:
return self._backlog >= self._capacity
def close(self, timeout: float | None = None) -> None:
"""Retire the workers once they have closed everything already queued.
A proxy that rebuilds its telemetry builds another fan-out, so workers that
outlive the one that started them are two more threads per reload, forever.
``timeout`` bounds how long the caller waits for that draining to finish. The
workers are daemons, so whatever is still flushing when it expires is dropped
by the interpreter rather than holding it open.
"""
with self._lock:
if self._closed:
return
self._closed = True
for _ in range(self._workers):
self._pending.put(None)
if timeout is None:
return
deadline: Final = time.monotonic() + timeout
for worker in self._threads:
worker.join(timeout=max(0.0, deadline - time.monotonic()))
def _drain_until_closed(self) -> None:
while True:
processor: SpanProcessor | None = self._pending.get() # rebind-ok: loop variable
if processor is None:
return
_shutdown_quietly(processor)
with self._lock:
self._backlog -= 1
_NO_ATTRIBUTES: Final[Mapping[str, AttributeValue]] = MappingProxyType({})
_DB_SYSTEM_KEYS: Final = frozenset({DB.SYSTEM_NAME, DB.SYSTEM_LEGACY})
# Keys on a database span that describe the proxy's own datastore: its host, its
# port, and its schema.
_DATASTORE_ENDPOINT_KEYS: Final = frozenset({Server.ADDRESS, Server.PORT, DB.NAMESPACE})
# A span carrying one of these describes the tenant's own call (the model call, the
# MCP call, the guardrail), so its error text is theirs to see. Every other span is
# the proxy's own work, whose error text names the operator's infrastructure.
_TENANT_OWNED_KEYS: Final = frozenset({GenAI.OPERATION_NAME, MCP.METHOD_NAME, LiteLLM.GUARDRAIL_NAME})
_PROXY_ERROR_TEXT_KEYS: Final = frozenset({Error.MESSAGE, Error.MESSAGE_LEGACY})
# A guardrail that never answered carries the exception it raised as its response,
# which names the operator's guardrail endpoint. The second spelling is the legacy
# status the request-level logger still maps.
_GUARDRAIL_UNREACHABLE_STATUSES: Final = frozenset({"guardrail_failed_to_respond", "failure"})
# Attribute prefixes the FastAPI instrumentor uses for headers the operator opted to
# capture (``OTEL_INSTRUMENTATION_HTTP_CAPTURE_HEADERS_SERVER_*``). The request
# side carries the caller's bearer token verbatim.
_CAPTURED_HEADER_PREFIXES: Final = ("http.request.header.", "http.response.header.")
# The instrumentor stamps the request URL on the server span with its query string,
# under the old convention and the new one, and litellm accepts a virtual key as a
# ``?key=`` query parameter.
_URL_KEYS: Final = frozenset({"http.url", "http.target", "url.full"})
_URL_QUERY_KEY: Final = "url.query"
class _TenantSpanView(ReadableSpan):
"""A ``ReadableSpan`` view for one destination, leaving the operator's own span alone."""
def __init__(
self,
inner: ReadableSpan,
resource: Resource,
attributes: Attributes,
events: Sequence[Event],
status: Status,
) -> None:
super().__init__(
name=inner.name,
context=inner.context,
parent=inner.parent,
resource=resource,
attributes=attributes,
events=events,
links=inner.links,
kind=inner.kind,
status=status,
start_time=inner.start_time,
end_time=inner.end_time,
instrumentation_scope=inner.instrumentation_scope,
)
def _is_database_span(attributes: Mapping[str, AttributeValue]) -> bool:
return any(key in attributes for key in _DB_SYSTEM_KEYS)
def _is_tenant_owned_span(attributes: Mapping[str, AttributeValue]) -> bool:
return any(key in attributes for key in _TENANT_OWNED_KEYS)
def _guardrail_unreachable(attributes: Mapping[str, AttributeValue]) -> bool:
return attributes.get(LiteLLM.GUARDRAIL_STATUS) in _GUARDRAIL_UNREACHABLE_STATUSES
def _tenant_visible(key: str, database: bool, owned: bool, unreachable_guardrail: bool) -> bool:
if key.startswith(_CAPTURED_HEADER_PREFIXES) or key in (LiteLLMError.STACK_TRACE, _URL_QUERY_KEY):
return False
if database and key in _DATASTORE_ENDPOINT_KEYS:
return False
if unreachable_guardrail and key == LiteLLM.GUARDRAIL_RESPONSE:
return False
return owned or key not in _PROXY_ERROR_TEXT_KEYS
def _without_query(key: str, value: AttributeValue) -> AttributeValue:
if key not in _URL_KEYS or not isinstance(value, str):
return value
return value.partition("?")[0]
def _same_attributes(kept: Mapping[str, AttributeValue], attributes: Mapping[str, AttributeValue]) -> bool:
return len(kept) == len(attributes) and all(kept[key] is value for key, value in attributes.items())
def _without_stack_trace(event: Event) -> Event:
attributes: Final = event.attributes or _NO_ATTRIBUTES
if ExceptionEvent.STACKTRACE not in attributes:
return event
return Event(
name=event.name,
attributes=MappingProxyType(
{key: value for key, value in attributes.items() if key != ExceptionEvent.STACKTRACE}
),
timestamp=event.timestamp,
)
def _for_destination(span: ReadableSpan, destination: "OtelDestination") -> ReadableSpan:
"""The view of ``span`` a tenant destination receives.
A span the tenant's own call produced keeps its error text. Every other span is
the proxy's own work (the request root, auth, the database), and its error text,
its events and its status description come off, since a Prisma failure there
spells out the operator's Postgres endpoint. A database span loses that endpoint
too, and a guardrail that failed to respond loses its response text, which is the
exception it raised and names the operator's guardrail endpoint. Stack traces walk
the operator's install and come off every span, as do the headers the operator
captures on the server span, whose request side holds the caller's bearer token,
and the query string of the request URL, which can hold the same key. The span
itself stays, so the tenant still gets the whole trace tree.
"""
extra: Final = destination.resource_attributes
attributes: Final = span.attributes or _NO_ATTRIBUTES
database: Final = _is_database_span(attributes)
owned: Final = _is_tenant_owned_span(attributes)
unreachable: Final = _guardrail_unreachable(attributes)
kept: Final = MappingProxyType(
{
key: _without_query(key, value)
for key, value in attributes.items()
if _tenant_visible(key, database, owned, unreachable)
}
)
recorded: Final = span.events
events: Final = tuple(_without_stack_trace(event) for event in recorded) if owned else ()
unchanged: Final = owned and _same_attributes(kept, attributes) and all(a is b for a, b in zip(events, recorded))
if not extra and unchanged:
return span
resource: Final = span.resource.merge(Resource(extra)) if extra else span.resource
status: Final = span.status if owned else Status(span.status.status_code)
return _TenantSpanView(span, resource, kept, events, status)
class TenantFanOutSpanProcessor(SpanProcessor):
"""Export every finished span to each destination this request resolved.
Destinations ride a request-scoped ``ContextVar`` set during auth, so concurrent
requests stay isolated. The forwarded view keeps the original trace and parent
ids, so the tenant gets the same tree the operator would have received.
Exactly one provider carries this processor, the one published as the OTel global
(see :func:`attach_tenant_fan_out`). That provider is the only one every span
passes through: the FastAPI server span, the auth span and the post-call database
spans are emitted on the global, while a second v2 logger's provider sees only
that logger's own gen-AI span. Attaching the fan-out per logger would hand a
tenant a one-span trace whenever its backend is not the global one, and two
copies of the model call whenever it is.
"""
def __init__(
self,
processor_factory: 'Callable[["OtelDestination"], SpanProcessor | None] | None' = None,
shutdown_drain_seconds: float = _SHUTDOWN_DRAIN_SECONDS,
operator_sinks: frozenset[_SinkKey] = frozenset(),
pending_drains: int = _MAX_PENDING_DRAINS,
drain_pool: _DrainPool | None = None,
) -> None:
self._operator_sinks: Final = operator_sinks
self._drain_seconds: Final = shutdown_drain_seconds
self._lock: Final = threading.Condition()
self._closed = False # guarded by ``_lock``: an unlocked read races the teardown it gates
self._build: Final = processor_factory if processor_factory is not None else _destination_processor
self._processors: OrderedDict[object, SpanProcessor] = OrderedDict() # mutable-ok: bounded LRU
self._retired: OrderedDict[int, SpanProcessor] = OrderedDict() # mutable-ok: drains as exports finish
self._exporting: dict[int, int] = {} # mutable-ok: per-processor in-flight export count
self._drain: Final = drain_pool if drain_pool is not None else _DrainPool(capacity=pending_drains)
def on_start(self, span: SDKSpan, parent_context: Context | None = None) -> None:
return None
def on_end(self, span: ReadableSpan) -> None:
suppressed: Final = suppressed_backends()
for destination in request_destinations():
if self._operator_already_writes(destination, suppressed):
continue
processor = self._acquire(destination) # rebind-ok: loop variable; pyright forbids Final in a loop
if processor is None:
continue
try:
processor.on_end(_for_destination(span, destination))
except Exception as exc: # noqa: BLE001 # one destination's failure must not cost the others their span
verbose_logger.debug("OTel V2 fan-out: forwarding to %s failed: %s", destination.endpoint, exc)
finally:
self._release(processor)
def _operator_already_writes(self, destination: "OtelDestination", suppressed: frozenset[str]) -> bool:
"""Whether the operator's own exporter is sending this span to the same account.
Only reachable under ``additive``, where nothing is suppressed: a team that
names the operator's own project would otherwise have every span written
there twice, once by the operator's exporter and once by the fan-out.
"""
return (
destination.callback_name not in suppressed
and _sink_key(destination.endpoint, destination.headers) in self._operator_sinks
)
def shutdown(self) -> None:
"""Close every destination processor, once the spans in flight have landed.
``on_end`` runs on whichever thread ends a span and can reach this fan-out
while the SDK is tearing the provider down, so closing blind would drop a
trace mid-forward and would hand the next caller a fresh exporter nothing
will ever close. Refusing new work and then waiting out the in-flight ones
keeps both from happening. A straggler past the bound is retired instead of
closed: the thread still exporting it closes it through the drain as soon as
its export returns, so no span is dropped mid-forward.
Every close then goes to the drain rather than running here. Closing a
destination processor flushes it over the network and the SDK joins its own
worker with no timeout of its own, so one tenant collector that answers but
never finishes a response would otherwise hold process teardown open for as
long as it likes. The drain's workers are daemons, and the whole teardown
shares one deadline.
"""
deadline: Final = time.monotonic() + self._drain_seconds
with self._lock:
self._closed = True
self._lock.wait_for(lambda: not self._exporting, timeout=self._drain_seconds)
live: Final = tuple((id(p), p) for p in (*self._processors.values(), *self._retired.values()))
closing: Final = tuple(p for ident, p in live if ident not in self._exporting)
self._processors.clear()
self._retired = OrderedDict( # mutable-ok: the same bounded map, keeping only what is still exporting
(ident, p) for ident, p in live if ident in self._exporting
)
for processor in closing:
self._drain.submit(processor)
self._drain.close(timeout=max(0.0, deadline - time.monotonic()))
def force_flush(self, timeout_millis: int = 30000) -> bool:
results: Final = tuple(self._flush_one(processor, timeout_millis) for processor in self._snapshot())
return all(results)
def _snapshot(self) -> tuple[SpanProcessor, ...]:
with self._lock:
return (*self._processors.values(), *self._retired.values())
@staticmethod
def _flush_one(processor: SpanProcessor, timeout_millis: int) -> bool:
try:
return processor.force_flush(timeout_millis)
except Exception: # noqa: BLE001 # one exporter's flush failure must not fail the whole flush
return False
def deliverable(self, destinations: Iterable["OtelDestination"]) -> tuple["OtelDestination", ...]:
"""The subset of ``destinations`` this fan-out can actually export to.
A destination whose exporter will not build (a protocol whose package is not
installed, a malformed endpoint) has to be dropped before the request anchors
it, not when its first span ends. By then the operator's own exporter has been
told to hold that backend's spans back for this request, so dropping there
loses the span outright instead of leaving it where it would have gone with no
override at all.
"""
return tuple(destination for destination in destinations if self._buildable(destination))
def _buildable(self, destination: "OtelDestination") -> bool:
"""Whether a processor for ``destination`` exists or can be built right now."""
with self._lock:
if self._closed:
return False
built: Final = self._cached_or_built_locked(destination, anchored=False)
drained: Final = self._drainable_locked()
for shed in drained:
self._drain.submit(shed)
return built is not None
def _acquire(self, destination: "OtelDestination") -> SpanProcessor | None:
"""The processor for ``destination``, marked busy until ``_release``.
The build happens under the same lock that reads the cache, so a cold cache
met by a burst of concurrent requests yields one exporter rather than one per
thread with all but the winner shed. Building an exporter opens no connection,
so the cost of holding the lock is a constructor, once per destination.
"""
with self._lock:
if self._closed:
return None
processor: Final = self._cached_or_built_locked(destination, anchored=True)
if processor is None:
return None
self._exporting[id(processor)] = self._exporting.get(id(processor), 0) + 1
drained: Final = self._drainable_locked()
for shed in drained:
self._drain.submit(shed)
return processor
def _cached_or_built_locked(self, destination: "OtelDestination", *, anchored: bool) -> SpanProcessor | None:
"""The cached processor for ``destination``, or a new one if the drain can take it.
Every build past the cache cap sheds one processor into the drain, so while the
shed ones are stuck closing against a collector that stopped answering, a
destination that is not yet anchored is refused rather than parked behind them:
``deliverable`` then leaves its spans with the operator's exporter until the
drain catches up. One the request already anchored is rebuilt regardless. The
operator's exporter has stood down for it, so refusing here would drop the span,
and other tenants' auths can evict it in the meantime, with that eviction being
what tips the drain over. Eviction holds while the drain is saturated, so such a
rebuild costs the cache one entry rather than shedding another processor, and
the total stays at one per destination in flight.
"""
key: Final = destination.cache_key()
if (cached := self._processors.get(key)) is not None:
self._processors.move_to_end(key)
self._retire_overflow_locked()
return cached
if not anchored and self._drain.saturated():
verbose_logger.debug("OTel V2 fan-out: drain saturated, not building for %s", destination.endpoint)
return None
return self._build_locked(destination, key)
def _build_locked(self, destination: "OtelDestination", key: object) -> SpanProcessor | None:
built: Final = self._build(destination)
if built is None:
return None
self._processors[key] = built
self._retire_overflow_locked()
return built
def _release(self, processor: SpanProcessor) -> None:
with self._lock:
remaining: Final = self._exporting.get(id(processor), 1) - 1
if remaining > 0:
self._exporting[id(processor)] = remaining
else:
self._exporting.pop(id(processor), None)
if not self._exporting:
self._lock.notify_all()
drained: Final = self._drainable_locked()
for retired in drained:
self._drain.submit(retired)
def _retire_overflow_locked(self) -> None:
"""Move the LRU processor out of the cache once it is past the cap, drain permitting.
Eviction is what feeds the drain, and a destination a request already anchored
is rebuilt on its next span, which would shed another one. While the shed ones
are stuck closing against a collector that stopped answering, evicting would
churn the cache at one more processor, and one more batch thread, per span.
Holding above the cap instead keeps the total at one processor per destination
in flight, since ``deliverable`` anchors no new destination while the drain is
saturated. Once it has room again, every hit and build trims one entry.
"""
if len(self._processors) <= _MAX_CACHED_DESTINATION_PROCESSORS or self._drain.saturated():
return
_, evicted = self._processors.popitem(last=False)
self._retired[id(evicted)] = evicted
def _drainable_locked(self) -> tuple[SpanProcessor, ...]:
"""Retired processors no thread is exporting through, removed from the list.
``on_end`` holds a processor across an export, so closing an evicted one there
drops the span it is holding. A retiree is out of the cache and can never be
handed out again, so once its export count reaches zero it stays there.
"""
idle: Final = tuple(key for key in self._retired if self._exporting.get(key, 0) == 0)
return tuple(self._retired.pop(key) for key in idle)
def _destination_processor(destination: "OtelDestination") -> SpanProcessor | None:
"""A batching OTLP processor aimed at ``destination``, or ``None`` if unbuildable.
A protocol that resolves to a headerless exporter is unbuildable too: the
console fallback would swallow the tenant's credentials and print its spans to
the proxy's stdout while the operator's exporter stands down for them.
"""
kind: Final = destination.protocol or "otlp_http"
if exporter_transport(kind) == "headerless":
verbose_logger.debug("OTel V2 fan-out: no OTLP transport for protocol %r at %s", kind, destination.endpoint)
return None
try:
spec: Final = ExporterSpec(
kind=kind,
endpoint=destination.endpoint,
headers=destination.header_string(),
owner=None,
)
return _processor_for(_exporter_from_spec(spec), use_simple=False)
except Exception as exc: # noqa: BLE001 # a malformed destination must not break the request or the other destinations
verbose_logger.debug("OTel V2 fan-out: no processor for %s: %s", destination.endpoint, exc)
return None
def _shutdown_quietly(processor: SpanProcessor) -> None:
try:
processor.shutdown()
except Exception as exc: # noqa: BLE001 # defensive: shedding a spare processor must not raise
verbose_logger.debug("OTel V2 fan-out: discarding processor failed: %s", exc)
class _OverriddenBackendFilter(SpanProcessor):
"""Hold a span back from ``owner``'s operator-level exporter when the request
pointed ``owner`` at a tenant's own account.
Wrapping is the only place this works: ``SynchronousMultiSpanProcessor.on_end``
ignores return values, so a sibling processor can never veto the export.
Under ``additive`` mode nothing is suppressed, so the wrapper passes every span
straight through and the operator keeps its copy.
"""
def __init__(self, inner: SpanProcessor, owner: str) -> None:
self._inner: Final = inner
self._owner: Final = owner
def on_start(self, span: SDKSpan, parent_context: Context | None = None) -> None:
self._inner.on_start(span, parent_context)
def on_end(self, span: ReadableSpan) -> None:
if self._owner in suppressed_backends():
return
self._inner.on_end(span)
def shutdown(self) -> None:
self._inner.shutdown()
def force_flush(self, timeout_millis: int = 30000) -> bool:
return self._inner.force_flush(timeout_millis)
def build_span_exporter(config: OpenTelemetryV2Config) -> SpanExporter:
"""Build a single exporter from the top-level config fields.
@ -201,7 +781,14 @@ def build_span_exporter(config: OpenTelemetryV2Config) -> SpanExporter:
``exporter`` / ``endpoint`` / ``headers`` fields. To configure multiple
exporters, populate ``config.exporters`` directly.
"""
return _exporter_from_spec(ExporterSpec(kind=config.exporter, endpoint=config.endpoint, headers=config.headers))
return _exporter_from_spec(
ExporterSpec(
kind=config.exporter,
endpoint=config.endpoint,
traces_endpoint=config.traces_endpoint,
headers=config.headers,
)
)
def _otlp_metrics_endpoint(endpoint: str | None) -> str | None:
@ -437,6 +1024,7 @@ def build_tracer_provider(
exporter: SpanExporter | None = None,
baggage_processor: SpanProcessor | None = None,
use_simple_processor: bool | None = None,
tenant_overrides: bool = False,
) -> TracerProvider:
"""Build the shared :class:`TracerProvider`.
@ -445,6 +1033,13 @@ def build_tracer_provider(
``config.exporters`` entry — this is what fans spans out to multiple
backends. ``exporter`` and ``use_simple_processor`` are explicit overrides:
pass a single exporter to attach exactly that one (used by tests).
``tenant_overrides`` wraps each owned exporter so a request that pointed that
backend at a key's or team's own account skips it. Every v2 logger's provider
wants it, since any of them may own the overridden backend; delivering to the
tenant is a separate job, done once by :func:`attach_tenant_fan_out`. The
per-tenant providers this same function builds must leave it off, or they would
filter out the very spans they exist to carry.
"""
provider: Final = TracerProvider(resource=build_resource(config))
if baggage_processor is None:
@ -461,15 +1056,107 @@ def build_tracer_provider(
if spec.requires_headers and not spec.headers:
continue
exp = _exporter_from_spec(spec)
processor = _processor_for(
exp,
(spec.use_simple_processor if spec.use_simple_processor is not None else use_simple_processor),
)
owner = spec.owner.value if spec.owner is not None else None
provider.add_span_processor(
_processor_for(
exp,
(spec.use_simple_processor if spec.use_simple_processor is not None else use_simple_processor),
)
_OverriddenBackendFilter(processor, owner) if tenant_overrides and owner is not None else processor
)
return provider
_FAN_OUT_ATTACH_LOCK: Final = threading.Lock()
def attach_tenant_fan_out(provider: TracerProvider, *configs: OpenTelemetryV2Config) -> None:
"""Give ``provider`` the fan-out that delivers spans to key/team destinations.
Called on the one provider published as the OTel global, and idempotent so a
second publish (a test, a re-initialized proxy) cannot double-export. Concurrent
first calls (requests racing to anchor before any publish) serialize on one lock
so exactly one fan-out lands. ``configs`` name the operator's own exporters, one
config per v2 logger since each keeps its own provider and still writes its
account, so an additive destination pointing at any of them is delivered once
rather than twice.
"""
with _FAN_OUT_ATTACH_LOCK:
if any(isinstance(processor, TenantFanOutSpanProcessor) for processor in _attached_processors(provider)):
return
provider.add_span_processor(TenantFanOutSpanProcessor(operator_sinks=operator_sink_keys(*configs)))
def deliverable_destinations(
destinations: Iterable["OtelDestination"],
provider: trace.TracerProvider | None = None,
) -> tuple["OtelDestination", ...]:
"""The destinations a request can anchor, given what is published to carry them.
Anchoring a destination is what tells the operator's own exporter to stand down
for that backend, so one nothing can deliver has to be dropped here: with no
fan-out attached, or with an exporter that will not build, the request keeps
exactly the routing it would have had without any override.
"""
fan_out: Final = next(
(
processor
for processor in _attached_processors(provider if provider is not None else trace.get_tracer_provider())
if isinstance(processor, TenantFanOutSpanProcessor)
),
None,
)
return fan_out.deliverable(destinations) if fan_out is not None else ()
def operator_sink_keys(*configs: OpenTelemetryV2Config) -> frozenset[_SinkKey]:
"""The accounts the operator's own exporters write to, in destination terms.
Every v2 logger's config counts, since each logger exports through its own
provider. An exporter with no endpoint of its own resolves one from the
environment at export time, so it has no comparable identity and is left out,
and so is one that never reaches the wire: a console kind ignores the endpoint,
and a header-gated spec with no credentials is skipped when the provider is built.
"""
return frozenset(
key
for config in configs
for spec in config.exporters
if _exports_to_the_wire(spec) and (key := _sink_key(spec.endpoint, parse_headers(spec.headers))) is not None
)
def _exports_to_the_wire(spec: ExporterSpec) -> bool:
"""Whether ``build_tracer_provider`` gives ``spec`` an exporter that sends OTLP."""
return exporter_transport(spec.kind) != "headerless" and not (spec.requires_headers and not spec.headers)
def _sink_key(endpoint: str | None, headers: Mapping[str, str]) -> "_SinkKey | None":
"""The account an exporter writes to, or ``None`` when it has no fixed one.
Normalized on the three counts that make one account look like two: the operator's
spec carries the signal path a tenant destination leaves for the exporter to
append, header names survive one round trip lowercased and the other not, and one
credential answers to more than one name (see :data:`_CREDENTIAL_ALIASES`).
"""
normalized: Final = _otlp_traces_endpoint(endpoint)
if normalized is None:
return None
return (normalized, tuple(sorted((_credential_name(name), value) for name, value in headers.items())))
def _credential_name(header: str) -> str:
"""The credential a header carries, under whichever name the backend spells it."""
normalized: Final = header.strip().lower().replace("-", "_")
return _CREDENTIAL_ALIASES.get(normalized, normalized)
def _attached_processors(provider: trace.TracerProvider) -> "tuple[SpanProcessor, ...]":
"""The processors already on ``provider``, or empty when the SDK hides them."""
multi: Final = getattr(provider, "_active_span_processor", None)
return tuple(getattr(multi, "_span_processors", ()))
def get_tracer(provider: TracerProvider, name: str = "litellm") -> Tracer:
# Stamp the instrumentation scope with the LiteLLM package version so every
# emitted span carries a deterministic ``scope.version`` (the standard OTel

16
litellm/integrations/otel/plumbing/routing.py
View file

@ -25,6 +25,7 @@ from opentelemetry.trace import Tracer
from litellm._logging import verbose_logger
from litellm.constants import OTEL_SERVICE_NAME_METADATA_KEYS
from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
from litellm.integrations.otel.plumbing.context import destination_backends
from litellm.integrations.otel.plumbing.providers import (
build_tracer_provider,
exporter_transport,
@ -231,10 +232,21 @@ class TenantTracerCache:
concurrent overflow eviction can't shut it down between selection and
the caller's span start. The caller must ``release`` it exactly once.
"""
# A backend with a destination is delivered by the fan-out processor, which
# carries the whole trace and already carries this tenant's credentials and
# service name. Routing here too would detach this span onto a second provider,
# so the tenant would get the request tree plus a stray one-span trace.
if self._callback_name is not None and self._callback_name in destination_backends():
return TenantRoute(tracer=default, detached=False)
credential_headers: Final = self._credential_headers(dynamic_params)
project_headers: Final = self._project_headers(auth_metadata)
service_name: Final = tenant_service_name(auth_metadata)
if not credential_headers and not project_headers and service_name is None:
tenant_account: Final = bool(credential_headers) or bool(project_headers)
# A service name on its own only relabels the operator's own backend, so moving
# the span to a second provider for it while some other backend has a
# destination would drop the model call out of the trace the fan-out delivers.
# The destination stamps the same service name itself.
if not tenant_account and (service_name is None or destination_backends()):
return TenantRoute(tracer=default, detached=False)
# A fixed per-integration region endpoint (New Relic us/eu), never a
# caller-supplied host; ``None`` keeps the preset's own endpoint.
@ -255,7 +267,7 @@ class TenantTracerCache:
_shutdown_provider(evicted)
return TenantRoute(
tracer=get_tracer(provider, self._tracer_name),
detached=bool(project_headers) or bool(credential_headers),
detached=tenant_account,
provider=provider,
)

1
litellm/integrations/otel/presets/agentops.py
View file

@ -39,6 +39,7 @@ class _AgentOpsSettings(BaseSettings):
def agentops_preset(
*,
config_overrides: OpenTelemetryV2Config | None = None,
allow_missing_credentials: bool = False,
) -> OpenTelemetryV2Config:
"""Build the AgentOps config without any network I/O.

6
litellm/integrations/otel/presets/arize.py
View file

@ -26,10 +26,12 @@ class _ArizeSettings(BaseSettings):
def arize_preset(
*,
config_overrides: OpenTelemetryV2Config | None = None,
allow_missing_credentials: bool = False,
) -> OpenTelemetryV2Config:
base: Final = config_overrides or OpenTelemetryV2Config()
mappers: Final = ensure_mappers(base.mapper_names, "openinference")
arize_cfg: Final = _V1ArizeLogger.get_arize_config()
headers: Final = _arize_headers(arize_cfg)
base: Final = config_overrides or OpenTelemetryV2Config()
return base.model_copy(
update={
"exporters": [
@ -41,7 +43,7 @@ def arize_preset(
owner=ExporterOwner.ARIZE_AX,
),
],
"mapper_names": ensure_mappers(base.mapper_names, "openinference"),
"mapper_names": mappers,
"resource_attributes": {
**base.resource_attributes,
**({"model_id": arize_cfg.project_name} if arize_cfg.project_name else {}),

14
litellm/integrations/otel/presets/base.py
View file

@ -18,6 +18,18 @@ class Preset(Protocol):
``config_overrides`` lets one preset layer onto another's config (or onto
test-supplied defaults); the factory calls presets with no arguments.
``allow_missing_credentials`` lets a credential-mandatory backend (langfuse and
weave) degrade to an exporter-less, mapper-only config instead of raising when the
operator set no env credentials of their own. That is a real
deployment: every team brings its own account and the operator keeps none, and
without it the whole V2 path silently falls back to the legacy integration, so
no team destination is ever reached. Credential-optional backends ignore it.
"""
def __call__(self, *, config_overrides: OpenTelemetryV2Config | None = None) -> OpenTelemetryV2Config: ...
def __call__(
self,
*,
config_overrides: OpenTelemetryV2Config | None = None,
allow_missing_credentials: bool = False,
) -> OpenTelemetryV2Config: ...

