scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-11-14 09:03:16 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 214

Merge litellm_internal_staging into litellm_mcp_spec_version_and_rfc9207_iss

Browse source 
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
This commit is contained in:
yassin 2026-09-02 15:46:48 +00:00
commit ace6117a72
3815 changed files with 231876 additions and 43922 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

234
.circleci/config.yml
View file

@ -430,7 +430,7 @@ jobs:
--verbose \
--command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
-vv \
--cov=./litellm \
--cov=./litellm --cov=./enterprise/litellm_enterprise \
--cov-report=xml \
--junitxml=test-results/junit.xml \
--durations=20 \
@ -504,7 +504,7 @@ jobs:
--verbose \
--command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
-vv \
--cov=./litellm \
--cov=./litellm --cov=./enterprise/litellm_enterprise \
--cov-report=xml \
--junitxml=test-results/junit.xml \
--durations=20 \
@ -631,7 +631,7 @@ jobs:
--verbose \
--command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
-v -x \
--cov=./litellm --cov-report=xml \
--cov=./litellm --cov=./enterprise/litellm_enterprise --cov-report=xml \
--junitxml=test-results/junit.xml \
--durations=5 \
-n 2"
@ -651,126 +651,6 @@ jobs:
- auth_ui_unit_tests_coverage.xml
- auth_ui_unit_tests_coverage
proxy_behavior_tests:
docker:
- *python312_image
- image: cimg/postgres:16.0@sha256:b125148bc76e8e8eee5eb3ad6020a3a14110a14e8192f1c645128afebe2e2f84
environment:
POSTGRES_USER: postgres
POSTGRES_PASSWORD: postgres
POSTGRES_DB: litellm_test
working_directory: ~/project
environment:
DATABASE_URL: "postgresql://postgres:postgres@localhost:5432/litellm_test"
steps:
- checkout
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
uv sync --frozen --all-groups --all-extras --python 3.12
- wait_for_service:
url: tcp://localhost:5432
timeout: "60"
- run:
name: Seed DB schema via prisma db push
command: |
uv run --no-sync prisma db push --schema litellm/proxy/schema.prisma --accept-data-loss
- run:
name: Generate Prisma Client
command: uv run --no-sync python -m prisma generate
- run:
name: Run proxy management behavior tests
command: |
mkdir -p test-results
uv run --no-sync python -m pytest tests/proxy_behavior \
-v --junitxml=test-results/junit.xml --durations=10
no_output_timeout: 15m
- store_test_results:
path: test-results
proxy_security_tests:
docker:
- *python312_image
- image: cimg/postgres:16.0@sha256:b125148bc76e8e8eee5eb3ad6020a3a14110a14e8192f1c645128afebe2e2f84
environment:
POSTGRES_USER: postgres
POSTGRES_PASSWORD: postgres
POSTGRES_DB: litellm_test
working_directory: ~/project
environment:
DATABASE_URL: "postgresql://postgres:postgres@localhost:5432/litellm_test"
steps:
- checkout
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
uv sync --frozen --all-groups --all-extras --python 3.12
- wait_for_service:
url: tcp://localhost:5432
timeout: "60"
- run:
name: Seed DB schema via prisma db push
command: |
uv run --no-sync prisma db push --schema litellm/proxy/schema.prisma --accept-data-loss
- run:
name: Generate Prisma Client
command: uv run --no-sync python -m prisma generate
- run:
name: Run proxy security tests
command: |
mkdir -p test-results
uv run --no-sync python -m pytest tests/proxy_security_tests \
-v --junitxml=test-results/junit.xml --durations=10
no_output_timeout: 15m
- store_test_results:
path: test-results
schema_migration_check:
docker:
- *python312_image
- image: cimg/postgres:16.0@sha256:b125148bc76e8e8eee5eb3ad6020a3a14110a14e8192f1c645128afebe2e2f84
environment:
POSTGRES_USER: postgres
POSTGRES_PASSWORD: postgres
POSTGRES_DB: litellm_test
working_directory: ~/project
environment:
# An empty database; the test applies every committed migration itself.
DATABASE_URL: "postgresql://postgres:postgres@localhost:5432/litellm_test"
steps:
- checkout
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
uv sync --frozen --all-groups --all-extras --python 3.12
- wait_for_service:
url: tcp://localhost:5432
timeout: "60"
- run:
name: Generate Prisma Client
command: uv run --no-sync python -m prisma generate
- run:
name: Check schema.prisma is in sync with committed migrations
command: |
mkdir -p test-results
uv run --no-sync python -m pytest tests/proxy_migration_tests \
-v --junitxml=test-results/junit.xml --durations=10
no_output_timeout: 15m
- store_test_results:
path: test-results
litellm_router_testing: # Runs all tests with the "router" keyword
docker:
- *python312_image
@ -858,7 +738,7 @@ jobs:
--verbose \
--command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
-v -x \
--cov=./litellm --cov-report=xml \
--cov=./litellm --cov=./enterprise/litellm_enterprise --cov-report=xml \
--junitxml=test-results/junit.xml \
--durations=5 \
-n 4"
@ -985,7 +865,7 @@ jobs:
--verbose \
--command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
-vv \
--cov=./litellm --cov-report=xml \
--cov=./litellm --cov=./enterprise/litellm_enterprise --cov-report=xml \
--junitxml=test-results/junit.xml \
--durations=20 \
-n 4 \
@ -1030,7 +910,7 @@ jobs:
--verbose \
--command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
-vv -x -s \
--cov=./litellm --cov-report=xml \
--cov=./litellm --cov=./enterprise/litellm_enterprise --cov-report=xml \
--junitxml=test-results/junit.xml \
--durations=5"
no_output_timeout: 15m
@ -1074,7 +954,7 @@ jobs:
--verbose \
--command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
-vv \
--cov=./litellm --cov-report=xml \
--cov=./litellm --cov=./enterprise/litellm_enterprise --cov-report=xml \
--junitxml=test-results/junit.xml \
--durations=5 \
-n 2 \
@ -1120,7 +1000,7 @@ jobs:
--verbose \
--command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
-vv -x -s \
--cov=./litellm --cov-report=xml \
--cov=./litellm --cov=./enterprise/litellm_enterprise --cov-report=xml \
--junitxml=test-results/junit.xml \
--durations=5 \
--retries 3 --retry-delay 5"
@ -1211,7 +1091,7 @@ jobs:
--verbose \
--command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
-vv -x \
--cov=./litellm --cov-report=xml \
--cov=./litellm --cov=./enterprise/litellm_enterprise --cov-report=xml \
--junitxml=test-results/junit.xml \
--durations=5 \
-n 4"
@ -1255,7 +1135,7 @@ jobs:
--verbose \
--command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
-vv -x \
--cov=./litellm --cov-report=xml \
--cov=./litellm --cov=./enterprise/litellm_enterprise --cov-report=xml \
--junitxml=test-results/junit.xml \
--durations=5 \
-n 4"
@ -1274,40 +1154,6 @@ jobs:
paths:
- search_coverage.xml
- search_coverage
litellm_mapped_enterprise_tests:
docker:
- *python312_image
working_directory: ~/project
resource_class: large
steps:
- checkout
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
uv sync --frozen --all-groups --all-extras --python 3.12
- setup_litellm_enterprise_pip
- run:
name: Run enterprise tests
command: |
uv run --no-sync python -m prisma generate
mkdir -p test-results
TEST_FILES=$(circleci tests glob "tests/enterprise/**/test_*.py")
echo "$TEST_FILES" | circleci tests run \
--verbose \
--command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
-v -x \
--junitxml=test-results/junit-enterprise.xml \
--durations=10 \
-n 4"
no_output_timeout: 15m
# Store test results
- store_test_results:
path: test-results
batches_testing:
docker:
- *python312_image
@ -1333,7 +1179,7 @@ jobs:
--verbose \
--command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
-vv -x -s \
--cov=./litellm --cov-report=xml \
--cov=./litellm --cov=./enterprise/litellm_enterprise --cov-report=xml \
--junitxml=test-results/junit.xml \
--durations=5 \
-n 2"
@ -1377,7 +1223,7 @@ jobs:
--verbose \
--command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
-vv -x -s \
--cov=./litellm --cov-report=xml \
--cov=./litellm --cov=./enterprise/litellm_enterprise --cov-report=xml \
--junitxml=test-results/junit.xml \
--durations=5 \
-n 2"
@ -1422,7 +1268,7 @@ jobs:
--verbose \
--command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
-vv -x \
--cov=./litellm --cov-report=xml \
--cov=./litellm --cov=./enterprise/litellm_enterprise --cov-report=xml \
--junitxml=test-results/junit.xml \
--durations=5 \
-n 4"
@ -1501,7 +1347,7 @@ jobs:
--verbose \
--command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
-vv \
--cov=./litellm --cov-report=xml \
--cov=./litellm --cov=./enterprise/litellm_enterprise --cov-report=xml \
-n 4 \
--junitxml=test-results/junit.xml \
--durations=5 \
@ -1546,7 +1392,7 @@ jobs:
--verbose \
--command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
-vv -x -s \
--cov=./litellm --cov-report=xml \
--cov=./litellm --cov=./enterprise/litellm_enterprise --cov-report=xml \
--junitxml=test-results/junit.xml \
--durations=5"
no_output_timeout: 15m
@ -1599,7 +1445,7 @@ jobs:
--verbose \
--command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
-vv -x -s \
--cov=./litellm --cov-report=xml \
--cov=./litellm --cov=./enterprise/litellm_enterprise --cov-report=xml \
--junitxml=test-results/junit.xml \
--durations=5 -n 2 \
--reruns 2 --reruns-delay 1"
@ -2575,45 +2421,6 @@ jobs:
- wait_for_service:
url: http://localhost:4000
timeout: "300"
# Add Ruby installation and testing before the existing Node.js and Python tests
- run:
name: Install Ruby and Bundler
command: |
# Clone RVM at pinned tag and verify the commit SHA matches the
# published tag before running its install script.
RVM_VERSION="1.29.12"
RVM_EXPECTED_SHA="6bfc9213c9d6914fe756f524eb034a403d51db81"
git clone --depth 1 --branch "$RVM_VERSION" https://github.com/rvm/rvm.git /tmp/rvm
RVM_ACTUAL_SHA="$(git -C /tmp/rvm rev-parse HEAD)"
if [ "$RVM_ACTUAL_SHA" != "$RVM_EXPECTED_SHA" ]; then
echo "RVM tag $RVM_VERSION resolved to $RVM_ACTUAL_SHA; expected $RVM_EXPECTED_SHA" >&2
exit 1
fi
# Import RVM signing keys (used by `rvm install` to verify Ruby tarballs)
gpg --keyserver hkp://keyserver.ubuntu.com --recv-keys 409B6B1796C275462A1703113804BB82D39DC0E3 7D2BAF1CF37B13E2069D6956105BD0E739499BDB
# Install RVM from the verified checkout. The install script
# sources `scripts/functions/installer` using paths relative to
# its own working directory, so it must be run from /tmp/rvm.
(cd /tmp/rvm && ./install --path "$HOME/.rvm")
source "$HOME/.rvm/scripts/rvm"
# Install Ruby 3.2.2 (RVM verifies the tarball PGP signature)
rvm install 3.2.2
rvm use 3.2.2 --default
# Install latest Bundler
gem install bundler
- run:
name: Run Ruby tests
command: |
source $HOME/.rvm/scripts/rvm
cd tests/pass_through_tests/ruby_passthrough_tests
bundle install
bundle exec rspec
no_output_timeout: 30m
# Install Node.js directly from nodejs.org with SHA256 verification,
# instead of piping NodeSource's setup_24.x apt-repo installer into
# sudo bash (which runs a mutable upstream script unattended).
@ -3105,12 +2912,6 @@ workflows:
filters: *main_branches
- auth_ui_unit_tests:
filters: *main_branches
- proxy_behavior_tests:
filters: *main_branches
- proxy_security_tests:
filters: *main_branches
- schema_migration_check:
filters: *main_branches
- build_docker_database_image:
filters: *main_branches
- e2e_ui_testing:
@ -3167,8 +2968,6 @@ workflows:
filters: *main_branches
- search_testing:
filters: *main_branches
- litellm_mapped_enterprise_tests:
filters: *main_branches
- batches_testing:
filters: *main_branches
- litellm_utils_testing:
@ -3191,7 +2990,6 @@ workflows:
- guardrails_testing
- ocr_testing
- search_testing
- litellm_mapped_enterprise_tests
- batches_testing
- litellm_utils_testing
- pass_through_unit_testing

9
.github/CODEOWNERS vendored
View file

@ -1,5 +1,10 @@
/ui/ @yuneng-jiang @ryan-crabbe-berri
/litellm/proxy/_experimental/out/ @yuneng-jiang @ryan-crabbe-berri
/ui/ @yuneng-berri @ryan-crabbe-berri
/litellm/proxy/_experimental/out/ @yuneng-berri @ryan-crabbe-berri
/ui/Dockerfile
/ui/nginx.conf
/ui/litellm-dashboard/src/lib/http/schema.d.ts
/ui/litellm-dashboard/tsconfig.tsbuildinfo
/model_prices_and_context_window.json @mateo-berri
/litellm/model_prices_and_context_window_backup.json @mateo-berri
/litellm-proxy-extras/litellm_proxy_extras/migrations/ @yuneng-berri @ryan-crabbe-berri
/.github/CODEOWNERS @yuneng-berri

30
.github/actions/cache-cargo-build/action.yml vendored Normal file
View file

@ -0,0 +1,30 @@
name: "Cache the Rust build"
description: >-
Cache the Cargo registry and target directory the root package's build needs,
so only the first job on a given Cargo.lock compiles the bridge from scratch.
litellm builds through maturin, which compiles litellm-rust/crates/python-bridge
in the dev profile for editable installs. `uv sync` therefore pays a full build
in every job that installs the workspace. Nothing caught it, because the uv cache
holds wheels uv downloads rather than wheels it builds, and a path dependency
whose source moves every commit could never hit that cache anyway. Cargo rebuilds
only what changed when its target directory survives, so a warm job pays for the
bridge crate alone.
The key namespace is separate from test-rust.yml's check and release caches. They
cache the same directory for different workloads, and a shared key would let
whichever ran first deny the others a save.
runs:
using: composite
steps:
- name: Restore the Cargo registry and target directory
uses: actions/cache@0057852bfaa89a56745cba8c7296529d2fc39830 # v4.3.0
with:
path: |
~/.cargo/registry
~/.cargo/git
litellm-rust/target
key: ${{ runner.os }}-maturin-dev-${{ hashFiles('litellm-rust/Cargo.lock') }}
restore-keys: |
${{ runner.os }}-maturin-dev-

46
.github/ci-coverage-allowlist.yml vendored
View file

@ -5,24 +5,21 @@ description: >-
test_paths:
- reason: >-
The caching suite in tests/local_testing, which runs nowhere. Every job that globs that
directory either deselects it (local_testing_part1 and part2 carry `-k "... and not caching
and not cache"`) or keeps only another keyword (langfuse, router, assistants), and no job
names these files the way redis_caching_unit_tests names test_dual_cache.py. Measured
2026-08-20 by collecting the directory under each job's own selector: 118 tests across
these eight files are selected by none of them. Listed so the gap is a decision rather
than an accident, and so the --slices guard has a baseline to ratchet down from. Revisit
when tests/local_testing is ported off CircleCI, where the keyless part of this suite
belongs in a real job
What is left of the caching suite in tests/local_testing that runs nowhere. Every job that
globs that directory either deselects it (local_testing_part1 and part2 carry `-k "... and
not caching and not cache"`) or keeps only another keyword (langfuse, router, assistants),
and no job names these files the way redis_caching_unit_tests names test_dual_cache.py.
The gap was eight files and 118 tests when measured 2026-08-20; the five keyless ones now
run in the caching-local shard, leaving these three. Measured 2026-08-21 with no provider
credentials and no Redis: test_caching.py needs both (37 of 65 fail without them),
test_disk_cache_unit_tests.py needs OPENAI_API_KEY for 2 of its 4, and
test_gcs_cache_unit_tests.py needs GCS credentials for all 4. They want the keyless/live
split that porting tests/local_testing off CircleCI will force, not a job that is red by
construction
paths:
- tests/local_testing/test_cache_preset_key.py
- tests/local_testing/test_caching.py
- tests/local_testing/test_caching_handler.py
- tests/local_testing/test_disk_cache_unit_tests.py
- tests/local_testing/test_gcs_cache_unit_tests.py
- tests/local_testing/test_prompt_caching.py
- tests/local_testing/test_responses_stream_cache_keys.py
- tests/local_testing/test_unit_test_caching.py
- reason: >-
The end-to-end suite runs against a deployed proxy from its own in-cluster rig rather than
from a pull request; it needs a live gateway and provider credentials no PR job holds
@ -48,16 +45,6 @@ test_paths:
choice it informed is settled
paths:
- tests/code_coverage_tests/test_aio_http_image_conversion.py
- reason: >-
The last file of a second mirror that sat beside tests/test_litellm and ran nowhere. Its
other 33 files landed in the real mirror during August 2026, 30 as moves and 3 by merging
their bodies into the live file of the same name. This one cannot follow either route yet:
its live twin was rewritten from 1268 lines to 9434, and of the 19 tests here 5 have no
counterpart while 25 assertions fail against today's code, so what survives that rewrite
is a judgement about the endpoints, not a merge. Revisit by deciding which of the five
behaviours still hold
paths:
- tests/litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py
- reason: >-
No job invokes this suite and its files mix pure transformation tests with ones driving live
vendor vector stores, so assigning them needs a per-file decision
@ -102,17 +89,6 @@ test_paths:
- tests/integration/sandbox/test_e2b_sandbox.py
- tests/integration/test_oci_integration.py
- tests/integration/test_oci_proxy_integration.py
- reason: >-
A unit test for the proxy-extras package that no job invokes, while the package's other
tests live under tests/proxy_migration_tests. Measured 2026-08-20: 24 of its 28 tests pass
and the 4 in TestMigrationSQLIdempotency fail, because 13 migrations from 2026-03 onward use
bare CREATE TABLE, ADD COLUMN, CREATE INDEX and ADD CONSTRAINT rather than the guarded forms
this file requires. It also matches those keywords inside SQL comments, so two further
migrations are reported that are in fact fine. Wiring it up means deciding what to do about
the 13 first, and they cannot simply be edited: Prisma checksums an applied migration, so a
changed one breaks migrate deploy for existing installs
paths:
- tests/litellm-proxy-extras/test_litellm_proxy_extras_utils.py
dockerfiles:
- reason: >-

5
.github/mutmut-coverage.rc vendored Normal file
View file

@ -0,0 +1,5 @@
# mutmut's gather_coverage() looks covered lines up by absolute path, so the
# repo's `relative_files = true` makes every lookup miss and mutmut generates
# zero mutants. Point COVERAGE_RCFILE here for mutation runs only.
[run]
relative_files = false

