scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-10-10 03:28:53 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 85

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

Browse source
This commit is contained in:
mateo-berri 2026-08-14 17:44:39 -07:00
commit 2c6409c7e6
599 changed files with 43777 additions and 17721 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

84
.circleci/config.yml
View file

@ -2744,84 +2744,6 @@ jobs:
file: ./coverage.xml
flags: circleci
ui_build:
docker:
- image: cimg/node:24.19@sha256:8966565f07189a67d64d6808a2b127f31dafae566508e3547f55640e1070bfad
auth:
username: ${DOCKERHUB_USERNAME}
password: ${DOCKERHUB_PASSWORD}
resource_class: medium+
working_directory: ~/project
steps:
- checkout
- skip_if_unrelated_changes:
category: client
- setup_google_dns
- restore_cache:
keys:
- ui-build-deps-v1-{{ checksum "ui/litellm-dashboard/package-lock.json" }}
- ui-build-deps-v1-
- restore_cache:
keys:
- ui-nextjs-cache-v1-{{ checksum "ui/litellm-dashboard/package-lock.json" }}
- ui-nextjs-cache-v1-
- run:
name: Install dependencies
command: |
cd ui/litellm-dashboard
npm ci
- save_cache:
key: ui-build-deps-v1-{{ checksum "ui/litellm-dashboard/package-lock.json" }}
paths:
- ui/litellm-dashboard/node_modules
- run:
name: Build UI
command: |
cd ui/litellm-dashboard
source ./build_ui.sh
- save_cache:
key: ui-nextjs-cache-v1-{{ checksum "ui/litellm-dashboard/package-lock.json" }}
paths:
- ui/litellm-dashboard/.next/cache
- persist_to_workspace:
root: .
paths:
- litellm/proxy/_experimental/out
ui_unit_tests:
docker:
- image: cimg/node:24.19@sha256:8966565f07189a67d64d6808a2b127f31dafae566508e3547f55640e1070bfad
auth:
username: ${DOCKERHUB_USERNAME}
password: ${DOCKERHUB_PASSWORD}
resource_class: xlarge
working_directory: ~/project
steps:
- checkout
- skip_if_unrelated_changes:
category: client
- setup_google_dns
- restore_cache:
keys:
- ui-unit-deps-v1-{{ checksum "ui/litellm-dashboard/package-lock.json" }}
- ui-unit-deps-v1-
- run:
name: Install dependencies
command: |
cd ui/litellm-dashboard
npm ci
- save_cache:
key: ui-unit-deps-v1-{{ checksum "ui/litellm-dashboard/package-lock.json" }}
paths:
- ui/litellm-dashboard/node_modules
- run:
name: Run UI unit tests (Vitest)
command: |
cd ui/litellm-dashboard
CI=true npm run test -- --run \
--pool forks --poolOptions.forks.maxForks=6
e2e_ui_testing:
docker:
- image: cimg/python:3.12-browsers@sha256:b432899af01c9a311bf74f4f22e9ada2e5306d4b1b4383f8d29e1228a5844ef2
@ -3181,12 +3103,6 @@ workflows:
filters: *main_branches
- litellm_router_unit_testing:
filters: *main_branches
- ui_build:
filters: *main_branches
- ui_unit_tests:
requires:
- ui_build
filters: *main_branches
- auth_ui_unit_tests:
filters: *main_branches
- proxy_behavior_tests:

106
.github/workflows/test-unit-proxy-legacy.yml vendored
View file

@ -1,106 +0,0 @@
name: "Unit Tests: Proxy Legacy Tests"
on:
pull_request:
branches:
- main
- litellm_internal_staging
- litellm_oss_staging
- "litellm_**"
push:
branches:
- main
- litellm_internal_staging
permissions:
contents: read
concurrency:
group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}
cancel-in-progress: ${{ github.event_name == 'pull_request' }}
jobs:
test:
runs-on: ubuntu-latest
timeout-minutes: 20
strategy:
fail-fast: false
matrix:
test-group:
- name: "auth-and-jwt"
path: "tests/proxy_unit_tests/test_[a-j]*.py"
- name: "key-generation"
path: "tests/proxy_unit_tests/test_[k-o]*.py"
- name: "proxy-config"
path: "tests/proxy_unit_tests/test_prisma*.py tests/proxy_unit_tests/test_prompt*.py tests/proxy_unit_tests/test_proxy_[c-r]*.py"
- name: "proxy-server"
path: "tests/proxy_unit_tests/test_proxy_server.py"
- name: "proxy-server-extras"
path: "tests/proxy_unit_tests/test_proxy_server_*.py tests/proxy_unit_tests/test_proxy_setting_guardrails.py"
- name: "proxy-utils"
path: "tests/proxy_unit_tests/test_proxy_utils.py"
- name: "proxy-token-counter"
path: "tests/proxy_unit_tests/test_proxy_token_counter.py"
- name: "proxy-response-and-misc"
path: "tests/proxy_unit_tests/test_[r-t]*.py"
- name: "proxy-user-auth-and-spend"
path: "tests/proxy_unit_tests/test_[u-z]*.py"
name: ${{ matrix.test-group.name }}
steps:
- uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
with:
persist-credentials: false
- name: Detect backend-relevant changes
id: changes
uses: ./.github/actions/detect-backend-changes
- name: Set up Python
uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
with:
python-version: "3.12"
- name: Set up uv
uses: ./.github/actions/setup-uv-with-retries
with:
version: "0.10.9"
- name: Cache uv dependencies
uses: actions/cache@0057852bfaa89a56745cba8c7296529d2fc39830 # v4.3.0
with:
path: |
~/.cache/uv
.venv
key: ${{ runner.os }}-uv-${{ hashFiles('uv.lock') }}
restore-keys: |
${{ runner.os }}-uv-
- name: Install dependencies
if: steps.changes.outputs.decision != 'skip'
run: |
.github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra google --extra proxy --extra semantic-router
- name: Cache Prisma binaries
if: steps.changes.outputs.decision != 'skip'
uses: ./.github/actions/cache-prisma-binaries
- name: Generate Prisma client
if: steps.changes.outputs.decision != 'skip'
run: |
uv run --no-sync prisma generate --schema litellm/proxy/schema.prisma
- name: Run tests - ${{ matrix.test-group.name }}
if: steps.changes.outputs.decision != 'skip'
env:
TEST_PATH: ${{ matrix.test-group.path }}
run: |
uv run --no-sync pytest ${TEST_PATH} \
--tb=short -vv \
--maxfail=10 \
-n 2 \
--reruns 1 \
--reruns-delay 1 \
--dist=loadscope \
--durations=20

2
CLAUDE.md
View file

@ -29,7 +29,7 @@ End-to-end tests belong in `tests/e2e/` and must follow the harness conventions
When creating PRs, don't set base to `main`. `litellm_internal_staging` is the default base branch and serves that purpose for both internal and external / OSS contributions
When writing a PR body, treat the comments and imperative instructions inside @.github/pull_request_template.md as rules to follow, not just layout. Agent harnesses may strip HTML comments from copies of that file injected into context, so read .github/pull_request_template.md from disk before writing a PR body to make sure you see every comment rule
When writing a PR body, treat the comments and imperative instructions inside .github/pull_request_template.md as rules to follow, not just layout. Agent harnesses may strip HTML comments from copies of that file injected into context, so read .github/pull_request_template.md from disk before writing a PR body to make sure you see every comment rule
Same applies for filing bug reports and feature requests, with .github/ISSUE_TEMPLATE/bug_report.yml and .github/ISSUE_TEMPLATE/feature_request.yml, respectively

2
backend/routes/allowlist.py
View file

@ -146,11 +146,13 @@ BACKEND_EXACT_PATHS: frozenset[str] = frozenset(
"/docs/oauth2-redirect",
"/redoc",
"/fallback/login",
"/mcp", # bare spelling of the aggregate MCP endpoint; /mcp/ prefix covers the rest
}
)
BACKEND_MOUNT_PATHS: frozenset[str] = frozenset(
{
"/swagger", # API documentation static assets belong to the backend
"/mcp", # lazily-mounted MCP sub-app serves on the backend component
}
)

20
basedpyright-code-budget.json
View file

@ -1,9 +1,9 @@
{
"reportAny": {
"limit": 23914
"limit": 22947
},
"reportArgumentType": {
"limit": 2580
"limit": 2579
},
"reportAssignmentType": {
"limit": 323
@ -24,7 +24,7 @@
"limit": 19
},
"reportExplicitAny": {
"limit": 7573
"limit": 7312
},
"reportFunctionMemberAccess": {
"limit": 7
@ -54,10 +54,10 @@
"limit": 0
},
"reportMissingParameterType": {
"limit": 5719
"limit": 5707
},
"reportMissingTypeArgument": {
"limit": 15657
"limit": 15642
},
"reportMissingTypeStubs": {
"limit": 40
@ -99,22 +99,22 @@
"limit": 0
},
"reportUnknownArgumentType": {
"limit": 44832
"limit": 44776
},
"reportUnknownLambdaType": {
"limit": 113
},
"reportUnknownMemberType": {
"limit": 39269
"limit": 39237
},
"reportUnknownParameterType": {
"limit": 19988
"limit": 19969
},
"reportUnknownVariableType": {
"limit": 30923
"limit": 30881
},
"reportUnnecessaryCast": {
"limit": 118
"limit": 117
},
"reportUnnecessaryComparison": {
"limit": 699

103
enterprise/litellm_enterprise/proxy/common_utils/check_batch_cost.py
View file

@ -3,7 +3,7 @@ Polls LiteLLM_ManagedObjectTable to check if the batch job is complete, and if t
"""
from datetime import datetime, timedelta, timezone
from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple
from typing import TYPE_CHECKING, Any, Dict, Final, List, Optional, Tuple
from litellm._logging import verbose_proxy_logger
from litellm._uuid import uuid
@ -23,6 +23,15 @@ if TYPE_CHECKING:
CHECK_BATCH_COST_USER_AGENT = "LiteLLM Proxy/CheckBatchCost"
TERMINAL_MANAGED_OBJECT_STATUSES: Final[Tuple[str, ...]] = (
"completed",
"complete",
"failed",
"expired",
"cancelled",
"stale_expired",
)
class CheckBatchCost:
def __init__(
@ -132,11 +141,11 @@ class CheckBatchCost:
in non-terminal states as 'stale_expired'. These will never complete and
should not be polled.
"""
cutoff = datetime.now(timezone.utc) - timedelta(days=MANAGED_OBJECT_STALENESS_CUTOFF_DAYS)
result = await self.prisma_client.db.litellm_managedobjecttable.update_many(
cutoff: Final = datetime.now(timezone.utc) - timedelta(days=MANAGED_OBJECT_STALENESS_CUTOFF_DAYS)
result: Final = await self.prisma_client.db.litellm_managedobjecttable.update_many(
where={
"file_purpose": "batch",
"status": {"not_in": ["completed", "complete", "failed", "expired", "cancelled", "stale_expired"]},
"status": {"not_in": list(TERMINAL_MANAGED_OBJECT_STATUSES)},
"created_at": {"lt": cutoff},
},
data={"status": "stale_expired"},
@ -147,6 +156,26 @@ class CheckBatchCost:
f"(older than {MANAGED_OBJECT_STALENESS_CUTOFF_DAYS} days) as stale_expired"
)
if not self._has_batch_processed_column:
return
# A row already in a terminal status is never rewritten by the sweep above, so
# without this it keeps a poll-page slot forever and starves newer batches.
retired: Final = await self.prisma_client.db.litellm_managedobjecttable.update_many(
where={
"file_purpose": "batch",
"batch_processed": False,
"status": {"in": ["complete", "completed"]},
"created_at": {"lt": cutoff},
},
data={"batch_processed": True},
)
if retired > 0:
verbose_proxy_logger.warning(
f"CheckBatchCost: gave up on {retired} completed managed objects older than "
f"{MANAGED_OBJECT_STALENESS_CUTOFF_DAYS} days that were never costed"
)
async def _fallback_find_jobs(self) -> list:
"""Query batch jobs without the batch_processed filter (for older schemas)."""
return await self.prisma_client.db.litellm_managedobjecttable.find_many(
@ -167,6 +196,68 @@ class CheckBatchCost:
order={"created_at": "asc"},
)
async def _retire_job(self, job: "LiteLLM_ManagedObjectTable", reason: str) -> None:
"""
Take a row that can never be costed out of the poll page. Leaving it selectable
would burn one of the MAX_OBJECTS_PER_POLL_CYCLE slots on every future cycle, and
once enough such rows accumulate no newer batch is ever reached. Older schemas
without batch_processed can only be excluded through the status filter.
"""
data: Final = (
{"batch_processed": True}
if self._has_batch_processed_column
else {"status": "stale_expired"}
)
try:
await self.prisma_client.db.litellm_managedobjecttable.update(
where={"id": job.id},
data=data,
)
except Exception as db_err:
verbose_proxy_logger.error(
f"CheckBatchCost: failed to retire uncostable job {job.id} ({reason}): {db_err}"
)
return
verbose_proxy_logger.warning(
f"CheckBatchCost: job {job.id} can never be costed ({reason}), "
"so it will no longer be polled"
)
@staticmethod
def _has_unified_id_without_model(job: "LiteLLM_ManagedObjectTable") -> bool:
"""A unified id that decodes but carries no model_id can never be routed."""
from litellm.proxy.openai_files_endpoints.common_utils import (
convert_b64_uid_to_unified_uid,
get_model_id_from_unified_batch_id,
)
decoded: Final = convert_b64_uid_to_unified_uid(job.unified_object_id)
return (
decoded != job.unified_object_id
and get_model_id_from_unified_batch_id(decoded) is None
)
@staticmethod
def _is_batch_gone_at_provider(error: Exception, batch_id: str) -> bool:
"""
A 404 naming the batch means the provider dropped its record of it, so no later
retrieve can ever succeed. A 404 about anything else, a renamed Azure deployment
or a fallback deployment that never saw this batch, is still fixable in config, so
it keeps retrying.
"""
import openai
from litellm.exceptions import NotFoundError
return isinstance(error, (NotFoundError, openai.NotFoundError)) and batch_id in str(error)
def _batch_deployment_exists(self, model_id: str) -> bool:
"""A 404 only proves the batch is gone when it came from the batch's own
deployment. Once that deployment leaves the router, default fallbacks can
silently send the retrieve to a provider that never saw the batch, so its
404 must not retire the row; the staleness sweep bounds it instead."""
return self.llm_router.get_deployment(model_id=model_id) is not None
@staticmethod
def _record_error(
prom_logger: Optional["PrometheusLogger"], error_type: str
@ -645,6 +736,8 @@ class CheckBatchCost:
for job in jobs:
routing = self._resolve_job_routing(job, prom_logger)
if routing is None:
if self._has_unified_id_without_model(job):
await self._retire_job(job, "unified object id has no model id")
continue
model_id, batch_id = routing
@ -667,6 +760,8 @@ class CheckBatchCost:
)
if prom_logger:
prom_logger.record_check_batch_cost_error("provider_retrieval_error")
if self._is_batch_gone_at_provider(e, batch_id) and self._batch_deployment_exists(model_id):
await self._retire_job(job, f"batch {batch_id} no longer exists at the provider")
continue
## RETRIEVE THE BATCH JOB OUTPUT FILE

4
helm/litellm-helm/templates/migrations-job.yaml
View file

@ -105,6 +105,10 @@ spec:
{{- toYaml . | nindent 8 }}
{{- end }}
restartPolicy: OnFailure
{{- with .Values.nodeSelector }}
nodeSelector:
{{- toYaml . | nindent 8 }}
{{- end }}
{{- with .Values.affinity }}
affinity:
{{- toYaml . | nindent 8 }}

24
helm/litellm-helm/tests/migrations-job_tests.yaml
View file

@ -290,3 +290,27 @@ tests:
value:
allowPrivilegeEscalation: false
readOnlyRootFilesystem: true
- it: should schedule onto the same nodes as the gateway
template: migrations-job.yaml
set:
migrationJob:
enabled: true
nodeSelector:
karpenter.sh/nodepool: litellm-e2e
tolerations:
- key: workload
operator: Equal
value: litellm-e2e
effect: NoSchedule
asserts:
- equal:
path: spec.template.spec.nodeSelector
value:
karpenter.sh/nodepool: litellm-e2e
- equal:
path: spec.template.spec.tolerations
value:
- key: workload
operator: Equal
value: litellm-e2e
effect: NoSchedule

4
helm/litellm/templates/backend/deployment.yaml
View file

@ -81,6 +81,10 @@ spec:
readinessProbe:
{{- toYaml . | nindent 12 }}
{{- end }}
{{- with .Values.backend.startupProbe }}
startupProbe:
{{- toYaml . | nindent 12 }}
{{- end }}
{{- with .Values.backend.lifecycle }}
lifecycle:
{{- toYaml . | nindent 12 }}

4
helm/litellm/templates/backend/hpa.yaml
View file

@ -30,4 +30,8 @@ spec:
type: Utilization
averageUtilization: {{ .Values.backend.hpa.targetMemoryUtilizationPercentage }}
{{- end }}
{{- with .Values.backend.hpa.behavior }}
behavior:
{{- toYaml . | nindent 4 }}
{{- end }}
{{- end }}

4
helm/litellm/templates/gateway/deployment.yaml
View file

@ -83,6 +83,10 @@ spec:
readinessProbe:
{{- toYaml . | nindent 12 }}
{{- end }}
{{- with .Values.gateway.startupProbe }}
startupProbe:
{{- toYaml . | nindent 12 }}
{{- end }}
{{- with .Values.gateway.lifecycle }}
lifecycle:
{{- toYaml . | nindent 12 }}

4
helm/litellm/templates/gateway/hpa.yaml
View file

@ -30,4 +30,8 @@ spec:
type: Utilization
averageUtilization: {{ .Values.gateway.hpa.targetMemoryUtilizationPercentage }}
{{- end }}
{{- with .Values.gateway.hpa.behavior }}
behavior:
{{- toYaml . | nindent 4 }}
{{- end }}
{{- end }}

4
helm/litellm/templates/ui/deployment.yaml
View file

@ -69,6 +69,10 @@ spec:
readinessProbe:
{{- toYaml . | nindent 12 }}
{{- end }}
{{- with .Values.ui.startupProbe }}
startupProbe:
{{- toYaml . | nindent 12 }}
{{- end }}
{{- with .Values.ui.lifecycle }}
lifecycle:
{{- toYaml . | nindent 12 }}

4
helm/litellm/templates/ui/hpa.yaml
View file

@ -30,4 +30,8 @@ spec:
type: Utilization
averageUtilization: {{ .Values.ui.hpa.targetMemoryUtilizationPercentage }}
{{- end }}
{{- with .Values.ui.hpa.behavior }}
behavior:
{{- toYaml . | nindent 4 }}
{{- end }}
{{- end }}

58
helm/litellm/tests/hpa_behavior_tests.yaml Normal file
View file

@ -0,0 +1,58 @@
suite: test HPA scaling behavior passthrough
templates:
- gateway/hpa.yaml
- backend/hpa.yaml
- ui/hpa.yaml
values:
- ./values/required.yaml
tests:
- it: HPA omits spec.behavior by default, so Kubernetes' default scaling applies
templates:
- gateway/hpa.yaml
- backend/hpa.yaml
asserts:
- isKind:
of: HorizontalPodAutoscaler
- notExists:
path: spec.behavior
- it: gateway HPA renders spec.behavior verbatim when configured
template: gateway/hpa.yaml
set:
gateway.hpa.behavior:
scaleDown:
stabilizationWindowSeconds: 300
policies:
- { type: Percent, value: 50, periodSeconds: 60 }
scaleUp:
stabilizationWindowSeconds: 0
selectPolicy: Max
policies:
- { type: Percent, value: 100, periodSeconds: 30 }
- { type: Pods, value: 2, periodSeconds: 30 }
asserts:
- equal:
path: spec.behavior
value:
scaleDown:
stabilizationWindowSeconds: 300
policies:
- { type: Percent, value: 50, periodSeconds: 60 }
scaleUp:
stabilizationWindowSeconds: 0
selectPolicy: Max
policies:
- { type: Percent, value: 100, periodSeconds: 30 }
- { type: Pods, value: 2, periodSeconds: 30 }
- it: behavior passthrough works on every autoscaled component (ui parity)
template: ui/hpa.yaml
set:
ui.hpa.enabled: true
ui.hpa.behavior:
scaleUp:
stabilizationWindowSeconds: 0
asserts:
- equal:
path: spec.behavior.scaleUp.stabilizationWindowSeconds
value: 0

27
helm/litellm/tests/probe_tests.yaml
View file

@ -104,3 +104,30 @@ tests:
periodSeconds: 15
timeoutSeconds: 4
failureThreshold: 3
- it: no startupProbe by default, so existing installs are unchanged
templates:
- gateway/deployment.yaml
- backend/deployment.yaml
asserts:
- notExists:
path: spec.template.spec.containers[0].startupProbe
- it: startupProbe renders verbatim when configured, gating a slow cold start
template: gateway/deployment.yaml
set:
gateway.startupProbe:
httpGet: { path: /health/readiness, port: http }
failureThreshold: 30
periodSeconds: 10
timeoutSeconds: 5
asserts:
- equal:
path: spec.template.spec.containers[0].startupProbe
value:
httpGet:
path: /health/readiness
port: http
failureThreshold: 30
periodSeconds: 10
timeoutSeconds: 5

24
helm/litellm/values.yaml
View file

@ -223,12 +223,28 @@ gateway:
initialDelaySeconds: 5
periodSeconds: 10
timeoutSeconds: 10
# Optional startupProbe. Empty by default, so existing installs are unchanged
# and liveness/readiness apply from container start. Set it to gate
# liveness/readiness until a slow cold start finishes — a high failureThreshold
# tolerates long first-boot times without a liveness-kill loop, e.g.:
# httpGet: { path: /health/readiness, port: http }
# failureThreshold: 30
# periodSeconds: 10
startupProbe: {}
hpa:
enabled: true
minReplicas: 1
maxReplicas: 10
targetCPUUtilizationPercentage: 70
targetMemoryUtilizationPercentage: 80
# Optional autoscaling/v2 scaling behavior (scaleUp / scaleDown policies and
# stabilization windows). Empty by default -> Kubernetes' default behavior.
# Rendered verbatim under spec.behavior, e.g.:
# scaleUp:
# stabilizationWindowSeconds: 0
# policies:
# - { type: Percent, value: 100, periodSeconds: 30 }
behavior: {}
# PodDisruptionBudget for the gateway pods. Set exactly one of
# `minAvailable` / `maxUnavailable` (minAvailable wins if both are set;
# enabling without either falls back to `maxUnavailable: 1`). Disabled by
@ -319,11 +335,15 @@ backend:
initialDelaySeconds: 5
periodSeconds: 10
timeoutSeconds: 10
# Optional startupProbe; same shape as gateway.startupProbe. Empty by default.
startupProbe: {}
hpa:
enabled: true
minReplicas: 1
maxReplicas: 4
targetCPUUtilizationPercentage: 70
# Optional autoscaling/v2 scaling behavior; same shape as gateway.hpa.behavior.
behavior: {}
# Same shape as gateway.pdb.
pdb:
enabled: false
@ -379,11 +399,15 @@ ui:
httpGet: { path: /, port: http }
initialDelaySeconds: 2
periodSeconds: 10
# Optional startupProbe; same shape as gateway.startupProbe. Empty by default.
startupProbe: {}
hpa:
enabled: false
minReplicas: 1
maxReplicas: 3
targetCPUUtilizationPercentage: 80
# Optional autoscaling/v2 scaling behavior; same shape as gateway.hpa.behavior.
behavior: {}
# Same shape as gateway.pdb.
pdb:
enabled: false

49
litellm-proxy-extras/litellm_proxy_extras/migrations/20260811172448_add_shadow_eval/migration.sql Normal file
View file

@ -0,0 +1,49 @@
-- CreateTable
CREATE TABLE "LiteLLM_ShadowEvalJob" (
"id" TEXT NOT NULL,
"api_key_id" TEXT NOT NULL,
"router_name" TEXT NOT NULL,
"judge_model" TEXT NOT NULL,
"shadow_percentage" DOUBLE PRECISION NOT NULL,
"max_turns" INTEGER NOT NULL,
"created_at" TIMESTAMP(3) NOT NULL DEFAULT CURRENT_TIMESTAMP,
"created_by" TEXT,
"ends_at" TIMESTAMP(3) NOT NULL,
"stopped_at" TIMESTAMP(3),
CONSTRAINT "LiteLLM_ShadowEvalJob_pkey" PRIMARY KEY ("id")
);
-- CreateTable
CREATE TABLE "LiteLLM_ShadowEvalAttempt" (
"id" TEXT NOT NULL,
"job_id" TEXT NOT NULL,
"request_id" TEXT NOT NULL,
"outcome" TEXT NOT NULL,
"tier" TEXT,
"real_model" TEXT,
"shadow_model" TEXT,
"confidence" DOUBLE PRECISION,
"judge_cost" DOUBLE PRECISION NOT NULL DEFAULT 0,
"error" TEXT,
"created_at" TIMESTAMP(3) NOT NULL DEFAULT CURRENT_TIMESTAMP,
CONSTRAINT "LiteLLM_ShadowEvalAttempt_pkey" PRIMARY KEY ("id")
);
-- CreateIndex
CREATE INDEX "LiteLLM_ShadowEvalJob_api_key_id_idx" ON "LiteLLM_ShadowEvalJob"("api_key_id");
-- CreateIndex
CREATE INDEX "LiteLLM_ShadowEvalJob_created_at_idx" ON "LiteLLM_ShadowEvalJob"("created_at");
-- CreateIndex
CREATE INDEX "LiteLLM_ShadowEvalAttempt_job_id_idx" ON "LiteLLM_ShadowEvalAttempt"("job_id");
-- One active job per key, enforced by the database rather than a read-then-create in the
-- start endpoint, which races against a concurrent start on another pod. Partial indexes
-- are not expressible in schema.prisma, so this lives here only. Active means not yet
-- stopped; the start endpoint stamps stopped_at on expired jobs before creating.
CREATE UNIQUE INDEX "LiteLLM_ShadowEvalJob_one_active_per_key"
ON "LiteLLM_ShadowEvalJob"("api_key_id") WHERE "stopped_at" IS NULL;

8
litellm-proxy-extras/litellm_proxy_extras/migrations/20260813180408_add_shadow_eval_direction/migration.sql Normal file
View file

@ -0,0 +1,8 @@
-- AlterTable
ALTER TABLE "LiteLLM_ShadowEvalJob" ADD COLUMN "baseline_model" TEXT,
ADD COLUMN "direction" TEXT NOT NULL DEFAULT 'forward';
DROP INDEX IF EXISTS "LiteLLM_ShadowEvalJob_one_active_per_key";
CREATE UNIQUE INDEX IF NOT EXISTS "LiteLLM_ShadowEvalJob_one_active_per_key_direction"
ON "LiteLLM_ShadowEvalJob"("api_key_id", "direction") WHERE "stopped_at" IS NULL;

43
litellm-proxy-extras/litellm_proxy_extras/schema.prisma
View file

@ -1450,6 +1450,49 @@ model LiteLLM_AutoRouterSession {
@@index([last_turn_at], map: "idx_autorouter_session_last_turn")
}
// Shadow eval: evaluation of an auto-router against a key's live traffic, in either
// direction. forward duplicates the requests the key did not route through the router
// through it, answering whether the key should adopt it; reverse duplicates the requests
// the router did serve against a fixed baseline model, answering whether a key already on
// it still benefits. Either way a sampled slice runs in a detached task and an LLM judge
// compares real vs shadow responses blind. The job row is immutable config plus
// stopped_at; every count, status, and spend figure is derived from the append-only
// attempt rows, so nothing can disagree across pods or stop races.
model LiteLLM_ShadowEvalJob {
id String @id @default(cuid())
api_key_id String // hashed virtual key whose traffic is shadowed
router_name String // the auto-router under evaluation, in either direction
direction String @default("forward") // forward | reverse
baseline_model String? // reverse only: the fixed model the router is judged against
judge_model String
shadow_percentage Float
max_turns Int // sample budget: judge at most this many turns
created_at DateTime @default(now())
created_by String?
ends_at DateTime
stopped_at DateTime?
@@index([api_key_id])
@@index([created_at])
}
// One row per sampled pipeline: a blind verdict (real | shadow | tie) or an error.
model LiteLLM_ShadowEvalAttempt {
id String @id @default(cuid())
job_id String
request_id String // the judged real request
outcome String // real | shadow | tie | error
tier String? // router's tier for the prompt, when classified
real_model String?
shadow_model String?
confidence Float?
judge_cost Float @default(0)
error String?
created_at DateTime @default(now())
@@index([job_id])
}
// ---------------------------------------------------------------------------
// Workflow Run Tracking
//

1
litellm/__init__.py
View file

@ -172,6 +172,7 @@ callbacks: List[
callback_settings: Dict[str, Dict[str, Any]] = {}
initialized_langfuse_clients: int = 0
langfuse_default_tags: Optional[List[str]] = None
langfuse_enable_update_trace_keys: bool = False
langsmith_batch_size: Optional[int] = None
prometheus_initialize_budget_metrics: Optional[bool] = False
prometheus_latency_buckets: Optional[List[float]] = None

23
litellm/a2a_protocol/litellm_completion_bridge/handler.py
View file

@ -10,7 +10,7 @@ A2A Streaming Events (in order):
4. Status update (kind: "status-update") - Final status "completed" with final=true
"""
from collections.abc import AsyncIterator, Mapping
from collections.abc import AsyncIterator, Callable, Coroutine, Mapping
from typing import Any, Final
import litellm
@ -54,7 +54,7 @@ class A2ACompletionBridgeHandler:
agent_extra_headers: Mapping[str, str] | None,
*,
stream: bool,
) -> Mapping[str, Any]:
) -> Mapping[str, object]:
# Extract message from params
message: Final = params.get("message", {})
@ -63,7 +63,7 @@ class A2ACompletionBridgeHandler:
# Get completion params
custom_llm_provider: Final = litellm_params.get("custom_llm_provider")
model: Final = litellm_params.get("model", "agent")
model: Final[str] = litellm_params.get("model", "agent")
# Build full model string if provider specified
# Skip prepending if model already starts with the provider prefix
@ -109,13 +109,16 @@ class A2ACompletionBridgeHandler:
return completion_params
@staticmethod
async def _acompletion(completion_params: Mapping[str, Any]) -> ModelResponse | CustomStreamWrapper:
return await litellm.acompletion(**completion_params)
async def _acompletion(completion_params: Mapping[str, object]) -> ModelResponse | CustomStreamWrapper:
acompletion_fn: Final[Callable[..., Coroutine[object, object, ModelResponse | CustomStreamWrapper]]] = vars(
litellm
)["acompletion"]
return await acompletion_fn(**completion_params)
@staticmethod
async def handle_non_streaming(
request_id: str,
params: dict[str, Any],
params: dict[str, object],
litellm_params: dict[str, Any],
api_base: str | None = None,
agent_extra_headers: dict[str, str] | None = None,
@ -296,8 +299,8 @@ class A2ACompletionBridgeHandler:
# Convenience functions that delegate to the class methods
async def handle_a2a_completion(
request_id: str,
params: dict[str, Any],
litellm_params: dict[str, Any],
params: dict[str, object],
litellm_params: dict[str, object],
api_base: str | None = None,
agent_extra_headers: dict[str, str] | None = None,
) -> dict[str, object]:
@ -313,8 +316,8 @@ async def handle_a2a_completion(
async def handle_a2a_completion_streaming(
request_id: str,
params: dict[str, Any],
litellm_params: dict[str, Any],
params: dict[str, object],
litellm_params: dict[str, object],
api_base: str | None = None,
agent_extra_headers: dict[str, str] | None = None,
) -> AsyncIterator[dict[str, object]]:

66
litellm/a2a_protocol/main.py
View file

@ -12,7 +12,8 @@ Provides standalone functions with @client decorator for LiteLLM logging integra
import asyncio
import datetime
import uuid
from collections.abc import AsyncIterator, Coroutine
from collections.abc import AsyncIterator, Coroutine, Mapping
from types import ModuleType
from typing import TYPE_CHECKING, Any, Final, Optional, cast
import litellm
@ -38,12 +39,15 @@ if TYPE_CHECKING:
SendMessageResponse,
SendStreamingMessageRequest,
SendStreamingMessageResponse,
SendStreamingMessageSuccessResponse,
Task,
)
from a2a.types.a2a_pb2 import SendMessageRequest as CoreSendMessageRequest
from a2a.types.a2a_pb2 import StreamResponse as CoreStreamResponse
# Runtime imports — requires a2a-sdk>=1.1.0
A2A_SDK_AVAILABLE = False
_a2a_conversions: Any = None
_a2a_conversions: ModuleType | None = None
try:
from a2a.client import Client, ClientCallContext, ClientConfig, create_client
@ -128,7 +132,7 @@ _A2A_COST_PARAM_KEYS: Final = ("cost_per_query", "input_cost_per_token", "output
def _set_litellm_params_on_logging_obj(
kwargs: dict[str, Any],
litellm_params: dict[str, Any],
litellm_params: Mapping[str, object],
) -> None:
"""
Merge the agent's pricing params into model_call_details["litellm_params"]
@ -150,7 +154,7 @@ def _set_litellm_params_on_logging_obj(
logging_obj.model_call_details["litellm_params"] = {**existing, **cost_params}
def _get_a2a_model_info(a2a_client: Any, kwargs: dict[str, Any]) -> str:
def _get_a2a_model_info(a2a_client: "A2AClientType", kwargs: dict[str, Any]) -> str:
"""
Extract agent info and set model/custom_llm_provider for cost tracking.
@ -179,7 +183,7 @@ def _get_a2a_model_info(a2a_client: Any, kwargs: dict[str, Any]) -> str:
return agent_name
def _get_a2a_client_agent_card(a2a_client: Any) -> Optional["AgentCard"]:
def _get_a2a_client_agent_card(a2a_client: "A2AClientType") -> Optional["AgentCard"]:
agent_card = cast(Optional["AgentCard"], getattr(a2a_client, "_litellm_agent_card", None))
if agent_card is not None:
return agent_card
@ -191,9 +195,9 @@ def _get_a2a_client_agent_card(a2a_client: Any) -> Optional["AgentCard"]:
async def _send_message_via_completion_bridge(
request: "SendMessageRequest",
custom_llm_provider: str,
custom_llm_provider: object,
api_base: str | None,
litellm_params: dict[str, Any],
litellm_params: dict[str, object],
agent_extra_headers: dict[str, str] | None = None,
) -> LiteLLMSendMessageResponse:
"""
@ -224,6 +228,20 @@ def _get_a2a_call_context(a2a_client: "A2AClientType") -> Optional["A2ACallConte
return getattr(a2a_client, "_litellm_call_context", None)
def _to_core_send_message_request(request: "SendMessageRequest") -> "CoreSendMessageRequest":
from a2a.compat.v0_3 import conversions
return conversions.to_core_send_message_request(request)
def _to_compat_stream_response(
event: "CoreStreamResponse", request_id: str | int
) -> "SendStreamingMessageSuccessResponse":
from a2a.compat.v0_3 import conversions
return conversions.to_compat_stream_response(event, request_id=request_id)
async def _send_message(a2a_client: "A2AClientType", request: "SendMessageRequest") -> "SendMessageResponse":
"""Send a non-streaming message via a2a-sdk 1.x and return JSON-RPC response."""
if _a2a_conversions is None:
@ -231,17 +249,14 @@ async def _send_message(a2a_client: "A2AClientType", request: "SendMessageReques
"The 'a2a' package is required for A2A agent invocation. Install it with: pip install a2a-sdk"
)
pb_request: Final = _a2a_conversions.to_core_send_message_request(request)
pb_request: Final = _to_core_send_message_request(request)
last_event = None
async for event in a2a_client.send_message(pb_request, context=_get_a2a_call_context(a2a_client)):
last_event = event
if last_event is None:
raise RuntimeError("A2A send_message failed: no response received from agent.")
stream_compat: Final = _a2a_conversions.to_compat_stream_response(
last_event,
request_id=request.id,
)
stream_compat: Final = _to_compat_stream_response(last_event, request_id=request.id)
result: Final = stream_compat.result
if not isinstance(result, (Message, Task)):
raise RuntimeError(
@ -306,12 +321,9 @@ async def _stream_messages(
"The 'a2a' package is required for A2A agent invocation. Install it with: pip install a2a-sdk"
)
pb_request: Final = _a2a_conversions.to_core_send_message_request(request)
pb_request: Final[CoreSendMessageRequest] = _a2a_conversions.to_core_send_message_request(request)
async for event in a2a_client.send_message(pb_request, context=_get_a2a_call_context(a2a_client)):
compat_chunk = _a2a_conversions.to_compat_stream_response(
event,
request_id=request.id,
)
compat_chunk = _to_compat_stream_response(event, request_id=request.id)
yield SendStreamingMessageResponse(root=compat_chunk)
@ -368,10 +380,10 @@ async def asend_message(
a2a_client: Optional["A2AClientType"] = None,
request: Optional["SendMessageRequest"] = None,
api_base: str | None = None,
litellm_params: dict[str, Any] | None = None,
litellm_params: dict[str, object] | None = None,
agent_id: str | None = None,
agent_extra_headers: dict[str, str] | None = None,
**kwargs: Any,
**kwargs: object,
) -> LiteLLMSendMessageResponse:
"""
Async: Send a message to an A2A agent.
@ -485,7 +497,7 @@ async def asend_message(
response: Final = LiteLLMSendMessageResponse.from_a2a_response(a2a_response, request_id=str(request.id))
# Calculate token usage from request and response
response_dict: Final = a2a_response.model_dump(mode="json", exclude_none=True)
response_dict: Final[dict[str, object]] = a2a_response.model_dump(mode="json", exclude_none=True)
(
prompt_tokens,
completion_tokens,
@ -516,7 +528,7 @@ def send_message(
a2a_client: "A2AClientType",
request: "SendMessageRequest",
**kwargs: Any,
) -> LiteLLMSendMessageResponse | Coroutine[Any, Any, LiteLLMSendMessageResponse]:
) -> LiteLLMSendMessageResponse | Coroutine[object, object, LiteLLMSendMessageResponse]:
"""
Sync: Send a message to an A2A agent.
@ -545,9 +557,9 @@ def _build_streaming_logging_obj(
request: "SendStreamingMessageRequest",
agent_name: str,
agent_id: str | None,
litellm_params: dict[str, Any] | None,
metadata: dict[str, Any] | None,
proxy_server_request: dict[str, Any] | None,
litellm_params: dict[str, object] | None,
metadata: dict[str, object] | None,
proxy_server_request: dict[str, object] | None,
) -> Logging:
"""Build logging object for streaming A2A requests."""
start_time: Final = datetime.datetime.now()
@ -588,10 +600,10 @@ async def asend_message_streaming(
a2a_client: Optional["A2AClientType"] = None,
request: Optional["SendStreamingMessageRequest"] = None,
api_base: str | None = None,
litellm_params: dict[str, Any] | None = None,
litellm_params: dict[str, object] | None = None,
agent_id: str | None = None,
metadata: dict[str, Any] | None = None,
proxy_server_request: dict[str, Any] | None = None,
metadata: dict[str, object] | None = None,
proxy_server_request: dict[str, object] | None = None,
agent_extra_headers: dict[str, str] | None = None,
**kwargs: object,
) -> AsyncIterator[Any]:

4
litellm/caching/redis_cache.py
View file

@ -1572,7 +1572,7 @@ class RedisCache(BaseCache):
async def _pipeline_rpush_helper(
self,
pipe: pipeline,
rpush_list: list[RedisPipelineRpushOperation],
rpush_list: Sequence[RedisPipelineRpushOperation],
) -> list[int]:
"""Helper function for pipeline rpush operations"""
for rpush_op in rpush_list:
@ -1588,7 +1588,7 @@ class RedisCache(BaseCache):
@_redis_circuit_breaker_guard
async def async_rpush_pipeline(
self,
rpush_list: list[RedisPipelineRpushOperation],
rpush_list: Sequence[RedisPipelineRpushOperation],
) -> list[int]:
"""
Use Redis Pipelines for bulk RPUSH operations

6
litellm/constants.py
View file

@ -1491,6 +1491,9 @@ SPEND_LOG_CLEANUP_MAX_CONSECUTIVE_BATCH_FAILURES = int(os.getenv("SPEND_LOG_CLEA
SPEND_LOG_CLEANUP_BATCH_FAILURE_BACKOFF_SECONDS: Final = float(
os.getenv("SPEND_LOG_CLEANUP_BATCH_FAILURE_BACKOFF_SECONDS", 0.5)
)
SPEND_LOG_CLEANUP_RUN_BUDGET_SECONDS: Final = float(os.getenv("SPEND_LOG_CLEANUP_RUN_BUDGET_SECONDS", "300"))
SPEND_LOG_CLEANUP_BATCH_TIMEOUT_SECONDS: Final = float(os.getenv("SPEND_LOG_CLEANUP_BATCH_TIMEOUT_SECONDS", "30"))
SPEND_LOG_CLEANUP_REMAINING_COUNT_CAP: Final = int(os.getenv("SPEND_LOG_CLEANUP_REMAINING_COUNT_CAP", "100000"))
TOOL_SPEND_TOP_TOOLS: Final = 100
SPEND_LOG_PARTITION_INTERVAL: Final = os.getenv("SPEND_LOG_PARTITION_INTERVAL", "day")
SPEND_LOG_PARTITION_PRECREATE_AHEAD: Final = int(os.getenv("SPEND_LOG_PARTITION_PRECREATE_AHEAD", 7))
@ -1742,6 +1745,9 @@ PTU_ROLLUP_LOCK_TTL_SECONDS: Final[int] = 900
# Furthest back the catch-up pass looks for unpriced PTU days when a deployment
# declares no ptu_effective_from, bounding the scan for an open-ended window.
PTU_ROLLUP_MAX_BACKFILL_DAYS: Final[int] = 90
# Deployments named in the lapsed-window alert before it is truncated, so a fleet-wide
# expiry cannot produce an alert too large for the channel delivering it.
PTU_LAPSED_ALERT_LIMIT: Final[int] = 10
# Slack allowed when deciding a sentinel row is stale. The row's updated_at and the
# run's cutoff are stamped by different hosts, so clock skew between them must not let
# one run delete a charge another just wrote. A stale row is hours old and a concurrent

61
litellm/google_genai/adapters/transformation.py
View file

@ -1,6 +1,8 @@
import json
from collections.abc import AsyncIterator, Iterator
from typing import Any, Final, cast
from typing import Any, Final, TypedDict, cast
from typing_extensions import ReadOnly
from litellm import verbose_logger
from litellm.litellm_core_utils.json_validation_rule import normalize_tool_schema
@ -28,6 +30,19 @@ from litellm.types.utils import (
)
class _GenAITextPart(TypedDict, total=False):
text: ReadOnly[str]
class _GenAISystemInstruction(TypedDict, total=False):
parts: ReadOnly[list[_GenAITextPart]]
class _GenAIPart(TypedDict, total=False):
text: ReadOnly[str]
functionCall: ReadOnly[dict[str, object]]
class GoogleGenAIStreamWrapper(AdapterCompletionStreamWrapper):
"""
Wrapper for streaming Google GenAI generate_content responses.
@ -36,9 +51,9 @@ class GoogleGenAIStreamWrapper(AdapterCompletionStreamWrapper):
sent_first_chunk: bool = False
# State tracking for accumulating partial tool calls
accumulated_tool_calls: dict[str, dict[str, Any]]
accumulated_tool_calls: dict[str, dict[str, str]]
def __init__(self, completion_stream: Any):
def __init__(self, completion_stream: object):
self.sent_first_chunk = False
self.accumulated_tool_calls = {}
self._returned_response = False
@ -85,7 +100,7 @@ class GoogleGenAIStreamWrapper(AdapterCompletionStreamWrapper):
# After the stream is exhausted, check for any remaining accumulated tool calls
if self.accumulated_tool_calls:
try:
parts: Final = []
parts: Final[list[_GenAIPart]] = []
for (
tool_call_index,
tool_call_data,
@ -94,7 +109,7 @@ class GoogleGenAIStreamWrapper(AdapterCompletionStreamWrapper):
# For tool calls with no arguments, accumulated_args will be "", which is not valid JSON.
# We default to an empty JSON object in this case.
parsed_args = json.loads(tool_call_data["arguments"] or "{}")
function_call_part = {
function_call_part: _GenAIPart = {
"functionCall": {
"name": tool_call_data["name"] or "undefined_tool_name",
"args": parsed_args,
@ -110,7 +125,7 @@ class GoogleGenAIStreamWrapper(AdapterCompletionStreamWrapper):
tool_call_data["arguments"],
)
if parts:
final_chunk: Final = {
final_chunk: Final[dict[str, object]] = {
"candidates": [
{
"content": {"parts": parts, "role": "model"},
@ -273,9 +288,9 @@ class GoogleGenAIAdapter:
def _add_generic_litellm_params_to_request(
self,
completion_request_dict: dict[str, Any],
completion_request_dict: dict[str, object],
litellm_params: GenericLiteLLMParams | None = None,
) -> dict:
) -> dict[str, object]:
"""Add generic litellm params to request. e.g add api_base, api_key, api_version, etc.
Args:
@ -295,7 +310,7 @@ class GoogleGenAIAdapter:
def translate_completion_output_params_streaming(
self,
completion_stream: Any,
completion_stream: object,
) -> AsyncIterator[bytes] | None:
"""Transform streaming completion output to Google GenAI format"""
google_genai_wrapper: Final = GoogleGenAIStreamWrapper(completion_stream=completion_stream)
@ -307,12 +322,12 @@ class GoogleGenAIAdapter:
tools: list[dict[str, Any]],
) -> list[ChatCompletionToolParam]:
"""Transform Google GenAI tools to OpenAI tools format"""
openai_tools: Final[list[dict[str, Any]]] = []
openai_tools: Final[list[dict[str, object]]] = []
for tool in tools:
if "functionDeclarations" in tool:
for func_decl in tool["functionDeclarations"]:
function_chunk: dict[str, Any] = {
function_chunk: dict[str, object] = {
"name": func_decl.get("name", ""),
}
@ -321,7 +336,7 @@ class GoogleGenAIAdapter:
if "parametersJsonSchema" in func_decl:
function_chunk["parameters"] = func_decl["parametersJsonSchema"]
openai_tool = {"type": "function", "function": function_chunk}
openai_tool: dict[str, object] = {"type": "function", "function": function_chunk}
openai_tools.append(openai_tool)
# normalize the tool schemas
@ -345,7 +360,7 @@ class GoogleGenAIAdapter:
def _transform_contents_to_messages(
self,
contents: list[dict[str, Any]],
system_instruction: dict[str, Any] | None = None,
system_instruction: _GenAISystemInstruction | None = None,
) -> list[AllMessageValues]:
"""Transform Google GenAI contents to OpenAI messages format"""
messages: Final[list[AllMessageValues]] = []
@ -461,7 +476,7 @@ class GoogleGenAIAdapter:
def translate_completion_to_generate_content(
self,
response: ModelResponse,
) -> dict[str, Any]:
) -> dict[str, object]:
"""
Transform litellm completion response to Google GenAI generate_content format
@ -490,7 +505,7 @@ class GoogleGenAIAdapter:
parts = [{"text": message_content}] if message_content else []
# Create Google GenAI format response
generate_content_response: Final[dict[str, Any]] = {
generate_content_response: Final[dict[str, object]] = {
"candidates": [
{
"content": {"parts": parts, "role": "model"},
@ -524,7 +539,7 @@ class GoogleGenAIAdapter:
self,
response: ModelResponse | ModelResponseStream,
wrapper: GoogleGenAIStreamWrapper,
) -> dict[str, Any] | None:
) -> dict[str, object] | None:
"""
Transform streaming litellm completion chunk to Google GenAI generate_content format
@ -560,7 +575,7 @@ class GoogleGenAIAdapter:
return None
# Create Google GenAI streaming format response
streaming_chunk: Final[dict[str, Any]] = {
streaming_chunk: Final[dict[str, object]] = {
"candidates": [
{
"content": {"parts": parts, "role": "model"},
@ -597,9 +612,9 @@ class GoogleGenAIAdapter:
def _transform_openai_message_to_google_genai_parts(
self,
message: Any,
) -> list[dict[str, Any]]:
) -> list[_GenAIPart]:
"""Transform OpenAI message to Google GenAI parts format"""
parts: Final[list[dict[str, Any]]] = []
parts: Final[list[_GenAIPart]] = []
# Add text content if present
if hasattr(message, "content") and message.content:
@ -614,7 +629,7 @@ class GoogleGenAIAdapter:
except json.JSONDecodeError:
args = {}
function_call_part = {
function_call_part: _GenAIPart = {
"functionCall": {
"name": tool_call.function.name or "undefined_tool_name",
"args": args,
@ -626,14 +641,14 @@ class GoogleGenAIAdapter:
def _transform_openai_delta_to_google_genai_parts_with_accumulation(
self, delta: Any, wrapper: GoogleGenAIStreamWrapper
) -> list[dict[str, Any]]:
) -> list[_GenAIPart]:
"""Transforms OpenAI delta to Google GenAI parts, accumulating streaming tool calls."""
# 1. Initialize wrapper state if it doesn't exist
if not hasattr(wrapper, "accumulated_tool_calls"):
wrapper.accumulated_tool_calls = {}
parts: Final[list[dict[str, Any]]] = []
parts: Final[list[_GenAIPart]] = []
if hasattr(delta, "content") and delta.content:
parts.append({"text": delta.content})
@ -686,7 +701,7 @@ class GoogleGenAIAdapter:
# The part will be created by a later chunk that brings the name.
if accumulated_name:
# If successful, create the part and clean up
function_call_part = {"functionCall": {"name": accumulated_name, "args": parsed_args}}
function_call_part: _GenAIPart = {"functionCall": {"name": accumulated_name, "args": parsed_args}}
parts.append(function_call_part)
# Remove the completed tool call from the accumulator