152
litellm/integrations/otel/presets/destinations.py Normal file
View file

@ -0,0 +1,152 @@
"""Map a key's or team's callback vars to the OTLP destination its traces export to.
Header building is delegated to each preset's existing ``*_dynamic_headers`` builder,
so a destination authenticates exactly the way the per-request tracer route already
did; only the endpoint and transport need a per-backend rule.
"""
import os
from collections.abc import Callable, Mapping
from functools import lru_cache
from types import MappingProxyType
from typing import Final
import litellm
from litellm._logging import verbose_logger
from litellm.integrations.otel.model.destination import OtelDestination
from litellm.litellm_core_utils.url_utils import is_url_destination_allowed_by_host
from litellm.types.utils import StandardCallbackDynamicParams
#: An endpoint plus the OTLP transport to reach it with, or ``None`` when the backend
#: names no destination. The transport is ``None`` where the backend has only one.
_Destination = tuple[str, str | None]
@lru_cache(maxsize=128)
def _warn_host_not_allowlisted(host: str) -> None:
"""Cached so one misconfigured team logs once rather than once per request."""
verbose_logger.warning(
"OTel V2: not exporting to key/team Langfuse host '%s'. Add it to "
"litellm_settings.provider_url_destination_allowed_hosts to permit it",
host,
)
def _langfuse_destination(params: StandardCallbackDynamicParams) -> "_Destination | None":
"""The tenant's own Langfuse host, else the operator's, else Langfuse US cloud.
A host the tenant named has to be allowlisted by the operator, the same way a
URL-valued ``model`` is: anyone who can mint a key can write it, and it becomes an
endpoint the proxy posts the request's whole trace to, carrying the tenant's own
credentials. The operator's own ``LANGFUSE_HOST`` is not checked, since an internal
collector there is a deployment choice.
"""
from litellm.integrations.langfuse.langfuse_otel import (
LANGFUSE_CLOUD_US_ENDPOINT,
LangfuseOtelLogger,
)
tenant_host: Final = params.get("langfuse_host") or None
host: Final = tenant_host or LangfuseOtelLogger._get_langfuse_otel_host() # pyright: ignore[reportPrivateUsage] # reuse the backend's own env host resolver rather than duplicating it
if not host:
return (LANGFUSE_CLOUD_US_ENDPOINT, None)
normalized: Final = host if host.startswith("http") else f"https://{host}"
endpoint: Final = f"{normalized.rstrip('/')}/api/public/otel"
if tenant_host is None:
return (endpoint, None)
if not is_url_destination_allowed_by_host(endpoint, litellm.provider_url_destination_allowed_hosts):
_warn_host_not_allowlisted(host)
return None
return (endpoint, None)
def _arize_destination(params: StandardCallbackDynamicParams) -> "_Destination | None":
from litellm.integrations.arize.arize import ArizeLogger
config: Final = ArizeLogger.get_arize_config()
return (config.endpoint, config.protocol)
def _weave_destination(params: StandardCallbackDynamicParams) -> "_Destination | None":
from litellm.integrations.weave.weave_otel import weave_otel_endpoint
return (weave_otel_endpoint(os.environ.get("WANDB_HOST")), None)
def _newrelic_destination(params: StandardCallbackDynamicParams) -> "_Destination | None":
from litellm.integrations.otel.presets.newrelic import newrelic_dynamic_endpoint
endpoint: Final = newrelic_dynamic_endpoint(params)
return (endpoint, None) if endpoint else None
#: Callback name -> destination resolver. A backend is destination-capable exactly
#: when it appears here AND in ``DYNAMIC_HEADERS_BY_CALLBACK``: without a header
#: builder the destination would carry no tenant credentials, and the exporter
#: would post the tenant's traffic to the operator's account.
_DESTINATION_BY_CALLBACK: Final[Mapping[str, Callable[[StandardCallbackDynamicParams], "_Destination | None"]]] = (
MappingProxyType(
{
"langfuse_otel": _langfuse_destination,
"arize": _arize_destination,
"weave_otel": _weave_destination,
"newrelic": _newrelic_destination,
}
)
)
#: Headers a destination must carry to authenticate. Several dynamic-header builders
#: gate each credential independently, so a half-configured backend yields a non-empty
#: but unusable header set; accepting it would suppress the operator's own exporter and
#: send the request's whole trace where it cannot be stored.
_REQUIRED_HEADERS_BY_CALLBACK: Final[Mapping[str, frozenset[str]]] = MappingProxyType(
{
"langfuse_otel": frozenset({"Authorization"}),
"arize": frozenset({"arize-space-id", "api_key"}),
"weave_otel": frozenset({"Authorization", "project_id"}),
"newrelic": frozenset({"api-key"}),
}
)
_NO_ATTRS: Final[Mapping[str, str]] = MappingProxyType({})
def destination_capable_backends() -> frozenset[str]:
"""Backends a key or team can point at its own account."""
from litellm.integrations.otel.presets import DYNAMIC_HEADERS_BY_CALLBACK
return frozenset(_DESTINATION_BY_CALLBACK) & frozenset(DYNAMIC_HEADERS_BY_CALLBACK)
def destination_for(
callback_name: str,
params: StandardCallbackDynamicParams,
service_name: str | None = None,
) -> OtelDestination | None:
"""The destination ``params`` names for ``callback_name``, or ``None``.
``None`` means the caller configured nothing usable for this backend, so the
request keeps the operator's global exporters. ``service_name`` is the key's or
team's ``otel_service_name``, which the per-request tracer route applies when the
backend is not overridden and the destination has to apply once it is.
"""
from litellm.integrations.otel.presets import DYNAMIC_HEADERS_BY_CALLBACK
header_builder: Final = DYNAMIC_HEADERS_BY_CALLBACK.get(callback_name)
destination_builder: Final = _DESTINATION_BY_CALLBACK.get(callback_name)
if header_builder is None or destination_builder is None:
return None
headers: Final = header_builder(params)
if not headers or not _REQUIRED_HEADERS_BY_CALLBACK[callback_name] <= frozenset(headers):
return None
resolved: Final = destination_builder(params)
if resolved is None:
return None
endpoint, protocol = resolved
return OtelDestination(
endpoint=endpoint,
headers=MappingProxyType(dict(headers)), # mutable-ok: MappingProxyType needs a concrete mapping to wrap
resource_attributes=MappingProxyType({"service.name": service_name}) if service_name else _NO_ATTRS,
callback_name=callback_name,
protocol=protocol,
)

23
litellm/integrations/otel/presets/langfuse.py
View file

@ -10,17 +10,32 @@ from litellm.integrations.otel.model.config import (
ExporterSpec,
OpenTelemetryV2Config,
)
from litellm.integrations.otel.presets.utils import ensure_mappers
from litellm.integrations.otel.presets.utils import (
credential_gated_exporters,
ensure_mappers,
)
from litellm.types.utils import StandardCallbackDynamicParams
def langfuse_preset(
*,
config_overrides: OpenTelemetryV2Config | None = None,
allow_missing_credentials: bool = False,
) -> OpenTelemetryV2Config:
cfg: Final = _V1Langfuse.get_langfuse_otel_config()
kind: Final = cfg.exporter if isinstance(cfg.exporter, str) else "otlp_http"
base: Final = config_overrides or OpenTelemetryV2Config()
mappers: Final = ensure_mappers(base.mapper_names, "langfuse")
try:
cfg: Final = _V1Langfuse.get_langfuse_otel_config()
except Exception:
if not allow_missing_credentials:
raise
return base.model_copy(
update={ # mutable-ok: pydantic model_copy takes a plain update mapping
"exporters": credential_gated_exporters(base.exporters, ExporterOwner.LANGFUSE_OTEL),
"mapper_names": mappers,
}
)
kind: Final = cfg.exporter if isinstance(cfg.exporter, str) else "otlp_http"
return base.model_copy(
update={
"exporters": [
@ -32,7 +47,7 @@ def langfuse_preset(
owner=ExporterOwner.LANGFUSE_OTEL,
),
],
"mapper_names": ensure_mappers(base.mapper_names, "langfuse"),
"mapper_names": mappers,
}
)

1
litellm/integrations/otel/presets/langtrace.py
View file

@ -9,6 +9,7 @@ from litellm.integrations.otel.presets.utils import ensure_mappers
def langtrace_preset(
*,
config_overrides: OpenTelemetryV2Config | None = None,
allow_missing_credentials: bool = False,
) -> OpenTelemetryV2Config:
"""Compose the Langtrace mapper on top of the customer's OTLP destination.

1
litellm/integrations/otel/presets/levo.py
View file

@ -13,6 +13,7 @@ from litellm.integrations.otel.model.config import (
def levo_preset(
*,
config_overrides: OpenTelemetryV2Config | None = None,
allow_missing_credentials: bool = False,
) -> OpenTelemetryV2Config:
cfg: Final = _V1Levo.get_levo_config()
base: Final = config_overrides or OpenTelemetryV2Config()

1
litellm/integrations/otel/presets/newrelic.py
View file

@ -44,6 +44,7 @@ class _NewRelicSettings(BaseSettings):
def newrelic_preset(
*,
config_overrides: OpenTelemetryV2Config | None = None,
allow_missing_credentials: bool = False,
) -> OpenTelemetryV2Config:
settings: Final = _NewRelicSettings()
base: Final = config_overrides or OpenTelemetryV2Config()

1
litellm/integrations/otel/presets/phoenix.py
View file

@ -60,6 +60,7 @@ def phoenix_project_headers(auth_metadata: Mapping[str, str] | None) -> Mapping[
def phoenix_preset(
*,
config_overrides: OpenTelemetryV2Config | None = None,
allow_missing_credentials: bool = False,
) -> OpenTelemetryV2Config:
cfg: Final = _V1Phoenix.get_arize_phoenix_config()
headers: Final = cfg.otlp_auth_headers if hasattr(cfg, "otlp_auth_headers") else None

31
litellm/integrations/otel/presets/utils.py
View file

@ -3,6 +3,8 @@
from collections.abc import Iterable
from typing import Final
from litellm.integrations.otel.model.config import ExporterOwner, ExporterSpec
def ensure_mappers(mapper_names: Iterable[str], *names: str) -> list[str]:
"""Return ``mapper_names`` with each of ``names`` appended if not already present.
@ -15,3 +17,32 @@ def ensure_mappers(mapper_names: Iterable[str], *names: str) -> list[str]:
if name not in result:
result.append(name)
return result
def credential_gated_exporters(
exporters: "Iterable[ExporterSpec]", owner: "ExporterOwner"
) -> "tuple[ExporterSpec, ...]":
"""``exporters`` with the operator's destination replaced by a header-gated one.
Used when a credential-mandatory backend is asked to build without the operator's
own credentials, so only key/team destinations receive spans. Two things have to
happen for that to mean "export nowhere": the placeholder console spec that
``OpenTelemetryV2Config`` folds in for an empty exporter list is dropped, or every
span would be printed to stdout, and the gated spec keeps the owner so the
override filter still recognises which backend this provider speaks for.
"""
return (
*(spec for spec in exporters if not is_unconfigured_placeholder(spec)),
ExporterSpec(owner=owner, requires_headers=True),
)
def is_unconfigured_placeholder(spec: "ExporterSpec") -> bool:
"""Whether ``spec`` is the one ``_normalize`` folds in when nothing was configured.
No field set is what says the operator asked for nothing: an exporter they did
configure survives, even ``OTEL_EXPORTER=console`` whose value matches the default,
and so does the gated spec this module appends, which would otherwise eat itself
when one preset layers onto another.
"""
return not spec.model_fields_set

21
litellm/integrations/otel/presets/weave.py
View file

@ -7,7 +7,10 @@ from litellm.integrations.otel.model.config import (
ExporterSpec,
OpenTelemetryV2Config,
)
from litellm.integrations.otel.presets.utils import ensure_mappers
from litellm.integrations.otel.presets.utils import (
credential_gated_exporters,
ensure_mappers,
)
from litellm.integrations.weave.weave_otel import (
_get_weave_authorization_header,
get_weave_otel_config,
@ -18,9 +21,21 @@ from litellm.types.utils import StandardCallbackDynamicParams
def weave_preset(
*,
config_overrides: OpenTelemetryV2Config | None = None,
allow_missing_credentials: bool = False,
) -> OpenTelemetryV2Config:
weave_cfg: Final = get_weave_otel_config()
base: Final = config_overrides or OpenTelemetryV2Config()
mappers: Final = ensure_mappers(base.mapper_names, "openinference", "weave")
try:
weave_cfg: Final = get_weave_otel_config()
except Exception:
if not allow_missing_credentials:
raise
return base.model_copy(
update={ # mutable-ok: pydantic model_copy takes a plain update mapping
"exporters": credential_gated_exporters(base.exporters, ExporterOwner.WEAVE_OTEL),
"mapper_names": mappers,
}
)
return base.model_copy(
update={
"exporters": [
@ -33,7 +48,7 @@ def weave_preset(
),
],
# Weave consumes OpenInference + a small Weave-specific overlay.
"mapper_names": ensure_mappers(base.mapper_names, "openinference", "weave"),
"mapper_names": mappers,
}
)

20
litellm/integrations/weave/weave_otel.py
View file

@ -117,6 +117,14 @@ def _get_weave_authorization_header(api_key: str) -> str:
return f"Basic {auth_header}"
def weave_otel_endpoint(host: str | None) -> str:
"""The OTLP traces endpoint for a self-managed ``host``, else Weave cloud."""
if not host:
return WEAVE_BASE_URL + WEAVE_OTEL_ENDPOINT
normalized: Final = host if host.startswith("http") else f"https://{host}"
return normalized.rstrip("/") + WEAVE_OTEL_ENDPOINT
def get_weave_otel_config() -> WeaveOtelConfig:
"""
Retrieves the Weave OpenTelemetry configuration based on environment variables.
@ -134,7 +142,6 @@ def get_weave_otel_config() -> WeaveOtelConfig:
"""
api_key: Final = os.getenv("WANDB_API_KEY")
project_id: Final = os.getenv("WANDB_PROJECT_ID")
host = os.getenv("WANDB_HOST")
if not api_key:
raise ValueError("WANDB_API_KEY must be set for Weave OpenTelemetry integration.")
@ -144,15 +151,8 @@ def get_weave_otel_config() -> WeaveOtelConfig:
"WANDB_PROJECT_ID must be set for Weave OpenTelemetry integration. Format: <entity>/<project_name>"
)
if host:
if not host.startswith("http"):
host = "https://" + host
# Self-managed instances use a different path
endpoint = host.rstrip("/") + WEAVE_OTEL_ENDPOINT
verbose_logger.debug("Using Weave OTEL endpoint from host: %s", endpoint)
else:
endpoint = WEAVE_BASE_URL + WEAVE_OTEL_ENDPOINT
verbose_logger.debug("Using Weave cloud endpoint: %s", endpoint)
endpoint: Final = weave_otel_endpoint(os.getenv("WANDB_HOST"))
verbose_logger.debug("Using Weave OTEL endpoint: %s", endpoint)
# Weave uses Basic auth with format: api:<WANDB_API_KEY>
auth_header: Final = _get_weave_authorization_header(api_key=api_key)

37
litellm/litellm_core_utils/core_helpers.py
View file

@ -1,10 +1,12 @@
# What is this?
## Helper utilities
import copy
import logging
from collections.abc import Iterable, Mapping
from typing import TYPE_CHECKING, Any, Final, Literal
import httpx
from pydantic import TypeAdapter, ValidationError
from litellm._logging import verbose_logger
from litellm.types.llms.openai import AllMessageValues, OpenAIChatCompletionFinishReason
@ -37,6 +39,41 @@ def safe_divide_seconds(seconds: float, denominator: float, default: float | Non
return float(seconds / denominator)
_DROP_PARAMS_BOOL: Final = TypeAdapter(bool)
def normalize_drop_params(value: object) -> bool | None:
if value is None or isinstance(value, bool):
return value
try:
return _DROP_PARAMS_BOOL.validate_python(value.strip() if isinstance(value, str) else value)
except ValidationError:
return None
def drop_params_flag(value: object, source: str, logger: logging.Logger) -> bool:
normalized: Final = normalize_drop_params(value)
if normalized is None and value is not None:
logger.warning("%s=%r is not a flag value, treating it as off", source, value)
return bool(normalized)
DROP_PARAMS_ENV_VAR: Final = "LITELLM_DROP_PARAMS"
def drop_params_env_flag(environ: Mapping[str, str], logger: logging.Logger) -> bool:
configured: Final = environ.get(DROP_PARAMS_ENV_VAR, "").strip()
if configured == "":
return False
normalized: Final = normalize_drop_params(configured)
if normalized is None:
logger.warning(
"%s=%r is not a flag value, treating it as on. Set it to true or false", DROP_PARAMS_ENV_VAR, configured
)
return True
return normalized
def safe_divide(
numerator: float,
denominator: float,

5
litellm/litellm_core_utils/get_litellm_params.py
View file

@ -2,6 +2,7 @@ from collections.abc import Mapping, MutableMapping
from types import MappingProxyType
from typing import Final
from litellm.litellm_core_utils.core_helpers import normalize_drop_params
from litellm.llms.openai.data_residency import infer_openai_data_residency
AWS_CREDENTIAL_KWARGS_KEYS: Final = frozenset(
@ -113,7 +114,7 @@ def get_litellm_params(
custom_prompt_dict: dict | None = None,
litellm_metadata: dict | None = None,
disable_add_transform_inline_image_block: bool | None = None,
drop_params: bool | None = None,
drop_params: bool | str | None = None,
prompt_id: str | None = None,
prompt_variables: dict | None = None,
async_call: bool | None = None,
@ -175,7 +176,7 @@ def get_litellm_params(
"custom_prompt_dict": custom_prompt_dict,
"litellm_metadata": litellm_metadata,
"disable_add_transform_inline_image_block": disable_add_transform_inline_image_block,
"drop_params": drop_params,
"drop_params": normalize_drop_params(drop_params),
"prompt_id": prompt_id,
"prompt_variables": prompt_variables,
"async_call": async_call,

104
litellm/litellm_core_utils/litellm_logging.py
View file

@ -48,6 +48,7 @@ from litellm.constants import (
)
from litellm.cost_calculator import (
RealtimeAPITokenUsageProcessor,
ResponsesWebSocketTokenUsageProcessor,
_select_model_name_for_cost_calc,
)
from litellm.exceptions import (
@ -201,6 +202,7 @@ if TYPE_CHECKING:
from mcp.types import EmbeddedResource, ImageContent, TextContent
from litellm.integrations.otel.logger import OpenTelemetryV2
from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
from litellm.llms.base_llm.passthrough.transformation import BasePassthroughConfig
try:
from litellm_enterprise.enterprise_callbacks.callback_controls import (
@ -447,6 +449,13 @@ def _provider_response_id(source: object) -> str | None:
return candidate if isinstance(candidate, str) and candidate else None
def mask_api_base_credentials(api_base: str) -> str:
if "key=" not in api_base:
return api_base
key_end: Final = api_base.find("key=") + 4
return api_base[:key_end] + "*" * 5 + api_base[-4:]
class Logging(LiteLLMLoggingBaseClass):
global \
supabaseClient, \
@ -1189,14 +1198,7 @@ class Logging(LiteLLMLoggingBaseClass):
return data
def _get_masked_api_base(self, api_base: str) -> str:
if "key=" in api_base:
# Find the position of "key=" in the string
key_index: Final = api_base.find("key=") + 4
# Mask the last 5 characters after "key="
masked_api_base = api_base[:key_index] + "*" * 5 + api_base[-4:]
else:
masked_api_base = api_base
return str(masked_api_base)
return str(mask_api_base_credentials(api_base))
def _pre_call(self, input, api_key, model=None, additional_args={}):
"""
@ -2028,6 +2030,17 @@ class Logging(LiteLLMLoggingBaseClass):
results=result,
)
elif self.call_type == CallTypes.aresponses_websocket.value and isinstance(result, list): # pyright: ignore[reportUnknownMemberType] # Logging.call_type is untyped
combined_ws_usage: Final = (
ResponsesWebSocketTokenUsageProcessor.collect_and_combine_usage_from_responses_ws_results(
results=result # pyright: ignore[reportUnknownArgumentType] # raw event dicts from the WS stream
)
)
logging_result = LiteLLMRealtimeStreamLoggingObject(
usage=combined_ws_usage,
results=result, # pyright: ignore[reportUnknownArgumentType] # raw event dicts from the WS stream
)
elif (
self.call_type == CallTypes.llm_passthrough_route.value
or self.call_type == CallTypes.allm_passthrough_route.value
@ -2938,6 +2951,19 @@ class Logging(LiteLLMLoggingBaseClass):
result._hidden_params["batch_successful_requests"] = batch_successful_requests # pyright: ignore[reportPrivateUsage] # rebind-ok: same result._hidden_params pattern as response_cost/batch_models above
result._hidden_params["batch_failed_requests"] = batch_failed_requests # pyright: ignore[reportPrivateUsage] # rebind-ok: same pattern as above
result.usage = batch_usage
batch_prompt_cost: Final = kwargs.get("batch_prompt_cost", None)
batch_completion_cost: Final = kwargs.get("batch_completion_cost", None)
if (
isinstance(batch_prompt_cost, float)
and isinstance(batch_completion_cost, float)
and isinstance(batch_cost, float)
):
self.set_cost_breakdown(
input_cost=batch_prompt_cost,
output_cost=batch_completion_cost,
total_cost=batch_cost,
cost_for_built_in_tools_cost_usd_dollar=0.0,
)
elif should_compute_batch_data:
batch_result: Final = await _handle_completed_batch(
@ -2953,6 +2979,12 @@ class Logging(LiteLLMLoggingBaseClass):
result._hidden_params["batch_successful_requests"] = batch_result.successful_requests # pyright: ignore[reportPrivateUsage] # rebind-ok: same pattern as above
result._hidden_params["batch_failed_requests"] = batch_result.failed_requests # pyright: ignore[reportPrivateUsage] # rebind-ok: same pattern as above
result.usage = batch_result.usage
self.set_cost_breakdown(
input_cost=batch_result.prompt_cost,
output_cost=batch_result.completion_cost,
total_cost=batch_result.cost,
cost_for_built_in_tools_cost_usd_dollar=0.0,
)
self.truncated_messages_for_logging = await truncate_base64_in_messages_async(
StandardLoggingPayloadSetup.append_system_prompt_messages(
@ -4819,31 +4851,83 @@ def _maybe_construct_otel_v2(callback_name: str, _in_memory_loggers: list[Custom
Returns ``None`` when V2 is off OR when there's no preset registered for
``callback_name`` — callers should then fall through to the legacy path.
A preset that needs operator credentials it cannot find is allowed to build
only when this request has a key/team destination for that backend and another
V2 logger is already registered to carry the fan-out. The resulting logger keeps
only its credential-gated exporter, while the registered logger owns operator
delivery. Without that carrier, a preset that raises or that ends up with nothing
but its gated exporter and the default console placeholder returns ``None``, so the
caller falls through to the legacy path exactly as before V2 landed.
"""
from litellm.integrations.otel.model.config import is_otel_v2_enabled
if not is_otel_v2_enabled():
return None
from litellm.integrations.otel.logger import OpenTelemetryV2, build_otel_v2_logger
from litellm.integrations.otel.plumbing.context import destination_backends
from litellm.integrations.otel.presets import PRESET_BY_CALLBACK
preset_fn: Final = PRESET_BY_CALLBACK.get(callback_name)
if preset_fn is None:
return None
serves_a_destination: Final = callback_name in destination_backends()
has_v2_logger: Final = any(isinstance(callback, OpenTelemetryV2) for callback in _in_memory_loggers)
carried: Final = serves_a_destination and has_v2_logger
for callback in _in_memory_loggers:
if isinstance(callback, OpenTelemetryV2) and getattr(callback, "callback_name", None) == callback_name:
if (
isinstance(callback, OpenTelemetryV2)
and getattr(callback, "callback_name", None) == callback_name
and (serves_a_destination or not _exports_nowhere(callback.config))
):
return callback
try:
config: Final = preset_fn()
built: Final = preset_fn(allow_missing_credentials=carried)
except Exception:
# If env vars are missing or the preset raises, defer to the legacy path
# so customers get the same error story they had before V2 landed.
return None
gated: Final = _is_credential_gated(built)
if gated and not carried and not _has_operator_exporter(built):
return None
config: Final = _only_the_gated_exporter(built) if gated and carried else built
if _exports_nowhere(config):
verbose_logger.warning(
"OTel V2: no operator credentials for '%s'; only key/team destinations will receive its traces",
callback_name,
)
v2_logger: Final = build_otel_v2_logger(config=config, callback_name=callback_name)
_in_memory_loggers.append(v2_logger)
return v2_logger
def _exports_nowhere(config: "OpenTelemetryV2Config") -> bool:
"""Whether every exporter in ``config`` is waiting on credentials it never got."""
return all(_is_gated(spec) for spec in config.exporters)
def _is_credential_gated(config: "OpenTelemetryV2Config") -> bool:
"""Whether the preset built without the operator's own credentials for its backend."""
return any(_is_gated(spec) for spec in config.exporters)
def _has_operator_exporter(config: "OpenTelemetryV2Config") -> bool:
"""Whether the operator configured somewhere real to export, beyond the default console placeholder."""
from litellm.integrations.otel.presets.utils import is_unconfigured_placeholder
return any(not _is_gated(spec) and not is_unconfigured_placeholder(spec) for spec in config.exporters)
def _only_the_gated_exporter(config: "OpenTelemetryV2Config") -> "OpenTelemetryV2Config":
return config.model_copy(
update={"exporters": [spec for spec in config.exporters if _is_gated(spec)]} # mutable-ok: model_copy update
)
def _is_gated(spec: "ExporterSpec") -> bool:
return spec.requires_headers and not spec.headers
def _maybe_auto_initialize_arize_phoenix(_in_memory_loggers: list[CustomLogger]) -> None:
"""
Auto-initialize ArizePhoenixLogger when Phoenix env vars are detected.