23
.github/pull_request_template.md vendored
View file

@ -1,7 +1,10 @@
<!-- The whole description's target audience is humans, not AI agents: write it in plain, simple,
everyday engineering language, extremely parsable and readable at a glance. This goes double for
the TLDR, User Flow, and Caveats sections -->
## TLDR
<!-- Fill in the bullets below and keep each one short and concrete: one line per bullet, roughly 10 words max
This section must be extremely human parsable, comprehensible, and readable: its target audience is humans, not AI agents -->
<!-- Fill in the bullets below and keep each one short and concrete: one line per bullet, roughly 10 words max -->
Problem this solves:
@ -110,8 +113,20 @@ If you're seeing a delay in your PR being merged, ping the LiteLLM Team on [Slac
## Caveats (if any)
<!-- Short bullet points, just like the TLDR: one line per bullet, roughly 10 words max
<!-- Group caveats under severity subheadings (### Severe, ### High, ### Medium, ### Low), with
short bullet points inside each, just like the TLDR: one line per bullet, roughly 10 words max
Call out known limitations, follow-up work, or anything a reviewer should watch out for
Include only the tiers that have caveats; drop the empty ones
- Severe: inherent to what the PR deliberately ships, there even when the code works as intended:
it can degrade or take down a running deployment (e.g. a slow or table-locking boot migration),
rewrite data by design, break an existing workflow on purpose, or change auth behavior. An
operator must plan around it before rollout
- High: an unintended hole: a correctness, security, data-loss, or backward-compatibility bug,
unsafe to ship as is
- Medium: a real gap someone can hit, but with a workaround or a narrow blast radius
- Low: anything else worth noting: naming, cleanup, an edge case nobody hits
Nest bullets as deep as helps: hierarchy beats one long line when it makes things clearer to a
human reader
Leave this section empty if there are none -->
## QA runbook
@ -134,6 +149,6 @@ Example checklists:
- [ ] Sanity check: this test makes sense to add and is not hand-wavey (e.g., assert actual expected spend instead of just spend > 0) or potentially flaky
-->
### Final Attestation
## Final Attestation
- [ ] The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

16
.github/scripts/assert_ci_coverage.py vendored
View file

@ -312,8 +312,23 @@ def _matchable_names(relative_path: str) -> frozenset[str]:
)
def _workflow_named_tokens() -> frozenset[str]:
"""Test tokens a GitHub Actions job names directly.
A CircleCI `-k` that deselects a file no longer means the file runs nowhere once a
workflow names it, so the slice check has to credit those the same way the census does.
"""
return _invoked_test_tokens(
scalar
for path in _config_files()
if path != CIRCLECI_CONFIG
for scalar in _scalars(yaml.safe_load(path.read_text(encoding="utf-8")), path.name)
)
def _deselected_everywhere(allowlist: Allowlist) -> tuple[Finding, ...]:
slices: Final = _slices()
named_by_workflow: Final = _workflow_named_tokens()
globbed: Final = tuple(
path
for path in _test_files()
@ -326,6 +341,7 @@ def _deselected_everywhere(allowlist: Allowlist) -> tuple[Finding, ...]:
)
for path in globbed
if not allowlist.covers_test(path)
and not any(_token_covers(token, path) for token in named_by_workflow)
and not any(slice_.claims(path, _matchable_names(path)) for slice_ in slices)
)

230
.github/scripts/close_duplicate_issues.py vendored
View file

@ -1,230 +0,0 @@
#!/usr/bin/env python3
"""
Detect and close duplicate GitHub issues using title similarity.
Modes:
--scan Compare all open issues against each other (batch)
--issue-number N Check a single issue against older open issues
Requires the `gh` CLI to be authenticated.
"""
import argparse
import difflib
import json
import re
import subprocess
import sys
def normalize_title(title: str) -> str:
"""Strip common prefixes, lowercase, and collapse whitespace."""
title = re.sub(
r"^\[?(bug|feature request|enhancement|question|docs)[:\]]?\s*",
"",
title,
flags=re.IGNORECASE,
)
return " ".join(title.lower().split())
def gh(*args: str) -> str:
"""Run a gh CLI command and return stdout."""
result = subprocess.run(
["gh", *args],
capture_output=True,
text=True,
check=True,
)
return result.stdout
def fetch_open_issues(repo: str | None) -> list[dict]:
"""Fetch all open issues (excluding PRs) via gh api --paginate."""
if repo:
endpoint = (
f"repos/{repo}/issues?state=open&per_page=100&sort=created&direction=asc"
)
else:
endpoint = "repos/{owner}/{repo}/issues?state=open&per_page=100&sort=created&direction=asc"
cmd = ["api", "--paginate", endpoint]
raw = gh(*cmd)
# gh --paginate concatenates JSON arrays, so we may get multiple arrays
issues = []
for line in raw.strip().splitlines():
line = line.strip()
if not line:
continue
parsed = json.loads(line)
if isinstance(parsed, list):
issues.extend(parsed)
else:
issues.append(parsed)
# Filter out pull requests (they also appear in the issues endpoint)
return [i for i in issues if "pull_request" not in i]
def close_as_duplicate(
issue_number: int, duplicate_of: int, repo: str | None, dry_run: bool
) -> None:
"""Close an issue as duplicate of another, adding a comment and label."""
repo_args = ["--repo", repo] if repo else []
if dry_run:
print(
f" [DRY RUN] Would close #{issue_number} as duplicate of #{duplicate_of}"
)
return
# Add comment
comment_body = (
f"Closing as duplicate of #{duplicate_of}.\n\n"
"If you believe this is not a duplicate, please reopen and add context "
"explaining how this differs."
)
gh("issue", "comment", str(issue_number), "--body", comment_body, *repo_args)
# Add label
gh("issue", "edit", str(issue_number), "--add-label", "duplicate", *repo_args)
# Close with not_planned reason
gh(
"api",
f"repos/{repo or '{owner}/{repo}'}/issues/{issue_number}",
"-X",
"PATCH",
"-f",
"state=closed",
"-f",
"state_reason=not_planned",
)
print(f" Closed #{issue_number} as duplicate of #{duplicate_of}")
def find_duplicate(
issue: dict, candidates: list[dict], threshold: float
) -> dict | None:
"""Return the first candidate whose normalized title is above threshold."""
norm = normalize_title(issue["title"])
for candidate in candidates:
if candidate["number"] == issue["number"]:
continue
cand_norm = normalize_title(candidate["title"])
ratio = difflib.SequenceMatcher(None, norm, cand_norm).ratio()
if ratio >= threshold:
return candidate
return None
def scan_all(
issues: list[dict], threshold: float, repo: str | None, dry_run: bool
) -> int:
"""Compare every issue against all older issues. Returns count of duplicates found."""
# Sort oldest first
issues.sort(key=lambda i: i["number"])
closed_count = 0
for idx, issue in enumerate(issues):
older = issues[:idx]
if not older:
continue
dup = find_duplicate(issue, older, threshold)
if dup:
ratio = difflib.SequenceMatcher(
None,
normalize_title(issue["title"]),
normalize_title(dup["title"]),
).ratio()
print(
f"#{issue['number']}: \"{issue['title']}\"\n"
f" -> duplicate of #{dup['number']}: \"{dup['title']}\" "
f"({ratio:.0%} similar)"
)
close_as_duplicate(issue["number"], dup["number"], repo, dry_run)
closed_count += 1
return closed_count
def check_single(
issue_number: int,
issues: list[dict],
threshold: float,
repo: str | None,
dry_run: bool,
) -> bool:
"""Check a single issue against all older open issues. Returns True if duplicate found."""
target = None
for i in issues:
if i["number"] == issue_number:
target = i
break
if target is None:
print(f"Issue #{issue_number} not found among open issues.")
return False
older = [i for i in issues if i["number"] < issue_number]
dup = find_duplicate(target, older, threshold)
if dup:
ratio = difflib.SequenceMatcher(
None,
normalize_title(target["title"]),
normalize_title(dup["title"]),
).ratio()
print(
f"#{target['number']}: \"{target['title']}\"\n"
f" -> duplicate of #{dup['number']}: \"{dup['title']}\" "
f"({ratio:.0%} similar)"
)
close_as_duplicate(issue_number, dup["number"], repo, dry_run)
return True
print(f"#{issue_number}: no duplicate found above threshold {threshold}")
return False
def main() -> None:
parser = argparse.ArgumentParser(
description="Detect and close duplicate GitHub issues"
)
mode = parser.add_mutually_exclusive_group(required=True)
mode.add_argument("--scan", action="store_true", help="Scan all open issues")
mode.add_argument("--issue-number", type=int, help="Check a single issue number")
parser.add_argument(
"--threshold", type=float, default=0.85, help="Similarity threshold (0-1)"
)
parser.add_argument(
"--close",
action="store_true",
help="Actually close duplicates (default is dry-run)",
)
parser.add_argument(
"--repo", type=str, help="Repository (owner/repo). Auto-detected if omitted."
)
args = parser.parse_args()
dry_run = not args.close
if dry_run:
print("=== DRY RUN MODE (pass --close to actually close issues) ===\n")
print("Fetching open issues...")
issues = fetch_open_issues(args.repo)
print(f"Found {len(issues)} open issues.\n")
if args.scan:
count = scan_all(issues, args.threshold, args.repo, dry_run)
print(f"\nTotal duplicates {'found' if dry_run else 'closed'}: {count}")
else:
found = check_single(
args.issue_number, issues, args.threshold, args.repo, dry_run
)
sys.exit(0 if found else 0) # Always exit 0; finding no dup is not an error
if __name__ == "__main__":
main()

198
.github/scripts/e2e_egress_sentinel.py vendored Executable file
View file

@ -0,0 +1,198 @@
"""Prove an e2e replay run makes zero outbound provider calls, by counting them.
`serve` pins each provider host (`--host`) to a local sink address in the hosts
file and binds a counting listener on that address, so any connection the proxy
or the record/replay edge opens to a real provider is redirected to the sink,
recorded as one line in `--hits-file`, and never leaves the box. The record and
replay edge only ever dials `127.0.0.1:<edge-port>` (a different host than the
pinned provider names), so in a clean replay the sink sees nothing; a single hit
means a provider call escaped the bundle. `assert-empty` turns that hit file into
the pass/fail check.
Stdlib only, so CI runs it under the system interpreter as root (binding :443 and
editing the hosts file both need root); `--sink-address`, `--port`, and
`--hosts-file` are injectable so it runs unprivileged against a temp hosts file on
a high port under test.
"""
# ruff: noqa: T201 # CLI script: its stdout/stderr progress and results are the interface
from __future__ import annotations
import argparse
import json
import os
import signal
import socket
import sys
import threading
import time
from dataclasses import dataclass
from pathlib import Path
from types import FrameType
from typing import Final
_BLOCK_BEGIN: Final = "# BEGIN e2e-egress-sentinel"
_BLOCK_END: Final = "# END e2e-egress-sentinel"
@dataclass(frozen=True, slots=True)
class ServeConfig:
hosts: tuple[str, ...]
sink_address: str
ports: tuple[int, ...]
hits_file: Path
hosts_file: Path
ready_file: Path | None
pid_file: Path | None
def _pin_block(sink_address: str, hosts: tuple[str, ...]) -> str:
lines = "\n".join(f"{sink_address}\t{host}" for host in hosts)
return f"\n{_BLOCK_BEGIN}\n{lines}\n{_BLOCK_END}\n"
def _install_pins(hosts_file: Path, sink_address: str, hosts: tuple[str, ...]) -> bytes:
original = hosts_file.read_bytes() if hosts_file.exists() else b""
hosts_file.write_bytes(original + _pin_block(sink_address, hosts).encode())
return original
def _restore_pins(hosts_file: Path, original: bytes) -> None:
hosts_file.write_bytes(original)
def _bind(sink_address: str, port: int) -> socket.socket:
listener = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
listener.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
listener.bind((sink_address, port))
listener.listen(128)
return listener
@dataclass(frozen=True, slots=True)
class _HitLog:
path: Path
_lock: threading.Lock
def record(self, *, port: int, peer: tuple[str, int]) -> None:
entry = json.dumps({"ts": time.time(), "port": port, "peer": list(peer)})
with self._lock:
with self.path.open("a", encoding="utf-8") as handle:
handle.write(entry + "\n")
def _serve_socket(listener: socket.socket, port: int, hits: _HitLog, stop: threading.Event) -> None:
while not stop.is_set():
try:
conn, peer = listener.accept()
except OSError:
return
hits.record(port=port, peer=(peer[0], peer[1]))
try:
conn.close()
except OSError:
pass
def serve(config: ServeConfig) -> int:
config.hits_file.write_text("", encoding="utf-8")
original_hosts = _install_pins(config.hosts_file, config.sink_address, config.hosts)
try:
listeners = tuple(_bind(config.sink_address, port) for port in config.ports)
except OSError as exc:
_restore_pins(config.hosts_file, original_hosts)
print(f"egress sentinel could not bind a sink: {exc}", file=sys.stderr)
return 1
stop = threading.Event()
hits = _HitLog(path=config.hits_file, _lock=threading.Lock())
threads = tuple(
threading.Thread(target=_serve_socket, args=(listener, port, hits, stop), daemon=True)
for listener, port in zip(listeners, config.ports)
)
for thread in threads:
thread.start()
def _handle(_signum: int, _frame: FrameType | None) -> None:
stop.set()
for listener in listeners:
try:
listener.close()
except OSError:
pass
signal.signal(signal.SIGTERM, _handle)
signal.signal(signal.SIGINT, _handle)
if config.pid_file is not None:
config.pid_file.write_text(str(os.getpid()), encoding="utf-8")
if config.ready_file is not None:
config.ready_file.write_text("ready", encoding="utf-8")
print(
f"egress sentinel up: pinned {', '.join(config.hosts)} to {config.sink_address} "
f"on port(s) {', '.join(str(p) for p in config.ports)}",
flush=True,
)
stop.wait()
_restore_pins(config.hosts_file, original_hosts)
if config.ready_file is not None and config.ready_file.exists():
config.ready_file.unlink()
if config.pid_file is not None and config.pid_file.exists():
config.pid_file.unlink()
return 0
def assert_empty(hits_file: Path) -> int:
if not hits_file.exists():
print(f"egress sentinel recorded no provider calls ({hits_file} absent): zero egress")
return 0
hits = [line for line in hits_file.read_text(encoding="utf-8").splitlines() if line.strip()]
if not hits:
print("egress sentinel recorded no provider calls: zero egress")
return 0
print(f"egress sentinel recorded {len(hits)} provider call(s); replay was not hermetic:", file=sys.stderr)
for line in hits:
print(f" {line}", file=sys.stderr)
return 1
def _serve_from_args(args: argparse.Namespace) -> int:
config = ServeConfig(
hosts=tuple(args.host),
sink_address=args.sink_address,
ports=tuple(args.port),
hits_file=Path(args.hits_file),
hosts_file=Path(args.hosts_file),
ready_file=Path(args.ready_file) if args.ready_file else None,
pid_file=Path(args.pid_file) if args.pid_file else None,
)
return serve(config)
def main(argv: tuple[str, ...]) -> int:
parser = argparse.ArgumentParser(description="count outbound provider calls during an e2e replay")
sub = parser.add_subparsers(dest="command", required=True)
serve_parser = sub.add_parser("serve", help="pin provider hosts and count connection attempts")
serve_parser.add_argument("--host", action="append", required=True, help="provider host to pin and watch")
serve_parser.add_argument("--sink-address", default="127.0.0.1")
serve_parser.add_argument("--port", action="append", type=int, default=None)
serve_parser.add_argument("--hits-file", required=True)
serve_parser.add_argument("--hosts-file", default="/etc/hosts")
serve_parser.add_argument("--ready-file", default=None)
serve_parser.add_argument("--pid-file", default=None)
assert_parser = sub.add_parser("assert-empty", help="exit non-zero if any provider call was recorded")
assert_parser.add_argument("--hits-file", required=True)
args = parser.parse_args(argv)
if args.command == "serve":
if args.port is None:
args.port = [443]
return _serve_from_args(args)
return assert_empty(Path(args.hits_file))
if __name__ == "__main__":
raise SystemExit(main(tuple(sys.argv[1:])))

55
.github/scripts/e2e_fetch_fixture_bundle.sh vendored Executable file
View file

@ -0,0 +1,55 @@
#!/usr/bin/env bash
set -euo pipefail
REPO="${1:-${GITHUB_REPOSITORY:?REPO required}}"
ARTIFACT_NAME="${2:-e2e-fixtures-bundle}"
BASE_BRANCH="${3:?base branch required}"
DEST_DIR="${4:?destination bundle dir required}"
: "${GH_TOKEN:?GH_TOKEN required to query and download artifacts}"
WORKDIR="$(mktemp -d)"
trap 'rm -rf "${WORKDIR}"' EXIT
echo "resolving newest non-expired '${ARTIFACT_NAME}' artifact on ${REPO}@${BASE_BRANCH}"
SELECTED="$(
gh api "repos/${REPO}/actions/artifacts" -X GET -f per_page=100 --paginate \
--jq ".artifacts[] | select(.name == \"${ARTIFACT_NAME}\" and .expired == false and .workflow_run.head_branch == \"${BASE_BRANCH}\") | {id, digest, created_at, run_id: .workflow_run.id, run_number: .workflow_run.run_number}" \
| jq -s 'sort_by(.created_at) | reverse | .[0] // empty'
)"
if [[ -z "${SELECTED}" ]]; then
echo "no usable '${ARTIFACT_NAME}' artifact on ${BASE_BRANCH}: the last record run produced none (a red Saturday), so there is nothing fresh to replay; failing loudly instead of replaying a stale bundle" >&2
exit 1
fi
RUN_ID="$(echo "${SELECTED}" | jq -r '.run_id')"
RUN_NUMBER="$(echo "${SELECTED}" | jq -r '.run_number')"
ARTIFACT_ID="$(echo "${SELECTED}" | jq -r '.id')"
GH_DIGEST="$(echo "${SELECTED}" | jq -r '.digest // "unknown"')"
CREATED_AT="$(echo "${SELECTED}" | jq -r '.created_at')"
echo "pinned bundle: run #${RUN_NUMBER} (run_id=${RUN_ID}, artifact_id=${ARTIFACT_ID}), recorded ${CREATED_AT}, github digest ${GH_DIGEST}"
gh run download "${RUN_ID}" --repo "${REPO}" -n "${ARTIFACT_NAME}" -D "${WORKDIR}"
TARBALL="$(find "${WORKDIR}" -name '*.tar.gz' -type f | head -n 1)"
if [[ -z "${TARBALL}" ]]; then
echo "downloaded artifact contained no tarball" >&2
exit 1
fi
SIDECAR="${TARBALL}.sha256"
if [[ ! -f "${SIDECAR}" ]]; then
echo "downloaded artifact has no ${SIDECAR}: cannot verify the bundle digest" >&2
exit 1
fi
echo "verifying bundle against its recorded sha256 digest"
( cd "$(dirname "${TARBALL}")" && sha256sum -c "$(basename "${SIDECAR}")" )
mkdir -p "${DEST_DIR}"
tar xzf "${TARBALL}" -C "${DEST_DIR}"
echo "extracted bundle into ${DEST_DIR}"
python3 -c "import json,sys; m=json.load(open(sys.argv[1])); print(' recorded_at', m['recorded_at'], 'harness', m['harness_version'], 'format_version', m['format_version'])" "${DEST_DIR}/manifest.json"