115
litellm/integrations/langfuse/langfuse.py
View file

@ -2,8 +2,9 @@
# On success, logs events to Langfuse
import os
import traceback
from collections.abc import Callable
from collections.abc import Callable, Iterable, Mapping
from datetime import datetime
from types import MappingProxyType
from typing import TYPE_CHECKING, Any, Final, cast
from packaging.version import Version
@ -30,6 +31,7 @@ from litellm.types.utils import (
ImageResponse,
ModelResponse,
RerankResponse,
StandardLoggingMetadata,
StandardLoggingPayload,
StandardLoggingPromptManagementMetadata,
TextCompletionResponse,
@ -46,6 +48,11 @@ else:
Langfuse = Any
_DENIED_STEERING_KEYS: Final = frozenset({"headers", "endpoint", "caching_groups", "previous_models"})
_NO_METADATA: Final[Mapping[str, Any]] = MappingProxyType({})
_REDACTED_PROXY_HEADERS: Final[frozenset[str]] = frozenset({"authorization", "cookie", "referer"})
def _extract_cache_read_input_tokens(usage_obj) -> int:
"""
Extract cache_read_input_tokens from usage object.
@ -75,6 +82,22 @@ def _extract_cache_read_input_tokens(usage_obj) -> int:
return cache_read_input_tokens
def _as_steering_flag(value: object) -> bool:
"""A string ``str_to_bool`` does not recognise falls back to its truthiness."""
if isinstance(value, str):
parsed: Final = str_to_bool(value)
return bool(value) if parsed is None else parsed
return bool(value)
def _as_steering_key_sequence(value: object) -> tuple[str, ...]:
if isinstance(value, str):
return tuple(key.strip() for key in value.split(",") if key.strip())
if isinstance(value, Iterable):
return tuple(str(key) for key in value)
return ()
def resolve_langfuse_credentials(
langfuse_public_key=None,
langfuse_secret=None,
@ -496,16 +519,14 @@ class LangFuseLogger:
else []
)
if standard_logging_object is None:
end_user_id = None
prompt_management_metadata: StandardLoggingPromptManagementMetadata | None = None
else:
end_user_id = standard_logging_object["metadata"].get("user_api_key_end_user_id", None)
prompt_management_metadata = cast(
StandardLoggingPromptManagementMetadata | None,
standard_logging_object["metadata"].get("prompt_management_metadata", None),
)
allowlisted_metadata: Final[StandardLoggingMetadata | dict[str, Any]] = (
standard_logging_object["metadata"] if standard_logging_object is not None else _NO_METADATA
)
end_user_id: Final = allowlisted_metadata.get("user_api_key_end_user_id", None)
prompt_management_metadata: Final[StandardLoggingPromptManagementMetadata | None] = cast(
StandardLoggingPromptManagementMetadata | None,
allowlisted_metadata.get("prompt_management_metadata", None),
)
# Clean Metadata before logging - never log raw metadata
# the raw metadata can contain circular references which leads to infinite recursion
@ -524,12 +545,7 @@ class LangFuseLogger:
tags.append(f"{key}:{value}")
# clean litellm metadata before logging
if key in [
"headers",
"endpoint",
"caching_groups",
"previous_models",
]:
if key in _DENIED_STEERING_KEYS:
continue
else:
clean_metadata[key] = value
@ -552,10 +568,13 @@ class LangFuseLogger:
# This allows continuing an existing trace while still returning the correct trace_id
if existing_trace_id is not None:
trace_id = existing_trace_id
update_trace_keys: Final = cast(list, clean_metadata.pop("update_trace_keys", []))
requested_trace_keys: Final = _as_steering_key_sequence(clean_metadata.pop("update_trace_keys", ()))
update_trace_keys: Final = (
requested_trace_keys if _as_steering_flag(litellm.langfuse_enable_update_trace_keys) else ()
)
debug: Final = clean_metadata.pop("debug_langfuse", None)
mask_input: Final = clean_metadata.pop("mask_input", False)
mask_output: Final = clean_metadata.pop("mask_output", False)
mask_input: Final = _as_steering_flag(clean_metadata.pop("mask_input", False))
mask_output: Final = _as_steering_flag(clean_metadata.pop("mask_output", False))
# Look for masking function in the dedicated location first (set by scrub_sensitive_keys_in_metadata)
# Fall back to metadata for backwards compatibility
masking_function: Final = litellm_params.get("_langfuse_masking_function") or clean_metadata.pop(
@ -614,19 +633,18 @@ class LangFuseLogger:
trace_params["output"] = output if not mask_output else "redacted-by-litellm"
if debug is True or (isinstance(debug, str) and debug.lower() == "true"):
if "metadata" in trace_params:
# log the raw_metadata in the trace
trace_params["metadata"]["metadata_passed_to_litellm"] = metadata
else:
trace_params["metadata"] = {"metadata_passed_to_litellm": metadata}
debug_metadata: Final = {
key: value for key, value in metadata.items() if isinstance(value, (str, int, float, bool))
}
trace_params["metadata"] = {
**(trace_params.get("metadata") or _NO_METADATA),
"metadata_passed_to_litellm": debug_metadata,
}
cost: Final = kwargs.get("response_cost", None)
verbose_logger.debug("trace: %s", cost)
clean_metadata["litellm_response_cost"] = cost
if standard_logging_object is not None:
hidden_params: Final = standard_logging_object.get("hidden_params", {})
clean_metadata["hidden_params"] = filter_exceptions_from_params(hidden_params)
hidden_params: Final = standard_logging_object.get("hidden_params") if standard_logging_object else None
if (
litellm.langfuse_default_tags is not None
@ -638,22 +656,24 @@ class LangFuseLogger:
tags.append(f"proxy_base_url:{proxy_base_url}")
api_base: Final = litellm_params.get("api_base", None)
if api_base:
clean_metadata["api_base"] = api_base
vertex_location: Final = kwargs.get("vertex_location", None)
if vertex_location:
clean_metadata["vertex_location"] = vertex_location
aws_region_name: Final = kwargs.get("aws_region_name", None)
if aws_region_name:
clean_metadata["aws_region_name"] = aws_region_name
candidate_enrichments: Final = (
("litellm_response_cost", cost, True),
("hidden_params", filter_exceptions_from_params(hidden_params), hidden_params is not None),
("api_base", api_base, bool(api_base)),
("vertex_location", vertex_location, bool(vertex_location)),
("aws_region_name", aws_region_name, bool(aws_region_name)),
("cache_hit", kwargs.get("cache_hit") or False, self._supports_tags() and "cache_hit" in kwargs),
)
enrichments: Final[Mapping[str, Any]] = {
key: value for key, value, include in candidate_enrichments if include
}
if self._supports_tags():
if "cache_hit" in kwargs:
if kwargs["cache_hit"] is None:
kwargs["cache_hit"] = False
clean_metadata["cache_hit"] = kwargs["cache_hit"]
if "cache_hit" in kwargs and kwargs["cache_hit"] is None:
kwargs["cache_hit"] = False # rebind-ok: pre-existing normalization other integrations rely on
if existing_trace_id is None:
trace_params.update({"tags": tags})
@ -666,13 +686,13 @@ class LangFuseLogger:
if headers:
for key, value in headers.items():
# these headers can leak our API keys and/or JWT tokens
if key.lower() not in ["authorization", "cookie", "referer"]:
if key.lower() not in _REDACTED_PROXY_HEADERS:
clean_headers[key] = value
trace: Final[StatefulTraceClient] = self.Langfuse.trace(**trace_params)
# Log provider specific information as a span
log_provider_specific_information_as_span(trace, clean_metadata)
log_provider_specific_information_as_span(trace, enrichments)
# Log guardrail information as a span
self._log_guardrail_information_as_span(
@ -745,7 +765,10 @@ class LangFuseLogger:
"output": output if not mask_output else "redacted-by-litellm",
"usage": usage,
"usage_details": usage_details,
"metadata": log_requester_metadata(clean_metadata),
"metadata": {
**log_requester_metadata(redact_user_api_key_info(metadata=allowlisted_metadata)),
**enrichments,
},
"level": level,
"version": clean_metadata.pop("version", None),
}
@ -1042,7 +1065,7 @@ def _add_prompt_to_generation_params(
def log_provider_specific_information_as_span(
trace,
clean_metadata,
clean_metadata: Mapping[str, Any],
):
"""
Logs provider-specific information as spans.
@ -1082,7 +1105,7 @@ def log_provider_specific_information_as_span(
)
def log_requester_metadata(clean_metadata: dict):
def log_requester_metadata(clean_metadata: Mapping[str, Any]):
returned_metadata: Final = {}
requester_metadata: Final = clean_metadata.get("requester_metadata") or {}
for k, v in clean_metadata.items():

10
litellm/integrations/langfuse/langfuse_otel.py
View file

@ -90,7 +90,6 @@ class LangfuseOtelLogger(OpenTelemetry):
"generation_name": LangfuseSpanAttributes.GENERATION_NAME,
"generation_id": LangfuseSpanAttributes.GENERATION_ID,
"parent_observation_id": LangfuseSpanAttributes.PARENT_OBSERVATION_ID,
"version": LangfuseSpanAttributes.GENERATION_VERSION,
"mask_input": LangfuseSpanAttributes.MASK_INPUT,
"mask_output": LangfuseSpanAttributes.MASK_OUTPUT,
"trace_user_id": LangfuseSpanAttributes.TRACE_USER_ID,
@ -99,13 +98,18 @@ class LangfuseOtelLogger(OpenTelemetry):
"trace_name": LangfuseSpanAttributes.TRACE_NAME,
"trace_id": LangfuseSpanAttributes.TRACE_ID,
"trace_metadata": LangfuseSpanAttributes.TRACE_METADATA,
"trace_version": LangfuseSpanAttributes.TRACE_VERSION,
"trace_release": LangfuseSpanAttributes.TRACE_RELEASE,
"trace_release": LangfuseSpanAttributes.RELEASE,
"existing_trace_id": LangfuseSpanAttributes.EXISTING_TRACE_ID,
"update_trace_keys": LangfuseSpanAttributes.UPDATE_TRACE_KEYS,
"debug_langfuse": LangfuseSpanAttributes.DEBUG_LANGFUSE,
}
version: Final = (
metadata.get("trace_version") if metadata.get("trace_version") is not None else metadata.get("version")
)
if version is not None:
safe_set_attribute(span, LangfuseSpanAttributes.VERSION.value, version)
for key, enum_attr in mapping.items():
if key in metadata and metadata[key] is not None:
value = metadata[key]

8
litellm/integrations/otel/presets/langfuse.py
View file

@ -42,9 +42,7 @@ def langfuse_dynamic_headers(params: StandardCallbackDynamicParams) -> dict[str,
public_key: Final = params.get("langfuse_public_key")
secret_key: Final = params.get("langfuse_secret_key")
if public_key and secret_key:
return {
"Authorization": _V1Langfuse._get_langfuse_authorization_header(
public_key=public_key, secret_key=secret_key
)
}
return _V1Langfuse._build_langfuse_otel_headers(
_V1Langfuse._get_langfuse_authorization_header(public_key=public_key, secret_key=secret_key)
)
return {}

85
litellm/integrations/rubrik.py
View file

@ -6,12 +6,13 @@ import random
import time
import uuid
from collections import Counter
from collections.abc import Mapping, Sequence
from collections.abc import Awaitable, Mapping, Sequence
from dataclasses import dataclass
from types import MappingProxyType
from typing import TYPE_CHECKING, Any, Final, Literal, Optional
from typing import TYPE_CHECKING, Any, Final, Literal, Optional, TypedDict
import httpx
from typing_extensions import Never, ReadOnly
from litellm._logging import verbose_logger
from litellm.integrations.custom_batch_logger import CustomBatchLogger
@ -48,7 +49,20 @@ _WEBHOOK_PATH_PROMPT_MODERATION: Final = "/v1/before_prompt/openai/v1"
_WEBHOOK_PATH_LOGGING_BATCH: Final = "/v1/litellm/batch"
_MAX_QUEUE_SIZE: Final = 10_000
_DROP_WARNING_INTERVAL_SECONDS: Final = 60.0
_EMPTY_MAPPING: Final[Mapping[str, Any]] = MappingProxyType({})
_EMPTY_MAPPING: Final[Mapping[str, Never]] = MappingProxyType({})
class _ServiceToolCall(TypedDict):
id: ReadOnly[str]
class _ServiceMessage(TypedDict, total=False):
content: ReadOnly[str]
tool_calls: ReadOnly[Sequence[_ServiceToolCall]]
class _ServiceChoice(TypedDict, total=False):
message: ReadOnly[_ServiceMessage]
class _MalformedToolBlockingResponseError(Exception):
@ -143,7 +157,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
else {"Content-Type": "application/json"}
)
self._periodic_flush_task: asyncio.Task[Any] | None = self._start_periodic_flush_task()
self._periodic_flush_task: asyncio.Task[None] | None = self._start_periodic_flush_task()
@classmethod
def get_supported_event_hooks(cls) -> list[GuardrailEventHooks]:
@ -191,7 +205,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
params={"timeout": httpx.Timeout(5.0, connect=2.0)},
)
def _start_periodic_flush_task(self) -> asyncio.Task[Any] | None:
def _start_periodic_flush_task(self) -> asyncio.Task[None] | None:
"""Start the periodic flush task only when an event loop is already running."""
try:
loop: Final = asyncio.get_running_loop()
@ -212,7 +226,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
Closing them here would close the shared connection pool for every
other logger instance; let LiteLLM manage their lifecycle instead.
"""
task: Final = getattr(self, "_periodic_flush_task", None)
task: Final[asyncio.Task[None] | None] = getattr(self, "_periodic_flush_task", None)
if task is not None:
task.cancel()
@ -253,7 +267,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
@staticmethod
async def _guarded(
coro: Any,
coro: Awaitable[GenericGuardrailAPIInputs],
inputs: GenericGuardrailAPIInputs,
label: str,
) -> GenericGuardrailAPIInputs:
@ -400,7 +414,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
request_data["_rubrik_logging_obj"] = logging_obj
@staticmethod
def _normalize_tool_calls(tool_calls: Any) -> tuple[ChatCompletionMessageToolCall, ...]:
def _normalize_tool_calls(tool_calls: Sequence[object]) -> tuple[ChatCompletionMessageToolCall, ...]:
"""Convert tool_calls from inputs to ChatCompletionMessageToolCall objects."""
return tuple(RubrikLogger._normalize_tool_call(tc) for tc in tool_calls)
@ -427,7 +441,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
raise TypeError(f"Cannot normalize tool_call of type {type(tc).__name__}: {tc!r}")
@staticmethod
def _join_texts(texts: Any) -> str:
def _join_texts(texts: Sequence[str] | None) -> str:
"""Join response text segments into the single content string the
webhook evaluates. Empty when there is no assistant text."""
if not texts:
@ -439,14 +453,14 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
tool_calls: Sequence[ChatCompletionMessageToolCall],
content: str,
request_id: str | None,
) -> Mapping[str, Any]:
) -> Mapping[str, object]:
"""Build an OpenAI ChatCompletion-format dict (assistant text + tool
calls) for the after_completion webhook.
``content`` is sent so the webhook can moderate the response text;
``None`` when the assistant produced no text (tool-call-only response).
"""
message: Final[dict[str, Any]] = {
message: Final[dict[str, object]] = {
"role": "assistant",
"content": content or None,
}
@ -467,7 +481,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
}
@staticmethod
def _flatten_messages_for_moderation(messages: Any) -> tuple[Mapping[str, Any], ...]:
def _flatten_messages_for_moderation(messages: Sequence[object] | None) -> tuple[Mapping[str, Any], ...]:
"""Collapse each message's content to a plain string for the webhook.
litellm normalizes Anthropic ``/v1/messages`` requests to OpenAI shape,
@ -506,8 +520,8 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
@staticmethod
def _build_prompt_moderation_payload(
inputs: GenericGuardrailAPIInputs,
request_data: Mapping[str, Any],
) -> Mapping[str, Any]:
request_data: Mapping[str, object],
) -> Mapping[str, object]:
"""Build the bare OpenAI request the before_prompt webhook consumes.
Unlike the after_completion envelope, this endpoint takes a raw OpenAI
@ -516,7 +530,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
``/v1/messages`` requests too. Optional fields are sent only when
present so the payload stays clean.
"""
payload: Final[dict[str, Any]] = {
payload: Final[dict[str, object]] = {
"model": inputs.get("model") or request_data.get("model") or "",
"messages": RubrikLogger._flatten_messages_for_moderation(inputs.get("structured_messages")),
}
@ -540,8 +554,8 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
@staticmethod
def _extract_request_data(
call_details: Mapping[str, Any],
request_data: Mapping[str, Any] | None,
) -> Mapping[str, Any]:
request_data: Mapping[str, object] | None,
) -> Mapping[str, object]:
"""Extract original request data from model_call_details for the
response moderation service envelope.
@ -576,7 +590,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
}
@staticmethod
def _sanitize_proxy_server_request(proxy_server_request: Any) -> Any:
def _sanitize_proxy_server_request(proxy_server_request: object) -> object:
"""Allowlist only routing fields (``url``, ``method``) when forwarding
``proxy_server_request`` to an external webhook, dropping inbound
``headers`` (Authorization, Cookie, x-api-key, ...) and the raw
@ -586,17 +600,18 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
return {key: proxy_server_request[key] for key in ("url", "method") if key in proxy_server_request}
@staticmethod
def _resolve_model(request_data: Mapping[str, Any], call_details: Mapping[str, Any]) -> str:
def _resolve_model(request_data: Mapping[str, object], call_details: Mapping[str, str]) -> str:
"""Get the model name for the ModifyResponseException."""
response: Final = request_data.get("response")
if response and hasattr(response, "model"):
return response.model or "unknown"
response_model: Final[str | None] = getattr(response, "model", None)
return response_model or "unknown"
return call_details.get("model", "unknown")
# -- Logging hooks ---------------------------------------------------------
@staticmethod
def _correlation_id(call_details: Mapping[str, Any], request_data: Mapping[str, Any] | None = None) -> str | None:
def _correlation_id(call_details: Mapping[str, str], request_data: Mapping[str, str] | None = None) -> str | None:
"""The id that joins a blocked request's two S3 logs by filename: the
moderation (``_blocking``) log and the failure (response) log.
@ -610,7 +625,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
return call_details.get("litellm_call_id") or (request_data or _EMPTY_MAPPING).get("litellm_call_id")
@classmethod
def _apply_correlation_id(cls, payload: dict[str, Any], source: Mapping[str, Any]) -> None:
def _apply_correlation_id(cls, payload: dict[str, object], source: Mapping[str, str]) -> None:
"""Pin ``payload["id"]`` to ``litellm_call_id`` in place so this log
shares its S3 filename id with the moderation (``_blocking``) and
failure logs for the same request -- for every provider.
@ -630,7 +645,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
payload["id"] = correlated
@staticmethod
def _prepend_system_prompt(payload: dict[str, Any], source: Mapping[str, Any]) -> None:
def _prepend_system_prompt(payload: dict[str, object], source: Mapping[str, object]) -> None:
"""Prepend ``source["system"]`` onto ``payload["messages"]``.
Builds a NEW messages list rather than mutating ``payload["messages"]``
@ -658,7 +673,9 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
exc_info=True,
)
async def _prepare_log_payload(self, kwargs: Mapping[str, Any], event_type: str) -> StandardLoggingPayload | None:
async def _prepare_log_payload(
self, kwargs: Mapping[str, object], event_type: str
) -> StandardLoggingPayload | None:
"""Shared logic for success logging (sampled)."""
if random.random() > self.sampling_rate:
verbose_logger.debug("Skipping Rubrik %s logging (sampling_rate=%s)", event_type, self.sampling_rate)
@ -697,7 +714,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
self._dropped_since_warning = 0
self._last_drop_warning_time = now
async def _enqueue_log_event(self, kwargs: Mapping[str, Any], event_type: str):
async def _enqueue_log_event(self, kwargs: Mapping[str, object], event_type: str):
try:
payload: Final = await self._prepare_log_payload(kwargs, event_type)
if payload is None:
@ -862,7 +879,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
base: Final = call_details.get("standard_logging_object")
if base is not None:
payload: dict = safe_deep_copy(base)
payload: dict[str, object] = safe_deep_copy(base)
else:
verbose_logger.debug(
"Rubrik: standard_logging_object not yet on model_call_details "
@ -908,7 +925,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
cls,
call_details: Mapping[str, Any],
user_api_key_dict: "UserAPIKeyAuth",
) -> dict[str, Any]:
) -> dict[str, object]:
# Convert datetime to a Unix float so json.dumps can serialize it.
# httpx's json= parameter uses stdlib json.dumps with no custom encoder.
_raw_start: Final = call_details.get("start_time")
@ -996,7 +1013,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
# -- Webhook services ------------------------------------------------------
async def _post_json(self, endpoint: str, payload: Mapping[str, Any], service_name: str) -> Mapping[str, Any]:
async def _post_json(self, endpoint: str, payload: Mapping[str, object], service_name: str) -> Mapping[str, Any]:
"""POST ``payload`` to a Rubrik webhook and return its dict response.
Raises:
@ -1010,7 +1027,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
headers=self._headers,
)
http_response.raise_for_status()
result: Final = http_response.json()
result: Final[object] = http_response.json()
if not isinstance(result, dict):
raise TypeError(
f"{service_name} returned non-dict JSON "
@ -1021,8 +1038,8 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
async def _post_to_response_moderation_endpoint(
self,
response_data: Mapping[str, Any],
request_data: Mapping[str, Any],
response_data: Mapping[str, object],
request_data: Mapping[str, object],
) -> Mapping[str, Any]:
"""Post the ``{request, response}`` envelope to the after_completion
webhook and return its (possibly rewritten) response.
@ -1039,7 +1056,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
"Response moderation service",
)
async def _post_to_prompt_moderation_endpoint(self, payload: Mapping[str, Any]) -> Mapping[str, Any]:
async def _post_to_prompt_moderation_endpoint(self, payload: Mapping[str, object]) -> Mapping[str, Any]:
"""Post a bare OpenAI request to the before_prompt webhook.
Returns ``{}`` (passthrough) or a synthetic chat.completion (block).
@ -1054,7 +1071,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
chat.completion whose ``choices[0].message.content`` is the refusal
explanation.
"""
choices: Final = service_response.get("choices")
choices: Final[Sequence[_ServiceChoice] | None] = service_response.get("choices")
if not choices:
return None
message: Final = choices[0].get("message") or _EMPTY_MAPPING
@ -1086,7 +1103,7 @@ class RubrikLogger(CustomGuardrail, CustomBatchLogger):
Expects service_response in OpenAI chat completion format:
{"choices": [{"message": {"tool_calls": [...], "content": "..."}}]}
"""
choices: Final = service_response.get("choices") or ()
choices: Final[Sequence[_ServiceChoice]] = service_response.get("choices") or ()
if not choices:
raise _MalformedToolBlockingResponseError("Response moderation service returned empty response")

622
litellm/integrations/shadow_eval_logger.py Normal file
View file

@ -0,0 +1,622 @@
"""Shadow Eval Logger: samples a shadowed key's successful chat requests, duplicates each
against the job's other arm in a detached task (the auto-router for a forward job, the
fixed baseline model for a reverse one), blind-judges real vs shadow, and appends one
``LiteLLM_ShadowEvalAttempt`` row (verdict or error) as the feature's only hot-path write.
Counts, status, and spend derive from those rows at read time, so nothing can disagree
across pods or stop races; the hook reads active jobs through a short-TTL cache."""
import asyncio
import hashlib
import random
from collections.abc import Callable, Mapping, Sequence
from dataclasses import dataclass
from datetime import datetime, timezone
from itertools import groupby
from operator import itemgetter
from types import MappingProxyType
from typing import TYPE_CHECKING, Final
from pydantic import BaseModel, ConfigDict, ValidationError, field_validator, model_validator
from litellm._logging import verbose_logger
from litellm.caching.in_memory_cache import InMemoryCache
from litellm.constants import INTERNAL_CALL_ORIGIN_METADATA_KEY
from litellm.integrations.custom_logger import CustomLogger
from litellm.litellm_core_utils.core_helpers import get_litellm_metadata_from_kwargs
from litellm.litellm_core_utils.internal_call_metadata import sanitized_forwardable_call_metadata
from litellm.litellm_core_utils.llm_judge import (
default_router_provider,
extract_text_from_content,
judge_acompletion,
parse_json_verdict,
)
from litellm.litellm_core_utils.redact_messages import should_redact_message_logging
from litellm.types.management_endpoints.auto_router_endpoints import ShadowEvalDirection
from litellm.types.utils import SHADOW_EVAL_JUDGE_CALL_ORIGIN, SHADOW_EVAL_ROUTER_CALL_ORIGIN
if TYPE_CHECKING:
from litellm.proxy.utils import PrismaClient
from litellm.router import Router
from litellm.types.utils import StandardLoggingPayload
# A job starting, stopping, or hitting its turn budget propagates to sampling within one
# TTL; the turn budget can overshoot by at most one TTL of in-flight samples per pod.
_JOBS_CACHE_TTL_SECONDS: Final = 10
# Concurrent shadow+judge pipelines per pod: a traffic spike turns into skipped samples
# rather than an unbounded task pileup.
_MAX_CONCURRENT_SHADOW_TASKS: Final = 16
# Total character budget for the judge's user prompt, however long the conversation and
# the two responses are, so the prompt can never overflow a judge model's context window.
_MAX_JUDGE_RESPONSE_CHARS: Final = 8_000
_MAX_JUDGE_PROMPT_CHARS: Final = 24_000
# The judge answers with a small JSON object; a tighter budget truncates the JSON
# mid-object and the attempt is lost to an error row.
JUDGE_MAX_OUTPUT_TOKENS: Final = 500
_MAX_ERROR_CHARS: Final = 500
_EMPTY_METADATA: Final[Mapping[str, object]] = MappingProxyType({})
_SAMPLED_CALL_TYPES: Final = frozenset({"completion", "acompletion"})
PAIRWISE_JUDGE_SYSTEM_PROMPT: Final = """You are an impartial quality judge comparing two responses to the same conversation.
The responses are labeled A and B in random order. You do not know which system produced which.
Criteria: correctness, completeness, clarity, conciseness.
Return ONLY valid JSON in this exact format, no other text:
{
"preference": "A" | "B" | "tie",
"confidence": <0.0 to 1.0>,
"reasoning": "<one sentence>"
}"""
class PairwiseVerdict(BaseModel):
"""The judge's blind A/B verdict, validated at the parse boundary."""
preference: str = "tie"
confidence: float = 0.0
def _sample_hits(request_id: str, job_id: str, percentage: float) -> bool:
"""Deterministically decide whether a request falls in the shadowed slice: hash-based
rather than random so retries sample the same way and pods agree without coordination."""
digest: Final = hashlib.sha256(f"{job_id}:{request_id}".encode()).digest()
bucket: Final = int.from_bytes(digest[:8], "big") / float(2**64)
return bucket * 100.0 < percentage
def _judge_call_cost(response: object) -> float:
"""Price a judge call, treating an unmapped judge model as free rather than fatal."""
import litellm
try:
return litellm.completion_cost(completion_response=response) or 0.0
except Exception: # noqa: BLE001 # unmapped judge model: the verdict still counts, cost stays 0
return 0.0
def _unmask_preference(raw_preference: str, real_is_a: bool) -> str:
"""Map the judge's blind A/B/tie verdict back to real/shadow/tie."""
normalized: Final = raw_preference.strip().lower()
if normalized == "a":
return "real" if real_is_a else "shadow"
if normalized == "b":
return "shadow" if real_is_a else "real"
return "tie"
def _judge_user_prompt(conversation: str, response_a: str, response_b: str) -> str:
"""The judge prompt under one total character budget: each response is capped, and
the conversation tail gets whatever budget the responses left over."""
a: Final = response_a[:_MAX_JUDGE_RESPONSE_CHARS]
b: Final = response_b[:_MAX_JUDGE_RESPONSE_CHARS]
conversation_budget: Final = _MAX_JUDGE_PROMPT_CHARS - len(a) - len(b)
return (
f"Conversation:\n{conversation[-conversation_budget:]}\n\n"
f"Response A:\n{a}\n\n"
f"Response B:\n{b}\n\n"
"Which response is better?"
)
async def _key_or_team_is_over_budget(metadata: Mapping[str, object]) -> bool:
"""Whether the shadowed key or its team is over budget, decided by the same owners
the request path uses, so counter keys and thresholds can never drift from auth's.
Advisory and fail-open: real traffic on an over-budget key is already rejected at
auth (so nothing reaches the success hook), and this gate only closes the race
where the key crosses its budget while a request is in flight.
"""
try:
from litellm.exceptions import BudgetExceededError
from litellm.proxy._types import UserAPIKeyAuth
from litellm.proxy.auth.auth_checks import (
_team_max_budget_check,
_virtual_key_max_budget_check,
get_team_object,
)
from litellm.proxy.proxy_server import prisma_client, proxy_logging_obj, user_api_key_cache
except ImportError:
return False
auth: Final = metadata.get("user_api_key_auth")
if not isinstance(auth, UserAPIKeyAuth):
return False
try:
await _virtual_key_max_budget_check(valid_token=auth, proxy_logging_obj=proxy_logging_obj)
if auth.team_id:
team: Final = await get_team_object(
team_id=auth.team_id,
prisma_client=prisma_client,
user_api_key_cache=user_api_key_cache,
check_cache_only=True,
)
await _team_max_budget_check(team_object=team, valid_token=auth, proxy_logging_obj=proxy_logging_obj)
except BudgetExceededError:
return True
except Exception as e: # noqa: BLE001 # advisory gate: a failed read must not block sampling
verbose_logger.debug("shadow_eval: budget read failed: %s", e)
return False
def _routing_decision(metadata: Mapping[str, object]) -> Mapping[str, object]:
"""The routing decision a pre-routing strategy wrote to a call's metadata, empty when
a plain model served it. Read off the sampled request for the control arm, and off the
shadow call's own write-back for the shadow arm."""
decision: Final = metadata.get("routing_decision")
return decision if isinstance(decision, Mapping) else _EMPTY_METADATA
def _routed_tier(metadata: Mapping[str, object]) -> str | None:
decision: Final = _routing_decision(metadata)
raw: Final = decision.get("tier_label") or decision.get("tier")
return str(raw) if raw is not None else None
def _request_was_routed_by(request_metadata: Mapping[str, object], router_name: str) -> bool:
"""Whether the router under evaluation served this request, which is what decides
the direction it belongs to. A forward job skips its own router's traffic, since
duplicating it would compare the router to itself: guaranteed ties, judge spend for
zero information. A reverse job samples exactly that traffic and nothing else."""
return _routing_decision(request_metadata).get("router_model_name") == router_name
@dataclass(frozen=True, slots=True)
class _CallFailure:
"""A shadow or judge call that produced no usable response. cost carries any judge
spend the failed attempt still billed, so job-level judge_spend never undercounts."""
error: str
cost: float = 0.0
@dataclass(frozen=True, slots=True)
class _ShadowResponse:
"""A successful shadow call, with what the attempt row records."""
text: str
model: str
tier: str | None
@dataclass(frozen=True, slots=True)
class _JudgeVerdict:
"""A parsed judge verdict, unmasked back to real/shadow/tie."""
preference: str
confidence: float
cost: float
class ActiveShadowEvalJob(BaseModel):
"""One active job as the sampling path needs it, validated straight off the untyped
job row: immutable config plus the attempt count as of the cache fill (the turn
budget's staleness is bounded by the cache TTL). Every way a row can be unsamplable
is a validation error here, so a bad row is skipped rather than sampled wrongly."""
model_config = ConfigDict(frozen=True, from_attributes=True)
id: str
router_name: str
direction: ShadowEvalDirection = "forward"
baseline_model: str | None = None
shadow_percentage: float
judge_model: str
max_turns: int
ends_at: datetime
attempts: int = 0
@field_validator("ends_at")
@classmethod
def _as_utc(cls, value: datetime) -> datetime:
return value.replace(tzinfo=timezone.utc) if value.tzinfo is None else value
@model_validator(mode="after")
def _baseline_model_matches_direction(self) -> "ActiveShadowEvalJob":
if (self.baseline_model is not None) != (self.direction == "reverse"):
raise ValueError("baseline_model is set for exactly the reverse jobs")
return self
@property
def shadow_target(self) -> str:
"""The model the duplicated arm calls: the router itself for a forward job, the
fixed baseline for a reverse one. Total because the validator above pins
baseline_model to reverse jobs and only those."""
return self.baseline_model or self.router_name
def _as_active_job(record: object, attempts: int) -> ActiveShadowEvalJob | None:
"""The sampling path's view of one job row, or None for a row it cannot sample: an
unknown direction, or a reverse job with no baseline model to duplicate against.
Failing closed here is what keeps the dispatch path total."""
try:
job: Final = ActiveShadowEvalJob.model_validate(record)
except ValidationError as e:
verbose_logger.debug("shadow_eval: skipping unsamplable job row: %s", e)
return None
return job.model_copy(update={"attempts": attempts})
_jobs_cache: Final = InMemoryCache(max_size_in_memory=4, default_ttl=_JOBS_CACHE_TTL_SECONDS)
_JOBS_CACHE_KEY: Final = "shadow_eval:active_jobs"
class ShadowEvalLogger(CustomLogger):
"""Fires blind pairwise shadow evaluations for keys with an active shadow-eval job."""
def __init__(
self,
router_provider: Callable[[], "Router | None"] | None = None,
prisma_provider: Callable[[], "PrismaClient | None"] | None = None,
jobs_cache: InMemoryCache | None = None,
) -> None:
"""Providers are callables so the proxy's lazily-initialized globals are resolved
at call time, not at logger construction."""
self._router_provider = router_provider or default_router_provider
self._prisma_provider = prisma_provider or _default_prisma_provider
self._jobs_cache = jobs_cache or _jobs_cache
self._inflight_shadow_tasks: int = 0
# Starts per job since the last cache fill, never decremented within a
# generation; the refill absorbs written rows and resets.
self._job_starts: dict[str, int] = {} # mutable-ok: per-generation counter
async def _active_jobs(self) -> Mapping[str, tuple[ActiveShadowEvalJob, ...]]:
"""Active jobs by api_key_id, cache-first. A key holds at most one job per
direction, so the value is a collection. A DB fault returns empty without
caching, so sampling pauses for that request and the next one retries."""
cached: Final = await self._jobs_cache.async_get_cache(_JOBS_CACHE_KEY)
if cached is not None:
return cached # pyright: ignore[reportReturnType] # cache stores exactly this mapping shape
prisma: Final = self._prisma_provider()
if prisma is None:
return _EMPTY_JOBS
try:
records: Final = await prisma.db.litellm_shadowevaljob.find_many(
where={ # mutable-ok: Prisma filter
"stopped_at": None,
"ends_at": {"gt": datetime.now(timezone.utc)}, # mutable-ok: Prisma filter
},
)
grouped: Final = (
await prisma.db.litellm_shadowevalattempt.group_by(
by=["job_id"],
count=True,
where={"job_id": {"in": [str(record.id) for record in records]}}, # mutable-ok: Prisma filter
)
if records
else ()
)
attempt_counts: Final = {str(row["job_id"]): int(row["_count"]["_all"]) for row in grouped or []}
by_key: Final = tuple(
sorted(
(
(str(record.api_key_id), job)
for record in records or []
if (job := _as_active_job(record, attempt_counts.get(str(record.id), 0))) is not None
),
key=itemgetter(0),
)
)
jobs: Final = MappingProxyType(
{key: tuple(job for _, job in group) for key, group in groupby(by_key, key=itemgetter(0))}
)
await self._jobs_cache.async_set_cache(_JOBS_CACHE_KEY, jobs)
self._job_starts = {} # rebind-ok: new generation, counts absorbed into the fill
return jobs
except Exception as e: # noqa: BLE001 # a DB blip must never break request logging
verbose_logger.debug("shadow_eval: active-job read failed: %s", e)
return _EMPTY_JOBS
#### hook ####
async def async_log_success_event(
self,
kwargs: Mapping[str, object],
response_obj: object,
start_time: object,
end_time: object,
) -> None:
try:
payload: Final[StandardLoggingPayload | None] = kwargs.get("standard_logging_object") # pyright: ignore[reportAssignmentType] # untyped callback kwargs
if payload is None:
return
raw_meta: Final = get_litellm_metadata_from_kwargs(dict(kwargs)) # mutable-ok: helper needs dict
request_metadata: Final = raw_meta if isinstance(raw_meta, Mapping) else _EMPTY_METADATA
if request_metadata.get(INTERNAL_CALL_ORIGIN_METADATA_KEY):
return # internal sub-call (our own shadow/judge, a classifier), not user traffic
# redaction rewrites logged content before callbacks run, so this hook
# only ever sees placeholders for a redacted request
if should_redact_message_logging(dict(kwargs)): # mutable-ok: predicate takes a plain dict
return
metadata: Final = payload.get("metadata") or _EMPTY_METADATA
api_key_hash: Final = metadata.get("user_api_key_hash")
if not api_key_hash:
return
request_id: Final = payload.get("id") or ""
if not request_id:
return
if payload.get("call_type") not in _SAMPLED_CALL_TYPES:
return # only known chat-shaped traffic is comparable; unknown or missing types fail closed
raw_messages: Final = kwargs.get("messages")
messages: Final = (
tuple(m for m in raw_messages if isinstance(m, Mapping)) if isinstance(raw_messages, Sequence) else ()
)
control_tier: Final = _routed_tier(request_metadata)
# A key can hold one job per direction, and a request routed by one job's
# router while bypassing the other's qualifies for both. Each is separately
# budgeted, so both fire.
for job in (await self._active_jobs()).get(str(api_key_hash), ()):
if datetime.now(timezone.utc) >= job.ends_at:
continue
if job.attempts + self._job_starts.get(job.id, 0) >= job.max_turns:
continue
if not _sample_hits(request_id, job.id, job.shadow_percentage):
continue
if _request_was_routed_by(request_metadata, job.router_name) != (job.direction == "reverse"):
continue
if self._inflight_shadow_tasks >= _MAX_CONCURRENT_SHADOW_TASKS:
return
self._job_starts[job.id] = self._job_starts.get(job.id, 0) + 1
self._inflight_shadow_tasks += 1
asyncio.create_task(
self._run_shadow_eval(
job=job,
request_id=request_id,
messages=messages,
response_obj=response_obj,
real_model=payload.get("model") or "",
control_tier=control_tier,
model_parameters=MappingProxyType(
dict(payload.get("model_parameters") or {}) # mutable-ok: frozen snapshot
),
parent_metadata=MappingProxyType(dict(request_metadata)), # mutable-ok: frozen snapshot
)
).add_done_callback(self._release_shadow_slot)
except Exception as e: # noqa: BLE001 # logging hooks must never fail the request
verbose_logger.debug("shadow_eval: failed to schedule task: %s", e)
def _release_shadow_slot(self, _task: "asyncio.Task[None]") -> None:
self._inflight_shadow_tasks -= 1
#### the detached pipeline: one attempt row per sampled request, verdict or error ####
async def _run_shadow_eval(
self,
job: ActiveShadowEvalJob,
request_id: str,
messages: Sequence[Mapping[str, object]],
response_obj: object,
real_model: str,
control_tier: str | None,
model_parameters: Mapping[str, object],
parent_metadata: Mapping[str, object],
) -> None:
"""Budget gate -> shadow call -> blind judge -> one attempt row. The prisma gate
sits above the dispatch so no provider spend happens without a place to record
the outcome, and the budget read lives here rather than in the success hook."""
prisma: Final = self._prisma_provider()
try:
if prisma is None:
return
real_text: Final = self._extract_response_text(response_obj)
if not real_text or not messages:
return
if await _key_or_team_is_over_budget(parent_metadata):
return
shadow: Final = await self._call_router_shadow(
job.shadow_target, messages, model_parameters, parent_metadata
)
if isinstance(shadow, _CallFailure):
await self._record_attempt(prisma, job, request_id, control_tier, outcome="error", error=shadow.error)
return
verdict: Final = await self._call_judge(
judge_model=job.judge_model,
messages=messages,
real_text=real_text,
shadow_text=shadow.text,
parent_metadata=parent_metadata,
)
if isinstance(verdict, _CallFailure):
await self._record_attempt(
prisma,
job,
request_id,
control_tier,
outcome="error",
error=verdict.error,
shadow=shadow,
judge_cost=verdict.cost,
)
return
await self._record_attempt(
prisma,
job,
request_id,
control_tier,
outcome=verdict.preference,
shadow=shadow,
real_model=real_model,
confidence=verdict.confidence,
judge_cost=verdict.cost,
)
except Exception as e: # noqa: BLE001 # detached task: record what happened, never raise
verbose_logger.debug("shadow_eval: pipeline failed for %s: %s", request_id, e)
await self._record_attempt(
prisma, job, request_id, control_tier, outcome="error", error=f"pipeline error: {e}"
)
@staticmethod
async def _record_attempt(
prisma: "PrismaClient | None",
job: ActiveShadowEvalJob,
request_id: str,
control_tier: str | None,
*,
outcome: str,
shadow: _ShadowResponse | None = None,
real_model: str = "",
confidence: float | None = None,
judge_cost: float = 0.0,
error: str | None = None,
) -> None:
if prisma is None:
return
try:
await prisma.db.litellm_shadowevalattempt.create(
data={ # mutable-ok: Prisma payload
"job_id": job.id,
"request_id": request_id,
"outcome": outcome,
"tier": control_tier if job.direction == "reverse" else (shadow.tier if shadow else None),
"real_model": real_model or None,
"shadow_model": shadow.model if shadow else None,
"confidence": confidence,
"judge_cost": judge_cost,
"error": error[:_MAX_ERROR_CHARS] if error else None,
}
)
except Exception as e: # noqa: BLE001 # a lost row degrades sample size, nothing can disagree with it
verbose_logger.debug("shadow_eval: attempt write failed for %s: %s", request_id, e)
async def _call_router_shadow(
self,
target_model: str,
messages: Sequence[Mapping[str, object]],
model_parameters: Mapping[str, object],
parent_metadata: Mapping[str, object],
) -> "_ShadowResponse | _CallFailure":
"""Send the prompt through the arm nobody was served: the auto-router under
evaluation, or a reverse job's fixed baseline model. The metadata carries the
shadowed key's identity (spend attribution) and receives a routing decision
write-back, which a plain baseline model simply never makes."""
router: Final = self._router_provider()
if router is None:
return _CallFailure("no router configured on this pod")
shadow_metadata: Final[dict[str, object]] = ( # mutable-ok: router writes its routing decision back
sanitized_forwardable_call_metadata(parent_metadata, SHADOW_EVAL_ROUTER_CALL_ORIGIN)
)
shadow_params: Final = { # mutable-ok: splatted as kwargs
k: v for k, v in model_parameters.items() if k not in ("stream", "metadata")
}
try:
response: Final = await router.acompletion(
model=target_model,
messages=messages, # pyright: ignore[reportArgumentType] # snapshot of the SDK's own message dicts
metadata=shadow_metadata,
num_retries=0,
fallbacks=[], # mutable-ok: SDK kwarg; a failed shadow is a recorded error, never a spend multiplier
**shadow_params,
)
except Exception as e: # noqa: BLE001 # provider errors become error rows, not crashes
verbose_logger.debug("shadow_eval: router call failed: %s", e)
return _CallFailure(f"shadow router call failed: {e}")
text: Final = self._extract_response_text(response)
if not text:
return _CallFailure("shadow router returned an empty response")
return _ShadowResponse(
text=text,
model=str(getattr(response, "model", None) or _routing_decision(shadow_metadata).get("routed_model") or ""),
tier=_routed_tier(shadow_metadata),
)
async def _call_judge(
self,
judge_model: str,
messages: Sequence[Mapping[str, object]],
real_text: str,
shadow_text: str,
parent_metadata: Mapping[str, object],
) -> "_JudgeVerdict | _CallFailure":
"""Blind pairwise judge with A/B labels randomized to cancel position bias."""
real_is_a: Final = random.random() < 0.5
response_a: Final = real_text if real_is_a else shadow_text
response_b: Final = shadow_text if real_is_a else real_text
conversation: Final = "\n".join(
f"{str(m.get('role', 'user')).upper()}: {extract_text_from_content(m.get('content'))}"
for m in messages
if m.get("content") is not None
)
judge_metadata: Final = sanitized_forwardable_call_metadata(parent_metadata, SHADOW_EVAL_JUDGE_CALL_ORIGIN)
judge_messages: Final = [ # mutable-ok: SDK takes a list
{"role": "system", "content": PAIRWISE_JUDGE_SYSTEM_PROMPT}, # mutable-ok: SDK message
{
"role": "user",
"content": _judge_user_prompt(conversation, response_a, response_b),
}, # mutable-ok: SDK message
]
try:
response: Final = await judge_acompletion(
self._router_provider(),
judge_model,
judge_messages, # pyright: ignore[reportArgumentType] # plain SDK message dicts
temperature=0,
max_tokens=JUDGE_MAX_OUTPUT_TOKENS,
metadata=judge_metadata,
)
except Exception as e: # noqa: BLE001 # judge outages become error rows, not crashes
verbose_logger.debug("shadow_eval: judge call failed: %s", e)
return _CallFailure(f"judge call failed: {e}")
try:
raw: Final = response["choices"][0]["message"]["content"] or ""
verdict: Final = PairwiseVerdict.model_validate(parse_json_verdict(raw))
except Exception as e: # noqa: BLE001 # malformed verdicts become error rows
verbose_logger.debug("shadow_eval: unparseable judge verdict: %s", e)
return _CallFailure(f"unparseable judge verdict: {e}", cost=_judge_call_cost(response))
return _JudgeVerdict(
preference=_unmask_preference(verdict.preference, real_is_a),
confidence=max(0.0, min(1.0, verdict.confidence)),
cost=_judge_call_cost(response),
)
@staticmethod
def _extract_response_text(response_obj: object) -> str:
"""Extract the assistant's text from a ModelResponse-shaped object or dict."""
try:
content: Final = (
response_obj["choices"][0]["message"]["content"]
if isinstance(response_obj, Mapping)
else response_obj.choices[0].message.content # pyright: ignore[reportAttributeAccessIssue] # duck-typed ModelResponse
)
except (AttributeError, KeyError, IndexError, TypeError):
return ""
return extract_text_from_content(content)
_EMPTY_JOBS: Final[Mapping[str, tuple[ActiveShadowEvalJob, ...]]] = MappingProxyType({})
def _default_prisma_provider() -> "PrismaClient | None":
try:
from litellm.proxy.proxy_server import prisma_client
except ImportError:
return None
return prisma_client

38
litellm/interactions/litellm_responses_transformation/handler.py
View file

@ -2,8 +2,8 @@
Handler for transforming interactions API requests to litellm.responses requests.
"""
from collections.abc import AsyncIterator, Coroutine, Iterator
from typing import Any, Final, cast
from collections.abc import AsyncIterator, Callable, Coroutine, Iterator
from typing import Any, Final
import litellm
from litellm.interactions.litellm_responses_transformation.streaming_iterator import (
@ -37,7 +37,7 @@ class LiteLLMResponsesInteractionsHandler:
) -> (
InteractionsAPIResponse
| Iterator[InteractionsAPIStreamingResponse]
| Coroutine[Any, Any, InteractionsAPIResponse | AsyncIterator[InteractionsAPIStreamingResponse]]
| Coroutine[object, object, InteractionsAPIResponse | AsyncIterator[InteractionsAPIStreamingResponse]]
):
"""
Handle Interactions API request by calling litellm.responses().
@ -55,13 +55,15 @@ class LiteLLMResponsesInteractionsHandler:
InteractionsAPIResponse or streaming iterator
"""
# Transform interactions request to responses request
responses_request = LiteLLMResponsesInteractionsConfig.transform_interactions_request_to_responses_request(
model=model,
input=input,
optional_params=optional_params,
custom_llm_provider=custom_llm_provider,
stream=stream,
**kwargs,
responses_request: Final = (
LiteLLMResponsesInteractionsConfig.transform_interactions_request_to_responses_request(
model=model,
input=input,
optional_params=optional_params,
custom_llm_provider=custom_llm_provider,
stream=stream,
**kwargs,
)
)
if _is_async:
@ -76,7 +78,10 @@ class LiteLLMResponsesInteractionsHandler:
# Call litellm.responses()
# Note: litellm.responses() returns Union[ResponsesAPIResponse, BaseResponsesAPIStreamingIterator]
# but the type checker may see it as a coroutine in some contexts
responses_response: Final = litellm.responses(
responses_fn: Final[Callable[..., ResponsesAPIResponse | BaseResponsesAPIStreamingIterator]] = vars(litellm)[
"responses"
]
responses_response: Final = responses_fn(
**responses_request,
)
@ -92,8 +97,7 @@ class LiteLLMResponsesInteractionsHandler:
)
# At this point, responses_response must be ResponsesAPIResponse (not streaming)
# Cast to satisfy type checker since we've already checked it's not a streaming iterator
responses_api_response: Final = cast(ResponsesAPIResponse, responses_response)
responses_api_response: Final = responses_response
# Transform responses response to interactions response
return LiteLLMResponsesInteractionsConfig.transform_responses_response_to_interactions_response(
@ -112,7 +116,10 @@ class LiteLLMResponsesInteractionsHandler:
"""Async handler for interactions API requests."""
# Call litellm.aresponses()
# Note: litellm.aresponses() returns Union[ResponsesAPIResponse, BaseResponsesAPIStreamingIterator]
responses_response: Final = await litellm.aresponses(
aresponses_fn: Final[
Callable[..., Coroutine[object, object, ResponsesAPIResponse | BaseResponsesAPIStreamingIterator]]
] = vars(litellm)["aresponses"]
responses_response: Final = await aresponses_fn(
**responses_request,
)
@ -128,8 +135,7 @@ class LiteLLMResponsesInteractionsHandler:
)
# At this point, responses_response must be ResponsesAPIResponse (not streaming)
# Cast to satisfy type checker since we've already checked it's not a streaming iterator
responses_api_response: Final = cast(ResponsesAPIResponse, responses_response)
responses_api_response: Final = responses_response
# Transform responses response to interactions response
return LiteLLMResponsesInteractionsConfig.transform_responses_response_to_interactions_response(

154
litellm/interactions/litellm_responses_transformation/transformation.py
View file

@ -2,12 +2,16 @@
Transformation utilities for bridging Interactions API to Responses API.
This module handles transforming between:
- Interactions API format (Google's format with Turn[], system_instruction, etc.)
- Interactions API format (Google's format with Step[]/Turn[], system_instruction, etc.)
- Responses API format (OpenAI's format with input[], instructions, etc.)
"""
from collections.abc import Mapping, Sequence
from types import MappingProxyType
from typing import Any, Final, cast
from pydantic import BaseModel
from litellm.types.interactions import (
InteractionInput,
InteractionsAPIOptionalRequestParams,
@ -19,6 +23,8 @@ from litellm.types.llms.openai import (
ResponsesAPIResponse,
)
_STEP_TYPE_ROLES: Final = MappingProxyType({"user_input": "user", "model_output": "assistant"})
class LiteLLMResponsesInteractionsConfig:
"""Configuration class for transforming between Interactions API and Responses API."""
@ -91,112 +97,94 @@ class LiteLLMResponsesInteractionsConfig:
Interactions API input can be:
- string: "Hello"
- Turn[]: [{"role": "user", "content": [...]}]
- Content object
- Step[]: [{"type": "user_input", "content": [...]}, {"type": "model_output", "content": [...]}]
- Turn[] (legacy): [{"role": "user", "content": [...]}]
- Content | Content[]: one user message worth of content parts
Responses API input is:
- string: "Hello"
- Message[]: [{"role": "user", "content": [...]}]
- Message[]: [{"role": "user", "content": [{"type": "input_text", ...}]}]
"""
if isinstance(input, str):
# ResponseInputParam accepts str
return cast(ResponseInputParam, input)
if isinstance(input, list):
# Turn[] format - convert to Responses API Message[] format
messages: Final = []
for turn in input:
if isinstance(turn, dict):
role = turn.get("role", "user")
content = turn.get("content", [])
transformed: Final = (
[
LiteLLMResponsesInteractionsConfig._transform_history_item(item)
for item in input
if LiteLLMResponsesInteractionsConfig._is_history_item(item)
]
if any(LiteLLMResponsesInteractionsConfig._is_history_item(item) for item in input)
else [
{
"role": "user",
"content": LiteLLMResponsesInteractionsConfig._transform_content_array(input, "user"),
}
]
)
return cast(ResponseInputParam, transformed)
# Transform content array
transformed_content = LiteLLMResponsesInteractionsConfig._transform_content_array(content)
messages.append(
{
"role": role,
"content": transformed_content,
}
)
elif isinstance(turn, Turn):
# Pydantic model
role = turn.role if hasattr(turn, "role") else "user"
content = turn.content if hasattr(turn, "content") else []
# Ensure content is a list for _transform_content_array
# Cast to List[Any] to handle various content types
if isinstance(content, list):
content_list: list[Any] = list(content)
elif content is not None:
content_list = [content]
else:
content_list = []
transformed_content = LiteLLMResponsesInteractionsConfig._transform_content_array(content_list)
messages.append(
{
"role": role,
"content": transformed_content,
}
)
return cast(ResponseInputParam, messages)
# Single content object - wrap in message
if isinstance(input, dict):
raw_content: Final = input.get("content")
content_items: Final = raw_content if isinstance(raw_content, list) else [input]
return cast(
ResponseInputParam,
[
{
"role": "user",
"content": LiteLLMResponsesInteractionsConfig._transform_content_array(
input.get("content", []) if isinstance(input.get("content"), list) else [input]
),
"content": LiteLLMResponsesInteractionsConfig._transform_content_array(content_items, "user"),
}
],
)
# Fallback: convert to string
return cast(ResponseInputParam, str(input))
@staticmethod
def _transform_content_array(content: list[Any]) -> list[dict[str, Any]]:
"""Transform Interactions API content array to Responses API format."""
if not isinstance(content, list):
# Single content item - wrap in array
content = [content]
def _is_history_item(item: object) -> bool:
if isinstance(item, Turn):
return True
return isinstance(item, dict) and ("role" in item or item.get("type") in _STEP_TYPE_ROLES)
transformed: Final[list[dict[str, Any]]] = []
for item in content:
if isinstance(item, dict):
# Already in dict format, pass through
transformed.append(item)
elif isinstance(item, str):
# Plain string - wrap in text format
transformed.append({"type": "text", "text": item})
else:
# Pydantic model or other - convert to dict
if hasattr(item, "model_dump"):
dumped = item.model_dump()
if isinstance(dumped, dict):
transformed.append(dumped)
else:
# Fallback: wrap in text format
transformed.append({"type": "text", "text": str(dumped)})
elif hasattr(item, "dict"):
dumped = item.dict()
if isinstance(dumped, dict):
transformed.append(dumped)
else:
# Fallback: wrap in text format
transformed.append({"type": "text", "text": str(dumped)})
else:
# Fallback: wrap in text format
transformed.append({"type": "text", "text": str(item)})
@staticmethod
def _transform_history_item(item: object) -> Mapping[str, object]:
raw: Final = item.model_dump(exclude_none=True) if isinstance(item, Turn) else item
fields: Final = raw if isinstance(raw, Mapping) else {}
role: Final = LiteLLMResponsesInteractionsConfig._responses_role(fields)
raw_content: Final = fields.get("content")
content_items: Final = (
raw_content if isinstance(raw_content, list) else [] if raw_content is None else [raw_content]
)
return {
"role": role,
"content": LiteLLMResponsesInteractionsConfig._transform_content_array(content_items, role),
}
return transformed
@staticmethod
def _responses_role(item: Mapping[str, object]) -> str:
step_role: Final = _STEP_TYPE_ROLES.get(str(item.get("type", "")))
if step_role is not None:
return step_role
raw_role: Final = str(item.get("role") or "user")
return "assistant" if raw_role == "model" else raw_role
@staticmethod
def _transform_content_array(content: Sequence[object], role: str) -> Sequence[Mapping[str, object]]:
"""Transform Interactions API content parts to Responses API parts for the given role."""
return [LiteLLMResponsesInteractionsConfig._transform_content_item(item, role) for item in content]
@staticmethod
def _transform_content_item(item: object, role: str) -> Mapping[str, object]:
text_type: Final = "output_text" if role == "assistant" else "input_text"
if isinstance(item, str):
return {"type": text_type, "text": item}
if isinstance(item, Mapping):
if item.get("type") == "text":
return {"type": text_type, "text": str(item.get("text", ""))}
return item
if isinstance(item, BaseModel):
return LiteLLMResponsesInteractionsConfig._transform_content_item(item.model_dump(exclude_none=True), role)
return {"type": text_type, "text": str(item)}
@staticmethod
def transform_responses_response_to_interactions_response(