13
litellm/litellm_core_utils/prompt_templates/image_handling.py
View file

@ -183,6 +183,7 @@ class _RemoteSource:
class RemoteMedia:
url: str
fields: Mapping[str, object]
part_type: str
_NO_FIELDS: Final[Mapping[str, object]] = MappingProxyType({})
@ -192,6 +193,10 @@ def inline_every_remote_url(_media: RemoteMedia) -> bool:
return True
def inline_remote_image_urls(media: RemoteMedia) -> bool:
return media.part_type == "image_url"
def _parse_remote_image(fields: Mapping[str, object]) -> _RemoteImage | None:
if fields.get("type") != "image_url":
return None
@ -223,11 +228,11 @@ def _parse_remote_part(part: object) -> _RemoteImage | _RemoteFile | _RemoteSour
def _remote_media(remote: _RemoteImage | _RemoteFile | _RemoteSource) -> RemoteMedia:
match remote:
case _RemoteImage(_, image_url, url):
return RemoteMedia(url, image_url if image_url is not None else _NO_FIELDS)
return RemoteMedia(url, image_url if image_url is not None else _NO_FIELDS, "image_url")
case _RemoteFile(_, file, url):
return RemoteMedia(url, file)
case _RemoteSource(_, source, url):
return RemoteMedia(url, source)
return RemoteMedia(url, file, "file")
case _RemoteSource(part, source, url):
return RemoteMedia(url, source, str(part.get("type")))
_PDF_FORMAT: Final = MappingProxyType({"format": "application/pdf"})

200
litellm/llms/anthropic/chat/handler.py
View file

@ -368,27 +368,92 @@ class AnthropicChatCompletion(BaseLLM):
if config is None:
raise ValueError(f"Provider config not found for model: {model} and provider: {custom_llm_provider}")
def build_request() -> tuple[dict, dict]: # mutable-ok: rewritten in place downstream
"""Translate the request the Python way, returning `(headers, data)`.
transform_params: Final = {**optional_params, "is_vertex_request": is_vertex_request}
def finish_request(request_data: dict) -> tuple[dict, dict]: # mutable-ok: rewritten in place downstream
"""Filter beta headers and emit pre_call, returning `(headers, data)`.
The pair stays mutable because the streaming path rewrites it in
place (`data["stream"] = True`) before sending.
Shared by the normal path and by the Rust path's fallback, which
builds it only when the Rust call did not serve the request.
place (`data["stream"] = True`) before sending. A Rust attempt that
declined already emitted pre_call for this request, so skip it there.
"""
request_data: Final = config.transform_request(
model=model,
messages=messages,
optional_params={**optional_params, "is_vertex_request": is_vertex_request},
litellm_params=litellm_params,
headers=headers,
)
return update_request_with_filtered_beta(
request_headers, data = update_request_with_filtered_beta(
headers=headers,
request_data=request_data,
provider=custom_llm_provider,
)
if not serves_via_rust:
logging_obj.pre_call(
input=messages,
api_key=api_key,
additional_args={
"complete_input_dict": data,
"api_base": api_base,
"headers": request_headers,
},
)
print_verbose(f"_is_function_call: {_is_function_call}")
return request_headers, data
async def acompletion_dispatch() -> "ModelResponse | CustomStreamWrapper":
"""Translate then send, so the provider config can inline remote media off the event loop."""
request_headers, data = finish_request(
await config.async_transform_request(
model=model,
messages=messages,
optional_params=transform_params,
litellm_params=litellm_params,
headers=headers,
)
)
if (
stream is True
): # if function call - fake the streaming (need complete blocks for output parsing in openai format)
print_verbose("makes async anthropic streaming POST request")
data["stream"] = stream
return await self.acompletion_stream_function(
model=model,
messages=messages,
data=data,
api_base=api_base,
custom_prompt_dict=custom_prompt_dict,
model_response=model_response,
print_verbose=print_verbose,
encoding=encoding,
api_key=api_key,
logging_obj=logging_obj,
optional_params=optional_params,
stream=stream,
_is_function_call=_is_function_call,
json_mode=json_mode,
litellm_params=litellm_params,
logger_fn=logger_fn,
headers=request_headers,
timeout=timeout,
client=(client if client is not None and isinstance(client, AsyncHTTPHandler) else None),
)
return await self.acompletion_function(
model=model,
messages=messages,
data=data,
api_base=api_base,
custom_prompt_dict=custom_prompt_dict,
model_response=model_response,
print_verbose=print_verbose,
encoding=encoding,
api_key=api_key,
provider_config=config,
logging_obj=logging_obj,
optional_params=optional_params,
stream=stream,
_is_function_call=_is_function_call,
litellm_params=litellm_params,
logger_fn=logger_fn,
headers=request_headers,
client=client,
json_mode=json_mode,
timeout=timeout,
)
# The Rust core owns the whole call for the subset it accepts, so ask
# before transforming: whichever path runs emits pre_call exactly once.
@ -424,35 +489,6 @@ class AnthropicChatCompletion(BaseLLM):
additional_args=rust_logging_args,
)
if acompletion is True:
async def python_fallback() -> "ModelResponse | CustomStreamWrapper":
# pre_call already fired for this request above. The Rust
# path only declines before the provider is called, so this
# is the same attempt continuing, not a second one.
fallback_headers, fallback_data = build_request()
return await self.acompletion_function(
model=model,
messages=messages,
data=fallback_data,
api_base=api_base,
custom_prompt_dict=custom_prompt_dict,
model_response=model_response,
print_verbose=print_verbose,
encoding=encoding,
api_key=api_key,
provider_config=config,
logging_obj=logging_obj,
optional_params=optional_params,
stream=stream,
_is_function_call=_is_function_call,
litellm_params=litellm_params,
logger_fn=logger_fn,
headers=fallback_headers,
client=client,
json_mode=json_mode,
timeout=timeout,
)
return rust_chat_completions_bridge.achat_completions_or_fallback(
model=model,
messages=messages,
@ -464,7 +500,7 @@ class AnthropicChatCompletion(BaseLLM):
extra_headers=headers,
timeout=timeout,
on_response=log_rust_post_call,
python_fallback=python_fallback,
python_fallback=acompletion_dispatch,
)
rust_response: Final = rust_chat_completions_bridge.chat_completions(
model=model,
@ -481,74 +517,18 @@ class AnthropicChatCompletion(BaseLLM):
if rust_response is not None:
return rust_response
headers, data = build_request()
## LOGGING
# Reaching here with `serves_via_rust` set means the Rust attempt
# declined at call time, before the provider was called, and already
# logged this request. That is the same attempt continuing.
if not serves_via_rust:
logging_obj.pre_call(
input=messages,
api_key=api_key,
additional_args={
"complete_input_dict": data,
"api_base": api_base,
"headers": headers,
},
)
print_verbose(f"_is_function_call: {_is_function_call}")
if acompletion is True:
if (
stream is True
): # if function call - fake the streaming (need complete blocks for output parsing in openai format)
print_verbose("makes async anthropic streaming POST request")
data["stream"] = stream
return self.acompletion_stream_function(
model=model,
messages=messages,
data=data,
api_base=api_base,
custom_prompt_dict=custom_prompt_dict,
model_response=model_response,
print_verbose=print_verbose,
encoding=encoding,
api_key=api_key,
logging_obj=logging_obj,
optional_params=optional_params,
stream=stream,
_is_function_call=_is_function_call,
json_mode=json_mode,
litellm_params=litellm_params,
logger_fn=logger_fn,
headers=headers,
timeout=timeout,
client=(client if client is not None and isinstance(client, AsyncHTTPHandler) else None),
)
else:
return self.acompletion_function(
model=model,
messages=messages,
data=data,
api_base=api_base,
custom_prompt_dict=custom_prompt_dict,
model_response=model_response,
print_verbose=print_verbose,
encoding=encoding,
api_key=api_key,
provider_config=config,
logging_obj=logging_obj,
optional_params=optional_params,
stream=stream,
_is_function_call=_is_function_call,
litellm_params=litellm_params,
logger_fn=logger_fn,
headers=headers,
client=client,
json_mode=json_mode,
timeout=timeout,
)
return acompletion_dispatch()
else:
headers, data = finish_request(
config.transform_request(
model=model,
messages=messages,
optional_params=transform_params,
litellm_params=litellm_params,
headers=headers,
)
)
## COMPLETION CALL
if (
stream is True

24
litellm/llms/anthropic/chat/transformation.py
View file

@ -26,6 +26,11 @@ from litellm.litellm_core_utils.core_helpers import map_finish_reason
from litellm.litellm_core_utils.prompt_templates.common_utils import (
sanitize_input_schema_for_anthropic,
)
from litellm.litellm_core_utils.prompt_templates.image_handling import (
RemoteMedia,
async_inline_remote_media,
inline_remote_image_urls,
)
from litellm.llms.base_llm.base_utils import type_to_response_format_param
from litellm.llms.base_llm.chat.transformation import BaseConfig, BaseLLMException
from litellm.types.llms.anthropic import (
@ -1840,6 +1845,25 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
break
return headers
def inlines_remote_media(self, media: RemoteMedia) -> bool:
return inline_remote_image_urls(media) and media.url.startswith("http://")
async def async_transform_request(
self,
model: str,
messages: list[AllMessageValues], # mutable-ok: BaseConfig signature
optional_params: dict[str, object], # mutable-ok: BaseConfig signature
litellm_params: dict[str, object], # mutable-ok: BaseConfig signature
headers: dict[str, object], # mutable-ok: BaseConfig signature
) -> dict[str, object]: # mutable-ok: BaseConfig signature
return self.transform_request(
model=model,
messages=await async_inline_remote_media(messages, should_inline=self.inlines_remote_media),
optional_params=optional_params,
litellm_params=litellm_params,
headers=headers,
)
def transform_request(
self,
model: str,

3
litellm/llms/base_llm/vector_store/transformation.py
View file

@ -121,6 +121,9 @@ class RouterVectorStoreEmbeddingExecutor:
class BaseVectorStoreConfig:
def validate_create_vector_store(self) -> None:
return None
def get_supported_openai_params(self, model: str) -> list[VECTOR_STORE_OPENAI_PARAMS]:
return []

21
litellm/llms/custom_httpx/llm_http_handler.py
View file

@ -9826,7 +9826,7 @@ class BaseLLMHTTPHandler:
vector_store_search_optional_params=vector_store_search_optional_params,
api_base=api_base,
litellm_logging_obj=logging_obj,
litellm_params=dict(litellm_params),
litellm_params=MappingProxyType(dict(litellm_params, timeout=timeout)),
extra_body=extra_body,
embedding_executor=embedding_executor,
)
@ -9871,6 +9871,12 @@ class BaseLLMHTTPHandler:
data=request_data,
timeout=timeout,
)
except httpx.TimeoutException:
raise vector_store_provider_config.get_error_class(
error_message="Vector store search exceeded the caller timeout.",
status_code=408,
headers=httpx.Headers(),
) from None
except Exception as e:
raise self._handle_error(e=e, provider_config=vector_store_provider_config)
@ -9955,7 +9961,7 @@ class BaseLLMHTTPHandler:
vector_store_search_optional_params=vector_store_search_optional_params,
api_base=api_base,
litellm_logging_obj=logging_obj,
litellm_params=dict(litellm_params),
litellm_params=MappingProxyType(dict(litellm_params, timeout=timeout)),
extra_body=extra_body,
embedding_executor=embedding_executor,
)
@ -10000,7 +10006,14 @@ class BaseLLMHTTPHandler:
url=url,
headers=headers,
data=request_data,
timeout=timeout,
)
except httpx.TimeoutException:
raise vector_store_provider_config.get_error_class(
error_message="Vector store search exceeded the caller timeout.",
status_code=408,
headers=httpx.Headers(),
) from None
except Exception as e:
raise self._handle_error(e=e, provider_config=vector_store_provider_config)
@ -10030,6 +10043,8 @@ class BaseLLMHTTPHandler:
else:
async_httpx_client = client
vector_store_provider_config.validate_create_vector_store()
headers: Final = vector_store_provider_config.validate_environment(
headers=extra_headers or {}, litellm_params=litellm_params
)
@ -10100,6 +10115,8 @@ class BaseLLMHTTPHandler:
else:
sync_httpx_client = client
vector_store_provider_config.validate_create_vector_store()
headers: Final = vector_store_provider_config.validate_environment(
headers=extra_headers or {}, litellm_params=litellm_params
)

303
litellm/llms/mongodb/common_utils.py
View file

@ -1,303 +0,0 @@
"""Shared helpers for the MongoDB integrations. pymongo lives in the optional ``mongodb`` extra,
so every import of it is deferred to call time."""
import asyncio
import threading
import weakref
from asyncio import AbstractEventLoop
from collections import OrderedDict
from collections.abc import Callable, Mapping
from dataclasses import dataclass
from types import MappingProxyType
from typing import TYPE_CHECKING, Final, TypeAlias, TypeVar
from litellm.exceptions import BadRequestError, ServiceUnavailableError, Timeout
if TYPE_CHECKING:
from pymongo import AsyncMongoClient, MongoClient
PYMONGO_INSTALL_HINT: Final = (
"The MongoDB vector store requires the 'pymongo' package. "
"Run 'pip install litellm[mongodb]' (or 'pip install pymongo') to install it."
)
MONGODB_PROVIDER: Final = "mongodb"
def config_error(message: str) -> BadRequestError:
"""400 rather than the 500 a bare ValueError becomes once litellm.exception_type wraps it."""
return BadRequestError(message=message, model=None, llm_provider=MONGODB_PROVIDER)
def timeout_error(message: str) -> Timeout:
return Timeout(message=message, model=None, llm_provider=MONGODB_PROVIDER)
def unavailable_error(message: str) -> ServiceUnavailableError:
"""litellm only retries 408, 409, 429 and 5xx, so a 400 here would make a failover permanent."""
return ServiceUnavailableError(message=message, model=None, llm_provider=MONGODB_PROVIDER)
DEFAULT_CONNECT_TIMEOUT_MS: Final = 10_000
DEFAULT_SOCKET_TIMEOUT_MS: Final = 30_000
DEFAULT_SERVER_SELECTION_TIMEOUT_MS: Final = 10_000
_MAX_CACHED_CLIENTS: Final = 32
_APP_NAME: Final = "litellm"
@dataclass(frozen=True, slots=True)
class MongoClientKey:
connection_string: str
connect_timeout_ms: int
socket_timeout_ms: int
server_selection_timeout_ms: int
SyncClientFactory: TypeAlias = Callable[..., "MongoClient"]
AsyncClientFactory: TypeAlias = Callable[..., "AsyncMongoClient"]
_K = TypeVar("_K")
_V = TypeVar("_V")
_AsyncClientCacheKey: TypeAlias = tuple[MongoClientKey, int]
# CPython recycles id() aggressively, so the id alone would hand a new loop a closed loop's client
_AsyncClientEntry: TypeAlias = tuple["weakref.ref[AbstractEventLoop]", "AsyncMongoClient"]
_SyncClientCache: TypeAlias = "OrderedDict[MongoClientKey, MongoClient]"
_AsyncClientCache: TypeAlias = "OrderedDict[_AsyncClientCacheKey, _AsyncClientEntry]"
_sync_clients: Final[_SyncClientCache] = OrderedDict() # mutable-ok: process-level client cache
_async_clients: Final[_AsyncClientCache] = OrderedDict() # mutable-ok: same cache, per loop
# async searches reach the sync client through executor threads, so both caches are shared state
_cache_lock: Final = threading.Lock()
def _store_bounded(cache: "OrderedDict[_K, _V]", cache_key: "_K", value: "_V") -> None:
"""Eviction only drops this cache's reference; an in-flight search keeps its client alive."""
with _cache_lock:
cache[cache_key] = value # mutable-ok: an LRU cache is mutable state by definition
cache.move_to_end(cache_key)
while len(cache) > _MAX_CACHED_CLIENTS:
cache.popitem(last=False)
def _mark_used(cache: "OrderedDict[_K, _V]", cache_key: "_K") -> None:
with _cache_lock:
if cache_key in cache:
cache.move_to_end(cache_key)
def import_sync_mongo_client() -> "type[MongoClient]":
try:
from pymongo import MongoClient as SyncMongoClient
except ImportError as e:
raise config_error(PYMONGO_INSTALL_HINT) from e
return SyncMongoClient
def import_async_mongo_client() -> "type[AsyncMongoClient]":
try:
from pymongo import AsyncMongoClient as AsyncMongoClientClass
except ImportError as e:
raise config_error(PYMONGO_INSTALL_HINT) from e
return AsyncMongoClientClass
def _client_kwargs(key: MongoClientKey) -> Mapping[str, object]:
return MappingProxyType(
{
"connectTimeoutMS": key.connect_timeout_ms,
"socketTimeoutMS": key.socket_timeout_ms,
"serverSelectionTimeoutMS": key.server_selection_timeout_ms,
"appname": _APP_NAME,
}
)
def get_sync_client(key: MongoClientKey, client_class: SyncClientFactory | None = None) -> "MongoClient":
cached: Final = _sync_clients.get(key)
if cached is not None:
_mark_used(_sync_clients, key)
return cached
build: Final = client_class if client_class is not None else import_sync_mongo_client()
client: Final = build(key.connection_string, **_client_kwargs(key))
_store_bounded(_sync_clients, key, client)
return client
def _purge_dead_loops() -> None:
"""A cached client holds its loop alive, so a closed loop's entry would pin that client and its
sockets for the life of the process."""
with _cache_lock:
for stale in tuple(
cache_key
for cache_key, (loop_ref, _) in _async_clients.items()
if (cached_loop := loop_ref()) is None or cached_loop.is_closed()
):
del _async_clients[stale]
def get_async_client(key: MongoClientKey, client_class: AsyncClientFactory | None = None) -> "AsyncMongoClient":
"""Async clients bind to the loop that created them, so the cache is keyed per loop."""
loop: Final = asyncio.get_running_loop()
loop_key: Final = (key, id(loop))
cached: Final = _async_clients.get(loop_key)
if cached is not None and cached[0]() is loop:
_mark_used(_async_clients, loop_key)
return cached[1]
_purge_dead_loops()
build: Final = client_class if client_class is not None else import_async_mongo_client()
client: Final = build(key.connection_string, **_client_kwargs(key))
_store_bounded(_async_clients, loop_key, (weakref.ref(loop), client))
return client
def reset_client_cache() -> None:
with _cache_lock:
_sync_clients.clear()
_async_clients.clear()
_AUTHENTICATION_FAILED_CODE: Final = 18
_UNAUTHORIZED_CODE: Final = 13
# Atlas reports a rejected user as code 8000 "AtlasError" where a self-managed mongod reports 18
_AUTHENTICATION_MESSAGE_MARKERS: Final = ("bad auth", "authentication failed", "not authorized")
_RESOLUTION_TIMEOUT_MARKERS: Final = ("resolution lifetime expired", "dns operation timed out")
_UNKNOWN_HOSTNAME_MARKERS: Final = ("dns query name does not exist", "name or service not known")
_CREDENTIAL_ESCAPING_MARKERS: Final = ("must be escaped according to rfc 3986", "bad database name")
def _index_hint(index_name: str, database: str, collection: str) -> str:
return (
f"No queryable MongoDB Vector Search index named '{index_name}' was found on "
f"'{database}.{collection}'. Confirm the index exists on that exact collection, that its "
"status is READY rather than still building, and that the vector store id matches the index name."
)
def missing_index_error(index_name: str, database: str, collection: str) -> BadRequestError:
"""$vectorSearch against a missing index, database or collection returns zero documents rather
than failing, so an empty result set is checked against the catalogue and reported as this."""
return config_error(
f"{_index_hint(index_name, database, collection)} A vector search against a database, "
"collection or index that does not exist returns no results rather than an error, so this "
"was reported as an empty result set by MongoDB."
)
def index_not_ready_error(index_name: str, database: str, collection: str, status: str) -> BadRequestError:
return config_error(
f"The MongoDB Vector Search index '{index_name}' on '{database}.{collection}' is not queryable "
f"yet; its status is {status}. Searches against it return no results until the build finishes."
)
def translate_mongo_error(error: Exception, index_name: str, database: str, collection: str) -> Exception:
"""Returns the exception to raise, so callers keep the driver error as ``__cause__``."""
try:
from pymongo.errors import (
ConfigurationError,
ConnectionFailure,
ExecutionTimeout,
InvalidOperation,
NetworkTimeout,
OperationFailure,
ServerSelectionTimeoutError,
)
except ImportError:
return error
if isinstance(error, ServerSelectionTimeoutError):
return timeout_error(
"Could not reach the MongoDB deployment before the timeout. On Atlas this is usually the "
"project's IP access list not containing this host, or a paused cluster. On a self-managed "
"deployment it is usually the host or port in the URI, or a firewall between this process "
f"and mongod. Either way it can also be an unresolvable hostname. Driver detail: {error}"
)
# ExecutionTimeout subclasses OperationFailure, so it has to be matched before it
if isinstance(error, (NetworkTimeout, ExecutionTimeout)):
return timeout_error(
f"The MongoDB vector search against '{database}.{collection}' timed out before returning. "
f"Driver detail: {error}"
)
# ServerSelectionTimeoutError and NetworkTimeout also subclass ConnectionFailure, so this only
# sees what those branches left
if isinstance(error, ConnectionFailure):
return unavailable_error(
f"The connection to '{database}.{collection}' was dropped or refused. That is usually a "
"replica set failover or a restarted node, so the search is worth retrying. If it keeps "
"happening: on Atlas the usual cause is a connection string with no username and password, "
"or a TLS failure, so confirm the URI is the one Atlas shows under Connect, Drivers; on a "
"self-managed deployment, check that mongod is listening on the host and port in the URI. "
f"Driver detail: {error}"
)
if isinstance(error, OperationFailure):
code: Final = error.code
detail: Final = str(error).lower()
if code in (_AUTHENTICATION_FAILED_CODE, _UNAUTHORIZED_CODE) or any(
marker in detail for marker in _AUTHENTICATION_MESSAGE_MARKERS
):
return config_error(
"MongoDB rejected the credentials in mongodb_connection_string, or the database user "
f"lacks read access to '{database}.{collection}'. Driver detail: {error.details}"
)
if "dimension" in detail:
return config_error(
"The query embedding does not match the vector dimensions the index was built for. "
"litellm_embedding_model must be the same model that produced the stored vectors. "
f"Driver detail: {error}"
)
if "is not indexed as vector" in detail:
return config_error(
"mongodb_embedding_field names a field the MongoDB Vector Search index does not cover. "
f"It must match the 'path' the index '{index_name}' was created on. Driver detail: {error}"
)
if "index" in detail and ("not found" in detail or "does not exist" in detail or "unknown" in detail):
return config_error(f"{_index_hint(index_name, database, collection)} Driver detail: {error}")
return config_error(
f"MongoDB rejected the vector search against '{database}.{collection}' using index "
f"'{index_name}'. Driver detail: {error}"
)
if isinstance(error, ConfigurationError):
configuration_detail: Final = str(error).lower()
if any(marker in configuration_detail for marker in _RESOLUTION_TIMEOUT_MARKERS):
return timeout_error(
"The DNS lookup for the cluster in mongodb_connection_string did not finish in time. "
"A mongodb+srv:// URI needs an SRV lookup before any connection is attempted, so this "
f"is DNS or the configured timeout, not MongoDB. Driver detail: {error}"
)
if any(marker in configuration_detail for marker in _UNKNOWN_HOSTNAME_MARKERS):
return config_error(
"The hostname in mongodb_connection_string does not exist in DNS. On Atlas, check the "
"cluster name against the URI shown under Connect, Drivers. On a self-managed deployment, "
f"check that the hostname resolves from this process. Driver detail: {error}"
)
if any(marker in configuration_detail for marker in _CREDENTIAL_ESCAPING_MARKERS):
return config_error(
"mongodb_connection_string could not be parsed. A username or password containing "
"'@', '/', ':' or '%' has to be percent-encoded per RFC 3986, so 'p@ss/word' becomes "
"'p%40ss%2Fword'. If the credentials are already encoded, check the database name in "
f"the URI path instead. Driver detail: {error}"
)
return config_error(
f"mongodb_connection_string is not a usable MongoDB connection string. Driver detail: {error}"
)
if isinstance(error, InvalidOperation):
return config_error(f"The MongoDB client was already closed or is unusable. Driver detail: {error}")
# An unreadable tlsCAFile or tlsCertificateKeyFile raises OSError, not a PyMongoError
if isinstance(error, OSError) and error.filename:
return config_error(
f"'{error.filename}', named by a TLS option in mongodb_connection_string, could not be read. "
"Check that tlsCAFile and tlsCertificateKeyFile point at files this process can open; inside "
f"a container that is the path in the container, not on the host. Driver detail: {error}"
)
# pymongo raises a plain ValueError, not a PyMongoError, for an unusable port
if isinstance(error, ValueError):
return config_error(
"The host and port in mongodb_connection_string could not be parsed. If the port is a "
"number between 0 and 65535, the cause is usually an unescaped ':' in the password, which "
f"has to be percent-encoded per RFC 3986 as '%3A'. Driver detail: {error}"
)
return error