36
.github/scripts/e2e_pack_fixture_bundle.sh vendored Executable file
View file

@ -0,0 +1,36 @@
#!/usr/bin/env bash
set -euo pipefail
if [[ $# -ne 2 ]]; then
echo "usage: $0 <bundle-dir> <out-tarball>" >&2
exit 2
fi
BUNDLE_DIR="$1"
OUT_TARBALL="$2"
MANIFEST="${BUNDLE_DIR}/manifest.json"
if [[ ! -f "${MANIFEST}" ]]; then
echo "no ${MANIFEST}: refusing to publish a bundle with no manifest (record produced nothing)" >&2
exit 1
fi
echo "packing fixture bundle from ${BUNDLE_DIR}"
python3 -c "import json,sys; m=json.load(open(sys.argv[1])); print(' format_version', m['format_version'], 'recorded_at', m['recorded_at'], 'harness', m['harness_version'])" "${MANIFEST}"
TEST_DIRS=$(find "${BUNDLE_DIR}" -mindepth 1 -maxdepth 1 -type d | wc -l | tr -d ' ')
if [[ "${TEST_DIRS}" -eq 0 ]]; then
echo "bundle at ${BUNDLE_DIR} has a manifest but no recorded interactions; refusing to publish an empty bundle" >&2
exit 1
fi
echo " ${TEST_DIRS} recorded test director(ies)"
mkdir -p "$(dirname "${OUT_TARBALL}")"
tar czf "${OUT_TARBALL}" -C "${BUNDLE_DIR}" .
OUT_DIR="$(cd "$(dirname "${OUT_TARBALL}")" && pwd)"
OUT_BASE="$(basename "${OUT_TARBALL}")"
( cd "${OUT_DIR}" && sha256sum "${OUT_BASE}" > "${OUT_BASE}.sha256" )
echo "wrote ${OUT_TARBALL} ($(du -h "${OUT_TARBALL}" | cut -f1)) and ${OUT_BASE}.sha256"
cat "${OUT_DIR}/${OUT_BASE}.sha256"

13
.github/workflows/_test-unit-base.yml vendored
View file

@ -27,7 +27,7 @@ on:
default: 20
job-timeout-minutes:
description: >-
Backstop for the whole job. Keep it >= `timeout-minutes` plus 35: 30 for
Backstop for the whole job. Keep it >= `timeout-minutes` plus 40: 35 for
the per-step ceilings on the setup steps below, and 5 for the runner
overhead the job clock charges but no step owns (job init, step
transitions, post-job cleanup). That headroom is what makes the test
@ -36,7 +36,7 @@ on:
arithmetic, so the sum is passed in rather than computed.
required: false
type: number
default: 55
default: 60
max-failures:
description: "Stop after this many failures"
required: false
@ -103,6 +103,11 @@ jobs:
restore-keys: |
${{ runner.os }}-uv-
- name: Cache the Rust build
if: steps.changes.outputs.decision != 'skip'
timeout-minutes: 5
uses: ./.github/actions/cache-cargo-build
- name: Install dependencies
if: steps.changes.outputs.decision != 'skip'
timeout-minutes: 8
@ -144,7 +149,7 @@ jobs:
--reruns "${RERUNS}" \
--reruns-delay 1 \
--durations=20 \
--cov=./litellm \
--cov=./litellm --cov=./enterprise/litellm_enterprise \
--cov-report=xml:coverage.xml \
--cov-config=pyproject.toml
else
@ -156,7 +161,7 @@ jobs:
--reruns-delay 1 \
--dist="${DIST}" \
--durations=20 \
--cov=./litellm \
--cov=./litellm --cov=./enterprise/litellm_enterprise \
--cov-report=xml:coverage.xml \
--cov-config=pyproject.toml
fi

69
.github/workflows/auto-close-duplicates.yml vendored Normal file
View file

@ -0,0 +1,69 @@
name: Auto-close duplicate issues
on:
schedule:
- cron: "0 9 * * *"
workflow_dispatch:
inputs:
dry_run:
description: Log which issues would close without closing anything
type: boolean
default: true
grace_period_days:
description: Days a duplicate notice must go unanswered before the close
type: number
default: 3
pull_request:
paths:
- .github/workflows/auto-close-duplicates.yml
- scripts/auto-close-duplicates.ts
- scripts/auto-close-duplicates.test.ts
permissions: {}
jobs:
test:
if: github.event_name == 'pull_request'
runs-on: ubuntu-latest
timeout-minutes: 5
permissions:
contents: read
steps:
- name: Checkout repository
uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
with:
persist-credentials: false
- name: Setup Bun
uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0
with:
bun-version: "1.4.0"
- name: Test the sweep
run: bun test scripts/auto-close-duplicates.test.ts
sweep:
if: github.event_name != 'pull_request' && github.repository == 'BerriAI/litellm'
runs-on: ubuntu-latest
timeout-minutes: 10
permissions:
contents: read
issues: write
steps:
- name: Checkout repository
uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
with:
persist-credentials: false
- name: Setup Bun
uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0
with:
# Exact version, never latest: the next step holds an issues: write token
bun-version: "1.4.0"
- name: Close unanswered duplicates, reopen ones the reporter answered
run: bun run scripts/auto-close-duplicates.ts
env:
GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
DRY_RUN: ${{ inputs.dry_run == true }}
GRACE_PERIOD_DAYS: ${{ inputs.grace_period_days }}

22
.github/workflows/check-ui-api-types.yml vendored
View file

@ -67,6 +67,10 @@ jobs:
restore-keys: |
${{ runner.os }}-uv-
- name: Cache the Rust build
if: steps.changes.outputs.relevant == 'true'
uses: ./.github/actions/cache-cargo-build
- name: Install backend dependencies
if: steps.changes.outputs.relevant == 'true'
run: .github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra google --extra proxy --extra semantic-router
@ -79,6 +83,24 @@ jobs:
if: steps.changes.outputs.relevant == 'true'
run: uv run --no-sync prisma generate --schema litellm/proxy/schema.prisma
- name: Regenerate the lazy OpenAPI snapshot
if: steps.changes.outputs.relevant == 'true'
run: uv run --no-sync python -m litellm.proxy._lazy_openapi_snapshot
- name: Fail if the lazy OpenAPI snapshot is stale
if: steps.changes.outputs.relevant == 'true'
run: |
if ! git diff --exit-code -- litellm/proxy/_lazy_openapi_snapshot.json; then
echo "::error file=litellm/proxy/_lazy_openapi_snapshot.json::The lazy OpenAPI snapshot is out of sync with the lazily loaded routes."
echo ""
echo "A lazily loaded route or model changed without regenerating the snapshot that /openapi.json serves for unloaded features."
echo "To fix, run from the repo root:"
echo " uv run python -m litellm.proxy._lazy_openapi_snapshot"
echo "then run npm run gen:api from ui/litellm-dashboard and commit both files."
exit 1
fi
echo "_lazy_openapi_snapshot.json is in sync with the lazily loaded routes."
- name: Set up Node.js
if: steps.changes.outputs.relevant == 'true'
uses: actions/setup-node@a0853c24544627f65ddf259abe73b1d18a591444 # v5.0.0

40
.github/workflows/check_duplicate_issues.yml vendored
View file

@ -1,12 +1,19 @@
name: Check Duplicate Issues
# Flagging only. "Auto-close duplicate issues" closes a flagged issue 3 days later,
# and only when its title is identical to an older open issue and nobody replied.
# The HTML marker below is the handshake between the two, so keep it in the template.
on:
issues:
types: [opened, edited]
permissions: {}
jobs:
check-duplicate:
runs-on: ubuntu-latest
timeout-minutes: 5
permissions:
issues: write
contents: read
@ -19,35 +26,12 @@ jobs:
threshold: 0.6
reaction: eyes
comment: |
**⚠️ Potential duplicate detected**
<!-- litellm:potential-duplicate candidates={{#issues}}{{number}},{{/issues}} -->
**Potential duplicate detected**
This issue appears similar to existing issue(s):
This looks similar to:
{{#issues}}
- [#{{number}}]({{html_url}}) - {{title}} ({{accuracy}}% similar)
- #{{number}} - {{title}}
{{/issues}}
Please review the linked issue(s) to see if they address your concern. If this is not a duplicate, please provide additional context to help us understand the difference.
- name: Checkout close script
if: github.event.action == 'opened'
uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
with:
sparse-checkout: .github/scripts
persist-credentials: false
- name: Set up Python
if: github.event.action == 'opened'
uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
with:
python-version: "3.12"
- name: Auto-close if high-confidence duplicate
if: github.event.action == 'opened'
env:
GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
run: |
python3 .github/scripts/close_duplicate_issues.py \
--issue-number ${{ github.event.issue.number }} \
--repo ${{ github.repository }} \
--threshold 0.85 \
--close
If this is a duplicate, add a thumbs-up reaction to the existing issue and follow along there. When the title is identical to an older open issue, this issue closes automatically in 3 days unless someone responds. If it is not a duplicate, comment here or add a thumbs-down reaction to this comment and it stays open.

22
.github/workflows/codspeed.yml vendored
View file

@ -12,6 +12,7 @@ on:
- "uv.lock"
- ".github/workflows/codspeed.yml"
- ".github/actions/setup-uv-with-retries/**"
- ".github/actions/cache-cargo-build/**"
pull_request:
branches:
- main
@ -23,6 +24,7 @@ on:
- "uv.lock"
- ".github/workflows/codspeed.yml"
- ".github/actions/setup-uv-with-retries/**"
- ".github/actions/cache-cargo-build/**"
# Allow CodSpeed to trigger backtest performance analysis
# in order to generate initial data
workflow_dispatch:
@ -55,6 +57,26 @@ jobs:
with:
version: "0.10.9"
- name: Cache the Rust build
uses: ./.github/actions/cache-cargo-build
# Build the wheel and resolve every dependency outside the CodSpeed
# runner: the same maturin build took 42 minutes inside `codspeed run`
# versus under 3 minutes as a plain step (LIT-6183)
- name: Build environment
run: >
env PYTEST_DISABLE_PLUGIN_AUTOLOAD=1
uv run --frozen --no-default-groups
--with pytest==8.3.5
--with pytest-codspeed==4.3.0
--with "mcp>=1.26.0,<2.0"
--with "a2a-sdk>=1.1.0,<2.0"
pytest
-p pytest_codspeed.plugin
tests/benchmarks/
--codspeed
--collect-only -q
- name: Run benchmarks
uses: CodSpeedHQ/action@1c8ae4843586d3ba879736b7f6b7b0c990757fab # v4.12.1
with:

237
.github/workflows/e2e_record_replay.yml vendored Normal file
View file

@ -0,0 +1,237 @@
name: "E2E Record and Replay"
on:
schedule:
- cron: "0 8 * * 6"
- cron: "0 8 * * 1-5"
workflow_dispatch:
inputs:
mode:
description: "record (hits real providers and publishes a fresh bundle) or replay (bundle only, zero provider egress)"
type: choice
options:
- record
- replay
default: record
permissions:
contents: read
jobs:
record:
name: "Record the e2e suite against real providers"
if: >-
(github.event_name != 'schedule' || github.repository == 'BerriAI/litellm') &&
(github.event.schedule == '0 8 * * 6' ||
(github.event_name == 'workflow_dispatch' && github.event.inputs.mode == 'record'))
runs-on: ubuntu-latest
timeout-minutes: 45
services:
postgres:
image: postgres:16.6
env:
POSTGRES_USER: llmproxy
POSTGRES_PASSWORD: dbpassword9090
POSTGRES_DB: litellm
ports:
- 5432:5432
options: >-
--health-cmd "pg_isready -U llmproxy"
--health-interval 5s
--health-timeout 5s
--health-retries 10
env:
DATABASE_URL: postgresql://llmproxy:dbpassword9090@localhost:5432/litellm
LITELLM_MASTER_KEY: sk-e2e-record-replay
LITELLM_LOCAL_MODEL_COST_MAP: "True"
OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
steps:
- uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
with:
persist-credentials: false
- 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 the Rust build
uses: ./.github/actions/cache-cargo-build
- name: Install dependencies
run: |
.github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra proxy
- name: Cache Prisma binaries
uses: ./.github/actions/cache-prisma-binaries
- name: Generate Prisma client
run: |
uv run --no-sync prisma generate --schema litellm/proxy/schema.prisma
- name: Start the proxy
run: |
nohup uv run --no-sync litellm --config tests/e2e/gateway/record_replay_ci_config.yml --port 4000 > proxy.log 2>&1 &
for _ in $(seq 1 90); do
if curl -fs http://localhost:4000/health/liveliness > /dev/null; then
exit 0
fi
sleep 2
done
echo "proxy never became live"
tail -n 100 proxy.log
exit 1
- name: Record the replayable e2e lane
env:
E2E_FIXTURE_MODE: record
run: |
uv run --no-sync pytest tests/e2e -m replayable --reruns 0 -v --tb=short -rA
- name: Pack the fixture bundle
run: |
.github/scripts/e2e_pack_fixture_bundle.sh tests/e2e/.fixtures "${RUNNER_TEMP}/bundle/e2e-fixtures.tar.gz"
- name: Publish the fixture bundle
uses: actions/upload-artifact@4cec3d8aa04e39d1a68397de0c4cd6fb9dce8ec1 # v4.6.1
with:
name: e2e-fixtures-bundle
path: |
${{ runner.temp }}/bundle/e2e-fixtures.tar.gz
${{ runner.temp }}/bundle/e2e-fixtures.tar.gz.sha256
if-no-files-found: error
retention-days: 30
- name: Show proxy log on failure
if: failure()
run: tail -n 300 proxy.log
replay:
name: "Replay the e2e suite from the pinned bundle with zero egress"
if: >-
(github.event_name != 'schedule' || github.repository == 'BerriAI/litellm') &&
(github.event.schedule == '0 8 * * 1-5' ||
(github.event_name == 'workflow_dispatch' && github.event.inputs.mode == 'replay'))
runs-on: ubuntu-latest
timeout-minutes: 45
permissions:
contents: read
actions: read
services:
postgres:
image: postgres:16.6
env:
POSTGRES_USER: llmproxy
POSTGRES_PASSWORD: dbpassword9090
POSTGRES_DB: litellm
ports:
- 5432:5432
options: >-
--health-cmd "pg_isready -U llmproxy"
--health-interval 5s
--health-timeout 5s
--health-retries 10
env:
DATABASE_URL: postgresql://llmproxy:dbpassword9090@localhost:5432/litellm
LITELLM_MASTER_KEY: sk-e2e-record-replay
LITELLM_LOCAL_MODEL_COST_MAP: "True"
GH_TOKEN: ${{ github.token }}
OPENAI_API_KEY: sk-replay-must-never-reach-a-provider
ANTHROPIC_API_KEY: sk-ant-replay-must-never-reach-a-provider
steps:
- uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
with:
persist-credentials: false
- 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 the Rust build
uses: ./.github/actions/cache-cargo-build
- name: Install dependencies
run: |
.github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra proxy
- name: Cache Prisma binaries
uses: ./.github/actions/cache-prisma-binaries
- name: Generate Prisma client
run: |
uv run --no-sync prisma generate --schema litellm/proxy/schema.prisma
- name: Fetch the pinned fixture bundle by digest
env:
BASE_BRANCH: ${{ github.ref_name }}
run: |
.github/scripts/e2e_fetch_fixture_bundle.sh \
"${GITHUB_REPOSITORY}" \
e2e-fixtures-bundle \
"${BASE_BRANCH}" \
tests/e2e/.fixtures
- name: Start the proxy
run: |
nohup uv run --no-sync litellm --config tests/e2e/gateway/record_replay_ci_config.yml --port 4000 > proxy.log 2>&1 &
for _ in $(seq 1 90); do
if curl -fs http://localhost:4000/health/liveliness > /dev/null; then
exit 0
fi
sleep 2
done
echo "proxy never became live"
tail -n 100 proxy.log
exit 1
- name: Start the egress sentinel
run: |
# shellcheck disable=SC2024 # the log redirect is deliberately the runner user's, so a later non-sudo cat can read it
sudo python3 .github/scripts/e2e_egress_sentinel.py serve \
--host api.openai.com \
--host api.anthropic.com \
--hits-file "${RUNNER_TEMP}/egress-hits.jsonl" \
--ready-file "${RUNNER_TEMP}/egress-ready" \
--pid-file "${RUNNER_TEMP}/egress.pid" \
> "${RUNNER_TEMP}/egress-sentinel.log" 2>&1 &
for _ in $(seq 1 30); do
if [[ -f "${RUNNER_TEMP}/egress-ready" ]]; then
cat "${RUNNER_TEMP}/egress-sentinel.log"
exit 0
fi
sleep 1
done
echo "egress sentinel never became ready"
cat "${RUNNER_TEMP}/egress-sentinel.log"
exit 1
- name: Replay the replayable e2e lane
env:
E2E_FIXTURE_MODE: replay
run: |
uv run --no-sync pytest tests/e2e -m replayable --reruns 0 -v --tb=short -rA
- name: Stop the egress sentinel and assert zero provider egress
if: always()
run: |
if [[ -f "${RUNNER_TEMP}/egress.pid" ]]; then
sudo kill -TERM "$(cat "${RUNNER_TEMP}/egress.pid")" 2>/dev/null || true
sleep 2
fi
python3 .github/scripts/e2e_egress_sentinel.py assert-empty --hits-file "${RUNNER_TEMP}/egress-hits.jsonl"
- name: Show proxy log on failure
if: failure()
run: tail -n 300 proxy.log