4
litellm/litellm_core_utils/core_helpers.py
View file

@ -1,7 +1,7 @@
# What is this?
## Helper utilities
import copy
from collections.abc import Iterable
from collections.abc import Iterable, Mapping
from typing import TYPE_CHECKING, Any, Final, Literal
import httpx
@ -181,7 +181,7 @@ def add_missing_spend_metadata_to_litellm_metadata(litellm_metadata: dict, metad
def get_metadata_variable_name_from_kwargs(
kwargs: dict,
kwargs: Mapping[str, object],
) -> Literal["metadata", "litellm_metadata"]:
"""
Helper to return what the "metadata" field should be called in the request data

15
litellm/litellm_core_utils/exception_mapping_utils.py
View file

@ -34,12 +34,16 @@ class ExceptionCheckers:
"""
@staticmethod
def is_error_str_rate_limit(error_str: str) -> bool:
def is_error_str_rate_limit(error_str: str, status_code: int | None = None) -> bool:
"""
Check if an error string indicates a rate limit error.
Args:
error_str: The error string to check
status_code: The HTTP status the provider returned, when known. Gates only the
bare-number branch: providers echo the request back in validation errors and
429 is an ordinary token id, so an echoed prompt can put a standalone 429 in
the body of a 400. The phrase branches stay ungated (#11455).
Returns:
True if the error indicates a rate limit, False otherwise
@ -47,8 +51,9 @@ class ExceptionCheckers:
if not isinstance(error_str, str):
return False
# Only treat 429 as a rate limit signal when it appears as a standalone token
if re.search(r"\b429\b", error_str):
# A standalone 429 counts unless the provider's own status says otherwise. The
# status is read off an arbitrary exception, so a non-integer means "unknown".
if re.search(r"\b429\b", error_str) and (not isinstance(status_code, int) or status_code == 429):
return True
_error_str_lower: Final = error_str.lower()
@ -280,7 +285,9 @@ def _map_openai_exception(
else:
exception_provider = custom_llm_provider[0].upper() + custom_llm_provider[1:] + "Exception"
if ExceptionCheckers.is_error_str_rate_limit(error_str):
if ExceptionCheckers.is_error_str_rate_limit(
error_str, status_code=getattr(original_exception, "status_code", None)
):
raise RateLimitError(
message=f"RateLimitError: {exception_provider} - {message}",
model=model,

94
litellm/litellm_core_utils/internal_call_metadata.py Normal file
View file

@ -0,0 +1,94 @@
"""Metadata a request forwards to the internal LLM sub-calls it triggers.
Internal features (the auto-router's classifier and embeddings, shadow eval's shadow and
judge calls) bill real provider spend that nobody typed a prompt for. That spend must land
on the same key/team/org/user as the request that caused it, so the sub-call carries the
caller's identity metadata, minus two things that must never be forwarded as-is:
* ``user_api_key_budget_reservation`` (and the reservation nested inside
``user_api_key_auth``) belongs to the parent completion. If a sub-call's cost callback
sees it, that callback finalizes the reservation and the parent's own callback then
skips incrementing the key/team budget counters, losing the parent's spend.
``user_api_key_auth`` itself is kept, sanitized, because model access-group filtering
needs it.
* The sub-call is stamped with ``INTERNAL_CALL_ORIGIN_METADATA_KEY`` so its spend log row
records that it is not traffic the caller sent.
"""
from __future__ import annotations
from collections.abc import Mapping
from typing import Final
from litellm.constants import INTERNAL_CALL_ORIGIN_METADATA_KEY
from litellm.types.utils import InternalCallOrigin
BUDGET_RESERVATION_METADATA_KEYS: Final = frozenset({"user_api_key_budget_reservation"})
_USER_API_KEY_AUTH_KEY: Final = "user_api_key_auth"
FORWARDABLE_IDENTITY_METADATA_KEYS: Final = frozenset(
{
"user_api_key",
"user_api_key_hash",
"user_api_key_alias",
"user_api_key_team_id",
"user_api_key_org_id",
"user_api_key_user_id",
"user_api_key_end_user_id",
_USER_API_KEY_AUTH_KEY,
}
)
"""The caller-identity subset a detached sub-call needs to be attributed and
budget-checked like the request that spawned it. Everything else on the parent's metadata
(routing decision, guardrail state, logging payload) describes the parent call and would
be a lie on a sub-call that runs after it returned."""
def sanitize_user_api_key_auth(auth: object) -> object:
"""Copy of the auth object with its budget reservation removed; the cost callback
falls back to reading the reservation from inside the auth object."""
if isinstance(auth, dict):
return {k: v for k, v in auth.items() if k != "budget_reservation"} # mutable-ok: SDK metadata value
reservation: Final[object] = getattr(auth, "budget_reservation", None)
model_copy: Final[object] = getattr(auth, "model_copy", None)
if reservation is not None and callable(model_copy):
return model_copy(update={"budget_reservation": None}) # mutable-ok: pydantic update payload
return auth
def _sanitized(parent_metadata: Mapping[str, object]) -> dict[str, object]: # mutable-ok: SDK metadata kwarg
return { # mutable-ok: SDK metadata kwarg
k: sanitize_user_api_key_auth(v) if k == _USER_API_KEY_AUTH_KEY else v
for k, v in parent_metadata.items()
if k not in BUDGET_RESERVATION_METADATA_KEYS
}
def forwarded_internal_call_metadata(
parent_metadata: Mapping[str, object] | None,
call_origin: InternalCallOrigin,
) -> dict[str, object]: # mutable-ok: SDK metadata kwarg
"""Parent metadata, minus its budget reservation, stamped with the sub-call's origin.
For sub-calls made inside the parent request (classifier, embeddings), where the
parent's full context still describes the call being made.
"""
if not parent_metadata:
return {} # mutable-ok: SDK metadata kwarg
return _sanitized(parent_metadata) | { # mutable-ok: SDK metadata kwarg
INTERNAL_CALL_ORIGIN_METADATA_KEY: call_origin
}
def sanitized_forwardable_call_metadata(
parent_metadata: Mapping[str, object],
call_origin: InternalCallOrigin,
) -> dict[str, object]: # mutable-ok: SDK metadata kwarg
"""Just the caller's identity, stamped with the sub-call's origin.
For sub-calls detached from the parent request (shadow eval), which outlive it and
must not inherit per-request state such as its routing decision or logging payload.
"""
identity: Final = {k: v for k, v in parent_metadata.items() if k in FORWARDABLE_IDENTITY_METADATA_KEYS}
return _sanitized(identity) | {INTERNAL_CALL_ORIGIN_METADATA_KEY: call_origin} # mutable-ok: SDK metadata kwarg

25
litellm/litellm_core_utils/llm_cost_calc/utils.py
View file

@ -743,6 +743,23 @@ def _get_regional_uplift_multiplier(model_info: ModelInfo, data_residency: str |
return 1.0
def get_provider_specific_geo_multiplier(model_info: ModelInfo, usage: Usage) -> float:
"""
Resolve the provider-specific regional pricing multiplier for the geo the
request was served from (``usage.inference_geo``), e.g. Anthropic's ``us: 1.1``
stored under ``provider_specific_entry``. The regional surcharge applies to
every token type, so per-type cost breakdowns must scale by it too.
Returns 1.0 when the request was served globally or the model carries no
multiplier for the geo.
"""
inference_geo: Final = getattr(usage, "inference_geo", None)
if not isinstance(inference_geo, str) or inference_geo.lower() in ("global", "not_available"):
return 1.0
provider_specific_entry: Final[dict[str, float]] = model_info.get("provider_specific_entry") or {}
return float(provider_specific_entry.get(inference_geo.lower(), 1.0))
def _resolve_reasoning_token_cost(
model_info: ModelInfo,
service_tier: str | None,
@ -1040,6 +1057,14 @@ def get_token_type_cost_breakdown(
cache_read_cost *= uplift
cache_creation_cost *= uplift
# Mirror the provider-specific geo uplift (e.g. Anthropic us: 1.1) the totals
# apply, so cache and reasoning line items stay reconciled with them.
geo_multiplier: Final = get_provider_specific_geo_multiplier(model_info=model_info, usage=usage)
if geo_multiplier != 1.0:
reasoning_cost *= geo_multiplier
cache_read_cost *= geo_multiplier
cache_creation_cost *= geo_multiplier
return TokenTypeCostBreakdown(
reasoning_cost=reasoning_cost,
cache_read_cost=cache_read_cost,

87
litellm/litellm_core_utils/llm_judge.py Normal file
View file

@ -0,0 +1,87 @@
"""Shared primitives for LLM-judge features (llm_as_a_judge guardrail, shadow eval)."""
from __future__ import annotations
import json
import re
from typing import TYPE_CHECKING, Final
import litellm
if TYPE_CHECKING:
from litellm import Router
from litellm.types.llms.openai import AllMessageValues
from litellm.types.utils import ModelResponse
JSON_FENCE_RE: Final = re.compile(r"```(?:json)?\s*(.*?)\s*```", re.DOTALL | re.IGNORECASE)
def default_router_provider() -> Router | None:
try:
from litellm.proxy.proxy_server import llm_router
except ImportError:
return None
return llm_router
def parse_json_verdict(raw: str) -> dict[str, object]: # mutable-ok: plain parsed-JSON payload
"""Parse a judge's JSON verdict, tolerating markdown fences and surrounding prose."""
text = raw.strip() # rebind-ok: progressively narrowed to the JSON payload
fenced: Final = JSON_FENCE_RE.search(text)
if fenced is not None:
text = fenced.group(1).strip() # rebind-ok: progressively narrowed to the JSON payload
parsed: object
try:
parsed = json.loads(text)
except json.JSONDecodeError:
start: Final = text.find("{")
end: Final = text.rfind("}")
if start == -1 or end <= start:
raise
parsed = json.loads(text[start : end + 1])
if not isinstance(parsed, dict):
raise ValueError("judge response is not a JSON object")
return {str(k): v for k, v in parsed.items()} # mutable-ok: plain parsed-JSON payload
def extract_text_from_content(content: object) -> str:
"""Return plain text from a message content field (str or multimodal list)."""
if isinstance(content, str):
return content
if isinstance(content, list):
return " ".join(
str(part.get("text", "")) for part in content if isinstance(part, dict) and part.get("type") == "text"
)
return ""
def router_resolves_model(router: Router | None, model: str) -> bool:
"""Whether the model name resolves through the proxy's router (configured deployment
or model-group alias), the same check the judge dispatch itself makes, so start-time
validation cannot accept a name the call path then fails on."""
return router is not None and bool(model in router.model_group_alias or router.get_model_list(model_name=model))
async def judge_acompletion(
router: Router | None,
judge_model: str,
messages: list[AllMessageValues], # mutable-ok: the SDK acompletion signature takes a list
**params: object,
) -> ModelResponse:
"""Dispatch a judge call through the proxy's router when the judge model is a
configured deployment (DB-stored credentials work), through the SDK for
provider-qualified public names. The router path never retries or falls back:
a failed judge call is the caller's counted failure, not a spend multiplier.
Sampling preferences are advisory: models that removed sampling params (e.g.
claude-sonnet-5) drop them instead of rejecting the judge call."""
if router_resolves_model(router, judge_model):
return await router.acompletion( # pyright: ignore[reportOptionalMemberAccess] # router_resolves_model implies router is not None
model=judge_model,
messages=messages,
num_retries=0,
fallbacks=[],
drop_params=True,
**params,
)
return await litellm.acompletion(model=judge_model, messages=messages, num_retries=0, drop_params=True, **params)

40
litellm/llms/anthropic/chat/transformation.py
View file

@ -1,6 +1,7 @@
import json
import re
import time
from collections.abc import Mapping, Sequence
from typing import TYPE_CHECKING, Any, Final, NoReturn, cast
import httpx
@ -2117,6 +2118,37 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
return False
return any(key in usage_object for key in ("cache_read_input_tokens", "cache_creation_input_tokens"))
@staticmethod
def _aggregate_cache_creation_token_details(
iterations: Sequence[Mapping[str, Any]],
) -> CacheCreationTokenDetails | None:
breakdowns: Final = tuple(c for c in (it.get("cache_creation") for it in iterations) if isinstance(c, Mapping))
if not breakdowns:
return None
detailed_5m: Final = sum(int(c.get("ephemeral_5m_input_tokens") or 0) for c in breakdowns)
detailed_1h: Final = sum(int(c.get("ephemeral_1h_input_tokens") or 0) for c in breakdowns)
total: Final = sum(int(it.get("cache_creation_input_tokens") or 0) for it in iterations)
undetailed: Final = max(total - detailed_5m - detailed_1h, 0)
return CacheCreationTokenDetails(
ephemeral_5m_input_tokens=detailed_5m + undetailed,
ephemeral_1h_input_tokens=detailed_1h,
)
@staticmethod
def _resolve_cache_creation_token_details(usage: Mapping[str, Any]) -> CacheCreationTokenDetails | None:
iterations: Final = usage.get("iterations")
if iterations:
aggregated: Final = AnthropicConfig._aggregate_cache_creation_token_details(iterations)
if aggregated is not None:
return aggregated
cache_creation: Final = usage.get("cache_creation")
if not isinstance(cache_creation, Mapping):
return None
return CacheCreationTokenDetails(
ephemeral_5m_input_tokens=cache_creation.get("ephemeral_5m_input_tokens"),
ephemeral_1h_input_tokens=cache_creation.get("ephemeral_1h_input_tokens"),
)
def calculate_usage(
self,
usage_object: dict,
@ -2132,7 +2164,7 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
_usage: Final = usage_object
cache_creation_input_tokens: int = 0
cache_read_input_tokens: int = 0
cache_creation_token_details: CacheCreationTokenDetails | None = None
cache_creation_token_details: Final = self._resolve_cache_creation_token_details(_usage)
web_search_requests: int | None = None
tool_search_requests: int | None = None
inference_geo: str | None = None
@ -2182,12 +2214,6 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
if tool_search_count > 0:
tool_search_requests = tool_search_count
if "cache_creation" in _usage and _usage["cache_creation"] is not None:
cache_creation_token_details = CacheCreationTokenDetails(
ephemeral_5m_input_tokens=_usage["cache_creation"].get("ephemeral_5m_input_tokens"),
ephemeral_1h_input_tokens=_usage["cache_creation"].get("ephemeral_1h_input_tokens"),
)
raw_input_tokens: Final = prompt_tokens - cache_read_input_tokens - cache_creation_input_tokens
prompt_tokens_details: Final = PromptTokensDetailsWrapper(
cached_tokens=cache_read_input_tokens,

37
litellm/llms/anthropic/common_utils.py
View file

@ -5,6 +5,7 @@ This file contains common utils for anthropic calls.
import copy
import re
from collections.abc import Mapping, Sequence
from datetime import datetime, timezone
from types import MappingProxyType
from typing import Any, Final, Literal
@ -12,6 +13,7 @@ import httpx
from pydantic import BaseModel, ConfigDict, TypeAdapter, ValidationError
import litellm
from litellm.constants import DEFAULT_MODEL_CREATED_AT_TIME
from litellm.litellm_core_utils.prompt_templates.common_utils import (
get_file_ids_from_messages,
)
@ -28,6 +30,7 @@ from litellm.types.llms.anthropic import (
AnthropicMcpServerTool,
)
from litellm.types.llms.openai import AllMessageValues
from litellm.types.proxy.model_listing import ModelInfoResponse
_BEDROCK_VERSION_SUFFIX_RE: Final = re.compile(r"-v\d+(?::\d+)?$")
_INFERENCE_PROFILE_MINOR_RE: Final = re.compile(r":\d+$")
@ -1221,3 +1224,37 @@ def process_anthropic_headers(headers: httpx.Headers | dict) -> dict:
additional_headers: Final = {**llm_response_headers, **openai_headers}
return additional_headers
def _anthropic_model_entry(model: ModelInfoResponse, created_at: str) -> Mapping[str, object]:
return { # mutable-ok: JSON response body, serialized by the route and never mutated
"type": "model",
"id": model["id"],
"display_name": model["id"],
"created_at": created_at,
"max_input_tokens": model.get("max_input_tokens"),
"max_tokens": model.get("max_output_tokens"),
}
def create_anthropic_model_list_response(models: Sequence[ModelInfoResponse]) -> Mapping[str, object]:
"""Build the Anthropic-native /v1/models envelope.
Clients that send an anthropic-version header parse the Anthropic Models API
shape (type/display_name/created_at plus has_more/first_id/last_id) and filter
the list themselves, so every model is returned here. The token limits carry
over from the OpenAI-shaped listing, named as the Messages API names them, and
are always present because the vendor shape declares them nullable, not optional
"""
created_at: Final = (
datetime.fromtimestamp(DEFAULT_MODEL_CREATED_AT_TIME, tz=timezone.utc).isoformat().replace("+00:00", "Z")
)
data: Final = [ # mutable-ok: JSON response body, serialized by the route and never mutated
_anthropic_model_entry(model, created_at) for model in models
]
return { # mutable-ok: JSON response body, serialized by the route and never mutated
"data": data,
"has_more": False,
"first_id": models[0]["id"] if models else None,
"last_id": models[-1]["id"] if models else None,
}

29
litellm/llms/anthropic/cost_calculation.py
View file

@ -13,6 +13,7 @@ from litellm.litellm_core_utils.llm_cost_calc.utils import (
_parse_prompt_tokens_details,
calculate_cache_writing_cost,
generic_cost_per_token,
get_provider_specific_geo_multiplier,
)
if TYPE_CHECKING:
@ -24,9 +25,10 @@ def _compute_cache_only_cost(model_info: "ModelInfo", usage: "Usage", service_ti
"""
Return only the cache-related portion of the prompt cost (cache read + cache write).
These costs must NOT be scaled by geo/speed multipliers because the old
These costs must NOT be scaled by the ``fast`` speed multiplier because the old
explicit ``fast/`` model entries carried unchanged cache rates while
multiplying only the regular input/output token costs.
multiplying only the regular input/output token costs. Regional pricing, by
contrast, uplifts every token type, so the geo multiplier does scale them.
"""
if usage.prompt_tokens_details is None:
return 0.0
@ -81,20 +83,19 @@ def cost_per_token(model: str, usage: "Usage", service_tier: str | None = None)
model_info: Final = litellm.get_model_info(model=model, custom_llm_provider="anthropic")
provider_specific_entry: Final[dict] = model_info.get("provider_specific_entry") or {}
multiplier = 1.0
if (
hasattr(usage, "inference_geo")
and usage.inference_geo
and usage.inference_geo.lower() not in ["global", "not_available"]
):
multiplier *= provider_specific_entry.get(usage.inference_geo.lower(), 1.0)
if hasattr(usage, "speed") and usage.speed == "fast":
multiplier *= provider_specific_entry.get("fast", 1.0)
geo_multiplier: Final = get_provider_specific_geo_multiplier(model_info=model_info, usage=usage)
speed_multiplier: Final = (
provider_specific_entry.get("fast", 1.0) if getattr(usage, "speed", None) == "fast" else 1.0
)
if multiplier != 1.0:
if speed_multiplier != 1.0:
cache_cost: Final = _compute_cache_only_cost(model_info=model_info, usage=usage, service_tier=service_tier)
prompt_cost = (prompt_cost - cache_cost) * multiplier + cache_cost
completion_cost *= multiplier
prompt_cost = (prompt_cost - cache_cost) * speed_multiplier + cache_cost
completion_cost *= speed_multiplier
if geo_multiplier != 1.0:
prompt_cost *= geo_multiplier
completion_cost *= geo_multiplier
except Exception:
pass

66
litellm/llms/anthropic/experimental_pass_through/context_management/editors/compact.py
View file

@ -13,8 +13,10 @@ Mirrors Anthropic's native ``compact_20260112`` for non-Anthropic providers:
"""
import re
from collections.abc import Mapping
from typing import TYPE_CHECKING, Any, Final, Literal, Optional, Union, cast
from collections.abc import Mapping, Sequence
from typing import TYPE_CHECKING, Any, Final, Literal, NotRequired, Optional, TypedDict, Union, cast
from typing_extensions import ReadOnly
import litellm
from litellm._logging import verbose_logger
@ -29,9 +31,8 @@ if TYPE_CHECKING:
from litellm.proxy._types import UserAPIKeyAuth
from litellm.router import Router
from litellm.types.llms.anthropic import (
AllAnthropicPassThroughMessageValues,
AllAnthropicToolsValues,
AnthopicMessagesAssistantMessageParam,
AnthropicMessagesUserMessageParam,
)
from litellm.types.llms.openai import ChatCompletionToolParam
from litellm.types.utils import ModelResponse
@ -534,7 +535,7 @@ def _augment_system_with_summary(
return [{"type": "text", "text": prefix.rstrip()}, *system]
def _resolve_trigger_tokens(edit_spec: dict[str, object]) -> tuple[int, list[str]]:
def _resolve_trigger_tokens(edit_spec: Mapping[str, object]) -> tuple[int, list[str]]:
"""Validate and resolve ``trigger.value``.
Raises ``AnthropicContextManagementError`` if the explicitly-supplied value
@ -568,7 +569,7 @@ def _resolve_trigger_tokens(edit_spec: dict[str, object]) -> tuple[int, list[str
return value, warnings
def _build_summary_prompt(edit_spec: dict[str, object], tools: list[dict[str, object]] | None) -> str:
def _build_summary_prompt(edit_spec: Mapping[str, object], tools: Sequence[Mapping[str, object]] | None) -> str:
custom: Final = edit_spec.get("instructions")
if isinstance(custom, str) and custom.strip():
return custom
@ -623,7 +624,7 @@ def _count_effective_tokens(
try:
openai_shape = adapter.translate_anthropic_messages_to_openai(
messages=cast(
"list[AnthropicMessagesUserMessageParam | AnthopicMessagesAssistantMessageParam]",
"list[AllAnthropicPassThroughMessageValues]",
messages_without_compaction,
)
)
@ -736,7 +737,7 @@ def _extract_summary_text(raw: str | None) -> str | None:
def _system_to_openai_message(
system: str | list[dict[str, Any]] | None,
) -> dict[str, Any] | None:
) -> dict[str, object] | None:
"""Translate Anthropic-shaped ``system`` to an OpenAI system message.
Accepts a bare string or a list of Anthropic content blocks; returns
@ -773,7 +774,7 @@ def _build_summary_messages(
try:
openai_messages = LiteLLMAnthropicMessagesAdapter().translate_anthropic_messages_to_openai(
messages=cast(
"list[AnthropicMessagesUserMessageParam | AnthopicMessagesAssistantMessageParam]",
"list[AllAnthropicPassThroughMessageValues]",
stripped,
)
)
@ -809,7 +810,7 @@ def _is_user_message(msg: object) -> bool:
return isinstance(msg, dict) and msg.get("role") == "user"
def _append_text_to_content(content: Any, extra_text: str) -> Any:
def _append_text_to_content(content: object, extra_text: str) -> object:
"""Append ``extra_text`` to an OpenAI-shape message ``content`` field.
Handles the two common shapes: ``str`` and ``list`` of content parts.
@ -820,10 +821,29 @@ def _append_text_to_content(content: Any, extra_text: str) -> Any:
if isinstance(content, str):
return f"{content}\n\n{extra_text}"
if isinstance(content, list):
return [*content, {"type": "text", "text": extra_text}]
appended: Final[list[object]] = [*content, {"type": "text", "text": extra_text}]
return appended
return [content, {"type": "text", "text": extra_text}]
class _SummaryCallUserKwarg(TypedDict, total=False):
user: ReadOnly[object]
class _SummaryCallRegionKwarg(TypedDict, total=False):
allowed_model_region: ReadOnly[str]
class _SummaryCallKwargs(TypedDict):
model: ReadOnly[str]
messages: ReadOnly[list[dict[str, object]]]
max_tokens: ReadOnly[int]
timeout: ReadOnly[float]
litellm_metadata: ReadOnly[Mapping[str, object]]
user: NotRequired[ReadOnly[object]]
allowed_model_region: NotRequired[ReadOnly[str]]
async def _call_summary_model(
*,
summary_model: str,
@ -860,22 +880,24 @@ async def _call_summary_model(
# the parent ``/v1/messages`` request. On timeout the caller catches the
# exception and surfaces ``applied_edits[0].error = "summary_call_failed"``,
# forwarding the request without compaction rather than hanging.
call_kwargs: Final[dict[str, Any]] = {
"model": summary_model,
"messages": summary_messages,
"max_tokens": max_tokens,
"timeout": COMPACT_SUMMARY_TIMEOUT_SECONDS,
"litellm_metadata": metadata,
}
# The end-user id must also travel as the top-level ``user`` kwarg: legacy
# limiter hooks and prometheus end-user tracking read it from there rather
# than from ``litellm_metadata``, so without it the summary tokens would not
# debit the caller's end-user counters.
end_user_id: Final = metadata.get("user_api_key_end_user_id")
if end_user_id:
call_kwargs["user"] = end_user_id
if allowed_model_region is not None:
call_kwargs["allowed_model_region"] = allowed_model_region
call_kwargs: Final[_SummaryCallKwargs] = {
"model": summary_model,
"messages": summary_messages,
"max_tokens": max_tokens,
"timeout": COMPACT_SUMMARY_TIMEOUT_SECONDS,
"litellm_metadata": metadata,
**(_SummaryCallUserKwarg(user=end_user_id) if end_user_id else _SummaryCallUserKwarg()),
**(
_SummaryCallRegionKwarg(allowed_model_region=allowed_model_region)
if allowed_model_region is not None
else _SummaryCallRegionKwarg()
),
}
if llm_router is not None and hasattr(llm_router, "acompletion"):
return await llm_router.acompletion(**call_kwargs)
return await litellm.acompletion(**call_kwargs)

13
litellm/llms/azure/azure.py
View file

@ -10,6 +10,7 @@ from openai import (
AsyncAzureOpenAI,
AsyncOpenAI,
AzureOpenAI,
BadRequestError,
OpenAI,
)
@ -37,6 +38,10 @@ from litellm.utils import (
from ...types.llms.openai import HttpxBinaryResponseContent
from ..base import BaseLLM
from ..openai.common_utils import (
build_output_token_limit_response,
is_output_token_limit_error,
)
from .common_utils import (
AzureOpenAIError,
BaseAzureLLM,
@ -147,6 +152,10 @@ class AzureChatCompletion(BaseAzureLLM, BaseLLM):
headers: Final = dict(raw_response.headers)
response: Final = raw_response.parse()
return headers, response
except BadRequestError as e:
if not is_output_token_limit_error(e):
raise
return build_output_token_limit_response(e=e, data=data, is_async=False)
except Exception as e:
raise e
@ -175,6 +184,10 @@ class AzureChatCompletion(BaseAzureLLM, BaseLLM):
time_delta: Final = round(end_time - start_time, 2)
e.message += f" - timeout value={timeout}, time taken={time_delta} seconds"
raise e
except BadRequestError as e:
if not is_output_token_limit_error(e):
raise
return build_output_token_limit_response(e=e, data=data, is_async=True)
except Exception as e:
raise e

31
litellm/llms/azure/common_utils.py
View file

@ -2,11 +2,12 @@ import asyncio
import hashlib
import json
import os
from collections.abc import Callable
from collections.abc import Callable, Mapping
from typing import Any, Final, Literal, NamedTuple, cast
import httpx
from openai import AsyncAzureOpenAI, AsyncOpenAI, AzureOpenAI, OpenAI
from typing_extensions import ReadOnly, TypedDict
import litellm
from litellm._logging import verbose_logger
@ -23,6 +24,22 @@ from litellm.utils import _add_path_to_api_base
azure_ad_cache: Final = DualCache()
class _AzureAdTokenJson(TypedDict, total=False):
access_token: ReadOnly[str]
expires_in: ReadOnly[int]
class _AzureV1ClientParams(TypedDict, total=False, extra_items=object):
base_url: ReadOnly[str]
class _AzureGatewayClientParams(TypedDict, total=False, extra_items=object):
api_version: ReadOnly[str]
base_url: ReadOnly[str]
max_retries: ReadOnly[int]
timeout: ReadOnly[float | httpx.Timeout]
class AzureOpenAIError(BaseLLMException):
def __init__(
self,
@ -220,7 +237,7 @@ def get_azure_ad_token_from_oidc(
message=req_token.text,
)
azure_ad_token_json: Final = req_token.json()
azure_ad_token_json: Final[_AzureAdTokenJson] = req_token.json()
azure_ad_token_access_token = azure_ad_token_json.get("access_token", None)
azure_ad_token_expires_in: Final = azure_ad_token_json.get("expires_in", None)
@ -486,7 +503,7 @@ class BaseAzureLLM(BaseOpenAILLM):
v1_api_key = _async_v1_api_key
v1_params: Final[dict[str, Any]] = {
v1_params: Final[_AzureV1ClientParams] = {
"api_key": v1_api_key,
"base_url": f"{api_base}/openai/v1/",
}
@ -643,7 +660,7 @@ class BaseAzureLLM(BaseOpenAILLM):
api_base += "/"
api_base += f"{model}"
azure_client_params: Final[dict[str, Any]] = {
azure_client_params: Final[_AzureGatewayClientParams] = {
"api_version": api_version,
"base_url": f"{api_base}",
"http_client": litellm.client_session,
@ -702,7 +719,7 @@ class BaseAzureLLM(BaseOpenAILLM):
@staticmethod
def _get_base_azure_url(
api_base: str | None,
litellm_params: GenericLiteLLMParams | dict[str, Any] | None,
litellm_params: GenericLiteLLMParams | Mapping[str, object] | None,
route: Literal["/openai/responses", "/openai/vector_stores"] | str,
default_api_version: str | Literal["latest", "preview"] | None = None,
) -> str:
@ -757,7 +774,9 @@ class BaseAzureLLM(BaseOpenAILLM):
return False
return api_version in {"preview", "latest", "v1"}
def _resolve_env_var(self, litellm_params: dict[str, Any], param_key: str, env_var_key: str) -> str | None:
def _resolve_env_var(
self, litellm_params: Mapping[str, str | None], param_key: str, env_var_key: str
) -> str | None:
"""Resolve the environment variable for a given parameter key.
The logic here is different from `params.get(key, os.getenv(env_var))` because

27
litellm/llms/azure_ai/vector_stores/transformation.py
View file

@ -37,9 +37,32 @@ class AzureAIVectorStoreConfig(BaseVectorStoreConfig, BaseAzureLLM):
super().__init__()
def get_vector_store_endpoints_by_type(self) -> VectorStoreIndexEndpoints:
"""
Every ``GET`` under ``/indexes/`` is a read: get details, stats, and the
document reads (GET-form search, ``$count``, point lookup, and the
GET forms of suggest and autocomplete).
``POST`` splits by endpoint. Search, suggest, autocomplete, and analyze
are query endpoints, so they read; ``/docs/index`` is the batch endpoint
carrying upload, merge, mergeOrUpload, and delete actions, so it writes.
Patterns stay literal rather than ``{placeholder}`` templates because the
matcher falls back to the substring before a ``{``, which here is always
``/indexes/``. The matcher is substring-based, so an index name may
itself contain a read fragment (an index named ``analyze*`` puts
``/analyze`` inside the batch-write path); writes are classified before
reads, so such a path demands the write grant rather than being
shadowed into a read.
"""
return {
"read": [("GET", "/docs/search"), ("POST", "/docs/search")],
"write": [("PUT", "/docs")],
"read": [
("GET", "/indexes/"),
("POST", "/docs/search"),
("POST", "/docs/suggest"),
("POST", "/docs/autocomplete"),
("POST", "/analyze"),
],
"write": [("POST", "/docs/index")],
}
def get_auth_credentials(self, litellm_params: dict) -> BaseVectorStoreAuthCredentials:

19
litellm/llms/base_llm/search/transformation.py
View file

@ -18,6 +18,16 @@ else:
LiteLLMLoggingObj = Any
_PERPLEXITY_UNIFIED_PARAMS: Final[frozenset[str]] = frozenset(
(
"max_results",
"search_domain_filter",
"country",
"max_tokens_per_page",
)
)
def _search_host(url: str) -> str:
return urlsplit(url).netloc.lower()
@ -96,7 +106,7 @@ class BaseSearchConfig:
return "POST"
@staticmethod
def get_supported_perplexity_optional_params() -> set:
def get_supported_perplexity_optional_params() -> frozenset[str]:
"""
Get the set of Perplexity unified search parameters.
These are the standard parameters that providers should transform from.
@ -104,12 +114,7 @@ class BaseSearchConfig:
Returns:
Set of parameter names that are part of the unified spec
"""
return {
"max_results",
"search_domain_filter",
"country",
"max_tokens_per_page",
}
return _PERPLEXITY_UNIFIED_PARAMS
def _assert_trusted_api_base_for_server_credential(
self,

6
litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py
View file

@ -19,9 +19,9 @@ from litellm.llms.bedrock.common_utils import (
convert_bedrock_invoke_output_format_to_inline_schema,
get_anthropic_beta_from_headers,
normalize_bedrock_opus_output_config_effort,
normalize_custom_field_on_tools,
normalize_tool_input_schema_types_for_bedrock_invoke,
pop_bedrock_invoke_output_config_format,
remove_custom_field_from_tools,
)
from litellm.types.llms.anthropic import ANTHROPIC_TOOL_SEARCH_BETA_HEADER
from litellm.types.llms.openai import AllMessageValues
@ -243,8 +243,8 @@ class AmazonAnthropicClaudeConfig(AmazonInvokeConfig, AnthropicConfig):
if "anthropic_version" not in anthropic_request:
anthropic_request["anthropic_version"] = self.anthropic_version
# Remove `custom` field from tools (Bedrock doesn't support it)
remove_custom_field_from_tools(anthropic_request)
# Hoist `custom.defer_loading` then drop `custom` (Bedrock doesn't support it)
normalize_custom_field_on_tools(anthropic_request)
normalize_tool_input_schema_types_for_bedrock_invoke(anthropic_request)
return anthropic_request

21
litellm/llms/bedrock/common_utils.py
View file

@ -176,13 +176,14 @@ def convert_bedrock_invoke_output_format_to_inline_schema(
request_body["messages"] = new_messages
def remove_custom_field_from_tools(request_body: dict) -> None:
def normalize_custom_field_on_tools(request_body: dict) -> None:
"""
Remove ``custom`` field from each tool in the request body.
Drop the ``custom`` field from each tool, first hoisting a boolean
``custom.defer_loading`` onto the top-level ``defer_loading`` flag that
Bedrock and Anthropic actually document, unless the tool already carries one.
Claude Code (v2.1.69+) sends ``custom: {defer_loading: true}`` on tool
definitions, which Anthropic's API accepts but Bedrock rejects with
``"Extra inputs are not permitted"``.
Claude Code (v2.1.69+) is reported to send ``custom: {defer_loading: true}`` on
tool definitions, which Bedrock rejects with ``"Extra inputs are not permitted"``.
Args:
request_body: The request dictionary to modify in-place.
@ -193,8 +194,14 @@ def remove_custom_field_from_tools(request_body: dict) -> None:
if not tools or not isinstance(tools, list):
return
for tool in tools:
if isinstance(tool, dict):
tool.pop("custom", None)
if not isinstance(tool, dict):
continue
custom: dict[str, object] | None = tool.pop("custom", None)
if not isinstance(custom, dict) or "defer_loading" in tool:
continue
deferred: object = custom.get("defer_loading")
if isinstance(deferred, bool):
tool["defer_loading"] = deferred
def normalize_json_schema_custom_types_to_object(schema: dict) -> None:

94
litellm/llms/bedrock/files/transformation.py
View file

@ -2,17 +2,18 @@ import base64
import json
import os
import time
from collections.abc import Iterable, Mapping, MutableMapping
from collections.abc import Iterable, Mapping, MutableMapping, Sequence
from functools import cache
from itertools import chain
from types import MappingProxyType
from typing import Any, Final
from typing import Any, Final, TypeAlias, TypedDict
from urllib.parse import unquote
import httpx
from httpx import Headers, Response
from openai.types.file_deleted import FileDeleted
from pydantic import BaseModel, ConfigDict, TypeAdapter
from typing_extensions import ReadOnly
from litellm._logging import verbose_logger
from litellm._uuid import uuid
@ -63,10 +64,39 @@ from ..common_utils import BedrockError, merge_bedrock_aws_request_params, resol
S3_SIGNED_GET_HEADERS_PARAM: Final = "_s3_signed_get_headers"
def _frozen_mapping(items: Iterable[tuple[str, Any]]) -> Mapping[str, Any]:
def _frozen_mapping(items: Iterable[tuple[str, object]]) -> Mapping[str, object]:
return MappingProxyType(dict(items))
_EmbeddingBatchInput: TypeAlias = (
str | int | float | Sequence[str] | Sequence[int] | Sequence[Sequence[int]] | Mapping[str, object]
)
class _OpenAIBatchRecordBody(TypedDict, total=False):
model: ReadOnly[str]
prompt: ReadOnly[str | Sequence[str] | Sequence[int] | Sequence[Sequence[int]]]
input: ReadOnly[_EmbeddingBatchInput]
metadata: ReadOnly[Mapping[str, object]]
class _OpenAIBatchRecord(TypedDict, total=False):
custom_id: ReadOnly[str]
url: ReadOnly[str]
body: ReadOnly[_OpenAIBatchRecordBody]
class _BedrockBatchRecord(TypedDict):
recordId: ReadOnly[str]
modelInput: ReadOnly[Mapping[str, object]]
class _S3UploadResponse(TypedDict, total=False):
Key: ReadOnly[str]
Bucket: ReadOnly[str]
ContentLength: ReadOnly[int]
# JSONL batch records are untyped json, so the `/v1/responses` fields are
# validated into their concrete Responses API types before being handed to the
# Responses-to-Chat bridge. Both adapters drop keys the Responses API doesn't
@ -231,7 +261,7 @@ class BedrockFilesConfig(BaseAWSLLM, BaseFilesConfig):
def _get_s3_object_name_from_batch_jsonl(
self,
openai_jsonl_content: list[dict[str, Any]],
openai_jsonl_content: Sequence[_OpenAIBatchRecord],
) -> str:
"""
Gets a unique S3 object name for the Bedrock batch processing job
@ -341,7 +371,7 @@ class BedrockFilesConfig(BaseAWSLLM, BaseFilesConfig):
OPENAI_RESPONSES_URL = "/v1/responses"
@staticmethod
def _classify_batch_record(openai_jsonl_record: Mapping[str, Any]) -> BedrockBatchRecordKind:
def _classify_batch_record(openai_jsonl_record: _OpenAIBatchRecord) -> BedrockBatchRecordKind:
"""
Decide which OpenAI endpoint shape an OpenAI batch JSONL line carries.
@ -484,7 +514,7 @@ class BedrockFilesConfig(BaseAWSLLM, BaseFilesConfig):
return value if isinstance(value, str) and value else None
@staticmethod
def _coerce_embedding_input_to_string(raw_input: Any, model: str = "") -> str:
def _coerce_embedding_input_to_string(raw_input: _EmbeddingBatchInput | None, model: str = "") -> str:
"""
Normalize an OpenAI /v1/embeddings `input` field into the single
string that Bedrock Titan v2 InvokeModel expects in `inputText`.
@ -541,8 +571,8 @@ class BedrockFilesConfig(BaseAWSLLM, BaseFilesConfig):
def _map_openai_embedding_to_bedrock_params(
self,
openai_request_body: dict[str, Any],
) -> dict[str, Any]:
openai_request_body: _OpenAIBatchRecordBody,
) -> dict[str, object]:
"""
Transform an OpenAI /v1/embeddings request body into the
Bedrock InvokeModel `modelInput` for embedding models that AWS
@ -588,7 +618,9 @@ class BedrockFilesConfig(BaseAWSLLM, BaseFilesConfig):
return dict(titan_config._transform_request(input=input_text, inference_params=inference_params))
@staticmethod
def _transform_text_completion_body_to_chat_body(openai_request_body: Mapping[str, Any]) -> Mapping[str, Any]:
def _transform_text_completion_body_to_chat_body(
openai_request_body: _OpenAIBatchRecordBody,
) -> Mapping[str, object]:
"""
Rewrite an OpenAI `/v1/completions` batch body as a Chat Completions body.
@ -610,7 +642,7 @@ class BedrockFilesConfig(BaseAWSLLM, BaseFilesConfig):
)
@staticmethod
def _transform_responses_body_to_chat_body(openai_request_body: Mapping[str, Any]) -> Mapping[str, Any]:
def _transform_responses_body_to_chat_body(openai_request_body: _OpenAIBatchRecordBody) -> Mapping[str, object]:
"""
Rewrite an OpenAI `/v1/responses` batch body as a Chat Completions body.
@ -631,23 +663,25 @@ class BedrockFilesConfig(BaseAWSLLM, BaseFilesConfig):
"Batch record for /v1/responses is missing required `input` field: "
f"model={openai_request_body.get('model', '')}"
)
chat_body: Final = LiteLLMCompletionResponsesConfig.transform_responses_api_request_to_chat_completion_request(
model=openai_request_body.get("model", ""),
input=_responses_input_adapter().validate_python(responses_input),
responses_api_request=_responses_request_adapter().validate_python(
_frozen_mapping(
(key, value) for key, value in openai_request_body.items() if key not in ("model", "input")
)
),
metadata=openai_request_body.get("metadata"),
chat_body: Final[Mapping[str, object]] = (
LiteLLMCompletionResponsesConfig.transform_responses_api_request_to_chat_completion_request(
model=openai_request_body.get("model", ""),
input=_responses_input_adapter().validate_python(responses_input),
responses_api_request=_responses_request_adapter().validate_python(
_frozen_mapping(
(key, value) for key, value in openai_request_body.items() if key not in ("model", "input")
)
),
metadata=openai_request_body.get("metadata"),
)
)
return _frozen_mapping((key, value) for key, value in chat_body.items() if key != "tools" or value)
@staticmethod
def _transform_batch_body_to_chat_body(
openai_request_body: Mapping[str, Any],
openai_request_body: _OpenAIBatchRecordBody,
record_kind: BedrockBatchRecordKind,
) -> Mapping[str, Any]:
) -> Mapping[str, object]:
"""
Normalize a non-embedding batch body to the Chat Completions shape the
per-provider Bedrock transformations expect.
@ -666,7 +700,7 @@ class BedrockFilesConfig(BaseAWSLLM, BaseFilesConfig):
self,
openai_request_body: Mapping[str, Any],
provider: str | None = None,
) -> dict[str, Any]:
) -> dict[str, object]:
"""
Transform OpenAI request body to Bedrock-compatible modelInput
parameters using existing transformation logic.
@ -677,7 +711,7 @@ class BedrockFilesConfig(BaseAWSLLM, BaseFilesConfig):
"""
from litellm.types.utils import LlmProviders
_model: Final = openai_request_body.get("model", "")
_model: Final[str] = openai_request_body.get("model", "")
messages: Final = openai_request_body.get("messages", [])
optional_params: Final = {k: v for k, v in openai_request_body.items() if k not in ["model", "messages"]}
@ -733,8 +767,8 @@ class BedrockFilesConfig(BaseAWSLLM, BaseFilesConfig):
}
def _transform_openai_jsonl_content_to_bedrock_jsonl_content(
self, openai_jsonl_content: list[dict[str, Any]]
) -> list[dict[str, Any]]:
self, openai_jsonl_content: Sequence[_OpenAIBatchRecord]
) -> list[_BedrockBatchRecord]:
"""
Transforms OpenAI JSONL content to Bedrock batch format
@ -1026,7 +1060,7 @@ class BedrockFilesConfig(BaseAWSLLM, BaseFilesConfig):
response_headers: Final = raw_response.headers
# Extract S3 object information from the response
# S3 PUT object returns ETag and other metadata in headers
content_length: Final = response_headers.get("Content-Length", "0")
content_length: Final[str] = response_headers.get("Content-Length", "0")
# Use the actual upload URL that was used for the S3 upload
upload_url: Final = litellm_params.get("upload_url")
@ -1224,7 +1258,9 @@ class BedrockJsonlFilesTransformation:
object_name: Final = self._get_s3_object_name(openai_jsonl_content=openai_jsonl_content)
return bedrock_jsonl_string, object_name
def _transform_openai_jsonl_content_to_bedrock_jsonl_content(self, openai_jsonl_content: list[dict[str, Any]]):
def _transform_openai_jsonl_content_to_bedrock_jsonl_content(
self, openai_jsonl_content: Sequence[_OpenAIBatchRecord]
):
"""
Delegate to the main BedrockFilesConfig transformation method
"""
@ -1233,7 +1269,7 @@ class BedrockJsonlFilesTransformation:
def _get_s3_object_name(
self,
openai_jsonl_content: list[dict[str, Any]],
openai_jsonl_content: Sequence[_OpenAIBatchRecord],
) -> str:
"""
Gets a unique S3 object name for the Bedrock batch processing job
@ -1285,7 +1321,7 @@ class BedrockJsonlFilesTransformation:
return content
def transform_s3_bucket_response_to_openai_file_object(
self, create_file_data: CreateFileRequest, s3_upload_response: dict[str, Any]
self, create_file_data: CreateFileRequest, s3_upload_response: _S3UploadResponse
) -> OpenAIFileObject:
"""
Transforms S3 Bucket upload file response to OpenAI FileObject

8
litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py
View file

@ -33,9 +33,9 @@ from litellm.llms.bedrock.common_utils import (
get_anthropic_beta_from_headers,
is_claude_4_5_on_bedrock,
normalize_bedrock_opus_output_config_effort,
normalize_custom_field_on_tools,
normalize_tool_input_schema_types_for_bedrock_invoke,
pop_bedrock_invoke_output_config_format,
remove_custom_field_from_tools,
)
from litellm.types.llms.anthropic import (
ANTHROPIC_BETA_HEADER_VALUES,
@ -749,11 +749,9 @@ class AmazonAnthropicClaudeMessagesConfig(
model,
)
# 5b. Remove `custom` field from tools (Bedrock doesn't support it)
# Claude Code sends `custom: {defer_loading: true}` on tool definitions,
# which causes Bedrock to reject the request with "Extra inputs are not permitted"
# 5b. Hoist `custom.defer_loading` then drop `custom` (Bedrock doesn't support it)
# Ref: https://github.com/BerriAI/litellm/issues/22847
remove_custom_field_from_tools(anthropic_messages_request)
normalize_custom_field_on_tools(anthropic_messages_request)
normalize_tool_input_schema_types_for_bedrock_invoke(anthropic_messages_request)
ensure_bedrock_anthropic_messages_tool_names(anthropic_messages_request)

3
litellm/llms/nimble/__init__.py Normal file
View file

@ -0,0 +1,3 @@
from litellm.llms.nimble.search.transformation import NimbleSearchConfig
__all__ = ("NimbleSearchConfig",)

3
litellm/llms/nimble/search/__init__.py Normal file
View file

@ -0,0 +1,3 @@
from litellm.llms.nimble.search.transformation import NimbleSearchConfig
__all__ = ("NimbleSearchConfig",)