469
litellm/llms/mongodb/vector_stores/transformation.py
View file

@ -1,37 +1,29 @@
"""MongoDB Vector Search has no HTTP query API, so this is a direct provider that runs the
``$vectorSearch`` aggregation through pymongo. ``vector_store_id`` is the search index name."""
from collections.abc import Callable, Mapping, Sequence
from collections.abc import Mapping, Sequence
from ipaddress import ip_address
from math import isfinite
from types import MappingProxyType
from typing import TYPE_CHECKING, Final, NoReturn
from typing import TYPE_CHECKING, Final, Literal, NoReturn
from urllib.parse import quote, urlsplit
import httpx
from pydantic import BaseModel, ConfigDict
from pydantic import BaseModel, ConfigDict, TypeAdapter, ValidationError
from litellm.exceptions import AuthenticationError, BadRequestError, ServiceUnavailableError, Timeout
from litellm.llms.base_llm.chat.transformation import BaseLLMException
from litellm.llms.base_llm.vector_store.transformation import (
BaseDirectVectorStoreConfig,
BaseQueryEmbeddingVectorStoreConfig,
LiteLLMVectorStoreEmbeddingExecutor,
VectorStoreEmbeddingExecutor,
)
from litellm.llms.mongodb.common_utils import (
DEFAULT_CONNECT_TIMEOUT_MS,
DEFAULT_SERVER_SELECTION_TIMEOUT_MS,
DEFAULT_SOCKET_TIMEOUT_MS,
MongoClientKey,
config_error,
get_async_client,
get_sync_client,
index_not_ready_error,
missing_index_error,
translate_mongo_error,
)
from litellm.secret_managers.main import get_secret_str
from litellm.types.router import GenericLiteLLMParams
from litellm.types.utils import EmbeddingResponse
from litellm.types.vector_stores import (
BaseVectorStoreAuthCredentials,
VectorStoreCreateOptionalRequestParams,
VectorStoreResultContent,
VectorStoreIndexEndpoints,
VectorStoreSearchOptionalRequestParams,
VectorStoreSearchResponse,
VectorStoreSearchResult,
)
if TYPE_CHECKING:
@ -39,26 +31,45 @@ if TYPE_CHECKING:
DEFAULT_EMBEDDING_FIELD_NAME: Final = "embedding"
DEFAULT_TEXT_FIELD_NAME: Final = "text"
SCORE_FIELD_NAME: Final = "score"
DEFAULT_MAX_NUM_RESULTS: Final = 10
MIN_MAX_NUM_RESULTS: Final = 1
MAX_MAX_NUM_RESULTS: Final = 50
NUM_CANDIDATES_MULTIPLIER: Final = 10
MIN_NUM_CANDIDATES: Final = 100
MAX_NUM_CANDIDATES: Final = 10_000
MAX_QUERY_CHARACTERS: Final = 32_000
_EMPTY_EMBEDDING_CONFIG: Final = MappingProxyType({})
_SEARCH_ONLY_MESSAGE: Final = (
"MongoDB vector store is search-only. Create the collection and its MongoDB Vector Search "
"index in MongoDB directly, then register it here by index name."
)
def config_error(message: str) -> BadRequestError:
return BadRequestError(message=message, model=None, llm_provider="mongodb")
class _Content(BaseModel):
model_config = ConfigDict(frozen=True, strict=True)
type: Literal["text"]
text: str
class _Result(BaseModel):
model_config = ConfigDict(frozen=True, strict=True, allow_inf_nan=False)
score: float | None
content: Sequence[_Content]
file_id: str | None
filename: str | None
class _SearchResponse(BaseModel):
model_config = ConfigDict(frozen=True, strict=True)
object: Literal["vector_store.search_results.page"]
search_query: str
data: Sequence[_Result]
class _MongoDBSearchParams(BaseModel):
"""Typed view over the vector store's litellm_params; unrelated keys are ignored."""
@ -66,7 +77,6 @@ class _MongoDBSearchParams(BaseModel):
litellm_embedding_model: str | None = None
litellm_embedding_config: Mapping[str, object] | None = None
mongodb_connection_string: str | None = None
mongodb_database: str | None = None
mongodb_collection: str | None = None
mongodb_text_field: str | None = None
@ -91,21 +101,6 @@ class _MongoDBSearchParams(BaseModel):
)
return self.litellm_embedding_model
def require_connection_string(self) -> str:
if not self.mongodb_connection_string:
raise config_error(
"mongodb_connection_string is required in litellm_params for the MongoDB vector store. "
"Example: mongodb+srv://<user>:<password>@<cluster>.mongodb.net for Atlas, or "
"mongodb://<user>:<password>@<host>:27017 for a self-managed deployment"
)
scheme: Final = self.mongodb_connection_string.split("://", 1)[0].lower()
if scheme not in ("mongodb", "mongodb+srv"):
raise config_error(
"mongodb_connection_string must start with 'mongodb://' or 'mongodb+srv://', "
f"got '{self.mongodb_connection_string.split('://', 1)[0]}://'"
)
return self.mongodb_connection_string
def require_database(self) -> str:
if not self.mongodb_database:
raise config_error(
@ -127,30 +122,28 @@ _MONGODB_PARAM_PREFIX: Final = "mongodb_"
_KNOWN_MONGODB_PARAMS: Final = frozenset(
name for name in _MongoDBSearchParams.model_fields if name.startswith(_MONGODB_PARAM_PREFIX)
)
_RESPONSE_ADAPTER: Final = TypeAdapter(VectorStoreSearchResponse)
class MongoDBVectorStoreConfig(BaseDirectVectorStoreConfig):
def __init__(
self,
embedding_executor: VectorStoreEmbeddingExecutor | None = None,
sync_client_factory: Callable[[MongoClientKey], object] | None = None,
async_client_factory: Callable[[MongoClientKey], object] | None = None,
) -> None:
super().__init__()
self.embedding_executor: Final[VectorStoreEmbeddingExecutor] = (
embedding_executor if embedding_executor is not None else LiteLLMVectorStoreEmbeddingExecutor()
)
self.sync_client_factory: Final[Callable[[MongoClientKey], object]] = (
sync_client_factory if sync_client_factory is not None else get_sync_client
)
self.async_client_factory: Final[Callable[[MongoClientKey], object]] = (
async_client_factory if async_client_factory is not None else get_async_client
)
class MongoDBVectorStoreConfig(BaseQueryEmbeddingVectorStoreConfig):
def __init__(self, embedding_executor: VectorStoreEmbeddingExecutor | None = None) -> None:
self.embedding_executor: Final = embedding_executor or LiteLLMVectorStoreEmbeddingExecutor()
def get_auth_credentials(self, litellm_params: Mapping[str, object]) -> BaseVectorStoreAuthCredentials:
return BaseVectorStoreAuthCredentials()
def get_vector_store_endpoints_by_type(self) -> VectorStoreIndexEndpoints:
return VectorStoreIndexEndpoints(read=[], write=[]) # mutable-ok: the TypedDict declares list fields
@staticmethod
def _reject_unknown_params(litellm_params: Mapping[str, object]) -> None:
"""Without this a mistyped mongodb_collection reads as 'mongodb_collection is required',
naming a key the reader can see they have set."""
if litellm_params.get("mongodb_connection_string") is not None:
raise config_error(
"MongoDB vector stores now use the BETA sidecar. Move mongodb_connection_string to "
"MONGODB_CONNECTION_STRING in the sidecar, remove it from LiteLLM, and configure api_base and api_key."
)
unknown: Final = sorted(
key for key in litellm_params if key.startswith(_MONGODB_PARAM_PREFIX) and key not in _KNOWN_MONGODB_PARAMS
)
@ -191,239 +184,203 @@ class MongoDBVectorStoreConfig(BaseDirectVectorStoreConfig):
return configured
return min(max(limit * NUM_CANDIDATES_MULTIPLIER, MIN_NUM_CANDIDATES), MAX_NUM_CANDIDATES)
@staticmethod
def _timeout_ms(timeout: float | httpx.Timeout | None) -> tuple[int, int]:
"""The connect and socket budgets pymongo is built with, in that order."""
if isinstance(timeout, httpx.Timeout):
return (
int((timeout.connect or DEFAULT_CONNECT_TIMEOUT_MS / 1000) * 1000),
int((timeout.read or DEFAULT_SOCKET_TIMEOUT_MS / 1000) * 1000),
def validate_environment(
self, headers: Mapping[str, object], litellm_params: GenericLiteLLMParams | None
) -> dict[str, object]: # mutable-ok: the shared HTTP handler requires writable headers
if litellm_params is None:
raise config_error("Configure api_base and api_key for the MongoDB BETA sidecar.")
self._reject_unknown_params(MappingProxyType(dict(litellm_params)))
api_key: Final = litellm_params.api_key or get_secret_str("MONGODB_SIDECAR_API_KEY")
if not api_key:
raise config_error("MongoDB sidecar api_key is required. Set api_key or MONGODB_SIDECAR_API_KEY.")
return {
**headers,
"Authorization": f"Bearer {api_key}",
"Content-Type": "application/json",
} # mutable-ok: writable HTTP headers
def get_complete_url(self, api_base: str | None, litellm_params: Mapping[str, object]) -> str:
if not api_base:
raise config_error("MongoDB sidecar api_base is required, for example http://127.0.0.1:8080.")
try:
parsed: Final = urlsplit(api_base)
valid: Final = parsed.scheme in ("http", "https") and bool(parsed.hostname) and parsed.port != 0
except ValueError:
raise config_error("MongoDB sidecar api_base must be a valid HTTP or HTTPS URL.") from None
if not valid or parsed.username or parsed.password or parsed.query or parsed.fragment:
raise config_error(
"MongoDB sidecar api_base must be an HTTP or HTTPS URL without credentials, query, or fragment."
)
if timeout is None:
return DEFAULT_CONNECT_TIMEOUT_MS, DEFAULT_SOCKET_TIMEOUT_MS
return min(int(float(timeout) * 1000), DEFAULT_CONNECT_TIMEOUT_MS), int(float(timeout) * 1000)
if parsed.scheme == "http":
try:
loopback: Final = ip_address(parsed.hostname or "").is_loopback
except ValueError:
raise config_error(
"MongoDB sidecar requires HTTPS. HTTP is supported only for a loopback IP such as 127.0.0.1."
) from None
if not loopback:
raise config_error(
"MongoDB sidecar requires HTTPS. HTTP is supported only for a loopback IP such as 127.0.0.1."
)
return api_base.rstrip("/")
@staticmethod
def _timeout_ms(value: object) -> int:
seconds: Final = value.read if isinstance(value, httpx.Timeout) else value
if seconds is None:
return 30_000
if not isinstance(seconds, (int, float)) or not isfinite(seconds) or seconds <= 0:
raise config_error("MongoDB search timeout must be a positive finite number.")
try:
return max(1, int(seconds * 1000))
except (ValueError, OverflowError):
raise config_error("MongoDB search timeout must be a positive finite number.") from None
@classmethod
def _client_key(cls, params: _MongoDBSearchParams, timeout: float | httpx.Timeout | None) -> MongoClientKey:
connect_ms, socket_ms = cls._timeout_ms(timeout)
return MongoClientKey(
connection_string=params.require_connection_string(),
connect_timeout_ms=connect_ms,
socket_timeout_ms=socket_ms,
server_selection_timeout_ms=min(connect_ms, DEFAULT_SERVER_SELECTION_TIMEOUT_MS),
)
def _params(
cls,
litellm_params: Mapping[str, object],
optional_params: VectorStoreSearchOptionalRequestParams,
extra_body: Mapping[str, object] | None,
) -> _MongoDBSearchParams:
cls._reject_unknown_params(litellm_params)
if extra_body:
raise config_error("MongoDB vector store does not support extra_body overrides.")
for unsupported in ("filters", "ranking_options", "rewrite_query"):
if optional_params.get(unsupported) is not None:
raise config_error(f"MongoDB vector store does not support the {unsupported} parameter.")
try:
params: Final = _MongoDBSearchParams.model_validate(litellm_params)
except ValidationError:
raise config_error(
"Invalid MongoDB vector-store configuration. Check the database, collection, fields, and candidate count."
) from None
params.require_database()
params.require_collection()
params.require_embedding_model()
cls._num_candidates(cls._limit(optional_params), params.mongodb_num_candidates)
cls._timeout_ms(litellm_params.get("timeout"))
return params
@classmethod
def _pipeline(
def _request(
cls,
vector_store_id: str,
query_vector: Sequence[float],
query_text: str,
params: _MongoDBSearchParams,
vector_store_search_optional_params: VectorStoreSearchOptionalRequestParams,
) -> Sequence[Mapping[str, object]]:
if vector_store_search_optional_params.get("filters") is not None:
optional_params: VectorStoreSearchOptionalRequestParams,
api_base: str,
embedding_response: EmbeddingResponse,
timeout: object,
) -> tuple[str, dict[str, object]]: # mutable-ok: the provider contract returns a writable JSON request body
if not embedding_response.data:
raise config_error(
"MongoDB vector store does not support the filters parameter yet. "
"Restrict the collection or the MongoDB Vector Search index definition instead."
"The embedding model returned no embedding for the search query. Check litellm_embedding_model."
)
if vector_store_search_optional_params.get("ranking_options") is not None:
raise config_error(
"MongoDB vector store does not support the ranking_options parameter yet. "
"Every result already carries the vectorSearchScore, so filter or re-rank "
"on that rather than having the threshold silently ignored."
)
if vector_store_search_optional_params.get("rewrite_query") is not None:
raise config_error(
"MongoDB vector store does not support the rewrite_query parameter. The query is "
"embedded exactly as sent; rewrite it before calling if you need that."
)
limit: Final = cls._limit(vector_store_search_optional_params)
search: Final = MappingProxyType(
{
"index": vector_store_id,
"path": params.embedding_field,
"queryVector": tuple(query_vector),
"numCandidates": cls._num_candidates(limit, params.mongodb_num_candidates),
"limit": limit,
}
)
projection: Final = MappingProxyType(
{params.text_field: 1, SCORE_FIELD_NAME: MappingProxyType({"$meta": "vectorSearchScore"})}
)
return [ # mutable-ok: pymongo rejects any non-list pipeline in common.validate_list
MappingProxyType({"$vectorSearch": search}),
MappingProxyType({"$project": projection}),
]
@classmethod
def _field_value(cls, document: Mapping[str, object], dotted_path: str) -> str | None:
"""None means absent, which is what separates a mistyped field from genuinely empty text."""
head, _, rest = dotted_path.partition(".")
if head not in document:
return None
value: Final = document[head]
if not rest:
return None if value is None else str(value)
return cls._field_value(value, rest) if isinstance(value, Mapping) else None
@classmethod
def _to_result(cls, document: Mapping[str, object], text_field: str) -> VectorStoreSearchResult:
document_id: Final = document.get("_id")
identifier: Final = None if document_id is None else str(document_id)
content: Final = [ # mutable-ok: VectorStoreSearchResult declares a list of content parts
VectorStoreResultContent(text=cls._field_value(document, text_field) or "", type="text")
]
raw_score: Final = document.get(SCORE_FIELD_NAME)
return VectorStoreSearchResult(
score=float(raw_score) if isinstance(raw_score, (int, float)) else None,
content=content,
file_id=identifier,
filename=identifier,
vector: Final = embedding_response.data[0]["embedding"]
if not vector or any(not isinstance(value, (float, int)) or not isfinite(value) for value in vector):
raise config_error("The embedding model must return a non-empty, finite query vector.")
limit: Final = cls._limit(optional_params)
return (
f"{api_base}/v1/vector_stores/{quote(vector_store_id, safe='')}/search",
{ # mutable-ok: JSON transport requires a dict
"query": query_text,
"query_vector": tuple(vector),
"mongodb_database": params.require_database(),
"mongodb_collection": params.require_collection(),
"mongodb_embedding_field": params.embedding_field,
"mongodb_text_field": params.text_field,
"mongodb_num_candidates": cls._num_candidates(limit, params.mongodb_num_candidates),
"max_num_results": limit,
"timeout_ms": cls._timeout_ms(timeout),
},
)
@classmethod
def _raise_for_missing_text_field(
cls, documents: Sequence[Mapping[str, object]], text_field: str, database: str, collection: str
) -> None:
"""$vectorSearch matches documents carrying no text, so a mistyped mongodb_text_field
returns well-scored results with empty content instead of failing."""
if documents and all(cls._field_value(document, text_field) is None for document in documents):
raise config_error(
f"None of the {len(documents)} matched documents in '{database}.{collection}' has a "
f"'{text_field}' field, so every result would carry empty text. Set mongodb_text_field "
"to the field holding the readable text; it accepts a dotted path such as metadata.body."
)
@classmethod
def _to_response(
cls, documents: Sequence[Mapping[str, object]], query_text: str, text_field: str
) -> VectorStoreSearchResponse:
return VectorStoreSearchResponse(
object="vector_store.search_results.page",
search_query=query_text,
data=[ # mutable-ok: VectorStoreSearchResponse declares data as a list
cls._to_result(document, text_field) for document in documents
],
)
@staticmethod
def _raise_for_unusable_index(
catalogue: Sequence[Mapping[str, object]], index_name: str, database: str, collection: str
) -> None:
"""mongod returns zero documents both for a query that matched nothing and for a missing
database, collection or index, so the catalogue decides which one happened."""
if not catalogue:
raise missing_index_error(index_name, database, collection)
entry: Final = catalogue[0]
if not entry.get("queryable"):
raise index_not_ready_error(index_name, database, collection, str(entry.get("status") or "unknown"))
@staticmethod
def _embedding_vector(embedding_response: EmbeddingResponse) -> Sequence[float]:
data: Final = embedding_response.data
if not data:
raise config_error(
"The embedding model returned no embedding for the search query, so there is nothing "
"to search MongoDB with. Check the embedding deployment named by litellm_embedding_model."
)
return data[0]["embedding"]
def execute_search_vector_store_request(
def transform_search_vector_store_request(
self,
vector_store_id: str,
query: str | Sequence[str],
vector_store_search_optional_params: VectorStoreSearchOptionalRequestParams,
api_base: str,
litellm_logging_obj: "LiteLLMLoggingObj",
litellm_params: Mapping[str, object],
extra_body: Mapping[str, object] | None = None,
embedding_executor: VectorStoreEmbeddingExecutor | None = None,
timeout: float | httpx.Timeout | None = None,
) -> VectorStoreSearchResponse:
self._reject_unknown_params(litellm_params)
params: Final = _MongoDBSearchParams.model_validate(litellm_params)
) -> tuple[str, dict[str, object]]: # mutable-ok: the provider contract returns a writable JSON request body
params: Final = self._params(litellm_params, vector_store_search_optional_params, extra_body)
query_text: Final = self._query_text(query)
key: Final = self._client_key(params, timeout)
database: Final = params.require_database()
collection: Final = params.require_collection()
embedding_response: Final = (embedding_executor or self.embedding_executor).embed(
params.require_embedding_model(),
response: Final = (embedding_executor or self.embedding_executor).embed(
params.require_embedding_model(), query_text, params.litellm_embedding_config or _EMPTY_EMBEDDING_CONFIG
)
return self._request(
vector_store_id,
query_text,
params.litellm_embedding_config or _EMPTY_EMBEDDING_CONFIG,
)
pipeline: Final = self._pipeline(
vector_store_id, self._embedding_vector(embedding_response), params, vector_store_search_optional_params
params,
vector_store_search_optional_params,
api_base,
response,
litellm_params.get("timeout"),
)
try:
client: Final = self.sync_client_factory(key)
target: Final = client[database][collection] # pyright: ignore[reportIndexIssue] # factory is typed as returning object so injected doubles are accepted
documents: Final = tuple(target.aggregate(pipeline))
except Exception as e:
raise translate_mongo_error(e, index_name=vector_store_id, database=database, collection=collection) from e
if not documents:
try:
catalogue: Final = tuple(target.list_search_indexes(vector_store_id))
except Exception as e:
raise translate_mongo_error(
e, index_name=vector_store_id, database=database, collection=collection
) from e
self._raise_for_unusable_index(catalogue, vector_store_id, database, collection)
self._raise_for_missing_text_field(documents, params.text_field, database, collection)
return self._to_response(documents, query_text, params.text_field)
async def aexecute_search_vector_store_request(
async def atransform_search_vector_store_request(
self,
vector_store_id: str,
query: str | Sequence[str],
vector_store_search_optional_params: VectorStoreSearchOptionalRequestParams,
api_base: str,
litellm_logging_obj: "LiteLLMLoggingObj",
litellm_params: Mapping[str, object],
extra_body: Mapping[str, object] | None = None,
embedding_executor: VectorStoreEmbeddingExecutor | None = None,
timeout: float | httpx.Timeout | None = None,
) -> VectorStoreSearchResponse:
self._reject_unknown_params(litellm_params)
params: Final = _MongoDBSearchParams.model_validate(litellm_params)
) -> tuple[str, dict[str, object]]: # mutable-ok: the provider contract returns a writable JSON request body
params: Final = self._params(litellm_params, vector_store_search_optional_params, extra_body)
query_text: Final = self._query_text(query)
key: Final = self._client_key(params, timeout)
database: Final = params.require_database()
collection: Final = params.require_collection()
embedding_response: Final = await (embedding_executor or self.embedding_executor).aembed(
params.require_embedding_model(),
response: Final = await (embedding_executor or self.embedding_executor).aembed(
params.require_embedding_model(), query_text, params.litellm_embedding_config or _EMPTY_EMBEDDING_CONFIG
)
return self._request(
vector_store_id,
query_text,
params.litellm_embedding_config or _EMPTY_EMBEDDING_CONFIG,
)
pipeline: Final = self._pipeline(
vector_store_id, self._embedding_vector(embedding_response), params, vector_store_search_optional_params
params,
vector_store_search_optional_params,
api_base,
response,
litellm_params.get("timeout"),
)
def transform_search_vector_store_response(
self, response: httpx.Response, litellm_logging_obj: "LiteLLMLoggingObj"
) -> VectorStoreSearchResponse:
try:
client: Final = self.async_client_factory(key)
target: Final = client[database][collection] # pyright: ignore[reportIndexIssue] # factory is typed as returning object so injected doubles are accepted
cursor: Final = await target.aggregate(pipeline)
documents: Final = [ # mutable-ok: an async comprehension cannot build a tuple directly
document async for document in cursor
]
except Exception as e:
raise translate_mongo_error(e, index_name=vector_store_id, database=database, collection=collection) from e
if not documents:
try:
index_cursor: Final = await target.list_search_indexes(vector_store_id)
catalogue: Final = [ # mutable-ok: an async comprehension cannot build a tuple directly
entry async for entry in index_cursor
]
except Exception as e:
raise translate_mongo_error(
e, index_name=vector_store_id, database=database, collection=collection
) from e
self._raise_for_unusable_index(catalogue, vector_store_id, database, collection)
self._raise_for_missing_text_field(documents, params.text_field, database, collection)
return self._to_response(documents, query_text, params.text_field)
validated: Final = _SearchResponse.model_validate_json(response.content)
return _RESPONSE_ADAPTER.validate_python(validated.model_dump())
except ValidationError:
raise ServiceUnavailableError(
message="MongoDB sidecar returned an invalid search response. Check the sidecar version and deployment.",
model=None,
llm_provider="mongodb",
) from None
def get_error_class(
self, error_message: str, status_code: int, headers: Mapping[str, object] | httpx.Headers
) -> BaseLLMException:
if status_code == 400:
raise config_error(error_message)
if status_code == 401:
raise AuthenticationError(message="MongoDB sidecar rejected api_key.", model=None, llm_provider="mongodb")
if status_code == 408:
raise Timeout(message=error_message, model=None, llm_provider="mongodb")
raise ServiceUnavailableError(
message="MongoDB sidecar is unavailable. Check its address, health, and logs.",
model=None,
llm_provider="mongodb",
)
def validate_create_vector_store(self) -> NoReturn:
raise config_error(_SEARCH_ONLY_MESSAGE)
def transform_create_vector_store_request(
self,
vector_store_create_optional_params: VectorStoreCreateOptionalRequestParams,
api_base: str,
self, vector_store_create_optional_params: VectorStoreCreateOptionalRequestParams, api_base: str
) -> NoReturn:
raise config_error(_SEARCH_ONLY_MESSAGE)

24
litellm/llms/ollama/completion/transformation.py
View file

@ -16,6 +16,10 @@ from litellm.litellm_core_utils.prompt_templates.factory import (
custom_prompt,
ollama_pt,
)
from litellm.litellm_core_utils.prompt_templates.image_handling import (
async_inline_remote_media,
inline_remote_image_urls,
)
from litellm.llms.base_llm.base_model_iterator import BaseModelResponseIterator
from litellm.llms.base_llm.chat.transformation import BaseConfig, BaseLLMException
from litellm.types.llms.openai import AllMessageValues, ChatCompletionUsageBlock
@ -344,6 +348,26 @@ class OllamaConfig(BaseConfig):
)
return model_response
@property
def uses_async_transform_request(self) -> bool:
return True
async def async_transform_request(
self,
model: str,
messages: list[AllMessageValues], # mutable-ok: BaseConfig signature
optional_params: dict[str, object], # mutable-ok: BaseConfig signature
litellm_params: dict[str, object], # mutable-ok: BaseConfig signature
headers: dict[str, object], # mutable-ok: BaseConfig signature
) -> dict[str, object]: # mutable-ok: BaseConfig signature
return self.transform_request(
model=model,
messages=await async_inline_remote_media(messages, should_inline=inline_remote_image_urls),
optional_params=optional_params,
litellm_params=litellm_params,
headers=headers,
)
def transform_request(
self,
model: str,

4
litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/transformation.py
View file

@ -5,6 +5,7 @@ from typing import TYPE_CHECKING, Final
import httpx
import litellm
from litellm.litellm_core_utils.prompt_templates.image_handling import RemoteMedia, inline_remote_image_urls
from litellm.llms.base_llm.chat.transformation import LiteLLMLoggingObj
from litellm.types.llms.openai import AllMessageValues
from litellm.types.utils import ModelResponse
@ -51,6 +52,9 @@ class VertexAIAnthropicConfig(AnthropicConfig):
def custom_llm_provider(self) -> str | None:
return "vertex_ai"
def inlines_remote_media(self, media: RemoteMedia) -> bool:
return inline_remote_image_urls(media)
def should_strip_billing_metadata(self) -> bool:
return True

8
litellm/llms/watsonx/chat/transformation.py
View file

@ -157,11 +157,9 @@ class IBMWatsonXChatConfig(IBMWatsonXMixin, OpenAIGPTConfig):
@staticmethod
async def aapply_prompt_template(model: str, messages: list[dict[str, str]]) -> str | None:
"""Apply prompt template (async version)"""
import litellm
from litellm.litellm_core_utils.prompt_templates.factory import (
ahf_chat_template,
custom_prompt,
hf_chat_template,
ibm_granite_pt,
mistral_instruct_pt,
)
@ -179,11 +177,7 @@ class IBMWatsonXChatConfig(IBMWatsonXMixin, OpenAIGPTConfig):
else:
hf_model = model
try:
# Use sync if cached, async if not
if hf_model in litellm.known_tokenizer_config:
result = hf_chat_template(model=hf_model, messages=messages)
else:
result = await ahf_chat_template(model=hf_model, messages=messages)
result = await ahf_chat_template(model=hf_model, messages=messages)
# Return result if it's truthy (not None and not empty string)
# The caller (_aconvert_watsonx_messages_core) will handle None/empty by falling back to default
if result:

12
litellm/llms/watsonx/completion/transformation.py
View file

@ -16,6 +16,7 @@ from ..common_utils import (
IBMWatsonXMixin,
WatsonXAIError,
_get_api_params,
aconvert_watsonx_messages_to_prompt,
convert_watsonx_messages_to_prompt,
)
@ -236,7 +237,11 @@ class IBMWatsonXAIConfig(IBMWatsonXMixin, BaseConfig):
**watsonx_auth_payload,
}
async def atransform_request(
@property
def uses_async_transform_request(self) -> bool:
return True
async def async_transform_request(
self,
model: str,
messages: list[AllMessageValues],
@ -244,11 +249,6 @@ class IBMWatsonXAIConfig(IBMWatsonXMixin, BaseConfig):
litellm_params: dict,
headers: dict,
) -> dict:
"""Async version of transform_request"""
from litellm.llms.watsonx.common_utils import (
aconvert_watsonx_messages_to_prompt,
)
provider: Final = model.split("/")[0]
prompt: Final = await aconvert_watsonx_messages_to_prompt(
model=model, messages=messages, provider=provider, custom_prompt_dict={}

1
litellm/models/managed_files.py
View file

@ -32,6 +32,7 @@ class LiteLLM_ManagedObjectTable(LiteLLMPydanticObjectBase):
file_object: LiteLLMBatch | LiteLLMFineTuningJob | ResponsesAPIResponse
created_by: str | None = None
team_id: str | None = None
org_id: str | None = None
class LiteLLM_ManagedVectorStoreTable(LiteLLMPydanticObjectBase):

66
litellm/proxy/_experimental/mcp_server/server.py
View file

@ -19,7 +19,7 @@ from typing import TYPE_CHECKING, Any, Final, Protocol
import httpx
from fastapi import FastAPI, HTTPException
from pydantic import AnyUrl, ConfigDict
from pydantic import AnyUrl, ConfigDict, TypeAdapter, ValidationError
from starlette.requests import Request as StarletteRequest
from starlette.responses import JSONResponse
from starlette.types import Message, Receive, Scope, Send
@ -108,9 +108,9 @@ _MAX_STATEFUL_SESSIONS_PER_OWNER: Final = 100
# prevents an authenticated client from forcing the proxy to buffer an
# arbitrarily large body just to make a routing decision.
_MCP_ROUTING_PEEK_MAX_BYTES: Final = 4096
# ASGI scope key holding the tracing span of the request carrying an MCP
# message, written on the request task and read back by the message handler.
# ASGI scope keys carrying OTel request state into a stateful MCP message handler.
_MCP_TRANSPORT_SPAN_SCOPE_KEY: Final = "litellm_otel_transport_span"
_MCP_DESTINATIONS_SCOPE_KEY: Final = "litellm_otel_request_destinations"
def _invalidate_byok_cred_cache(user_id: str, server_id: str) -> None:
@ -328,18 +328,17 @@ def _otel_publish_transport_span_on_scope(scope: Scope) -> None:
scope[_MCP_TRANSPORT_SPAN_SCOPE_KEY] = span
def _otel_transport_span_from_message(req_ctx: object) -> object:
"""The tracing span of the HTTP request that carried this MCP message.
Read off that request's ASGI scope, reached through the ``Request`` the
streamable-HTTP transport attaches to each message, so it is this message's
transport and not whichever request happens to have touched the session last.
Returns whatever the scope holds; the otel plumbing validates it."""
def _otel_value_from_message_scope(req_ctx: object, key: str) -> object:
request: Final = getattr(req_ctx, "request", None)
scope: Final = getattr(request, "scope", None)
if not isinstance(scope, Mapping):
return None
return scope.get(_MCP_TRANSPORT_SPAN_SCOPE_KEY)
return scope.get(key)
def _otel_transport_span_from_message(req_ctx: object) -> object:
"""The tracing span of the HTTP request that carried this MCP message."""
return _otel_value_from_message_scope(req_ctx, _MCP_TRANSPORT_SPAN_SCOPE_KEY)
def _otel_set_mcp_transport_span(span: object) -> object:
@ -372,6 +371,44 @@ def _otel_reset_mcp_transport_span(token: object) -> None:
return
def _otel_publish_request_destinations_on_scope(scope: Scope) -> None:
try:
from litellm.integrations.otel.plumbing.context import request_destinations
scope[_MCP_DESTINATIONS_SCOPE_KEY] = request_destinations()
except ImportError:
return
def _otel_set_mcp_request_destinations(req_ctx: object) -> object:
destinations: Final = _otel_value_from_message_scope(req_ctx, _MCP_DESTINATIONS_SCOPE_KEY)
if not isinstance(destinations, tuple):
return None
try:
from litellm.integrations.otel.model.destination import OtelDestination
from litellm.integrations.otel.plumbing.context import set_request_destinations
destination_adapter: Final[TypeAdapter[tuple[OtelDestination, ...]]] = TypeAdapter(
tuple[OtelDestination, ...],
config=ConfigDict(revalidate_instances="always"),
)
validated_destinations: Final = destination_adapter.validate_python(destinations, strict=True)
return set_request_destinations(validated_destinations)
except (ImportError, ValidationError):
return None
def _otel_reset_mcp_request_destinations(token: object) -> None:
if token is None:
return
try:
from litellm.integrations.otel.plumbing.context import reset_request_destinations
reset_request_destinations(token)
except ImportError:
return
def _proxy_exception_to_http_exception(exc: ProxyException) -> HTTPException:
"""Map a ``ProxyException`` to an ``HTTPException`` that preserves its real
status code and headers.
@ -763,10 +800,12 @@ if MCP_AVAILABLE:
_session_reset_token = active_mcp_session_var.set(req_ctx.session)
_trace_token = None
_transport_token = None
_destinations_token = None
try:
_trace_token = _otel_set_mcp_trace_carrier(_mcp_meta_trace_carrier(req_ctx))
_transport_token = _otel_set_mcp_transport_span(_otel_transport_span_from_message(req_ctx))
_destinations_token = _otel_set_mcp_request_destinations(req_ctx)
# Get user authentication from context variable
(
user_api_key_auth,
@ -828,6 +867,7 @@ if MCP_AVAILABLE:
# This prevents the HTTP stream from failing and allows the client to get a response
return []
finally:
_otel_reset_mcp_request_destinations(_destinations_token)
_otel_reset_mcp_transport_span(_transport_token)
_otel_reset_mcp_trace_carrier(_trace_token)
if _session_reset_token is not None:
@ -1021,10 +1061,12 @@ if MCP_AVAILABLE:
_session_reset_token = active_mcp_session_var.set(req_ctx.session)
_trace_token = None
_transport_token = None
_destinations_token = None
try:
_trace_token = _otel_set_mcp_trace_carrier(_mcp_meta_trace_carrier(req_ctx))
_transport_token = _otel_set_mcp_transport_span(_otel_transport_span_from_message(req_ctx))
_destinations_token = _otel_set_mcp_request_destinations(req_ctx)
# Validate arguments
(
user_api_key_auth,
@ -1163,6 +1205,7 @@ if MCP_AVAILABLE:
return response
finally:
_otel_reset_mcp_request_destinations(_destinations_token)
_otel_reset_mcp_transport_span(_transport_token)
_otel_reset_mcp_trace_carrier(_trace_token)
if _session_reset_token is not None:
@ -4493,6 +4536,7 @@ if MCP_AVAILABLE:
async def _dispatch() -> None:
_otel_publish_transport_span_on_scope(scope)
_otel_publish_request_destinations_on_scope(scope)
auth_user: Final = _set_or_update_auth_context(
user_api_key_auth=user_api_key_auth,
mcp_auth_header=mcp_auth_header,