39
.github/workflows/image-scan.yml vendored
View file

@ -17,10 +17,14 @@ on:
- backend/Dockerfile
- backend/main.py
- docker/component_entrypoint.sh
- docker/entrypoint.sh
- litellm/proxy/prisma_migration.py
- litellm-proxy-extras/**
- tests/proxy_migration_tests/**
- uv.lock
- ui/litellm-dashboard/package-lock.json
- ui/Dockerfile
- ui/nginx.conf
- .github/workflows/image-scan.yml
schedule:
- cron: "41 6 * * *"
@ -76,7 +80,7 @@ jobs:
LITELLM_IMAGE: litellm-image-scan:${{ github.sha }}
run: |
python -m pip install "pytest==9.0.3"
python -m pytest tests/proxy_migration_tests/test_offline_image_migration.py -v
python -m pytest tests/proxy_migration_tests/test_offline_image_migration.py tests/proxy_migration_tests/test_image_bedrock_realtime_extra.py -v
# Scans the whole shipped artifact: OS/apk plus every language package
# baked into the image, including ones no lockfile declares (e.g. prisma's
@ -120,7 +124,7 @@ jobs:
LITELLM_IMAGE: litellm-runtime-scan:${{ github.sha }}
run: |
python -m pip install "pytest==9.0.3"
python -m pytest tests/proxy_migration_tests/test_offline_image_migration.py -v
python -m pytest tests/proxy_migration_tests/test_offline_image_migration.py tests/proxy_migration_tests/test_image_bedrock_realtime_extra.py -v
migrations-image:
name: migrations-image
@ -181,7 +185,36 @@ jobs:
LITELLM_COMPONENT_PORT: "4000"
run: |
python -m pip install "pytest==9.0.3"
python -m pytest tests/proxy_migration_tests/test_component_image_serves_offline.py -v
python -m pytest tests/proxy_migration_tests/test_component_image_serves_offline.py tests/proxy_migration_tests/test_image_bedrock_realtime_extra.py -v
ui-image:
name: ui-image
runs-on: ubuntu-latest
if: >-
github.event_name != 'pull_request' ||
github.event.pull_request.head.repo.full_name == github.repository
timeout-minutes: 30
permissions:
contents: read
steps:
- uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
with:
persist-credentials: false
- name: Build UI image
run: docker build -f ui/Dockerfile -t litellm-ui-scan:${{ github.sha }} .
- name: Set up Python
uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
with:
python-version: "3.12"
- name: Verify the UI serves offline as an arbitrary uid with a read-only root fs
env:
LITELLM_IMAGE: litellm-ui-scan:${{ github.sha }}
run: |
python -m pip install "pytest==9.0.3"
python -m pytest tests/proxy_migration_tests/test_ui_image_serves_offline.py -v
backend-image:
name: backend-image

13
.github/workflows/mutation-test.yml vendored
View file

@ -53,6 +53,9 @@ jobs:
restore-keys: |
${{ runner.os }}-uv-
- name: Cache the Rust build
uses: ./.github/actions/cache-cargo-build
- name: Install dependencies
run: |
.github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra google --extra proxy --extra semantic-router --extra saml
@ -84,11 +87,20 @@ jobs:
run: |
uv pip uninstall pytest-retry || true
# Ends before the job's own deadline so a run that outlasts the budget is
# still followed by the report and upload steps. mutmut saves after every
# mutant result, to mutants/<source path>.meta, so an interrupted run
# still scores the mutants it finished and export-cicd-stats can read
# them; a cancelled job skips those steps and publishes nothing at all.
- name: Run mutmut
timeout-minutes: 300
env:
# Make the mutants/ sandbox win over site-packages on sys.path so the
# trampolined files are imported instead of the installed copy.
PYTHONPATH: ${{ github.workspace }}/mutants
# Without this mutmut finds no covered lines and generates 0 mutants.
# See the file itself for why.
COVERAGE_RCFILE: ${{ github.workspace }}/.github/mutmut-coverage.rc
run: |
set -o pipefail
mkdir -p mutants
@ -127,6 +139,7 @@ jobs:
mutmut-run.log
mutants/mutmut-stats.json
mutants/mutmut-cicd-stats.json
mutants/**/*.meta
mutants/litellm/proxy/management_endpoints/**/*.py
if-no-files-found: warn
retention-days: 14

3
.github/workflows/publish-basedpyright-base-counts.yml vendored
View file

@ -43,6 +43,9 @@ jobs:
with:
version: "0.10.9"
- name: Cache the Rust build
uses: ./.github/actions/cache-cargo-build
- name: Cache Prisma binaries
uses: ./.github/actions/cache-prisma-binaries

68
.github/workflows/sync-together-ai-models.yml vendored Normal file
View file

@ -0,0 +1,68 @@
name: Sync Together AI model registry
on:
schedule:
- cron: "30 6 * * *"
workflow_dispatch:
permissions:
contents: write
pull-requests: write
jobs:
sync_together_ai_models:
if: github.repository == 'BerriAI/litellm'
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
with:
ref: litellm_internal_staging
persist-credentials: false
- name: Set up uv
uses: ./.github/actions/setup-uv-with-retries
with:
version: "0.10.9"
- name: Look for an already-open sync PR
id: existing
run: |
open_pr="$(gh pr list --repo "$GITHUB_REPOSITORY" --state open --limit 1000 --json headRefName \
--jq '[.[].headRefName | select(startswith("litellm_together_registry_sync_"))] | first // empty')"
echo "open_pr=$open_pr" >> "$GITHUB_OUTPUT"
if [ -n "$open_pr" ]; then
echo "An open sync PR already exists on branch $open_pr; skipping this run."
fi
env:
GH_TOKEN: ${{ secrets.GH_TOKEN || github.token }}
- name: Run the sync
if: steps.existing.outputs.open_pr == ''
run: |
uv run --frozen python scripts/sync_together_ai_models.py --write --pr-body-file "$RUNNER_TEMP/pr_body.md"
env:
TOGETHER_API_KEY: ${{ secrets.TOGETHER_API_KEY }}
- name: Regenerate the JSON schema
if: steps.existing.outputs.open_pr == ''
run: |
uv run --frozen python ci_cd/generate_model_prices_schema.py
- name: Create a pull request when the registry changed
if: steps.existing.outputs.open_pr == ''
run: |
if git diff --quiet; then
echo "Registry already in sync; no PR needed."
exit 0
fi
branch="litellm_together_registry_sync_$(date +'%Y-%m-%d')"
git config user.name "github-actions[bot]"
git config user.email "41898282+github-actions[bot]@users.noreply.github.com"
git checkout -b "$branch"
git add model_prices_and_context_window.json \
litellm/model_prices_and_context_window_backup.json \
model_prices_and_context_window.schema.json
git commit -m "feat(models): sync together_ai model registry $(date +'%Y-%m-%d')"
gh auth setup-git
git push origin "$branch"
gh pr create --title "feat(models): sync together_ai model registry" \
--body-file "$RUNNER_TEMP/pr_body.md" \
--head "$branch" \
--base litellm_internal_staging
env:
GH_TOKEN: ${{ secrets.GH_TOKEN || github.token }}

6
.github/workflows/test-code-quality.yml vendored
View file

@ -56,6 +56,9 @@ jobs:
restore-keys: |
${{ runner.os }}-uv-
- name: Cache the Rust build
uses: ./.github/actions/cache-cargo-build
- name: Install dependencies
run: uv sync --frozen --all-groups --all-extras
@ -128,6 +131,9 @@ jobs:
- name: check_e2e_no_raw_requests
run: uv run --no-sync python ./tests/code_coverage_tests/check_e2e_no_raw_requests.py
- name: check_migrations_no_data_rewrites
run: uv run --no-sync python ./tests/code_coverage_tests/check_migrations_no_data_rewrites.py
- name: memory_test
run: uv run --no-sync python ./tests/code_coverage_tests/memory_test.py

15
.github/workflows/test-linting.yml vendored
View file

@ -67,6 +67,17 @@ jobs:
with:
version: "0.10.9"
- name: Cache uv dependencies
if: steps.changes.outputs.decision != 'skip'
uses: actions/cache@0057852bfaa89a56745cba8c7296529d2fc39830 # v4.3.0
with:
path: |
~/.cache/uv
.venv
key: ${{ runner.os }}-uv-lint-${{ hashFiles('uv.lock') }}
restore-keys: |
${{ runner.os }}-uv-lint-
- name: Clean Python cache
if: steps.changes.outputs.decision != 'skip'
run: |
@ -78,6 +89,10 @@ jobs:
run: |
uv lock --check || (echo "❌ uv.lock is out of sync with pyproject.toml. Run 'uv lock' locally and commit the result." && exit 1)
- name: Cache the Rust build
if: steps.changes.outputs.decision != 'skip'
uses: ./.github/actions/cache-cargo-build
- name: Install dependencies
if: steps.changes.outputs.decision != 'skip'
run: |

6
.github/workflows/test-mcp.yml vendored
View file

@ -47,6 +47,10 @@ jobs:
with:
version: "0.10.9"
- name: Cache the Rust build
if: steps.changes.outputs.decision != 'skip'
uses: ./.github/actions/cache-cargo-build
- name: Install dependencies
if: steps.changes.outputs.decision != 'skip'
run: |
@ -56,4 +60,4 @@ jobs:
- name: Run MCP tests
if: steps.changes.outputs.decision != 'skip'
run: |
uv run --no-sync pytest tests/mcp_tests -x -vv -n 4 --cov=./litellm --cov-report=xml --durations=5
uv run --no-sync pytest tests/mcp_tests -x -vv -n 4 --cov=./litellm --cov=./enterprise/litellm_enterprise --cov-report=xml --durations=5

145
.github/workflows/test-postgres.yml vendored Normal file
View file

@ -0,0 +1,145 @@
name: "Postgres Tests"
on:
pull_request:
branches:
- main
- litellm_internal_staging
- litellm_oss_staging
- "litellm_**"
push:
branches:
- main
- litellm_internal_staging
workflow_dispatch:
permissions:
contents: read
concurrency:
group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}
cancel-in-progress: ${{ github.event_name == 'pull_request' }}
jobs:
postgres:
name: ${{ matrix.shard }}
runs-on: ubuntu-latest
timeout-minutes: ${{ matrix.job-timeout-minutes }}
permissions:
contents: read
services:
postgres:
image: postgres:16@sha256:e17e86066e5ef83e0952a9347f5c792b7ece00972e2aa787a6986f471b3dd3d5
env:
POSTGRES_USER: postgres
POSTGRES_PASSWORD: postgres
POSTGRES_DB: litellm_test
ports:
- 5432:5432
options: >-
--health-cmd pg_isready
--health-interval 10s
--health-timeout 5s
--health-retries 10
strategy:
fail-fast: false
matrix:
include:
- shard: proxy-behavior
test-path: "tests/proxy_behavior"
seed: db-push
workers: 0
timeout-minutes: 25
job-timeout-minutes: 50
- shard: proxy-security
test-path: "tests/proxy_security_tests"
seed: db-push
workers: 0
timeout-minutes: 15
job-timeout-minutes: 40
- shard: schema-migration
test-path: "tests/proxy_migration_tests"
seed: none
workers: 0
timeout-minutes: 20
job-timeout-minutes: 45
env:
DATABASE_URL: "postgresql://postgres:postgres@localhost:5432/litellm_test"
steps:
- uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
timeout-minutes: 3
with:
persist-credentials: false
- name: Detect relevant changes
id: changes
timeout-minutes: 2
uses: ./.github/actions/detect-changes
- name: Set up Python
if: steps.changes.outputs.decision != 'skip'
timeout-minutes: 3
uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
with:
python-version: "3.12"
- name: Set up uv
if: steps.changes.outputs.decision != 'skip'
timeout-minutes: 3
uses: ./.github/actions/setup-uv-with-retries
with:
version: "0.10.9"
- name: Cache uv dependencies
if: steps.changes.outputs.decision != 'skip'
timeout-minutes: 5
uses: actions/cache@0057852bfaa89a56745cba8c7296529d2fc39830 # v4.3.0
with:
path: |
~/.cache/uv
.venv
key: ${{ runner.os }}-uv-postgres-${{ hashFiles('uv.lock') }}
restore-keys: |
${{ runner.os }}-uv-postgres-
- name: Install dependencies
if: steps.changes.outputs.decision != 'skip'
timeout-minutes: 12
run: |
.github/scripts/uv_sync_with_retries.sh --frozen --all-groups --all-extras
- name: Cache Prisma binaries
if: steps.changes.outputs.decision != 'skip'
timeout-minutes: 3
uses: ./.github/actions/cache-prisma-binaries
- name: Generate Prisma client
if: steps.changes.outputs.decision != 'skip'
timeout-minutes: 5
run: |
uv run --no-sync prisma generate --schema litellm/proxy/schema.prisma
- name: Seed database schema
if: steps.changes.outputs.decision != 'skip' && matrix.seed != 'none'
timeout-minutes: 10
run: |
uv run --no-sync prisma db push --schema litellm/proxy/schema.prisma --accept-data-loss
- name: Run tests
if: steps.changes.outputs.decision != 'skip'
timeout-minutes: ${{ matrix.timeout-minutes }}
env:
TEST_PATH: ${{ matrix.test-path }}
WORKERS: ${{ matrix.workers }}
run: |
if [ "${WORKERS}" = "0" ]; then
uv run --no-sync pytest ${TEST_PATH:?} -vv --tb=short --durations=10
else
uv run --no-sync pytest ${TEST_PATH:?} -vv --tb=short --durations=10 -n "${WORKERS}"
fi

77
.github/workflows/test-redis-compat.yml vendored Normal file
View file

@ -0,0 +1,77 @@
name: "Unit Tests: Redis Client Version Compatibility"
on:
pull_request:
branches:
- main
- litellm_internal_staging
- litellm_oss_staging
- "litellm_**"
paths:
- "litellm/_redis.py"
- "litellm/_redis_credential_provider.py"
- "tests/test_litellm/test_redis.py"
- "tests/test_litellm/caching/test_redis_connection_pool.py"
- ".github/workflows/test-redis-compat.yml"
- "pyproject.toml"
- "uv.lock"
permissions:
contents: read
concurrency:
group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
cancel-in-progress: true
jobs:
redis-compat:
name: "redis-py ${{ matrix.redis-version }}"
runs-on: ubuntu-latest
timeout-minutes: 15
strategy:
fail-fast: false
matrix:
# 5.3.1 is the version pinned in uv.lock (redisvl caps it below 6); the
# newer legs prove the inspect.signature introspection in litellm/_redis.py
# keeps extracting kwargs on the redis-py releases people actually run now.
# Only the exact release 6.0.0 is skipped: rq (pulled by the proxy extra)
# specifies `redis != 6`, which excludes 6.0.0 alone, so 6.4.0 stands in
# for the 6.x line.
redis-version: ["5.3.1", "6.4.0", "7.4.1", "8.0.1"]
steps:
- uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
with:
persist-credentials: false
- 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: Install dependencies
run: |
.github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra google --extra proxy --extra semantic-router
- name: Pin redis-py to the matrix version
env:
REDIS_VERSION: ${{ matrix.redis-version }}
run: |
uv pip install "redis==${REDIS_VERSION:?}"
uv run --no-sync python -c "import redis; assert redis.__version__ == '${REDIS_VERSION:?}', redis.__version__; print('redis-py', redis.__version__)"
- name: Run redis unit tests
run: |
uv run --no-sync pytest \
tests/test_litellm/test_redis.py \
tests/test_litellm/caching/test_redis_connection_pool.py \
--tb=short -vv \
--reruns 2 \
--reruns-delay 1 \
--durations=20

5
.github/workflows/test-terraform-provider.yml vendored
View file

@ -88,6 +88,9 @@ jobs:
restore-keys: |
${{ runner.os }}-uv-
- name: Cache the Rust build
uses: ./.github/actions/cache-cargo-build
- name: Install dependencies
run: |
.github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra google --extra proxy --extra semantic-router
@ -111,4 +114,4 @@ jobs:
- name: Audit provider endpoints against the schema
working-directory: terraform/provider
run: go run ./tools/endpointaudit -provider-dir ./litellm -spec "${RUNNER_TEMP}/openapi.json"
run: go run ./tools/endpointaudit -provider-dir ./litellm -spec "${RUNNER_TEMP}/openapi.json" -coverage-allowlist ./tools/endpointaudit/coverage_allowlist.txt

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

@ -67,6 +67,10 @@ jobs:
restore-keys: |
${{ runner.os }}-uv-
- name: Cache the Rust build
if: steps.changes.outputs.decision != 'skip'
uses: ./.github/actions/cache-cargo-build
- name: Install dependencies
if: steps.changes.outputs.decision != 'skip'
run: |

56
.github/workflows/test-unit.yml vendored
View file

@ -55,7 +55,7 @@ jobs:
workers: 2
reruns: 1
timeout-minutes: 20
job-timeout-minutes: 55
job-timeout-minutes: 60
- shard: enterprise-routing
artifact-name: enterprise-routing
@ -67,7 +67,7 @@ jobs:
workers: 2
reruns: 2
timeout-minutes: 20
job-timeout-minutes: 55
job-timeout-minutes: 60
- shard: integrations
artifact-name: integrations
@ -75,7 +75,7 @@ jobs:
workers: 2
reruns: 3
timeout-minutes: 20
job-timeout-minutes: 55
job-timeout-minutes: 60
- shard: Vertex AI
artifact-name: llm-vertex-ai
@ -83,7 +83,7 @@ jobs:
workers: 1
reruns: 2
timeout-minutes: 20
job-timeout-minutes: 55
job-timeout-minutes: 60
- shard: All Other Providers
artifact-name: llm-other-providers
@ -91,7 +91,7 @@ jobs:
workers: 2
reruns: 2
timeout-minutes: 20
job-timeout-minutes: 55
job-timeout-minutes: 60
- shard: misc
artifact-name: misc
@ -103,6 +103,7 @@ jobs:
tests/test_litellm/completion_extras
tests/test_litellm/compression
tests/test_litellm/containers
tests/test_litellm/endpoints
tests/test_litellm/experimental_mcp_client
tests/test_litellm/models
tests/test_litellm/repositories
@ -122,7 +123,7 @@ jobs:
workers: 2
reruns: 2
timeout-minutes: 20
job-timeout-minutes: 55
job-timeout-minutes: 60
- shard: proxy-auth
artifact-name: proxy-auth
@ -134,13 +135,14 @@ jobs:
workers: 2
reruns: 2
timeout-minutes: 20
job-timeout-minutes: 55
job-timeout-minutes: 60
- shard: proxy-endpoints
artifact-name: proxy-endpoints
test-path: >-
tests/test_litellm/proxy/analytics_endpoints
tests/test_litellm/proxy/management_endpoints
tests/test_litellm/proxy/list_api
tests/test_litellm/proxy/memory
tests/test_litellm/proxy/guardrails
tests/test_litellm/proxy/management_helpers
@ -164,14 +166,15 @@ jobs:
tests/test_litellm/proxy/public_endpoints
tests/test_litellm/proxy/prompts
tests/test_litellm/proxy/rag_endpoints
tests/test_litellm/proxy/rerank_endpoints
tests/test_litellm/proxy/realtime_endpoints
tests/test_litellm/proxy/ui_crud_endpoints
tests/test_litellm/proxy/config_resolvers
tests/test_litellm/proxy/utils
workers: 2
workers: 4
reruns: 2
timeout-minutes: 20
job-timeout-minutes: 55
job-timeout-minutes: 60
- shard: proxy-server
artifact-name: proxy-server
@ -179,7 +182,7 @@ jobs:
workers: 4
reruns: 2
timeout-minutes: 60
job-timeout-minutes: 95
job-timeout-minutes: 100
- shard: proxy-infra
artifact-name: proxy-infra
@ -195,10 +198,39 @@ jobs:
tests/test_litellm/proxy/types_utils
tests/test_litellm/proxy/logging_endpoints
tests/test_litellm/proxy/test_*.py
workers: 4
reruns: 2
timeout-minutes: 20
job-timeout-minutes: 60
- shard: caching-local
artifact-name: caching-local
test-path: >-
tests/local_testing/test_cache_preset_key.py
tests/local_testing/test_caching_handler.py
tests/local_testing/test_prompt_caching.py
tests/local_testing/test_responses_stream_cache_keys.py
tests/local_testing/test_unit_test_caching.py
workers: 2
reruns: 2
timeout-minutes: 20
job-timeout-minutes: 55
job-timeout-minutes: 60
- shard: proxy-extras
artifact-name: proxy-extras
test-path: "tests/litellm-proxy-extras"
workers: 2
reruns: 2
timeout-minutes: 20
job-timeout-minutes: 60
- shard: enterprise-package
artifact-name: enterprise-package
test-path: "tests/enterprise"
workers: 4
reruns: 2
timeout-minutes: 20
job-timeout-minutes: 60
- shard: responses-caching-types
artifact-name: responses-caching-types
@ -209,7 +241,7 @@ jobs:
workers: 2
reruns: 2
timeout-minutes: 20
job-timeout-minutes: 55
job-timeout-minutes: 60
uses: ./.github/workflows/_test-unit-base.yml
with:
test-path: ${{ matrix.test-path }}