264
litellm/llms/nimble/search/transformation.py Normal file
View file

@ -0,0 +1,264 @@
"""
Calls Nimble's /v2/search endpoint to search the web.
Nimble API Reference: https://docs.nimbleway.com/api-reference/search/search
"""
from __future__ import annotations
from collections.abc import Mapping
from types import MappingProxyType
from typing import TYPE_CHECKING, Final
import httpx
from pydantic import BaseModel, ConfigDict, TypeAdapter, ValidationError
from litellm.llms.base_llm.chat.transformation import BaseLLMException
from litellm.llms.base_llm.search.transformation import (
BaseSearchConfig,
SearchResponse,
SearchResult,
)
from litellm.secret_managers.main import get_secret_str
if TYPE_CHECKING:
from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
_NIMBLE_DOCS_URL: Final = "https://docs.nimbleway.com/api-reference/search/search"
class _NimbleResult(BaseModel):
"""One entry of Nimble's `results` array. Every field is optional so a single degraded
result degrades to empty strings instead of failing the whole call."""
model_config = ConfigDict(extra="ignore", frozen=True)
title: str | None = None
url: str | None = None
content: str | None = None
description: str | None = None
# Free-form per Nimble's schema, so an unexpected shape must not fail the search.
additional_data: object = None
class _NimbleSearchResponse(BaseModel):
"""Nimble's /v2/search response envelope."""
model_config = ConfigDict(extra="ignore", frozen=True)
# Required: a search with no hits returns `[]`, so a null or absent `results` means the
# body is not a search response and must not be reported as a successful empty search.
results: tuple[_NimbleResult, ...]
class _AdditionalData(BaseModel):
"""The slice of a result's free-form `additional_data` that maps onto SearchResult."""
model_config = ConfigDict(extra="ignore", frozen=True)
publish_date: str | None = None
class _ErrorEnvelope(BaseModel):
"""Nimble reports errors as either `{"detail": ...}` (validation) or
`{"success": "false", "task_id": ..., "message": ...}` (collection)."""
model_config = ConfigDict(extra="ignore", frozen=True)
detail: str | None = None
message: str | None = None
_DomainListAdapter: Final = TypeAdapter(tuple[str, ...])
_NOTHING: Final[Mapping[str, object]] = MappingProxyType({})
def _optional(key: str, value: object) -> Mapping[str, object]:
"""A one-entry mapping to spread into a payload, or nothing when the value is absent."""
return MappingProxyType({key: value}) if value is not None else _NOTHING
class NimbleSearchConfig(BaseSearchConfig):
NIMBLE_API_BASE = "https://sdk.nimbleway.com/v2"
@staticmethod
def ui_friendly_name() -> str:
return "Nimble"
def validate_environment(
self,
headers: dict[str, str], # mutable-ok: BaseSearchConfig.validate_environment signature
api_key: str | None = None,
api_base: str | None = None,
**kwargs: object, # kwargs-ok: BaseSearchConfig.validate_environment signature
) -> dict[str, str]: # mutable-ok: the http handler passes this straight to httpx as headers
"""
Validate environment and return headers.
Returns a new dict rather than mutating ``headers``: the http handler calls this
a second time after ``litellm/search/main.py`` already did, so it has to be idempotent.
"""
resolved_api_key: Final = self.resolve_server_api_key(
caller_api_key=api_key,
caller_api_base=api_base,
key_env_vars=("NIMBLE_API_KEY",),
base_env_var="NIMBLE_API_BASE",
default_api_base=self.NIMBLE_API_BASE,
)
if not resolved_api_key:
raise ValueError("NIMBLE_API_KEY is not set. Set `NIMBLE_API_KEY` environment variable.")
return { # mutable-ok: httpx requires a plain dict of headers
**headers,
"Authorization": f"Bearer {resolved_api_key}",
"Content-Type": "application/json",
# Nimble's client-attribution header: names the calling software, nothing else.
"X-Client-Source": "litellm",
}
def get_complete_url(
self,
api_base: str | None,
optional_params: dict[str, object], # mutable-ok: BaseSearchConfig.get_complete_url signature
data: dict[str, object] | list[dict[str, object]] | None = None, # mutable-ok: base signature
**kwargs: object, # kwargs-ok: BaseSearchConfig.get_complete_url signature
) -> str:
resolved_base: Final = (api_base or get_secret_str("NIMBLE_API_BASE") or self.NIMBLE_API_BASE).rstrip("/")
if resolved_base.endswith("/search"):
return resolved_base
return f"{resolved_base}/search"
def transform_search_request(
self,
query: str | list[str], # mutable-ok: BaseSearchConfig.transform_search_request signature
optional_params: dict[str, object], # mutable-ok: base signature
**kwargs: object, # kwargs-ok: BaseSearchConfig.transform_search_request signature
) -> dict[str, object]: # mutable-ok: the http handler passes this straight to httpx as the JSON body
"""
Transform Search request to Nimble API format.
Nimble already uses the Perplexity unified spec's names, so this is close to a pass-through:
- query -> query (a list is joined with spaces; Nimble takes a single string)
- max_results -> max_results (sent unclamped so Nimble's own 1-100 validation reports the error)
- country -> country, upper-cased to the ISO form Nimble documents
- search_domain_filter -> include_domains, with `-`-prefixed entries going to exclude_domains
- max_tokens_per_page -> dropped (no Nimble equivalent)
Everything else is forwarded as-is, so the rest of Nimble's surface stays reachable
without LiteLLM tracking it.
"""
unified_params: Final = self.get_supported_perplexity_optional_params()
country: Final = optional_params.get("country")
# Spread after the derived domain filters so an explicitly supplied `include_domains`
# or `exclude_domains` wins over anything read out of `search_domain_filter`.
passthrough: Final = MappingProxyType(
{param: value for param, value in optional_params.items() if param not in unified_params}
)
return { # mutable-ok: httpx requires a plain dict for the JSON body
**_domain_filters(optional_params.get("search_domain_filter")),
**passthrough,
"query": " ".join(query) if isinstance(query, list) else query,
**_optional("max_results", optional_params.get("max_results")),
**_optional("country", country.upper() if isinstance(country, str) else None),
}
def transform_search_response(
self,
raw_response: httpx.Response,
logging_obj: LiteLLMLoggingObj,
**kwargs: object, # kwargs-ok: BaseSearchConfig.transform_search_response signature
) -> SearchResponse:
"""
Transform Nimble API response to LiteLLM unified SearchResponse format.
`date` carries only the absolute `publish_date`. News results often carry a relative
`publish_date_raw` ("1 day ago") instead, which is not a date, so the whole
`additional_data` object rides through as an extra on `SearchResult` and nothing is lost.
Nimble ranks results itself via metadata.position, so the order is preserved as received.
A body that does not match the documented schema raises an attributed error rather than
being reported as a successful empty search. Parsing the response bytes rather than
`.json()` covers the non-JSON case through that same path.
"""
try:
parsed: Final = _NimbleSearchResponse.model_validate_json(raw_response.content)
except ValidationError as e:
raise self.get_error_class(
error_message=f"response does not match the documented /v2/search schema: {e}",
status_code=raw_response.status_code,
headers=dict(raw_response.headers), # mutable-ok: BaseSearchConfig.get_error_class signature
)
return SearchResponse(
results=[ # mutable-ok: SearchResponse.results is declared list[SearchResult]
SearchResult(
title=result.title or "",
url=result.url or "",
snippet=result.content or result.description or "",
date=_publish_date(result.additional_data),
last_updated=None,
**_optional("additional_data", result.additional_data),
)
for result in parsed.results
],
object="search",
)
def get_error_class(
self,
error_message: str,
status_code: int,
headers: dict[str, str], # mutable-ok: BaseSearchConfig.get_error_class signature
) -> Exception:
detail: Final = _unwrap_error_detail(error_message).rstrip(". ")
return BaseLLMException(
status_code=status_code,
message=f"Nimble Search: {detail}. See {_NIMBLE_DOCS_URL} for details.",
headers=headers,
)
def _unwrap_error_detail(error_message: str) -> str:
"""
Surface the human-readable message inside Nimble's error envelopes.
Falls back to the raw body for anything else (CDN HTML pages, plain text, other shapes).
"""
try:
body: Final = _ErrorEnvelope.model_validate_json(error_message)
except ValidationError:
return error_message
return body.detail or body.message or error_message
def _domain_filters(search_domain_filter: object) -> Mapping[str, object]:
"""
Split the unified `search_domain_filter` into Nimble's include/exclude lists.
Follows the Perplexity unified spec, where a `-` prefix means "exclude this domain".
Anything that is not a list of strings is ignored rather than raising, since it only
ever narrows a search that is otherwise valid.
"""
try:
domains: Final = _DomainListAdapter.validate_python(search_domain_filter)
except ValidationError:
return _NOTHING
return MappingProxyType(
{
key: value
for key, value in (
("include_domains", tuple(d for d in domains if d and not d.startswith("-"))),
("exclude_domains", tuple(d[1:] for d in domains if d.startswith("-") and len(d) > 1)),
)
if value
}
)
def _publish_date(additional_data: object) -> str | None:
try:
return _AdditionalData.model_validate(additional_data).publish_date
except ValidationError:
return None

82
litellm/llms/openai/common_utils.py
View file

@ -7,16 +7,25 @@ import inspect
import json
import os
import ssl
import time
import uuid
from collections.abc import AsyncIterator, Iterator, Mapping
from typing import TYPE_CHECKING, Any, Final, Literal, NamedTuple, Optional
import httpx
import openai
from openai import AsyncAzureOpenAI, AsyncOpenAI, AzureOpenAI, OpenAI
from openai.types.chat import ChatCompletion, ChatCompletionChunk, ChatCompletionMessage
from openai.types.chat.chat_completion import Choice
from openai.types.chat.chat_completion_chunk import Choice as ChunkChoice
from openai.types.chat.chat_completion_chunk import ChoiceDelta
from openai.types.completion_usage import CompletionUsage
if TYPE_CHECKING:
from aiohttp import ClientSession
import litellm
from litellm.litellm_core_utils.token_counter import token_counter
from litellm.llms.base_llm.chat.transformation import BaseLLMException
from litellm.llms.custom_httpx.http_handler import (
_DEFAULT_TTL_FOR_HTTPX_CLIENTS,
@ -111,6 +120,79 @@ def drop_params_from_unprocessable_entity_error(
return new_data
_OUTPUT_TOKEN_LIMIT_ERROR_MARKER: Final[str] = (
"could not finish the message because max_tokens or model output limit was reached"
)
def is_output_token_limit_error(e: openai.BadRequestError) -> bool:
"""
True when OpenAI/Azure rejected a chat request because the output budget could not fit a single visible token.
GPT-5.x turns that case into a 400 while returning a length-truncated 200 for marginally larger budgets, so the
match has to stay pinned to the full provider sentence to avoid swallowing genuine bad requests.
"""
return _OUTPUT_TOKEN_LIMIT_ERROR_MARKER in e.message.lower()
def _output_token_limit_completion(model: str, prompt_tokens: int) -> ChatCompletion:
return ChatCompletion(
id=f"chatcmpl-{uuid.uuid4()}",
choices=(
Choice(
index=0,
finish_reason="length",
message=ChatCompletionMessage(role="assistant", content=""),
),
),
created=int(time.time()),
model=model,
object="chat.completion",
usage=CompletionUsage(completion_tokens=0, prompt_tokens=prompt_tokens, total_tokens=prompt_tokens),
)
def _output_token_limit_chunk(model: str) -> ChatCompletionChunk:
return ChatCompletionChunk(
id=f"chatcmpl-{uuid.uuid4()}",
choices=(
ChunkChoice(
index=0,
finish_reason="length",
delta=ChoiceDelta(role="assistant", content=""),
),
),
created=int(time.time()),
model=model,
object="chat.completion.chunk",
)
def _iter_once(chunk: ChatCompletionChunk) -> Iterator[ChatCompletionChunk]:
yield chunk
async def _aiter_once(chunk: ChatCompletionChunk) -> AsyncIterator[ChatCompletionChunk]:
yield chunk
def build_output_token_limit_response(
e: openai.BadRequestError, data: Mapping[str, object], is_async: bool
) -> tuple[httpx.Headers, ChatCompletion | Iterator[ChatCompletionChunk] | AsyncIterator[ChatCompletionChunk]]:
"""Synthesize the length-truncated response the provider itself returns for slightly larger output budgets.
The provider billed the prompt it processed but sends no usage object with the 400, so the prompt is estimated
the way every other usage-less path estimates it: reporting zero would spend input tokens against no budget.
"""
model: Final[str] = str(data.get("model", ""))
messages: Final = data.get("messages")
prompt_tokens: Final = token_counter(model=model, messages=messages) if isinstance(messages, list) else 0
if not data.get("stream"):
return e.response.headers, _output_token_limit_completion(model, prompt_tokens)
chunk: Final = _output_token_limit_chunk(model)
return e.response.headers, (_aiter_once(chunk) if is_async else _iter_once(chunk))
class BaseOpenAILLM:
"""
Base class for OpenAI LLMs for getting their httpx clients and SSL verification settings

7
litellm/llms/openai/cost_calculation.py
View file

@ -109,15 +109,16 @@ def cost_per_second(model: str, custom_llm_provider: str | None, duration: float
prompt_cost = 0.0
completion_cost = 0.0
## Speech / Audio cost calculation
if "output_cost_per_second" in model_info and model_info["output_cost_per_second"] is not None:
output_cost_per_second: Final = model_info.get("output_cost_per_second")
if output_cost_per_second is not None and output_cost_per_second > 0:
verbose_logger.debug(
"For model=%s - output_cost_per_second: %s; duration: %s",
model,
model_info.get("output_cost_per_second"),
output_cost_per_second,
duration,
)
## COST PER SECOND ##
completion_cost = model_info["output_cost_per_second"] * duration
completion_cost = output_cost_per_second * duration
elif "input_cost_per_second" in model_info and model_info["input_cost_per_second"] is not None:
verbose_logger.debug(
"For model=%s - input_cost_per_second: %s; duration: %s",

10
litellm/llms/openai/openai.py
View file

@ -46,7 +46,9 @@ from .chat.o_series_transformation import OpenAIOSeriesConfig
from .common_utils import (
BaseOpenAILLM,
OpenAIError,
build_output_token_limit_response,
drop_params_from_unprocessable_entity_error,
is_output_token_limit_error,
)
openaiOSeriesConfig: Final = OpenAIOSeriesConfig()
@ -436,6 +438,10 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM):
time_delta: Final = round(end_time - start_time, 2)
e.message += f" - timeout value={timeout}, time taken={time_delta} seconds"
raise e
except openai.BadRequestError as e:
if not is_output_token_limit_error(e):
raise
return build_output_token_limit_response(e=e, data=data, is_async=True)
except Exception as e:
raise e
@ -469,6 +475,10 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM):
return headers, response
except OpenAIError:
raise
except openai.BadRequestError as e:
if not is_output_token_limit_error(e):
raise
return build_output_token_limit_response(e=e, data=data, is_async=False)
except Exception as e:
if raw_response is not None:
raise Exception(

55
litellm/llms/runwayml/videos/transformation.py
View file

@ -1,8 +1,10 @@
from collections.abc import Mapping, Sequence
from datetime import datetime
from typing import TYPE_CHECKING, Any, Final
from typing import TYPE_CHECKING, Any, Final, Literal
import httpx
from httpx._types import RequestFiles
from typing_extensions import ReadOnly, TypedDict
import litellm
from litellm.constants import RUNWAYML_DEFAULT_API_VERSION
@ -31,6 +33,29 @@ else:
LiteLLMLoggingObj = Any
class _RunwayTaskResponse(TypedDict, total=False):
id: ReadOnly[str]
status: ReadOnly[str]
createdAt: ReadOnly[str]
completedAt: ReadOnly[str]
output: ReadOnly[Sequence[str] | str]
failureCode: ReadOnly[str]
failure: ReadOnly[str]
progress: ReadOnly[int]
class _VideoObjectData(TypedDict, extra_items=object):
id: ReadOnly[str]
object: ReadOnly[Literal["video"]]
status: ReadOnly[str]
created_at: ReadOnly[int]
def _parse_runway_task_response(raw_response: httpx.Response) -> _RunwayTaskResponse:
response_data: Final[_RunwayTaskResponse] = raw_response.json()
return response_data
class RunwayMLVideoConfig(BaseVideoConfig):
"""
Configuration class for RunwayML video generation.
@ -78,7 +103,7 @@ class RunwayMLVideoConfig(BaseVideoConfig):
- size -> ratio (convert "WIDTHxHEIGHT" to "WIDTH:HEIGHT")
- seconds -> duration (convert to integer)
"""
mapped_params: Final[dict[str, Any]] = {}
mapped_params: Final[dict[str, object]] = {}
# Handle input_reference parameter - map to promptImage
if "input_reference" in video_create_optional_params:
@ -180,7 +205,7 @@ class RunwayMLVideoConfig(BaseVideoConfig):
}
"""
# Build the request data
request_data: Final[dict[str, Any]] = {
request_data: Final[dict[str, object]] = {
"model": model,
"promptText": prompt,
}
@ -189,7 +214,7 @@ class RunwayMLVideoConfig(BaseVideoConfig):
request_data.update(video_create_optional_request_params)
# RunwayML uses JSON body, no files multipart
files_list: Final[list[tuple[str, Any]]] = []
files_list: Final[RequestFiles] = []
# Append the specific endpoint for video generation
full_api_base: Final = f"{api_base}/image_to_video"
@ -216,10 +241,10 @@ class RunwayMLVideoConfig(BaseVideoConfig):
We map this to OpenAI VideoObject format.
"""
response_data: Final = raw_response.json()
response_data: Final = _parse_runway_task_response(raw_response)
# Map RunwayML task response to VideoObject format
video_data: Final[dict[str, Any]] = {
video_data: Final[_VideoObjectData] = {
"id": response_data.get("id", ""),
"object": "video",
"status": self._map_runway_status(response_data.get("status", "pending")),
@ -326,7 +351,7 @@ class RunwayMLVideoConfig(BaseVideoConfig):
# Get task status to retrieve video URL
url: Final = f"{api_base}/tasks/{encoded_video_id}"
params: Final[dict[str, Any]] = {}
params: Final[dict[str, str]] = {}
return url, params
@ -421,7 +446,7 @@ class RunwayMLVideoConfig(BaseVideoConfig):
api_base: str,
litellm_params: GenericLiteLLMParams,
headers: dict,
extra_body: dict[str, Any] | None = None,
extra_body: Mapping[str, object] | None = None,
) -> tuple[str, dict]:
"""
Transform the video remix request for RunwayML API.
@ -448,7 +473,7 @@ class RunwayMLVideoConfig(BaseVideoConfig):
after: str | None = None,
limit: int | None = None,
order: str | None = None,
extra_query: dict[str, Any] | None = None,
extra_query: Mapping[str, object] | None = None,
) -> tuple[str, dict]:
"""
Transform the video list request for RunwayML API.
@ -484,7 +509,7 @@ class RunwayMLVideoConfig(BaseVideoConfig):
# Construct the URL for task cancellation
url: Final = f"{api_base}/tasks/{encoded_video_id}/cancel"
data: Final[dict[str, Any]] = {}
data: Final[dict[str, str]] = {}
return url, data
@ -494,7 +519,7 @@ class RunwayMLVideoConfig(BaseVideoConfig):
logging_obj: LiteLLMLoggingObj,
) -> VideoObject:
"""Transform the RunwayML video delete/cancel response."""
response_data: Final = raw_response.json()
response_data: Final = _parse_runway_task_response(raw_response)
video_obj: Final = VideoObject(
id=response_data.get("id", ""),
@ -524,7 +549,7 @@ class RunwayMLVideoConfig(BaseVideoConfig):
url: Final = f"{api_base}/tasks/{encoded_video_id}"
# Empty dict for GET request (no body)
data: Final[dict[str, Any]] = {}
data: Final[dict[str, str]] = {}
return url, data
@ -537,10 +562,10 @@ class RunwayMLVideoConfig(BaseVideoConfig):
"""
Transform the RunwayML video status retrieve response.
"""
response_data: Final = raw_response.json()
response_data: Final = _parse_runway_task_response(raw_response)
# Map RunwayML task response to VideoObject format
video_data: Final[dict[str, Any]] = {
video_data: Final[_VideoObjectData] = {
"id": response_data.get("id", ""),
"object": "video",
"status": self._map_runway_status(response_data.get("status", "pending")),
@ -572,7 +597,7 @@ class RunwayMLVideoConfig(BaseVideoConfig):
return video_obj
def transform_video_create_character_request(self, name, video, api_base, litellm_params, headers):
def transform_video_create_character_request(self, name, video: object, api_base, litellm_params, headers):
raise NotImplementedError("video create character is not supported for RunwayML")
def transform_video_create_character_response(self, raw_response, logging_obj):

71
litellm/llms/vertex_ai/files/transformation.py
View file

@ -5,12 +5,13 @@ import json
import os
import re
import time
from collections.abc import Callable, Iterable, Iterator
from typing import Any, Final
from collections.abc import Callable, Iterable, Iterator, Mapping
from typing import Any, Final, TypedDict
import httpx
from httpx import Headers, Response
from openai.types.file_deleted import FileDeleted
from typing_extensions import ReadOnly
import litellm
from litellm._uuid import uuid
@ -50,6 +51,7 @@ from litellm.types.llms.openai import (
HttpxBinaryResponseContent,
OpenAICreateFileRequestOptionalParams,
OpenAIFileObject,
OpenAIFilesPurpose,
PathLike,
)
from litellm.types.llms.vertex_ai import GcsBucketResponse
@ -62,6 +64,46 @@ _GCP_LABEL_VALUE_MAX_LEN: Final = 63
_CUSTOM_ID_RAW_LABEL_PREFIX: Final = "b32_"
class _GcsObjectMetadataJson(TypedDict, total=False):
purpose: ReadOnly[OpenAIFilesPurpose]
class _GcsObjectJson(TypedDict, total=False):
id: ReadOnly[str]
name: ReadOnly[str]
size: ReadOnly[str]
timeCreated: ReadOnly[str]
metadata: ReadOnly[_GcsObjectMetadataJson]
class _VertexBatchRowRequest(TypedDict, total=False):
labels: ReadOnly[Mapping[str, object]]
class _VertexBatchRow(TypedDict, total=False):
request: ReadOnly[_VertexBatchRowRequest]
status: ReadOnly[str]
processed_time: ReadOnly[str]
class _OpenAIBatchOutputError(TypedDict):
code: ReadOnly[str]
message: ReadOnly[str]
class _OpenAIBatchOutputResponse(TypedDict):
status_code: ReadOnly[int]
request_id: ReadOnly[str]
body: ReadOnly[Mapping[str, object]]
class _OpenAIBatchOutputRow(TypedDict):
id: ReadOnly[str]
custom_id: ReadOnly[str]
response: ReadOnly[_OpenAIBatchOutputResponse | None]
error: ReadOnly[_OpenAIBatchOutputError | None]
def _sanitize_gcp_label_value(value: str) -> str:
"""
Sanitize a string to meet GCP label value constraints.
@ -106,7 +148,7 @@ def _decode_gcp_label_value_chunks(values: list[str]) -> str | None:
return None
def _set_litellm_batch_custom_id_labels(labels: dict[str, str], custom_id: Any) -> None:
def _set_litellm_batch_custom_id_labels(labels: dict[str, str], custom_id: object) -> None:
"""
Store OpenAI batch custom_id for Vertex batch correlation.
@ -122,7 +164,7 @@ def _set_litellm_batch_custom_id_labels(labels: dict[str, str], custom_id: Any)
labels[f"litellm_custom_id_raw_{index}"] = raw_label_chunk
def _get_litellm_batch_custom_id_from_labels(labels: dict[str, Any]) -> str:
def _get_litellm_batch_custom_id_from_labels(labels: Mapping[str, object]) -> str:
"""Prefer encoded custom_id when present (see _set_litellm_batch_custom_id_labels)."""
raw: Final = labels.get("litellm_custom_id_raw")
if raw:
@ -186,7 +228,7 @@ def _iter_openai_jsonl_lines(openai_file_content: FileTypes) -> Iterator[str]:
``str.splitlines()`` + ``line.strip()`` for ``\\n`` / ``\\r\\n`` delimited
JSONL.
"""
content: Any = openai_file_content
content: FileTypes | str = openai_file_content
if isinstance(content, tuple):
content = content[1]
@ -246,6 +288,11 @@ def _iter_openai_jsonl_entries(
yield json.loads(line)
def _parse_vertex_batch_output_row(line: str) -> _VertexBatchRow:
row: Final[_VertexBatchRow] = json.loads(line)
return row
class _OpenAIToVertexBatchUploadStream(BaseFileUploadStream):
"""Streams an OpenAI batch JSONL upload as Vertex-wrapped JSONL one row at a
time, so the transformed payload is never held in full.
@ -463,7 +510,7 @@ class VertexAIFilesConfig(VertexBase, BaseFilesConfig):
"""
Transform VertexAI File upload response into OpenAI-style FileObject
"""
response_json: Final = raw_response.json()
response_json: Final[GcsBucketResponse] = raw_response.json()
try:
response_object: Final = GcsBucketResponse(**response_json)
@ -523,7 +570,7 @@ class VertexAIFilesConfig(VertexBase, BaseFilesConfig):
logging_obj: LiteLLMLoggingObj,
litellm_params: dict,
) -> OpenAIFileObject:
response_json: Final = raw_response.json()
response_json: Final[_GcsObjectJson] = raw_response.json()
gcs_id = response_json.get("id", "")
gcs_id = "/".join(gcs_id.split("/")[:-1]) if gcs_id else ""
return OpenAIFileObject(
@ -682,7 +729,7 @@ class VertexAIFilesConfig(VertexBase, BaseFilesConfig):
# discriminating fields. Anything else (e.g. a binary file whose
# first line is not valid UTF-8/JSON) raises and falls through to the
# passthrough below, leaving the content untouched.
first_row: Final = json.loads(first_line)
first_row: Final = _parse_vertex_batch_output_row(first_line)
is_vertex_batch_output: Final = (
"request" in first_row
and "response" in first_row
@ -723,7 +770,7 @@ class VertexAIFilesConfig(VertexBase, BaseFilesConfig):
for line in itertools.chain([first_line], lines):
try:
openai_output = self._transform_single_vertex_batch_output_to_openai(
vertex_output=json.loads(line),
vertex_output=_parse_vertex_batch_output_row(line),
vertex_gemini_config=vertex_gemini_config,
logging_obj=batch_transform_logging_obj,
mock_httpx_response=mock_httpx_response,
@ -742,18 +789,18 @@ class VertexAIFilesConfig(VertexBase, BaseFilesConfig):
def _transform_single_vertex_batch_output_to_openai(
self,
vertex_output: dict[str, Any],
vertex_output: _VertexBatchRow,
vertex_gemini_config: VertexGeminiConfig,
logging_obj: Logging,
mock_httpx_response: httpx.Response,
) -> dict[str, Any]:
) -> _OpenAIBatchOutputRow:
"""
Transform a single Vertex AI batch output line to OpenAI format.
Uses the existing VertexGeminiConfig transformation for the response.
"""
# Extract custom_id from request labels (prefer raw for OpenAI round-trip)
request_data: Final = vertex_output.get("request", {})
labels: Final = request_data.get("labels", {}) or {}
labels: Final[Mapping[str, object]] = request_data.get("labels", {}) or {}
custom_id: Final = _get_litellm_batch_custom_id_from_labels(labels)
# Check if there's an error

71
litellm/llms/vertex_ai/videos/transformation.py
View file

@ -7,10 +7,12 @@ Based on: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/model-refer
import base64
import time
from typing import TYPE_CHECKING, Any, Final, cast
from collections.abc import Sequence
from typing import TYPE_CHECKING, Any, Final, TypedDict, cast
import httpx
from httpx._types import RequestFiles
from typing_extensions import ReadOnly
from litellm.constants import DEFAULT_GOOGLE_VIDEO_DURATION_SECONDS
from litellm.images.utils import ImageEditRequestUtils
@ -40,11 +42,37 @@ else:
BaseLLMException = Any
class _VeoVideo(TypedDict, total=False):
gcsUri: ReadOnly[str]
bytesBase64Encoded: ReadOnly[str]
mimeType: ReadOnly[str]
class _VeoOperationResponse(TypedDict, total=False):
videos: ReadOnly[Sequence[_VeoVideo]]
class _VeoOperationMetadata(TypedDict, total=False):
createTime: ReadOnly[str]
class _VeoOperation(TypedDict, total=False):
name: ReadOnly[str]
done: ReadOnly[bool]
metadata: ReadOnly[_VeoOperationMetadata]
response: ReadOnly[_VeoOperationResponse]
def _parse_veo_operation(raw_response: httpx.Response) -> _VeoOperation:
operation: Final[_VeoOperation] = raw_response.json()
return operation
def _build_vertex_video_usage_from_request_data(
request_data: dict[str, Any] | None,
) -> dict[str, Any]:
) -> dict[str, float | str]:
"""Build usage metadata (duration, resolution) for video cost calculation."""
usage_data: Final[dict[str, Any]] = {}
usage_data: Final[dict[str, float | str]] = {}
if not request_data:
return usage_data
@ -125,7 +153,7 @@ class VertexAIVideoConfig(BaseVideoConfig, VertexBase):
video_create_optional_params: VideoCreateOptionalRequestParams,
model: str,
drop_params: bool,
) -> dict[str, Any]:
) -> dict[str, object]:
"""
Map OpenAI-style parameters to Veo format.
@ -135,7 +163,7 @@ class VertexAIVideoConfig(BaseVideoConfig, VertexBase):
- size → aspectRatio (e.g., "1280x720" → "16:9")
- seconds → durationSeconds (defaults to 4 seconds if not provided)
"""
mapped_params: Final[dict[str, Any]] = {}
mapped_params: Final[dict[str, object]] = {}
# Map input_reference to image (will be processed in transform_video_create_request)
if "input_reference" in video_create_optional_params:
@ -289,7 +317,7 @@ class VertexAIVideoConfig(BaseVideoConfig, VertexBase):
}
"""
# Build instance with prompt
instance_dict: Final[dict[str, Any]] = {"prompt": prompt}
instance_dict: Final[dict[str, object]] = {"prompt": prompt}
params_copy: Final = video_create_optional_request_params.copy()
# Check if user wants to provide full instance dict
@ -324,13 +352,13 @@ class VertexAIVideoConfig(BaseVideoConfig, VertexBase):
# {"parameters": {"parameters": {...}}} ← wrong
# {"parameters": {...}} ← correct
nested_params: Final = params_copy.pop("parameters", None)
vertex_params: Final[dict[str, Any]] = {}
vertex_params: Final[dict[str, object]] = {}
if isinstance(nested_params, dict):
vertex_params.update(nested_params)
vertex_params.update(params_copy)
# Build request data directly (TypedDict doesn't have model_dump)
request_data: Final[dict[str, Any]] = {"instances": [instance_dict]}
request_data: Final[dict[str, object]] = {"instances": [instance_dict]}
# Only add parameters if there are any
if vertex_params:
@ -363,7 +391,7 @@ class VertexAIVideoConfig(BaseVideoConfig, VertexBase):
- status: "processing"
- usage: includes duration_seconds and optional video_resolution for cost calculation
"""
response_data: Final = raw_response.json()
response_data: Final = _parse_veo_operation(raw_response)
operation_name: Final = response_data.get("name")
if not operation_name:
@ -441,7 +469,7 @@ class VertexAIVideoConfig(BaseVideoConfig, VertexBase):
}
}
"""
response_data: Final = raw_response.json()
response_data: Final = _parse_veo_operation(raw_response)
operation_name: Final = response_data.get("name", "")
is_done: Final = response_data.get("done", False)
@ -513,7 +541,7 @@ class VertexAIVideoConfig(BaseVideoConfig, VertexBase):
Extracts the base64 encoded video from the response and decodes it to bytes.
"""
response_data: Final = raw_response.json()
response_data: Final = _parse_veo_operation(raw_response)
if not response_data.get("done", False):
raise ValueError(
@ -548,7 +576,7 @@ class VertexAIVideoConfig(BaseVideoConfig, VertexBase):
api_base: str,
litellm_params: GenericLiteLLMParams,
headers: dict,
extra_body: dict[str, Any] | None = None,
extra_body: dict[str, object] | None = None,
) -> tuple[str, dict]:
"""
Video remix is not supported by Veo API.
@ -574,7 +602,7 @@ class VertexAIVideoConfig(BaseVideoConfig, VertexBase):
after: str | None = None,
limit: int | None = None,
order: str | None = None,
extra_query: dict[str, Any] | None = None,
extra_query: dict[str, object] | None = None,
) -> tuple[str, dict]:
"""
Video list is not supported by Veo API.
@ -615,7 +643,7 @@ class VertexAIVideoConfig(BaseVideoConfig, VertexBase):
"""Video delete is not supported."""
raise NotImplementedError("Video delete is not supported by Vertex AI Veo.")
def transform_video_create_character_request(self, name, video, api_base, litellm_params, headers):
def transform_video_create_character_request(self, name, video: object, api_base, litellm_params, headers):
raise NotImplementedError("video create character is not supported for Vertex AI")
def transform_video_create_character_response(self, raw_response, logging_obj):
@ -649,7 +677,7 @@ class VertexAIVideoConfig(BaseVideoConfig, VertexBase):
api_base: str,
litellm_params: GenericLiteLLMParams,
headers: dict,
extra_body: dict[str, Any] | None = None,
extra_body: dict[str, object] | None = None,
prefetched_source_data: dict[str, Any] | None = None,
) -> tuple[str, dict]:
"""
@ -667,12 +695,13 @@ class VertexAIVideoConfig(BaseVideoConfig, VertexBase):
if not prefetched_source_data.get("done", False):
raise ValueError("Source video generation is not complete yet. Check the video status before editing.")
videos: Final = prefetched_source_data.get("response", {}).get("videos", [])
source_response: Final[_VeoOperationResponse] = prefetched_source_data.get("response", {})
videos: Final = source_response.get("videos", [])
if not videos:
raise ValueError("No videos found in the completed operation. Cannot edit.")
source_video: Final = videos[0]
video_input: Final[dict[str, Any]] = {}
video_input: Final[dict[str, str]] = {}
if "gcsUri" in source_video:
video_input["gcsUri"] = source_video["gcsUri"]
elif "bytesBase64Encoded" in source_video:
@ -684,13 +713,13 @@ class VertexAIVideoConfig(BaseVideoConfig, VertexBase):
operation_name: Final = extract_original_video_id(video_id)
model: Final = self.extract_model_from_operation_name(operation_name) or ""
instance_dict: Final[dict[str, Any]] = {"prompt": prompt, "video": video_input}
request_data: Final[dict[str, Any]] = {"instances": [instance_dict]}
instance_dict: Final[dict[str, object]] = {"prompt": prompt, "video": video_input}
request_data: Final[dict[str, object]] = {"instances": [instance_dict]}
if extra_body:
extra_body_copy: Final = dict(extra_body)
nested_params: Final = extra_body_copy.pop("parameters", None)
vertex_params: Final[dict[str, Any]] = {}
vertex_params: Final[dict[str, object]] = {}
if isinstance(nested_params, dict):
vertex_params.update(nested_params)
vertex_params.update(extra_body_copy)
@ -716,7 +745,7 @@ class VertexAIVideoConfig(BaseVideoConfig, VertexBase):
usage includes duration_seconds and optional video_resolution from the
edit request parameters for cost calculation.
"""
response_data: Final = raw_response.json()
response_data: Final = _parse_veo_operation(raw_response)
operation_name: Final = response_data.get("name")
if not operation_name:

7
litellm/main.py
View file

@ -5616,7 +5616,12 @@ def completion(
elif custom_llm_provider == "hosted_vllm":
response = _complete_hosted_vllm(_dispatch_ctx)
elif (
model in litellm.open_ai_chat_completion_models
# A known OpenAI model name only decides the route when nothing else
# resolved a provider. get_llm_provider() already maps these names to
# "openai", so a different value here was asked for explicitly (or came
# from a register_model entry), and the provider config built for it
# would be handed to the OpenAI handler.
(model in litellm.open_ai_chat_completion_models and custom_llm_provider in (None, "openai"))
or custom_llm_provider == "custom_openai"
or custom_llm_provider == "deepinfra"
or custom_llm_provider == "perplexity"

708
litellm/model_prices_and_context_window_backup.json
View file

@ -6164,7 +6164,10 @@
"supports_response_schema": true,
"supports_system_messages": true,
"supports_tool_choice": true,
"supports_vision": true
"supports_vision": true,
"supports_none_reasoning_effort": true,
"supports_xhigh_reasoning_effort": true,
"supports_minimal_reasoning_effort": true
},
"azure/us/gpt-5.4": {
"cache_read_input_token_cost": 2.8e-07,
@ -6199,7 +6202,10 @@
"supports_response_schema": true,
"supports_system_messages": true,
"supports_tool_choice": true,
"supports_vision": true
"supports_vision": true,
"supports_none_reasoning_effort": true,
"supports_xhigh_reasoning_effort": true,
"supports_minimal_reasoning_effort": true
},
"azure/eu/gpt-5.4": {
"cache_read_input_token_cost": 2.8e-07,
@ -6234,7 +6240,10 @@
"supports_response_schema": true,
"supports_system_messages": true,
"supports_tool_choice": true,
"supports_vision": true
"supports_vision": true,
"supports_none_reasoning_effort": true,
"supports_xhigh_reasoning_effort": true,
"supports_minimal_reasoning_effort": true
},
"azure/gpt-5.4-2026-03-05": {
"cache_read_input_token_cost": 2.5e-07,
@ -6276,7 +6285,10 @@
"supports_response_schema": true,
"supports_system_messages": true,
"supports_tool_choice": true,
"supports_vision": true
"supports_vision": true,
"supports_none_reasoning_effort": true,
"supports_xhigh_reasoning_effort": true,
"supports_minimal_reasoning_effort": true
},
"azure/us/gpt-5.4-2026-03-05": {
"cache_read_input_token_cost": 2.8e-07,
@ -6312,7 +6324,10 @@
"supports_response_schema": true,
"supports_system_messages": true,
"supports_tool_choice": true,
"supports_vision": true
"supports_vision": true,
"supports_none_reasoning_effort": true,
"supports_xhigh_reasoning_effort": true,
"supports_minimal_reasoning_effort": true
},
"azure/eu/gpt-5.4-2026-03-05": {
"cache_read_input_token_cost": 2.8e-07,
@ -6348,7 +6363,10 @@
"supports_response_schema": true,
"supports_system_messages": true,
"supports_tool_choice": true,
"supports_vision": true
"supports_vision": true,
"supports_none_reasoning_effort": true,
"supports_xhigh_reasoning_effort": true,
"supports_minimal_reasoning_effort": true
},
"azure/gpt-5.4-pro": {
"cache_read_input_token_cost": 3e-06,
@ -7301,8 +7319,8 @@
"supports_tool_choice": true,
"supports_vision": true,
"supports_web_search": true,
"supports_none_reasoning_effort": false,
"supports_xhigh_reasoning_effort": false
"supports_none_reasoning_effort": true,
"supports_xhigh_reasoning_effort": true
},
"azure/gpt-5.4-mini-2026-03-17": {
"cache_read_input_token_cost": 7.5e-08,
@ -7337,8 +7355,8 @@
"supports_tool_choice": true,
"supports_vision": true,
"supports_web_search": true,
"supports_none_reasoning_effort": false,
"supports_xhigh_reasoning_effort": false
"supports_none_reasoning_effort": true,
"supports_xhigh_reasoning_effort": true
},
"azure/gpt-5.4-nano": {
"cache_read_input_token_cost": 2e-08,
@ -7372,8 +7390,8 @@
"supports_tool_choice": true,
"supports_vision": true,
"supports_web_search": true,
"supports_none_reasoning_effort": false,
"supports_xhigh_reasoning_effort": false
"supports_none_reasoning_effort": true,
"supports_xhigh_reasoning_effort": true
},
"azure/gpt-5.4-nano-2026-03-17": {
"cache_read_input_token_cost": 2e-08,
@ -7408,8 +7426,8 @@
"supports_tool_choice": true,
"supports_vision": true,
"supports_web_search": true,
"supports_none_reasoning_effort": false,
"supports_xhigh_reasoning_effort": false
"supports_none_reasoning_effort": true,
"supports_xhigh_reasoning_effort": true
},
"azure/gpt-image-1": {
"cache_read_input_token_cost": 1.25e-06,
@ -8712,6 +8730,268 @@
"/v1/images/generations"
]
},
"azure_ai/FW-DeepSeek-V3.2": {
"cache_read_input_token_cost": 3.1e-07,
"input_cost_per_token": 6.2e-07,
"litellm_provider": "azure_ai",
"max_input_tokens": 163840,
"max_output_tokens": 163840,
"max_tokens": 163840,
"mode": "chat",
"output_cost_per_token": 1.85e-06,
"source": "https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/fireworks/",
"supports_function_calling": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_tool_choice": true
},
"azure_ai/FW-DeepSeek-V4-Pro": {
"cache_read_input_token_cost": 1.65e-07,
"input_cost_per_token": 1.925e-06,
"litellm_provider": "azure_ai",
"max_input_tokens": 1000000,
"max_output_tokens": 384000,
"max_tokens": 384000,
"mode": "chat",
"output_cost_per_token": 3.828e-06,
"source": "https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/fireworks/",
"supports_function_calling": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_tool_choice": true
},
"azure_ai/FW-GLM-5": {
"cache_read_input_token_cost": 2.2e-07,
"input_cost_per_token": 1.1e-06,
"litellm_provider": "azure_ai",
"max_input_tokens": 200000,
"max_output_tokens": 128000,
"max_tokens": 128000,
"mode": "chat",
"output_cost_per_token": 3.52e-06,
"source": "https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/fireworks/",
"supports_function_calling": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_tool_choice": true
},
"azure_ai/FW-GLM-5.1": {
"cache_read_input_token_cost": 2.86e-07,
"input_cost_per_token": 1.54e-06,
"litellm_provider": "azure_ai",
"max_input_tokens": 202800,
"max_output_tokens": 131072,
"max_tokens": 131072,
"mode": "chat",
"output_cost_per_token": 4.84e-06,
"source": "https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/fireworks/",
"supports_function_calling": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_tool_choice": true
},
"azure_ai/FW-GLM-5.2": {
"cache_read_input_token_cost": 1.5e-07,
"input_cost_per_token": 1.54e-06,
"litellm_provider": "azure_ai",
"max_input_tokens": 1048576,
"max_output_tokens": 131072,
"max_tokens": 131072,
"mode": "chat",
"output_cost_per_token": 4.84e-06,
"source": "https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/fireworks/",
"supports_function_calling": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_tool_choice": true
},
"azure_ai/FW-GLM-5.2-Fast": {
"cache_read_input_token_cost": 2.1e-07,
"input_cost_per_token": 2.1e-06,
"litellm_provider": "azure_ai",
"max_input_tokens": 1048576,
"max_output_tokens": 131072,
"max_tokens": 131072,
"mode": "chat",
"output_cost_per_token": 6.6e-06,
"source": "https://docs.fireworks.ai/serverless/pricing",
"supports_function_calling": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_tool_choice": true
},
"azure_ai/FW-Inkling": {
"cache_read_input_token_cost": 1.7e-07,
"input_cost_per_token": 1e-06,
"litellm_provider": "azure_ai",
"max_input_tokens": 1048576,
"max_output_tokens": 1048576,
"max_tokens": 1048576,
"mode": "chat",
"output_cost_per_token": 4.05e-06,
"source": "https://fireworks.ai/models/fireworks/inkling",
"supported_modalities": [
"text"
],
"supported_output_modalities": [
"text"
],
"supports_function_calling": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_tool_choice": true
},
"azure_ai/FW-Kimi-K2.5": {
"cache_read_input_token_cost": 1.1e-07,
"input_cost_per_token": 6.6e-07,
"litellm_provider": "azure_ai",
"max_input_tokens": 262144,
"max_output_tokens": 262144,
"max_tokens": 262144,
"mode": "chat",
"output_cost_per_token": 3.3e-06,
"source": "https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/fireworks/",
"supported_modalities": [
"text",
"image"
],
"supported_output_modalities": [
"text"
],
"supports_function_calling": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_tool_choice": true,
"supports_vision": true
},
"azure_ai/FW-Kimi-K2.6": {
"cache_read_input_token_cost": 1.76e-07,
"input_cost_per_token": 1.045e-06,
"litellm_provider": "azure_ai",
"max_input_tokens": 262144,
"max_output_tokens": 262144,
"max_tokens": 262144,
"mode": "chat",
"output_cost_per_token": 4.4e-06,
"source": "https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/fireworks/",
"supported_modalities": [
"text",
"image"
],
"supported_output_modalities": [
"text"
],
"supports_function_calling": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_tool_choice": true,
"supports_vision": true
},
"azure_ai/FW-Kimi-K2.7-Code": {
"cache_read_input_token_cost": 2.1e-07,
"input_cost_per_token": 1.05e-06,
"litellm_provider": "azure_ai",
"max_input_tokens": 262144,
"max_output_tokens": 262144,
"max_tokens": 262144,
"mode": "chat",
"output_cost_per_token": 4.4e-06,
"source": "https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/fireworks/",
"supported_modalities": [
"text",
"image"
],
"supported_output_modalities": [
"text"
],
"supports_function_calling": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_tool_choice": true,
"supports_vision": true
},
"azure_ai/FW-Kimi-K3": {
"cache_read_input_token_cost": 3.3e-07,
"input_cost_per_token": 3.3e-06,
"litellm_provider": "azure_ai",
"max_input_tokens": 1048576,
"max_output_tokens": 131072,
"max_tokens": 131072,
"mode": "chat",
"output_cost_per_token": 1.65e-05,
"source": "https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/introducing-kimi-k3-through-fireworks-ai-on-microsoft-foundry/4540187",
"supported_modalities": [
"text",
"image"
],
"supported_output_modalities": [
"text"
],
"supports_function_calling": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_tool_choice": true,
"supports_vision": true
},
"azure_ai/FW-MiniMax-M2.5": {
"cache_read_input_token_cost": 3.3e-08,
"input_cost_per_token": 3.3e-07,
"litellm_provider": "azure_ai",
"max_input_tokens": 1000000,
"max_output_tokens": 1000000,
"max_tokens": 1000000,
"mode": "chat",
"output_cost_per_token": 1.32e-06,
"source": "https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/fireworks/",
"supports_function_calling": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_tool_choice": true
},
"azure_ai/FW-MiniMax-M3": {
"cache_read_input_token_cost": 6.6e-08,
"input_cost_per_token": 3.3e-07,
"litellm_provider": "azure_ai",
"max_input_tokens": 512000,
"max_output_tokens": 512000,
"max_tokens": 512000,
"mode": "chat",
"output_cost_per_token": 1.32e-06,
"source": "https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/fireworks/",
"supported_modalities": [
"text",
"image"
],
"supported_output_modalities": [
"text"
],
"supports_function_calling": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_tool_choice": true,
"supports_vision": true
},
"azure_ai/FW-Nemotron-3-Ultra-NVFP4": {
"cache_read_input_token_cost": 1.19e-07,
"input_cost_per_token": 6e-07,
"litellm_provider": "azure_ai",
"max_input_tokens": 262144,
"max_output_tokens": 262144,
"max_tokens": 262144,
"mode": "chat",
"output_cost_per_token": 2.4e-06,
"source": "https://fireworks.ai/models/fireworks/nemotron-3-ultra-nvfp4",
"supported_modalities": [
"text"
],
"supported_output_modalities": [
"text"
],
"supports_function_calling": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_tool_choice": true
},
"azure_ai/MAI-Image-2.5": {
"input_cost_per_image_token": 8e-06,
"input_cost_per_token": 5e-06,
@ -9329,6 +9609,24 @@
"supports_tool_choice": true,
"supports_web_search": true
},
"azure_ai/grok-4.3": {
"cache_read_input_token_cost": 2e-07,
"input_cost_per_token": 1.25e-06,
"litellm_provider": "azure_ai",
"max_input_tokens": 200000,
"max_output_tokens": 200000,
"max_tokens": 200000,
"mode": "chat",
"output_cost_per_token": 2.5e-06,
"source": "https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/introducing-grok-4-3-on-microsoft-foundry-latest-generation-agentic-capabilities/4517096",
"supports_function_calling": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_response_schema": true,
"supports_tool_choice": true,
"supports_vision": true,
"supports_web_search": true
},
"azure_ai/grok-4-fast-non-reasoning": {
"input_cost_per_token": 2e-07,
"output_cost_per_token": 5e-07,
@ -15997,6 +16295,14 @@
"notes": "TinyFish Search API"
}
},
"nimble/search": {
"input_cost_per_query": 0.005,
"litellm_provider": "nimble",
"mode": "search",
"metadata": {
"notes": "Nimble Search API pay-as-you-go list price: $5 per 1,000 searches, up to 100 results per search. Volume plans price differently."
}
},
"elevenlabs/scribe_v1": {
"input_cost_per_second": 6.11e-05,
"litellm_provider": "elevenlabs",
@ -19021,6 +19327,60 @@
},
"web_search_billing_unit": "per_query"
},
"vertex_ai/gemini-3.7-flash": {
"cache_read_input_token_cost": 7.5e-08,
"cache_read_input_token_cost_flex": 3.75e-08,
"input_cost_per_token": 7.5e-07,
"input_cost_per_token_batches": 3.75e-07,
"input_cost_per_token_flex": 3.75e-07,
"litellm_provider": "vertex_ai",
"max_input_tokens": 1048576,
"max_output_tokens": 65536,
"max_tokens": 65536,
"mode": "chat",
"output_cost_per_reasoning_token": 3.75e-06,
"output_cost_per_token": 3.75e-06,
"output_cost_per_token_batches": 1.875e-06,
"output_cost_per_token_flex": 1.875e-06,
"source": "https://cloud.google.com/vertex-ai/generative-ai/pricing",
"supported_endpoints": [
"/v1/chat/completions",
"/v1/completions",
"/v1/batch"
],
"supported_modalities": [
"text",
"image",
"audio",
"video"
],
"supported_output_modalities": [
"text"
],
"supports_audio_input": true,
"supports_function_calling": true,
"supports_parallel_function_calling": true,
"supports_pdf_input": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_response_schema": true,
"supports_system_messages": true,
"supports_tool_choice": true,
"supports_url_context": true,
"supports_video_input": true,
"supports_vision": true,
"supports_web_search": true,
"supports_native_streaming": true,
"input_cost_per_token_priority": 1.35e-06,
"output_cost_per_token_priority": 6.75e-06,
"cache_read_input_token_cost_priority": 1.35e-07,
"search_context_cost_per_query": {
"search_context_size_low": 0.014,
"search_context_size_medium": 0.014,
"search_context_size_high": 0.014
},
"web_search_billing_unit": "per_query"
},
"vertex_ai/gemini-3.1-pro-preview": {
"cache_read_input_token_cost": 2e-07,
"cache_read_input_token_cost_above_200k_tokens": 4e-07,
@ -20696,6 +21056,63 @@
},
"web_search_billing_unit": "per_query"
},
"gemini/gemini-3.7-flash": {
"cache_read_input_token_cost": 7.5e-08,
"cache_read_input_token_cost_flex": 3.75e-08,
"input_cost_per_token": 7.5e-07,
"input_cost_per_token_batches": 3.75e-07,
"input_cost_per_token_flex": 3.75e-07,
"litellm_provider": "gemini",
"max_input_tokens": 1048576,
"max_output_tokens": 65536,
"max_tokens": 65536,
"mode": "chat",
"output_cost_per_reasoning_token": 3.75e-06,
"output_cost_per_token": 3.75e-06,
"output_cost_per_token_batches": 1.875e-06,
"output_cost_per_token_flex": 1.875e-06,
"rpm": 2000,
"source": "https://ai.google.dev/pricing/gemini-3",
"supported_endpoints": [
"/v1/chat/completions",
"/v1/completions",
"/v1/batch"
],
"supported_modalities": [
"text",
"image",
"audio",
"video"
],
"supported_output_modalities": [
"text"
],
"supports_audio_output": false,
"supports_audio_input": true,
"supports_function_calling": true,
"supports_parallel_function_calling": true,
"supports_pdf_input": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_response_schema": true,
"supports_system_messages": true,
"supports_tool_choice": true,
"supports_url_context": true,
"supports_video_input": true,
"supports_vision": true,
"supports_web_search": true,
"supports_native_streaming": true,
"tpm": 800000,
"input_cost_per_token_priority": 1.35e-06,
"output_cost_per_token_priority": 6.75e-06,
"cache_read_input_token_cost_priority": 1.35e-07,
"search_context_cost_per_query": {
"search_context_size_low": 0.014,
"search_context_size_medium": 0.014,
"search_context_size_high": 0.014
},
"web_search_billing_unit": "per_query"
},
"gemini/gemini-omni-flash-preview": {
"input_cost_per_audio_token": 1.5e-06,
"input_cost_per_token": 1.5e-06,
@ -21031,6 +21448,61 @@
},
"web_search_billing_unit": "per_query"
},
"gemini-3.7-flash": {
"cache_read_input_token_cost": 7.5e-08,
"cache_read_input_token_cost_flex": 3.75e-08,
"input_cost_per_token": 7.5e-07,
"input_cost_per_token_batches": 3.75e-07,
"input_cost_per_token_flex": 3.75e-07,
"litellm_provider": "vertex_ai-language-models",
"max_input_tokens": 1048576,
"max_output_tokens": 65536,
"max_tokens": 65536,
"mode": "chat",
"output_cost_per_reasoning_token": 3.75e-06,
"output_cost_per_token": 3.75e-06,
"output_cost_per_token_batches": 1.875e-06,
"output_cost_per_token_flex": 1.875e-06,
"source": "https://ai.google.dev/pricing/gemini-3",
"supported_endpoints": [
"/v1/chat/completions",
"/v1/completions",
"/v1/batch"
],
"supported_modalities": [
"text",
"image",
"audio",
"video"
],
"supported_output_modalities": [
"text"
],
"supports_audio_output": false,
"supports_audio_input": true,
"supports_function_calling": true,
"supports_parallel_function_calling": true,
"supports_pdf_input": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_response_schema": true,
"supports_system_messages": true,
"supports_tool_choice": true,
"supports_url_context": true,
"supports_video_input": true,
"supports_vision": true,
"supports_web_search": true,
"supports_native_streaming": true,
"input_cost_per_token_priority": 1.35e-06,
"output_cost_per_token_priority": 6.75e-06,
"cache_read_input_token_cost_priority": 1.35e-07,
"search_context_cost_per_query": {
"search_context_size_low": 0.014,
"search_context_size_medium": 0.014,
"search_context_size_high": 0.014
},
"web_search_billing_unit": "per_query"
},
"gemini/gemini-2.5-pro-preview-tts": {
"cache_read_input_token_cost": 1.25e-07,
"cache_read_input_token_cost_above_200k_tokens": 2.5e-07,
@ -24703,7 +25175,10 @@
"supports_response_schema": true,
"supports_system_messages": true,
"supports_tool_choice": true,
"supports_vision": true
"supports_vision": true,
"supports_none_reasoning_effort": true,
"supports_xhigh_reasoning_effort": true,
"supports_minimal_reasoning_effort": true
},
"gpt-5.4-pro": {
"cache_read_input_token_cost": 3e-06,
@ -26120,11 +26595,12 @@
"supports_vision": true
},
"groq/llama-3.1-8b-instant": {
"deprecation_date": "2026-08-16",
"input_cost_per_token": 5e-08,
"litellm_provider": "groq",
"max_input_tokens": 128000,
"max_output_tokens": 8192,
"max_tokens": 8192,
"max_input_tokens": 131072,
"max_output_tokens": 131072,
"max_tokens": 131072,
"mode": "chat",
"output_cost_per_token": 8e-08,
"supports_function_calling": true,
@ -26132,9 +26608,10 @@
"supports_tool_choice": true
},
"groq/llama-3.3-70b-versatile": {
"deprecation_date": "2026-08-16",
"input_cost_per_token": 5.9e-07,
"litellm_provider": "groq",
"max_input_tokens": 128000,
"max_input_tokens": 131072,
"max_output_tokens": 32768,
"max_tokens": 32768,
"mode": "chat",
@ -26155,7 +26632,28 @@
"supports_response_schema": false,
"supports_tool_choice": true
},
"groq/meta-llama/llama-prompt-guard-2-22m": {
"input_cost_per_token": 3e-08,
"litellm_provider": "groq",
"max_input_tokens": 512,
"max_output_tokens": 512,
"max_tokens": 512,
"mode": "chat",
"output_cost_per_token": 3e-08,
"source": "https://console.groq.com/docs/models"
},
"groq/meta-llama/llama-prompt-guard-2-86m": {
"input_cost_per_token": 4e-08,
"litellm_provider": "groq",
"max_input_tokens": 512,
"max_output_tokens": 512,
"max_tokens": 512,
"mode": "chat",
"output_cost_per_token": 4e-08,
"source": "https://console.groq.com/docs/model/meta-llama/llama-prompt-guard-2-86m"
},
"groq/meta-llama/llama-guard-4-12b": {
"deprecation_date": "2026-03-05",
"input_cost_per_token": 2e-07,
"litellm_provider": "groq",
"max_input_tokens": 8192,
@ -26165,6 +26663,7 @@
"output_cost_per_token": 2e-07
},
"groq/meta-llama/llama-4-maverick-17b-128e-instruct": {
"deprecation_date": "2026-03-09",
"input_cost_per_token": 2e-07,
"litellm_provider": "groq",
"max_input_tokens": 131072,
@ -26178,6 +26677,7 @@
"supports_vision": true
},
"groq/meta-llama/llama-4-scout-17b-16e-instruct": {
"deprecation_date": "2026-07-17",
"input_cost_per_token": 1.1e-07,
"litellm_provider": "groq",
"max_input_tokens": 131072,
@ -26191,6 +26691,7 @@
"supports_vision": true
},
"groq/moonshotai/kimi-k2-instruct-0905": {
"deprecation_date": "2026-04-15",
"input_cost_per_token": 1e-06,
"output_cost_per_token": 3e-06,
"cache_read_input_token_cost": 5e-07,
@ -26208,8 +26709,8 @@
"input_cost_per_token": 1.5e-07,
"litellm_provider": "groq",
"max_input_tokens": 131072,
"max_output_tokens": 32766,
"max_tokens": 32766,
"max_output_tokens": 65536,
"max_tokens": 65536,
"mode": "chat",
"output_cost_per_token": 6e-07,
"search_context_cost_per_query": {
@ -26229,8 +26730,8 @@
"input_cost_per_token": 7.5e-08,
"litellm_provider": "groq",
"max_input_tokens": 131072,
"max_output_tokens": 32768,
"max_tokens": 32768,
"max_output_tokens": 65536,
"max_tokens": 65536,
"mode": "chat",
"output_cost_per_token": 3e-07,
"search_context_cost_per_query": {
@ -26265,7 +26766,26 @@
"supports_tool_choice": true,
"supports_web_search": true
},
"groq/canopylabs/orpheus-v1-english": {
"input_cost_per_character": 2.2e-05,
"litellm_provider": "groq",
"max_input_tokens": 4000,
"max_output_tokens": 50000,
"max_tokens": 50000,
"mode": "audio_speech",
"source": "https://console.groq.com/docs/model/canopylabs/orpheus-v1-english"
},
"groq/canopylabs/orpheus-arabic-saudi": {
"input_cost_per_character": 4e-05,
"litellm_provider": "groq",
"max_input_tokens": 4000,
"max_output_tokens": 50000,
"max_tokens": 50000,
"mode": "audio_speech",
"source": "https://console.groq.com/docs/models"
},
"groq/playai-tts": {
"deprecation_date": "2025-12-31",
"input_cost_per_character": 5e-05,
"litellm_provider": "groq",
"max_input_tokens": 10000,
@ -26273,7 +26793,23 @@
"max_tokens": 10000,
"mode": "audio_speech"
},
"groq/qwen/qwen3.6-27b": {
"input_cost_per_token": 6e-07,
"litellm_provider": "groq",
"max_input_tokens": 131072,
"max_output_tokens": 16384,
"max_tokens": 16384,
"mode": "chat",
"output_cost_per_token": 3e-06,
"source": "https://console.groq.com/docs/model/qwen/qwen3.6-27b",
"supports_function_calling": true,
"supports_reasoning": true,
"supports_response_schema": false,
"supports_tool_choice": true,
"supports_vision": true
},
"groq/qwen/qwen3-32b": {
"deprecation_date": "2026-07-17",
"input_cost_per_token": 2.9e-07,
"litellm_provider": "groq",
"max_input_tokens": 131000,
@ -27484,6 +28020,93 @@
"max_tokens": 131072,
"mode": "chat",
"output_cost_per_token": 4.25e-06,
"search_context_cost_per_query": {
"search_context_size_high": 0.0025,
"search_context_size_low": 0.0025,
"search_context_size_medium": 0.0025
},
"source": "https://dev.meta.ai/docs/getting-started/pricing-rate-limits",
"supported_endpoints": [
"/v1/chat/completions",
"/v1/responses",
"/v1/messages"
],
"supported_modalities": [
"text",
"image",
"video"
],
"supported_output_modalities": [
"text"
],
"supports_function_calling": true,
"supports_minimal_reasoning_effort": true,
"supports_parallel_function_calling": true,
"supports_pdf_input": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_response_schema": true,
"supports_system_messages": true,
"supports_tool_choice": true,
"supports_vision": true,
"supports_web_search": true,
"supports_xhigh_reasoning_effort": true
},
"meta/muse-spark-1.2": {
"cache_read_input_token_cost": 1.5e-07,
"input_cost_per_token": 1.25e-06,
"litellm_provider": "meta",
"max_input_tokens": 1048576,
"max_output_tokens": 131072,
"max_tokens": 131072,
"mode": "chat",
"output_cost_per_token": 4.25e-06,
"search_context_cost_per_query": {
"search_context_size_high": 0.0025,
"search_context_size_low": 0.0025,
"search_context_size_medium": 0.0025
},
"source": "https://dev.meta.ai/docs/getting-started/pricing-rate-limits",
"supported_endpoints": [
"/v1/chat/completions",
"/v1/responses",
"/v1/messages"
],
"supported_modalities": [
"text",
"image",
"video"
],
"supported_output_modalities": [
"text"
],
"supports_function_calling": true,
"supports_minimal_reasoning_effort": true,
"supports_parallel_function_calling": true,
"supports_pdf_input": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_response_schema": true,
"supports_system_messages": true,
"supports_tool_choice": true,
"supports_vision": true,
"supports_web_search": true,
"supports_xhigh_reasoning_effort": true
},
"meta/muse-spark-1.2-contributor": {
"cache_read_input_token_cost": 2e-09,
"input_cost_per_token": 1e-07,
"litellm_provider": "meta",
"max_input_tokens": 1048576,
"max_output_tokens": 131072,
"max_tokens": 131072,
"mode": "chat",
"output_cost_per_token": 2e-07,
"search_context_cost_per_query": {
"search_context_size_high": 0.0025,
"search_context_size_low": 0.0025,
"search_context_size_medium": 0.0025
},
"source": "https://dev.meta.ai/docs/getting-started/pricing-rate-limits",
"supported_endpoints": [
"/v1/chat/completions",
@ -40662,6 +41285,27 @@
"supports_vision": true,
"supports_web_search": true
},
"xai/grok-4.6": {
"cache_read_input_token_cost": 5e-07,
"cache_read_input_token_cost_above_200k_tokens": 1e-06,
"input_cost_per_token": 2e-06,
"input_cost_per_token_above_200k_tokens": 4e-06,
"litellm_provider": "xai",
"max_input_tokens": 500000,
"max_output_tokens": 500000,
"max_tokens": 500000,
"mode": "chat",
"output_cost_per_token": 6e-06,
"output_cost_per_token_above_200k_tokens": 1.2e-05,
"source": "https://docs.x.ai/developers/models",
"supports_function_calling": true,
"supports_prompt_caching": true,
"supports_reasoning": true,
"supports_response_schema": true,
"supports_tool_choice": true,
"supports_vision": true,
"supports_web_search": true
},
"xai/grok-beta": {
"input_cost_per_token": 5e-06,
"litellm_provider": "xai",
@ -45826,11 +46470,15 @@
},
"bedrock_mantle/openai.gpt-5.6-sol": {
"input_cost_per_token": 5.5e-06,
"input_cost_per_token_above_272k_tokens": 1.1e-05,
"cache_creation_input_token_cost": 6.875e-06,
"cache_creation_input_token_cost_above_272k_tokens": 1.375e-05,
"cache_read_input_token_cost": 5.5e-07,
"cache_read_input_token_cost_above_272k_tokens": 1.1e-06,
"output_cost_per_token": 3.3e-05,
"output_cost_per_token_above_272k_tokens": 4.95e-05,
"litellm_provider": "bedrock_mantle",
"max_input_tokens": 272000,
"max_input_tokens": 1000000,
"max_output_tokens": 128000,
"max_tokens": 128000,
"mode": "responses",
@ -45854,11 +46502,15 @@
},
"bedrock_mantle/openai.gpt-5.6-terra": {
"input_cost_per_token": 2.2e-06,
"input_cost_per_token_above_272k_tokens": 4.4e-06,
"cache_creation_input_token_cost": 2.75e-06,
"cache_creation_input_token_cost_above_272k_tokens": 5.5e-06,
"cache_read_input_token_cost": 2.2e-07,
"cache_read_input_token_cost_above_272k_tokens": 4.4e-07,
"output_cost_per_token": 1.32e-05,
"output_cost_per_token_above_272k_tokens": 1.98e-05,
"litellm_provider": "bedrock_mantle",
"max_input_tokens": 272000,
"max_input_tokens": 1000000,
"max_output_tokens": 128000,
"max_tokens": 128000,
"mode": "responses",
@ -45882,11 +46534,15 @@
},
"bedrock_mantle/openai.gpt-5.6-luna": {
"input_cost_per_token": 2.2e-07,
"input_cost_per_token_above_272k_tokens": 4.4e-07,
"cache_creation_input_token_cost": 2.75e-07,
"cache_creation_input_token_cost_above_272k_tokens": 5.5e-07,
"cache_read_input_token_cost": 2.2e-08,
"cache_read_input_token_cost_above_272k_tokens": 4.4e-08,
"output_cost_per_token": 1.32e-06,
"output_cost_per_token_above_272k_tokens": 1.98e-06,
"litellm_provider": "bedrock_mantle",
"max_input_tokens": 272000,
"max_input_tokens": 1000000,
"max_output_tokens": 128000,
"max_tokens": 128000,
"mode": "responses",