10
litellm/proxy/_types.py
View file

@ -835,6 +835,14 @@ class LiteLLMRoutes(enum.Enum):
"/team/daily/activity/aggregated",
"/team/spend/by_user",
"/team/{team_id}/members/me",
# POST/GET the team's logging callbacks, and DELETE one of them. Every
# handler calls _verify_team_access, which admits only a proxy admin, an
# org admin for the team, or an admin of this team.
#
# team_id is a free-form string, so it spells these with the same path
# converter the router uses; the gate matches that converter.
"/team/{team_id:path}/callback",
"/team/{team_id:path}/callback/{callback_name}",
"/model/new",
"/model/update",
"/model/delete",
@ -3584,7 +3592,9 @@ class AllCallbacks(LiteLLMPydanticObjectBase):
ui_callback_name="OpenTelemetry",
litellm_callback_params=[
"OTEL_EXPORTER",
"OTEL_EXPORTER_OTLP_PROTOCOL",
"OTEL_ENDPOINT",
"OTEL_TRACES_ENDPOINT",
"OTEL_HEADERS",
],
)

62
litellm/proxy/agent_endpoints/a2a_endpoints.py
View file

@ -144,31 +144,32 @@ def _validate_push_notification_url(url: str) -> None:
raise HTTPException(status_code=400, detail=str(e)) from e
def _caller_identity_headers(user_api_key_dict: UserAPIKeyAuth) -> dict[str, str]:
headers: Final[dict[str, str]] = {}
if user_api_key_dict.user_id:
headers["X-LiteLLM-User-Id"] = user_api_key_dict.user_id
if user_api_key_dict.team_id:
headers["X-LiteLLM-Team-Id"] = user_api_key_dict.team_id
return headers
def _caller_identity_headers(user_api_key_dict: UserAPIKeyAuth) -> Mapping[str, str]:
return MappingProxyType(
{
name: value
for name, value in (
("X-LiteLLM-User-Id", user_api_key_dict.user_id),
("X-LiteLLM-Team-Id", user_api_key_dict.team_id),
)
if value
}
)
def _forwarding_headers(
user_api_key_dict: UserAPIKeyAuth,
caller_identity: Mapping[str, str],
request_data: Mapping[str, object],
agent_extra_headers: Mapping[str, str] | None,
) -> Mapping[str, str] | None:
sanitized: Final = (
{k: v for k, v in agent_extra_headers.items() if not k.lower().startswith("x-litellm-")}
if agent_extra_headers
else None
) -> dict[str, str] | None:
passthrough: Final = tuple(
(name, value)
for name, value in (agent_extra_headers.items() if agent_extra_headers else ())
if not name.lower().startswith("x-litellm-")
)
merged: Final = merge_agent_headers(dynamic_headers=sanitized, static_headers=None) or {}
identity: Final = _caller_identity_headers(user_api_key_dict)
trace_id: Final = request_data.get("litellm_trace_id")
if trace_id:
identity["X-LiteLLM-Trace-Id"] = str(trace_id)
merged.update(identity)
trace: Final = (("X-LiteLLM-Trace-Id", str(trace_id)),) if trace_id else ()
merged: Final = dict((*passthrough, *caller_identity.items(), *trace))
return merged or None
@ -755,6 +756,7 @@ async def invoke_agent_a2a(
ProxyBaseLLMRequestProcessing,
)
caller_identity: Final = _caller_identity_headers(user_api_key_dict)
processor: Final = ProxyBaseLLMRequestProcessing(data=body)
data, logging_obj = await processor.common_processing_pre_call_logic(
request=request,
@ -793,9 +795,13 @@ async def invoke_agent_a2a(
if header_name:
dynamic_headers[header_name] = val
agent_extra_headers = merge_agent_headers(
dynamic_headers=dynamic_headers or None,
static_headers=static_headers or None,
agent_extra_headers = _forwarding_headers(
caller_identity=caller_identity,
request_data=data,
agent_extra_headers=merge_agent_headers(
dynamic_headers=dynamic_headers or None,
static_headers=static_headers or None,
),
)
# Databricks App endpoints require a short-lived OAuth M2M token rather
@ -942,12 +948,7 @@ async def invoke_agent_a2a(
"method": method,
"params": params,
}
caller_headers: Final = _forwarding_headers(
user_api_key_dict=user_api_key_dict,
request_data=data,
agent_extra_headers=agent_extra_headers,
)
result = await _forward_jsonrpc(agent_url, forward_body, extra_headers=caller_headers)
result = await _forward_jsonrpc(agent_url, forward_body, extra_headers=agent_extra_headers)
if method == "agent/getAuthenticatedExtendedCard":
card: Final = result.get("result")
if isinstance(card, dict):
@ -988,16 +989,11 @@ async def invoke_agent_a2a(
"method": method,
"params": params,
}
sse_caller_headers: Final = _forwarding_headers(
user_api_key_dict=user_api_key_dict,
request_data=data,
agent_extra_headers=agent_extra_headers,
)
return await _forward_jsonrpc_sse(
agent_url,
forward_body,
request_id=request_id,
extra_headers=sse_caller_headers,
extra_headers=agent_extra_headers,
proxy_logging_obj=proxy_logging_obj,
user_api_key_dict=user_api_key_dict,
request_data=data,

20
litellm/proxy/auth/route_checks.py
View file

@ -497,10 +497,22 @@ class RouteChecks:
def _placeholder_to_regex(match: re.Match) -> str:
placeholder: Final = match.group(0).strip("{}")
if placeholder.endswith(":path"):
# allow "/" in the placeholder value, but don't eat the route suffix after ":"
return r"[^:]+"
return r"[^/]+"
if not placeholder.endswith(":path"):
return r"[^/]+"
# A ":path" placeholder takes whatever the router's own path
# converter takes, slashes and colons alike, so an id spelled with
# either (or both) still matches the template it was mounted under.
#
# Unless the template puts a ":" literal of its own after the
# placeholder: the Google routes end in ":generateContent" and
# friends, and there the value has to stop before that suffix
# rather than swallow it and match a different verb.
#
# "[\s\S]" rather than ".", because "." stops at a newline and the
# path converter does not: a %0A anywhere in the value would leave
# the route unmatched here while still reaching the handler, which
# turns this gate into a bypass for the lists built on it.
return r"[^:]+" if ":" in match.string[match.end() :] else r"[\s\S]+"
pattern = re.sub(r"\{[^}]+\}", _placeholder_to_regex, pattern)
# Anchor the pattern to match the entire string

40
litellm/proxy/auth/user_api_key_auth.py
View file

@ -2038,7 +2038,7 @@ async def _user_api_key_auth_builder(
fallback_spend=team_member_spend,
max_budget=team_member_budget,
)
if team_member_spend > team_member_budget:
if team_member_spend >= team_member_budget:
_entity_id: Final = f"{valid_token.user_id}:{valid_token.team_id}"
raise litellm.BudgetExceededError(
current_cost=team_member_spend,
@ -2837,6 +2837,43 @@ async def _authorize_authenticated_request(
@tracer.wrap()
def _seed_request_destinations(user_api_key_dict: UserAPIKeyAuth, request: Request | None = None) -> None:
"""Anchor the OTLP destinations this key or team overrides its traces to.
Called inside the ``auth`` phase span so that span reaches the tenant's account
as well, and on the request task so the ``ContextVar`` is inherited by the logging
tasks that close the LLM span. Best-effort: trace routing must never fail auth.
``request`` carries the headers, so a backend this request disabled with
``x-litellm-disable-callbacks`` resolves to no destination.
Only destinations the published fan-out can build are anchored. Anchoring one is
what tells the operator's exporter to hold that backend's spans back under
``override``, so an unbuildable one would leave the span with nowhere to go.
The ``postgres`` spans under ``auth`` close before this runs, because they are the
reads that resolve the identity being read here. They never reach the tenant's
account, and they are never withheld from the operator's backend, whichever mode
is set.
"""
try:
from litellm.integrations.otel.logger import fan_out_provider
from litellm.integrations.otel.plumbing.context import set_request_destinations
from litellm.integrations.otel.plumbing.providers import deliverable_destinations
from litellm.proxy.litellm_pre_call_utils import (
resolve_tenant_otel_destinations,
)
set_request_destinations(
deliverable_destinations(
resolve_tenant_otel_destinations(user_api_key_dict, _safe_get_request_headers(request)),
fan_out_provider(),
)
)
except Exception as exc: # noqa: BLE001 # telemetry routing is best-effort and must never break authentication
verbose_proxy_logger.debug("OTel V2: tenant destination resolution failed: %s", exc)
async def user_api_key_auth(
request: Request,
api_key: str = fastapi.Security(api_key_header),
@ -2883,6 +2920,7 @@ async def user_api_key_auth(
raise body_parse_exception
raise
user_api_key_auth_obj.budget_reservation = None
_seed_request_destinations(user_api_key_auth_obj, request)
# A body that never parsed is authenticated (so the trace carries identity
# and this ``auth`` span) but not authorized: there is no model to check it

85
litellm/proxy/common_utils/callback_config_validation.py
View file

@ -44,6 +44,91 @@ def _langfuse_environment_error(callback_vars: Mapping[str, str]) -> str | None:
return None
# Which credential family a dynamic variable belongs to. The families are the
# integrations that share one account: every langfuse_* variable configures the
# same Langfuse project whether it rides the classic callback or the OTel one,
# and every dd_* variable configures the same Datadog account.
_VAR_FAMILIES: Final[Mapping[str, str]] = MappingProxyType(
{
"arize_": "Arize",
"dd_": "Datadog",
"gcs_": "GCS",
"humanloop_": "Humanloop",
"langfuse_": "Langfuse",
"langsmith_": "LangSmith",
"newrelic_": "New Relic",
"posthog_": "PostHog",
"wandb_": "Weights & Biases",
"weave_": "Weights & Biases",
}
)
def _family_of(var: str) -> str | None:
"""The credential family ``var`` configures, or ``None`` if it configures none.
``turn_off_message_logging`` and friends belong to no backend, so they carry
no credentials anyone could redirect.
"""
return next((family for prefix, family in _VAR_FAMILIES.items() if var.startswith(prefix)), None)
def cross_entry_family_error(
callback_vars: Mapping[str, str] | None,
stored_vars_by_entry: Sequence[Mapping[str, str]],
) -> str | None:
"""Reject an entry that changes what a family another entry holds resolves to.
Every stored entry's variables are flattened into one dict before a request
reads them, and the flattened dict is what the exporter authenticates and
addresses with. So an entry naming only a destination is enough to redirect
credentials that were written somewhere else: a host on a second entry pairs
with the key from the first, and the request carries that key to the new
host.
Two rules together keep the flattened dict out of the caller's hands. A
variable the family already configures has to keep the value it has, so
nothing already in use can be moved. A variable the family does not yet
configure may only carry a value the family already holds, which is what lets
the same credential go in under its other spelling (``langfuse_secret`` and
``langfuse_secret_key`` are one key) without anything here having to list the
spellings. Between them, no value the caller chose can enter the family, and
repeating the family as it stands is still allowed -- that is how one
integration gets registered for both the success and the failure event.
A team admin who does want to move a family deletes the entry holding it
first, which reveals nothing.
Only the writers this endpoint newly admits are held to this, because a proxy
admin already holds every credential the proxy has.
``stored_vars_by_entry`` has to arrive decrypted; the credential values are
encrypted at rest and ciphertext never equals the plaintext coming in.
"""
if not callback_vars:
return None
stored_by_var: Final = {
var: value for entry in stored_vars_by_entry for var, value in entry.items() if _family_of(var) is not None
}
family_values: Final = frozenset(
(family, value)
for entry in stored_vars_by_entry
for var, value in entry.items()
if (family := _family_of(var)) is not None
)
held_families: Final = frozenset(family for family, _ in family_values)
return next(
(
f"{family} is already configured by another callback entry on this team. "
f"Remove that entry before setting {var} here."
for var, value, family in ((v, callback_vars[v], _family_of(v)) for v in callback_vars)
if family in held_families
and (stored_by_var[var] != value if var in stored_by_var else (family, value) not in family_values)
),
None,
)
def logging_metadata_config_error(metadata: Mapping[str, object] | None) -> str | None:
"""Validate every ``logging`` entry of a team/key metadata payload."""
if not metadata:

4
litellm/proxy/common_utils/callback_utils.py
View file

@ -10,8 +10,10 @@ import litellm
from litellm import get_secret
from litellm._logging import verbose_proxy_logger
from litellm.constants import (
CLIENT_OUTPUT_CEILING_METADATA_KEY,
CONSUMED_REQUEST_TAGS_METADATA_KEY,
PRE_CALL_EXECUTED_GUARDRAILS_KEY,
ROUTING_REQUEST_TAGS_METADATA_KEY,
SESSION_DEPLOYMENT_AFFINITY_TTL_METADATA_KEY,
)
from litellm.integrations.custom_logger import CustomLogger
@ -507,6 +509,8 @@ LITELLM_PROXY_INTERNAL_METADATA_KEYS: Final = frozenset(
PRE_CALL_EXECUTED_GUARDRAILS_KEY,
SESSION_DEPLOYMENT_AFFINITY_TTL_METADATA_KEY,
CONSUMED_REQUEST_TAGS_METADATA_KEY,
CLIENT_OUTPUT_CEILING_METADATA_KEY,
ROUTING_REQUEST_TAGS_METADATA_KEY,
"disable_global_guardrails",
"disable_global_guardrail",
"opted_out_global_guardrails",

2
litellm/proxy/db/db_spend_update_writer.py
View file

@ -1063,7 +1063,7 @@ class DBSpendUpdateWriter:
await enqueue_spend_logs(prisma_client, (payload,))
if payload.get("call_type") in RESPONSES_SESSION_CALL_TYPES:
request_spend_log_flush()
request_spend_log_flush(prisma_client)
else:
verbose_proxy_logger.debug("prisma_client is None. Skipping writing spend logs to db.")

10
litellm/proxy/db/db_url_settings.py
View file

@ -64,6 +64,7 @@ DISABLE_PREPARED_STATEMENTS_ENV_VAR: Final = "DATABASE_DISABLE_PREPARED_STATEMEN
DisablePreparedStatementsFlag = Annotated[
bool, BeforeValidator(partial(token_auth_flag_enabled, env_var=DISABLE_PREPARED_STATEMENTS_ENV_VAR))
]
MAX_IDLE_CONNECTION_LIFETIME_ENV_VAR: Final = "DATABASE_MAX_IDLE_CONNECTION_LIFETIME"
# schema.prisma pins `provider = "postgresql"`, so these are the only schemes
# Prisma can actually connect with.
@ -217,6 +218,9 @@ class DatabaseURLSettings(BaseSettings):
disable_prepared_statements: DisablePreparedStatementsFlag = Field(
default=False, validation_alias=DISABLE_PREPARED_STATEMENTS_ENV_VAR
)
max_idle_connection_lifetime: int | None = Field(
default=None, validation_alias=MAX_IDLE_CONNECTION_LIFETIME_ENV_VAR
)
# Writer
database_url: str | None = Field(default=None, validation_alias="DATABASE_URL")
@ -453,6 +457,12 @@ class DatabaseURLSettings(BaseSettings):
if url:
os.environ[env_var] = add_missing_query_params(url, MappingProxyType({"pgbouncer": "true"}))
lifetime_params: Final = idle_lifetime_params(self.max_idle_connection_lifetime)
for env_var in ("DATABASE_URL", "DIRECT_URL"):
url = os.environ.get(env_var)
if url:
os.environ[env_var] = add_missing_query_params(url, lifetime_params)
# The reader inherits the writer's connection params (pool size, timeouts,
# pgbouncer mode). Without this the reader pool ignores the configured cap
# and falls back to Prisma's `num_physical_cpus * 2 + 1` default.

6
litellm/proxy/db/prisma_client.py
View file

@ -16,6 +16,7 @@ from datetime import datetime, timedelta
from typing import Any, Final, Protocol
from litellm._logging import verbose_proxy_logger
from litellm.proxy.db.db_url_settings import add_missing_query_params, connection_params_from_url
from litellm.proxy.db.token_auth import (
DEFAULT_POSTGRES_PORT,
DatabaseTokenAuth,
@ -438,7 +439,10 @@ class PrismaWrapper:
return None
endpoint: Final = self._iam_endpoint if self._iam_endpoint is not None else self._endpoint_from_env()
db_url: Final = endpoint.build_url(mint_database_token(auth, endpoint))
db_url: Final = add_missing_query_params(
endpoint.build_url(mint_database_token(auth, endpoint)),
connection_params_from_url(os.environ.get(self._db_url_env_var, "")),
)
os.environ[self._db_url_env_var] = db_url
return db_url

147
litellm/proxy/litellm_pre_call_utils.py
View file

@ -10,6 +10,7 @@ from types import MappingProxyType
from typing import TYPE_CHECKING, Any, Final, cast
from fastapi import HTTPException, Request
from pydantic import TypeAdapter
from pydantic import ValidationError as PydanticValidationError
from starlette.datastructures import Headers
@ -18,14 +19,17 @@ from litellm._logging import verbose_logger, verbose_proxy_logger
from litellm._service_logger import ServiceLogging
from litellm._uuid import uuid
from litellm.constants import (
CLIENT_OUTPUT_CEILING_METADATA_KEY,
CONSUMED_REQUEST_TAGS_METADATA_KEY,
INTERNAL_CALL_ORIGIN_METADATA_KEY,
LITELLM_PROXY_MASTER_KEY_ALIAS,
OTEL_SERVICE_NAME_METADATA_KEYS,
PRE_CALL_EXECUTED_GUARDRAILS_KEY,
ROUTING_REQUEST_TAGS_METADATA_KEY,
SESSION_DEPLOYMENT_AFFINITY_TTL_METADATA_KEY,
SESSION_ID_GENERATED_METADATA_KEY,
SESSION_ID_OMITTED_METADATA_KEY,
X_LITELLM_DISABLE_CALLBACKS,
)
from litellm.litellm_core_utils.credential_accessor import CredentialAccessor
from litellm.litellm_core_utils.initialize_dynamic_callback_params import (
@ -157,6 +161,7 @@ from litellm.types.utils import (
CustomPricingLiteLLMParams,
LlmProviders,
ProviderSpecificHeader,
StandardCallbackDynamicParams,
StandardLoggingUserAPIKeyMetadata,
SupportedCacheControls,
)
@ -170,6 +175,7 @@ _ENABLE_TEAM_STALE_ALIAS_BYPASS: bool | None = None
if TYPE_CHECKING:
from litellm.integrations.otel.model.destination import OtelDestination
from litellm.proxy.proxy_server import ProxyConfig as _ProxyConfig
from litellm.types.proxy.policy_engine import Policy, PolicyMatchContext
@ -289,6 +295,7 @@ _UNTRUSTED_METADATA_CONTROL_FIELDS: Final = (
GATEWAY_INJECTED_CACHE_METADATA_KEY,
SESSION_DEPLOYMENT_AFFINITY_TTL_METADATA_KEY,
CONSUMED_REQUEST_TAGS_METADATA_KEY,
ROUTING_REQUEST_TAGS_METADATA_KEY,
INTERNAL_CALL_ORIGIN_METADATA_KEY,
"standard_logging_object",
"proxy_server_request",
@ -325,7 +332,9 @@ _CLIENT_PRICING_METADATA_FIELDS: Final = frozenset({"model_info", "standard_logg
# ``attempted_fallbacks`` and ``original_model_group`` are written by the router
# and read by spend logs as fact; a client value has no legitimate meaning and no
# key or team setting keeps it, so the strip is never gated.
_ROUTER_RESERVED_METADATA_FIELDS: Final = frozenset({"attempted_fallbacks", "original_model_group"})
_ROUTER_RESERVED_METADATA_FIELDS: Final = frozenset(
{"attempted_fallbacks", "original_model_group", CLIENT_OUTPUT_CEILING_METADATA_KEY}
)
_ALLOW_CLIENT_PRICING_OVERRIDE_METADATA_KEY: Final = "allow_client_pricing_override"
# Request fields whose value, when URL-valued, becomes the outbound destination
@ -970,6 +979,142 @@ def _get_dynamic_logging_metadata(
return callback_settings_obj
_TENANT_OTEL_PARAMS: Final = TypeAdapter(StandardCallbackDynamicParams)
def _tenant_otel_params(callback_vars: Mapping[str, str]) -> StandardCallbackDynamicParams:
try:
return _TENANT_OTEL_PARAMS.validate_python(callback_vars)
except PydanticValidationError:
return StandardCallbackDynamicParams()
_NO_REQUEST_HEADERS: Final[Mapping[str, str]] = MappingProxyType({})
def _dynamically_disabled_backends(
user_api_key_dict: UserAPIKeyAuth,
request_headers: Mapping[str, str] | None,
) -> frozenset[str]:
"""The callbacks this request turned off, read the way dispatch reads them.
Same sources, precedence, and premium gate ``EnterpriseCallbackControls`` applies
before it skips a callback: the ``x-litellm-disable-callbacks`` header wins over the
key's stored list, team settings are not a source, and a non-premium proxy honours
neither. A destination has to agree with that decision, or a backend the key turned
off would still be exported to, now through the fan-out instead of the callback.
"""
from litellm.proxy.proxy_server import premium_user
if litellm.allow_dynamic_callback_disabling is not True or not premium_user:
return frozenset()
header: Final = (request_headers if request_headers is not None else _NO_REQUEST_HEADERS).get(
X_LITELLM_DISABLE_CALLBACKS
)
if header is not None:
return frozenset(name.strip().lower() for name in header.split(","))
metadata: Final = user_api_key_dict.metadata
disabled: Final = metadata.get("litellm_disabled_callbacks") if metadata else None
if not isinstance(disabled, list):
return frozenset()
return frozenset(name.lower() for name in disabled if isinstance(name, str))
def resolve_tenant_otel_destinations(
user_api_key_dict: UserAPIKeyAuth,
request_headers: Mapping[str, str] | None = None,
) -> "tuple[OtelDestination, ...]":
"""The OTLP destinations this request's key or team config overrides its traces to.
Key settings win over team settings outright, the same precedence
``_get_dynamic_logging_metadata`` applies, so one caller never exports the same
backend to two accounts. An empty key-level list counts as configured, since that
is what disabling a key's callbacks writes. Returns empty when OTEL V2 is off, when
neither level named a destination-capable backend, or when the config is
incomplete, and the request then keeps the operator's own exporters.
Two entries naming the same backend merge their ``callback_vars`` last-wins, the
way ``convert_key_logging_metadata_to_callback`` merges them, so the destination
and the per-request tracer routing cannot read one config two ways.
A ``failure``-only entry is skipped: a destination is resolved during auth, before
the request has an outcome, so honouring the filter would mean holding every span
back until the call finishes. Those entries keep today's behaviour instead, where
the tenant's credentials reach the backend through per-request tracer routing and
the operator's exporter is left alone.
A backend the request disabled dynamically, through the key's
``litellm_disabled_callbacks`` or the ``x-litellm-disable-callbacks`` header in
``request_headers``, resolves to no destination, so the fan-out never carries the
request tree to that account and the operator's exporter is never suppressed for
it. That leaves the request exactly where it stood before destinations existed:
the OTel V2 logger itself is not on the disable list's class registry, so its own
span still routes to the tenant's credentials the way it did then.
"""
from litellm.integrations.otel.model.config import is_otel_v2_enabled
from litellm.integrations.otel.presets.destinations import destination_for
if not is_otel_v2_enabled():
return ()
key_entries: Final = KeyAndTeamLoggingSettings.get_key_dynamic_logging_settings(user_api_key_dict)
entries: Final = (
key_entries
if key_entries is not None
else KeyAndTeamLoggingSettings.get_team_dynamic_logging_settings(user_api_key_dict)
)
if not entries:
return ()
disabled: Final = _dynamically_disabled_backends(user_api_key_dict, request_headers)
callbacks: Final = tuple(
callback
for item in entries
if (callback := _get_validated_callback_metadata(item=item, source="otel-destination")) is not None
if callback.callback_type != "failure"
if callback.callback_name.lower() not in disabled
)
return tuple(
destination
for name in dict.fromkeys(callback.callback_name for callback in callbacks)
if (
destination := destination_for(
name,
_tenant_otel_params(
MappingProxyType(
{
var: value
for callback in callbacks
if callback.callback_name == name
for var, value in callback.callback_vars.items()
}
)
),
_tenant_service_name(user_api_key_dict),
)
)
is not None
)
def _tenant_service_name(user_api_key_dict: UserAPIKeyAuth) -> str | None:
"""The ``service.name`` this key or team configured, the key winning over its team.
Same fields and same precedence the request-metadata build applies, read straight
off the auth object because destinations resolve during auth, before that metadata
is assembled.
"""
sources: Final = (user_api_key_dict.metadata, user_api_key_dict.team_metadata)
return next(
(
stripped
for source in sources
if source
for field in OTEL_SERVICE_NAME_METADATA_KEYS
if isinstance(value := source.get(field), str) and (stripped := value.strip())
),
None,
)
def clean_headers(
headers: Headers,
litellm_key_header_name: str | None = None,