3
.github/workflows/weekly_load_anomaly.yml vendored
View file

@ -47,6 +47,9 @@ jobs:
with:
version: "0.10.9"
- name: Cache the Rust build
uses: ./.github/actions/cache-cargo-build
- name: Install dependencies
run: |
.github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra proxy

2
.gitignore vendored
View file

@ -3,6 +3,8 @@
tests/e2e/.fixtures/
.venv-typecheck
.venv_policy_test
.venv-mutmut
mutants/
.env
.claude
CLAUDE.local.md

11
CLAUDE.md
View file

@ -23,6 +23,8 @@ When adding new features, add meaningful tests. Don't add tests that don't check
Same thing for bug fixes. The tests should make it so that this specific bug can never happen again without failing tests (i.e., regression)
Never test structure of code only function of it
`tests/test_litellm/` mirrors `litellm/` in a parallel path (see `tests/test_litellm/readme.md`). Name tests `test_<filename>.py`, but always match the existing test file in the directory you touch — many provider dirs use longer descriptive names (e.g. `test_anthropic_chat_transformation.py`) to avoid ambiguity across sibling folders. For bug fixes, extend the existing mapped test file rather than creating a new one. Only create a new test file for a new feature (provider, endpoint, or transformation module) that has no mapped test yet, following that directory's naming convention (or `test_<filename>.py` if you're the first test there). One focused regression test beats many shallow ones
End-to-end tests belong in `tests/e2e/` and must follow the harness conventions documented in that directory's `CLAUDE.md`
@ -37,13 +39,14 @@ If you're resolving a linear ticket, in the "## Linear ticket" section of the PR
Never use `pytest` commands or the like as "Screenshots / Proof of Fix". We prefer curl'ing a live proxy instance running on localhost:4000 (I like to run it with `python litellm/proxy/proxy_cli.py --config litellm/proxy/dev_config.yaml --detailed_debug --reload --use_v2_migration_resolver 2>&1 | tee litellm.log`; the Admin UI dev server is `npm run dev` in `ui/litellm-dashboard`, served on port 3000) and showing both the command run and the output. Also, it should hit real LLM provider APIs, not mocks, and cost real $$$ because that is the most realistic test. The proof of fix should be exactly what the end user / customer would see / do. The run logs in PR #27703 is a prime example of how to do it (not a huge fan of using a python test script that future me and the team will have no visibility into; I prefer just curl commands or a short list of bash commands (e.g., using `for`)). If it's a UI thing, just tell me which URLs to go to (e.g., http://localhost:4000/ui/?page=logs), where to click, what fields to fill out, etc. along with the other commands to run in an ordered list, and I'll do it myself and post the screenshots after you make the PR
If you ever make public-facing PR descriptions, comments, issues, commit messages, etc., always follow these guidelines to sound less AI-y:
If you ever write any human-facing text (pull requests, issues, commit messages, discussion posts, github comments, release notes, docs, etc.), always follow these guidelines to sound less AI-y:
- don't use emojis
- don't use "—". Instead, reach for ",", ".", conjunction words, ":", ";", etc. in descending order of preference: vary among them, weighted toward the front of the list, and skip "," where it would cause a comma splice or the sentence is getting long. Overusing any one of them, ";" especially, also feels AI-y. A word cap does not penalize you for adding more sentences: when writing under tight word budgets, prefer a period split or a conjunction over ";", and keep to at most one ";" per message
- don't use the pattern "It's not X, it's Y", "You're not X, you're Y", etc.
- don't use bulleted or numbered lists unless it would be nonsensical not to. Instead, prefer prose
- unless explicitly asked, don't use bulleted or numbered lists unless it would be nonsensical not to. Instead, prefer prose
- don't add a trailing "." at the end of paragraphs (just like this file). That means every paragraph, not just the last one (of the markdown file, PR description, GitHub comment, etc.). Rule of thumb: if you're adding new line(s) before the next sentence, don't add a "."
- don't use →. Instead, prefer not to use arrows, and if need be, use -> instead
- use plain, simple, everyday engineering language: the common phrase engineers actually say over rare compact phrasing, in grammatically complete sentences. When explicitly asked to use bullets or ordered lists and structure legitimately helps the reader, prefer nested bullets (any depth is fine) over dense lines in a flat structure
Don't hesitate to use values in .env to get needed API keys and other secrets, as long as you never add them to conversation history, commit them, or include them in GitHub issues / PRs
@ -65,6 +68,8 @@ Commit and push your work when you're done without asking
When referencing or running models (coding, QA'ing, writing docs, writing tests, etc.), use the latest model in that model family unless otherwise specified; treat your training knowledge, memories, configs, and tests as stale, and determine the family's latest with model_prices_and_context_window.json or the web
Always pull before starting any work. The checkout or worktree may be sitting on a stale branch
If you're an internal contributor, when creating a new PR, the typical flow is to branch off litellm_internal_staging and create a branch prefixed with litellm_. Do not create a branch prefixed with claude/ and generally do not have / in your branch names
Do not add `Co-Authored-By: Claude` or any Claude attribution to commit messages. Never use a `claude/` prefix or put a `/` in a branch name. Do not add "Generated with Claude Code" (or any similar attribution) to PR descriptions or comments. Do not create a new PR/branch off the existing PR to fix/add something that is related and could've just been committed directly to the existing PR's branch
@ -79,6 +84,8 @@ Do not put names of customers or customer company names in code, PR descriptions
CI supply-chain safety: Never pipe a remote script into a shell (`curl ... | bash`, `wget ... | sh`); download the artifact to a file, verify its SHA-256 checksum, then install. Pin every external tool to a specific version with a full URL (not `latest` or `stable`). Verify checksums for all downloaded binaries, using the provider's official `.sha256` / `.sha256sum` sidecar when available. These rules apply to every download in CI
Prisma migrations apply synchronously at proxy boot, before it serves traffic, so a migration must only change schema, never rewrite rows. No `UPDATE`, `DELETE` or `MERGE`, and no `INSERT ... SELECT`: on a spend-log-sized table any of those is minutes of downtime plus a doubled heap that plain autovacuum won't give back. `tests/code_coverage_tests/check_migrations_no_data_rewrites.py` enforces this. When a rewrite is genuinely bounded and has to ship inside the migration, mark the statement `-- data-migration-ok: <what bounds it>`
Follow these coding conventions for new/updated code (a three-line fix in a legacy file shouldn't trigger huge drive-by refactors):
- Composition over inheritance

23
Dockerfile
View file

@ -1,10 +1,10 @@
# syntax=docker/dockerfile:1.7
# Base image for building
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
# Runtime image
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
# Pinned by digest like the other base images; bump explicitly on Node upgrades.
ARG UI_BUILD_IMAGE=node:24.19-alpine3.24@sha256:d32cdf619f63fe0471182d08996dd516c6275bb5fd31ae06e55a570bd9e1ad43
@ -40,8 +40,8 @@ COPY --from=uvbin /uvx /usr/local/bin/uvx
RUN apk add --no-cache \
bash \
gcc \
python3 \
python3-dev \
python-3.13 \
python-3.13-dev \
rust \
openssl \
openssl-dev \
@ -51,6 +51,7 @@ RUN apk add --no-cache \
ENV UV_PROJECT_ENVIRONMENT=/app/.venv \
UV_LINK_MODE=copy \
UV_PYTHON_DOWNLOADS=0 \
PATH="/app/.venv/bin:${PATH}"
# Copy dependency metadata first for layer caching
@ -65,7 +66,8 @@ RUN uv sync --frozen --no-install-project --no-install-workspace --no-default-gr
--extra extra_proxy \
--extra semantic-router \
--extra saml \
--python python3
--extra bedrock-realtime \
--python python3.13
# Copy full source tree
COPY . .
@ -86,7 +88,8 @@ RUN uv sync --frozen --no-default-groups --no-editable \
--extra extra_proxy \
--extra semantic-router \
--extra saml \
--python python3
--extra bedrock-realtime \
--python python3.13
RUN HOME=/opt/prisma XDG_CACHE_HOME=/opt/prisma/.cache PRISMA_BINARY_CACHE_DIR=/opt/prisma/binaries \
npm_config_cache=/root/.npm \
@ -100,8 +103,14 @@ FROM $LITELLM_RUNTIME_IMAGE AS runtime
USER root
# The base image only configures Chainguard's authenticated apk repo, which
# requires an enterprise subscription. Add the public Wolfi repo so `apk add`
# also works for anyone installing extra packages into a running container.
# https://github.com/BerriAI/litellm/issues/33518
RUN echo "https://packages.wolfi.dev/os" >> /etc/apk/repositories
# node (without npm) is required by the prisma CLI at runtime
RUN apk add --no-cache bash openssl tzdata nodejs python3 libsndfile
RUN apk add --no-cache bash openssl tzdata nodejs python-3.13 libsndfile
WORKDIR /app
ENV PATH="/app/.venv/bin:${PATH}" \

2
README.md
View file

@ -354,6 +354,8 @@ curl -X POST 'http://0.0.0.0:4000/v1/chat/completions' \
| [Petals (`petals`)](https://docs.litellm.ai/docs/providers/petals) | ✅ | ✅ | ✅ | | | | | | | |
| [Pinstripes (`pinstripes`)](https://docs.litellm.ai/docs/providers/pinstripes) | ✅ | ✅ | ✅ | | | | | | | |
| [Predibase (`predibase`)](https://docs.litellm.ai/docs/providers/predibase) | ✅ | ✅ | ✅ | | | | | | | |
| [Qwen AI Platform (`qwen_ai_platform`)](https://docs.litellm.ai/docs/providers/qwencloud) | ✅ | ✅ | ✅ | ✅ | ✅ | | | | | ✅ |
| [QwenCloud (`qwencloud`)](https://docs.litellm.ai/docs/providers/qwencloud) | ✅ | ✅ | ✅ | ✅ | ✅ | | | | | ✅ |
| [Recraft (`recraft`)](https://docs.litellm.ai/docs/providers/recraft) | | | | | ✅ | | | | | |
| [Replicate (`replicate`)](https://docs.litellm.ai/docs/providers/replicate) | ✅ | ✅ | ✅ | | | | | | | |
| [Sagemaker Chat (`sagemaker_chat`)](https://docs.litellm.ai/docs/providers/aws_sagemaker) | ✅ | ✅ | ✅ | | | | | | | |

14
backend/Dockerfile
View file

@ -1,5 +1,5 @@
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
FROM $UV_IMAGE AS uvbin
@ -16,7 +16,7 @@ COPY --from=uvbin /uv /uvx /usr/local/bin/
# instead of nodeenv downloading one whose dynamic deps may not be in Wolfi
# (e.g. Node 26.2.0 needs libatomic). Retry for transient apk.cgr.dev flakes.
RUN for i in 1 2 3; do \
apk add --no-cache bash gcc python3 python3-dev openssl openssl-dev libsndfile nodejs npm && break; \
apk add --no-cache bash gcc python-3.13 python-3.13-dev openssl openssl-dev libsndfile nodejs npm && break; \
[ $i = 3 ] && { echo "apk add failed after 3 retries" >&2; exit 1; }; \
sleep 5; \
done
@ -46,7 +46,8 @@ RUN --mount=type=cache,target=/root/.cache/uv \
--extra proxy-runtime \
--extra extra_proxy \
--extra semantic-router \
--python python3
--extra saml \
--python python3.13
# Stage 2 — copy source and install the project + workspace members.
COPY . .
@ -57,7 +58,8 @@ RUN --mount=type=cache,target=/root/.cache/uv \
--extra proxy-runtime \
--extra extra_proxy \
--extra semantic-router \
--python python3
--extra saml \
--python python3.13
RUN HOME=/opt/prisma XDG_CACHE_HOME=/opt/prisma/.cache PRISMA_BINARY_CACHE_DIR=/opt/prisma/binaries \
npm_config_cache=/root/.npm \
@ -71,7 +73,7 @@ FROM $LITELLM_RUNTIME_IMAGE AS runtime
USER root
RUN for i in 1 2 3; do \
apk add --no-cache bash openssl tzdata python3 libsndfile libatomic && break; \
apk add --no-cache bash openssl tzdata python-3.13 libsndfile libatomic && break; \
[ $i = 3 ] && { echo "apk add failed after 3 retries" >&2; exit 1; }; \
sleep 5; \
done

52
basedpyright-code-budget.json
View file

@ -1,36 +1,36 @@
{
"reportAny": {
"limit": 19955
"limit": 14076
},
"reportArgumentType": {
"limit": 2566
"limit": 2216
},
"reportAssignmentType": {
"limit": 320
"limit": 319
},
"reportAttributeAccessIssue": {
"limit": 488
"limit": 480
},
"reportCallIssue": {
"limit": 114
"limit": 112
},
"reportConstantRedefinition": {
"limit": 40
},
"reportDeprecated": {
"limit": 213
"limit": 211
},
"reportDuplicateImport": {
"limit": 19
},
"reportExplicitAny": {
"limit": 6049
"limit": 4128
},
"reportFunctionMemberAccess": {
"limit": 7
},
"reportGeneralTypeIssues": {
"limit": 154
"limit": 101
},
"reportIncompatibleMethodOverride": {
"limit": 56
@ -42,10 +42,10 @@
"limit": 12
},
"reportIndexIssue": {
"limit": 35
"limit": 25
},
"reportInvalidTypeForm": {
"limit": 35
"limit": 34
},
"reportInvalidTypeVarUse": {
"limit": 2
@ -54,10 +54,10 @@
"limit": 0
},
"reportMissingParameterType": {
"limit": 5663
"limit": 5601
},
"reportMissingTypeArgument": {
"limit": 15555
"limit": 15306
},
"reportMissingTypeStubs": {
"limit": 40
@ -72,7 +72,7 @@
"limit": 0
},
"reportOptionalMemberAccess": {
"limit": 1061
"limit": 0
},
"reportOptionalOperand": {
"limit": 0
@ -84,46 +84,46 @@
"limit": 56
},
"reportPrivateUsage": {
"limit": 1823
"limit": 1808
},
"reportRedeclaration": {
"limit": 8
},
"reportReturnType": {
"limit": 213
"limit": 181
},
"reportTypedDictNotRequiredAccess": {
"limit": 26
"limit": 24
},
"reportUndefinedVariable": {
"limit": 0
},
"reportUnknownArgumentType": {
"limit": 44655
"limit": 44364
},
"reportUnknownLambdaType": {
"limit": 109
},
"reportUnknownMemberType": {
"limit": 39011
"limit": 38350
},
"reportUnknownParameterType": {
"limit": 19885
"limit": 19626
},
"reportUnknownVariableType": {
"limit": 30569
"limit": 29890
},
"reportUnnecessaryCast": {
"limit": 117
"limit": 111
},
"reportUnnecessaryComparison": {
"limit": 699
"limit": 692
},
"reportUnnecessaryContains": {
"limit": 5
},
"reportUnnecessaryIsInstance": {
"limit": 836
"limit": 826
},
"reportUntypedBaseClass": {
"limit": 0
@ -135,12 +135,12 @@
"limit": 21
},
"reportUnusedFunction": {
"limit": 139
"limit": 138
},
"reportUnusedImport": {
"limit": 545
"limit": 543
},
"reportUnusedVariable": {
"limit": 146
"limit": 137
}
}

18
ci_cd/generate_model_prices_schema.py
View file

@ -27,6 +27,7 @@ EXTRA_BOOLEAN_KEYS = frozenset(
"uses_embed_content",
"use_openai_responses_path",
"bedrock_converse_supports_strict_tools",
"thinking_always_on",
}
)
@ -72,6 +73,11 @@ ARRAY_KEYS: dict[str, JsonSchema] = {
"description": "Output modalities the model can produce.",
"items": {"type": "string", "enum": ["text", "image", "audio", "video", "code"]},
},
"reasoning_effort_levels": {
"type": "array",
"description": "Exact reasoning_effort levels this deployment accepts; wins over supports_* flags.",
"items": {"type": "string", "enum": ["none", "minimal", "low", "medium", "high", "xhigh", "max"]},
},
"supported_regions": {
"type": "array",
"description": "Cloud regions the model is available in ('global' or region ids).",
@ -156,6 +162,9 @@ COST_DESCRIPTIONS: dict[str, str] = {
"input_cost_per_token": "USD per prompt token.",
"output_cost_per_token": "USD per generated token.",
"output_cost_per_reasoning_token": "USD per reasoning/thinking token, when billed separately.",
"google_maps_grounding_cost_per_query": (
"USD per Grounding with Google Maps request; billed per query or per prompt per web_search_billing_unit."
),
"cache_creation_input_token_cost": "USD per token written to the provider's prompt cache.",
"cache_read_input_token_cost": "USD per prompt token served from the provider's prompt cache.",
"input_cost_per_token_batches": "USD per prompt token via the provider's batch API.",
@ -211,6 +220,15 @@ def string_key_schemas(modes: tuple) -> dict[str, JsonSchema]:
"description": "Highest reasoning effort the Bedrock output_config accepts for this model.",
"enum": ["low", "medium", "high", "max", "xhigh"],
},
"default_reasoning_effort": {
"type": "string",
"description": (
"Reasoning effort the provider applies when the request omits reasoning_effort. "
"Gates whether a non-default temperature or the top_p/logprobs sampling params are "
"accepted, which hold only when the effort resolves to 'none'."
),
"enum": ["none", "minimal", "low", "medium", "high", "xhigh"],
},
"comment": STRING,
"audio_transcription_config": STRING,
}

2
codecov.yaml
View file

@ -25,6 +25,8 @@ flag_management:
carryforward: false
- name: proxy-db-schema-migration
carryforward: false
- name: circleci
carryforward: false
component_management:
individual_components:

5
db_scripts/partition_spend_logs.sql
View file

@ -10,6 +10,11 @@
-- partitioned, so existing installs are unaffected until you run this.
--
-- IMPORTANT
-- * After partitioning, `prisma db push` (including the proxy's
-- --use_prisma_db_push startup mode) is NOT supported: it tries to rewrite
-- the primary key back to ("request_id"), which Postgres rejects on a
-- partitioned table. The proxy detects this and exits with guidance.
-- Use the default startup path (`prisma migrate deploy`) instead.
-- * Test on a staging copy first and take a backup.
-- * Postgres cannot convert a populated table to partitioned in place, so this
-- renames the old table aside and creates a fresh partitioned table.