4
litellm/proxy/_experimental/mcp_server/auth/user_api_key_auth_mcp.py
View file

@ -1190,7 +1190,7 @@ class MCPRequestHandler:
DEPRECATED: This method is deprecated in favor of server-specific auth headers using the format x-mcp-{{server_alias}}-{{header_name}} instead.
"""
mcp_client_side_auth_header_name: Final[str] = MCPRequestHandler._get_mcp_client_side_auth_header_name()
mcp_client_side_auth_header_name: Final[str] = MCPRequestHandler.get_mcp_client_side_auth_header_name()
auth_header: Final = headers.get(mcp_client_side_auth_header_name)
if auth_header:
verbose_logger.warning(
@ -1265,7 +1265,7 @@ class MCPRequestHandler:
return oauth2_headers
@staticmethod
def _get_mcp_client_side_auth_header_name() -> str:
def get_mcp_client_side_auth_header_name() -> str:
"""
Get the header name used to pass the MCP auth header to the MCP server

107
litellm/proxy/_experimental/mcp_server/db.py
View file

@ -552,13 +552,20 @@ async def get_all_mcp_servers(
) -> list[LiteLLM_MCPServerTable]:
"""
Returns mcp servers from the db, optionally filtered by approval_status.
Pass approval_status=None to return all servers regardless of approval state.
Pass approval_status=None to return every server except drafts, which back the admin OAuth
session flow, are addressable only by their own server_id, and must never appear in a listing.
NULL approval_status predates the approval workflow, so those rows are kept explicitly rather
than dropped by a bare inequality, which SQL evaluates as NULL and would silently hide them.
"""
try:
where: Final[prisma_db_types.LiteLLM_MCPServerTableWhereInput] = {}
if approval_status is not None:
where["approval_status"] = approval_status
mcp_servers: Final = await _db_find_mcp_server_rows(prisma_client, where if where else {})
where: Final[prisma_db_types.LiteLLM_MCPServerTableWhereInput] = (
{"approval_status": approval_status}
if approval_status is not None
# mutable-ok: prisma where-inputs must be plain dicts, and both `NOT` and `not` drop
# NULL rows (measured), so the OR is the only NULL-preserving way to exclude drafts
else {"OR": [{"approval_status": None}, {"approval_status": {"not": MCPApprovalStatus.draft}}]}
)
mcp_servers: Final = await _db_find_mcp_server_rows(prisma_client, where)
tables: Final = [LiteLLM_MCPServerTable.model_validate(mcp_server.model_dump()) for mcp_server in mcp_servers]
for table in tables:
@ -814,6 +821,96 @@ async def create_mcp_server(
return new_mcp_server
async def create_draft_mcp_server(
prisma_client: PrismaClient,
data: NewMCPServerRequest,
touched_by: str,
ttl_seconds: int,
server_id: str | None = None,
) -> LiteLLM_MCPServerTable:
"""
Persist a short-lived draft row backing the admin OAuth "Authorize & Fetch Token" flow.
The draft lives in the database rather than in process memory so that the /register,
/authorize and /token legs resolve it whichever worker or replica accepts each request.
Writing is strictly create-if-absent. Any existing row for the id is returned untouched, which
covers both a live draft for this same session and a real server the edit form is
re-authorizing against its own id, where writing a draft would collide on the primary key.
Each click of Authorize mints a fresh id, so nothing is lost by never overwriting, and it is
what makes concurrent callers sharing one id safe rather than mutually destructive.
"""
draft_id: Final = server_id or data.server_id or str(uuid.uuid4())
await _prune_expired_draft_mcp_servers(prisma_client, ttl_seconds)
existing: Final = await _db_find_mcp_server_row(prisma_client, draft_id)
if existing is not None:
# Already usable by every worker, whether it is a live draft for this same session or a
# real server the edit form is re-authorizing. Either way there is nothing to write, and
# not writing is what keeps concurrent callers for one server_id from racing each other.
return LiteLLM_MCPServerTable.model_validate(existing.model_dump())
draft_payload: Final = data.model_copy(update={"server_id": draft_id, "approval_status": MCPApprovalStatus.draft})
try:
return await create_mcp_server(prisma_client, draft_payload, touched_by)
except Exception:
# Lost the create race: the read above and this create are two statements, not one. The
# winner wrote a draft for this same session, so adopt it rather than failing a caller
# whose session is in fact ready. Anything else still raises.
raced: Final = await _db_find_mcp_server_row(prisma_client, draft_id)
if raced is None or raced.approval_status != MCPApprovalStatus.draft:
raise
return LiteLLM_MCPServerTable.model_validate(raced.model_dump())
async def _prune_expired_draft_mcp_servers(prisma_client: PrismaClient, ttl_seconds: int) -> None:
"""Drop drafts already past ``ttl_seconds``, so abandoned OAuth sessions do not accumulate.
Runs on each draft write rather than on a schedule, mirroring the in-memory cache this
replaces, which pruned on every store. Expired drafts are unreadable by then anyway, so the
only thing at stake is row count, and the work is bounded by how often admins authorize.
"""
cutoff: Final = datetime.now(timezone.utc) - timedelta(seconds=max(1, ttl_seconds))
# Age is filtered here rather than in the query: the draft set is bounded by how many OAuth
# authorizations are in flight, so it is a handful of rows even on a busy proxy.
drafts: Final = await _db_find_mcp_server_rows(
prisma_client,
where={"approval_status": MCPApprovalStatus.draft},
)
for row in drafts:
# A row without a timestamp has no age to judge, so leave it rather than guess it is stale.
# Two workers sweeping the same row is harmless: prisma's delete returns None for a row
# that is already gone rather than raising, so the loser of that race is a no-op.
if row.updated_at is not None and row.updated_at < cutoff:
await delete_mcp_server(prisma_client, row.server_id)
async def get_draft_mcp_server(
prisma_client: PrismaClient, server_id: str, ttl_seconds: int
) -> LiteLLM_MCPServerTable | None:
"""
Return the draft row for ``server_id`` if it has not yet aged past ``ttl_seconds``, else None.
Age is enforced in the query rather than by a sweeper so an expired draft is unreadable the
moment it lapses, regardless of which process last ran a cleanup.
"""
cutoff: Final = datetime.now(timezone.utc) - timedelta(seconds=max(1, ttl_seconds))
draft_rows: Final = await _db_find_mcp_server_rows(
prisma_client,
where={
"server_id": server_id,
"approval_status": MCPApprovalStatus.draft,
"updated_at": {"gte": cutoff},
},
)
if not draft_rows:
return None
table: Final = LiteLLM_MCPServerTable.model_validate(draft_rows[0].model_dump())
decrypt_global_env_var_values(table.env_vars)
return table
async def update_mcp_server(
prisma_client: PrismaClient,
data: UpdateMCPServerRequest,

26
litellm/proxy/_experimental/mcp_server/mcp_server_manager.py
View file

@ -118,6 +118,7 @@ from litellm.proxy._experimental.mcp_server.utils import (
is_short_mcp_tool_prefix_enabled,
iter_known_server_prefixes,
iter_known_tool_name_spellings,
logging_safe_mcp_headers,
match_known_server_prefix,
match_known_tool_name,
merge_mcp_headers,
@ -1598,6 +1599,9 @@ class MCPServerManager:
manual_token_url,
)
use_issuer_anchor = _uses_issuer_anchor(manual_issuer, is_discovery_auth_type or obo_needs_discovery)
configured_authorization_url = manual_authorization_url
configured_token_url = manual_token_url
configured_registration_url = manual_registration_url
manual_authorization_url, manual_token_url, manual_registration_url = _endpoints_yield_to_issuer(
manual_issuer,
is_discovery_auth_type,
@ -1724,6 +1728,9 @@ class MCPServerManager:
authorization_url=resolved_authorization_url,
token_url=resolved_token_url,
registration_url=resolved_registration_url,
configured_authorization_url=configured_authorization_url,
configured_token_url=configured_token_url,
configured_registration_url=configured_registration_url,
token_endpoint_auth_method=server_config.get("token_endpoint_auth_method", None),
# TODO: utility fn the default values
transport=server_config.get("transport", MCPTransport.http),
@ -2169,6 +2176,9 @@ class MCPServerManager:
is_discovery_auth_type
or self._obo_needs_endpoint_discovery(auth_type, token_exchange_endpoint, manual_token_url),
)
configured_authorization_url: Final = manual_authorization_url
configured_token_url: Final = manual_token_url
configured_registration_url: Final = manual_registration_url
manual_authorization_url, manual_token_url, manual_registration_url = _endpoints_yield_to_issuer(
manual_issuer,
is_discovery_auth_type,
@ -2221,6 +2231,9 @@ class MCPServerManager:
authorization_url=manual_authorization_url or getattr(gated_oauth_metadata, "authorization_url", None),
token_url=manual_token_url or getattr(gated_oauth_metadata, "token_url", None),
registration_url=manual_registration_url or getattr(gated_oauth_metadata, "registration_url", None),
configured_authorization_url=configured_authorization_url,
configured_token_url=configured_token_url,
configured_registration_url=configured_registration_url,
token_endpoint_auth_method=(
credentials_dict.get("token_endpoint_auth_method") if credentials_dict else None
),
@ -4603,6 +4616,7 @@ class MCPServerManager:
),
"user_api_key_hash": (getattr(user_api_key_auth, "api_key_hash", None) if user_api_key_auth else None),
"incoming_bearer_token": incoming_bearer_token,
"headers": logging_safe_mcp_headers(raw_headers),
}
# Create MCP request object for processing
@ -5856,9 +5870,9 @@ class MCPServerManager:
args=getattr(server, "args", None) or [],
env=getattr(server, "env", None) or {},
issuer=server.issuer,
authorization_url=server.authorization_url,
token_url=server.token_url,
registration_url=server.registration_url,
authorization_url=server.configured_authorization_url or server.authorization_url,
token_url=server.configured_token_url or server.token_url,
registration_url=server.configured_registration_url or server.registration_url,
oauth2_flow=server.oauth2_flow,
dcr_bridge=server.dcr_bridge,
token_exchange_endpoint=server.token_exchange_endpoint,
@ -5966,9 +5980,9 @@ class MCPServerManager:
args=getattr(server, "args", None) or [],
env=getattr(server, "env", None) or {},
issuer=server.issuer,
authorization_url=server.authorization_url,
token_url=server.token_url,
registration_url=server.registration_url,
authorization_url=server.configured_authorization_url or server.authorization_url,
token_url=server.configured_token_url or server.token_url,
registration_url=server.configured_registration_url or server.registration_url,
oauth2_flow=server.oauth2_flow,
token_exchange_endpoint=server.token_exchange_endpoint,
audience=server.audience,

99
litellm/proxy/_experimental/mcp_server/sampling_handler.py
View file

@ -1042,100 +1042,15 @@ def _build_sampling_request(
raw_headers: dict[str, str] | None = None,
client_ip: str | None = None,
) -> "Request":
"""Build a synthetic FastAPI Request for sampling sub-calls.
"""The synthetic FastAPI Request for sampling sub-calls, carrying the original
MCP connection's headers and client IP."""
from litellm.proxy._experimental.mcp_server.utils import build_synthetic_mcp_request
Converts the original MCP connection's HTTP headers into ASGI
scope format so that ``add_litellm_data_to_request`` can apply
header-dependent guardrails, tag-based routing, trace correlation,
and ``forward_llm_provider_auth_headers``.
Key fields populated:
- **headers**: All original HTTP headers are forwarded (except
hop-by-hop: content-length, transfer-encoding). This ensures
``traceparent``, ``authorization``, ``user-agent``, and
``x-litellm-api-key`` are visible to pre-call utils.
- **client**: The ASGI ``(host, port)`` tuple so that
``request.client.host`` returns the real client IP for
IP-based routing and guardrails.
- **server**: Derived from the running proxy's ``server_host``
/ ``server_port`` when available, avoiding the misleading
``127.0.0.1:0`` placeholder.
- **x-forwarded-for**: Injected from ``client_ip`` if the
original headers don't already carry it, as a fallback for
IP attribution.
"""
from fastapi import Request
# --- Build ASGI headers ---
_scope_headers: Final[list[tuple[bytes, bytes]]] = [(b"content-type", b"application/json")]
# Hop-by-hop headers that must NOT be forwarded into the
# synthetic request (they describe the original HTTP framing,
# not the logical request).
_HOP_BY_HOP: Final = frozenset(
{
"content-length",
"transfer-encoding",
"connection",
"keep-alive",
"upgrade",
"te",
"trailer",
}
return build_synthetic_mcp_request(
path="/mcp/sampling/createMessage",
raw_headers=raw_headers,
client_ip=client_ip,
)
if raw_headers:
for hdr_name, hdr_value in raw_headers.items():
_key = hdr_name.lower()
# Skip content-type (already set), x-forwarded-for (use resolved
# client_ip instead to prevent spoofing), and hop-by-hop headers
if _key in {"content-type", "x-forwarded-for"} or _key in _HOP_BY_HOP:
continue
_scope_headers.append(
(
_key.encode("latin-1", errors="replace"),
hdr_value.encode("utf-8"),
)
)
# Inject x-forwarded-for from captured client_ip if the
# original headers don't already carry it
if client_ip and not any(h[0] == b"x-forwarded-for" for h in _scope_headers):
_scope_headers.append((b"x-forwarded-for", client_ip.encode("utf-8")))
# --- Derive server (host, port) from the running proxy ---
_server_host = "127.0.0.1"
_server_port = 4000 # LiteLLM default
try:
from litellm.proxy import proxy_server
_proxy_host: Final[str | None] = getattr(proxy_server, "server_host", None)
_proxy_port: Final[str | int | None] = getattr(proxy_server, "server_port", None)
if _proxy_host:
_server_host = str(_proxy_host)
if _proxy_port:
_server_port = int(_proxy_port)
except (ImportError, AttributeError, TypeError, ValueError):
pass
# --- Build ASGI client tuple for request.client.host ---
_client_tuple = None
if client_ip:
_client_tuple = (client_ip, 0)
scope: Final[dict[str, object]] = {
"type": "http",
"method": "POST",
"path": "/mcp/sampling/createMessage",
"scheme": "http",
"server": (_server_host, _server_port),
"query_string": b"",
"root_path": "",
"headers": _scope_headers,
}
if _client_tuple is not None:
scope["client"] = _client_tuple
return Request(scope=scope)
async def _build_completion_kwargs(

36
litellm/proxy/_experimental/mcp_server/server.py
View file

@ -58,9 +58,11 @@ from litellm.proxy._experimental.mcp_server.utils import (
LITELLM_MCP_SERVER_VERSION,
MCPMissingUserEnvVarsError,
add_server_prefix_to_name,
build_synthetic_mcp_request,
extract_mcp_tool_result_error_message,
get_server_prefix,
iter_known_server_prefixes,
logging_safe_mcp_headers,
match_known_tool_name,
)
from litellm.proxy._types import (
@ -860,11 +862,11 @@ if MCP_AVAILABLE:
name: str,
arguments: dict[str, object],
user_api_key_auth: UserAPIKeyAuth,
raw_headers: Mapping[str, str] | None = None,
client_ip: str | None = None,
) -> LiteLLMLoggingObj | None:
"""Run the pre-call pipeline (guardrails + logging setup) for a virtual
mcp_tool_call so the SSE path spend-logs like the REST path."""
from fastapi import Request
from litellm.proxy.common_request_processing import (
ProxyBaseLLMRequestProcessing,
)
@ -874,13 +876,10 @@ if MCP_AVAILABLE:
proxy_logging_obj,
)
request: Final = Request(
scope={
"type": "http",
"method": "POST",
"path": "/mcp/tools/call",
"headers": [(b"content-type", b"application/json")],
}
request: Final = build_synthetic_mcp_request(
path="/mcp/tools/call",
raw_headers=raw_headers,
client_ip=client_ip,
)
_, virtual_logging_obj = await ProxyBaseLLMRequestProcessing(
data={"name": name, "arguments": arguments}
@ -952,7 +951,11 @@ if MCP_AVAILABLE:
assert user_api_key_auth is not None # guaranteed by the flag check above
virtual_logging_obj: Final = await _build_virtual_call_logging_obj(
name=name, arguments=args, user_api_key_auth=user_api_key_auth
name=name,
arguments=args,
user_api_key_auth=user_api_key_auth,
raw_headers=raw_headers,
client_ip=client_ip,
)
return await handle_mcp_tool_call(
tool_name=args.get("tool_name", ""),
@ -979,7 +982,6 @@ if MCP_AVAILABLE:
Raises:
HTTPException: If tool not found or arguments missing
"""
from fastapi import Request
from mcp.server.lowlevel.server import request_ctx
from mcp.types import CallToolResult
@ -1041,13 +1043,10 @@ if MCP_AVAILABLE:
body_data["litellm_trace_id"] = chain_id
body_data["litellm_session_id"] = chain_id
request: Final = Request(
scope={
"type": "http",
"method": "POST",
"path": "/mcp/tools/call",
"headers": [(b"content-type", b"application/json")],
}
request: Final = build_synthetic_mcp_request(
path="/mcp/tools/call",
raw_headers=raw_headers,
client_ip=_client_ip,
)
if user_api_key_auth is not None:
data = await add_litellm_data_to_request(
@ -1905,6 +1904,7 @@ if MCP_AVAILABLE:
"litellm_trace_id": effective_litellm_trace_id,
"metadata": {
"spend_logs_metadata": spend_logs_metadata,
"headers": logging_safe_mcp_headers(raw_headers),
**({"tags": request_tags} if request_tags else {}),
},
# Provide a small input payload for standard logging

274
litellm/proxy/_experimental/mcp_server/utils.py
View file

@ -7,12 +7,38 @@ import importlib
import json
import os
import re
import typing
from collections.abc import Iterable, Iterator, Mapping, MutableMapping, MutableSequence
from typing import Any, Final
from collections.abc import Set as AbstractSet
from typing import Any, Final, Protocol
from urllib.parse import quote
from litellm.types.mcp_server.mcp_server_manager import MCPServer
if typing.TYPE_CHECKING:
from fastapi import Request
class _McpServerLike(Protocol):
@property
def server_id(self) -> str: ...
@property
def server_name(self) -> str | None: ...
@property
def alias(self) -> str | None: ...
@property
def short_prefix(self) -> str | None: ...
class McpServerPayloadLike(Protocol):
alias: str | None
@property
def server_name(self) -> str | None: ...
@property
def tool_name_to_display_name(self) -> Mapping[str, str] | None: ...
# Constants
#
# NOTE: The environment-backed values below are read once, when this module is
@ -102,7 +128,7 @@ def compute_short_server_prefix(server_id: str, attempt: int = 0) -> str:
# at the end so the first emitted char comes from the high-order
# bits of the digest (which is the position we constrain to be
# alphabetic).
chars: Final = []
chars: Final[list[str]] = []
for position in range(SHORT_MCP_TOOL_PREFIX_LENGTH):
is_first_char = position == SHORT_MCP_TOOL_PREFIX_LENGTH - 1
alphabet = _BASE52_ALPHA_ALPHABET if is_first_char else _BASE62_ALPHABET
@ -176,34 +202,34 @@ def lookup_mcp_server_auth_in_headers(
MCP_TOOL_ALLOWLIST_ENFORCED_KEY: Final = "tool_allowlist_enforced"
def _parse_mcp_info_dict(mcp_info: Any) -> dict[str, Any] | None:
def _parse_mcp_info_dict(mcp_info: object) -> Mapping[str, object] | None:
if mcp_info is None:
return None
if isinstance(mcp_info, dict):
return mcp_info
if isinstance(mcp_info, str):
try:
parsed: Final = json.loads(mcp_info)
parsed: Final[object] = json.loads(mcp_info)
except (ValueError, TypeError):
return None
return parsed if isinstance(parsed, dict) else None
return None
def is_server_tool_allowlist_enforced(mcp_server: Any) -> bool:
def is_server_tool_allowlist_enforced(mcp_server: object) -> bool:
mcp_info: Final = _parse_mcp_info_dict(getattr(mcp_server, "mcp_info", None))
if not mcp_info:
return False
return bool(mcp_info.get(MCP_TOOL_ALLOWLIST_ENFORCED_KEY))
def server_applies_tool_allowlist(mcp_server: Any) -> bool:
def server_applies_tool_allowlist(mcp_server: object) -> bool:
"""Whether server-level allowed_tools whitelist filtering is active."""
allowed_tools: Final = getattr(mcp_server, "allowed_tools", None) or []
allowed_tools: Final[object] = getattr(mcp_server, "allowed_tools", None) or []
return is_server_tool_allowlist_enforced(mcp_server) or bool(allowed_tools)
def validate_and_normalize_mcp_server_payload(payload: Any) -> None:
def validate_and_normalize_mcp_server_payload(payload: McpServerPayloadLike) -> None:
"""
Validate and normalize MCP server payload fields (server_name, alias, and
tool_name_to_display_name).
@ -233,8 +259,8 @@ def validate_and_normalize_mcp_server_payload(payload: Any) -> None:
validate_tool_display_names(payload.tool_name_to_display_name)
# Alias normalization and defaulting
alias = getattr(payload, "alias", None)
server_name: Final = getattr(payload, "server_name", None)
alias: str | None = getattr(payload, "alias", None)
server_name: Final[str | None] = getattr(payload, "server_name", None)
if not alias and server_name:
alias = normalize_server_name(server_name)
@ -257,7 +283,7 @@ def add_server_prefix_to_name(name: str, server_name: str) -> str:
)
def get_server_prefix(server: Any) -> str:
def get_server_prefix(server: object) -> str:
"""Return the prefix for a server.
When the short-prefix mode is enabled (``LITELLM_USE_SHORT_MCP_TOOL_PREFIX``)
@ -270,23 +296,26 @@ def get_server_prefix(server: Any) -> str:
alias if present, else server_name, else server_id.
"""
if is_short_mcp_tool_prefix_enabled():
cached: Final = getattr(server, "short_prefix", None)
cached: Final[str | None] = getattr(server, "short_prefix", None)
if cached:
return cached
server_id: Final = getattr(server, "server_id", None)
server_id: Final[str | None] = getattr(server, "server_id", None)
if server_id:
return compute_short_server_prefix(server_id)
if hasattr(server, "alias") and server.alias:
return server.alias
if hasattr(server, "server_name") and server.server_name:
return server.server_name
alias: Final[str | None] = getattr(server, "alias", None)
if alias:
return alias
server_name: Final[str | None] = getattr(server, "server_name", None)
if server_name:
return server_name
if hasattr(server, "server_id"):
return server.server_id
fallback_server_id: Final[str] = getattr(server, "server_id", "")
return fallback_server_id
return ""
def iter_known_server_prefixes(server: Any) -> Iterator[str]:
def iter_known_server_prefixes(server: _McpServerLike) -> Iterator[str]:
"""Yield every prefix form that may appear in tool names for ``server``.
Always includes the *current* prefix returned by ``get_server_prefix``.
@ -304,7 +333,7 @@ def iter_known_server_prefixes(server: Any) -> Iterator[str]:
yield from _emit(get_server_prefix(server))
yield from _emit(getattr(server, "short_prefix", None))
server_id: Final = getattr(server, "server_id", None)
server_id: Final[str | None] = getattr(server, "server_id", None)
if server_id:
try:
yield from _emit(compute_short_server_prefix(server_id))
@ -397,7 +426,7 @@ def match_known_server_prefix(name: str, known_prefixes: Iterable[str]) -> tuple
return None
def strip_known_server_prefix(name: str, server: Any | None) -> str:
def strip_known_server_prefix(name: str, server: _McpServerLike | None) -> str:
"""Strip ``server``'s registered prefix from a prefixed tool/resource name.
Unlike :func:`split_server_prefix_from_name`, which guesses the boundary at
@ -420,7 +449,7 @@ def strip_known_server_prefix(name: str, server: Any | None) -> str:
def is_tool_name_prefixed(
tool_name: str,
known_server_prefixes: set | None = None,
known_server_prefixes: AbstractSet[str] | None = None,
) -> bool:
"""
Check if tool name has a known MCP server prefix.
@ -640,7 +669,7 @@ def parse_admin_env_vars(
if raw is None:
continue
if hasattr(raw, "model_dump"):
entry = raw.model_dump()
entry: Mapping[str, object] = raw.model_dump()
elif isinstance(raw, dict):
entry = raw
else:
@ -837,3 +866,202 @@ def set_mcp_tool_result_structured_content(result: object, value: object) -> boo
return True
except (AttributeError, TypeError, ValueError):
return False
_HOP_BY_HOP_HEADERS: Final = frozenset(
{
"content-length",
"transfer-encoding",
"connection",
"keep-alive",
"upgrade",
"te",
"trailer",
}
)
_SYNTHETIC_REQUEST_EXCLUDED_HEADERS: Final = _HOP_BY_HOP_HEADERS | frozenset(
{"content-type", "host", "x-forwarded-for"}
)
_SYNTHETIC_REQUEST_SERVER: Final = ("127.0.0.1", 4000)
_MCP_SERVER_AUTH_HEADER_PREFIX: Final = "x-mcp-"
def _custom_litellm_key_header_name() -> str | None:
"""``general_settings.litellm_key_header_name``, the deployment's custom header name for
the proxy virtual key, so it is stripped from observability copies like the standard ones."""
try:
from litellm.proxy.proxy_server import general_settings
except ImportError:
return None
return general_settings.get("litellm_key_header_name") if general_settings else None
def _mcp_client_side_auth_header_name() -> str:
"""The header name the client passes the upstream MCP credential in, falling back to the
default when ``general_settings`` is unavailable (the SDK, outside a running proxy)."""
from .auth.user_api_key_auth_mcp import MCPRequestHandler
try:
return MCPRequestHandler.get_mcp_client_side_auth_header_name()
except ImportError:
return MCPRequestHandler.LITELLM_MCP_AUTH_HEADER_NAME
def _identity_header_names() -> frozenset[str]:
"""Lowercased header names the deployment reads the caller's identity out of. A name here
is a claim about who the caller is rather than a secret, and ``get_user_from_headers``
resolves it off the request this module reconstructs, so dropping one would lose end user
attribution on the MCP paths that leave ``end_user_id`` unset at connect time.
``user_header_mappings`` is accepted as a bare mapping as well as a list of them, matching
``get_internal_user_header_from_mapping`` and ``get_customer_user_header_from_mapping``.
Iterating the bare form without normalizing yields its keys, which would silently exempt
nothing."""
try:
from litellm.proxy.proxy_server import general_settings
except ImportError:
return frozenset()
if not general_settings:
return frozenset()
user_header: Final = general_settings.get("user_header_name")
configured: Final = general_settings.get("user_header_mappings")
mappings: Final = configured if isinstance(configured, list) else (configured,) if configured else ()
mapped: Final = (mapping.get("header_name") for mapping in mappings if isinstance(mapping, Mapping))
return frozenset(name.lower() for name in (user_header, *mapped) if isinstance(name, str) and name)
def _forwarded_upstream_header_names() -> frozenset[str]:
"""Lowercased header names that a configured MCP server forwards upstream through its
``extra_headers`` allowlist. The names are chosen by the admin, so no prefix rule can
recognize them, and a caller supplied value under one of them is an upstream credential.
``authorization`` is left out because ``clean_headers`` already strips it, and claiming it
here would change which header ``authenticated_with_header`` resolves to on the oauth
passthrough config, which lists it in ``extra_headers`` by design. Identity headers are
left out for the same reason: naming one in ``extra_headers`` forwards the caller's
identity upstream, it does not turn that identity into a secret."""
try:
from .mcp_server_manager import global_mcp_server_manager
except ImportError:
return frozenset()
exempt: Final = _identity_header_names() | frozenset({"authorization"})
return frozenset(
name.lower()
for server in global_mcp_server_manager.get_registry().values()
for name in (server.extra_headers or ())
if name.lower() not in exempt
)
def _upstream_credential_headers(header_names: Iterable[str]) -> frozenset[str]:
"""Lowercased names of the headers in ``header_names`` that carry an upstream MCP
credential rather than request context: the configured client side auth header, any
header name a configured server forwards upstream via ``extra_headers``, and the
per-server ``x-mcp-{alias}-{header}`` family. ``clean_headers`` only knows the
credential headers of the chat completions path, so these are dropped on top of it.
"""
from .auth.user_api_key_auth_mcp import MCPRequestHandler
non_credential: Final = frozenset(
{
MCPRequestHandler.LITELLM_MCP_SERVERS_HEADER_NAME.lower(),
MCPRequestHandler.LITELLM_MCP_ACCESS_GROUPS_HEADER_NAME.lower(),
}
)
client_side_auth: Final = _mcp_client_side_auth_header_name().lower()
forwarded_upstream: Final = _forwarded_upstream_header_names()
return frozenset(
name
for name in (raw_name.lower() for raw_name in header_names)
if name == client_side_auth
or name in forwarded_upstream
or (name.startswith(_MCP_SERVER_AUTH_HEADER_PREFIX) and name not in non_credential)
)
def build_synthetic_mcp_request(
*,
path: str,
raw_headers: Mapping[str, str] | None = None,
client_ip: str | None = None,
) -> "Request":
"""A synthetic FastAPI ``Request`` carrying the MCP connection's HTTP headers.
The MCP protocol transports do not hand a per-call ``Request`` to the tool
handlers, so one is reconstructed from the connection's ``raw_headers``. That
lets ``add_litellm_data_to_request`` derive ``metadata.headers``,
``proxy_server_request``, header-based tags, guardrails and trace correlation
exactly as on the chat completions path. Hop-by-hop headers describe the
original HTTP framing rather than the logical request, so they are dropped, and
``x-forwarded-for`` comes from the resolved ``client_ip`` to avoid spoofing. ``host`` is
dropped for the same reason: it is what ``Request.url`` is built from, so forwarding it
would let a caller choose the URL every logging callback records. Upstream
MCP credentials and the deployment's proxy key header, including a custom
``litellm_key_header_name``, are dropped so they cannot reach a callback or a guardrail
through the derived metadata even when a caller omits ``general_settings``.
"""
from fastapi import Request
custom_key_header: Final = _custom_litellm_key_header_name()
excluded: Final = (
_SYNTHETIC_REQUEST_EXCLUDED_HEADERS
| _upstream_credential_headers(raw_headers.keys() if raw_headers else ())
| (frozenset({custom_key_header.lower()}) if custom_key_header else frozenset())
)
forwarded: Final = tuple(
(
name.lower().encode("latin-1", errors="replace"),
value.encode("utf-8", errors="replace"),
)
for name, value in (raw_headers.items() if raw_headers else ())
if name.lower() not in excluded
)
xff: Final = ((b"x-forwarded-for", client_ip.encode("utf-8")),) if client_ip else ()
return Request(
scope={
"type": "http",
"method": "POST",
"path": path,
"scheme": "http",
"server": _SYNTHETIC_REQUEST_SERVER,
"query_string": b"",
"root_path": "",
"headers": ((b"content-type", b"application/json"), *forwarded, *xff),
**({"client": (client_ip, 0)} if client_ip else {}),
}
)
def logging_safe_mcp_headers(raw_headers: Mapping[str, str] | None) -> Mapping[str, str]:
"""The MCP request's client headers, sanitized the way the chat completions path
sanitizes them before they reach a logging callback or a guardrail: proxy key
headers stripped, including the custom key header name the deployment configured,
upstream MCP credentials dropped, and credential-bearing values masked.
Client-controlled behaviour flags (``litellm-disable-message-redaction``) are dropped
too: these headers are read back out of the metadata to change proxy behaviour, so
leaving one in place would let any MCP client turn off the redaction an admin
configured. This path carries no key or team object to authorize an opt-out with, so
it always strips them. ``host`` goes too, so that a caller cannot name the deployment in
the guardrail payload and the spend row the way it could once name the request URL."""
from starlette.datastructures import Headers
from litellm.proxy.litellm_pre_call_utils import (
UNTRUSTED_REQUEST_HEADER_CONTROL_FIELDS,
clean_headers,
redact_credential_headers,
)
excluded: Final = (
_upstream_credential_headers(raw_headers.keys() if raw_headers else ())
| UNTRUSTED_REQUEST_HEADER_CONTROL_FIELDS
| frozenset({"host"})
)
cleaned: Final = clean_headers(
Headers(raw_headers),
litellm_key_header_name=_custom_litellm_key_header_name(),
)
return redact_credential_headers({name: value for name, value in cleaned.items() if name.lower() not in excluded})

19
litellm/proxy/_types.py
View file

@ -1283,6 +1283,9 @@ class MCPApprovalStatus(str, enum.Enum):
pending_review = "pending_review"
active = "active"
rejected = "rejected"
# Short-lived row backing the admin OAuth "Authorize & Fetch Token" flow. Never served: the
# registry loader and every listing exclude it, so it is reachable only by its own server_id.
draft = "draft"
from litellm.models.mcp_server import ( # noqa: E402
@ -2511,6 +2514,22 @@ class ConfigGeneralSettings(LiteLLMPydanticObjectBase):
None,
description="If True and LiteLLM_SpendLogs has been converted to a range-partitioned table (db_scripts/partition_spend_logs.sql), retention cleanup drops expired partitions instead of deleting rows, and pre-creates upcoming partitions. Default is False.",
)
maximum_spend_logs_cleanup_batch_size: int | None = Field(
None,
description="Rows deleted per DELETE statement by the spend log cleanup job. Defaults to 1000.",
)
maximum_spend_logs_cleanup_max_batches: int | None = Field(
None,
description="Maximum DELETE statements the spend log cleanup job issues per table per run. Defaults to 500.",
)
maximum_spend_logs_cleanup_run_budget: str | None = Field(
None,
description="Wall-clock budget for one spend log cleanup run (e.g. '5m'), shared across every table it prunes. A run that hits the budget stops and the next run resumes from where it left off. Defaults to '5m'.",
)
maximum_spend_logs_cleanup_batch_timeout: str | None = Field(
None,
description="Postgres statement_timeout and lock_timeout applied to each spend log cleanup delete batch (e.g. '30s'), so cleanup cannot hold row locks or a connection indefinitely. Defaults to '30s'.",
)
mcp_internal_ip_ranges: list[str] | None = Field(
None,
description="Custom CIDR ranges that define internal/private networks for MCP access control. When set, only these ranges are treated as internal. Defaults to RFC 1918 private ranges (10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16, 127.0.0.0/8).",

114
litellm/proxy/auth/auth_checks.py
View file

@ -14,6 +14,7 @@ import math
import re
import time
from collections.abc import Iterator, Mapping, Sequence
from types import MappingProxyType
from typing import TYPE_CHECKING, Any, Final, Literal, Optional, Protocol, cast
from fastapi import HTTPException, Request, status
@ -65,6 +66,7 @@ from litellm.proxy.auth.budget_throttle import (
should_throttle_budget_exceeded,
)
from litellm.proxy.auth.route_checks import RouteChecks
from litellm.proxy.common_utils.auth_cache_invalidation_pubsub import publish_auth_cache_invalidation
from litellm.proxy.common_utils.cache_pydantic_utils import CacheCodec
from litellm.proxy.common_utils.http_parsing_utils import (
_safe_get_request_headers,
@ -375,6 +377,16 @@ def _is_model_cost_zero(model: str | list[str] | None, llm_router: Router | None
zero_cost_cache[model_name] = False
return False
if _has_ptu_flat_cost(model_name, llm_router):
verbose_proxy_logger.debug(
"Model %s prices reserved PTU capacity as a flat cost, so its zero per-token "
"rate is not a free model (enforce budget)",
safe_name,
)
if zero_cost_cache is not None:
zero_cost_cache[model_name] = False
return False
verbose_proxy_logger.debug(
"Model %s has zero cost explicitly configured (input: %s, output: %s)",
safe_name,
@ -393,6 +405,24 @@ def _is_model_cost_zero(model: str | list[str] | None, llm_router: Router | None
return True
_NO_MODEL_INFO: Final[Mapping[str, object]] = MappingProxyType({})
def _has_ptu_flat_cost(model: str, llm_router: "Router") -> bool:
"""Whether any deployment in the model group bills reserved PTU capacity as a flat cost.
Such a deployment carries an explicit zero per-token price so the flat cost is not charged
twice, which otherwise reads here as a free model and waives every budget check for it.
"""
for deployment in llm_router.model_list:
if deployment.get("model_name") != model:
continue
model_info = deployment.get("model_info") or _NO_MODEL_INFO
if model_info.get("ptu_count") is not None and model_info.get("cost_per_ptu_per_hour") is not None:
return True
return False
def _is_cost_explicitly_configured(model: str, llm_router: "Router") -> bool:
"""
Check if any deployment in the model group has cost fields explicitly
@ -2016,6 +2046,44 @@ async def _cache_team_object(
)
async def delete_cache_team_object(
team_id: str,
team_alias: str | None,
user_api_key_cache: UserApiKeyCache,
proxy_logging_obj: ProxyLogging | None,
) -> None:
"""
Evict both keys `_cache_team_object` writes.
`get_team_object` reads the id key and the JWT `team_alias_jwt_field` path reads the alias key,
so leaving either behind keeps a deleted team resolvable for auth until its TTL expires.
Mirrors `delete_cached_project_object`: evicting locally only reaches the worker handling the
delete, so every key is also broadcast to drop the other workers' in-memory copies.
Eviction is best-effort, matching `_cache_team_object`. `delete_team` calls this after the team
rows are already gone, so letting an unreachable cache backend raise here would fail a request
whose delete has committed.
"""
keys: Final = (f"team_id:{team_id}", *((f"team_alias:{team_alias}",) if team_alias else ()))
for key in keys:
try:
user_api_key_cache.delete_cache(key=key)
## UPDATE REDIS CACHE ##
if proxy_logging_obj is not None:
await proxy_logging_obj.internal_usage_cache.dual_cache.async_delete_cache(key=key)
except Exception as e: # noqa: BLE001 # best-effort invalidation: any cache backend error must not abort the delete
verbose_proxy_logger.warning(
"Failed to invalidate cached team entry %s on delete; "
"a deleted team may be served until its TTL expires: %s",
key,
e,
)
await publish_auth_cache_invalidation(cache_key=key)
async def _cache_key_object(
hashed_token: str,
user_api_key_obj: UserAPIKeyAuth,
@ -2051,6 +2119,44 @@ async def _delete_cache_key_object(
await proxy_logging_obj.internal_usage_cache.dual_cache.async_delete_cache(key=key)
async def delete_cache_key_objects(
hashed_tokens: Sequence[str],
user_api_key_cache: UserApiKeyCache,
proxy_logging_obj: ProxyLogging | None,
) -> None:
"""
Evict a batch of key objects, for callers that delete keys in bulk rather than through
`/key/delete`. Auth resolves a cached key object without re-reading its team, so a key left
cached after its row is gone keeps buying access until its TTL expires.
Evicting locally only reaches this worker, so each token is also broadcast: a deleted key left
in a peer worker's in-memory cache still authenticates there until its TTL expires.
Best-effort per key: the rows are already deleted by the time this runs, so an unreachable
cache backend must not abort the caller partway through its own cascade.
"""
results: Final = await asyncio.gather(
*(
_delete_cache_key_object(
hashed_token=hashed_token,
user_api_key_cache=user_api_key_cache,
proxy_logging_obj=proxy_logging_obj,
)
for hashed_token in hashed_tokens
),
return_exceptions=True,
)
for hashed_token, result in zip(hashed_tokens, results):
if isinstance(result, BaseException):
verbose_proxy_logger.warning(
"Failed to evict cached key entry for %s; a deleted key may authenticate until its TTL expires: %s",
hashed_token,
result,
)
await publish_auth_cache_invalidation(cache_key=hashed_token)
@log_db_metrics
async def _get_team_db_check(
team_id: str, prisma_client: PrismaClient, team_id_upsert: bool | None = None
@ -2556,6 +2662,8 @@ class ExperimentalUIJWTToken:
user_info: LiteLLM_UserTable,
team_id: str | None = None,
team_alias: str | None = None,
team_models: Sequence[str] | None = None,
team_model_aliases: Mapping[str, str] | None = None,
max_budget: float | None = None,
) -> str:
"""
@ -2568,6 +2676,8 @@ class ExperimentalUIJWTToken:
user_info: User information from the database
team_id: Team ID for the user (optional, uses user's team if available)
team_alias: Team alias for the selected team, if available
team_models: Model allowlist granted by the selected team
team_model_aliases: Team model aliases for the selected team
Returns:
Encrypted JWT token string
@ -2606,7 +2716,9 @@ class ExperimentalUIJWTToken:
user_id=user_info.user_id,
team_id=_team_id,
team_alias=team_alias,
models=user_info.models,
team_models=list(team_models) if team_models is not None else [],
team_model_aliases=dict(team_model_aliases) if team_model_aliases is not None else None,
models=[] if _team_id is not None else user_info.models,
max_parallel_requests=None,
user_role=LitellmUserRoles(user_info.user_role),
is_session_token=True,

5
litellm/proxy/batches_endpoints/endpoints.py
View file

@ -715,7 +715,7 @@ async def list_batches(
operation_context="batch listing",
)
data.update(credentials)
prepare_data_with_credentials(data=data, credentials=credentials)
response = await litellm.alist_batches(
custom_llm_provider=credentials["custom_llm_provider"],
@ -948,9 +948,10 @@ async def cancel_batch(
# SCENARIO 3: Fallback to custom_llm_provider (uses env variables)
else:
body_custom_llm_provider = data.pop("custom_llm_provider", None)
custom_llm_provider: Final = (
provider
or data.pop("custom_llm_provider", None)
or body_custom_llm_provider
or get_custom_llm_provider_from_request_headers(request=request)
or get_custom_llm_provider_from_request_query(request=request)
or "openai"

11
litellm/proxy/client/cli/README.md
View file

@ -36,6 +36,17 @@ The base URL is resolved in this order of precedence:
3. `base_url` from `~/.litellm/config.json`
4. `http://localhost:4000`
### Hiding commands from the listings
Deployments that hand `lite` to end users often want to advertise only part of it. Store the commands to keep out of the listings, comma separated:
```bash
lite config set hidden_commands codex,opencode
lite config unset hidden_commands # list everything again
```
Hidden commands drop out of both `lite --help` and the interactive shell's "Available commands" block, and stay runnable so existing scripts keep working
## Global Options
- `--version`, `-v`: Print the LiteLLM Proxy client and server version and exit.