56
litellm/proxy/management_endpoints/team_callback_endpoints.py
View file

@ -20,6 +20,7 @@ from litellm.proxy._types import (
LiteLLM_AuditLogs,
LiteLLM_TeamTable,
LitellmTableNames,
LitellmUserRoles,
ProxyErrorTypes,
ProxyException,
TeamCallbackDeleteResponse,
@ -28,7 +29,10 @@ from litellm.proxy._types import (
UserAPIKeyAuth,
)
from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
from litellm.proxy.common_utils.callback_config_validation import callback_config_error
from litellm.proxy.common_utils.callback_config_validation import (
callback_config_error,
cross_entry_family_error,
)
from litellm.proxy.common_utils.callback_utils import (
_CALLBACK_VAR_ENCRYPTED_PREFIX,
decrypt_callback_vars,
@ -230,6 +234,22 @@ def _callback_error(status_code: int, message: str) -> HTTPException:
)
def _unknown_team_error(team_id: str, user_api_key_dict: UserAPIKeyAuth, status_code: int) -> HTTPException:
"""Report an unknown team without telling an unauthorized caller that it is unknown.
These routes are reachable by any authenticated caller so that a team admin can
get as far as _verify_team_access. A distinct "does not exist" would therefore let
any valid key probe which team ids exist, so a caller who could not have managed
the team either way gets the same 403 body _verify_team_access raises.
"""
if user_api_key_dict.user_role == LitellmUserRoles.PROXY_ADMIN:
return _callback_error(status_code, f"Team id = {team_id} does not exist.")
return HTTPException(
status_code=status.HTTP_403_FORBIDDEN,
detail="You do not have access to this team",
)
@router.post(
"/team/{team_id:path}/callback",
tags=["team management"],
@ -304,10 +324,7 @@ async def add_team_callbacks(
# Check if team_id exists already
_existing_team = await prisma_client.get_data(team_id=team_id, table_name="team", query_type="find_unique")
if _existing_team is None:
raise HTTPException(
status_code=400,
detail={"error": f"Team id = {team_id} does not exist. Please use a different team id."},
)
raise _unknown_team_error(team_id, user_api_key_dict, status.HTTP_400_BAD_REQUEST)
# IDOR guard: only proxy admins / org admins / team admins of THIS
# team may write callback credentials. Without this, any
@ -326,6 +343,28 @@ async def add_team_callbacks(
if team_callback_settings is None or not isinstance(team_callback_settings, list):
team_callback_settings = []
# One entry has to own a credential family end to end. The entries are
# flattened into one dict before a request reads them, so an entry
# naming only a destination would pair with a key written on another
# entry and carry it to that destination -- a key a team admin can read
# back nowhere. Repeating a value the owning entry already stores is
# fine, which is how one integration covers both events. Proxy admins
# are exempt: they already hold every credential the proxy has.
if user_api_key_dict.user_role != LitellmUserRoles.PROXY_ADMIN:
# Decrypted, because the check compares the incoming values against
# the stored ones and the credentials are encrypted at rest.
decrypted_logging: Final = decrypt_callback_vars(team_metadata).get("logging")
stored_entries: Final = decrypted_logging if isinstance(decrypted_logging, list) else ()
stored_entry_vars: Final = [ # mutable-ok: read-only input to the check, never stored
entry.get("callback_vars") or {} for entry in stored_entries
]
family_error: Final = cross_entry_family_error(data.callback_vars, stored_entry_vars)
if family_error is not None:
raise HTTPException(
status_code=status.HTTP_400_BAD_REQUEST,
detail=family_error,
)
## check if it already exists, for the same callback event
for callback in team_callback_settings:
if (
@ -452,7 +491,7 @@ async def delete_team_callback(
team_id=team_id, table_name="team", query_type="find_unique"
)
if _existing_team is None:
raise _callback_error(404, f"Team id = {team_id} does not exist.")
raise _unknown_team_error(team_id, user_api_key_dict, status.HTTP_404_NOT_FOUND)
# IDOR guard: only proxy admins / org admins / team admins of THIS team may
# deregister its callbacks, otherwise any authenticated key holder could
@ -726,10 +765,7 @@ async def get_team_callbacks(
# Check if team_id exists
_existing_team = await prisma_client.get_data(team_id=team_id, table_name="team", query_type="find_unique")
if _existing_team is None:
raise HTTPException(
status_code=404,
detail={"error": f"Team id = {team_id} does not exist."},
)
raise _unknown_team_error(team_id, user_api_key_dict, status.HTTP_404_NOT_FOUND)
# IDOR guard: callback metadata holds third-party API credentials
# (Langfuse / Langsmith / GCS). Only proxy admins / org admins /

35
litellm/proxy/prometheus_cleanup.py
View file

@ -8,10 +8,13 @@ from __future__ import annotations
import glob
import os
import re
from typing import Final
from litellm._logging import verbose_proxy_logger
_LIVE_GAUGE_PID: Final = re.compile(r"gauge_live[a-z]*_(\d+)\.db$")
def wipe_directory(directory: str) -> None:
"""Delete all .db files in the directory. Called once before workers fork."""
@ -38,3 +41,35 @@ def mark_worker_exit(worker_pid: int) -> None:
verbose_proxy_logger.info("Prometheus cleanup: marked worker %s as dead", worker_pid)
except Exception as e:
verbose_proxy_logger.warning("Failed to mark prometheus worker %s as dead: %s", worker_pid, e)
def _is_running(pid: int) -> bool:
try:
os.kill(pid, 0)
except ProcessLookupError:
return False
except PermissionError:
return True
return True
def mark_dead_workers(directory: str) -> tuple[int, ...]:
"""Drop the live-gauge files of workers that no longer exist and return their pids.
Uvicorn's multi-worker supervisor has no exit hook, so a replacement worker calls this at startup; without it
a crashed worker's in-flight gauges stay in the aggregate forever.
"""
owners: Final = frozenset(
int(match.group(1))
for match in map(_LIVE_GAUGE_PID.search, glob.glob(os.path.join(directory, "gauge_live*_*.db")))
if match is not None
)
dead: Final = tuple(sorted(pid for pid in owners if pid != os.getpid() and not _is_running(pid)))
if not dead:
return dead
from prometheus_client import multiprocess
for pid in dead:
multiprocess.mark_process_dead(pid, path=directory)
verbose_proxy_logger.info("Prometheus cleanup: marked dead workers %s in %s", dead, directory)
return dead

5
litellm/proxy/prometheus_metrics_server.py
View file

@ -30,6 +30,7 @@ from litellm.integrations.prometheus_metrics_endpoint import make_metrics_asgi_a
from litellm.llms.custom_httpx.http_handler import HTTPHandler
METRICS_PATH: Final = "/metrics"
HEALTH_PATH: Final = "/health"
PID_HEADER: Final = "x-litellm-metrics-pid"
_PARENT_POLL_INTERVAL_SECONDS: Final = 1.0
_STARTUP_TIMEOUT_SECONDS: Final = 30.0
@ -77,6 +78,10 @@ def build_metrics_app(multiproc_dir: str) -> FastAPI:
app: Final = FastAPI(title="LiteLLM Prometheus metrics", docs_url=None, redoc_url=None, openapi_url=None)
app.mount(METRICS_PATH, _add_pid_header(make_metrics_asgi_app(registry)))
@app.get(HEALTH_PATH)
def health() -> dict[str, str]:
return {"status": "healthy", "multiproc_dir": multiproc_dir}
return app

57
litellm/proxy/proxy_server.py
View file

@ -278,6 +278,7 @@ from litellm.litellm_core_utils.asyncify import asyncify
from litellm.litellm_core_utils.audio_utils.utils import resolve_speech_media_type
from litellm.litellm_core_utils.core_helpers import (
_get_parent_otel_span_from_kwargs,
drop_params_flag,
get_litellm_metadata_from_kwargs,
)
from litellm.litellm_core_utils.credential_accessor import CredentialAccessor
@ -631,6 +632,7 @@ from litellm.proxy.pass_through_endpoints.pass_through_endpoints import (
from litellm.proxy.pass_through_endpoints.pass_through_endpoints import (
router as pass_through_router,
)
from litellm.proxy.prometheus_cleanup import mark_dead_workers, mark_worker_exit
from litellm.proxy.public_endpoints import router as public_endpoints_router
from litellm.proxy.public_endpoints.public_v1 import router as public_v1_router
from litellm.proxy.rag_endpoints.endpoints import router as rag_router
@ -1059,6 +1061,10 @@ async def proxy_startup_event(app: FastAPI) -> AsyncGenerator[None, None]:
init_verbose_loggers()
prometheus_multiproc_dir: Final = os.environ.get("PROMETHEUS_MULTIPROC_DIR")
if prometheus_multiproc_dir:
mark_dead_workers(prometheus_multiproc_dir)
## RUN WORKER STARTUP HOOKS (e.g., gflags initialization) ##
_startup_hooks_env: Final = os.environ.get("LITELLM_WORKER_STARTUP_HOOKS", "")
if _startup_hooks_env:
@ -1365,6 +1371,9 @@ async def proxy_startup_event(app: FastAPI) -> AsyncGenerator[None, None]:
await proxy_shutdown_event(worker_heartbeat=worker_heartbeat)
if prometheus_multiproc_dir:
mark_worker_exit(os.getpid())
def _generate_stable_operation_id(route: "APIRoute") -> str:
operation_id = re.sub(r"\W", "_", f"{route.name}{route.path_format}")
@ -2574,10 +2583,11 @@ async def _repair_stale_spend_counter(counter_key: str, db_spend: float) -> None
)
async def reseed_spend_counter_from_db(counter_key: str) -> None:
async def reseed_spend_counter_from_db(counter_key: str) -> bool:
"""Recover a counter that the reservation reconcile found in an inconsistent
state (missing, or where applying the reconcile delta would drive it
negative) by reseeding it from the DB instead of deleting it.
negative) by reseeding it from the DB instead of deleting it. Returns
whether a DB row was found and the counter was reseeded.
The DB row is a LAGGING authoritative floor, not post-request truth: the
entity .spend column is flushed in batches (every PROXY_BATCH_WRITE_AT), so
@ -2592,8 +2602,9 @@ async def reseed_spend_counter_from_db(counter_key: str) -> None:
"""
db_spend: Final = await SpendCounterReseed.from_db(prisma_client=prisma_client, counter_key=counter_key)
if db_spend is None:
return
return False
await _repair_stale_spend_counter(counter_key=counter_key, db_spend=db_spend)
return True
async def _floor_spend_from_db(
@ -2655,21 +2666,14 @@ async def _authoritative_floor_spend(
return db_spend
async def _read_spend_counter_estimate(counter_key: str, fallback_spend: float) -> tuple[float, bool]:
"""Return (spend, authoritative). ``authoritative`` is True when the value
came from Redis or a fresh DB read (cross-pod truth), False when it came
from the per-pod in-memory copy or the caller's fallback. Only the
fail-closed path reads the flag; normal callers ignore it."""
# 1. Redis first (cross-pod authoritative). On clean miss, skip
# in-memory: per-pod in-memory only has this pod's writes, so it
# would mask cross-pod increments.
redis_clean_miss = False
async def read_spend_counter_cache_value(counter_key: str) -> tuple[float | None, bool]:
"""Return (value, authoritative) for the live counter, None when absent. A clean
Redis miss is final: the per-pod in-memory copy outlives the Redis TTL and only
holds this pod's writes, so it is consulted only when Redis is unreachable."""
if spend_counter_cache.redis_cache is not None:
try:
val = await spend_counter_cache.redis_cache.async_get_cache(key=counter_key)
if val is not None:
return float(val), True
redis_clean_miss = True
redis_val: Final = await spend_counter_cache.redis_cache.async_get_cache(key=counter_key)
return (float(redis_val) if redis_val is not None else None), True
except Exception as e:
verbose_proxy_logger.debug(
"get_current_spend: Redis read failed for %s, falling back to in-memory: %s",
@ -2677,13 +2681,20 @@ async def _read_spend_counter_estimate(counter_key: str, fallback_spend: float)
e,
)
# 2. In-memory only when Redis is unreachable.
if not redis_clean_miss:
val = spend_counter_cache.in_memory_cache.get_cache(key=counter_key)
if val is not None:
return float(val), False
in_memory_val: Final = spend_counter_cache.in_memory_cache.get_cache(key=counter_key)
return (float(in_memory_val) if in_memory_val is not None else None), False
# 3. Reseed from DB - fallback_spend lags cross-pod, would allow bypass.
async def _read_spend_counter_estimate(counter_key: str, fallback_spend: float) -> tuple[float, bool]:
"""Return (spend, authoritative). ``authoritative`` is True when the value
came from Redis or a fresh DB read (cross-pod truth), False when it came
from the per-pod in-memory copy or the caller's fallback. Only the
fail-closed path reads the flag; normal callers ignore it."""
cached_val, cached_authoritative = await read_spend_counter_cache_value(counter_key=counter_key)
if cached_val is not None:
return cached_val, cached_authoritative
# Reseed from DB - fallback_spend lags cross-pod, would allow bypass.
db_spend: Final = await SpendCounterReseed.coalesced(
prisma_client=prisma_client,
spend_counter_cache=spend_counter_cache,
@ -5509,6 +5520,8 @@ class ProxyConfig:
parse_budget_reset_time(value)
setattr(litellm, key, value)
elif key == "drop_params":
litellm.drop_params = drop_params_flag(value, "litellm_settings.drop_params", verbose_proxy_logger)
else:
verbose_proxy_logger.debug(
"%s setting litellm.%s=%s%s",

1
litellm/proxy/schema.prisma
View file

@ -1036,6 +1036,7 @@ model LiteLLM_ManagedObjectTable { // for batches or finetuning jobs which use t
created_at DateTime @default(now())
created_by String?
team_id String?
org_id String? // creating key's organization at submission time; CheckBatchCost bills org spend against it
api_key String?
request_tags Json? @default("[]")
updated_at DateTime @updatedAt

28
litellm/proxy/spend_tracking/budget_reservation.py
View file

@ -905,13 +905,13 @@ async def _set_reserved_entry_actual_cost(
increment=adjustment,
)
elif reseed_on_inconsistent:
# Post-call reconcile / release: the counter was flushed or reseeded
# between reservation and reconcile (Redis restart / cross-pod reset),
# so the optimistic delta no longer applies. Recover by reseeding from
# the DB's lagging authoritative floor rather than deleting the counter
# and failing open — deleting it is what left budgets unenforced after a
# Redis reload.
await reseed_spend_counter_from_db(counter_key=counter_key)
# Post-call reconcile / release: the counter was flushed, expired or reseeded
# between reservation and reconcile, so the optimistic delta no longer applies.
# Reseed from the DB floor (which cannot include this request's cost yet) and
# add the settled cost, since increment_spend_counters skips reserved keys.
reseeded: Final = await reseed_spend_counter_from_db(counter_key=counter_key)
if reseeded and actual_cost > 0:
await _increment_spend_counter_cache(counter_key=counter_key, increment=actual_cost)
else:
# Pre-call admission resize: the in-flight reservation cost is not yet
# persisted, so the DB floor would discard it. Keep the original
@ -925,18 +925,16 @@ async def _counter_can_apply_adjustment(
counter_key: str,
adjustment: float,
) -> bool:
from litellm.proxy.proxy_server import spend_counter_cache
from litellm.proxy.proxy_server import read_spend_counter_cache_value
current_value: Final = await spend_counter_cache.async_get_cache(key=counter_key)
try:
current_value, _ = await read_spend_counter_cache_value(counter_key=counter_key)
except (TypeError, ValueError):
return False
if current_value is None:
return False
try:
current_float: Final = float(current_value)
except (TypeError, ValueError):
return False
return not (adjustment < 0 and current_float + adjustment < -1e-12)
return not (adjustment < 0 and current_value + adjustment < -1e-12)
async def _release_applied_entries_best_effort(

27
litellm/proxy/spend_tracking/spend_tracking_utils.py
View file

@ -590,6 +590,7 @@ def get_logging_payload(kwargs, response_obj, start_time, end_time) -> SpendLogs
metadata=metadata,
standard_logging_payload=standard_logging_payload,
omit_when_missing=_omits_session_id_when_missing(metadata),
batch_trace_session_id=_get_batch_trace_session_id(call_type=call_type, request_id=id),
),
request_duration_ms=_get_request_duration_ms(start_time, end_time),
status=_get_status_for_spend_log(
@ -628,20 +629,44 @@ def _omits_session_id_when_missing(metadata: Mapping[str, object] | None) -> boo
return general_settings.get("missing_session_id") == "omit"
_BATCH_TRACE_CALL_TYPES: Final = frozenset(
{
CallTypes.create_batch.value,
CallTypes.acreate_batch.value,
CallTypes.retrieve_batch.value,
CallTypes.aretrieve_batch.value,
}
)
def _get_batch_trace_session_id(call_type: str | None, request_id: str | None) -> str | None:
"""A batch's create row and its poller-written cost row both derive their request id
from the same batch id (the cost row appends BATCH_COST_REQUEST_ID_SUFFIX), so using
that id as the session groups the batch lifecycle into one trace on the logs UI. The
poller builds its own logging context, so per-request trace ids can never link them."""
if call_type not in _BATCH_TRACE_CALL_TYPES or not request_id:
return None
return request_id.removesuffix(BATCH_COST_REQUEST_ID_SUFFIX)
def _get_session_id_for_spend_log(
kwargs: Mapping[str, object],
metadata: Mapping[str, object] | None,
standard_logging_payload: StandardLoggingPayload | None,
omit_when_missing: bool,
batch_trace_session_id: str | None = None,
) -> str | None:
"""Under `omit` only `metadata.session_id`, the key Langfuse reads, counts as a session; `litellm_session_id` may
be a copied trace id."""
be a copied trace id. Batch call types carry a deterministic session derived from the batch id, which outranks
the per-request trace ids because those differ between the create call and the cost poller's row."""
if omit_when_missing:
session_id: Final = metadata.get("session_id") if metadata else None
return str(session_id) if session_id else None
from litellm._uuid import uuid
if batch_trace_session_id is not None:
return batch_trace_session_id
if standard_logging_payload is not None and standard_logging_payload.get("trace_id") is not None:
return str(standard_logging_payload.get("trace_id"))
if kwargs.get("litellm_trace_id") is not None:

33
litellm/proxy/utils.py
View file

@ -182,6 +182,9 @@ from litellm.types.mcp import (
)
from litellm.types.proxy.policy_engine.pipeline_types import PipelineExecutionResult
from litellm.types.utils import LLMResponseTypes, LoggedLiteLLMParams
from litellm.utils import (
_add_custom_logger_callback_to_specific_event, # pyright: ignore[reportPrivateUsage] # only string-to-logger helper
)
if TYPE_CHECKING:
from mcp.types import CallToolResult
@ -1082,6 +1085,14 @@ class ProxyLogging:
litellm.logging_callback_manager.add_litellm_async_success_callback(callback)
litellm.logging_callback_manager.add_litellm_async_failure_callback(callback)
# Runs after load_config applied every litellm_settings key: logger __init__s read e.g. s3_callback_params
success_callbacks: Final = tuple(cb for cb in litellm.success_callback if isinstance(cb, str))
failure_callbacks: Final = tuple(cb for cb in litellm.failure_callback if isinstance(cb, str))
for callback in success_callbacks:
_add_custom_logger_callback_to_specific_event(callback, "success")
for callback in failure_callbacks:
_add_custom_logger_callback_to_specific_event(callback, "failure")
async def update_request_status(self, litellm_call_id: str, status: Literal["success", "fail"]):
# only use this if slack alerting is being used
if self.alerting is None:
@ -3889,7 +3900,7 @@ class _StaleReadEngine:
class PrismaClient:
spend_log_transactions: list = []
_spend_log_transactions_lock = asyncio.Lock()
spend_log_flush_requested: ClassVar[asyncio.Event] = asyncio.Event()
spend_log_flush_requested: "asyncio.Event | None" = None
spend_log_queue_bytes: ClassVar[int] = 0
spend_logs_queue_monitor_task: "asyncio.Task[None] | None" = None
tool_usage_transactions: list["ToolUsageTransaction"] = []
@ -6615,23 +6626,27 @@ async def enqueue_spend_logs(
)
def request_spend_log_flush() -> None:
"""Wake the queue monitor now rather than leaving the rows for its next poll.
def request_spend_log_flush(prisma_client: PrismaClient) -> None:
"""Wake this client's queue monitor now rather than leaving the rows for its next poll.
The Responses API hands the client an id it can chain from straight away, and that
lookup reads the DB, so the row cannot sit in this worker's queue for a poll interval.
Repeated requests coalesce into the monitor's next pass, so the batching holds.
A request made before the monitor is running is dropped, and loses nothing: the
monitor reads the queue on its first pass, before it ever waits on a request.
"""
PrismaClient.spend_log_flush_requested.set()
flush_requested: Final = prisma_client.spend_log_flush_requested
if flush_requested is not None:
flush_requested.set()
async def _wait_for_spend_log_flush_request(interval: float) -> bool:
async def _wait_for_spend_log_flush_request(flush_requested: asyncio.Event, interval: float) -> bool:
"""Wait out ``interval``, returning early and True when a flush was requested."""
try:
await asyncio.wait_for(PrismaClient.spend_log_flush_requested.wait(), timeout=interval)
await asyncio.wait_for(flush_requested.wait(), timeout=interval)
except asyncio.TimeoutError:
return False
PrismaClient.spend_log_flush_requested.clear()
flush_requested.clear()
return True
@ -7058,6 +7073,8 @@ async def _monitor_spend_logs_queue(
max_backoff: Final = 30.0 # Maximum backoff interval in seconds
backoff_multiplier: Final = 1.5 # Exponential backoff multiplier
current_interval = base_interval
flush_requested: Final = asyncio.Event()
prisma_client.spend_log_flush_requested = flush_requested # rebind-ok: the client owns its monitor's flush signal
verbose_proxy_logger.info(
"Starting spend logs queue monitor (threshold: %s, poll_interval: %ss)", threshold, base_interval
@ -7096,7 +7113,7 @@ async def _monitor_spend_logs_queue(
# Exponential backoff when no logs to process
current_interval = min(current_interval * backoff_multiplier, max_backoff)
if await _wait_for_spend_log_flush_request(current_interval):
if await _wait_for_spend_log_flush_request(flush_requested, current_interval):
current_interval = base_interval
except Exception as e:
spend_log_error("Error in spend logs queue monitor: %s", str(e), exc=e)

5
litellm/responses/main.py
View file

@ -17,6 +17,7 @@ from litellm.completion_extras.litellm_responses_transformation.transformation i
from litellm.constants import request_timeout
from litellm.integrations.anthropic_cache_control_hook import CARRY_UNMATCHED_MESSAGE_POINTS
from litellm.litellm_core_utils.asyncify import run_async_function
from litellm.litellm_core_utils.core_helpers import normalize_drop_params
from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
from litellm.litellm_core_utils.prompt_templates.common_utils import (
update_responses_input_with_model_file_ids,
@ -1257,7 +1258,7 @@ def responses(
responses_api_provider_config=responses_api_provider_config,
response_api_optional_params=response_api_optional_params,
allowed_openai_params=allowed_openai_params,
drop_params=request_drop_params if isinstance(request_drop_params, bool) else None,
drop_params=normalize_drop_params(request_drop_params),
)
litellm_logging_obj.update_from_kwargs(
@ -2085,7 +2086,7 @@ def compact_responses(
responses_api_provider_config=responses_api_provider_config,
response_api_optional_params=response_api_optional_params,
allowed_openai_params=None,
drop_params=request_drop_params if isinstance(request_drop_params, bool) else None,
drop_params=normalize_drop_params(request_drop_params),
)
# Pre Call logging