17
docker/Dockerfile.database
View file

@ -1,10 +1,10 @@
# syntax=docker/dockerfile:1.7
# Base image for building
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
# Runtime image
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
# Pinned by digest like the other base images; bump explicitly on Node upgrades.
ARG UI_BUILD_IMAGE=node:24.19-alpine3.24@sha256:d32cdf619f63fe0471182d08996dd516c6275bb5fd31ae06e55a570bd9e1ad43
@ -39,8 +39,8 @@ COPY --from=uvbin /uvx /usr/local/bin/uvx
RUN apk add --no-cache \
bash \
gcc \
python3 \
python3-dev \
python-3.13 \
python-3.13-dev \
openssl \
openssl-dev \
nodejs \
@ -49,6 +49,7 @@ RUN apk add --no-cache \
ENV UV_PROJECT_ENVIRONMENT=/app/.venv \
UV_LINK_MODE=copy \
UV_PYTHON_DOWNLOADS=0 \
PATH="/app/.venv/bin:${PATH}"
# Copy dependency metadata first for layer caching
@ -63,7 +64,8 @@ RUN uv sync --frozen --no-install-project --no-install-workspace --no-default-gr
--extra extra_proxy \
--extra semantic-router \
--extra saml \
--python python3
--extra bedrock-realtime \
--python python3.13
# Copy full source tree
COPY . .
@ -84,7 +86,8 @@ RUN uv sync --frozen --no-default-groups --no-editable \
--extra extra_proxy \
--extra semantic-router \
--extra saml \
--python python3
--extra bedrock-realtime \
--python python3.13
RUN HOME=/opt/prisma XDG_CACHE_HOME=/opt/prisma/.cache PRISMA_BINARY_CACHE_DIR=/opt/prisma/binaries \
npm_config_cache=/root/.npm \
@ -98,7 +101,7 @@ FROM $LITELLM_RUNTIME_IMAGE AS runtime
USER root
# node (without npm) is required by the prisma CLI at runtime
RUN apk add --no-cache bash openssl tzdata nodejs python3 libsndfile
RUN apk add --no-cache bash openssl tzdata nodejs python-3.13 libsndfile
WORKDIR /app
ENV PATH="/app/.venv/bin:${PATH}" \

20
docker/Dockerfile.non_root
View file

@ -1,8 +1,8 @@
# syntax=docker/dockerfile:1.7
# Base images
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
ARG PROXY_EXTRAS_SOURCE=published
ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
# Pinned by digest like the other base images; bump explicitly on Node upgrades.
@ -37,8 +37,8 @@ COPY --from=uvbin /uvx /usr/local/bin/uvx
RUN for i in 1 2 3; do \
apk add --no-cache \
python3 \
python3-dev \
python-3.13 \
python-3.13-dev \
gcc \
rust \
bash \
@ -52,6 +52,7 @@ RUN for i in 1 2 3; do \
ENV UV_PROJECT_ENVIRONMENT=/app/.venv \
UV_LINK_MODE=copy \
UV_PYTHON_DOWNLOADS=0 \
PATH="/app/.venv/bin:${PATH}" \
LITELLM_NON_ROOT=true \
XDG_CACHE_HOME=/app/.cache
@ -69,7 +70,8 @@ RUN --mount=type=cache,target=/app/.cache/uv,id=litellm-uv-cache \
--extra extra_proxy \
--extra semantic-router \
--extra saml \
--python python3
--extra bedrock-realtime \
--python python3.13
# Copy full source tree
COPY . .
@ -96,7 +98,8 @@ RUN --mount=type=cache,target=/app/.cache/uv,id=litellm-uv-cache \
--extra extra_proxy \
--extra semantic-router \
--extra saml \
--python python3 \
--extra bedrock-realtime \
--python python3.13 \
--no-sources-package litellm-proxy-extras; \
else \
uv sync --frozen --no-default-groups --no-editable \
@ -105,7 +108,8 @@ RUN --mount=type=cache,target=/app/.cache/uv,id=litellm-uv-cache \
--extra extra_proxy \
--extra semantic-router \
--extra saml \
--python python3; \
--extra bedrock-realtime \
--python python3.13; \
fi
RUN HOME=/opt/prisma XDG_CACHE_HOME=/opt/prisma/.cache PRISMA_BINARY_CACHE_DIR=/opt/prisma/binaries \
@ -124,7 +128,7 @@ RUN for i in 1 2 3; do \
apk upgrade --no-cache && break || sleep 5; \
done && \
for i in 1 2 3; do \
apk add --no-cache python3 bash openssl tzdata libsndfile nodejs && break || sleep 5; \
apk add --no-cache python-3.13 bash openssl tzdata libsndfile nodejs && break || sleep 5; \
done
# Copy only what runtime needs. The application is installed inside the venv;

1
enterprise/enterprise_hooks/banned_keywords.py
View file

@ -21,6 +21,7 @@ from fastapi import HTTPException
class _ENTERPRISE_BannedKeywords(CustomLogger):
enforces_request_content: bool = True
# Class variables or attributes
def __init__(self):
banned_keywords_list = litellm.banned_keywords_list

1
enterprise/enterprise_hooks/blocked_user_list.py
View file

@ -18,6 +18,7 @@ from fastapi import HTTPException
class _ENTERPRISE_BlockedUserList(CustomLogger):
enforces_request_content: bool = True
# Class variables or attributes
def __init__(self, prisma_client: Optional[PrismaClient]):
self.prisma_client = prisma_client

101
enterprise/litellm_enterprise/proxy/audit_logging_endpoints.py
View file

@ -7,7 +7,7 @@ GET - /audit/{id} - Get audit log by id
GET - /audit - Get all audit logs
"""
from typing import Any, Dict, List, Optional
from typing import TYPE_CHECKING, Final
#### AUDIT LOGGING ####
from fastapi import APIRouter, Depends, HTTPException, Query
@ -18,11 +18,16 @@ from litellm_enterprise.types.proxy.audit_logging_endpoints import (
from litellm.proxy._types import CommonProxyErrors, UserAPIKeyAuth
from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
from litellm.repositories.prisma_protocols import TableActions
from litellm.repositories.table_repositories import AuditLogRepository
if TYPE_CHECKING:
from prisma import models as prisma_models
router = APIRouter()
def _build_json_field_or_condition(json_key: str, value: str) -> Dict[str, Any]:
def _build_json_field_or_condition(json_key: str, value: str) -> dict[str, object]:
"""
Build an OR condition that matches a value inside a JSON column at the
given key, checking both before_value and updated_values.
@ -53,33 +58,33 @@ async def get_audit_logs(
page: int = Query(1, ge=1),
page_size: int = Query(10, ge=1, le=100),
# Filter parameters
changed_by: Optional[str] = Query(
changed_by: str | None = Query(
None, description="Filter by user or system that performed the action"
),
changed_by_api_key: Optional[str] = Query(
changed_by_api_key: str | None = Query(
None, description="Filter by API key hash that performed the action"
),
action: Optional[str] = Query(
action: str | None = Query(
None, description="Filter by action type (create, update, delete)"
),
table_name: Optional[str] = Query(
table_name: str | None = Query(
None, description="Filter by table name that was modified"
),
object_id: Optional[str] = Query(
object_id: str | None = Query(
None, description="Filter by ID of the object that was modified"
),
start_date: Optional[str] = Query(None, description="Filter logs after this date"),
end_date: Optional[str] = Query(None, description="Filter logs before this date"),
object_team_id: Optional[str] = Query(
start_date: str | None = Query(None, description="Filter logs after this date"),
end_date: str | None = Query(None, description="Filter logs before this date"),
object_team_id: str | None = Query(
None,
description="Filter by team_id present in before_value or updated_values JSON (PostgreSQL only)",
),
object_key_hash: Optional[str] = Query(
object_key_hash: str | None = Query(
None,
description="Filter by token (key hash) present in before_value or updated_values JSON (PostgreSQL only)",
),
# Sorting parameters
sort_by: Optional[str] = Query(
sort_by: str | None = Query(
None,
description="Column to sort by (e.g. 'updated_at', 'action', 'table_name')",
),
@ -101,46 +106,37 @@ async def get_audit_logs(
detail={"message": CommonProxyErrors.db_not_connected_error.value},
)
# Build filter conditions
where_conditions: Dict[str, Any] = {}
if changed_by:
where_conditions["changed_by"] = changed_by
if changed_by_api_key:
where_conditions["changed_by_api_key"] = changed_by_api_key
if action:
where_conditions["action"] = action
if table_name:
where_conditions["table_name"] = table_name
if object_id:
where_conditions["object_id"] = object_id
if start_date or end_date:
date_filter: Dict[str, Any] = {}
if start_date:
date_filter["gte"] = start_date
if end_date:
date_filter["lte"] = end_date
where_conditions["updated_at"] = date_filter
date_filter: Final[dict[str, str]] = {
**({"gte": start_date} if start_date else {}),
**({"lte": end_date} if end_date else {}),
}
# JSON field filters (PostgreSQL only) — each filter is AND'd with the
# others, but checks both before_value and updated_values internally (OR).
if object_team_id:
where_conditions["AND"] = where_conditions.get("AND", []) + [
_build_json_field_or_condition("team_id", object_team_id)
]
if object_key_hash:
where_conditions["AND"] = where_conditions.get("AND", []) + [
_build_json_field_or_condition("token", object_key_hash)
]
json_field_conditions: Final[list[dict[str, object]]] = [
*([_build_json_field_or_condition("team_id", object_team_id)] if object_team_id else []),
*([_build_json_field_or_condition("token", object_key_hash)] if object_key_hash else []),
]
# Build sort conditions
order_by: Dict[str, Any] = {}
if sort_by and isinstance(sort_by, str):
order_by[sort_by] = sort_order
else:
order_by["updated_at"] = sort_order # Default sort by updated_at
# Build filter conditions
where_conditions: Final[dict[str, object]] = {
**({"changed_by": changed_by} if changed_by else {}),
**({"changed_by_api_key": changed_by_api_key} if changed_by_api_key else {}),
**({"action": action} if action else {}),
**({"table_name": table_name} if table_name else {}),
**({"object_id": object_id} if object_id else {}),
**({"updated_at": date_filter} if start_date or end_date else {}),
**({"AND": json_field_conditions} if json_field_conditions else {}),
}
order_by: Final[dict[str, str]] = (
{sort_by: sort_order} if sort_by and isinstance(sort_by, str) else {"updated_at": sort_order}
)
audit_log_table: Final[TableActions["prisma_models.LiteLLM_AuditLog"]] = AuditLogRepository(prisma_client).table
# Get paginated results
audit_logs = await prisma_client.db.litellm_auditlog.find_many(
audit_logs: Final = await audit_log_table.find_many(
where=where_conditions,
order=order_by,
skip=(page - 1) * page_size,
@ -148,13 +144,14 @@ async def get_audit_logs(
)
# Get total count for pagination
total_count = await prisma_client.db.litellm_auditlog.count(where=where_conditions)
total_pages = -(-total_count // page_size) # Ceiling division
total_count: Final = await audit_log_table.count(where=where_conditions)
total_pages: Final = -(-total_count // page_size) # Ceiling division
# Return paginated response
return PaginatedAuditLogResponse(
audit_logs=[
AuditLogResponse(**audit_log.model_dump()) for audit_log in audit_logs
AuditLogResponse.model_validate(audit_log.model_dump())
for audit_log in audit_logs
]
if audit_logs
else [],
@ -198,8 +195,10 @@ async def get_audit_log_by_id(
detail={"message": CommonProxyErrors.db_not_connected_error.value},
)
audit_log_table: Final[TableActions["prisma_models.LiteLLM_AuditLog"]] = AuditLogRepository(prisma_client).table
# Get the audit log by ID
audit_log = await prisma_client.db.litellm_auditlog.find_unique(where={"id": id})
audit_log: Final = await audit_log_table.find_unique(where={"id": id})
if audit_log is None:
raise HTTPException(
@ -207,4 +206,4 @@ async def get_audit_log_by_id(
)
# Convert to response model
return AuditLogResponse(**audit_log.model_dump())
return AuditLogResponse.model_validate(audit_log.model_dump())

81
enterprise/litellm_enterprise/proxy/common_utils/check_batch_cost.py
View file

@ -2,9 +2,10 @@
Polls LiteLLM_ManagedObjectTable to check if the batch job is complete, and if the cost has been tracked.
"""
from dataclasses import replace as dataclasses_replace
from datetime import datetime, timedelta, timezone
from types import MappingProxyType
from typing import TYPE_CHECKING, Any, Dict, Final, List, Optional, Tuple
from typing import TYPE_CHECKING, Final, List, Literal, Optional, Tuple, cast
from litellm._logging import verbose_proxy_logger
from litellm._uuid import uuid
@ -14,6 +15,8 @@ from litellm.constants import (
)
if TYPE_CHECKING:
from prisma import models as prisma_models
from litellm.integrations.prometheus import PrometheusLogger
from litellm.proxy._types import LiteLLM_ManagedObjectTable
from litellm.proxy.utils import PrismaClient, ProxyLogging
@ -84,7 +87,7 @@ class CheckBatchCost:
return
self.batch_processed_support_confirmed = True
async def _get_user_info(self, batch_id: str, user_id: Optional[str]) -> Dict[str, Any]:
async def _get_user_info(self, batch_id: str, user_id: Optional[str]) -> dict[str, str | None]:
"""
Look up user email and key alias by user_id for enriching the S3 callback metadata.
Returns a dict with user_api_key_user_email and user_api_key_alias (both may be None).
@ -94,8 +97,10 @@ class CheckBatchCost:
if not user_id:
return {}
try:
user_row = await self.prisma_client.db.litellm_usertable.find_unique(
where={"user_id": user_id}
user_row: prisma_models.LiteLLM_UserTable | None = (
await self.prisma_client.db.litellm_usertable.find_unique(
where={"user_id": user_id}
)
)
if user_row is None:
return {}
@ -112,8 +117,10 @@ class CheckBatchCost:
if not api_key:
return None
try:
key_row = await self.prisma_client.db.litellm_verificationtoken.find_unique(
where={"token": api_key}
key_row: prisma_models.LiteLLM_VerificationToken | None = (
await self.prisma_client.db.litellm_verificationtoken.find_unique(
where={"token": api_key}
)
)
return getattr(key_row, "key_alias", None) if key_row is not None else None
except Exception as e:
@ -125,8 +132,10 @@ class CheckBatchCost:
if not team_id:
return None
try:
team_row = await self.prisma_client.db.litellm_teamtable.find_unique(
where={"team_id": team_id}
team_row: prisma_models.LiteLLM_TeamTable | None = (
await self.prisma_client.db.litellm_teamtable.find_unique(
where={"team_id": team_id}
)
)
return getattr(team_row, "team_alias", None) if team_row is not None else None
except Exception as e:
@ -135,7 +144,7 @@ class CheckBatchCost:
async def _build_creator_attribution_metadata(
self, job: "LiteLLM_ManagedObjectTable", batch_id: str
) -> Dict[str, Any]:
) -> dict[str, object]:
"""
Rebuild the spend-tracking metadata for the key, team, and tags that created the
batch so the batch-cost spend log is attributed the same way a non-batch request
@ -149,7 +158,7 @@ class CheckBatchCost:
team_id = getattr(job, "team_id", None)
request_tags = getattr(job, "request_tags", None)
metadata: Dict[str, Any] = {
metadata: dict[str, object] = {
"user_api_key_user_id": job.created_by,
"user_api_key": api_key,
"user_api_key_team_id": team_id,
@ -351,7 +360,7 @@ class CheckBatchCost:
return isinstance(error, (NotFoundError, openai.NotFoundError)) and output_file_id in str(error)
async def _finalize_unbilled_terminal_job(
self, job: "LiteLLM_ManagedObjectTable", response: "LiteLLMBatch"
self, job: "prisma_models.LiteLLM_ManagedObjectTable", response: "LiteLLMBatch"
) -> None:
"""Persist a terminal batch that has nothing billable, converting any raw
provider file ids to managed ids, and take it out of the poll page."""
@ -624,6 +633,7 @@ class CheckBatchCost:
later poll.
"""
from litellm.batches.batch_utils import (
count_error_file_failed_requests,
_get_file_content_as_dictionary,
calculate_batch_cost_and_usage,
)
@ -759,16 +769,33 @@ class CheckBatchCost:
model_id=model_id,
deployment_model=litellm_model_name,
)
batch_cost, batch_usage, batch_models = (
await calculate_batch_cost_and_usage(
file_content_dictionary=file_content_as_dict,
custom_llm_provider=llm_provider, # type: ignore
model_name=model_name,
model_info=deployment_model_info,
batch_file_provider: Final = cast(
Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic"], llm_provider
)
output_file_result: Final = await calculate_batch_cost_and_usage(
file_content_dictionary=file_content_as_dict,
custom_llm_provider=batch_file_provider,
model_name=model_name,
model_info=deployment_model_info,
)
error_file_failed_requests: Final = await count_error_file_failed_requests(
response,
custom_llm_provider=batch_file_provider,
litellm_params={
**credentials,
"_litellm_internal_model_credentials": MappingProxyType(dict(credentials)),
},
)
batch_result: Final = (
output_file_result
if not error_file_failed_requests
else dataclasses_replace(
output_file_result,
failed_requests=output_file_result.failed_requests + error_file_failed_requests,
)
)
logging_obj = LiteLLMLogging(
model=batch_models[0],
model=batch_result.models[0],
messages=[{"role": "user", "content": "<retrieve_batch>"}],
stream=False,
call_type="aretrieve_batch",
@ -800,9 +827,11 @@ class CheckBatchCost:
try:
await logging_obj.async_success_handler(
result=response,
batch_cost=batch_cost,
batch_usage=batch_usage,
batch_models=batch_models,
batch_cost=batch_result.cost,
batch_usage=batch_result.usage,
batch_models=batch_result.models,
batch_successful_requests=batch_result.successful_requests,
batch_failed_requests=batch_result.failed_requests,
)
except Exception:
await self._release_job_claim(job)
@ -966,6 +995,16 @@ class CheckBatchCost:
)
elif response.status in PROVIDER_TERMINAL_BATCH_STATUSES:
from litellm.proxy.openai_files_endpoints.common_utils import (
_completed_batch_safe_to_retire,
)
if response.status in ("completed", "complete") and not _completed_batch_safe_to_retire(response):
verbose_proxy_logger.info(
f"CheckBatchCost: batch {batch_id} is completed but its output file id "
f"has not appeared yet; leaving job {job.id} for the next poll cycle"
)
continue
await self._finalize_unbilled_terminal_job(job, response)
# Record polling run metrics (always, even if nothing was processed)