103
litellm/proxy/client/cli/commands/agents.py
View file

@ -1,5 +1,6 @@
import os
import shutil
import subprocess
import sys
from collections.abc import Callable, Mapping, Sequence
from typing import Final
@ -142,8 +143,95 @@ def verify_proxy_key(
)
def _exec(path: str, args: Sequence[str], env: Mapping[str, str]) -> None:
os.execvpe(path, list(args), dict(env))
_WINDOWS_SHIM_SUFFIXES: Final[frozenset[str]] = frozenset({".cmd", ".bat"})
_CMD_PERCENT_GUARD: Final = "%%cd:~,%"
_CMD_LINE_BREAKS: Final = ("\r", "\n")
def _double_trailing_backslashes(segment: str) -> str:
bare: Final = segment.rstrip("\\")
return bare + "\\" * 2 * (len(segment) - len(bare))
def _quote_for_cmd(token: str) -> str:
"""Quote one token so both parsers that read it see the original text.
Follows the algorithm the Rust standard library settled on for batch files
after CVE-2024-24576. Two parsers see this token: cmd.exe, which ends a
quoted string on a lone `"` and so wants an embedded one doubled, and the
shim's own interpreter, which re-splits `%*` under C runtime rules where a
backslash escapes the quote that follows it, so every backslash run standing
before a quote is doubled. Quoting cannot stop cmd expanding `%VAR%`, so each
`%` is prefixed with `%%cd:~,`: the zero-length substring of the always
defined `cd` expands to nothing and leaves no `%` pair for cmd to match.
"""
escaped: Final = '""'.join(_double_trailing_backslashes(part) for part in token.split('"'))
return '"' + escaped.replace("%", _CMD_PERCENT_GUARD) + '"'
def _windows_command(path: str, args: Sequence[str]) -> str | tuple[str, ...]:
"""Build what CreateProcess runs, routing batch shims through cmd.exe.
npm installs Claude Code as `claude.cmd`, which PATHEXT lets shutil.which
resolve but CreateProcess refuses to run (WinError 193), so a shim has to go
through the command processor. cmd.exe does not follow the C runtime quoting
that subprocess would apply to an argument list, and it would split on `&` or
`|` in a forwarded argument, so the shim case is emitted as one verbatim
command line with every token quoted. Every switch is load-bearing: `/s`
makes cmd strip only the outer pair, leaving each token quoted and its
metacharacters inert, `/e:on` keeps the command extensions that the percent
guard is built out of, `/v:off` keeps `!` from expanding, and `/d` keeps a
machine's AutoRun commands out of the launch. argv[0] carries the
caller-facing name on POSIX; Windows needs the resolved path there.
Raises AgentRunError for an argument holding a line break, which cmd would
read as the end of the command line and silently drop the rest of.
"""
rest: Final = tuple(args[1:])
if os.path.splitext(path)[1].lower() not in _WINDOWS_SHIM_SUFFIXES:
return (path, *rest)
if any(brk in token for token in rest for brk in _CMD_LINE_BREAKS):
raise AgentRunError(
f"Cannot pass an argument containing a line break to `{os.path.basename(path)}` on "
"Windows: cmd.exe ends the command line there, so the agent would silently lose it."
)
inner: Final = " ".join(_quote_for_cmd(token) for token in (path, *rest))
return f'cmd.exe /d /e:on /v:off /s /c "{inner}"'
def _spawn_and_wait(command: str | Sequence[str], env: Mapping[str, str]) -> int:
return subprocess.run(command, env=dict(env), check=False).returncode
def _replace_process(
path: str,
args: Sequence[str],
env: Mapping[str, str],
*,
execvpe: Callable[..., None] = os.execvpe,
) -> None:
execvpe(path, list(args), dict(env))
def _hand_off(
path: str,
args: Sequence[str],
env: Mapping[str, str],
*,
platform: str = sys.platform,
replace: Callable[[str, Sequence[str], Mapping[str, str]], None] = _replace_process,
spawn: Callable[[str | Sequence[str], Mapping[str, str]], int] = _spawn_and_wait,
) -> None:
"""Replace this process with the agent; on Windows, run it as a child instead.
os.exec* has no process-replacement semantics on Windows: the C runtime
spawns a detached child and terminates the parent, so the shell reclaims the
console and the agent's TUI never gets one. Windows therefore waits on the
child and exits with its status.
"""
if platform.startswith("win"):
raise SystemExit(spawn(_windows_command(path, args), env))
replace(path, list(args), dict(env))
def _restore_controlling_terminal() -> None:
@ -175,13 +263,14 @@ def run_agent(
base_env: Mapping[str, str] | None = None,
which: Callable[[str], str | None] = shutil.which,
verify: Callable[[str, str], None] = verify_proxy_key,
launcher: Callable[[str, Sequence[str], Mapping[str, str]], None] = _exec,
launcher: Callable[[str, Sequence[str], Mapping[str, str]], None] = _hand_off,
reattach_terminal: Callable[[], None] | None = None,
) -> None:
"""Validate, wire the environment, and hand off to the agent.
On success this replaces the current process and never returns. Raises
AgentRunError for missing binaries, an unreachable proxy, or a rejected key.
On success this never returns: POSIX replaces the current process, Windows
waits on the agent and exits with its status. Raises AgentRunError for
missing binaries, an unreachable proxy, or a rejected key.
reattach_terminal, when given, runs just before handoff to restore stdin.
"""
if not command:
@ -277,9 +366,9 @@ def _make_agent_command(binary: str, display_name: str) -> click.Command:
return _command
def agent_commands() -> list[click.Command]:
def agent_commands() -> tuple[click.Command, ...]:
"""Build one top-level command per known agent, e.g. `lite claude`."""
return [_make_agent_command(binary, name) for binary, (name, _profiles) in _KNOWN_AGENTS.items()]
return tuple(_make_agent_command(binary, name) for binary, (name, _profiles) in _KNOWN_AGENTS.items())
__all__ = [

59
litellm/proxy/client/cli/commands/config.py
View file

@ -1,8 +1,9 @@
import json
import os
import sys
from collections.abc import Mapping
from collections.abc import Callable, Mapping
from pathlib import Path
from types import MappingProxyType
from typing import Final
from urllib.parse import urlparse
@ -11,7 +12,7 @@ from pydantic import TypeAdapter
from .private_json import write_private_json
ALLOWED_CONFIG_KEYS: Final[tuple[str, ...]] = ("base_url",)
HIDDEN_COMMANDS_KEY: Final = "hidden_commands"
_config_adapter: Final[TypeAdapter[Mapping[str, str]]] = TypeAdapter(Mapping[str, str])
@ -49,6 +50,48 @@ def get_config_value(key: str) -> str | None:
return load_config().get(key)
def parse_hidden_commands(raw: str | None) -> frozenset[str]:
"""Split a stored `hidden_commands` value, e.g. "codex, opencode"."""
return frozenset(name.strip() for name in (raw or "").split(",") if name.strip())
def hidden_command_names() -> frozenset[str]:
"""Top-level commands the operator chose to keep out of `lite`'s listings."""
return parse_hidden_commands(get_config_value(HIDDEN_COMMANDS_KEY))
def _normalize_base_url(value: str) -> str:
parsed: Final = urlparse(value)
if parsed.scheme not in ("http", "https") or not parsed.netloc:
raise click.UsageError("base_url must be a full http:// or https:// URL including a host")
if "?" in value or "#" in value:
raise click.UsageError("base_url must not include a query string or fragment")
return value.rstrip("/")
def _normalize_hidden_commands(value: str) -> str:
names: Final = parse_hidden_commands(value)
if not names:
raise click.UsageError(
f"{HIDDEN_COMMANDS_KEY} must be a comma-separated list of command names, e.g. "
f"`lite config set {HIDDEN_COMMANDS_KEY} codex,opencode`. To list everything again, "
f"run `lite config unset {HIDDEN_COMMANDS_KEY}`"
)
if any(" " in name for name in names):
raise click.UsageError(f"{HIDDEN_COMMANDS_KEY} entries must be single command names, without spaces")
return ",".join(sorted(names))
_NORMALIZERS: Final[Mapping[str, Callable[[str], str]]] = MappingProxyType(
{
"base_url": _normalize_base_url,
HIDDEN_COMMANDS_KEY: _normalize_hidden_commands,
}
)
ALLOWED_CONFIG_KEYS: Final[tuple[str, ...]] = tuple(_NORMALIZERS)
@click.group(name="config")
def config_commands() -> None:
"""Manage persistent CLI configuration (~/.litellm/config.json)"""
@ -59,17 +102,11 @@ def config_commands() -> None:
@click.argument("value")
def set_config(key: str, value: str) -> None:
"""Set a config KEY to VALUE (e.g. `lite config set base_url https://your-proxy.example.com`)"""
if key not in ALLOWED_CONFIG_KEYS:
normalizer: Final = _NORMALIZERS.get(key)
if normalizer is None:
raise click.UsageError(f"Unknown config key '{key}'. Allowed keys: {', '.join(ALLOWED_CONFIG_KEYS)}")
if key == "base_url":
parsed: Final = urlparse(value)
if parsed.scheme not in ("http", "https") or not parsed.netloc:
raise click.UsageError("base_url must be a full http:// or https:// URL including a host")
if "?" in value or "#" in value:
raise click.UsageError("base_url must not include a query string or fragment")
normalized_value: Final = value.rstrip("/")
normalized_value: Final = normalizer(value)
save_config({**load_config(), key: normalized_value})
click.echo(f"Set {key} = {normalized_value} in {get_config_file_path()}")

8
litellm/proxy/client/cli/interface.py
View file

@ -74,8 +74,9 @@ def styled_prompt():
def show_commands():
"""Display available commands."""
"""Display available commands, minus any the operator chose to hide."""
from .commands.agents import agent_commands
from .commands.config import hidden_command_names
commands = [
("login", "Authenticate with the LiteLLM proxy server"),
@ -96,9 +97,12 @@ def show_commands():
("quit", "Exit the interactive session"),
]
hidden: Final = hidden_command_names()
click.echo("Available commands:")
for cmd, description in commands:
click.echo(f" {cmd:<20} {description}")
if cmd not in hidden:
click.echo(f" {cmd:<20} {description}")
click.echo()

18
litellm/proxy/client/cli/main.py
View file

@ -12,7 +12,7 @@ from .commands.agents import agent_commands
from .commands.auth import auth_group, get_stored_api_key, login, logout, whoami
from .commands.autoroute.commands import autoroute_group
from .commands.chat import chat
from .commands.config import config_commands, get_config_value
from .commands.config import config_commands, get_config_value, hidden_command_names
from .commands.credentials import credentials
from .commands.encryption import encryption
from .commands.http import http
@ -43,7 +43,21 @@ def print_version(base_url: str, api_key: str | None):
click.echo(f"Could not retrieve server version: {e}")
@click.group(invoke_without_command=True)
class HideConfiguredCommandsGroup(click.Group):
"""Group that omits operator-hidden commands from listings, still running them.
Deployments hand `lite` to users who should only see a curated subset of
commands (`lite config set hidden_commands codex,opencode`). Filtering the
listing rather than dropping the commands keeps anyone's existing scripts
working.
"""
def list_commands(self, ctx: click.Context) -> list[str]:
hidden: Final = hidden_command_names()
return [name for name in super().list_commands(ctx) if name not in hidden]
@click.group(cls=HideConfiguredCommandsGroup, invoke_without_command=True)
@click.option(
"--version",
"-v",

83
litellm/proxy/common_utils/callback_utils.py
View file

@ -1,7 +1,10 @@
import copy
import os
from collections.abc import Callable, Iterable
from typing import TYPE_CHECKING, Any, Final, Optional
from dataclasses import dataclass
from typing import TYPE_CHECKING, Any, Final, Literal, NoReturn, Optional, TypeAlias
from typing_extensions import assert_never
import litellm
from litellm import get_secret
@ -50,6 +53,66 @@ if TYPE_CHECKING:
from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLogging
@dataclass(frozen=True, slots=True)
class _CallbackResolvedToClass:
entry: str
loaded: type
tag: Literal["resolved_to_class"] = "resolved_to_class"
@dataclass(frozen=True, slots=True)
class _CallbackNotDispatchable:
entry: str
loaded: object
tag: Literal["not_dispatchable"] = "not_dispatchable"
_CallbackLoadError: TypeAlias = _CallbackResolvedToClass | _CallbackNotDispatchable
def _classify_loaded_callback(entry: str, loaded: object) -> CustomLogger | Callable[..., object] | _CallbackLoadError:
"""
Decide whether what a ``litellm_settings.callbacks`` dotted path resolved to can be dispatched.
A dotted path only ever runs as a ``CustomLogger`` instance or as a callback function. Anything
else (most commonly a class instead of an instance) used to load without complaint and then be
skipped on every request, with no log line and no error.
"""
if isinstance(loaded, CustomLogger) or (callable(loaded) and not isinstance(loaded, type)):
return loaded
if isinstance(loaded, type):
return _CallbackResolvedToClass(entry=entry, loaded=loaded)
return _CallbackNotDispatchable(entry=entry, loaded=loaded)
def _raise_callback_load_error(error: _CallbackLoadError) -> NoReturn:
"""The one edge that raises: map a load error onto config load's failure contract."""
match error:
case _CallbackResolvedToClass():
module_path: Final = error.entry.rsplit(".", 1)[0] if "." in error.entry else error.entry
raise ValueError(
f"litellm_settings.callbacks entry '{error.entry}' resolved to the class "
f"{error.loaded.__module__}.{error.loaded.__qualname__}, which is neither a "
"CustomLogger instance nor a callable, so the proxy would never run it."
f" Point it at an instance instead, e.g. add `proxy_handler_instance = {error.loaded.__name__}()` to "
f'{module_path} and set `callbacks: ["{module_path}.proxy_handler_instance"]`.'
)
case _CallbackNotDispatchable():
raise ValueError(
f"litellm_settings.callbacks entry '{error.entry}' resolved to "
f"{type(error.loaded).__name__} {error.loaded!r}, which is neither a "
"CustomLogger instance nor a callable, so the proxy would never run it."
)
assert_never(error)
def _loaded_callback_or_raise(entry: str, loaded: object) -> CustomLogger | Callable[..., object]:
resolved: Final = _classify_loaded_callback(entry=entry, loaded=loaded)
if isinstance(resolved, _CallbackResolvedToClass | _CallbackNotDispatchable):
_raise_callback_load_error(resolved)
return resolved
def initialize_callbacks_on_proxy(
value: Any,
premium_user: bool,
@ -305,9 +368,12 @@ def initialize_callbacks_on_proxy(
"%s attempting to import custom calback=%s %s", blue_color_code, callback, reset_color_code
)
imported_list.append(
get_instance_fn(
value=callback,
config_file_path=config_file_path,
_loaded_callback_or_raise(
entry=callback,
loaded=get_instance_fn(
value=callback,
config_file_path=config_file_path,
),
)
)
if isinstance(litellm.callbacks, list):
@ -321,9 +387,12 @@ def initialize_callbacks_on_proxy(
PrometheusLogger._mount_metrics_endpoint()
else:
litellm.callbacks = [
get_instance_fn(
value=value,
config_file_path=config_file_path,
_loaded_callback_or_raise(
entry=value,
loaded=get_instance_fn(
value=value,
config_file_path=config_file_path,
),
)
]
verbose_proxy_logger.debug("%s Initialized Callbacks - %s %s", blue_color_code, litellm.callbacks, reset_color_code)

11
litellm/proxy/common_utils/sse_keepalive.py
View file

@ -21,6 +21,17 @@ def _coerce_interval(ping_interval_seconds: float | str | None) -> float | None:
return interval
def keepalive_ping_has_fired(elapsed_seconds: float, ping_interval_seconds: float | str | None) -> bool:
"""Whether a keepalive ping has already gone out, which flushes the response headers.
A caller that discovers a failure after that point cannot raise its way to the client, since
the status line is already on the wire. With pings disabled nothing flushes early, so a raise
still carries its real status.
"""
interval: Final = _coerce_interval(ping_interval_seconds)
return interval is not None and elapsed_seconds >= interval
def wrap_sse_stream_with_keepalive_pings(
stream: AsyncGenerator[str, None],
ping_interval_seconds: float | str | None,

12
litellm/proxy/db/autorouter_session_rollup.py
View file

@ -24,6 +24,7 @@ from itertools import groupby
from typing import TYPE_CHECKING, Final, NamedTuple
from litellm._logging import verbose_proxy_logger
from litellm.constants import INTERNAL_CALL_ORIGIN_METADATA_KEY
from litellm.proxy._types import DB_RETRY_SAFE_ERROR_TYPES
if TYPE_CHECKING:
@ -180,12 +181,17 @@ def build_autorouter_turn_transaction(
The routing_decision record is what says a request was auto-routed at all, so a
request without one (including the auto-router's own classifier sub-calls) never
reaches the rollup. Failed requests served nothing and are excluded. Cache facts
are derived from the payload's own usage record through the savings owner, never
handed in beside it.
reaches the rollup. Internal sub-calls that DO carry one (a shadow eval's duplicate
of a request through the router) are excluded by their internal_call_origin stamp:
they are not traffic a user sent, so counting them would manufacture sessions and
savings in the adoption metrics. Failed requests served nothing and are excluded.
Cache facts are derived from the payload's own usage record through the savings
owner, never handed in beside it.
"""
if payload.get("status") != "success":
return None
if metadata.get(INTERNAL_CALL_ORIGIN_METADATA_KEY):
return None
routing_decision: Final = metadata.get("routing_decision")
if not isinstance(routing_decision, Mapping) or not routing_decision:
return None

75
litellm/proxy/db/create_views.py
View file

@ -1,15 +1,50 @@
from typing import Any, Final
from typing import Any, Final, Protocol
from litellm import verbose_logger
_db = Any
class SupportsExecuteRaw(Protocol):
"""The one database operation create_view_tolerating_race needs.
Narrower than the `_db = Any` the rest of this module still uses, so the
helper's contract is checkable at its call sites without retyping every
function here.
"""
async def execute_raw(self, query: str, *args: object) -> int: ...
# Markers that indicate a view/relation does not yet exist in the database.
# Keeping these in one place avoids repeating the check across all view blocks
# and prevents overly broad matches (e.g. bare 'undefined' would also match
# 'undefined function' or 'column undefined_col referenced in query').
_VIEW_NOT_FOUND_MARKERS: Final = ("does not exist", "no such table", "undefined table")
# Markers for the inverse condition: another replica created the view between
# our existence probe and our CREATE.
_VIEW_ALREADY_EXISTS_MARKERS: Final = ("already exists", "duplicate object", "duplicate table")
async def create_view_tolerating_race(db: SupportsExecuteRaw, view_name: str, ddl: str) -> None:
"""
Create a view, treating "a concurrent creator won" as success.
Every replica booting against the same fresh database observes the view as
absent and issues the CREATE; Postgres fails all but one with a
duplicate-object error. The desired end state is still reached, so losing
that race is success. Without this, the loser's exception propagates out of
a detached startup task and the remaining views are never created.
"""
try:
await db.execute_raw(ddl)
verbose_logger.debug("%s Created!", view_name)
except Exception as e:
if not any(marker in str(e).lower() for marker in _VIEW_ALREADY_EXISTS_MARKERS):
raise
verbose_logger.debug("%s already created by a concurrent replica", view_name)
async def create_missing_views(db: _db):
"""
@ -34,7 +69,10 @@ async def create_missing_views(db: _db):
if not any(marker in error_msg for marker in _VIEW_NOT_FOUND_MARKERS):
raise
# If an error occurs, the view does not exist, so create it
await db.execute_raw("""
await create_view_tolerating_race(
db,
"LiteLLM_VerificationTokenView",
"""
CREATE VIEW "LiteLLM_VerificationTokenView" AS
SELECT
v.*,
@ -46,9 +84,8 @@ async def create_missing_views(db: _db):
FROM "LiteLLM_VerificationToken" v
LEFT JOIN "LiteLLM_TeamTable" t ON v.team_id = t.team_id
LEFT JOIN "LiteLLM_ProjectTable" p ON v.project_id = p.project_id;
""")
verbose_logger.debug("LiteLLM_VerificationTokenView Created!")
""",
)
try:
await db.query_raw("""SELECT 1 FROM "MonthlyGlobalSpend" LIMIT 1""")
@ -69,9 +106,7 @@ async def create_missing_views(db: _db):
GROUP BY
DATE("startTime");
"""
await db.execute_raw(query=sql_query)
verbose_logger.debug("MonthlyGlobalSpend Created!")
await create_view_tolerating_race(db, "MonthlyGlobalSpend", sql_query)
try:
await db.query_raw("""SELECT 1 FROM "Last30dKeysBySpend" LIMIT 1""")
@ -100,9 +135,7 @@ async def create_missing_views(db: _db):
ORDER BY
total_spend DESC;
"""
await db.execute_raw(query=sql_query)
verbose_logger.debug("Last30dKeysBySpend Created!")
await create_view_tolerating_race(db, "Last30dKeysBySpend", sql_query)
try:
await db.query_raw("""SELECT 1 FROM "Last30dModelsBySpend" LIMIT 1""")
@ -126,9 +159,7 @@ async def create_missing_views(db: _db):
ORDER BY
total_spend DESC;
"""
await db.execute_raw(query=sql_query)
verbose_logger.debug("Last30dModelsBySpend Created!")
await create_view_tolerating_race(db, "Last30dModelsBySpend", sql_query)
try:
await db.query_raw("""SELECT 1 FROM "MonthlyGlobalSpendPerKey" LIMIT 1""")
verbose_logger.debug("MonthlyGlobalSpendPerKey Exists!")
@ -150,9 +181,7 @@ async def create_missing_views(db: _db):
DATE("startTime"),
api_key;
"""
await db.execute_raw(query=sql_query)
verbose_logger.debug("MonthlyGlobalSpendPerKey Created!")
await create_view_tolerating_race(db, "MonthlyGlobalSpendPerKey", sql_query)
try:
await db.query_raw("""SELECT 1 FROM "MonthlyGlobalSpendPerUserPerKey" LIMIT 1""")
verbose_logger.debug("MonthlyGlobalSpendPerUserPerKey Exists!")
@ -176,9 +205,7 @@ async def create_missing_views(db: _db):
"user",
api_key;
"""
await db.execute_raw(query=sql_query)
verbose_logger.debug("MonthlyGlobalSpendPerUserPerKey Created!")
await create_view_tolerating_race(db, "MonthlyGlobalSpendPerUserPerKey", sql_query)
try:
await db.query_raw("""SELECT 1 FROM "DailyTagSpend" LIMIT 1""")
@ -197,9 +224,7 @@ async def create_missing_views(db: _db):
FROM "LiteLLM_SpendLogs" s
GROUP BY individual_request_tag, DATE(s."startTime");
"""
await db.execute_raw(query=sql_query)
verbose_logger.debug("DailyTagSpend Created!")
await create_view_tolerating_race(db, "DailyTagSpend", sql_query)
try:
await db.query_raw("""SELECT 1 FROM "Last30dTopEndUsersSpend" LIMIT 1""")
@ -218,9 +243,7 @@ async def create_missing_views(db: _db):
ORDER BY total_spend DESC
LIMIT 100;
"""
await db.execute_raw(query=sql_query)
verbose_logger.debug("Last30dTopEndUsersSpend Created!")
await create_view_tolerating_race(db, "Last30dTopEndUsersSpend", sql_query)
async def should_create_missing_views(db: _db) -> bool:

91
litellm/proxy/db/db_spend_update_writer.py
View file

@ -21,6 +21,7 @@ from litellm.caching import RedisCache
from litellm.constants import (
DB_DAILY_TAG_SPEND_UPDATE_JOB_NAME,
DB_SPEND_UPDATE_JOB_NAME,
INTERNAL_CALL_ORIGIN_METADATA_KEY,
)
from litellm.litellm_core_utils.safe_json_loads import safe_json_loads
from litellm.proxy._types import (
@ -860,6 +861,8 @@ class DBSpendUpdateWriter:
):
verbose_proxy_logger.debug("acquired lock for spend updates")
uncommitted: dict[str, Any] = {} # mutable-ok: tracks popped categories still needing commit
try:
(
db_spend_update_transactions,
@ -870,6 +873,15 @@ class DBSpendUpdateWriter:
daily_agent_spend_update_transactions,
) = await self.redis_update_buffer.get_all_transactions_from_redis_buffer_pipeline()
uncommitted = { # mutable-ok: drives which popped categories still need re-queuing
"db_spend_update_transactions": db_spend_update_transactions,
"daily_spend_update_transactions": daily_spend_update_transactions,
"daily_team_spend_update_transactions": daily_team_spend_update_transactions,
"daily_org_spend_update_transactions": daily_org_spend_update_transactions,
"daily_end_user_spend_update_transactions": daily_end_user_spend_update_transactions,
"daily_agent_spend_update_transactions": daily_agent_spend_update_transactions,
}
if db_spend_update_transactions is not None:
verbose_proxy_logger.info(
"Spend tracking - committing spend updates from Redis to DB: "
@ -889,6 +901,7 @@ class DBSpendUpdateWriter:
proxy_logging_obj=proxy_logging_obj,
db_spend_update_transactions=db_spend_update_transactions,
)
uncommitted.pop("db_spend_update_transactions", None)
if daily_spend_update_transactions is not None:
await DBSpendUpdateWriter.update_daily_user_spend(
@ -897,6 +910,8 @@ class DBSpendUpdateWriter:
proxy_logging_obj=proxy_logging_obj,
daily_spend_transactions=daily_spend_update_transactions,
)
uncommitted.pop("daily_spend_update_transactions", None)
if daily_team_spend_update_transactions is not None:
await DBSpendUpdateWriter.update_daily_team_spend(
n_retry_times=n_retry_times,
@ -904,6 +919,7 @@ class DBSpendUpdateWriter:
proxy_logging_obj=proxy_logging_obj,
daily_spend_transactions=daily_team_spend_update_transactions,
)
uncommitted.pop("daily_team_spend_update_transactions", None)
if daily_org_spend_update_transactions is not None:
await DBSpendUpdateWriter.update_daily_org_spend(
@ -912,6 +928,7 @@ class DBSpendUpdateWriter:
proxy_logging_obj=proxy_logging_obj,
daily_spend_transactions=daily_org_spend_update_transactions,
)
uncommitted.pop("daily_org_spend_update_transactions", None)
if daily_end_user_spend_update_transactions is not None:
await DBSpendUpdateWriter.update_daily_end_user_spend(
@ -920,6 +937,8 @@ class DBSpendUpdateWriter:
proxy_logging_obj=proxy_logging_obj,
daily_spend_transactions=daily_end_user_spend_update_transactions,
)
uncommitted.pop("daily_end_user_spend_update_transactions", None)
if daily_agent_spend_update_transactions is not None:
await DBSpendUpdateWriter.update_daily_agent_spend(
n_retry_times=n_retry_times,
@ -927,14 +946,20 @@ class DBSpendUpdateWriter:
proxy_logging_obj=proxy_logging_obj,
daily_spend_transactions=daily_agent_spend_update_transactions,
)
uncommitted.pop("daily_agent_spend_update_transactions", None)
except Exception as e:
spend_log_error(
"Spend tracking - failed to commit spend updates from Redis to DB. "
"Data already popped from Redis may be lost. Error: %s",
"Re-queuing uncommitted transactions to Redis for retry on next tick. Error: %s",
str(e),
exc=e,
)
finally:
to_restore = { # mutable-ok: transient kwargs payload consumed immediately below
name: txns for name, txns in uncommitted.items() if txns is not None
}
if to_restore:
await self.redis_update_buffer.restore_transactions_to_redis(**to_restore)
await self.pod_lock_manager.release_lock(
cronjob_id=DB_SPEND_UPDATE_JOB_NAME,
)
@ -1084,21 +1109,15 @@ class DBSpendUpdateWriter:
):
verbose_proxy_logger.debug("acquired lock for daily tag spend updates")
try:
daily_tag_spend_update_transactions: Final = (
await self.redis_update_buffer.get_all_daily_tag_spend_update_transactions_from_redis_buffer()
await self._drain_and_commit_daily_tag_spend_from_redis(
prisma_client=prisma_client,
n_retry_times=n_retry_times,
proxy_logging_obj=proxy_logging_obj,
)
if daily_tag_spend_update_transactions:
await DBSpendUpdateWriter.update_daily_tag_spend(
n_retry_times=n_retry_times,
prisma_client=prisma_client,
proxy_logging_obj=proxy_logging_obj,
daily_spend_transactions=daily_tag_spend_update_transactions,
)
except Exception as e:
spend_log_error(
"Spend tracking - failed to commit daily tag spend updates from Redis to DB. "
"Data already popped from Redis may be lost. Error: %s",
"Re-queuing to Redis for retry on next tick. Error: %s",
str(e),
exc=e,
)
@ -1107,6 +1126,37 @@ class DBSpendUpdateWriter:
cronjob_id=DB_DAILY_TAG_SPEND_UPDATE_JOB_NAME,
)
async def _drain_and_commit_daily_tag_spend_from_redis(
self,
prisma_client: PrismaClient,
n_retry_times: int,
proxy_logging_obj: ProxyLogging,
) -> None:
"""
Drain the Redis tag spend buffer and commit it, restoring the drained transactions if the commit fails.
The drain is destructive, so a failed commit must push the transactions back for the next tick
or their spend is lost permanently.
"""
daily_tag_spend_update_transactions: Final = (
await self.redis_update_buffer.get_all_daily_tag_spend_update_transactions_from_redis_buffer()
)
if not daily_tag_spend_update_transactions:
return
try:
await DBSpendUpdateWriter.update_daily_tag_spend(
n_retry_times=n_retry_times,
prisma_client=prisma_client,
proxy_logging_obj=proxy_logging_obj,
daily_spend_transactions=daily_tag_spend_update_transactions,
)
except Exception:
await self.redis_update_buffer.restore_transactions_to_redis(
daily_tag_spend_update_transactions=daily_tag_spend_update_transactions,
)
raise
async def _flush_tool_discovery_queue(
self,
prisma_client: PrismaClient,
@ -1606,9 +1656,6 @@ class DBSpendUpdateWriter:
)
except Exception as e:
if "transactions_to_process" in locals():
for key in transactions_to_process:
daily_spend_transactions.pop(key, None)
_raise_failed_update_spend_exception(e=e, start_time=start_time, proxy_logging_obj=proxy_logging_obj)
@staticmethod
@ -1794,6 +1841,7 @@ class DBSpendUpdateWriter:
if call_type:
endpoint = ROUTE_ENDPOINT_MAPPING.get(call_type, None)
is_internal_call: Final = bool(_metadata.get(INTERNAL_CALL_ORIGIN_METADATA_KEY))
cache_read_input_tokens: Final = extract_cache_read_tokens(usage_obj)
compression_saved_tokens: Final = extract_compression_saved_tokens(_metadata)
savings_spend: Final = compute_savings_spend(
@ -1818,15 +1866,20 @@ class DBSpendUpdateWriter:
prompt_tokens=payload["prompt_tokens"],
completion_tokens=payload["completion_tokens"],
spend=payload["spend"],
api_requests=1,
successful_requests=1 if request_status == "success" else 0,
failed_requests=1 if request_status != "success" else 0,
# Internal sub-calls (auto-router classifier, shadow eval's shadow and
# judge) bill real spend and tokens to the key, but they are not
# requests the caller made: counting them inflates request-volume
# readers, and an auto-router savings figure computed on a shadow
# duplicate credits savings for traffic no user sent.
api_requests=0 if is_internal_call else 1,
successful_requests=1 if not is_internal_call and request_status == "success" else 0,
failed_requests=1 if not is_internal_call and request_status != "success" else 0,
cache_read_input_tokens=cache_read_input_tokens,
cache_creation_input_tokens=extract_cache_creation_tokens(usage_obj),
compression_saved_tokens=compression_saved_tokens,
compression_savings_spend=savings_spend.compression,
prompt_caching_savings_spend=savings_spend.prompt_caching,
autorouter_savings_spend=savings_spend.autorouter,
autorouter_savings_spend=0.0 if is_internal_call else savings_spend.autorouter,
)
return daily_transaction
except Exception as e:

56
litellm/proxy/db/db_transaction_queue/redis_update_buffer.py
View file

@ -6,8 +6,11 @@ This is to prevent deadlocks and improve reliability
import asyncio
import json
from collections.abc import Mapping
from typing import TYPE_CHECKING, Any, Final, cast
from redis.exceptions import RedisError
from litellm._logging import verbose_proxy_logger
from litellm.caching import RedisCache
from litellm.constants import (
@ -372,6 +375,59 @@ class RedisUpdateBuffer:
if daily_txns:
await daily_queue.update_queue.put(daily_txns)
async def restore_transactions_to_redis(
self,
db_spend_update_transactions: DBSpendUpdateTransactions | None = None,
daily_spend_update_transactions: Mapping[str, BaseDailySpendTransaction] | None = None,
daily_team_spend_update_transactions: Mapping[str, BaseDailySpendTransaction] | None = None,
daily_org_spend_update_transactions: Mapping[str, BaseDailySpendTransaction] | None = None,
daily_end_user_spend_update_transactions: Mapping[str, BaseDailySpendTransaction] | None = None,
daily_agent_spend_update_transactions: Mapping[str, BaseDailySpendTransaction] | None = None,
daily_tag_spend_update_transactions: Mapping[str, BaseDailySpendTransaction] | None = None,
) -> None:
"""
Re-push transactions that were popped from Redis but not committed to the DB.
The leader drains the buffers with a destructive ``lpop`` before committing to
the database. When a commit fails after its retries are exhausted, the popped
transactions must be pushed back so a later scheduler tick can retry them;
otherwise the aggregated spend is lost permanently. The re-pushed payloads use
the same JSON encoding as the store path, so the next drain parses them normally.
"""
if self.redis_cache is None:
return
restore_configs: Final = (
(db_spend_update_transactions, REDIS_UPDATE_BUFFER_KEY),
(daily_spend_update_transactions, REDIS_DAILY_SPEND_UPDATE_BUFFER_KEY),
(daily_team_spend_update_transactions, REDIS_DAILY_TEAM_SPEND_UPDATE_BUFFER_KEY),
(daily_org_spend_update_transactions, REDIS_DAILY_ORG_SPEND_UPDATE_BUFFER_KEY),
(daily_end_user_spend_update_transactions, REDIS_DAILY_END_USER_SPEND_UPDATE_BUFFER_KEY),
(daily_agent_spend_update_transactions, REDIS_DAILY_AGENT_SPEND_UPDATE_BUFFER_KEY),
(daily_tag_spend_update_transactions, REDIS_DAILY_TAG_SPEND_UPDATE_BUFFER_KEY),
)
rpush_list: Final = tuple(
RedisPipelineRpushOperation(key=redis_key, values=(safe_dumps(transactions),))
for transactions, redis_key in restore_configs
if transactions
)
if len(rpush_list) == 0:
return
try:
await self.redis_cache.async_rpush_pipeline(rpush_list=rpush_list)
verbose_proxy_logger.info(
"Spend tracking - restored %d uncommitted transaction set(s) to Redis for retry on next tick.",
len(rpush_list),
)
except RedisError as e:
verbose_proxy_logger.error(
"Spend tracking - failed to restore uncommitted transactions to Redis. "
"These spend updates are lost. Error: %s",
str(e),
)
@staticmethod
def _number_of_transactions_to_store_in_redis(
db_spend_update_transactions: DBSpendUpdateTransactions,

453
litellm/proxy/db/db_transaction_queue/spend_log_cleanup.py
View file

@ -1,22 +1,60 @@
import asyncio
import time
from dataclasses import dataclass
from datetime import datetime, timedelta, timezone
from typing import Final
from typing import Final, Literal, TypeAlias
from pydantic import BaseModel, TypeAdapter
from litellm._logging import verbose_proxy_logger
from litellm.caching import RedisCache
from litellm.constants import (
SPEND_LOG_CLEANUP_BATCH_FAILURE_BACKOFF_SECONDS,
SPEND_LOG_CLEANUP_BATCH_SIZE,
SPEND_LOG_CLEANUP_BATCH_TIMEOUT_SECONDS,
SPEND_LOG_CLEANUP_JOB_NAME,
SPEND_LOG_CLEANUP_MAX_CONSECUTIVE_BATCH_FAILURES,
SPEND_LOG_CLEANUP_REMAINING_COUNT_CAP,
SPEND_LOG_CLEANUP_RUN_BUDGET_SECONDS,
SPEND_LOG_RUN_LOOPS,
)
from litellm.litellm_core_utils.duration_parser import duration_in_seconds
from litellm.proxy.db.db_transaction_queue.spend_log_cleanup_metrics import (
RunOutcome,
SpendLogCleanupMetrics,
)
from litellm.proxy.db.db_transaction_queue.spend_logs_partition_manager import (
RemainingTimeoutMs,
SpendLogsPartitionManager,
)
from litellm.proxy.utils import PrismaClient
StopReason: TypeAlias = Literal["exhausted", "budget_exhausted", "batch_cap_reached", "aborted"]
@dataclass(frozen=True, slots=True)
class TableCleanupResult:
"""Outcome of pruning one table, so the caller can report why a run ended."""
rows_deleted: int
stop_reason: StopReason
class _RemainingRow(BaseModel):
"""One row of the capped outstanding-rows probe, validated out of prisma's untyped result."""
remaining: int
_REMAINING_ROWS: Final = TypeAdapter(list[_RemainingRow])
SPEND_LOG_CLEANUP_BOUND_SETTINGS: Final = (
"maximum_spend_logs_cleanup_batch_size",
"maximum_spend_logs_cleanup_max_batches",
"maximum_spend_logs_cleanup_run_budget",
"maximum_spend_logs_cleanup_batch_timeout",
)
class SpendLogCleanup:
"""
@ -26,6 +64,24 @@ class SpendLogCleanup:
dropping whole partitions (instant, frees disk immediately). Otherwise it
falls back to deleting logs in batches.
Uses PodLockManager to ensure only one pod runs cleanup in multi-pod deployments.
Every run is bounded so it can never monopolise the database: a wall-clock
budget shared across all tables, a per-table batch cap, and a Postgres
statement/lock timeout on every statement the job issues, deletes and the
outstanding-rows probe alike. A run that hits a bound stops cleanly and the
next run resumes from where it left off, because the cutoff is recomputed
and deleted rows are gone.
The budget is a hard wall clock, not an advisory one. Every statement this
job issues, deletes, the outstanding-rows probe and partition DDL alike, is
issued with a timeout clamped to the budget that is still left, so one
started just under the deadline is cancelled by Postgres at the deadline
rather than running a further batch timeout past it. No statement is issued
at all once the budget is spent, which is why the probe is skipped on that
path. Partition DDL additionally carries a lock_timeout, because it takes an
ACCESS EXCLUSIVE lock and would otherwise queue behind a long-running reader
for as long as that reader lives; a partition this run cannot get is left
for the next one.
"""
def __init__(
@ -34,17 +90,88 @@ class SpendLogCleanup:
redis_cache: RedisCache | None = None,
partition_manager: SpendLogsPartitionManager | None = None,
):
self.batch_size = SPEND_LOG_CLEANUP_BATCH_SIZE
self.retention_seconds: int | None = None
self.partition_manager = partition_manager or SpendLogsPartitionManager()
from litellm.proxy.proxy_server import general_settings as default_settings
self.general_settings = general_settings or default_settings
self._refresh_bounds()
from litellm.proxy.proxy_server import proxy_logging_obj
pod_lock_manager: Final = proxy_logging_obj.db_spend_update_writer.pod_lock_manager
self.pod_lock_manager = pod_lock_manager
verbose_proxy_logger.info("SpendLogCleanup initialized with batch size: %s", self.batch_size)
verbose_proxy_logger.info(
"SpendLogCleanup initialized: batch_size=%s max_batches=%s run_budget=%ss batch_timeout=%ss",
self.batch_size,
self.max_batches,
self.run_budget_seconds,
self.batch_timeout_seconds,
)
def _refresh_bounds(self) -> None:
"""
Re-read every bound in SPEND_LOG_CLEANUP_BOUND_SETTINGS from settings.
The scheduler holds one long-lived instance, so a bound captured at
construction would never reflect a dashboard change. general_settings is
the same dict the periodic config reload mutates in place, so reading it
per run is what makes these knobs live. Every bound falls back to its
shipped default, so clearing a field restores that default.
"""
self.batch_size: int = self._positive_int_setting(
"maximum_spend_logs_cleanup_batch_size", SPEND_LOG_CLEANUP_BATCH_SIZE
)
self.max_batches: int = self._positive_int_setting(
"maximum_spend_logs_cleanup_max_batches", SPEND_LOG_RUN_LOOPS
)
self.run_budget_seconds: float = self._duration_setting(
"maximum_spend_logs_cleanup_run_budget", SPEND_LOG_CLEANUP_RUN_BUDGET_SECONDS
)
self.batch_timeout_seconds: float = self._duration_setting(
"maximum_spend_logs_cleanup_batch_timeout", SPEND_LOG_CLEANUP_BATCH_TIMEOUT_SECONDS
)
def _positive_int_setting(self, setting_name: str, default: int) -> int:
"""
Read a positive-integer knob, falling back to the default when unset or unusable.
"""
raw: Final = self.general_settings.get(setting_name)
if raw is None:
return default
try:
parsed: Final = int(raw)
except (TypeError, ValueError):
verbose_proxy_logger.warning("Invalid %s value: %s, using default %s", setting_name, raw, default)
return default
if parsed <= 0:
verbose_proxy_logger.warning("%s must be positive, got %s, using default %s", setting_name, parsed, default)
return default
return parsed
def _duration_setting(self, setting_name: str, default_seconds: float) -> float:
"""
Read a duration knob (e.g. '5m'), falling back to the default when unset or unusable.
The knob must never be able to remove the bound it exists to enforce, so
anything the parser rejects (including the non-finite spellings 'inf' and
'nan') and anything non-positive falls back rather than being honoured.
"""
raw: Final = self.general_settings.get(setting_name)
if raw is None:
return default_seconds
try:
parsed: Final = float(duration_in_seconds(str(raw)))
except (ValueError, TypeError) as e:
verbose_proxy_logger.warning(
"Invalid %s value: %s (%s), using default %ss", setting_name, raw, e, default_seconds
)
return default_seconds
if parsed <= 0:
verbose_proxy_logger.warning(
"%s must be a positive duration, got %s, using default %ss", setting_name, raw, default_seconds
)
return default_seconds
return parsed
def _retention_seconds_for(self, setting_name: str) -> int | None:
"""
@ -78,6 +205,91 @@ class SpendLogCleanup:
self.retention_seconds = self._retention_seconds_for("maximum_spend_logs_retention_period")
return self.retention_seconds is not None
def _timeout_ms(self, deadline: float) -> int:
"""
The per-statement bound in milliseconds: the batch timeout, or whatever
is left of the run budget, whichever is smaller.
Clamping to the remaining budget is what makes the budget a real
wall-clock bound rather than an advisory one. Postgres offers no "stop
at time T", only a per-statement duration, so a statement issued just
under the deadline would otherwise run a full batch timeout past it, and
with several tables those overruns stack.
Interpolating this into SQL is safe by construction: an int cannot carry
SQL, and SET does not accept a bind parameter.
"""
remaining_ms: Final = int((deadline - time.monotonic()) * 1000)
return max(1, min(int(self.batch_timeout_seconds * 1000), remaining_ms))
def _remaining_timeout_ms(self, deadline: float) -> RemainingTimeoutMs:
"""
The per-statement bound for work this job delegates, as a callable.
Partition maintenance issues one statement per partition, so handing it a
number would bound each statement by the budget that was left before the
FIRST one and never by what remains. Re-evaluating per statement is what
makes the loop itself bounded, and None tells the callee to stop rather
than issue a statement it has no budget for.
"""
def remaining() -> int | None:
return None if time.monotonic() >= deadline else self._timeout_ms(deadline)
return remaining
async def _execute_delete_batch(
self, prisma_client: PrismaClient, delete_sql: str, cutoff_date: datetime, deadline: float
) -> int | None:
"""
Run one delete batch under a Postgres statement and lock timeout.
The timeouts are what actually bound the work: a Prisma transaction
timeout cannot interrupt a statement that is already executing, so
without these a single batch blocked behind a lock would hold its
connection, and the row locks it already took, indefinitely. SET LOCAL
scopes both to this transaction so the pooled connection is unaffected.
Returns the row count, or None when the driver returned something that
is not a row count. That is a contract violation rather than a transient
fault, so the caller stops instead of retrying.
"""
timeout_ms: Final = self._timeout_ms(deadline)
async with prisma_client.db.tx() as tx:
await tx.execute_raw(f"SET LOCAL statement_timeout = {timeout_ms}")
await tx.execute_raw(f"SET LOCAL lock_timeout = {timeout_ms}")
deleted_result: Final = await tx.execute_raw(delete_sql, cutoff_date, self.batch_size)
return deleted_result if isinstance(deleted_result, int) else None
async def _count_remaining(
self, prisma_client: PrismaClient, cutoff_date: datetime, table_name: str, time_column: str, deadline: float
) -> int | None:
"""
Count expired rows still outstanding, stopping at a cap.
An uncapped COUNT(*) over an expired backlog would itself be the kind of
long scan this job exists to avoid, so the probe reads at most
SPEND_LOG_CLEANUP_REMAINING_COUNT_CAP index entries. A result equal to
the cap means "at least this many".
"""
count_sql: Final = f"""
SELECT count(*)::int AS remaining FROM (
SELECT 1 FROM "{table_name}"
WHERE "{time_column}" < $1::timestamptz
LIMIT $2
) capped
"""
try:
async with prisma_client.db.tx() as tx:
await tx.execute_raw(f"SET LOCAL statement_timeout = {self._timeout_ms(deadline)}")
rows: Final = _REMAINING_ROWS.validate_python(
await tx.query_raw(count_sql, cutoff_date, SPEND_LOG_CLEANUP_REMAINING_COUNT_CAP)
)
except Exception as e: # noqa: BLE001 - an observability probe must never fail the cleanup run
verbose_proxy_logger.warning("Could not count remaining %s rows: %s", table_name, e)
return None
return rows[0].remaining if rows else None
async def _delete_old_rows_batched(
self,
prisma_client: PrismaClient,
@ -85,10 +297,14 @@ class SpendLogCleanup:
table_name: str,
key_columns: tuple[str, ...],
time_column: str,
) -> int:
deadline: float,
) -> TableCleanupResult:
"""
Helper method to delete a table's rows older than the cutoff in batches.
Returns the total number of rows deleted.
Delete a table's rows older than the cutoff in batches.
Stops at whichever bound is reached first: the backlog running out, the
shared wall-clock deadline, the per-table batch cap, or too many
consecutive batch failures.
"""
key_list: Final = ", ".join(f'"{col}"' for col in key_columns)
delete_sql: Final = f"""
@ -103,23 +319,46 @@ class SpendLogCleanup:
run_count = 0
consecutive_failures = 0
while True:
if run_count > SPEND_LOG_RUN_LOOPS:
if time.monotonic() >= deadline:
verbose_proxy_logger.info(
"Run budget exhausted during %s cleanup after %d rows; the next run resumes from here",
table_name,
total_deleted,
)
return await self._finish_table(
prisma_client, cutoff_date, table_name, time_column, total_deleted, "budget_exhausted", deadline
)
if run_count >= self.max_batches:
verbose_proxy_logger.info(
"Max batches reached for %s cleanup, remaining rows will be deleted in next run", table_name
)
break
# Step 1: Find rows and delete them in one go without fetching to application
# Delete in batches, limited by self.batch_size
try:
deleted_result = await prisma_client.db.execute_raw(
delete_sql,
cutoff_date,
self.batch_size,
return await self._finish_table(
prisma_client, cutoff_date, table_name, time_column, total_deleted, "batch_cap_reached", deadline
)
# Find rows and delete them in one go without fetching to application
batch_started_at = time.monotonic()
try:
batch_result = await self._execute_delete_batch(prisma_client, delete_sql, cutoff_date, deadline)
except Exception as batch_exc:
if time.monotonic() >= deadline:
# The statement timeout was clamped to the budget that was
# left, so this batch was cancelled by the deadline itself.
# That is the bound working, not a database fault, and
# counting it would both inflate the failure metric and push
# every budget-exhausted run toward the abort threshold.
verbose_proxy_logger.info(
"Run budget exhausted mid-batch during %s cleanup after %d rows; "
"the next run resumes from here",
table_name,
total_deleted,
)
return await self._finish_table(
prisma_client, cutoff_date, table_name, time_column, total_deleted, "budget_exhausted", deadline
)
# A single batch failure (e.g. Prisma/DB timeout) must not abort
# the whole run — subsequent batches may still succeed.
consecutive_failures += 1
SpendLogCleanupMetrics.record_batch_failure(table_name)
verbose_proxy_logger.exception(
"%s cleanup batch failed "
"(run_count=%d, consecutive_failures=%d, batch_size=%d, "
@ -140,28 +379,31 @@ class SpendLogCleanup:
consecutive_failures,
total_deleted,
)
break
return await self._finish_table(
prisma_client, cutoff_date, table_name, time_column, total_deleted, "aborted", deadline
)
await asyncio.sleep(SPEND_LOG_CLEANUP_BATCH_FAILURE_BACKOFF_SECONDS)
continue
consecutive_failures = 0
deleted_count = 0
if isinstance(deleted_result, int):
deleted_count = deleted_result
else:
if batch_result is None:
verbose_proxy_logger.error(
"Unexpected execute_raw return type for %s cleanup: %s; aborting cleanup to avoid infinite loop",
"Unexpected execute_raw return type for %s cleanup; aborting cleanup to avoid infinite loop",
table_name,
type(deleted_result),
)
break
return await self._finish_table(
prisma_client, cutoff_date, table_name, time_column, total_deleted, "aborted", deadline
)
consecutive_failures = 0
deleted_count = batch_result
SpendLogCleanupMetrics.record_batch(table_name, deleted_count, time.monotonic() - batch_started_at)
verbose_proxy_logger.info("Deleted %s %s rows in this batch", deleted_count, table_name)
if deleted_count == 0:
verbose_proxy_logger.info("No more %s rows to delete. Total deleted: %s", table_name, total_deleted)
break
return await self._finish_table(
prisma_client, cutoff_date, table_name, time_column, total_deleted, "exhausted", deadline
)
total_deleted += deleted_count
run_count += 1
@ -169,18 +411,49 @@ class SpendLogCleanup:
# Add a small sleep to prevent overwhelming the database
await asyncio.sleep(0.1)
return total_deleted
async def _finish_table(
self,
prisma_client: PrismaClient,
cutoff_date: datetime,
table_name: str,
time_column: str,
rows_deleted: int,
stop_reason: StopReason,
deadline: float,
) -> TableCleanupResult:
"""
Publish how much of this table is still outstanding, then report the run's result.
async def _delete_old_logs(self, prisma_client: PrismaClient, cutoff_date: datetime) -> int:
The probe is skipped once the budget is spent. It is the one piece of
work that would otherwise be ISSUED after the deadline, and every table
exits through here, including the ones a spent run never started, so
keeping it would put one more statement per table past the bound. A run
that ends this way already reports "budget_exhausted", which tells an
operator the backlog was not drained; the gauge simply keeps its value
from the last run that finished inside its budget.
"""
if time.monotonic() >= deadline:
return TableCleanupResult(rows_deleted=rows_deleted, stop_reason=stop_reason)
remaining: Final = await self._count_remaining(prisma_client, cutoff_date, table_name, time_column, deadline)
if remaining is not None:
SpendLogCleanupMetrics.set_rows_remaining(table_name, remaining)
return TableCleanupResult(rows_deleted=rows_deleted, stop_reason=stop_reason)
async def _delete_old_logs(
self, prisma_client: PrismaClient, cutoff_date: datetime, deadline: float
) -> TableCleanupResult:
return await self._delete_old_rows_batched(
prisma_client,
cutoff_date,
table_name="LiteLLM_SpendLogs",
key_columns=("request_id", "startTime"),
time_column="startTime",
deadline=deadline,
)
async def _delete_old_tool_index_rows(self, prisma_client: PrismaClient, cutoff_date: datetime) -> int:
async def _delete_old_tool_index_rows(
self, prisma_client: PrismaClient, cutoff_date: datetime, deadline: float
) -> TableCleanupResult:
# SpendLogToolIndex rows are derived from spend logs, so they expire on the
# same cutoff; rows older than retention point at already-deleted logs.
return await self._delete_old_rows_batched(
@ -189,17 +462,87 @@ class SpendLogCleanup:
table_name="LiteLLM_SpendLogToolIndex",
key_columns=("request_id", "tool_name"),
time_column="start_time",
deadline=deadline,
)
async def _delete_old_autorouter_session_rows(self, prisma_client: PrismaClient, cutoff_date: datetime) -> int:
async def _delete_old_autorouter_session_rows(
self, prisma_client: PrismaClient, cutoff_date: datetime, deadline: float
) -> TableCleanupResult:
return await self._delete_old_rows_batched(
prisma_client,
cutoff_date,
table_name="LiteLLM_AutoRouterSession",
key_columns=("api_key", "session_id", "router_name"),
time_column="last_turn_at",
deadline=deadline,
)
async def _clean_spend_log_tables(
self, prisma_client: PrismaClient, deadline: float
) -> tuple[TableCleanupResult, ...]:
"""
Prune the spend logs and the tool index rows derived from them.
When the table is range-partitioned, whole expired partitions are dropped
first because that reclaims disk immediately. Expired rows can still sit in
the DEFAULT partition (backfill, coverage gaps) or in a partition that spans
the cutoff, so retention still deletes those stragglers row-wise.
"""
cutoff_date: Final = datetime.now(timezone.utc) - timedelta(seconds=float(self.retention_seconds or 0))
verbose_proxy_logger.info("Removing logs older than %s", cutoff_date.isoformat())
# Partition maintenance is DDL taking an ACCESS EXCLUSIVE lock, so it is
# only STARTED while the run still has budget, and each statement carries
# the same timeouts the batches do. Without those, a DROP would queue
# behind any long-running reader for as long as that reader lives, which
# is the one way this job could still outlast its budget without bound.
remaining_timeout_ms: Final = self._remaining_timeout_ms(deadline)
if time.monotonic() >= deadline:
verbose_proxy_logger.info("Run budget already spent, skipping partition maintenance this run")
elif self.general_settings.get(
"use_spend_logs_partitioning", False
) and await self.partition_manager.is_partitioned(prisma_client, remaining_timeout_ms):
await self.partition_manager.ensure_partitions(prisma_client, remaining_timeout_ms)
dropped: Final = await self.partition_manager.drop_partitions_older_than(
prisma_client, cutoff_date, remaining_timeout_ms
)
verbose_proxy_logger.info("Dropped %d expired spend-log partitions: %s", len(dropped), dropped)
logs_result: Final = await self._delete_old_logs(prisma_client, cutoff_date, deadline)
verbose_proxy_logger.info("Deleted %s logs", logs_result.rows_deleted)
index_result: Final = await self._delete_old_tool_index_rows(prisma_client, cutoff_date, deadline)
verbose_proxy_logger.info("Deleted %s expired tool index rows", index_result.rows_deleted)
return (logs_result, index_result)
async def _clean_session_rollup(
self, prisma_client: PrismaClient, retention_seconds: int, deadline: float
) -> tuple[TableCleanupResult, ...]:
"""
Prune auto-router session rollup rows, which carry their own retention horizon.
"""
session_cutoff: Final = datetime.now(timezone.utc) - timedelta(seconds=float(retention_seconds))
sessions_result: Final = await self._delete_old_autorouter_session_rows(prisma_client, session_cutoff, deadline)
verbose_proxy_logger.info("Deleted %s expired auto-router session rollup rows", sessions_result.rows_deleted)
return (sessions_result,)
@staticmethod
def _run_outcome(results: tuple[TableCleanupResult, ...]) -> RunOutcome:
"""
Report the most operationally significant reason the run stopped.
A bound that was hit matters more than a table that simply ran dry, so
those win over "completed", and an abort wins over everything.
"""
reasons: Final = frozenset(result.stop_reason for result in results)
if "aborted" in reasons:
return "aborted"
if "budget_exhausted" in reasons:
return "budget_exhausted"
if "batch_cap_reached" in reasons:
return "batch_cap_reached"
return "completed"
async def cleanup_old_spend_logs(self, prisma_client: PrismaClient) -> None:
"""
Main cleanup function. Deletes old spend logs in batches.
@ -209,16 +552,19 @@ class SpendLogCleanup:
lock_acquired = False
try:
verbose_proxy_logger.info("Cleanup job triggered at %s", datetime.now())
self._refresh_bounds()
delete_spend_logs: Final = self._should_delete_spend_logs()
autorouter_retention_seconds: Final = self._retention_seconds_for(
"maximum_autorouter_session_retention_period"
)
if not delete_spend_logs and autorouter_retention_seconds is None:
SpendLogCleanupMetrics.record_run("skipped_disabled")
return
if delete_spend_logs and self.retention_seconds is None:
verbose_proxy_logger.error("Retention seconds is None, cannot proceed with cleanup")
SpendLogCleanupMetrics.record_run("skipped_disabled")
return
# If we have a pod lock manager, try to acquire the lock
@ -235,43 +581,23 @@ class SpendLogCleanup:
if not lock_acquired:
verbose_proxy_logger.info("Another pod is already running cleanup")
SpendLogCleanupMetrics.record_run("skipped_locked")
return
if delete_spend_logs and self.retention_seconds is not None:
cutoff_date: Final = datetime.now(timezone.utc) - timedelta(seconds=float(self.retention_seconds))
verbose_proxy_logger.info("Removing logs older than %s", cutoff_date.isoformat())
deadline: Final = time.monotonic() + self.run_budget_seconds
if self.general_settings.get(
"use_spend_logs_partitioning", False
) and await self.partition_manager.is_partitioned(prisma_client):
await self.partition_manager.ensure_partitions(prisma_client)
dropped: Final = await self.partition_manager.drop_partitions_older_than(prisma_client, cutoff_date)
verbose_proxy_logger.info(
"Dropped %d expired spend-log partitions: %s",
len(dropped),
dropped,
)
# DROP only reclaims whole expired partitions. Expired rows can
# still sit in the DEFAULT partition (backfill, coverage gaps)
# or in a partition that spans the cutoff, so retention must
# also delete those stragglers row-wise.
total_deleted = await self._delete_old_logs(prisma_client, cutoff_date)
verbose_proxy_logger.info(
"Deleted %s expired logs not covered by dropped partitions", total_deleted
)
else:
total_deleted = await self._delete_old_logs(prisma_client, cutoff_date)
verbose_proxy_logger.info("Deleted %s logs", total_deleted)
spend_log_results: Final = (
await self._clean_spend_log_tables(prisma_client, deadline)
if delete_spend_logs and self.retention_seconds is not None
else ()
)
session_results: Final = (
await self._clean_session_rollup(prisma_client, autorouter_retention_seconds, deadline)
if autorouter_retention_seconds is not None
else ()
)
index_deleted: Final = await self._delete_old_tool_index_rows(prisma_client, cutoff_date)
verbose_proxy_logger.info("Deleted %s expired tool index rows", index_deleted)
if autorouter_retention_seconds is not None:
session_cutoff: Final = datetime.now(timezone.utc) - timedelta(
seconds=float(autorouter_retention_seconds)
)
sessions_deleted: Final = await self._delete_old_autorouter_session_rows(prisma_client, session_cutoff)
verbose_proxy_logger.info("Deleted %s expired auto-router session rollup rows", sessions_deleted)
SpendLogCleanupMetrics.record_run(self._run_outcome(spend_log_results + session_results))
except Exception as e:
# .exception() captures the traceback; str(e) alone on a Prisma/DB
@ -281,6 +607,7 @@ class SpendLogCleanup:
type(e).__name__,
e,
)
SpendLogCleanupMetrics.record_run("aborted")
return # Return after error handling
finally:
# Only release the lock if it was actually acquired