238
litellm/router.py
View file

@ -21,7 +21,16 @@ import time
import traceback
import weakref
from collections import defaultdict
from collections.abc import AsyncGenerator, AsyncIterator, Callable, Generator, Iterator, Mapping, Sequence
from collections.abc import (
AsyncGenerator,
AsyncIterator,
Callable,
Generator,
Iterator,
Mapping,
MutableMapping,
Sequence,
)
from functools import lru_cache, partial
from types import MappingProxyType
from typing import TYPE_CHECKING, Any, Final, Literal, Optional, TypeAlias, TypeVar, Union, cast
@ -45,12 +54,15 @@ from litellm.caching.caching import (
RedisClusterCache,
)
from litellm.constants import (
CLIENT_OUTPUT_CEILING_METADATA_KEY,
CONSUMED_REQUEST_TAGS_METADATA_KEY,
DEFAULT_AUTO_ROUTER_MAX_INPUT_CHARS,
DEFAULT_HEALTH_CHECK_INTERVAL,
DEFAULT_HEALTH_CHECK_STALENESS_MULTIPLIER,
DEFAULT_MAX_LRU_CACHE_SIZE,
INTERNAL_CALL_ORIGIN_METADATA_KEY,
OUTPUT_TOKEN_CEILING_PARAMS,
ROUTING_REQUEST_TAGS_METADATA_KEY,
RUNTIME_UPDATABLE_ROUTER_SETTINGS,
SESSION_DEPLOYMENT_AFFINITY_TTL_METADATA_KEY,
)
@ -648,6 +660,18 @@ class FallbackAwareStreamWrapper(CustomStreamWrapper):
self.fallback_headers_adopted = True
def as_output_cap(value: object) -> int | None:
"""A client-sent output cap coerced to an int: ints, floats and numeric strings, never bools
or negatives."""
if isinstance(value, bool) or not isinstance(value, (int, float, str)):
return None
try:
cap: Final = int(float(value))
except (ValueError, OverflowError):
return None
return cap if cap >= 0 else None
class Router:
model_names: set = set()
cache_responses: bool | None = False
@ -955,7 +979,6 @@ class Router:
DEFAULT_HEALTH_CHECK_INTERVAL * DEFAULT_HEALTH_CHECK_STALENESS_MULTIPLIER
)
self.health_state_cache = DeploymentHealthCache(cache=self.cache, staleness_threshold=float(_staleness))
self.failed_calls = InMemoryCache() # cache to track failed call per deployment, if num failed calls within 1 minute > allowed fails, then add it to cooldown
if num_retries is not None:
self.num_retries = num_retries
@ -1523,9 +1546,33 @@ class Router:
self._override_selectors[strategy] = self._build_strategy_selector(
strategy=strategy,
routing_strategy_args={},
register_callbacks=False,
)
return self._override_selectors[strategy]
def _override_selector_pre_call_check(
self, strategy: str | None, selector: RouterStrategySelector | None, deployment: dict
) -> None:
"""
Override selectors are not in `litellm.callbacks`, so the pre-call check that
`routing_strategy_pre_call_checks` runs for the router's own selectors (rpm
accounting for `usage-based-routing-v2`) runs here, for the overriding request only.
"""
if selector is None or strategy is None or selector is not self._override_selectors.get(strategy):
return
selector.pre_call_check(deployment)
async def _async_override_selector_pre_call_check(
self,
strategy: str | None,
selector: RouterStrategySelector | None,
deployment: dict,
parent_otel_span: Span | None,
) -> None:
if selector is None or strategy is None or selector is not self._override_selectors.get(strategy):
return
await selector.async_pre_call_check(deployment, parent_otel_span)
def _get_routing_context(
self, model: str, request_kwargs: dict | None = None
) -> tuple[str | None, RouterStrategySelector | None]:
@ -3609,6 +3656,13 @@ class Router:
effective_model_info: Final = kwargs.get("model_info") or deployment.get("model_info") or MappingProxyType({})
self._set_failed_deployment_id_on_exception(exception, MappingProxyType({"model_info": effective_model_info}))
@staticmethod
def _stamp_retry_skip_deployment_id(exception: Exception, kwargs: Mapping[str, object]) -> None:
effective_model_info: Final = kwargs.get("model_info")
deployment_id: Final = effective_model_info.get("id") if isinstance(effective_model_info, Mapping) else None
if isinstance(deployment_id, str) and deployment_id:
exception.retry_skip_deployment_id = deployment_id # pyright: ignore[reportAttributeAccessIssue] # dynamic stamp, read by _deployment_ids_to_skip_on_retry
def _update_kwargs_with_default_litellm_params(
self, kwargs: dict, metadata_variable_name: str | None = "metadata"
) -> None:
@ -3726,6 +3780,11 @@ class Router:
refund_stale_reservation_before_retry(self.cache, kwargs)
set_io_token_rate_limit_request_kwargs(kwargs, store_in_context=deployment_has_io_token_limits(deployment))
kwargs[metadata_variable_name].setdefault(
ROUTING_REQUEST_TAGS_METADATA_KEY,
tuple(_get_tags_from_request_kwargs(kwargs, metadata_variable_name=metadata_variable_name)),
)
## DEPLOYMENT-LEVEL TAGS
deployment_tags: Final = deployment.get("litellm_params", {}).get("tags")
if deployment_tags:
@ -4306,6 +4365,7 @@ class Router:
model=model,
messages=[{"role": "user", "content": "prompt"}],
specific_deployment=kwargs.pop("specific_deployment", None),
request_kwargs=kwargs,
)
self._update_kwargs_with_deployment(deployment=deployment, kwargs=kwargs)
data: Final = deployment["litellm_params"].copy()
@ -4336,6 +4396,7 @@ class Router:
verbose_router_logger.info("litellm.image_generation(model=%s)\x1b[31m Exception %s\x1b[0m", model_name, e)
if model_name is not None:
self.fail_calls[model_name] += 1
self._stamp_retry_skip_deployment_id(e, kwargs)
raise e
async def aimage_generation(self, prompt: str, model: str, **kwargs):
@ -4420,6 +4481,7 @@ class Router:
verbose_router_logger.info("litellm.aimage_generation(model=%s)\x1b[31m Exception %s\x1b[0m", model_name, e)
if model_name is not None:
self.fail_calls[model_name] += 1
self._stamp_retry_skip_deployment_id(e, kwargs)
raise e
async def atranscription(self, file: FileTypes, model: str, **kwargs):
@ -4524,6 +4586,7 @@ class Router:
verbose_router_logger.info("litellm.atranscription(model=%s)\x1b[31m Exception %s\x1b[0m", model_name, e)
if model_name is not None:
self.fail_calls[model_name] += 1
self._stamp_retry_skip_deployment_id(e, kwargs)
raise e
async def aspeech(self, model: str, input: str, voice: str | None = None, **kwargs):
@ -4638,6 +4701,7 @@ class Router:
verbose_router_logger.info("litellm.aspeech(model=%s)\x1b[31m Exception %s\x1b[0m", model_name, e)
if model_name is not None:
self.fail_calls[model_name] += 1
self._stamp_retry_skip_deployment_id(e, kwargs)
raise e
async def arerank(self, model: str, **kwargs):
@ -4696,6 +4760,7 @@ class Router:
verbose_router_logger.info("litellm.arerank(model=%s)\x1b[31m Exception %s\x1b[0m", model_name, e)
if model_name is not None:
self.fail_calls[model_name] += 1
self._stamp_retry_skip_deployment_id(e, kwargs)
raise e
def text_completion(
@ -4830,6 +4895,7 @@ class Router:
verbose_router_logger.info("litellm.atext_completion(model=%s)\x1b[31m Exception %s\x1b[0m", model, e)
if model is not None:
self.fail_calls[model] += 1
self._stamp_retry_skip_deployment_id(e, kwargs)
raise e
async def aadapter_completion(
@ -4920,6 +4986,7 @@ class Router:
verbose_router_logger.info("litellm.aadapter_completion(model=%s)\x1b[31m Exception %s\x1b[0m", model, e)
if model is not None:
self.fail_calls[model] += 1
self._stamp_retry_skip_deployment_id(e, kwargs)
raise e
async def _asearch_with_fallbacks(self, original_function: Callable, **kwargs):
@ -5702,6 +5769,7 @@ class Router:
model=model,
input=input,
specific_deployment=kwargs.pop("specific_deployment", None),
request_kwargs=kwargs,
)
self._update_kwargs_with_deployment(deployment=deployment, kwargs=kwargs)
data: Final = deployment["litellm_params"].copy()
@ -5740,6 +5808,7 @@ class Router:
verbose_router_logger.info("litellm.embedding(model=%s)\x1b[31m Exception %s\x1b[0m", model_name, e)
if model_name is not None:
self.fail_calls[model_name] += 1
self._stamp_retry_skip_deployment_id(e, kwargs)
raise e
async def aembedding(
@ -5827,6 +5896,7 @@ class Router:
verbose_router_logger.info("litellm.aembedding(model=%s)\x1b[31m Exception %s\x1b[0m", model_name, e)
if model_name is not None:
self.fail_calls[model_name] += 1
self._stamp_retry_skip_deployment_id(e, kwargs)
raise e
#### FILES API ####
@ -6200,6 +6270,7 @@ class Router:
)
if model is not None:
self.fail_calls[model] += 1
self._stamp_retry_skip_deployment_id(e, kwargs)
raise e
async def aretrieve_batch(
@ -6420,6 +6491,7 @@ class Router:
)
if model is not None:
self.fail_calls[model] += 1
self._stamp_retry_skip_deployment_id(e, kwargs)
raise e
async def alist_batches(
@ -7537,7 +7609,9 @@ class Router:
@staticmethod
def _deployment_ids_to_skip_on_retry(exception: Exception, already_skipped: object) -> tuple[str, ...]:
failed_deployment_id: Final[str | None] = getattr(exception, "failed_deployment_id", None)
failed_deployment_id: Final[str | None] = getattr(exception, "retry_skip_deployment_id", None) or getattr(
exception, "failed_deployment_id", None
)
status_code: Final = getattr(exception, "status_code", None)
if not failed_deployment_id or not isinstance(status_code, int):
return ()
@ -9368,6 +9442,12 @@ class Router:
#### VALIDATE MODEL ########
# Check if this is a prompt management model before validating as LLM provider
litellm_model: Final = deployment.litellm_params.model
if isinstance(deployment.litellm_params.drop_params, str):
verbose_router_logger.warning(
"model=%s drop_params=%r is not a flag value, treating it as unset",
deployment.model_name,
deployment.litellm_params.drop_params,
)
is_prompt_management_model = False
if "/" in litellm_model:
@ -12606,7 +12686,83 @@ class Router:
request_kwargs.pop(carrier, None)
@staticmethod
def _drop_client_effort_carriers_a_tier_pin_supersedes(
def _tier_ceiling_under_the_surface_name(
tier_litellm_params: Mapping[str, object], responses_call: bool
) -> Mapping[str, object]:
"""``max_tokens``, ``max_completion_tokens`` and ``max_output_tokens`` are one
ceiling under three names, and each surface reads exactly one of them: the
Responses bridge builds its internal ``max_tokens`` from ``max_output_tokens``
and would overwrite the tier's, chat and /v1/messages never read
``max_output_tokens``, and litellm already renames ``max_tokens`` to
``max_completion_tokens`` for the OpenAI models that require it. Collapse
whatever the tier carries onto the surface's own name, preferring a value the
operator already wrote under that name."""
surface_key: Final = "max_output_tokens" if responses_call else "max_tokens"
carried: Final = tuple(
key
for key in (surface_key, "max_tokens", "max_completion_tokens", "max_output_tokens")
if key in tier_litellm_params
)
if not carried:
return tier_litellm_params
return MappingProxyType(
{
**{k: v for k, v in tier_litellm_params.items() if k not in OUTPUT_TOKEN_CEILING_PARAMS},
surface_key: tier_litellm_params[carried[0]],
}
)
def _pin_tier_params_onto_request(
self,
model: str,
tier_litellm_params: Mapping[str, object] | None,
request_kwargs: dict,
responses_call: bool,
) -> bool:
"""Apply a routing strategy's per-tier litellm_params on top of the request and report
whether they pinned an output ceiling, so the caller can hand the request its own ceiling
back on a routing pass that pins none."""
if not tier_litellm_params:
return False
accepted_tier_params: Final = self._tier_params_the_target_accepts(model, tier_litellm_params, request_kwargs)
surface_tier_params: Final = self._tier_ceiling_under_the_surface_name(
accepted_tier_params, responses_call=responses_call
)
self._drop_client_carriers_a_tier_pin_supersedes(request_kwargs, surface_tier_params)
request_kwargs.update(surface_tier_params)
return not OUTPUT_TOKEN_CEILING_PARAMS.isdisjoint(surface_tier_params)
@staticmethod
def _restore_client_ceiling_no_tier_pins(request_kwargs: MutableMapping[str, object]) -> None:
"""A model-group fallback re-enters routing with the kwargs an earlier auto-router pass
already rewrote, so a ceiling sized for that pass's tier would ride onto a group no tier
chose. When this pass pins none, hand the request back exactly the carriers the caller
sent, which the first pinning pass stamped. The stamp lives in a metadata bucket a
caller can also write, so the proxy strips the key at ingestion and this read takes
nothing but the three ceiling carriers as integers: no other key ever reaches kwargs."""
stamped: Final = next(
(
bucket.get(CLIENT_OUTPUT_CEILING_METADATA_KEY)
for bucket in (request_kwargs.get("metadata"), request_kwargs.get("litellm_metadata"))
if isinstance(bucket, dict) and CLIENT_OUTPUT_CEILING_METADATA_KEY in bucket
),
None,
)
if not isinstance(stamped, dict):
return
callers_ceiling: Final = MappingProxyType(
{
carrier: cap
for carrier, value in stamped.items()
if carrier in OUTPUT_TOKEN_CEILING_PARAMS and (cap := as_output_cap(value)) is not None
}
)
for carrier in OUTPUT_TOKEN_CEILING_PARAMS:
request_kwargs.pop(carrier, None)
request_kwargs.update(callers_ceiling)
@staticmethod
def _drop_client_carriers_a_tier_pin_supersedes(
request_kwargs: dict[str, object],
tier_litellm_params: Mapping[str, object],
) -> None:
@ -12616,7 +12772,22 @@ class Router:
the ``reasoning_effort`` alias, so a pinned effort only reaches the wire
if the client's other encodings are removed before the merge. Non-effort
fields a carrier also holds (``output_config.format``,
``reasoning.summary``) are kept."""
``reasoning.summary``) are kept. An output ceiling has the same shape:
``max_tokens``, ``max_completion_tokens`` and ``max_output_tokens`` are
one setting under three names, and a provider handed two of them either
rejects the request or picks one by iteration order."""
if not OUTPUT_TOKEN_CEILING_PARAMS.isdisjoint(tier_litellm_params):
_, metadata_bucket = get_or_create_metadata_bucket(request_kwargs)
metadata_bucket.setdefault(
CLIENT_OUTPUT_CEILING_METADATA_KEY,
{
carrier: request_kwargs[carrier]
for carrier in OUTPUT_TOKEN_CEILING_PARAMS
if carrier in request_kwargs
},
)
for carrier in OUTPUT_TOKEN_CEILING_PARAMS:
request_kwargs.pop(carrier, None)
if "reasoning_effort" not in tier_litellm_params:
return
request_kwargs.pop("thinking", None)
@ -12657,6 +12828,7 @@ class Router:
# Execute Pre-Routing Hooks
# this hook can modify the model, messages before the routing decision is made
#########################################################
responses_call: Final = input is not None and messages is None
pre_routing_hook_response: Final = await self.async_pre_routing_hook(
model=model,
request_kwargs=request_kwargs,
@ -12668,12 +12840,14 @@ class Router:
model = pre_routing_hook_response.model
messages = pre_routing_hook_response.messages
record_pre_routing_selection(request_kwargs, model)
if pre_routing_hook_response.litellm_params:
accepted_tier_params: Final = self._tier_params_the_target_accepts(
model, pre_routing_hook_response.litellm_params, request_kwargs
)
self._drop_client_effort_carriers_a_tier_pin_supersedes(request_kwargs, accepted_tier_params)
request_kwargs.update(accepted_tier_params)
tier_pins_ceiling: Final = self._pin_tier_params_onto_request(
model=model,
tier_litellm_params=pre_routing_hook_response.litellm_params if pre_routing_hook_response else None,
request_kwargs=request_kwargs,
responses_call=responses_call,
)
if not tier_pins_ceiling:
self._restore_client_ceiling_no_tier_pins(request_kwargs)
#########################################################
# Resolve the strategy and logger AFTER the pre-routing hook, since
@ -12690,10 +12864,16 @@ class Router:
parent_otel_span=parent_otel_span,
)
if isinstance(healthy_deployments, dict):
await self._async_override_selector_pre_call_check(
strategy, strategy_selector, healthy_deployments, parent_otel_span
)
return healthy_deployments
# When encrypted content affinity pins to a specific deployment,
if request_kwargs.get("_encrypted_content_affinity_pinned") and len(healthy_deployments) == 1:
await self._async_override_selector_pre_call_check(
strategy, strategy_selector, healthy_deployments[0], parent_otel_span
)
return healthy_deployments[0]
start_time: Final = time.time()
@ -12719,6 +12899,9 @@ class Router:
parent_otel_span=parent_otel_span,
)
raise exception
await self._async_override_selector_pre_call_check(
strategy, strategy_selector, deployment, parent_otel_span
)
verbose_router_logger.info(
"get_available_deployment for model: %s, Selected deployment: %s for model: %s",
model,
@ -12773,6 +12956,7 @@ class Router:
parent_otel_span: Final = _get_parent_otel_span_from_kwargs(request_kwargs)
# 1. Execute pre-routing hook
responses_call: Final = input is not None and messages is None
pre_routing_hook_response: Final = await self.async_pre_routing_hook(
model=model,
request_kwargs=request_kwargs,
@ -12784,12 +12968,14 @@ class Router:
model = pre_routing_hook_response.model
messages = pre_routing_hook_response.messages
record_pre_routing_selection(request_kwargs, model)
if pre_routing_hook_response.litellm_params:
accepted_tier_params: Final = self._tier_params_the_target_accepts(
model, pre_routing_hook_response.litellm_params, request_kwargs
)
self._drop_client_effort_carriers_a_tier_pin_supersedes(request_kwargs, accepted_tier_params)
request_kwargs.update(accepted_tier_params)
tier_pins_ceiling: Final = self._pin_tier_params_onto_request(
model=model,
tier_litellm_params=pre_routing_hook_response.litellm_params if pre_routing_hook_response else None,
request_kwargs=request_kwargs,
responses_call=responses_call,
)
if not tier_pins_ceiling:
self._restore_client_ceiling_no_tier_pins(request_kwargs)
# 2. Get healthy deployments
healthy_deployments: Final = await self.async_get_healthy_deployments(
@ -12801,6 +12987,8 @@ class Router:
parent_otel_span=parent_otel_span,
)
strategy, strategy_selector = self._get_routing_context(model, request_kwargs)
# 3. If specific deployment returned, verify if it supports pass-through
if isinstance(healthy_deployments, dict):
if (healthy_deployments.get("model_info") or {}).get("blocked") is True:
@ -12811,6 +12999,9 @@ class Router:
)
litellm_params: Final = healthy_deployments.get("litellm_params", {})
if litellm_params.get("use_in_pass_through"):
await self._async_override_selector_pre_call_check(
strategy, strategy_selector, healthy_deployments, parent_otel_span
)
return healthy_deployments
else:
raise litellm.BadRequestError(
@ -12831,7 +13022,6 @@ class Router:
# 5. Apply load balancing strategy
start_time: Final = time.perf_counter()
strategy, strategy_selector = self._get_routing_context(model, request_kwargs)
if strategy == "simple-shuffle":
return simple_shuffle(
llm_router_instance=self,
@ -12855,6 +13045,9 @@ class Router:
parent_otel_span=parent_otel_span,
)
raise exception
await self._async_override_selector_pre_call_check(
strategy, strategy_selector, deployment, parent_otel_span
)
verbose_router_logger.info(
"async_get_available_deployment_for_pass_through model: %s, selected deployment: %s",
@ -13430,6 +13623,7 @@ class Router:
specific_deployment=specific_deployment,
request_kwargs=request_kwargs,
)
strategy, strategy_selector = self._get_routing_context(model, request_kwargs)
if isinstance(healthy_deployments, dict):
if (healthy_deployments.get("model_info") or {}).get("blocked") is True:
@ -13438,6 +13632,7 @@ class Router:
model=model,
llm_provider="",
)
self._override_selector_pre_call_check(strategy, strategy_selector, healthy_deployments)
return healthy_deployments
parent_otel_span: Final[Span | None] = _get_parent_otel_span_from_kwargs(request_kwargs)
@ -13513,7 +13708,6 @@ class Router:
cooldown_list=_cooldown_list,
)
strategy, strategy_selector = self._get_routing_context(model, request_kwargs)
if strategy == "simple-shuffle":
# if users pass rpm or tpm, we do a random weighted pick - based on rpm/tpm
############## Check 'weight' param set for weighted pick #################
@ -13545,6 +13739,7 @@ class Router:
enable_pre_call_checks=self.enable_pre_call_checks,
cooldown_list=_cooldown_list,
)
self._override_selector_pre_call_check(strategy, strategy_selector, deployment)
verbose_router_logger.info(
"get_available_deployment for model: %s, Selected deployment: %s for model: %s",
model,
@ -13588,6 +13783,8 @@ class Router:
specific_deployment=specific_deployment,
)
strategy, strategy_selector = self._get_routing_context(model, request_kwargs)
# 2. If the returned is a specific deployment (Dict), verify and return directly
if isinstance(healthy_deployments, dict):
if (healthy_deployments.get("model_info") or {}).get("blocked") is True:
@ -13598,6 +13795,7 @@ class Router:
)
litellm_params: Final = healthy_deployments.get("litellm_params", {})
if litellm_params.get("use_in_pass_through"):
self._override_selector_pre_call_check(strategy, strategy_selector, healthy_deployments)
return healthy_deployments
else:
# Specific deployment does not support pass-through
@ -13657,7 +13855,6 @@ class Router:
)
# 6. Apply load balancing strategy
strategy, strategy_selector = self._get_routing_context(model, request_kwargs)
if strategy == "simple-shuffle":
return simple_shuffle(
llm_router_instance=self,
@ -13689,6 +13886,7 @@ class Router:
enable_pre_call_checks=self.enable_pre_call_checks,
cooldown_list=_cooldown_list,
)
self._override_selector_pre_call_check(strategy, strategy_selector, deployment)
verbose_router_logger.info(
"get_available_deployment_for_pass_through model: %s, selected deployment: %s",