10
enterprise/litellm_enterprise/proxy/common_utils/check_responses_cost.py
View file

@ -1,6 +1,8 @@
"""
Polls LiteLLM_ManagedObjectTable to check if the response is complete.
Cost tracking is handled automatically by the get-responses call.
Cost tracking is handled by the get-responses call, which prices normally only because the
poll stamps itself with BACKGROUND_RESPONSE_COST_POLL_CALL_ORIGIN; user-facing reads of the
same route are non-inference and free.
"""
from datetime import datetime, timedelta, timezone
@ -9,12 +11,14 @@ from typing import TYPE_CHECKING, Dict, Optional, cast
import litellm
from litellm._logging import verbose_proxy_logger
from litellm.constants import (
INTERNAL_CALL_ORIGIN_METADATA_KEY,
MANAGED_OBJECT_STALENESS_CUTOFF_DAYS,
MAX_OBJECTS_PER_POLL_CYCLE,
STALE_OBJECT_CLEANUP_BATCH_SIZE,
)
from litellm.responses.utils import ResponsesAPIRequestUtils
from litellm.types.llms.openai import ResponsesAPIResponse
from litellm.types.utils import BACKGROUND_RESPONSE_COST_POLL_CALL_ORIGIN
if TYPE_CHECKING:
from litellm.proxy.utils import PrismaClient, ProxyLogging
@ -113,7 +117,8 @@ class CheckResponsesCost:
Check if background responses are complete and track their cost.
- Get all status="queued" or "in_progress" and file_purpose="response" jobs
- Query the provider to check if response is complete
- Cost is automatically tracked by the get-responses call
- Cost is tracked by the get-responses call, billed because the poll is stamped
with BACKGROUND_RESPONSE_COST_POLL_CALL_ORIGIN
- Mark responses in a terminal state as complete in the database
"""
try:
@ -153,6 +158,7 @@ class CheckResponsesCost:
# Prepare metadata with model information for cost tracking
litellm_metadata = {
"user_api_key_user_id": job.created_by or "default-user-id",
INTERNAL_CALL_ORIGIN_METADATA_KEY: BACKGROUND_RESPONSE_COST_POLL_CALL_ORIGIN,
}
# Add model information if available

188
enterprise/litellm_enterprise/proxy/hooks/managed_files.py
View file

@ -36,6 +36,7 @@ from litellm.llms.base_llm.managed_resources.isolation import (
build_list_page,
build_owner_filter,
can_access_resource,
resolve_resource_owner_id,
)
from litellm.proxy._types import (
CallTypes,
@ -45,6 +46,8 @@ from litellm.proxy._types import (
UserAPIKeyAuth,
)
from litellm.proxy.openai_files_endpoints.common_utils import (
FILE_LIST_CONTINUATION_CHUNK_SIZE,
MAX_FILE_LIST_LIMIT,
_is_base64_encoded_unified_file_id,
apply_unified_file_ids,
ensure_batch_response_managed_file_ids,
@ -54,6 +57,8 @@ from litellm.proxy.openai_files_endpoints.common_utils import (
map_raw_file_ids_to_unified,
normalize_mime_type_for_provider,
resolve_managed_output_file_model_name,
validate_file_list_limit,
validate_file_list_purpose,
)
from litellm.proxy.pass_through_endpoints.llm_provider_handlers.batch_attribution import (
request_tags_from_metadata,
@ -63,9 +68,9 @@ from litellm.types.llms.openai import ( # pyright: ignore[reportAttributeAccess
AsyncCursorPage,
ChatCompletionFileObject,
CreateFileRequest,
FileListPage,
FileObject,
OpenAIFileObject,
OpenAIFilesPurpose,
ResponsesAPIResponse,
)
from litellm.types.utils import (
@ -144,7 +149,14 @@ class _ManagedFileRow(Protocol):
class _ManagedFileTableActions(Protocol):
async def find_first(self, where: Mapping[str, object]) -> Optional[_ManagedFileRow]: ...
async def find_many(self, where: Mapping[str, object]) -> Sequence[_ManagedFileRow]: ...
async def find_many(
self,
where: Mapping[str, object],
take: int = ...,
order: Union[Mapping[str, str], Sequence[Mapping[str, str]]] = ...,
cursor: Mapping[str, str] = ...,
skip: int = ...,
) -> Sequence[_ManagedFileRow]: ...
async def upsert(self, where: Mapping[str, str], data: Mapping[str, Mapping[str, object]]) -> _ManagedFileRow: ...
@ -170,6 +182,10 @@ class _ManagedObjectTableActions(Protocol):
async def update_many(self, where: Mapping[str, object], data: Mapping[str, object]) -> int: ...
class _SchedulerWithJobLookup(Protocol):
def get_job(self, job_id: str) -> object: ...
class _CursorPageArgs(TypedDict, total=False):
cursor: Mapping[str, str]
skip: int
@ -211,7 +227,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
file_object=file_object,
model_mappings=model_mappings,
flat_model_file_ids=list(model_mappings.values()),
created_by=user_api_key_dict.user_id,
created_by=resolve_resource_owner_id(user_api_key_dict),
team_id=user_api_key_dict.team_id,
updated_by=user_api_key_dict.user_id,
)
@ -227,7 +243,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
"unified_file_id": file_id,
"model_mappings": json.dumps(model_mappings),
"flat_model_file_ids": list(model_mappings.values()),
"created_by": user_api_key_dict.user_id,
"created_by": resolve_resource_owner_id(user_api_key_dict),
"team_id": user_api_key_dict.team_id,
"updated_by": user_api_key_dict.user_id,
}
@ -331,7 +347,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
"file_object": file_object.model_dump_json(),
"model_object_id": model_object_id,
"file_purpose": file_purpose,
"created_by": user_api_key_dict.user_id,
"created_by": resolve_resource_owner_id(user_api_key_dict),
"team_id": user_api_key_dict.team_id,
"updated_by": user_api_key_dict.user_id,
"status": file_object.status,
@ -462,19 +478,56 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
)
page_size: Final = min(limit or 20, 100)
cursor_args: _CursorPageArgs = {"cursor": {"unified_object_id": after}, "skip": 1} if after else {}
batches = await _managed_object_table(self.prisma_client).find_many(
where=where_clause,
take=page_size + 1,
order=[{"created_at": "desc"}, {"unified_object_id": "desc"}],
**cursor_args,
matches: Final = await self._collect_listed_batches(
where_clause=where_clause,
after=after,
wanted=page_size + 1,
user_api_key_dict=user_api_key_dict,
)
return build_list_page(list(matches[:page_size]), has_more=len(matches) > page_size)
has_more = len(batches) > page_size
async def _collect_listed_batches(
self,
where_clause: Mapping[str, object],
after: Optional[str],
wanted: int,
user_api_key_dict: UserAPIKeyAuth,
) -> tuple[LiteLLMBatch, ...]:
"""Read chunks newest-first until ``wanted`` batches survive parsing and
file-id resolution or the caller's rows run out, so a run of rows that will
not parse refills the page instead of emptying it. The first chunk is
``wanted`` rows, so a healthy page still costs one query; a scan that has to
continue widens to ``FILE_LIST_CONTINUATION_CHUNK_SIZE`` like ``afile_list``,
and every chunk advances the keyset cursor, so the walk ends once the
caller's rows are exhausted."""
matches: tuple[LiteLLMBatch, ...] = () # rebind-ok: accumulates survivors across chunks
cursor_id: Optional[str] = after # rebind-ok: keyset cursor advances to each chunk's last row
chunk_size: int = wanted # rebind-ok: widens once a scan has to continue past the first chunk
while len(matches) < wanted:
cursor_args: _CursorPageArgs = {"cursor": {"unified_object_id": cursor_id}, "skip": 1} if cursor_id else {}
chunk = await _managed_object_table(self.prisma_client).find_many(
where=where_clause,
take=chunk_size,
order=[{"created_at": "desc"}, {"unified_object_id": "desc"}],
**cursor_args,
)
matches = matches + await self._resolve_listed_rows(
rows=chunk, wanted=wanted - len(matches), user_api_key_dict=user_api_key_dict
)
if len(chunk) < chunk_size:
break
cursor_id = chunk[-1].unified_object_id
chunk_size = max(chunk_size, FILE_LIST_CONTINUATION_CHUNK_SIZE)
return matches
async def _resolve_listed_rows(
self,
rows: "Sequence[PrismaManagedObjectRow]",
wanted: int,
user_api_key_dict: UserAPIKeyAuth,
) -> tuple[LiteLLMBatch, ...]:
parsed_rows: Final = tuple(
(row, batch_obj) for row in batches[:page_size] if (batch_obj := _parse_managed_batch_row(row)) is not None
(row, batch_obj) for row in rows if (batch_obj := _parse_managed_batch_row(row)) is not None
)
unified_id_by_raw_id: Final = await map_raw_file_ids_to_unified(
raw_file_ids=frozenset(
@ -485,19 +538,19 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
),
prisma_client=self.prisma_client,
)
resolved_batches: Final = [
await self._resolve_listed_batch(
resolved: Final[list[LiteLLMBatch]] = [] # mutable-ok: resolution stops as soon as the page is full
for row, batch_obj in parsed_rows:
if len(resolved) == wanted:
break
resolved_batch = await self._resolve_listed_batch(
row=row,
batch_obj=batch_obj,
unified_id_by_raw_id=unified_id_by_raw_id,
user_api_key_dict=user_api_key_dict,
)
for row, batch_obj in parsed_rows
]
return build_list_page(
[batch_obj for batch_obj in resolved_batches if batch_obj is not None],
has_more=has_more,
)
if resolved_batch is not None:
resolved.append(resolved_batch)
return tuple(resolved)
async def _resolve_listed_batch(
self,
@ -804,7 +857,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
file_ids.append(file_id)
return file_ids
def get_file_ids_from_responses_input(self, input: Union[str, List[Dict[str, Any]]]) -> List[str]:
def get_file_ids_from_responses_input(self, input: Union[str, List[Dict[str, object]]]) -> List[str]:
"""
Gets file ids from responses API input.
@ -829,7 +882,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
# Check for direct input_file type
if item.get("type") == "input_file":
file_id = item.get("file_id")
if file_id:
if isinstance(file_id, str) and file_id:
file_ids.append(file_id)
# Check for input_file in content array
@ -838,7 +891,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
for content_item in content:
if isinstance(content_item, dict) and content_item.get("type") == "input_file":
file_id = content_item.get("file_id")
if file_id:
if isinstance(file_id, str) and file_id:
file_ids.append(file_id)
return file_ids
@ -1178,7 +1231,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
# Handle both output_file_id and error_file_id
for file_attr in ["output_file_id", "error_file_id"]:
file_id_value = getattr(response, file_attr, None)
file_id_value: str | None = getattr(response, file_attr, None)
if file_id_value and model_id:
decoded_output_file_id = _is_base64_encoded_unified_file_id(file_id_value)
if decoded_output_file_id and "llm_output_file_id," in decoded_output_file_id:
@ -1365,12 +1418,76 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
async def afile_list(
self,
purpose: Optional[OpenAIFilesPurpose],
purpose: Optional[str],
litellm_parent_otel_span: Optional[Span],
user_api_key_dict: UserAPIKeyAuth,
limit: Optional[int] = None,
after: Optional[str] = None,
**data: Dict,
) -> List[OpenAIFileObject]:
"""Handled in files_endpoints.py"""
return []
) -> FileListPage:
"""List the managed files the caller owns, newest first.
Pagination is keyset based on ``unified_file_id`` so a key that owns
every file on the proxy still reads one bounded page at a time.
``purpose`` is applied after parsing, because the managed file table
keeps it inside the ``file_object`` blob instead of a column, and rows
whose blob will not parse drop out there too, so a chunk of rows can
yield fewer matches than the page holds. Successive chunks are read
until the page is full or the caller's rows run out, which keeps
``data`` non-empty while matches remain and its last id usable as the
next cursor. A first chunk that fills the page costs one query; once a
scan has to continue past it, the chunk widens to
``FILE_LIST_CONTINUATION_CHUNK_SIZE``, so the walk costs one query per
that many rows instead of one per page. That bound is per query, not
per request: the work is still linear in the rows the caller owns, and
a filter matching nothing reads every one of them, with no index
covering either the owner filter or the sort.
"""
validate_file_list_limit(limit)
validate_file_list_purpose(purpose)
owner_filter: Final = build_owner_filter(user_api_key_dict)
if owner_filter is None:
return FileListPage(**build_list_page([]))
if after:
cursor_row = await _managed_file_table(self.prisma_client).find_first(
where={**owner_filter, "unified_file_id": after}
)
if cursor_row is None:
raise ProxyException(
message=f"Invalid 'after' cursor: no file found with id '{after}'.",
type="invalid_request_error",
param="after",
code=400,
openai_code="invalid_value",
)
page_size: Final = min(limit or MAX_FILE_LIST_LIMIT, MAX_FILE_LIST_LIMIT)
matches: Final[List[OpenAIFileObject]] = []
cursor_id = after
chunk_size = page_size + 1
while len(matches) <= page_size:
cursor_args: _CursorPageArgs = {"cursor": {"unified_file_id": cursor_id}, "skip": 1} if cursor_id else {}
chunk = await _managed_file_table(self.prisma_client).find_many(
where=owner_filter,
take=chunk_size,
order=[{"created_at": "desc"}, {"unified_file_id": "desc"}],
**cursor_args,
)
matches.extend(
parsed_file_object.model_copy(update={"id": row.unified_file_id})
for row in chunk
if (parsed_file_object := _parse_managed_file_object(row.file_object, row.unified_file_id)) is not None
and (purpose is None or parsed_file_object.purpose == purpose)
)
if len(chunk) < chunk_size:
break
cursor_id = chunk[-1].unified_file_id
chunk_size = max(chunk_size, FILE_LIST_CONTINUATION_CHUNK_SIZE)
return FileListPage(**build_list_page(matches[:page_size], has_more=len(matches) > page_size))
def _is_batch_polling_enabled(self) -> bool:
"""
@ -1383,7 +1500,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
import litellm.proxy.proxy_server as proxy_server_module
# Check if the scheduler has the batch cost checking job registered
scheduler = getattr(proxy_server_module, "scheduler", None)
scheduler: Final[_SchedulerWithJobLookup | None] = getattr(proxy_server_module, "scheduler", None)
if scheduler is None:
return False
@ -1429,7 +1546,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
)
MAX_MATCHES_TO_RETURN = 10
batches = await self.prisma_client.db.litellm_managedobjecttable.find_many(
batches = await _managed_object_table(self.prisma_client).find_many(
where={
"file_purpose": "batch",
"batch_processed": False,
@ -1439,11 +1556,14 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
order={"created_at": "desc"},
)
referencing_batches = []
referencing_batches: Final[list[dict[str, object]]] = []
for batch in batches:
try:
# Parse the batch file_object to check for file references
batch_data = json.loads(batch.file_object) if isinstance(batch.file_object, str) else batch.file_object
decoded_file_object = _decode_json_blob(batch.file_object)
batch_data: Mapping[str, object] = (
decoded_file_object if isinstance(decoded_file_object, Mapping) else {}
)
# Extract file IDs from batch
# Batches typically reference the unified file ID in input_file_id

207
enterprise/litellm_enterprise/proxy/management_endpoints/project_endpoints.py
View file