122
litellm/proxy/db/db_transaction_queue/spend_log_cleanup_metrics.py Normal file
View file

@ -0,0 +1,122 @@
"""
Prometheus metrics for the spend-log retention cleanup job.
The job runs in the background on a single elected pod, so its cost is invisible
from request-path metrics. These instruments make a run's database footprint
observable: how much it deleted, how long each batch took, how much work is
still outstanding, and why a run stopped.
``prometheus_client`` is an optional dependency, so every recorder degrades to a
no-op when it is absent.
"""
from typing import TYPE_CHECKING, Final, Literal, TypeAlias
from litellm._logging import verbose_proxy_logger
if TYPE_CHECKING:
# aliased so the annotations below cannot be mistaken for collections.Counter
from prometheus_client import Counter as PrometheusCounter
from prometheus_client import Gauge as PrometheusGauge
from prometheus_client import Histogram as PrometheusHistogram
RunOutcome: TypeAlias = Literal[
"completed",
"budget_exhausted",
"batch_cap_reached",
"skipped_locked",
"skipped_disabled",
"aborted",
]
_BATCH_DURATION_BUCKETS: Final = (0.005, 0.025, 0.1, 0.25, 0.5, 1.0, 2.5, 5.0, 10.0, 30.0, 60.0)
_TABLE_LABEL: Final = ("table",)
_OUTCOME_LABEL: Final = ("outcome",)
class SpendLogCleanupMetrics:
"""
Lazily-registered Prometheus instruments for the retention cleanup job.
Registration is deferred to first use so that importing this module never
touches the Prometheus registry, which keeps it safe to import from the
proxy regardless of whether Prometheus is a configured callback.
"""
_initialized: bool = False
rows_deleted: "PrometheusCounter | None" = None
batch_duration: "PrometheusHistogram | None" = None
rows_remaining: "PrometheusGauge | None" = None
batch_failures: "PrometheusCounter | None" = None
runs: "PrometheusCounter | None" = None
@classmethod
def _ensure_initialized(cls) -> None:
if cls._initialized:
return
cls._initialized = True
try:
# prometheus_client is an optional extra, so it is resolved here rather
# than at module import: this module is reachable from proxy startup
# regardless of whether Prometheus is a configured callback.
from prometheus_client import Counter, Gauge, Histogram
cls.rows_deleted = Counter(
"litellm_spend_log_cleanup_rows_deleted_total",
"Rows deleted by the spend-log retention cleanup job",
labelnames=_TABLE_LABEL,
)
cls.batch_duration = Histogram(
"litellm_spend_log_cleanup_batch_duration_seconds",
"Wall-clock duration of one retention cleanup delete batch",
labelnames=_TABLE_LABEL,
buckets=_BATCH_DURATION_BUCKETS,
)
cls.rows_remaining = Gauge(
"litellm_spend_log_cleanup_rows_remaining",
"Expired rows still awaiting deletion, counted only up to "
"SPEND_LOG_CLEANUP_REMAINING_COUNT_CAP so the probe itself cannot scan a "
"large table; a value equal to that cap means at least that many remain",
labelnames=_TABLE_LABEL,
multiprocess_mode="livemax",
)
cls.batch_failures = Counter(
"litellm_spend_log_cleanup_batch_failures_total",
"Retention cleanup delete batches that raised",
labelnames=_TABLE_LABEL,
)
cls.runs = Counter(
"litellm_spend_log_cleanup_runs_total",
"Retention cleanup runs, labelled by why the run ended",
labelnames=_OUTCOME_LABEL,
)
except Exception as e: # noqa: BLE001 - a metrics problem must never fail the cleanup run
# Covers the extra being absent, a duplicate registration (repeated
# imports under a test runner), and registry misconfiguration alike.
verbose_proxy_logger.warning("Could not register spend-log cleanup metrics: %s", e)
@classmethod
def record_batch(cls, table_name: str, rows_deleted: int, duration_seconds: float) -> None:
cls._ensure_initialized()
if cls.rows_deleted is not None:
cls.rows_deleted.labels(table=table_name).inc(rows_deleted)
if cls.batch_duration is not None:
cls.batch_duration.labels(table=table_name).observe(duration_seconds)
@classmethod
def record_batch_failure(cls, table_name: str) -> None:
cls._ensure_initialized()
if cls.batch_failures is not None:
cls.batch_failures.labels(table=table_name).inc()
@classmethod
def set_rows_remaining(cls, table_name: str, remaining: int) -> None:
cls._ensure_initialized()
if cls.rows_remaining is not None:
cls.rows_remaining.labels(table=table_name).set(remaining)
@classmethod
def record_run(cls, outcome: RunOutcome) -> None:
cls._ensure_initialized()
if cls.runs is not None:
cls.runs.labels(outcome=outcome).inc()

128
litellm/proxy/db/db_transaction_queue/spend_logs_partition_manager.py
View file

@ -14,8 +14,9 @@ keeps the batched-DELETE path, so existing deployments are untouched.
"""
import re
from collections.abc import Callable
from datetime import date, datetime, timedelta, timezone
from typing import Final
from typing import TYPE_CHECKING, Final, TypeAlias
from litellm._logging import verbose_proxy_logger
from litellm.constants import (
@ -23,8 +24,23 @@ from litellm.constants import (
SPEND_LOG_PARTITION_PRECREATE_AHEAD,
)
if TYPE_CHECKING:
from litellm.proxy.utils import PrismaClient
SPEND_LOGS_TABLE: Final = "LiteLLM_SpendLogs"
RemainingTimeoutMs: TypeAlias = Callable[[], "int | None"]
"""
The per-statement bound in milliseconds, or None once the caller's budget is
spent.
Injected rather than passed as a number so it is re-evaluated before EVERY
statement: a value read once at entry would let a loop issue N statements each
bounded by the budget that was left before the first of them, which is not a
bound on the loop at all. The caller owns the policy; this module only asks how
much time it may still use.
"""
PartitionInterval = str # "day" | "week" | "month"
VALID_PARTITION_INTERVALS: Final = {"day", "week", "month"}
@ -116,21 +132,26 @@ class SpendLogsPartitionManager:
self.interval = interval
self.precreate_ahead = precreate_ahead
async def is_partitioned(self, prisma_client) -> bool:
async def is_partitioned(self, prisma_client: "PrismaClient", remaining_timeout_ms: RemainingTimeoutMs) -> bool:
budget_ms: Final = remaining_timeout_ms()
if budget_ms is None:
return False
try:
rows: Final = await prisma_client.db.query_raw(
"""
SELECT EXISTS (
SELECT 1
FROM pg_partitioned_table pt
JOIN pg_class c ON c.oid = pt.partrelid
JOIN pg_namespace n ON n.oid = c.relnamespace
WHERE c.relname = $1
AND n.nspname = current_schema()
) AS partitioned
""",
SPEND_LOGS_TABLE,
)
async with prisma_client.db.tx() as tx:
await tx.execute_raw(f"SET LOCAL statement_timeout = {budget_ms}")
rows: Final = await tx.query_raw(
"""
SELECT EXISTS (
SELECT 1
FROM pg_partitioned_table pt
JOIN pg_class c ON c.oid = pt.partrelid
JOIN pg_namespace n ON n.oid = c.relnamespace
WHERE c.relname = $1
AND n.nspname = current_schema()
) AS partitioned
""",
SPEND_LOGS_TABLE,
)
except Exception as e:
verbose_proxy_logger.warning(
"Could not determine if %s is partitioned, assuming it is not: %s",
@ -140,7 +161,25 @@ class SpendLogsPartitionManager:
return False
return bool(rows and rows[0].get("partitioned"))
async def ensure_partitions(self, prisma_client) -> list[str]:
@staticmethod
async def _execute_bounded_ddl(prisma_client: "PrismaClient", statement: str, timeout_ms: int) -> None:
"""
Run one DDL statement under a Postgres statement and lock timeout.
Partition DDL takes an ACCESS EXCLUSIVE lock, so an unbounded statement
queues behind any long-running reader for as long as that reader lives,
and the caller's run budget cannot cut it short. lock_timeout bounds the
wait for the lock and statement_timeout bounds the work itself, so a
partition this run cannot get is simply left for the next one.
"""
async with prisma_client.db.tx() as tx:
await tx.execute_raw(f"SET LOCAL statement_timeout = {timeout_ms}")
await tx.execute_raw(f"SET LOCAL lock_timeout = {timeout_ms}")
await tx.execute_raw(statement)
async def ensure_partitions(
self, prisma_client: "PrismaClient", remaining_timeout_ms: RemainingTimeoutMs
) -> list[str]:
"""
Ensure the current and upcoming partitions exist, returning the names
now present. CREATE TABLE IF NOT EXISTS is a no-op for partitions that
@ -150,42 +189,61 @@ class SpendLogsPartitionManager:
for name, lower, upper in upcoming_partitions(
datetime.now(timezone.utc).date(), self.interval, self.precreate_ahead
):
budget_ms = remaining_timeout_ms()
if budget_ms is None:
verbose_proxy_logger.info("Run budget spent, leaving the remaining partitions for the next run")
break
try:
await prisma_client.db.execute_raw(
await self._execute_bounded_ddl(
prisma_client,
f'CREATE TABLE IF NOT EXISTS "{name}" '
f'PARTITION OF "{SPEND_LOGS_TABLE}" '
f"FOR VALUES FROM ('{lower.isoformat()}') TO ('{upper.isoformat()}')"
f"FOR VALUES FROM ('{lower.isoformat()}') TO ('{upper.isoformat()}')",
budget_ms,
)
ensured.append(name)
except Exception as e:
verbose_proxy_logger.warning("Failed to ensure spend-log partition %s: %s", name, e)
return ensured
async def _list_partitions(self, prisma_client) -> list[tuple[str, datetime | None]]:
rows: Final = await prisma_client.db.query_raw(
"""
SELECT c.relname AS name,
pg_get_expr(c.relpartbound, c.oid) AS bound
FROM pg_inherits i
JOIN pg_class c ON c.oid = i.inhrelid
JOIN pg_class p ON p.oid = i.inhparent
JOIN pg_namespace n ON n.oid = p.relnamespace
WHERE p.relname = $1
AND n.nspname = current_schema()
""",
SPEND_LOGS_TABLE,
)
async def _list_partitions(
self, prisma_client: "PrismaClient", timeout_ms: int
) -> list[tuple[str, datetime | None]]:
async with prisma_client.db.tx() as tx:
await tx.execute_raw(f"SET LOCAL statement_timeout = {timeout_ms}")
rows: Final = await tx.query_raw(
"""
SELECT c.relname AS name,
pg_get_expr(c.relpartbound, c.oid) AS bound
FROM pg_inherits i
JOIN pg_class c ON c.oid = i.inhrelid
JOIN pg_class p ON p.oid = i.inhparent
JOIN pg_namespace n ON n.oid = p.relnamespace
WHERE p.relname = $1
AND n.nspname = current_schema()
""",
SPEND_LOGS_TABLE,
)
return [(row["name"], parse_partition_upper_bound(row.get("bound") or "")) for row in rows]
async def drop_partitions_older_than(self, prisma_client, cutoff: datetime) -> list[str]:
async def drop_partitions_older_than(
self, prisma_client: "PrismaClient", cutoff: datetime, remaining_timeout_ms: RemainingTimeoutMs
) -> list[str]:
"""DROP every partition whose whole range is older than `cutoff`."""
list_budget_ms: Final = remaining_timeout_ms()
if list_budget_ms is None:
return []
cutoff_naive: Final = cutoff.astimezone(timezone.utc).replace(tzinfo=None)
partitions: Final = await self._list_partitions(prisma_client)
partitions: Final = await self._list_partitions(prisma_client, list_budget_ms)
to_drop: Final = select_partitions_to_drop(partitions, cutoff_naive)
dropped: Final[list[str]] = []
for name in to_drop:
budget_ms = remaining_timeout_ms()
if budget_ms is None:
verbose_proxy_logger.info("Run budget spent, leaving the remaining partitions for the next run")
break
try:
await prisma_client.db.execute_raw(f'DROP TABLE IF EXISTS "{name}"')
await self._execute_bounded_ddl(prisma_client, f'DROP TABLE IF EXISTS "{name}"', budget_ms)
dropped.append(name)
except Exception as e:
verbose_proxy_logger.warning("Failed to drop spend-log partition %s: %s", name, e)

14
litellm/proxy/db/routing_prisma_wrapper.py
View file

@ -103,6 +103,17 @@ class RoutingPrismaWrapper:
def reader(self) -> PrismaWrapper:
return self._reader
@property
def read_target(self) -> PrismaWrapper:
"""The wrapper `_TOP_LEVEL_READ_METHODS` dispatch to right now.
Callers that need to reason about the engine a read actually ran on
(e.g. recovering from prepared statements that went stale on it) must
consult this rather than `writer`, and `__getattr__` routes through it
so the two cannot drift apart.
"""
return self._writer if self._reader_unavailable else self._reader
@property
def reader_unavailable(self) -> bool:
return self._reader_unavailable
@ -254,8 +265,7 @@ class RoutingPrismaWrapper:
def __getattr__(self, name: str) -> Any:
if name in _TOP_LEVEL_READ_METHODS:
target: Final = self._writer if self._reader_unavailable else self._reader
return getattr(target, name)
return getattr(self.read_target, name)
writer_attr: Final = getattr(self._writer, name)
# Per-model action accessors are non-callable instances that expose
# both `find_many` and `create`. Methods like execute_raw / batch_ /

125
litellm/proxy/guardrails/anthropic_sse.py Normal file
View file

@ -0,0 +1,125 @@
"""Anthropic SSE <-> ModelResponse conversion for guardrail streaming hooks.
`/v1/messages` streams reach a guardrail's `async_post_call_streaming_iterator_hook` as raw SSE
frames rather than chunk objects, which `stream_chunk_builder` cannot assemble. These helpers let a
hook scan such a stream, and re-emit it when the guardrail rewrote the response.
"""
from __future__ import annotations
import json
from collections.abc import Mapping, Sequence
from typing import Final
from litellm.types.utils import Choices, ModelResponse
def is_raw_sse_stream(all_chunks: Sequence[object]) -> bool:
return any(isinstance(chunk, (str, bytes)) for chunk in all_chunks)
def _joined_sse_stream(all_chunks: Sequence[object]) -> str | None:
raw: Final = b"".join(
chunk if isinstance(chunk, bytes) else chunk.encode("utf-8")
for chunk in all_chunks
if isinstance(chunk, (str, bytes))
)
try:
return raw.decode("utf-8")
except UnicodeDecodeError:
return None
def _anthropic_message_start(sse_stream: str) -> Mapping[str, object] | None:
from litellm.proxy.pass_through_endpoints.llm_provider_handlers.anthropic_passthrough_logging_handler import (
AnthropicPassthroughLoggingHandler,
)
return next(
(
message
for event in AnthropicPassthroughLoggingHandler._split_sse_chunk_into_events(sse_stream) # pyright: ignore[reportPrivateUsage] # same parser the assembler uses
if (event_data := AnthropicPassthroughLoggingHandler._extract_sse_data(event)) is not None # pyright: ignore[reportPrivateUsage] # same parser the assembler uses; a private import beats forking SSE parsing
and event_data.get("type") == "message_start"
and isinstance(message := event_data.get("message"), dict)
),
None,
)
def assemble_anthropic_sse_stream(
all_chunks: Sequence[object], *, restore_identity: bool = False
) -> ModelResponse | None:
"""Assemble raw Anthropic SSE frames into a ModelResponse.
``restore_identity`` stamps the upstream message id and model onto the result, which the
assembler does not carry through. It is off by default so callers that re-emit the assembled
response keep the wire shape they had before this helper was shared. The writes land on a
freshly built object that is unreachable from caller state until returned.
"""
from litellm.proxy.pass_through_endpoints.llm_provider_handlers.anthropic_passthrough_logging_handler import (
AnthropicPassthroughLoggingHandler,
)
sse_stream: Final = _joined_sse_stream(all_chunks)
if sse_stream is None:
return None
message_start: Final = _anthropic_message_start(sse_stream)
if message_start is None:
return None
model: Final = message_start.get("model") if restore_identity else None
try:
assembled: Final = AnthropicPassthroughLoggingHandler._build_complete_streaming_response( # pyright: ignore[reportPrivateUsage] # the only SSE-to-ModelResponse assembler; reimplementing it here would fork the parser
all_chunks=(sse_stream,),
litellm_logging_obj=None, # pyright: ignore[reportArgumentType] # only forwarded to stream_chunk_builder, which accepts None
model=model if isinstance(model, str) else "",
)
except Exception: # noqa: BLE001 # stream_chunk_builder re-raises every assembly failure as litellm.APIError
return None
if not isinstance(assembled, ModelResponse):
return None
if not restore_identity:
return assembled
message_id: Final = message_start.get("id")
if isinstance(message_id, str):
assembled.id = message_id
if isinstance(model, str) and model:
assembled.model = model
return assembled
def model_response_text(response: ModelResponse) -> str:
"""Assistant text of a response, used to detect whether a guardrail rewrote it."""
return "".join(
choice.message.content
for choice in response.choices
if isinstance(choice, Choices) # pyright: ignore[reportUnnecessaryIsInstance] # runtime choices can be StreamingChoices
and isinstance(choice.message.content, str)
)
def anthropic_sse_error_frames(message: str) -> tuple[bytes, ...]:
"""Anthropic error event, for a failure discovered after the response headers were flushed.
Once a keepalive ping has been sent a raise cannot reach the client, so the failure has to
travel as a frame.
"""
body: Final = json.dumps(message)
return (
f'event: error\ndata: {{"type": "error", "error": {{"type": "guardrail_error", '
f'"message": {body}}}}}\n\n'.encode(),
)
def anthropic_sse_chunks_from_response(assembled: ModelResponse) -> tuple[bytes, ...]:
from litellm.llms.anthropic.experimental_pass_through.adapters.transformation import (
LiteLLMAnthropicMessagesAdapter,
)
from litellm.llms.anthropic.experimental_pass_through.messages.fake_stream_iterator import (
FakeAnthropicMessagesStreamIterator,
)
anthropic_response: Final = LiteLLMAnthropicMessagesAdapter().translate_openai_response_to_anthropic(
response=assembled
)
return tuple(FakeAnthropicMessagesStreamIterator(response=anthropic_response).chunks)

65
litellm/proxy/guardrails/guardrail_hooks/bedrock_guardrails.py
View file

@ -14,6 +14,7 @@ import copy
import json
import re
import sys
import time
from collections.abc import AsyncGenerator, Mapping, Sequence
from datetime import datetime, timezone
from itertools import accumulate, groupby
@ -30,6 +31,7 @@ from litellm.constants import BEDROCK_APPLY_GUARDRAIL_CHUNK_BUDGET_CHARS
from litellm.exceptions import ModifyResponseException
from litellm.integrations.custom_guardrail import CustomGuardrail
from litellm.litellm_core_utils.core_helpers import redact_nested_match_and_regex_keys
from litellm.llms.anthropic.chat.guardrail_translation.handler import AnthropicMessagesHandler
from litellm.llms.base_llm.guardrail_translation.utils import (
effective_scan_only_tool_results_for_guardrail,
)
@ -39,6 +41,15 @@ from litellm.llms.custom_httpx.http_handler import (
httpxSpecialProvider,
)
from litellm.proxy._types import UserAPIKeyAuth
from litellm.proxy.common_request_processing import _serialize_http_exception_detail
from litellm.proxy.common_utils.sse_keepalive import keepalive_ping_has_fired
from litellm.proxy.guardrails.anthropic_sse import (
anthropic_sse_chunks_from_response,
anthropic_sse_error_frames,
assemble_anthropic_sse_stream,
is_raw_sse_stream,
model_response_text,
)
from litellm.secret_managers.main import get_secret_str
from litellm.types.guardrails import BedrockChecksConfigModel, GuardrailEventHooks
from litellm.types.llms.openai import AllMessageValues, ChatCompletionUserMessage
@ -2578,14 +2589,21 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
from litellm.types.utils import TextCompletionResponse
# Collect all chunks to process them together
started_at: Final = time.monotonic()
all_chunks: Final[list[ModelResponseStream]] = []
async for chunk in response:
all_chunks.append(chunk)
assembled_model_response: ModelResponse | TextCompletionResponse | None = stream_chunk_builder(
chunks=all_chunks,
# /v1/messages arrives as SSE frames, which stream_chunk_builder cannot assemble
raw_sse: Final = is_raw_sse_stream(all_chunks)
assembled_model_response: ModelResponse | TextCompletionResponse | None = (
assemble_anthropic_sse_stream(all_chunks, restore_identity=True)
if raw_sse
else stream_chunk_builder(chunks=all_chunks)
)
if isinstance(assembled_model_response, ModelResponse):
pre_guardrail_text: Final = model_response_text(assembled_model_response)
_pre_block_response: Final = assembled_model_response
####################################################################
########## 1. Make Bedrock Apply Guardrail API request ##########
#
@ -2609,7 +2627,32 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
request_data=request_data,
logging_event_type=GuardrailEventHooks.post_call,
)
except HTTPException as block_exc:
block_detail: Final = block_exc.detail
# A policy block is the only 400 carrying a structured detail; a service failure
# either details a plain string or reports a non-400 status. Re-raising a service
# failure keeps its real status, but only while the headers are unflushed: past the
# first keepalive ping the raise reaches nobody, so it has to travel as a frame too
is_block: Final = raw_sse and block_exc.status_code == 400 and isinstance(block_detail, Mapping)
headers_flushed: Final = keepalive_ping_has_fired(
time.monotonic() - started_at, litellm.anthropic_sse_ping_interval_seconds
)
if not raw_sse or (not is_block and not headers_flushed):
raise
block_message, _ = _serialize_http_exception_detail(block_detail)
for error_frame in anthropic_sse_error_frames(
block_message if is_block else f"{block_exc.status_code}: {block_message}"
):
yield error_frame
return
except ModifyResponseException as e:
if raw_sse:
e.model = _pre_block_response.model or e.model # rebind-ok: exc.model defaults to the guardrail
if e.original_response is None:
e.original_response = _pre_block_response # rebind-ok: the block builder reads usage off this
for block_chunk in AnthropicMessagesHandler().build_block_sse_chunks(e, stream_started=False):
yield block_chunk
return
# Preserve upstream usage from the LLM call we already
# consumed. Non-streaming blocks carry it via
# ModifyResponseException.original_response +
@ -2642,11 +2685,29 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
#########################################################################
########## 3. Return the (potentially masked) chunks ##########
#########################################################################
if raw_sse:
for sse_chunk in (
anthropic_sse_chunks_from_response(assembled_model_response)
if model_response_text(assembled_model_response) != pre_guardrail_text
else all_chunks
):
yield sse_chunk
return
mock_response: Final = MockResponseIterator(model_response=assembled_model_response)
# Return the reconstructed stream
async for chunk in mock_response:
yield chunk
elif raw_sse:
# Forwarding an unscannable stream would silently disable the guardrail, so fail closed.
# A raise cannot reach the client once a keepalive ping has flushed the headers, so the
# refusal travels as a frame, matching how a block is delivered above
for error_frame in anthropic_sse_error_frames(
f"{self.guardrail_name}: streamed response could not be assembled for scanning, blocking it"
):
yield error_frame
return
else:
for chunk in all_chunks:
yield chunk

32
litellm/proxy/guardrails/guardrail_hooks/hiddenlayer/hiddenlayer.py
View file

@ -11,6 +11,7 @@ import requests
from fastapi import HTTPException
from httpx import HTTPStatusError
from requests.auth import HTTPBasicAuth
from typing_extensions import ReadOnly
from litellm._logging import verbose_proxy_logger
from litellm.integrations.custom_guardrail import (
@ -55,6 +56,26 @@ class _HiddenlayerResponse(TypedDict, total=False):
modified_data: Mapping[str, _HiddenlayerModifiedSide]
class _LoggedCallMetadata(TypedDict, total=False):
headers: ReadOnly[Mapping[str, str]]
class _LoggedCallLitellmParams(TypedDict, total=False):
metadata: ReadOnly[_LoggedCallMetadata]
class _HiddenlayerOutputMessage(TypedDict, total=False):
content: ReadOnly[str | Sequence[Mapping[str, str]]]
class _HiddenlayerChoiceMessage(TypedDict, total=False):
content: ReadOnly[str]
class _HiddenlayerChoice(TypedDict, total=False):
message: ReadOnly[_HiddenlayerChoiceMessage]
def is_saas(host: str) -> bool:
"""Checks whether the connection is to the SaaS platform"""
@ -155,7 +176,10 @@ class HiddenlayerGuardrail(CustomGuardrail):
# from the logger object on the response from the model.
headers = request_data.get("proxy_server_request", {}).get("headers", {})
if not headers and logging_obj and logging_obj.model_call_details:
headers = logging_obj.model_call_details.get("litellm_params", {}).get("metadata", {}).get("headers", {})
logged_litellm_params: Final[_LoggedCallLitellmParams] = logging_obj.model_call_details.get(
"litellm_params", {}
)
headers = logged_litellm_params.get("metadata", {}).get("headers", {})
hl_request_metadata["requester_id"] = headers.get("hl-requester-id") or "LiteLLM"
project_id: Final = headers.get("hl-project-id")
@ -408,7 +432,8 @@ class HiddenlayerGuardrailV2(CustomGuardrail):
if input_type == "request":
inputs["structured_messages"] = output
for message in output.get("messages", []):
modified_messages: Final[Sequence[_HiddenlayerOutputMessage]] = output.get("messages", [])
for message in modified_messages:
content = message.get("content", "")
if isinstance(content, list):
text_parts = [
@ -422,7 +447,8 @@ class HiddenlayerGuardrailV2(CustomGuardrail):
inputs["texts"] = new_texts
elif input_type == "response" and inputs.get("texts"):
inputs["texts"] = [output.get("choices", [{}])[-1].get("message", {}).get("content", "")]
redacted_choices: Final[Sequence[_HiddenlayerChoice]] = output.get("choices", [{}])
inputs["texts"] = [redacted_choices[-1].get("message", {}).get("content", "")]
elif input_type == "response" and inputs.get("tool_calls"):
inputs["tool_calls"] = output

83
litellm/proxy/guardrails/guardrail_hooks/llm_as_a_judge/__init__.py
View file

@ -1,16 +1,20 @@
"""LLM-as-a-Judge guardrail: uses an LLM to score responses against weighted criteria."""
import json
import re
from collections.abc import Callable
from datetime import datetime
from typing import TYPE_CHECKING, Any, Final, Literal, Optional, cast
from typing import TYPE_CHECKING, Any, Final, Literal, Optional
from fastapi import HTTPException
import litellm
from litellm._logging import verbose_logger
from litellm.integrations.custom_guardrail import CustomGuardrail
from litellm.litellm_core_utils.llm_judge import (
default_router_provider,
extract_text_from_content,
judge_acompletion,
parse_json_verdict,
)
from litellm.types.guardrails import GuardrailEventHooks, SupportedGuardrailIntegrations
from litellm.types.utils import GenericGuardrailAPIInputs, GuardrailStatus
@ -32,50 +36,9 @@ Return ONLY valid JSON in this exact format:
_VALID_ON_FAILURE: Final = frozenset({"block", "log"})
def _default_router_provider() -> "Router | None":
try:
from litellm.proxy.proxy_server import llm_router
except ImportError:
return None
return llm_router
_JSON_FENCE_RE: Final = re.compile(r"```(?:json)?\s*(.*?)\s*```", re.DOTALL | re.IGNORECASE)
def _parse_judge_verdict(raw: str) -> dict[str, Any]:
"""Parse the judge's JSON verdict, tolerating markdown fences and surrounding prose."""
text = raw.strip()
fenced: Final = _JSON_FENCE_RE.search(text)
if fenced is not None:
text = fenced.group(1).strip()
parsed: object
try:
parsed = json.loads(text)
except json.JSONDecodeError:
start: Final = text.find("{")
end: Final = text.rfind("}")
if start == -1 or end <= start:
raise
parsed = json.loads(text[start : end + 1])
if not isinstance(parsed, dict):
raise ValueError("judge response is not a JSON object")
return cast(dict[str, Any], parsed) # cast-ok: narrowed to dict by the isinstance guard above
def _extract_text_from_content(content: Any) -> str:
"""Return plain text from a message content field (str or multimodal list)."""
if isinstance(content, str):
return content
if isinstance(content, list):
parts: Final = []
for part in content:
if isinstance(part, dict) and part.get("type") == "text":
parts.append(part.get("text", ""))
return " ".join(parts)
return ""
_default_router_provider: Final = default_router_provider
_parse_judge_verdict: Final = parse_json_verdict
_extract_text_from_content: Final = extract_text_from_content
def _get_litellm_param(
@ -168,25 +131,13 @@ class LLMAsAJudgeGuardrail(CustomGuardrail):
"content": _build_judge_prompt(self.criteria, messages, response_text),
},
]
router: Final = self._router_provider()
if router is not None and (
self.judge_model in router.model_group_alias or router.get_model_list(model_name=self.judge_model)
):
response = await router.acompletion(
model=self.judge_model,
messages=judge_messages,
response_format={"type": "json_object"},
temperature=0,
num_retries=0,
fallbacks=[],
)
else:
response = await litellm.acompletion(
model=self.judge_model,
messages=judge_messages,
response_format={"type": "json_object"},
temperature=0,
)
response: Final = await judge_acompletion(
self._router_provider(),
self.judge_model,
judge_messages,
response_format={"type": "json_object"},
temperature=0,
)
raw: Final = response.choices[0].message.content or "{}"
return _parse_judge_verdict(raw)

74
litellm/proxy/guardrails/guardrail_hooks/tool_permission.py
View file

@ -17,6 +17,11 @@ from litellm.proxy._types import UserAPIKeyAuth
from litellm.proxy.common_utils.callback_utils import (
add_guardrail_to_applied_guardrails_header,
)
from litellm.proxy.guardrails.anthropic_sse import (
anthropic_sse_chunks_from_response,
assemble_anthropic_sse_stream,
is_raw_sse_stream,
)
from litellm.types.guardrails import GuardrailEventHooks, LitellmParams
from litellm.types.proxy.guardrails.guardrail_hooks.tool_permission import (
PermissionError,
@ -870,7 +875,7 @@ class ToolPermissionGuardrail(CustomGuardrail):
all_chunks.append(chunk)
assembled_model_response: Final[ModelResponse | TextCompletionResponse | None] = (
stream_chunk_builder(chunks=all_chunks) if not self._is_raw_sse_stream(all_chunks) else None
stream_chunk_builder(chunks=all_chunks) if not is_raw_sse_stream(all_chunks) else None
)
if isinstance(assembled_model_response, ModelResponse):
denied_tools = self._check_assembled_stream(assembled_model_response)
@ -883,9 +888,9 @@ class ToolPermissionGuardrail(CustomGuardrail):
yield chunk
return
anthropic_response: Final = self._assemble_anthropic_stream(all_chunks)
anthropic_response: Final = assemble_anthropic_sse_stream(all_chunks)
if anthropic_response is None:
if self._is_raw_sse_stream(all_chunks):
if is_raw_sse_stream(all_chunks):
raise GuardrailRaisedException(
guardrail_name=self.guardrail_name,
message=(
@ -904,13 +909,9 @@ class ToolPermissionGuardrail(CustomGuardrail):
return
self._modify_response_with_permission_errors(anthropic_response, anthropic_denials)
for sse_chunk in self._rewritten_anthropic_sse_chunks(anthropic_response):
for sse_chunk in anthropic_sse_chunks_from_response(anthropic_response):
yield sse_chunk
@staticmethod
def _is_raw_sse_stream(all_chunks: Sequence[Any]) -> bool:
return any(isinstance(chunk, (str, bytes)) for chunk in all_chunks)
def _check_assembled_stream(
self, assembled: ModelResponse
) -> tuple[tuple[ChatCompletionMessageToolCall, PermissionError], ...]:
@ -924,60 +925,3 @@ class ToolPermissionGuardrail(CustomGuardrail):
if not denied_tools:
verbose_proxy_logger.debug("Tool Permission Guardrail Post-Call Hook: All tools allowed")
return denied_tools
@staticmethod
def _joined_sse_stream(all_chunks: Sequence[Any]) -> str | None:
raw: Final = b"".join(
chunk if isinstance(chunk, bytes) else chunk.encode("utf-8")
for chunk in all_chunks
if isinstance(chunk, (str, bytes))
)
try:
return raw.decode("utf-8")
except UnicodeDecodeError:
return None
@staticmethod
def _has_anthropic_message_start(sse_stream: str) -> bool:
from litellm.proxy.pass_through_endpoints.llm_provider_handlers.anthropic_passthrough_logging_handler import (
AnthropicPassthroughLoggingHandler,
)
return any(
(event_data := AnthropicPassthroughLoggingHandler._extract_sse_data(event)) is not None # pyright: ignore[reportPrivateUsage] # same parser the assembler uses; a private import beats forking SSE parsing
and event_data.get("type") == "message_start"
for event in AnthropicPassthroughLoggingHandler._split_sse_chunk_into_events(sse_stream) # pyright: ignore[reportPrivateUsage] # same parser the assembler uses
)
@staticmethod
def _assemble_anthropic_stream(all_chunks: Sequence[Any]) -> ModelResponse | None:
from litellm.proxy.pass_through_endpoints.llm_provider_handlers.anthropic_passthrough_logging_handler import (
AnthropicPassthroughLoggingHandler,
)
sse_stream: Final = ToolPermissionGuardrail._joined_sse_stream(all_chunks)
if sse_stream is None or not ToolPermissionGuardrail._has_anthropic_message_start(sse_stream):
return None
try:
assembled = AnthropicPassthroughLoggingHandler._build_complete_streaming_response( # pyright: ignore[reportPrivateUsage] # the only SSE-to-ModelResponse assembler; reimplementing it here would fork the parser
all_chunks=(sse_stream,),
litellm_logging_obj=None, # pyright: ignore[reportArgumentType] # only forwarded to stream_chunk_builder, which accepts None
model="",
)
except (AttributeError, TypeError, ValueError, json.JSONDecodeError):
return None
return assembled if isinstance(assembled, ModelResponse) else None
@staticmethod
def _rewritten_anthropic_sse_chunks(assembled: ModelResponse) -> tuple[bytes, ...]:
from litellm.llms.anthropic.experimental_pass_through.adapters.transformation import (
LiteLLMAnthropicMessagesAdapter,
)
from litellm.llms.anthropic.experimental_pass_through.messages.fake_stream_iterator import (
FakeAnthropicMessagesStreamIterator,
)
anthropic_response: Final = LiteLLMAnthropicMessagesAdapter().translate_openai_response_to_anthropic(
response=assembled
)
return tuple(FakeAnthropicMessagesStreamIterator(response=anthropic_response).chunks)

38
litellm/proxy/guardrails/guardrail_registry.py
View file

@ -2,9 +2,10 @@
import importlib
import os
from collections.abc import Callable, Iterator, Mapping
from datetime import datetime, timezone
from itertools import chain, count
from typing import Any, Final, Literal, Optional, cast
from typing import Any, Final, Literal, Optional, Protocol, cast
from pydantic import ValidationError
@ -59,6 +60,13 @@ from .guardrail_initializers import (
initialize_tool_permission,
)
class _GuardrailRowLike(Protocol):
@property
def guardrail_id(self) -> str: ...
def __iter__(self) -> Iterator[tuple[str, object]]: ...
guardrail_initializer_registry: Final = {
SupportedGuardrailIntegrations.BEDROCK.value: initialize_bedrock,
SupportedGuardrailIntegrations.LAKERA.value: initialize_lakera,
@ -125,7 +133,9 @@ def get_guardrail_initializer_from_hooks():
# Check for guardrail_initializer_registry dictionary
if hasattr(module, "guardrail_initializer_registry"):
registry = getattr(module, "guardrail_initializer_registry")
registry: Mapping[str, Callable[..., CustomGuardrail]] | None = getattr(
module, "guardrail_initializer_registry", None
)
if isinstance(registry, dict):
discovered_initializers.update(registry)
verbose_proxy_logger.debug(
@ -135,7 +145,7 @@ def get_guardrail_initializer_from_hooks():
# Check for standalone initialize_guardrail function (fallback for directory-based guardrails)
elif hasattr(module, "initialize_guardrail"):
# For directories with just initialize_guardrail, use the directory name as the key
initialize_fn = getattr(module, "initialize_guardrail")
initialize_fn: Callable[..., CustomGuardrail] | None = getattr(module, "initialize_guardrail", None)
discovered_initializers[item] = initialize_fn
verbose_proxy_logger.debug("Found initialize_guardrail function in %s", module_path)
@ -206,7 +216,9 @@ def get_guardrail_class_from_hooks():
# Check for guardrail_initializer_registry dictionary
if hasattr(module, "guardrail_class_registry"):
registry = getattr(module, "guardrail_class_registry")
registry: Mapping[str, type[CustomGuardrail]] | None = getattr(
module, "guardrail_class_registry", None
)
if isinstance(registry, dict):
discovered_classes.update(registry)
@ -275,7 +287,7 @@ class GuardrailRegistry:
guardrail_info: Final[str] = safe_dumps(guardrail.get("guardrail_info", {}))
# Create guardrail in DB
created_guardrail: Final = await GuardrailsRepository(prisma_client).table.create(
created_guardrail: Final[_GuardrailRowLike] = await GuardrailsRepository(prisma_client).table.create(
data={
"guardrail_name": guardrail_name,
"litellm_params": litellm_params,
@ -321,7 +333,7 @@ class GuardrailRegistry:
guardrail_info: Final[str] = safe_dumps(guardrail.get("guardrail_info", {}))
# Update in DB
updated_guardrail: Final = await GuardrailsRepository(prisma_client).table.update(
updated_guardrail: Final[_GuardrailRowLike] = await GuardrailsRepository(prisma_client).table.update(
where={"guardrail_id": guardrail_id},
data={
"guardrail_name": guardrail_name,
@ -482,7 +494,7 @@ class InMemoryGuardrailHandler:
custom_guardrail_callback = initializer(litellm_params, guardrail)
elif isinstance(guardrail_type, str) and "." in guardrail_type:
custom_guardrail_callback = self.initialize_custom_guardrail(
guardrail=cast(dict, guardrail),
guardrail=guardrail,
guardrail_type=guardrail_type,
litellm_params=litellm_params,
config_file_path=config_file_path,
@ -512,7 +524,7 @@ class InMemoryGuardrailHandler:
"skip_tool_message_in_guardrail are enabled together, which excludes every message from "
"scanning, so no request content would ever be scanned. Remove one of the two."
)
configured_run_in_parallel: Final = getattr(litellm_params, "run_in_parallel", None)
configured_run_in_parallel: Final[bool | None] = getattr(litellm_params, "run_in_parallel", None)
if configured_run_in_parallel is not None:
custom_guardrail_callback.run_in_parallel = bool(configured_run_in_parallel)
@ -532,7 +544,7 @@ class InMemoryGuardrailHandler:
def initialize_custom_guardrail(
self,
guardrail: dict,
guardrail: Guardrail,
guardrail_type: str,
litellm_params: LitellmParams,
config_file_path: str | None = None,
@ -550,7 +562,9 @@ class InMemoryGuardrailHandler:
guardrail_type,
)
_guardrail_class: Final = get_instance_fn(guardrail_type, config_file_path=config_file_path)
_guardrail_class: Final[Callable[..., CustomGuardrail]] = get_instance_fn(
guardrail_type, config_file_path=config_file_path
)
mode: Final = litellm_params.mode
if mode is None:
@ -683,8 +697,8 @@ class InMemoryGuardrailHandler:
@staticmethod
def _normalize_litellm_params_for_comparison(
params: Any | None,
) -> dict[str, Any] | None:
params: LitellmParams | Mapping[str, object] | None,
) -> Mapping[str, object] | None:
"""
Render litellm_params to a canonical dict so an in-memory LitellmParams and
the raw dict loaded from the DB compare equal when they describe the same

8
litellm/proxy/hooks/parallel_request_limiter_v3.py
View file

@ -24,7 +24,7 @@ from typing import (
from litellm import DualCache
from litellm._logging import verbose_proxy_logger
from litellm.constants import DYNAMIC_RATE_LIMIT_ERROR_THRESHOLD_PER_MINUTE
from litellm.constants import DYNAMIC_RATE_LIMIT_ERROR_THRESHOLD_PER_MINUTE, INTERNAL_CALL_ORIGIN_METADATA_KEY
from litellm.integrations.custom_logger import CustomLogger
from litellm.litellm_core_utils.prompt_templates.common_utils import (
get_str_from_messages,
@ -2991,6 +2991,7 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger):
rate_limit_type: Literal["output", "input", "total"],
) -> list[RedisPipelineIncrementOperation]:
"""Build Redis pipeline increment ops for TPM / parallel-request counters."""
from litellm.litellm_core_utils.core_helpers import get_litellm_metadata_from_kwargs
from litellm.proxy.common_utils.callback_utils import (
get_model_group_from_litellm_kwargs,
)
@ -2998,6 +2999,11 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger):
# Get metadata from standard_logging_object - this correctly handles both
# 'metadata' and 'litellm_metadata' fields from litellm_params
standard_logging_object: Final = kwargs.get("standard_logging_object") or {}
request_metadata: Final = get_litellm_metadata_from_kwargs(kwargs)
if request_metadata.get(INTERNAL_CALL_ORIGIN_METADATA_KEY):
# Internal sub-calls bill spend to the caller but are not the caller's
# traffic; charging them here would let background evals eat TPM headroom.
return []
standard_logging_metadata: Final = standard_logging_object.get("metadata") or {}
model_group: Final = get_model_group_from_litellm_kwargs(kwargs)

4
litellm/proxy/litellm_pre_call_utils.py
View file

@ -274,7 +274,7 @@ _UNTRUSTED_METADATA_CONTROL_FIELDS: Final = (
PRE_CALL_EXECUTED_GUARDRAILS_KEY,
)
_UNTRUSTED_REQUEST_HEADER_CONTROL_FIELDS: Final = frozenset(
UNTRUSTED_REQUEST_HEADER_CONTROL_FIELDS: Final = frozenset(
{
"litellm-disable-message-redaction",
}
@ -355,7 +355,7 @@ def _strip_untrusted_request_header_controls(
return
for header_name in list(headers.keys()):
if isinstance(header_name, str) and header_name.lower() in _UNTRUSTED_REQUEST_HEADER_CONTROL_FIELDS:
if isinstance(header_name, str) and header_name.lower() in UNTRUSTED_REQUEST_HEADER_CONTROL_FIELDS:
if allow_client_message_redaction_opt_out:
continue
headers.pop(header_name, None)

20
litellm/proxy/management_endpoints/access_group_endpoints.py
View file

@ -14,11 +14,11 @@ from litellm.proxy.auth.auth_checks import (
_cache_access_object,
_cache_key_object,
_cache_team_object,
_delete_cache_access_object,
_get_team_object_from_cache,
)
from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
from litellm.proxy.db.exception_handler import PrismaDBExceptionHandler
from litellm.proxy.management_helpers.access_group_team_sync import invalidate_access_group_cache
from litellm.proxy.utils import get_prisma_client_or_throw
from litellm.repositories.table_repositories import AccessGroupRepository
from litellm.types.access_group import (
@ -146,22 +146,6 @@ async def _cache_access_group_record(record: _AccessGroupRecord) -> None:
)
async def _invalidate_cache_access_group(access_group_id: str) -> None:
"""
Invalidate (delete) an access group entry from both in-memory and Redis caches.
Uses a lazy import of user_api_key_cache and proxy_logging_obj from proxy_server
to avoid circular imports, following the same pattern as key_management_endpoints.
"""
from litellm.proxy.proxy_server import proxy_logging_obj, user_api_key_cache
await _delete_cache_access_object(
access_group_id=access_group_id,
user_api_key_cache=user_api_key_cache,
proxy_logging_obj=proxy_logging_obj,
)
# ---------------------------------------------------------------------------
# DB sync helpers (called inside a Prisma transaction)
# ---------------------------------------------------------------------------
@ -595,7 +579,7 @@ async def delete_access_group(
from litellm.proxy.proxy_server import proxy_logging_obj, user_api_key_cache
await _invalidate_cache_access_group(access_group_id)
await invalidate_access_group_cache(access_group_id)
await _patch_team_caches_remove_access_group(
affected_team_ids, access_group_id, user_api_key_cache, proxy_logging_obj
)