12
litellm/router_strategy/complexity_router/README.md
View file

@ -207,17 +207,27 @@ custom_dimensions:
- name: sqlMigration
weight: 0.7
patterns: ['\b(create|alter|drop)\s{1,4}table\b']
- name: dataPipeline
weight: 0.4
scoring_mode: match_count
keywords: [airflow, dbt, snowflake]
```
Each dimension contributes its weight once when any matcher hits the current ask. Repeated matches do not increase it. The built-in score and tier boundaries are unchanged, and the total score is not renormalized. Keywords use the existing case-insensitive word-boundary and CJK rules. Regexes search the first 2048 characters case-insensitively and compile during configuration validation and router initialization, never per request
`scoring_mode` is optional and defaults to `binary`, the behavior above. `match_count` grades the dimension by how many distinct matchers hit: none scores 0 and emits no signal, one scores half the weight, two or more score the full weight. Repeated occurrences of one matcher never raise the count, keywords are distinct case-insensitively, patterns are distinct by source, and a keyword and a pattern are always distinct from each other. Matching stops as soon as the selected mode's maximum is reached, so a binary dimension still stops at its first hit. Existing configurations without the field keep binary scoring and the same tuning fingerprint, so the field only counts as a tuning change when set to `match_count`
### Weights through the API versus the dashboard
The API and YAML store exactly the weights written. A `dimension_weights` map and inline custom weights are read literally, missing recognized built-in names score zero, and nothing renormalizes the vector, so a total other than 1 is legal and scores accordingly. The dashboard's heuristic scoring editor is the one place that rebalances: editing one weight there holds it and redistributes the remainder across the other active dimensions in the draft, then Save sends the resulting explicit values, which the backend stores and scores as written. Opening a router, applying a preset, editing matchers, changing `scoring_mode`, or saving unrelated fields never normalizes existing weights
Only `heuristic`, `heuristic_first` and `hybrid` accept custom dimensions. Each name must be a unique ASCII identifier starting with a letter, at most 64 characters, and cannot reuse a built-in dimension name or a key in `dimension_weights`. Set its weight inline, greater than zero and at most one
Patterns are checked at configuration time against a grammar whose worst case stays a few milliseconds on 2048 characters. Every quantifier needs an explicit upper bound of at most 64 and must repeat a single character or character class, so `\s{1,4}` is accepted while `\s+`, `(a|aa){0,12}` and `(?:ab){0,64}` are refused. Backreferences, lookarounds, atomic groups and possessive quantifiers are refused as well. Each pattern is then costed: alternation branches and repeat lengths multiply the ways the engine can retry, and every later piece of the pattern is charged once per path that can reach it, so `a?a?a?a?a?a?a?a?` followed by a long fixed tail is refused even though each quantifier is small. The budget is 2048 work units per pattern and 8192 across the router. An invalid or over-budget pattern fails the write with a message naming the pattern and the rule it broke
Limits are 16 dimensions, 32 combined keywords/patterns per dimension, 256 characters per matcher and 4096 matcher characters per dimension. Matching runs inline on the request path with no timeout and no worker thread, because the grammar is what bounds the cost. These are routing hints, not security enforcement rules
The existing heuristic-v1 tuning quota covers custom dimensions and their weights: one changed router without an auto-router license, unlimited with the entitlement. Omitting `custom_dimensions` preserves existing scoring. Routing decisions and spend logs include signals such as `custom (sqlMigration)` without recording the configured pattern or matched text. The field is configured through YAML or the model API; this change adds no dashboard editor
The existing heuristic-v1 tuning quota covers custom dimensions, their weights and their scoring mode: one changed router without an auto-router license, unlimited with the entitlement. Omitting `custom_dimensions` preserves existing scoring. Routing decisions and spend logs include signals such as `custom (sqlMigration)` without recording the configured pattern or matched text. The field is configured through YAML, the model API, or the dashboard's heuristic scoring editor
## Usage

95
litellm/router_strategy/complexity_router/complexity_router.py
View file

@ -20,7 +20,7 @@ import random
import re
import time
from collections.abc import Callable, Iterator, Mapping, Sequence
from itertools import accumulate, islice, takewhile
from itertools import accumulate, chain, islice, takewhile
from threading import Lock
from types import MappingProxyType
from typing import TYPE_CHECKING, Any, Final, Literal, NamedTuple, cast
@ -31,6 +31,7 @@ from litellm._logging import verbose_router_logger
from litellm.constants import (
EMPTY_MAPPING,
INTERNAL_CALL_ORIGIN_METADATA_KEY,
OUTPUT_TOKEN_CEILING_PARAMS,
RETURN_RAW_MODEL_NAME_METADATA_KEY,
SESSION_ID_GENERATED_METADATA_KEY,
)
@ -82,6 +83,7 @@ from .config import (
ClassificationRubric,
ComplexityRouterConfig,
ComplexityTier,
CustomDimension,
TierDefinition,
)
from .stall_detector import detect_stalled_task
@ -879,6 +881,15 @@ class DimensionScore:
self.signal = signal
class _CustomDimensionMatchers(NamedTuple):
"""One custom dimension's distinct matchers and the number of hits that saturates its score."""
dimension: CustomDimension
keywords: tuple[str, ...]
patterns: tuple[re.Pattern[str], ...]
saturation: int
class KeywordOverride(NamedTuple):
"""A keyword_tier_rules match: the winning tier and, on the lexical path, the keyword that fired."""
@ -1121,7 +1132,12 @@ class ComplexityRouter(CustomLogger):
)
self.simple_keywords = self.config.simple_keywords or DEFAULT_SIMPLE_KEYWORDS
self._custom_dimensions = tuple(
(dimension, tuple(re.compile(pattern, re.IGNORECASE) for pattern in dimension.patterns))
_CustomDimensionMatchers(
dimension,
tuple(dict.fromkeys(keyword.lower() for keyword in dimension.keywords)),
tuple(re.compile(pattern, re.IGNORECASE) for pattern in dict.fromkeys(dimension.patterns)),
2 if dimension.scoring_mode == "match_count" else 1,
)
for dimension in self.config.custom_dimensions
)
if self.config.has_custom_tiers:
@ -1325,15 +1341,26 @@ class ComplexityRouter(CustomLogger):
score: Final = score_high if match_count >= high_threshold else score_low
return DimensionScore(name, score, f"{signal_label} ({detail})"), match_count
def _count_custom_hits(self, matchers: _CustomDimensionMatchers, user_text: str, scanned: str) -> int:
hits: Final = chain(
(self._keyword_matches(user_text, keyword) for keyword in matchers.keywords),
(pattern.search(scanned) is not None for pattern in matchers.patterns),
)
return sum(islice((1 for hit in hits if hit), matchers.saturation))
def _score_custom_dimensions(self, prompt: str, user_text: str) -> tuple[tuple[DimensionScore, float], ...]:
if not self._custom_dimensions:
return ()
scanned: Final = prompt[:CUSTOM_PATTERN_SCAN_CHARS]
return tuple(
(DimensionScore(dimension.name, 1.0, f"custom ({dimension.name})"), dimension.weight)
for dimension, patterns in self._custom_dimensions
if any(self._keyword_matches(user_text, keyword) for keyword in dimension.keywords)
or any(pattern.search(scanned) is not None for pattern in patterns)
(
DimensionScore(
matchers.dimension.name, hits / matchers.saturation, f"custom ({matchers.dimension.name})"
),
matchers.dimension.weight,
)
for matchers in self._custom_dimensions
if (hits := self._count_custom_hits(matchers, user_text, scanned))
)
def _score_multi_step(self, text: str) -> DimensionScore:
@ -2084,11 +2111,15 @@ class ComplexityRouter(CustomLogger):
raise ValueError(f"No model configured for tier {tier_key} and no default_model set")
def _litellm_params_for_model(self, tier: ComplexityTier | str | None, model: str) -> Mapping[str, object]:
if tier is None:
return MappingProxyType({})
entries: Final = self.config.tier_model_configs.get(_tier_name(tier), ())
entries: Final = self.config.tier_model_configs.get(_tier_name(tier), ()) if tier is not None else ()
entry: Final = next((candidate for candidate in entries if candidate.model_name == model), None)
return entry.litellm_params if entry is not None else MappingProxyType({})
explicit: Final = entry.litellm_params if entry is not None else MappingProxyType({})
if not self.config.max_tokens_from_tier_model or not OUTPUT_TOKEN_CEILING_PARAMS.isdisjoint(explicit):
return explicit
ceiling: Final = self._group_output_ceiling(model)
if ceiling is None:
return explicit
return MappingProxyType({**explicit, "max_tokens": ceiling})
@staticmethod
def _pick_from_tier_value(model: str | Sequence[str], tier_key: str) -> str:
@ -2423,12 +2454,15 @@ class ComplexityRouter(CustomLogger):
return name if self.config.has_custom_tiers else ComplexityTier(name)
def _deployment_window(self, group: str, deployment: Mapping[str, object]) -> int | None:
return self._deployment_limit(group, deployment, "max_input_tokens")
def _deployment_limit(
self, group: str, deployment: Mapping[str, object], key: Literal["max_input_tokens", "max_output_tokens"]
) -> int | None:
from litellm.litellm_core_utils.get_llm_provider_logic import declared_authenticating_provider
deployment_model_info: Final = deployment.get("model_info")
declared: Final = (
deployment_model_info.get("max_input_tokens") if isinstance(deployment_model_info, Mapping) else None
)
declared: Final = deployment_model_info.get(key) if isinstance(deployment_model_info, Mapping) else None
if isinstance(declared, int):
return declared
litellm_params: Final = deployment.get("litellm_params")
@ -2445,18 +2479,34 @@ class ComplexityRouter(CustomLogger):
deployment=cast(dict, deployment), # cast-ok: router deployments are plain dicts
received_model_name=group,
)
window: Final = model_info.get("max_input_tokens")
limit: Final = model_info.get(key)
except Exception: # noqa: BLE001 # best-effort: an unmappable deployment must not hide the others
return None
return window if isinstance(window, int) else None
return limit if isinstance(limit, int) else None
def _group_deployments(self, group: str) -> Sequence[Mapping[str, object]]:
list_models: Final = getattr(self.litellm_router_instance, "get_model_list", None)
deployments: Final = list_models(model_name=group) if callable(list_models) else None
return tuple(deployments) if isinstance(deployments, list) else ()
def _group_output_ceiling(self, group: str) -> int | None:
"""Smallest max_output_tokens across the group's deployments, or None when any deployment
declares none: the core router picks within the group without a fit check, and a ceiling
above an unmapped member's real limit is a provider 400 on that member."""
deployments: Final = self._group_deployments(group)
ceilings: Final = tuple(
ceiling
for deployment in deployments
if (ceiling := self._deployment_limit(group, deployment, "max_output_tokens")) is not None
)
return min(ceilings) if ceilings and len(ceilings) == len(deployments) else None
def _group_window_facts(self, group: str) -> tuple[int | None, bool]:
"""(smallest declared context window across the group's deployments, whether any deployment
declares none). The core router picks a deployment within the group without a fit check, so
the group is only as safe as its smallest member."""
list_models: Final = getattr(self.litellm_router_instance, "get_model_list", None)
deployments: Final = list_models(model_name=group) if callable(list_models) else None
if not isinstance(deployments, list) or not deployments:
deployments: Final = self._group_deployments(group)
if not deployments:
return (None, True)
windows: Final = tuple(
window for deployment in deployments if (window := self._deployment_window(group, deployment)) is not None
@ -3505,6 +3555,7 @@ class ComplexityRouter(CustomLogger):
ComplexityTier.MEDIUM, messages, resolved_messages, request_kwargs
)
fallback_tier: Final = None if default_model_first else ComplexityTier.MEDIUM
default_tier_params: Final = self._litellm_params_for_model(fallback_tier, routed_model)
return PreRoutingHookResponse(
model=routed_model,
messages=messages if has_original_messages else None,
@ -3513,7 +3564,9 @@ class ComplexityRouter(CustomLogger):
cause="default_fallback",
tier=fallback_tier,
conversation_continuing=conversation_continuing,
tier_litellm_params=default_tier_params,
),
litellm_params=default_tier_params,
)
ask: Final = user_message or ""
@ -3540,6 +3593,7 @@ class ComplexityRouter(CustomLogger):
_tier_name(plan_floor),
routed_model,
)
plan_tier_params: Final = self._litellm_params_for_model(plan_floor, routed_model)
return PreRoutingHookResponse(
model=routed_model,
messages=messages if has_original_messages else None,
@ -3551,7 +3605,9 @@ class ComplexityRouter(CustomLogger):
matched_keyword=plan_mode_sentinel,
escalation_keyword=escalation_keyword,
escalated=False,
tier_litellm_params=plan_tier_params,
),
litellm_params=plan_tier_params,
)
override: Final = await self._resolve_keyword_tier_override(ask, request_kwargs)
@ -3644,6 +3700,7 @@ class ComplexityRouter(CustomLogger):
outcome.signals,
fallback_model,
)
fallback_tier_params: Final = self._litellm_params_for_model(None, fallback_model)
return PreRoutingHookResponse(
model=fallback_model,
messages=messages if has_original_messages else None,
@ -3654,7 +3711,9 @@ class ComplexityRouter(CustomLogger):
signals=outcome.signals,
escalation_keyword=escalation_keyword,
escalated=False,
tier_litellm_params=fallback_tier_params,
),
litellm_params=fallback_tier_params,
)
if self.config.adaptive:
# hard_floor rather than a hard pick, and passed whenever the sentinel is present

27
litellm/router_strategy/complexity_router/config.py
View file

@ -667,6 +667,14 @@ class CustomDimension(BaseModel):
weight: float = Field(gt=0, le=1, allow_inf_nan=False)
keywords: tuple[Annotated[str, Field(min_length=1, max_length=256)], ...] = Field(default=(), max_length=32)
patterns: tuple[Annotated[str, Field(min_length=1, max_length=256)], ...] = Field(default=(), max_length=32)
scoring_mode: Literal["binary", "match_count"] = Field(
default="binary",
description=(
"'binary' scores 1 when any matcher hits. 'match_count' scores 0.5 when one distinct matcher hits and 1 "
"when two or more do; repeated occurrences of one matcher never raise it. Keywords are distinct "
"case-insensitively, patterns by source, and a keyword and a pattern are always distinct from each other."
),
)
@model_validator(mode="after")
def _validate_matchers(self) -> "CustomDimension":
@ -794,8 +802,9 @@ class ComplexityRouterConfig(BaseModel):
default=(),
max_length=16,
description=(
"Named binary dimensions added to the heuristic-v1 score. Each contributes its inline weight once "
"when any keyword matches the current ask or a case-insensitive regex matches its first 2048 characters. "
"Named dimensions added to the heuristic-v1 score. Each contributes its inline weight once "
"when any keyword matches the current ask or a case-insensitive regex matches its first 2048 characters; "
"scoring_mode 'match_count' instead grades half weight for one distinct matcher and full for two or more. "
"Regex quantifiers repeat one character or class at most 64 times. Unbounded quantifiers, repeated groups, "
"backreferences and lookarounds are rejected. Conservative work limits include alternation paths, "
"repeat lengths and subsequent matching: 2048 units per pattern, 8192 across the router. "
@ -1080,6 +1089,20 @@ class ComplexityRouterConfig(BaseModel):
"wording the built-ins don't cover, or after a client release changes its strings."
),
)
max_tokens_from_tier_model: bool = Field(
default=True,
description=(
"Set max_tokens on every routed request to the output ceiling of the tier model it "
"lands on, replacing whatever the caller sent. A caller behind an auto-router cannot "
"pick one value that fits every tier: the smallest tier's ceiling starves a bigger "
"tier's thinking budget, and a bigger tier's ceiling is rejected by the smallest. The "
"ceiling is the smallest max_output_tokens across the tier model's deployments, read "
"from each deployment's model_info and then the model cost map; a tier model with a "
"deployment whose ceiling is unknown keeps the caller's value. A max_tokens, "
"max_completion_tokens or max_output_tokens in the tier's own litellm_params still "
"wins. Set false to forward the caller's value unchanged."
),
)
route_housekeeping_to_cheapest_tier: bool = Field(
default=True,
description=(

11
litellm/router_strategy/lowest_cost.py
View file

@ -39,7 +39,7 @@ class LowestCostLoggingHandler(CustomLogger):
# ------------
"""
{
{model_group}_map: {
cost_map:{model_group}: {
id: {
f"{date:hour:minute}" : {"tpm": 34, "rpm": 3}
}
@ -50,7 +50,7 @@ class LowestCostLoggingHandler(CustomLogger):
current_hour: Final = datetime.now().strftime("%H")
current_minute: Final = datetime.now().strftime("%M")
precise_minute: Final = f"{current_date}-{current_hour}-{current_minute}"
cost_key: Final = f"{model_group}_map"
cost_key: Final = f"cost_map:{model_group}"
total_tokens = 0
@ -112,15 +112,14 @@ class LowestCostLoggingHandler(CustomLogger):
# ------------
"""
{
{model_group}_map: {
cost_map:{model_group}: {
id: {
"cost": [..]
f"{date:hour:minute}" : {"tpm": 34, "rpm": 3}
}
}
}
"""
cost_key: Final = f"{model_group}_map"
cost_key: Final = f"cost_map:{model_group}"
current_date: Final = datetime.now().strftime("%Y-%m-%d")
current_hour: Final = datetime.now().strftime("%H")
@ -176,7 +175,7 @@ class LowestCostLoggingHandler(CustomLogger):
"""
Returns a deployment with the lowest cost
"""
cost_key: Final = f"{model_group}_map"
cost_key: Final = f"cost_map:{model_group}"
request_count_dict: Final = await self.router_cache.async_get_cache(key=cost_key) or {}

3
litellm/router_strategy/simple_shuffle.py
View file

@ -41,8 +41,7 @@ def simple_shuffle(
############## Check if 'weight' or 'rpm' or 'tpm' param set for a weighted pick #################
for weight_by in ["weight", "rpm", "tpm"]:
weight = healthy_deployments[0].get("litellm_params").get(weight_by, None)
if weight is not None:
if any(m["litellm_params"].get(weight_by) is not None for m in healthy_deployments):
weights = [m["litellm_params"].get(weight_by, 0) for m in healthy_deployments]
verbose_router_logger.debug("\nweight %s", weights)
total_weight = sum(weights)

11
litellm/router_strategy/tag_based_routing.py
View file

@ -13,7 +13,7 @@ from types import MappingProxyType
from typing import TYPE_CHECKING, Any, Final, Literal, Protocol, overload
from litellm._logging import verbose_logger
from litellm.constants import CONSUMED_REQUEST_TAGS_METADATA_KEY
from litellm.constants import CONSUMED_REQUEST_TAGS_METADATA_KEY, ROUTING_REQUEST_TAGS_METADATA_KEY
from litellm.litellm_core_utils.core_helpers import get_metadata_variable_name_from_kwargs
from litellm.types.router import ConsumedRequestTagsStamp, DeploymentTypedDict, RouterErrors
@ -461,7 +461,10 @@ def _request_tags_after_router_consumption(metadata: object, model: str) -> Sequ
if not isinstance(metadata, Mapping):
return None
typed_metadata: Final[Mapping[str, object]] = metadata
request_tags: Final = _tags_in_metadata(typed_metadata)
request_tags: Final = _tags_in_metadata(
typed_metadata,
key=ROUTING_REQUEST_TAGS_METADATA_KEY if ROUTING_REQUEST_TAGS_METADATA_KEY in typed_metadata else "tags",
)
stamp: Final = typed_metadata.get(CONSUMED_REQUEST_TAGS_METADATA_KEY)
if not isinstance(stamp, ConsumedRequestTagsStamp) or stamp.model_group != model:
return request_tags
@ -646,7 +649,7 @@ async def get_deployments_for_tag(
return healthy_deployments
def _tags_in_metadata(metadata: object) -> list[str]:
def _tags_in_metadata(metadata: object, key: str = "tags") -> list[str]:
"""
Tags out of a metadata bucket the caller controls the shape of.
@ -657,7 +660,7 @@ def _tags_in_metadata(metadata: object) -> list[str]:
if not isinstance(metadata, Mapping):
return []
typed_metadata: Final[Mapping[str, object]] = metadata
tags: Final = typed_metadata.get("tags")
tags: Final = typed_metadata.get(key)
if isinstance(tags, str) or not isinstance(tags, Sequence):
return []
typed_tags: Final[Sequence[object]] = tags

12
litellm/router_utils/auto_router_tuning_baseline.py
View file

@ -52,7 +52,17 @@ def tuning_fingerprint(complexity_router_config: object) -> str | None:
supplied: Final = ((_TUNING_FIELD_SET - frozenset(("tier_model_configs",))) & frozenset(raw)) | (
frozenset(("tier_model_configs",)) if validated.tier_model_configs else frozenset()
)
payload: Final = validated.model_dump(mode="json", include=supplied)
payload: Final = validated.model_dump(
mode="json",
include=supplied,
exclude={
"custom_dimensions": {
index: {"scoring_mode"}
for index, dimension in enumerate(validated.custom_dimensions)
if dimension.scoring_mode == "binary"
}
},
)
return hashlib.sha256(json.dumps(payload, sort_keys=True, separators=(",", ":")).encode()).hexdigest()

33
litellm/router_utils/cooldown_handlers.py
View file

@ -14,6 +14,7 @@ from typing import TYPE_CHECKING, Any, Final
import litellm
from litellm._logging import verbose_router_logger
from litellm.caching.dual_cache import DualCache
from litellm.constants import (
DEFAULT_COOLDOWN_TIME_SECONDS,
DEFAULT_FAILURE_THRESHOLD_MINIMUM_REQUESTS,
@ -558,9 +559,12 @@ def should_cooldown_based_on_allowed_fails_policy(
When *allowed_fails_override* / *cooldown_time_override* are supplied they
take precedence over the router-level values (used by deployment-level overrides).
The counter lives in the router's shared ``DualCache`` (Redis when configured), so
every worker process increments the same key and the threshold applies fleet-wide.
When *cache_key_suffix* is supplied the fail counter is keyed as
``{deployment}:{cache_key_suffix}`` so that different exception types are
tracked independently per deployment.
``deployment:{deployment}:allowed_fails:{cache_key_suffix}`` so that different
exception types are tracked independently per deployment.
Returns:
- True if fails exceed the allowed limit (should cooldown)
@ -584,16 +588,25 @@ def should_cooldown_based_on_allowed_fails_policy(
else (litellm_router_instance.cooldown_time or DEFAULT_COOLDOWN_TIME_SECONDS)
)
cache_key: Final = f"{deployment}:{cache_key_suffix}" if cache_key_suffix else deployment
current_fails: Final = litellm_router_instance.failed_calls.get_cache(key=cache_key) or 0
updated_fails: Final = current_fails + 1
base_key: Final = f"deployment:{deployment}:allowed_fails"
cache_key: Final = f"{base_key}:{cache_key_suffix}" if cache_key_suffix else base_key
updated_fails: Final = _increment_allowed_fails(
cache=litellm_router_instance.cache, cache_key=cache_key, ttl=cooldown_time
)
return updated_fails > allowed_fails
if updated_fails > allowed_fails:
return True
else:
litellm_router_instance.failed_calls.set_cache(key=cache_key, value=updated_fails, ttl=cooldown_time)
return False
def _increment_allowed_fails(cache: DualCache, cache_key: str, ttl: float) -> int:
"""
Return the fleet-wide fail count. ``DualCache.increment_cache`` bumps the in-memory tier
before Redis and re-raises a Redis error, so a Redis outage degrades to this worker's own count.
"""
try:
return cache.increment_cache(key=cache_key, value=1, ttl=ttl)
except Exception as e: # noqa: BLE001 # a Redis outage must not stop failing deployments from cooling down
verbose_router_logger.warning("allowed_fails counter fell back to this worker's in-memory count: %s", e)
local_fails: Final = cache.get_cache(key=cache_key, local_only=True)
return local_fails if isinstance(local_fails, int) else 0
def _is_allowed_fails_set_on_router(

15
litellm/types/router.py
View file

@ -12,7 +12,9 @@ import httpx
from pydantic import BaseModel, ConfigDict, Field, field_validator, model_validator
from typing_extensions import Protocol, ReadOnly, Required, TypedDict, runtime_checkable
from litellm._logging import verbose_logger
from litellm._uuid import uuid
from litellm.litellm_core_utils.core_helpers import normalize_drop_params
if TYPE_CHECKING:
from litellm.router import Router
@ -314,6 +316,7 @@ class GenericLiteLLMParams(CredentialLiteLLMParams, CustomPricingLiteLLMParams):
timeout: float | str | httpx.Timeout | None = None # if str, pass in as os.environ/
stream_timeout: float | str | None = None # timeout when making stream=True calls, if str, pass in as os.environ/
max_retries: int | None = None
drop_params: bool | str | None = None
organization: str | None = None # for openai orgs
configurable_clientside_auth_params: CONFIGURABLE_CLIENTSIDE_AUTH_PARAMS = None
litellm_credential_name: str | None = None
@ -404,6 +407,18 @@ class GenericLiteLLMParams(CredentialLiteLLMParams, CustomPricingLiteLLMParams):
return filtered
return data
@field_validator("drop_params", mode="before")
@classmethod
def coerce_drop_params(cls, value: object) -> bool | str | None:
normalized: Final = normalize_drop_params(value)
if normalized is not None:
return normalized
if isinstance(value, str):
return value
if value is not None:
verbose_logger.warning("drop_params=%r is not a flag value, treating it as unset", value)
return None
def __contains__(self, key) -> bool:
# Define custom behavior for the 'in' operator
return hasattr(self, key)

1
litellm/types/utils.py
View file

@ -580,6 +580,7 @@ CallTypesLiteral = Literal[
"search",
"asearch",
"_arealtime",
"_aresponses_websocket",
"create_batch",
"acreate_batch",
"create_file",

122
litellm/utils.py
View file

@ -80,6 +80,7 @@ from litellm.constants import (
PROVIDERS_THAT_AUTHENTICATE_ON_PROVIDER_INFO,
TOOL_CHOICE_OBJECT_TOKEN_COUNT,
)
from litellm.litellm_core_utils.core_helpers import normalize_drop_params
from litellm.litellm_core_utils.fallback_generalizations import (
match_capability_generalizations,
)
@ -3239,7 +3240,7 @@ def get_optional_params_transcription(
passed_params.pop("OPENAI_TRANSCRIPTION_PARAMS")
custom_llm_provider = passed_params.pop("custom_llm_provider")
drop_params = passed_params.pop("drop_params")
drop_params = normalize_drop_params(passed_params.pop("drop_params"))
special_params: Final[Mapping[str, object]] = passed_params.pop("kwargs")
for k, v in special_params.items():
passed_params[k] = v
@ -3347,7 +3348,7 @@ def get_optional_params_image_gen(
model = passed_params.pop("model", None)
custom_llm_provider = passed_params.pop("custom_llm_provider")
provider_config = passed_params.pop("provider_config", None)
drop_params = passed_params.pop("drop_params", None)
drop_params = normalize_drop_params(passed_params.pop("drop_params", None))
additional_drop_params = passed_params.pop("additional_drop_params", None)
special_params: Final[Mapping[str, object]] = passed_params.pop("kwargs")
for k, v in special_params.items():
@ -3475,7 +3476,7 @@ def get_optional_params_embeddings(
custom_llm_provider = passed_params.pop("custom_llm_provider", None)
special_params: Final = passed_params.pop("kwargs")
drop_params = passed_params.pop("drop_params", None)
drop_params = normalize_drop_params(passed_params.pop("drop_params", None))
additional_drop_params = passed_params.pop("additional_drop_params", None)
allowed_openai_params = passed_params.pop("allowed_openai_params", None) or []
# Remove function objects from passed_params to avoid JSON serialization errors
@ -4202,6 +4203,7 @@ def get_optional_params(
base_model: str | None = None,
**kwargs,
):
drop_params = normalize_drop_params(drop_params) # rebind-ok: config and DB deployments pass "true" as a string
passed_params: Final = locals().copy()
special_params: Final = passed_params.pop("kwargs")
# Remove base_model from passed_params so it doesn't interfere with
@ -4279,20 +4281,20 @@ def get_optional_params(
model=model,
non_default_params=non_default_params,
optional_params=optional_params,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "anthropic_text":
optional_params = litellm.AnthropicTextConfig().map_openai_params(
model=model,
non_default_params=non_default_params,
optional_params=optional_params,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
optional_params = litellm.AnthropicTextConfig().map_openai_params(
model=model,
non_default_params=non_default_params,
optional_params=optional_params,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "cohere_chat" or custom_llm_provider == "cohere":
@ -4301,14 +4303,14 @@ def get_optional_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "triton":
optional_params = litellm.TritonConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=drop_params if drop_params is not None else False,
drop_params=bool(drop_params),
)
elif custom_llm_provider == "maritalk":
@ -4316,35 +4318,35 @@ def get_optional_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "replicate":
optional_params = litellm.ReplicateConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "predibase":
optional_params = litellm.PredibaseConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "huggingface":
optional_params = litellm.HuggingFaceChatConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "together_ai":
optional_params = litellm.TogetherAIChatConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "vertex_ai" and (
model in litellm.vertex_chat_models
@ -4358,7 +4360,7 @@ def get_optional_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "gemini":
@ -4366,21 +4368,21 @@ def get_optional_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "vertex_ai_beta" or (custom_llm_provider == "vertex_ai" and "gemini" in model):
optional_params = litellm.VertexGeminiConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif litellm.VertexAIAnthropicConfig.is_supported_model(model=model, custom_llm_provider=custom_llm_provider):
optional_params = litellm.VertexAIAnthropicConfig().map_openai_params(
model=model,
non_default_params=non_default_params,
optional_params=optional_params,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "vertex_ai":
if model in litellm.vertex_mistral_models:
@ -4389,35 +4391,35 @@ def get_optional_params(
model=model,
non_default_params=non_default_params,
optional_params=optional_params,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
else:
optional_params = litellm.MistralConfig().map_openai_params(
model=model,
non_default_params=non_default_params,
optional_params=optional_params,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif model in litellm.vertex_ai_ai21_models:
optional_params = litellm.VertexAIAi21Config().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif provider_config is not None:
optional_params = provider_config.map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
else: # use generic openai-like param mapping
optional_params = litellm.VertexAILlama3Config().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "sagemaker":
@ -4426,7 +4428,7 @@ def get_optional_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "bedrock":
BedrockModelInfo: Final = getattr(sys.modules[__name__], "BedrockModelInfo")
@ -4437,14 +4439,14 @@ def get_optional_params(
model=model,
non_default_params=non_default_params,
optional_params=optional_params,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif bedrock_route == "openai":
optional_params = litellm.AmazonBedrockOpenAIConfig().map_openai_params(
model=model,
non_default_params=non_default_params,
optional_params=optional_params,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif "anthropic" in bedrock_base_model and bedrock_route == "invoke":
if bedrock_base_model in litellm.AmazonAnthropicConfig.get_legacy_anthropic_model_names():
@ -4452,21 +4454,21 @@ def get_optional_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
else:
optional_params = litellm.AmazonAnthropicClaudeConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif provider_config is not None:
optional_params = provider_config.map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
if bedrock_route == "claude_platform":
optional_params = BedrockModelInfo.map_claude_platform_auth_params(
@ -4477,28 +4479,28 @@ def get_optional_params(
model=model,
non_default_params=non_default_params,
optional_params=optional_params,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "ollama":
optional_params = litellm.OllamaConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "ollama_chat":
optional_params = litellm.OllamaChatConfig().map_openai_params(
model=model,
non_default_params=non_default_params,
optional_params=optional_params,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "nlp_cloud":
optional_params = litellm.NLPCloudConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "petals":
@ -4506,35 +4508,35 @@ def get_optional_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "deepinfra":
optional_params = litellm.DeepInfraConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "perplexity" and provider_config is not None:
optional_params = provider_config.map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "mistral" or custom_llm_provider == "codestral":
optional_params = litellm.MistralConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "text-completion-codestral":
optional_params = litellm.CodestralTextCompletionConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "text-completion-inception":
@ -4542,7 +4544,7 @@ def get_optional_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "databricks":
@ -4550,21 +4552,21 @@ def get_optional_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "nvidia_nim":
optional_params = litellm.NvidiaNimConfig().map_openai_params(
model=model,
non_default_params=non_default_params,
optional_params=optional_params,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "cerebras":
optional_params = litellm.CerebrasConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "xai":
optional_params = litellm.XAIChatConfig().map_openai_params(
@ -4577,77 +4579,77 @@ def get_optional_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "fireworks_ai":
optional_params = litellm.FireworksAIConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "volcengine":
optional_params = litellm.VolcEngineConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "hosted_vllm":
optional_params = litellm.HostedVLLMChatConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "vllm":
optional_params = litellm.VLLMConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "groq":
optional_params = litellm.GroqChatConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "bedrock_mantle":
optional_params = litellm.BedrockMantleChatConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "deepseek":
optional_params = litellm.DeepSeekChatConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "tencent":
optional_params = litellm.TencentChatConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "openrouter":
optional_params = litellm.OpenrouterConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "watsonx":
optional_params = litellm.IBMWatsonXChatConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
# WatsonX-text param check
for param in passed_params:
@ -4660,21 +4662,21 @@ def get_optional_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "openai":
optional_params = litellm.OpenAIConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "nebius":
optional_params = litellm.NebiusConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif custom_llm_provider == "azure":
_azure_detection_model: Final = base_model or model
@ -4683,14 +4685,14 @@ def get_optional_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=_azure_detection_model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif litellm.AzureOpenAIGPT5Config.is_model_gpt_5_model(model=_azure_detection_model):
optional_params = litellm.AzureOpenAIGPT5Config().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=_azure_detection_model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
else:
verbose_logger.debug(
@ -4709,21 +4711,21 @@ def get_optional_params(
optional_params=optional_params,
model=_azure_detection_model,
api_version=api_version,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
elif provider_config is not None:
optional_params = provider_config.map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
else: # assume passing in params for openai-like api
optional_params = litellm.OpenAILikeChatConfig().map_openai_params(
non_default_params=non_default_params,
optional_params=optional_params,
model=model,
drop_params=(drop_params if drop_params is not None and isinstance(drop_params, bool) else False),
drop_params=bool(drop_params),
)
# if user passed in non-default kwargs for specific providers/models, pass them along
optional_params = add_provider_specific_params_to_optional_params(

7
pyproject.toml
View file

@ -1,6 +1,6 @@
[project]
name = "litellm"
version = "1.101.0"
version = "1.102.0"
description = "Library to easily interface with LLM API providers"
readme = "README.md"
requires-python = ">=3.10, <3.15"
@ -67,7 +67,7 @@ proxy = [
"azure-identity>=1.25.2,<2.0",
"azure-storage-blob>=12.28.0,<13.0",
"mcp>=1.28.1,<2.0",
"litellm-proxy-extras==0.4.94",
"litellm-proxy-extras==0.4.95",
"litellm-enterprise==0.1.65",
"RestrictedPython>=8.5,<9.0",
"rich>=13.9.4,<14.0",
@ -114,7 +114,6 @@ caching = ["diskcache>=5.6.3,<6.0"]
mcp = ["mcp>=1.28.1,<2.0"]
# Driver for the MongoDB Atlas vector store; Atlas Vector Search has no HTTP query API.
# The floor is 4.9 because that is the release AsyncMongoClient landed in.
mongodb = ["pymongo>=4.9,<5.0"]
# SAML SSO for the admin UI. python3-saml pulls in xmlsec/lxml, whose wheels
# bundle the native libxmlsec1/libxml2 libraries, so no system packages are
# required. Kept out of the base `proxy` extra so it stays optional.
@ -328,7 +327,7 @@ members = ["enterprise", "litellm-proxy-extras"]
profile = "black"
[tool.commitizen]
version = "1.101.0"
version = "1.102.0"
version_files = [
"pyproject.toml:^version",
]

1
schema.prisma
View file

@ -1036,6 +1036,7 @@ model LiteLLM_ManagedObjectTable { // for batches or finetuning jobs which use t
created_at DateTime @default(now())
created_by String?
team_id String?
org_id String? // creating key's organization at submission time; CheckBatchCost bills org spend against it
api_key String?
request_tags Json? @default("[]")
updated_at DateTime @updatedAt

16
terraform/litellm/aws/README.md
View file

@ -242,6 +242,22 @@ this with `litellm_license`. To tune the export cadence, set
`LITELLM_BILLING_METRICS_EXPORT_INTERVAL_MS` through `gateway_extra_env` /
`backend_extra_env`
### Prometheus metrics sidecar
`gateway_metrics_port` adds a `metrics` sidecar
(`python -m litellm.proxy.prometheus_metrics_server`) to the gateway task that
aggregates the workers' samples over a shared task volume, so a scrape never
runs on an inference worker. The ALB never routes to that port and the tasks
security group only opens it to `gateway_metrics_scrape_cidrs`. Needs
`gateway_image` v1.101.0 or newer. See
[Prometheus metrics](https://docs.litellm.ai/docs/proxy/prometheus) for the
metrics themselves.
```hcl
gateway_metrics_port = 4001
gateway_metrics_scrape_cidrs = ["10.0.0.0/16"]
```
## Tenant deployment
Every resource the stack creates is named `${tenant}-litellm-${env}` (or

Some files were not shown because too many files have changed in this diff Show more