368
litellm/proxy/management_endpoints/auto_router_endpoints.py
View file

@ -13,6 +13,7 @@ from pydantic import BaseModel, TypeAdapter
from litellm._logging import verbose_proxy_logger
from litellm.exceptions import BudgetExceededError
from litellm.litellm_core_utils.llm_judge import router_resolves_model
from litellm.proxy._types import (
CommonProxyErrors,
LiteLLM_TeamTable,
@ -39,11 +40,16 @@ from litellm.types.management_endpoints.auto_router_endpoints import (
AutoRouterRoutingTestRequest,
AutoRouterRoutingTestResponse,
RequestComplexityRouterConfig,
ShadowEvalJobResponse,
ShadowEvalResult,
ShadowEvalSlice,
StartShadowEvalRequest,
)
if TYPE_CHECKING:
from fastapi import APIRouter, Depends, HTTPException, Query, status
from litellm.proxy.utils import PrismaClient
from litellm.router import Router
else:
try:
@ -388,14 +394,7 @@ async def get_auto_router_benchmarks(
"""
from litellm.proxy.proxy_server import prisma_client
if user_api_key_dict.user_role not in (
LitellmUserRoles.PROXY_ADMIN,
LitellmUserRoles.PROXY_ADMIN_VIEW_ONLY,
):
raise HTTPException(
status_code=403,
detail="Only proxy admin roles can view auto-router benchmarks across the deployment",
)
_require_admin_viewer(user_api_key_dict, "view auto-router benchmarks across the deployment")
if prisma_client is None:
raise HTTPException(status_code=500, detail=CommonProxyErrors.db_not_connected_error.value)
@ -430,3 +429,356 @@ async def get_auto_router_benchmarks(
totals=_benchmark_totals(_summed_agg_row(rows)),
groups=groups,
)
# ---------------------------------------------------------------------------
# Shadow eval: pre-adoption evaluation of an auto-router against live traffic.
# The job row is immutable config plus stopped_at; status, counts, spend, and errors
# are derived from the append-only attempt rows, so reads here are aggregations
# bounded by each job's max_turns through the attempt table's job_id index.
# ---------------------------------------------------------------------------
def _require_admin_viewer(user_api_key_dict: UserAPIKeyAuth, action: str) -> None:
if user_api_key_dict.user_role not in (
LitellmUserRoles.PROXY_ADMIN,
LitellmUserRoles.PROXY_ADMIN_VIEW_ONLY,
):
raise HTTPException(status_code=403, detail=f"Only proxy admin roles can {action}")
def _require_admin_writer(user_api_key_dict: UserAPIKeyAuth, action: str) -> None:
if user_api_key_dict.user_role != LitellmUserRoles.PROXY_ADMIN:
raise HTTPException(status_code=403, detail=f"Only a proxy admin can {action}")
def _is_configured_pre_routing_strategy(llm_router: "Router", router_name: str) -> bool:
return any(
router_name in registry
for registry in (
llm_router.auto_routers,
llm_router.complexity_routers,
llm_router.adaptive_routers,
llm_router.quality_routers,
)
)
def _validate_plain_model(llm_router: "Router | None", model: str, field_name: str) -> None:
"""Reject a model the dispatch path cannot resolve, at start rather than as a silently
growing error count once the job is already sampling and billing. Both the judge and a
reverse job's baseline must be plain models: an auto-router in either slot would
re-route per turn, so the comparison would have no fixed arm to attribute results to."""
if llm_router is not None and _is_configured_pre_routing_strategy(llm_router, model):
raise HTTPException(
status_code=400,
detail=f"{field_name} '{model}' is an auto-router; it must be a plain model",
)
if router_resolves_model(llm_router, model):
return
import litellm
try:
litellm.get_llm_provider(model=model)
except Exception as e:
raise HTTPException(
status_code=400,
detail=(
f"{field_name} '{model}' is neither a model configured on this proxy nor a "
"provider-qualified public model name (e.g. 'anthropic/claude-sonnet-5')"
),
) from e
def _is_unique_violation(error: Exception) -> bool:
"""Whether a Prisma create failed on a unique index. One active job per key and
direction lives in a partial unique index (raw SQL in the migration; schema.prisma
cannot express partial indexes), so the read-then-create check above it is advisory:
two concurrent starts pass the read, and the loser must surface as the same 409
rather than a 500."""
try:
from prisma.errors import UniqueViolationError
except ImportError:
return "unique constraint" in str(error).lower() or "P2002" in str(error)
return isinstance(error, UniqueViolationError)
class _AttemptAggRow(BaseModel):
grp: str
turn_count: int
real_wins: int
shadow_wins: int
ties: int
avg_confidence: float | None
_ATTEMPT_AGG_ROWS: Final = TypeAdapter(list[_AttemptAggRow])
_ATTEMPT_AGG_SELECT: Final = """
COUNT(*)::int AS turn_count,
COUNT(*) FILTER (WHERE outcome = 'real')::int AS real_wins,
COUNT(*) FILTER (WHERE outcome = 'shadow')::int AS shadow_wins,
COUNT(*) FILTER (WHERE outcome = 'tie')::int AS ties,
AVG(confidence)::float AS avg_confidence
FROM "LiteLLM_ShadowEvalAttempt"
WHERE job_id = $1 AND outcome != 'error'
GROUP BY 1
"""
_ATTEMPT_AGG_BY_TIER_SQL: Final = "SELECT COALESCE(tier, 'UNCLASSIFIED') AS grp," + _ATTEMPT_AGG_SELECT
_ATTEMPT_AGG_BY_MODEL_SQL: Final = "SELECT COALESCE(real_model, 'unknown') AS grp," + _ATTEMPT_AGG_SELECT
_SWEEP_FINISHED_JOBS_SQL: Final = """
UPDATE "LiteLLM_ShadowEvalJob" j SET stopped_at = NOW()
WHERE j.api_key_id = $1 AND j.stopped_at IS NULL
AND (
j.ends_at <= NOW()
OR (SELECT COUNT(*) FROM "LiteLLM_ShadowEvalAttempt" a WHERE a.job_id = j.id) >= j.max_turns
)
"""
_ATTEMPT_TOTALS_SQL: Final = """
SELECT
COUNT(*) FILTER (WHERE outcome != 'error')::int AS judged_count,
COUNT(*) FILTER (WHERE outcome = 'error')::int AS error_count,
COALESCE(SUM(judge_cost), 0)::float AS judge_spend
FROM "LiteLLM_ShadowEvalAttempt"
WHERE job_id = $1
"""
class _AttemptTotalsRow(BaseModel):
judged_count: int
error_count: int
judge_spend: float
_ATTEMPT_TOTALS_ROWS: Final = TypeAdapter(list[_AttemptTotalsRow])
def _pct_of(numerator: int, denominator: int) -> float:
return _pct(numerator, denominator)
def _slices(rows: Sequence[_AttemptAggRow]) -> tuple[ShadowEvalSlice, ...]:
return tuple(
ShadowEvalSlice(
group=row.grp,
turn_count=row.turn_count,
real_win_rate_pct=_pct_of(row.real_wins, row.turn_count),
shadow_win_rate_pct=_pct_of(row.shadow_wins, row.turn_count),
tie_rate_pct=_pct_of(row.ties, row.turn_count),
avg_judge_confidence=round(row.avg_confidence or 0.0, 3),
)
for row in sorted(rows, key=lambda r: r.turn_count, reverse=True)
)
async def _shadow_eval_results(prisma_client: "PrismaClient", job_id: str) -> ShadowEvalResult | None:
"""Both stratifications of one job's verdicts. Tier answers "where does the router do
well"; the model stratification groups by whichever model served the real arm, so it
answers "which of the models this key uses today would the router beat" forward, and
"for the turns the router sent to X, did X beat the baseline" in reverse. Reads are
bounded by the job's own attempts (<= max_turns) via the job_id index."""
by_tier: Final = _ATTEMPT_AGG_ROWS.validate_python(
await prisma_client.db.query_raw(_ATTEMPT_AGG_BY_TIER_SQL, job_id) or ()
)
if not by_tier:
return None
by_model: Final = _ATTEMPT_AGG_ROWS.validate_python(
await prisma_client.db.query_raw(_ATTEMPT_AGG_BY_MODEL_SQL, job_id) or ()
)
total_turns: Final = sum(r.turn_count for r in by_tier)
return ShadowEvalResult(
by_tier=_slices(by_tier),
by_current_model=_slices(by_model),
overall_shadow_win_rate_pct=_pct_of(sum(r.shadow_wins for r in by_tier), total_turns),
overall_tie_rate_pct=_pct_of(sum(r.ties for r in by_tier), total_turns),
)
@router.post(
"/auto_router/shadow_eval/start",
tags=("auto router",),
dependencies=(Depends(user_api_key_auth),),
response_model=ShadowEvalJobResponse,
status_code=status.HTTP_201_CREATED,
)
async def start_shadow_eval(
data: StartShadowEvalRequest,
user_api_key_dict: Annotated[UserAPIKeyAuth, Depends(user_api_key_auth)],
) -> ShadowEvalJobResponse:
"""
Start a shadow eval: duplicate a sampled slice of a key's live traffic against a second
arm, judge the two responses blind, and stratify win rates by tier and by the model that
served the real arm.
A forward job answers whether the key should adopt router_name: it samples the requests
the router did not serve and duplicates them through it. A reverse job answers whether a
key already on the router still gains from it: it samples the requests the router did
serve and duplicates them against baseline_model. A key can hold one active job per
direction, so both questions can run at once.
Shadow responses are never served to users. The job samples until it has judged
max_turns turns, reaches the end of its window, or is stopped; sampling changes
propagate to pods within about 10 seconds. Shadow and judge calls bill to the
shadowed key but are excluded from request counts and auto-router adoption metrics.
"""
from litellm.proxy.proxy_server import llm_router, prisma_client
_require_admin_writer(user_api_key_dict, "start a shadow eval")
if prisma_client is None:
raise HTTPException(status_code=500, detail=CommonProxyErrors.db_not_connected_error.value)
if llm_router is None or not _is_configured_pre_routing_strategy(llm_router, data.router_name):
raise HTTPException(status_code=400, detail=f"'{data.router_name}' is not a configured auto-router")
_validate_plain_model(llm_router, data.judge_model, "judge_model")
if data.baseline_model is not None:
_validate_plain_model(llm_router, data.baseline_model, "baseline_model")
key_row: Final = await prisma_client.db.litellm_verificationtoken.find_unique(
where={"token": data.api_key_id} # mutable-ok: Prisma filter
)
if key_row is None:
raise HTTPException(
status_code=400,
detail=(
f"api_key_id '{data.api_key_id}' is not a key on this proxy; pass the key's token hash, "
"the value the key list and key info endpoints report"
),
)
# A job that expired or exhausted its turn budget stopped sampling on its own, but
# still holds its slot in the per-key, per-direction partial unique index until
# stamped; free it so a new eval can start. Sweeping both directions is deliberate.
await prisma_client.db.execute_raw(_SWEEP_FINISHED_JOBS_SQL, data.api_key_id)
active: Final = await prisma_client.db.litellm_shadowevaljob.find_first(
where={ # mutable-ok: Prisma filter
"api_key_id": data.api_key_id,
"direction": data.direction,
"stopped_at": None,
},
)
if active is not None:
raise HTTPException(
status_code=409,
detail=f"Key already has an active {data.direction} shadow eval job ({active.id}). Stop it first.",
)
now: Final = datetime.now(timezone.utc)
try:
job: Final = await prisma_client.db.litellm_shadowevaljob.create(
data={ # mutable-ok: Prisma payload
"api_key_id": data.api_key_id,
"router_name": data.router_name,
"direction": data.direction,
"baseline_model": data.baseline_model,
"judge_model": data.judge_model,
"shadow_percentage": data.shadow_percentage,
"max_turns": data.max_turns,
"created_by": user_api_key_dict.user_id,
"ends_at": now + timedelta(days=data.duration_days),
}
)
except Exception as e:
if not _is_unique_violation(e):
raise
raise HTTPException(
status_code=409,
detail=(
f"Key already has an active {data.direction} shadow eval job (started concurrently). Stop it first."
),
) from e
return ShadowEvalJobResponse.model_validate(job, from_attributes=True)
@router.get(
"/auto_router/shadow_eval",
tags=("auto router",),
dependencies=(Depends(user_api_key_auth),),
response_model=list[ShadowEvalJobResponse],
)
async def list_shadow_eval_jobs(
user_api_key_dict: Annotated[UserAPIKeyAuth, Depends(user_api_key_auth)],
api_key_id: Annotated[str | None, Query(description="Filter to jobs shadowing this key")] = None,
limit: Annotated[int, Query(ge=1, le=200, description="Newest jobs to return")] = 50,
) -> tuple[ShadowEvalJobResponse, ...]:
"""List shadow eval jobs, newest first. Counts and results ride the detail endpoint only."""
from litellm.proxy.proxy_server import prisma_client
_require_admin_viewer(user_api_key_dict, "view shadow evals")
if prisma_client is None:
raise HTTPException(status_code=500, detail=CommonProxyErrors.db_not_connected_error.value)
records: Final = await prisma_client.db.litellm_shadowevaljob.find_many(
where={"api_key_id": api_key_id} if api_key_id else {}, # mutable-ok: Prisma filter
order={"created_at": "desc"}, # mutable-ok: Prisma order
take=limit,
)
return tuple(ShadowEvalJobResponse.model_validate(record, from_attributes=True) for record in records or ())
@router.get(
"/auto_router/shadow_eval/{job_id}",
tags=("auto router",),
dependencies=(Depends(user_api_key_auth),),
response_model=ShadowEvalJobResponse,
)
async def get_shadow_eval_job(
job_id: str,
user_api_key_dict: Annotated[UserAPIKeyAuth, Depends(user_api_key_auth)],
) -> ShadowEvalJobResponse:
"""One job with derived counts, judge spend, latest error, and stratified results."""
from litellm.proxy.proxy_server import prisma_client
_require_admin_viewer(user_api_key_dict, "view shadow evals")
if prisma_client is None:
raise HTTPException(status_code=500, detail=CommonProxyErrors.db_not_connected_error.value)
record: Final = await prisma_client.db.litellm_shadowevaljob.find_unique(
where={"id": job_id} # mutable-ok: Prisma filter
)
if record is None:
raise HTTPException(status_code=404, detail=f"No shadow eval job {job_id}")
totals: Final = _ATTEMPT_TOTALS_ROWS.validate_python(
await prisma_client.db.query_raw(_ATTEMPT_TOTALS_SQL, job_id) or ()
)
latest_error: Final = await prisma_client.db.litellm_shadowevalattempt.find_first(
where={"job_id": job_id, "outcome": "error"}, # mutable-ok: Prisma filter
order={"created_at": "desc"}, # mutable-ok: Prisma order
)
return ShadowEvalJobResponse.model_validate(record, from_attributes=True).model_copy(
update={ # mutable-ok: pydantic update payload
"judged_count": totals[0].judged_count if totals else 0,
"error_count": totals[0].error_count if totals else 0,
"judge_spend": round(totals[0].judge_spend, 6) if totals else 0.0,
"last_error": latest_error.error if latest_error else None,
"results": await _shadow_eval_results(prisma_client, job_id),
}
)
@router.post(
"/auto_router/shadow_eval/{job_id}/stop",
tags=("auto router",),
dependencies=(Depends(user_api_key_auth),),
response_model=ShadowEvalJobResponse,
)
async def stop_shadow_eval_job(
job_id: str,
user_api_key_dict: Annotated[UserAPIKeyAuth, Depends(user_api_key_auth)],
) -> ShadowEvalJobResponse:
"""Stop an active shadow eval job. Attempts are kept; sampling halts within ~10s."""
from litellm.proxy.proxy_server import prisma_client
_require_admin_writer(user_api_key_dict, "stop a shadow eval")
if prisma_client is None:
raise HTTPException(status_code=500, detail=CommonProxyErrors.db_not_connected_error.value)
record: Final = await prisma_client.db.litellm_shadowevaljob.find_unique(
where={"id": job_id} # mutable-ok: Prisma filter
)
if record is None:
raise HTTPException(status_code=404, detail=f"No shadow eval job {job_id}")
current: Final = ShadowEvalJobResponse.model_validate(record, from_attributes=True)
if current.status != "running":
raise HTTPException(status_code=400, detail=f"Job {job_id} is already {current.status}")
updated: Final = await prisma_client.db.litellm_shadowevaljob.update(
where={"id": job_id}, # mutable-ok: Prisma filter
data={"stopped_at": datetime.now(timezone.utc)}, # mutable-ok: Prisma payload
)
return ShadowEvalJobResponse.model_validate(updated, from_attributes=True)

118
litellm/proxy/management_endpoints/customer_endpoints.py
View file

@ -10,12 +10,19 @@ All /customer management endpoints
"""
#### END-USER/CUSTOMER MANAGEMENT ####
from collections.abc import Mapping, Sequence
from datetime import datetime, timedelta
from typing import Final
from typing import TYPE_CHECKING, Final, Protocol, TypeVar, overload
import fastapi
from fastapi import APIRouter, Depends, HTTPException, Request
from pydantic import BaseModel
from pydantic import BaseModel, TypeAdapter
if TYPE_CHECKING:
from prisma.models import LiteLLM_BudgetTable as PrismaBudgetRow
from prisma.models import LiteLLM_EndUserTable as PrismaEndUserRow
from litellm.proxy.utils import PrismaClient
import litellm
from litellm._logging import verbose_proxy_logger
@ -41,6 +48,54 @@ from litellm.types.proxy.management_endpoints.customer_endpoints import (
UnblockUsersResponse,
)
_RowT_co: Final = TypeVar("_RowT_co", covariant=True)
_STR_OBJECT_DICT: Final = TypeAdapter(dict[str, object])
if TYPE_CHECKING:
class _TableOps(Protocol[_RowT_co]):
async def find_first(
self,
where: Mapping[str, object] | None = None,
include: Mapping[str, bool] | None = None,
) -> _RowT_co | None: ...
async def find_many(
self,
where: Mapping[str, object] | None = None,
include: Mapping[str, bool] | None = None,
) -> Sequence[_RowT_co]: ...
async def create(
self,
data: Mapping[str, object],
include: Mapping[str, bool] | None = None,
) -> _RowT_co: ...
async def update(
self,
where: Mapping[str, object],
data: Mapping[str, object],
include: Mapping[str, bool] | None = None,
) -> _RowT_co | None: ...
async def upsert(
self,
where: Mapping[str, object],
data: Mapping[str, Mapping[str, object]],
) -> _RowT_co: ...
async def delete_many(self, where: Mapping[str, object]) -> int: ...
@overload
def _typed_table(repo: EndUserRepository) -> "_TableOps[PrismaEndUserRow]": ...
@overload
def _typed_table(repo: BudgetRepository) -> "_TableOps[PrismaBudgetRow]": ...
def _typed_table(repo: EndUserRepository | BudgetRepository) -> object:
return repo.table
router: Final = APIRouter()
@ -89,7 +144,7 @@ async def block_user(data: BlockUsers):
records: Final = []
if prisma_client is not None:
for id in data.user_ids:
record = await EndUserRepository(prisma_client).table.upsert(
record = await _typed_table(EndUserRepository(prisma_client)).upsert(
where={"user_id": id},
data={
"create": {"user_id": id, "blocked": True},
@ -184,7 +239,7 @@ def new_budget_request(data: NewCustomerRequest) -> BudgetNewRequest | None:
budget_kv_pairs[field_name] = value
if budget_kv_pairs:
budget_request: Final = BudgetNewRequest(**budget_kv_pairs)
budget_request: Final = BudgetNewRequest.model_validate(budget_kv_pairs)
validate_budget_duration(budget_request.budget_duration)
if budget_request.budget_reset_at is None and budget_request.budget_duration is not None:
budget_request.budget_reset_at = datetime.utcnow() + timedelta(
@ -195,10 +250,10 @@ def new_budget_request(data: NewCustomerRequest) -> BudgetNewRequest | None:
async def _handle_customer_object_permission_update(
non_default_values: dict,
non_default_values: dict[str, object],
end_user_table_data_typed: LiteLLM_EndUserTable | None,
update_end_user_table_data: dict,
prisma_client,
update_end_user_table_data: dict[str, object],
prisma_client: "PrismaClient",
) -> None:
"""
Handle object permission updates for customer endpoints.
@ -344,13 +399,13 @@ async def new_end_user(
},
)
new_end_user_obj: dict = {}
new_end_user_obj: dict[str, object] = {}
## CREATE BUDGET ## if set
_new_budget: Final = new_budget_request(data)
if _new_budget is not None:
try:
budget_record: Final = await BudgetRepository(prisma_client).table.create(
budget_record: Final = await _typed_table(BudgetRepository(prisma_client)).create(
data={
**_new_budget.model_dump(exclude_unset=True),
"created_by": user_api_key_dict.user_id or litellm_proxy_admin_name,
@ -364,16 +419,18 @@ async def new_end_user(
elif data.budget_id is not None:
new_end_user_obj["budget_id"] = data.budget_id
_user_data: Final = data.dict(exclude_none=True)
_user_data: Final = _STR_OBJECT_DICT.validate_python(data.dict(exclude_none=True))
for k, v in _user_data.items():
if k not in BudgetNewRequest.model_fields:
new_end_user_obj[k] = v
## Handle Object Permission - MCP Servers, Vector Stores etc.
new_end_user_obj = await _set_object_permission(
data_json=new_end_user_obj,
prisma_client=prisma_client,
new_end_user_obj = _STR_OBJECT_DICT.validate_python(
await _set_object_permission(
data_json=new_end_user_obj,
prisma_client=prisma_client,
)
)
# Ensure object_permission is not in the data being sent to create
@ -386,7 +443,7 @@ async def new_end_user(
new_end_user_obj.pop("object_permission", None)
## WRITE TO DB ##
end_user_record: Final = await EndUserRepository(prisma_client).table.create(
end_user_record: Final = await _typed_table(EndUserRepository(prisma_client)).create(
data=new_end_user_obj,
include={"litellm_budget_table": True, "object_permission": True},
)
@ -442,7 +499,7 @@ async def end_user_info(
detail={"error": CommonProxyErrors.db_not_connected_error.value},
)
user_info: Final = await EndUserRepository(prisma_client).table.find_first(
user_info: Final = await _typed_table(EndUserRepository(prisma_client)).find_first(
where={"user_id": end_user_id},
include={"litellm_budget_table": True, "object_permission": True},
)
@ -535,13 +592,13 @@ async def update_end_user(
from litellm.proxy.proxy_server import litellm_proxy_admin_name, prisma_client
try:
data_json: Final[dict] = data.json()
data_json: Final = _STR_OBJECT_DICT.validate_python(data.json())
# get the row from db
if prisma_client is None:
raise Exception("Not connected to DB!")
# get non default values for key
non_default_values: Final = {}
non_default_values: Final = dict[str, object]()
for k, v in data_json.items():
if v is not None and v not in (
[],
@ -551,7 +608,7 @@ async def update_end_user(
non_default_values[k] = v
## Get end user table data ##
end_user_table_data: Final = await EndUserRepository(prisma_client).table.find_first(
end_user_table_data: Final = await _typed_table(EndUserRepository(prisma_client)).find_first(
where={"user_id": data.user_id}, include={"litellm_budget_table": True}
)
@ -563,14 +620,14 @@ async def update_end_user(
param="user_id",
)
end_user_table_data_typed: Final = LiteLLM_EndUserTable(**end_user_table_data.model_dump())
end_user_table_data_typed: Final = LiteLLM_EndUserTable.model_validate(end_user_table_data.model_dump())
## Get budget table data ##
end_user_budget_table: Final = end_user_table_data_typed.litellm_budget_table
## Get all params for budget table ##
budget_table_data: Final = {}
update_end_user_table_data: Final = {}
budget_table_data: Final = dict[str, object]()
update_end_user_table_data: Final = dict[str, object]()
for k, v in non_default_values.items():
# budget_id is for linking to existing budget, not for creating new budget
if k == "budget_id":
@ -593,7 +650,7 @@ async def update_end_user(
if budget_table_data:
if end_user_budget_table is None:
## Create new budget ##
budget_table_data_record = await BudgetRepository(prisma_client).table.create(
budget_table_data_record = await _typed_table(BudgetRepository(prisma_client)).create(
data={
**budget_table_data,
"created_by": user_api_key_dict.user_id or litellm_proxy_admin_name,
@ -605,7 +662,7 @@ async def update_end_user(
update_end_user_table_data["budget_id"] = budget_table_data_record.budget_id
else:
## Update existing budget ##
budget_table_data_record = await BudgetRepository(prisma_client).table.update(
budget_table_data_record = await _typed_table(BudgetRepository(prisma_client)).update(
where={"budget_id": end_user_budget_table.budget_id},
data=budget_table_data,
)
@ -625,7 +682,7 @@ async def update_end_user(
if data.user_id is not None and len(data.user_id) > 0:
update_end_user_table_data["user_id"] = data.user_id
verbose_proxy_logger.debug("In update customer, user_id condition block.")
response: Final = await EndUserRepository(prisma_client).table.update(
response: Final = await _typed_table(EndUserRepository(prisma_client)).update(
where={"user_id": data.user_id},
data=update_end_user_table_data,
include={"litellm_budget_table": True, "object_permission": True},
@ -688,7 +745,7 @@ async def delete_end_user(
verbose_proxy_logger.debug("/customer/delete: Received data = %s", data)
if data.user_ids is not None and isinstance(data.user_ids, list) and len(data.user_ids) > 0:
# First check if all users exist
existing_users: Final = await EndUserRepository(prisma_client).table.find_many(
existing_users: Final = await _typed_table(EndUserRepository(prisma_client)).find_many(
where={"user_id": {"in": data.user_ids}}
)
existing_user_ids: Final = {user.user_id for user in existing_users}
@ -703,7 +760,7 @@ async def delete_end_user(
)
# All users exist, proceed with deletion
response: Final = await EndUserRepository(prisma_client).table.delete_many(
response: Final = await _typed_table(EndUserRepository(prisma_client)).delete_many(
where={"user_id": {"in": data.user_ids}}
)
verbose_proxy_logger.debug("received response from updating prisma client. response=%s", response)
@ -764,7 +821,7 @@ async def list_end_user(
detail={"error": CommonProxyErrors.db_not_connected_error.value},
)
response: Final = await EndUserRepository(prisma_client).table.find_many(
response: Final = await _typed_table(EndUserRepository(prisma_client)).find_many(
include={"litellm_budget_table": True, "object_permission": True}
)
@ -827,11 +884,10 @@ async def get_customer_daily_activity(
exclude_end_user_ids_list = exclude_end_user_ids.split(",") if exclude_end_user_ids else None
# Fetch organization aliases for metadata
where_condition: Final = {}
where_condition: Final = dict[str, object]()
if end_user_ids_list:
where_condition["user_id"] = {"in": list(end_user_ids_list)}
end_user_aliases: Final = await EndUserRepository(prisma_client).table.find_many(where=where_condition)
end_user_alias_metadata: Final = {e.user_id: {"alias": e.alias} for e in end_user_aliases}
end_user_aliases: Final = await _typed_table(EndUserRepository(prisma_client)).find_many(where=where_condition)
# Query daily activity for organizations
return await get_daily_activity(
@ -839,7 +895,7 @@ async def get_customer_daily_activity(
table_name="litellm_dailyenduserspend",
entity_id_field="end_user_id",
entity_id=end_user_ids_list,
entity_metadata_field=end_user_alias_metadata,
entity_metadata_field={e.user_id: {"alias": e.alias} for e in end_user_aliases},
exclude_entity_ids=exclude_end_user_ids_list,
start_date=start_date,
end_date=end_date,

4
litellm/proxy/management_endpoints/internal_user_endpoints.py
View file

@ -2311,7 +2311,7 @@ async def delete_user(
fetch_all_teams = await TeamRepository(prisma_client).table.find_many(where={"team_id": {"in": user_row.teams}})
teams_to_update = []
for team in fetch_all_teams:
is_member_in_team, new_team_members = _cleanup_members_with_roles(
removed_team_members, new_team_members = _cleanup_members_with_roles(
existing_team_row=LiteLLM_TeamTable.model_validate(team.model_dump()),
data=TeamMemberDeleteRequest(
team_id=team.team_id,
@ -2319,7 +2319,7 @@ async def delete_user(
user_email=user_row.user_email,
),
)
if is_member_in_team:
if removed_team_members:
_db_new_team_members: list[dict] = [m.model_dump() for m in new_team_members]
team.members_with_roles = json.dumps(_db_new_team_members)
teams_to_update.append(team)

84
litellm/proxy/management_endpoints/key_management_endpoints.py
View file

@ -88,6 +88,11 @@ from litellm.proxy.management_endpoints.common_utils import (
from litellm.proxy.management_endpoints.model_management_endpoints import (
_add_model_to_db,
)
from litellm.proxy.management_helpers.access_group_key_sync import (
sync_key_access_group_membership,
sync_key_regeneration_access_group_membership,
sync_key_update_access_group_membership,
)
from litellm.proxy.management_helpers.key_settings_audit import with_settings_updated_at
from litellm.proxy.management_helpers.object_permission_utils import (
_set_object_permission,
@ -888,17 +893,24 @@ async def _common_key_generation_helper(
if litellm.default_key_generate_params is not None:
for elem in data:
key, value = elem
if value is None and key in [
"max_budget",
"user_id",
"team_id",
"max_parallel_requests",
"tpm_limit",
"rpm_limit",
"budget_duration",
"duration",
]:
setattr(data, key, litellm.default_key_generate_params.get(key, None))
if (
value is None
and (key != "budget_duration" or key not in data.model_fields_set)
and key
in [
"max_budget",
"user_id",
"team_id",
"max_parallel_requests",
"tpm_limit",
"rpm_limit",
"budget_duration",
"duration",
]
):
default_value = litellm.default_key_generate_params.get(key)
if default_value is not None:
setattr(data, key, default_value)
elif key == "models" and value == []:
setattr(data, key, litellm.default_key_generate_params.get(key, []))
elif key == "metadata" and value == {}:
@ -2340,6 +2352,17 @@ async def _process_single_key_update(
proxy_logging_obj=proxy_logging_obj,
)
# After the key's own cache entry is dropped, so a failure here cannot leave the key
# authenticating against the access groups it just lost.
await sync_key_update_access_group_membership(
prisma_client=prisma_client,
key_token=_hash_token_if_needed(
_resolve_token_to_update(data=update_key_request, existing_key_row=existing_key_row)
),
data=update_key_request,
existing_key_row=existing_key_row,
)
# Trigger async hook
asyncio.create_task(
KeyManagementEventHooks.async_key_updated_hook(
@ -2821,6 +2844,15 @@ async def update_key_fn(
proxy_logging_obj=proxy_logging_obj,
)
# After the key's own cache entry is dropped, so a failure here cannot leave the key
# authenticating against the access groups it just lost.
await sync_key_update_access_group_membership(
prisma_client=prisma_client,
key_token=_hash_token_if_needed(key),
data=data,
existing_key_row=existing_key_row,
)
if data.spend is not None:
from litellm.proxy.proxy_server import spend_counter_cache
@ -3764,7 +3796,7 @@ async def generate_key_helper_fn(
auto_rotate: bool | None = None,
rotation_interval: str | None = None,
router_settings: dict | None = None,
access_group_ids: list | None = None,
access_group_ids: list[str] | None = None,
budget_limits: list | None = None, # multiple concurrent budget windows
):
from litellm.proxy.proxy_server import premium_user, prisma_client
@ -3972,6 +4004,14 @@ async def generate_key_helper_fn(
create_key_response: Final = await prisma_client.insert_data(data=key_data, table_name="key")
key_data["token_id"] = getattr(create_key_response, "token", None)
created_token_hash: Final = getattr(create_key_response, "token", None)
if isinstance(created_token_hash, str):
await sync_key_access_group_membership(
prisma_client=prisma_client,
key_token=created_token_hash,
previous_access_group_ids=None,
updated_access_group_ids=access_group_ids,
)
key_data["litellm_budget_table"] = getattr(create_key_response, "litellm_budget_table", None)
key_data["created_at"] = getattr(create_key_response, "created_at", None)
key_data["updated_at"] = getattr(create_key_response, "updated_at", None)
@ -4189,6 +4229,7 @@ async def delete_verification_tokens(
deleted_tokens = [key.token for key in authorized_keys]
if len(deleted_tokens) != len(tokens):
failed_tokens = [token for token in tokens if token not in deleted_tokens]
else:
raise Exception("DB not connected. prisma_client is None")
except Exception as e:
@ -4204,6 +4245,16 @@ async def delete_verification_tokens(
hashed_token = hash_token(cast(str, key))
user_api_key_cache.delete_cache(hashed_token)
# After credential invalidation, so a failure here can never keep a deleted key alive.
for deleted_key in authorized_keys:
if deleted_key.token is not None:
await sync_key_access_group_membership(
prisma_client=prisma_client,
key_token=deleted_key.token,
previous_access_group_ids=deleted_key.access_group_ids,
updated_access_group_ids=None,
)
return {
"deleted_keys": deleted_tokens,
"failed_tokens": failed_tokens,
@ -4719,6 +4770,15 @@ async def _execute_virtual_key_regeneration(
proxy_logging_obj=proxy_logging_obj,
)
# After credential invalidation, so a failure here can never keep the old key alive.
await sync_key_regeneration_access_group_membership(
prisma_client=prisma_client,
previous_key_token=hashed_api_key,
new_key_token=new_token_hash,
data=data,
existing_key_row=key_in_db,
)
response: Final = GenerateKeyResponse.model_validate(updated_token_dict)
asyncio.create_task(
KeyManagementEventHooks.async_key_rotated_hook(

109
litellm/proxy/management_endpoints/mcp_management_endpoints.py
View file

@ -22,7 +22,7 @@ import os
from collections.abc import Iterable
from dataclasses import dataclass
from datetime import datetime, timedelta, timezone
from typing import Any, Final, Literal
from typing import TYPE_CHECKING, Any, Final, Literal
from fastapi import (
APIRouter,
@ -50,6 +50,7 @@ from litellm.constants import LITELLM_PROXY_ADMIN_NAME
from litellm.proxy._experimental.mcp_server.utils import (
LITELLM_MCP_SERVER_DESCRIPTION,
LITELLM_MCP_SERVER_NAME,
McpServerPayloadLike,
build_env_var_setup_url,
collect_env_var_references,
get_server_prefix,
@ -91,6 +92,9 @@ def does_mcp_server_exist(mcp_server_records: Iterable[Any], mcp_server_id: str)
DEFAULT_MCP_REGISTRY_VERSION: Final = "1.0.0"
if TYPE_CHECKING:
from litellm.proxy.utils import PrismaClient
try:
importlib.import_module("mcp")
except ImportError as e:
@ -114,11 +118,13 @@ if MCP_AVAILABLE:
from litellm.proxy._experimental.mcp_server.db import (
approve_mcp_server,
create_draft_mcp_server,
create_mcp_server,
delete_mcp_server,
delete_user_credential,
delete_user_env_vars,
get_all_mcp_servers_for_user,
get_draft_mcp_server,
get_mcp_server,
get_mcp_servers,
get_mcp_submissions,
@ -196,7 +202,7 @@ if MCP_AVAILABLE:
server: MCPServer
expires_at: datetime
def _validate_mcp_server_name_fields(payload: Any) -> None:
def _validate_mcp_server_name_fields(payload: McpServerPayloadLike) -> None:
candidates: Final[list[tuple[str, str | None]]] = []
server_name: Final = getattr(payload, "server_name", None)
@ -223,7 +229,7 @@ if MCP_AVAILABLE:
detail={"error": error_messages_text},
)
def validate_and_normalize_mcp_server_payload(payload: Any) -> None:
def validate_and_normalize_mcp_server_payload(payload: McpServerPayloadLike) -> None:
_base_validate_and_normalize_mcp_server_payload(payload)
_validate_mcp_server_name_fields(payload)
@ -466,19 +472,68 @@ if MCP_AVAILABLE:
verbose_proxy_logger.debug("Invalid temporary MCP server payload in Redis cache: %s", e)
return None
def _get_prisma_client_or_none() -> "PrismaClient | None":
"""Non-throwing counterpart to ``get_prisma_client_or_throw`` for paths that degrade
gracefully: a proxy configured without a database keeps the in-memory OAuth session."""
from litellm.proxy.proxy_server import prisma_client
return prisma_client
async def _persist_draft_mcp_server(
payload: NewMCPServerRequest,
server_id: str,
created_by: str,
) -> None:
"""Write the draft row that makes the OAuth session resolvable from any worker.
A failure here is raised, not swallowed: without the shared row the flow degrades to
the per-process cache and fails intermittently, which is the defect being fixed.
"""
prisma_client: Final = _get_prisma_client_or_none()
if prisma_client is None:
return
await create_draft_mcp_server(
prisma_client,
payload,
created_by,
ttl_seconds=TEMPORARY_MCP_SERVER_TTL_SECONDS,
server_id=server_id,
)
async def _get_draft_mcp_server_as_mcp_server(server_id: str) -> MCPServer | None:
"""Resolve a database-backed draft, which is the only lookup that works across workers."""
prisma_client: Final = _get_prisma_client_or_none()
if prisma_client is None:
return None
draft: Final = await get_draft_mcp_server(
prisma_client, server_id, ttl_seconds=TEMPORARY_MCP_SERVER_TTL_SECONDS
)
if draft is None:
return None
return await global_mcp_server_manager.build_mcp_server_from_table(draft)
async def get_cached_temporary_mcp_server(
server_id: str,
) -> MCPServer | None:
_prune_expired_temporary_mcp_servers()
entry: Final = _temporary_mcp_servers.get(server_id)
if entry is None:
redis_server: Final = await _get_temporary_mcp_server_from_redis(server_id)
if redis_server is None:
return None
# Intentionally avoid repopulating local cache from Redis to prevent
# extending effective lifetime beyond the remaining Redis TTL.
return redis_server
return entry.server
if entry is not None:
return entry.server
# A miss here means either an expired session or, on a multi-worker or multi-replica
# proxy, that a different process served /session. The draft row is shared, so it
# resolves the second case; the in-memory hit above still serves single-process
# deployments with no database configured.
draft_server: Final = await _get_draft_mcp_server_as_mcp_server(server_id)
if draft_server is not None:
return draft_server
redis_server: Final = await _get_temporary_mcp_server_from_redis(server_id)
if redis_server is None:
return None
# Intentionally avoid repopulating local cache from Redis to prevent
# extending effective lifetime beyond the remaining Redis TTL.
return redis_server
def _redact_mcp_credentials(
mcp_server: LiteLLM_MCPServerTable,
@ -708,12 +763,36 @@ if MCP_AVAILABLE:
payload_dict["credentials"] = inherited_credentials
return NewMCPServerRequest.model_validate(payload_dict)
async def _resolve_session_server_id(payload: NewMCPServerRequest) -> str:
"""Decide the id an OAuth session runs under.
A caller-supplied id is honoured only when it names a server that really exists, which is
the edit form re-authorizing a saved server against its own id. Anything else gets a fresh
id, so two concurrent sessions can never land on one id and silently adopt each other's
URL or client credentials. Without a database there is nothing shared to collide over, so
the supplied id is kept and behaviour is unchanged.
"""
supplied: Final = payload.server_id
if not supplied:
return str(uuid.uuid4())
if global_mcp_server_manager.get_mcp_server_by_id(supplied) is not None:
return supplied
prisma_client: Final = _get_prisma_client_or_none()
if prisma_client is None:
return supplied
# A draft is another session's row, not a saved server, so re-supplying an id this
# endpoint previously handed back must not let a later session adopt its configuration.
existing: Final = await get_mcp_server(prisma_client, supplied)
if existing is None or existing.approval_status == MCPApprovalStatus.draft:
return str(uuid.uuid4())
return supplied
def _build_temporary_mcp_server_record(
payload: NewMCPServerRequest,
created_by: str | None,
server_id: str,
) -> LiteLLM_MCPServerTable:
now: Final = datetime.utcnow()
server_id: Final = payload.server_id or str(uuid.uuid4())
server_name: Final = payload.server_name or payload.alias or server_id
return LiteLLM_MCPServerTable(
server_id=server_id,
@ -1543,6 +1622,7 @@ if MCP_AVAILABLE:
temp_record: Final = _build_temporary_mcp_server_record(
payload_with_credentials,
created_by,
await _resolve_session_server_id(payload_with_credentials),
)
try:
@ -1554,6 +1634,11 @@ if MCP_AVAILABLE:
temporary_server,
ttl_seconds=TEMPORARY_MCP_SERVER_TTL_SECONDS,
)
await _persist_draft_mcp_server(
payload_with_credentials,
temp_record.server_id,
created_by,
)
await _cache_temporary_mcp_server_in_redis(
temporary_server,
ttl_seconds=TEMPORARY_MCP_SERVER_TTL_SECONDS,

186
litellm/proxy/management_endpoints/model_management_endpoints.py
View file

@ -68,6 +68,7 @@ from litellm.repositories.team_repository import TeamRepository
from litellm.router import Router
from litellm.router_strategy.complexity_router import (
DEFAULT_CLASSIFIER_CONTEXT_WINDOW_SIZE,
ClassificationRubric,
ComplexityRouterConfig,
ComplexityTier,
classification_system_prompt,
@ -88,6 +89,7 @@ from litellm.types.router import (
ModelInfo,
updateDeployment,
)
from litellm.types.utils import CustomPricingLiteLLMParams
from litellm.utils import get_utc_datetime
router: Final = APIRouter()
@ -241,6 +243,7 @@ def _raise_on_strategy_router_write_violation(
_PTU_MODEL_INFO_FIELDS: Final = ("ptu_count", "cost_per_ptu_per_hour", "ptu_effective_from", "ptu_effective_to")
_PTU_PRICED_PAIR: Final = frozenset({"ptu_count", "cost_per_ptu_per_hour"})
def _explicitly_cleared_ptu_fields(model_info: ModelInfo | None) -> frozenset[str]:
@ -264,9 +267,10 @@ def _merged_ptu_model_info(*, db_model: Deployment, patch_data: updateDeployment
A PTU invariant holds over the deployment as it will exist, not over whichever subset
of fields a caller happened to send.
"""
empty: Final[Mapping[str, object]] = MappingProxyType({})
stored: Final = db_model.model_info.model_dump(exclude_none=True) if db_model.model_info else empty
incoming: Final = patch_data.model_info.model_dump(exclude_none=True) if patch_data.model_info else empty
stored: Final = db_model.model_info.model_dump(exclude_none=True) if db_model.model_info else _EMPTY_MODEL_INFO
incoming: Final = (
patch_data.model_info.model_dump(exclude_none=True) if patch_data.model_info else _EMPTY_MODEL_INFO
)
cleared: Final = _explicitly_cleared_ptu_fields(patch_data.model_info)
return MappingProxyType({k: v for k, v in {**stored, **incoming}.items() if k not in cleared})
@ -338,6 +342,140 @@ def _validate_ptu_model_info(model_info: Mapping[str, object]) -> None:
)
# The six mirrored pricing fields plus the three remaining fields
# Router._inherit_builtin_cache_pricing back-fills from the public cost map. An unset field is
# what that back-fill targets, so a field left out here is one a PTU deployment still bills.
_PTU_ZEROED_PRICING_FIELDS: Final = SPECIAL_MODEL_INFO_PARAMS + (
"cache_creation_input_token_cost_above_1hr",
"cache_creation_input_token_cost_above_200k_tokens",
"cache_read_input_token_cost_above_200k_tokens",
)
_PTU_ZEROED_PRICING: Final[Mapping[str, float]] = MappingProxyType(dict.fromkeys(_PTU_ZEROED_PRICING_FIELDS, 0.0))
_NO_PRICING_OVERRIDE: Final[Mapping[str, float]] = MappingProxyType({})
_EMPTY_MODEL_INFO: Final[Mapping[str, object]] = _NO_PRICING_OVERRIDE
# Rate fields only. CustomPricingLiteLLMParams also carries settings that are not charges
# (an embedding's output_vector_size, the regional uplift multipliers), and zeroing one of
# those would destroy the deployment's configuration rather than stop a charge.
_CUSTOM_PRICING_FIELDS: Final = frozenset(f for f in CustomPricingLiteLLMParams.model_fields if "cost" in f)
def _is_nonzero_price(value: object) -> bool:
return isinstance(value, (int, float)) and not isinstance(value, bool) and value != 0
def _is_zero_price(value: object) -> bool:
return isinstance(value, (int, float)) and not isinstance(value, bool) and value == 0
def _raise_if_ptu_deployment_is_priced(*, model_info: Mapping[str, object], supplied: Mapping[str, object]) -> None:
"""Refuse a rate the caller supplies for a deployment that bills reserved capacity.
Separate from the zeroing so the team-model path can run it before it touches the team, whose
ACL write autocommits: a refusal raised after it would leave the team changed and the
deployment row never written.
"""
if not is_ptu_cost_attribution_enabled():
return
if model_info.get("ptu_count") is None or model_info.get("cost_per_ptu_per_hour") is None:
return
priced: Final = tuple(sorted(field for field in _CUSTOM_PRICING_FIELDS if _is_nonzero_price(supplied.get(field))))
if not priced:
return
raise HTTPException(
status_code=400,
detail=(
f"A PTU deployment bills by reserved capacity, so {', '.join(priced)} cannot be charged on "
"top of it. Send 0 or no value, or remove ptu_count and cost_per_ptu_per_hour to bill per token."
),
)
def _ptu_zeroed_pricing(
*,
model_info: Mapping[str, object],
litellm_params: Mapping[str, object],
supplied: Mapping[str, object],
) -> Mapping[str, float]:
"""The pricing a PTU deployment must carry, empty unless one is being stored.
Reserved capacity is already billed by the flat cost the rollup writes, so charging the
traffic it serves bills the same tokens twice. Left unset the rate falls back to the public
cost map, which makes the double charge the default rather than an opt-in.
Only a price the caller supplies is refused. A non-zero price already on the row is zeroed
instead, so a deployment priced through a path this rule does not cover heals on its next
save rather than rejecting every later edit of a field that has nothing to do with pricing.
``supplied`` is the caller's litellm_params alone, because that is the blob a price is
authored on. model_info's copy is written by the server, both by the mirror in
``Deployment.__init__`` and by the cost-map defaults /model/info fills in, so a client that
round-trips a model_info blob sends back prices it never chose.
"""
if not is_ptu_cost_attribution_enabled():
return _NO_PRICING_OVERRIDE
if model_info.get("ptu_count") is None or model_info.get("cost_per_ptu_per_hour") is None:
return _NO_PRICING_OVERRIDE
_raise_if_ptu_deployment_is_priced(model_info=model_info, supplied=supplied)
stored: Final = frozenset(
field
for field in _CUSTOM_PRICING_FIELDS
if _is_nonzero_price(model_info.get(field)) or _is_nonzero_price(litellm_params.get(field))
)
if not stored:
return _PTU_ZEROED_PRICING
return MappingProxyType({**_PTU_ZEROED_PRICING, **dict.fromkeys(stored, 0.0)})
def _ptu_pricing_delta(
*,
stored_model_info: Mapping[str, object],
model_info: Mapping[str, object],
litellm_params: Mapping[str, object],
patch: updateDeployment,
) -> tuple[Mapping[str, float], frozenset[str]]:
"""The pricing a patch must write into both blobs, and the pricing it must drop from them.
A patch that takes the deployment off PTU takes the zeroed pricing with it, since the zeros
exist only to stop the double charge. Left behind they would serve the deployment for free.
Reading the stored row rather than the patch alone keeps that release off a deployment that
never carried PTU config, whose zero price is a rate its operator chose. A zero the patch
itself carries is released with the rest, because the dashboard echoes the whole stored
blob on every save, so a supplied zero cannot be told apart from the one this rule wrote.
The release spans every field the zeroing could have written, not just the mirrored ones, or
a rate zeroed on the way in (per-second, per-character tiers) would bill nothing forever.
"""
supplied: Final = patch.litellm_params.model_dump(exclude_none=True) if patch.litellm_params else _EMPTY_MODEL_INFO
zeroed: Final = _ptu_zeroed_pricing(model_info=model_info, litellm_params=litellm_params, supplied=supplied)
if zeroed:
return zeroed, frozenset()
was_ptu: Final = any(stored_model_info.get(field) is not None for field in _PTU_PRICED_PAIR)
if not was_ptu or not _explicitly_cleared_ptu_fields(patch.model_info) & _PTU_PRICED_PAIR:
return _NO_PRICING_OVERRIDE, frozenset()
return _NO_PRICING_OVERRIDE, frozenset(
field
for field in _CUSTOM_PRICING_FIELDS.union(_PTU_ZEROED_PRICING_FIELDS)
if _is_zero_price(model_info.get(field)) or _is_zero_price(litellm_params.get(field))
)
def _ptu_priced_deployment(model_params: Deployment) -> Deployment:
"""``model_params`` with PTU pricing applied, or itself when it configures no PTU."""
model_info: Final = model_params.model_info.model_dump(exclude_none=True)
litellm_params: Final = model_params.litellm_params.model_dump(exclude_none=True)
override: Final = _ptu_zeroed_pricing(model_info=model_info, litellm_params=litellm_params, supplied=litellm_params)
if not override:
return model_params
return model_params.model_copy(
update=MappingProxyType(
{
"litellm_params": model_params.litellm_params.model_copy(update=override),
"model_info": model_params.model_info.model_copy(update=override),
}
)
)
def _parse_ptu_datetime(value: object) -> datetime.datetime | None:
"""``value`` as a datetime, parsing an ISO string, else None."""
if isinstance(value, datetime.datetime):
@ -403,6 +541,19 @@ def update_db_model(db_model: Deployment, updated_patch: updateDeployment) -> Pr
merged_model_info.pop(field, None)
_validate_ptu_model_info(merged_model_info)
ptu_pricing, ptu_released = _ptu_pricing_delta(
stored_model_info=db_model.model_info.model_dump(exclude_none=True)
if db_model.model_info
else _EMPTY_MODEL_INFO,
model_info=merged_model_info,
litellm_params=merged_litellm_params,
patch=updated_patch,
)
merged_model_info.update(ptu_pricing)
merged_litellm_params.update(ptu_pricing)
for field in ptu_released:
merged_model_info.pop(field, None)
merged_litellm_params.pop(field, None)
# convert to prisma compatible format
@ -862,6 +1013,12 @@ async def _update_team_model_in_db(
if patch_data.model_info is not None:
_raise_if_ptu_cost_attribution_disabled(patch_data.model_info.model_dump(exclude_none=True))
_validate_ptu_model_info(_merged_ptu_model_info(db_model=db_model, patch_data=patch_data))
_raise_if_ptu_deployment_is_priced(
model_info=_merged_ptu_model_info(db_model=db_model, patch_data=patch_data),
supplied=(
patch_data.litellm_params.model_dump(exclude_none=True) if patch_data.litellm_params else _EMPTY_MODEL_INFO
),
)
patch_team_id: Final = patch_data.model_info.team_id if patch_data.model_info else None
@ -1588,6 +1745,7 @@ async def add_new_model(
incoming_model_info: Final = model_params.model_info.model_dump(exclude_none=True)
_raise_if_ptu_cost_attribution_disabled(incoming_model_info)
_validate_ptu_model_info(incoming_model_info)
priced_model_params: Final = _ptu_priced_deployment(model_params)
if store_model_in_db is True:
"""
@ -1601,13 +1759,13 @@ async def add_new_model(
_original_litellm_model_name: Final = model_params.model_name
if model_params.model_info.team_id is None:
model_response = await _add_model_to_db(
model_params=model_params,
model_params=priced_model_params,
user_api_key_dict=user_api_key_dict,
prisma_client=prisma_client,
)
else:
model_response = await _add_team_model_to_db(
model_params=model_params,
model_params=priced_model_params,
user_api_key_dict=user_api_key_dict,
prisma_client=prisma_client,
)
@ -1619,9 +1777,9 @@ async def add_new_model(
if "slack" in _alerting:
# send notification - new model added
await proxy_logging_obj.slack_alerting_instance.model_added_alert(
model_name=model_params.model_name,
model_name=priced_model_params.model_name,
litellm_model_name=_original_litellm_model_name,
passed_model_info=model_params.model_info,
passed_model_info=priced_model_params.model_info,
)
except Exception as e:
verbose_proxy_logger.exception("Exception in add_new_model: %s", e)
@ -2025,19 +2183,23 @@ def _labeled_tiers_from_query(tier_labels: str | None) -> tuple[tuple[Complexity
async def get_auto_router_classifier_default_prompt(
context_window_size: int = DEFAULT_CLASSIFIER_CONTEXT_WINDOW_SIZE,
tier_labels: str | None = None,
classification_rubric: ClassificationRubric | None = None,
) -> AutoRouterClassifierDefaultPromptResponse:
"""
Get the default classifier system prompt, so the dashboard's prompt editor can prefill it.
The prompt's closing line depends on whether prior conversation turns are quoted to the
classifier, and its tier bullets are named by the router's tier_labels, so the caller passes both
to get the text that router would actually send rather than a rubric it does not use.
classifier, its tier bullets are named by the router's tier_labels, and its calibration examples
come from the router's classification rubric, so the caller passes all three to get the text that router
would actually send rather than a rubric it does not use.
Parameters:
- context_window_size: int - The router's classifier_context_window_size. Defaults to the
built-in default.
- tier_labels: str | None - The router's tier_labels as a JSON object of canonical tier name to
display name, e.g. `{"SIMPLE": "Cheap"}`. Omit or pass an empty object for the default names.
- classification_rubric: ClassificationRubric | None - The router's
classifier_llm_config.classification_rubric. Omit for the default.
"""
if context_window_size < 0:
raise ProxyException(
@ -2050,9 +2212,11 @@ async def get_auto_router_classifier_default_prompt(
labeled_tiers: Final = _labeled_tiers_from_query(tier_labels)
return AutoRouterClassifierDefaultPromptResponse(
system_prompt=(
classification_system_prompt(context_window_size)
classification_system_prompt(context_window_size, classification_rubric=classification_rubric)
if labeled_tiers is None
else classification_system_prompt(context_window_size, labeled_tiers=labeled_tiers)
else classification_system_prompt(
context_window_size, labeled_tiers=labeled_tiers, classification_rubric=classification_rubric
)
)
)

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