@ -12,9 +12,10 @@ Endpoints for /project operations
import json
from collections.abc import Sequence
from typing import TYPE_CHECKING
from typing import TYPE_CHECKING, Final
from fastapi import APIRouter, Depends, HTTPException, Request
from pydantic import TypeAdapter
from litellm._logging import verbose_proxy_logger
from litellm._uuid import uuid
@ -26,37 +27,50 @@ from litellm.proxy.management_helpers.utils import (
management_endpoint_wrapper,
)
from litellm.proxy.utils import PrismaClient, handle_exception_on_proxy
from litellm.repositories.budget_repository import BudgetRepository
from litellm.repositories.object_permission_repository import ObjectPermissionRepository
from litellm.repositories.prisma_protocols import TableActions
from litellm.repositories.project_repository import ProjectRepository
from litellm.repositories.team_repository import TeamRepository
from litellm.repositories.user_repository import UserRepository
from litellm.repositories.verification_token_repository import VerificationTokenRepository
if TYPE_CHECKING:
from prisma import models as prisma_models
from prisma.actions import (
LiteLLM_ProjectTableActions,
LiteLLM_TeamTableActions,
LiteLLM_VerificationTokenActions,
)
from litellm import Router
router = APIRouter()
def _team_table(prisma_client: PrismaClient) -> "LiteLLM_TeamTableActions[prisma_models.LiteLLM_TeamTable]":
team_table: LiteLLM_TeamTableActions[prisma_models.LiteLLM_TeamTable] = prisma_client.db.litellm_teamtable
return team_table
_OBJECT_PERMISSION_PAYLOAD: Final = TypeAdapter(dict[str, object])
def _project_table(prisma_client: PrismaClient) -> "LiteLLM_ProjectTableActions[prisma_models.LiteLLM_ProjectTable]":
project_table: LiteLLM_ProjectTableActions[prisma_models.LiteLLM_ProjectTable] = (
prisma_client.db.litellm_projecttable
)
return project_table
def _team_table(prisma_client: PrismaClient) -> TableActions["prisma_models.LiteLLM_TeamTable"]:
return TeamRepository(prisma_client).table
def _project_table(prisma_client: PrismaClient) -> TableActions["prisma_models.LiteLLM_ProjectTable"]:
return ProjectRepository(prisma_client).table
def _verification_token_table(
prisma_client: PrismaClient,
) -> "LiteLLM_VerificationTokenActions[prisma_models.LiteLLM_VerificationToken]":
verification_token_table: LiteLLM_VerificationTokenActions[prisma_models.LiteLLM_VerificationToken] = (
prisma_client.db.litellm_verificationtoken
)
return verification_token_table
) -> TableActions["prisma_models.LiteLLM_VerificationToken"]:
return VerificationTokenRepository(prisma_client).table
def _budget_table(prisma_client: PrismaClient) -> TableActions["prisma_models.LiteLLM_BudgetTable"]:
return BudgetRepository(prisma_client).table
def _object_permission_table(
prisma_client: PrismaClient,
) -> TableActions["prisma_models.LiteLLM_ObjectPermissionTable"]:
return ObjectPermissionRepository(prisma_client).table
def _user_table(prisma_client: PrismaClient) -> TableActions["prisma_models.LiteLLM_UserTable"]:
return UserRepository(prisma_client).table
def _jsonified(prisma_client: PrismaClient, payload: dict[str, object]) -> dict[str, object]:
@ -205,6 +219,114 @@ def _check_team_project_limits(
)
def _project_models_missing_positive_quota(
models: list[str] | None,
rpm_limits: Mapping[str, object] | None,
tpm_limits: Mapping[str, object] | None,
) -> list[str]:
"""Return the models that lack a positive `rpm` AND `tpm` quota.
A valid quota is a positive integer; null, zero, and negative are rejected
because downstream rate limiters treat a non-positive limit as immediately
exhausted (every request blocked).
"""
def _is_positive(value: object) -> bool:
return isinstance(value, int) and not isinstance(value, bool) and value > 0
rpm = rpm_limits or {}
tpm = tpm_limits or {}
return [model for model in (models or []) if not _is_positive(rpm.get(model)) or not _is_positive(tpm.get(model))]
def _router_access_group_names(llm_router: "Router | None") -> frozenset[str]:
return frozenset(llm_router.get_model_access_groups()) if llm_router is not None else frozenset()
def _project_models_expanding_at_request_time(
models: Sequence[str] | None, access_group_names: frozenset[str]
) -> tuple[str, ...]:
"""Entries project auth expands to many concrete models (`all-proxy-models`, `*` patterns,
access groups). The rate limiter looks quotas up by the exact requested model name, so a
quota keyed on one of these entries is never applied."""
return tuple(
model
for model in (models or ())
if model == SpecialModelNames.all_proxy_models.value or "*" in model or model in access_group_names
)
def _raise_on_project_models_expanding_at_request_time(
models: Sequence[str] | None, access_group_names: frozenset[str]
) -> None:
expanding: Final = _project_models_expanding_at_request_time(models, access_group_names)
if not expanding:
return
raise HTTPException(
status_code=400,
detail={
"error": f"models {list(expanding)} expand to multiple models at request time, so a per-model rpm/tpm quota cannot be enforced for them while 'enforce_project_model_quota' is enabled. List concrete model names instead."
},
)
def _raise_on_missing_project_model_quota(
data: NewProjectRequest | UpdateProjectRequest, access_group_names: frozenset[str] = frozenset()
) -> None:
"""Require a positive `rpm`/`tpm` quota for every model on project CREATE.
`model_rpm_limit`/`model_tpm_limit` are relocated into `metadata` by the request
model's `set_model_info` validator, so they are read from there.
Only invoked when `general_settings.enforce_project_model_quota` is enabled
(default off), so it is opt-in and does not change behavior for existing users.
"""
_raise_on_project_models_expanding_at_request_time(data.models, access_group_names)
metadata = data.metadata or {}
missing = _project_models_missing_positive_quota(
data.models, metadata.get("model_rpm_limit"), metadata.get("model_tpm_limit")
)
if not missing:
return
raise HTTPException(
status_code=400,
detail={
"error": f"models {missing} added to project without a positive rpm/tpm quota. Set a positive model_rpm_limit and model_tpm_limit for each model."
},
)
def _raise_on_missing_project_model_quota_on_update(
data: UpdateProjectRequest, existing_project: object, access_group_names: frozenset[str] = frozenset()
) -> None:
"""Require a positive `rpm`/`tpm` quota over the RESULTING state on project UPDATE.
`/project/update` replaces `models` and `metadata` when they are provided, so the
check runs on what the project WILL look like: a partial update that doesn't touch
models/quota keeps the existing values, while one that adds a model or clears a
model's quota must leave every resulting model with a positive limit.
Only invoked when `general_settings.enforce_project_model_quota` is enabled
(default off), so it is opt-in and does not change behavior for existing users.
"""
resulting_models = data.models if data.models is not None else (getattr(existing_project, "models", None) or [])
resulting_metadata = (
data.metadata if data.metadata is not None else (getattr(existing_project, "metadata", None) or {})
)
_raise_on_project_models_expanding_at_request_time(resulting_models, access_group_names)
missing = _project_models_missing_positive_quota(
resulting_models, resulting_metadata.get("model_rpm_limit"), resulting_metadata.get("model_tpm_limit")
)
if not missing:
return
raise HTTPException(
status_code=400,
detail={
"error": f"models {missing} would be left on the project without a positive rpm/tpm quota. Set a positive model_rpm_limit and model_tpm_limit for each model."
},
)
async def _create_budget_for_project(
data: NewProjectRequest,
user_id: str | None,
@ -219,7 +341,7 @@ async def _create_budget_for_project(
new_budget = _jsonified(prisma_client, budget_row.model_dump(exclude_none=True))
_budget: prisma_models.LiteLLM_BudgetTable = await prisma_client.db.litellm_budgettable.create(
_budget: Final = await _budget_table(prisma_client).create(
data={
**new_budget,
"created_by": user_id or litellm_proxy_admin_name,
@ -242,10 +364,8 @@ async def _set_project_object_permission(
return None
if data.object_permission is not None:
created_object_permission: prisma_models.LiteLLM_ObjectPermissionTable = (
await prisma_client.db.litellm_objectpermissiontable.create(
data=data.object_permission.model_dump(exclude_none=True),
)
created_object_permission: Final = await _object_permission_table(prisma_client).create(
data=data.object_permission.model_dump(exclude_none=True),
)
del data.object_permission
return created_object_permission.object_permission_id
@ -352,7 +472,9 @@ async def new_project(
```
"""
from litellm.proxy.proxy_server import (
general_settings,
litellm_proxy_admin_name,
llm_router,
premium_user,
prisma_client,
)
@ -399,6 +521,10 @@ async def new_project(
data=data,
)
# Opt-in (default off): require rpm/tpm for every model added to the project.
if general_settings.get("enforce_project_model_quota", False):
_raise_on_missing_project_model_quota(data, _router_access_group_names(llm_router))
# Check if user has permission to create projects for this team
# only team admins can create projects for their team
has_permission = await _check_user_permission_for_project(
@ -470,10 +596,8 @@ async def new_project(
new_project_row = _remove_budget_fields_from_project_data(new_project_row)
verbose_proxy_logger.info(f"new_project_row: {json.dumps(new_project_row, indent=2)}")
response: prisma_models.LiteLLM_ProjectTable = await prisma_client.db.litellm_projecttable.create(
data={
**new_project_row, # type: ignore
},
response: Final = await _project_table(prisma_client).create(
data={**new_project_row},
include={"litellm_budget_table": True},
)
@ -538,7 +662,9 @@ async def update_project(
```
"""
from litellm.proxy.proxy_server import (
general_settings,
litellm_proxy_admin_name,
llm_router,
premium_user,
prisma_client,
user_api_key_cache,
@ -642,6 +768,12 @@ async def update_project(
data=data,
)
# Opt-in (default off): require rpm/tpm for every model the update would leave on the project.
if general_settings.get("enforce_project_model_quota", False):
_raise_on_missing_project_model_quota_on_update(
data, existing_project, _router_access_group_names(llm_router)
)
# Prepare update data
update_data = _jsonified(prisma_client, data.model_dump(exclude_none=True, exclude={"project_id"}))
update_data["updated_by"] = user_api_key_dict.user_id or litellm_proxy_admin_name
@ -652,7 +784,7 @@ async def update_project(
if budget_updates and existing_project.budget_id:
# Update existing budget
await prisma_client.db.litellm_budgettable.update(
await _budget_table(prisma_client).update(
where={"budget_id": existing_project.budget_id},
data={
**budget_updates,
@ -667,18 +799,17 @@ async def update_project(
if "object_permission" in update_data:
object_permission_data = update_data.pop("object_permission")
if object_permission_data:
object_permission_payload: Final = _OBJECT_PERMISSION_PAYLOAD.validate_python(object_permission_data)
if existing_project.object_permission_id:
# Update existing permission
await prisma_client.db.litellm_objectpermissiontable.update(
await _object_permission_table(prisma_client).update(
where={"object_permission_id": existing_project.object_permission_id},
data=object_permission_data,
data=object_permission_payload,
)
else:
# Create new permission
created_permission: prisma_models.LiteLLM_ObjectPermissionTable = (
await prisma_client.db.litellm_objectpermissiontable.create(
data=object_permission_data,
)
created_permission: Final = await _object_permission_table(prisma_client).create(
data=object_permission_payload,
)
update_data["object_permission_id"] = created_permission.object_permission_id
@ -694,7 +825,7 @@ async def update_project(
update_data = _remove_budget_fields_from_project_data(update_data)
# Update project
updated_project: prisma_models.LiteLLM_ProjectTable | None = await prisma_client.db.litellm_projecttable.update(
updated_project: Final = await _project_table(prisma_client).update(
where={"project_id": data.project_id},
data=update_data,
include={"litellm_budget_table": True, "object_permission": True},
@ -934,7 +1065,7 @@ async def list_projects(
# Look up the user's team memberships via the reverse-index on
# LiteLLM_UserTable.teams (maintained by team_member_add alongside
# members_with_roles). This avoids a full scan of all team rows.
user_record: prisma_models.LiteLLM_UserTable | None = await prisma_client.db.litellm_usertable.find_unique(
user_record: Final = await _user_table(prisma_client).find_unique(
where={"user_id": user_api_key_dict.user_id},
)
user_team_ids: list[str] = user_record.teams if user_record is not None and user_record.teams else []

4
enterprise/pyproject.toml
View file

@ -1,6 +1,6 @@
[project]
name = "litellm-enterprise"
version = "0.1.58"
version = "0.1.63"
description = "Package for LiteLLM Enterprise features"
readme = "README.md"
requires-python = ">=3.9"
@ -26,7 +26,7 @@ required-version = ">=0.10.9"
module-root = ""
[tool.commitizen]
version = "0.1.58"
version = "0.1.63"
version_files = [
"pyproject.toml:^version",
"../pyproject.toml:litellm-enterprise==",

12
gateway/Dockerfile
View file

@ -1,5 +1,5 @@
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:e624c5d5e42382ce7165ddafcbbf8e6769a24cbd02ea6114b880b05ae5ba2a8d
ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
FROM $UV_IMAGE AS uvbin
@ -16,7 +16,7 @@ COPY --from=uvbin /uv /uvx /usr/local/bin/
# instead of nodeenv downloading one whose dynamic deps may not be in Wolfi
# (e.g. Node 26.2.0 needs libatomic). Retry for transient apk.cgr.dev flakes.
RUN for i in 1 2 3; do \
apk add --no-cache bash gcc python3 python3-dev openssl openssl-dev libsndfile nodejs npm && break; \
apk add --no-cache bash gcc python-3.13 python-3.13-dev openssl openssl-dev libsndfile nodejs npm && break; \
[ $i = 3 ] && { echo "apk add failed after 3 retries" >&2; exit 1; }; \
sleep 5; \
done
@ -47,7 +47,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
--extra extra_proxy \
--extra semantic-router \
--extra bedrock-realtime \
--python python3
--python python3.13
# Stage 2 — copy source and install the project + workspace members.
COPY . .
@ -59,7 +59,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
--extra extra_proxy \
--extra semantic-router \
--extra bedrock-realtime \
--python python3
--python python3.13
RUN HOME=/opt/prisma XDG_CACHE_HOME=/opt/prisma/.cache PRISMA_BINARY_CACHE_DIR=/opt/prisma/binaries \
npm_config_cache=/root/.npm \
@ -73,7 +73,7 @@ FROM $LITELLM_RUNTIME_IMAGE AS runtime
USER root
RUN for i in 1 2 3; do \
apk add --no-cache bash openssl tzdata python3 libsndfile libatomic && break; \
apk add --no-cache bash openssl tzdata python-3.13 libsndfile libatomic && break; \
[ $i = 3 ] && { echo "apk add failed after 3 retries" >&2; exit 1; }; \
sleep 5; \
done

1
gateway/routes/allowlist.py
View file

@ -86,6 +86,7 @@ GATEWAY_PATH_PREFIXES: tuple[str, ...] = (
"/comprehendmedical",
"/cohere/",
"/gemini/",
"/gigachat/",
"/google/",
"/vertex_ai/",
"/vertex-ai/",

4
helm/litellm-helm/README.md
View file

@ -26,7 +26,7 @@ If `db.useStackgresOperator` is used (not yet implemented):
| `replicaCount` | The number of LiteLLM Proxy pods to be deployed | `1` |
| `masterkeySecretName` | The name of the Kubernetes Secret that contains the Master API Key for LiteLLM. If not specified, use the generated secret name. | N/A |
| `masterkeySecretKey` | The key within the Kubernetes Secret that contains the Master API Key for LiteLLM. If not specified, use `masterkey` as the key. | N/A |
| `masterkey` | The Master API Key for LiteLLM. If not specified, a random key in the `sk-...` format is generated. | N/A |
| `masterkey` | The Master API Key for LiteLLM. If not specified, a random key in the `sk-...` format is generated on first install and reused on upgrades. | N/A |
| `environmentSecrets` | An optional array of Secret object names. The keys and values in these secrets will be presented to the LiteLLM proxy pod as environment variables. See below for an example Secret object. | `[]` |
| `environmentConfigMaps` | An optional array of ConfigMap object names. The keys and values in these configmaps will be presented to the LiteLLM proxy pod as environment variables. See below for an example Secret object. | `[]` |
| `image.repository` | LiteLLM Proxy image repository | `ghcr.io/berriai/litellm` |
@ -212,6 +212,8 @@ service, the **Proxy Endpoint** should be set to `http://<RELEASE>-litellm:4000`
The **Proxy Key** is the value specified for `masterkey` or, if a `masterkey`
was not provided to the helm command line, the `masterkey` is a randomly
generated string in the `sk-...` format stored in the `<RELEASE>-litellm-masterkey` Kubernetes Secret.
The key is generated once on the first install; later `helm upgrade` runs reuse the
value already in that Secret, so upgrading never rotates the master key.
```bash
kubectl -n litellm get secret <RELEASE>-litellm-masterkey -o jsonpath="{.data.masterkey}"

6
helm/litellm-helm/templates/secret-masterkey.yaml
View file

@ -1,9 +1,11 @@
{{- if not .Values.masterkeySecretName }}
{{ $masterkey := (.Values.masterkey | default (printf "sk-%s" (randAlphaNum 18))) }}
{{- $secretName := printf "%s-masterkey" (include "litellm.fullname" .) }}
{{- $existing := lookup "v1" "Secret" .Release.Namespace $secretName }}
{{- $masterkey := .Values.masterkey | default (dig "data" "masterkey" "" $existing | b64dec) | default (printf "sk-%s" (randAlphaNum 18)) }}
apiVersion: v1
kind: Secret
metadata:
name: {{ include "litellm.fullname" . }}-masterkey
name: {{ $secretName }}
data:
masterkey: {{ $masterkey | b64enc }}
type: Opaque

47
helm/litellm-helm/tests/masterkey-secret_tests.yaml
View file

@ -15,6 +15,53 @@ tests:
# Note: The masterkey is generated as "sk-<18-random-chars>" in plain text,
# but stored as base64 encoded in Kubernetes secret (requirement).
# "sk-" base64 encodes to "c2st", so we check for "^c2st" pattern.
- it: should reuse the master key already stored in the cluster instead of generating a new one on upgrade
template: secret-masterkey.yaml
set:
masterkeySecretName: ""
kubernetesProvider:
scheme:
"v1/Secret":
gvr:
version: "v1"
resource: "secrets"
namespaced: true
objects:
- kind: Secret
apiVersion: v1
metadata:
name: RELEASE-NAME-litellm-masterkey
namespace: NAMESPACE
data:
masterkey: c2stZXhpc3Rpbmcta2V5
asserts:
- equal:
path: data.masterkey
value: c2stZXhpc3Rpbmcta2V5
- it: should let an explicit masterkey value override the one already stored in the cluster
template: secret-masterkey.yaml
set:
masterkeySecretName: ""
masterkey: sk-explicit
kubernetesProvider:
scheme:
"v1/Secret":
gvr:
version: "v1"
resource: "secrets"
namespaced: true
objects:
- kind: Secret
apiVersion: v1
metadata:
name: RELEASE-NAME-litellm-masterkey
namespace: NAMESPACE
data:
masterkey: c2stZXhpc3Rpbmcta2V5
asserts:
- equal:
path: data.masterkey
value: c2stZXhwbGljaXQ=
- it: should not create a secret if masterkeySecretName is set
template: secret-masterkey.yaml
set:

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

@ -7,6 +7,10 @@ metadata:
{{- include "litellm.commonLabels" . | nindent 4 }}
app.kubernetes.io/component: backend
spec:
{{- with .Values.backend.strategy }}
strategy:
{{- toYaml . | nindent 4 }}
{{- end }}
selector:
matchLabels:
{{- include "litellm.backend.selectorLabels" . | nindent 6 }}

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

@ -7,6 +7,10 @@ metadata:
{{- include "litellm.commonLabels" . | nindent 4 }}
app.kubernetes.io/component: gateway
spec:
{{- with .Values.gateway.strategy }}
strategy:
{{- toYaml . | nindent 4 }}
{{- end }}
selector:
matchLabels:
{{- include "litellm.gateway.selectorLabels" . | nindent 6 }}

148
helm/litellm/templates/ingress.yaml
View file

@ -5,6 +5,41 @@
{{- $gatewayPort := .Values.gateway.service.port -}}
{{- $backendPort := .Values.backend.service.port -}}
{{- $uiPort := .Values.ui.service.port -}}
{{/*
Backends addressable from ingress.extraPaths, keyed by the `service` field.
*/}}
{{- $extraPathBackends := dict
"gateway" (dict "name" $gatewayName "port" $gatewayPort)
"backend" (dict "name" $backendName "port" $backendPort)
"ui" (dict "name" $uiName "port" $uiPort)
-}}
{{/*
UI paths (Next.js static export).
/ui/* is where the SPA serves its login + dashboard routes (e.g. /ui/login).
Without it, /ui/* falls into the catch-all → backend → 404.
The App Router (output: "export", basePath: "") emits the RSC/flight payload
for every route as a ROOT-level <route>.txt (/index.txt, /teams.txt,
/__next._tree.txt, ...). The client router fetches these on every soft
navigation / prefetch as <route>.txt?_rsc=<hash> (the query string is
irrelevant to path matching). They are not under /ui, /_next, or
/litellm-asset-prefix, so without /*.txt they fall to the backend catch-all
→ 404 → client-side navigation never settles and the login flow spins in an
infinite redirect loop (/ ⇄ /ui/login). ui/nginx.conf already serves *.txt
from the export; the rule only routes the request to it. Needs an ingress
controller whose ImplementationSpecific path is a wildcard pattern
(AWS ALB: `*` = 0+ chars); this chart targets the AWS Load Balancer
Controller.
*/}}
{{- $uiPaths := list
(dict "path" "/" "pathType" "Exact")
(dict "path" "/favicon.ico" "pathType" "Exact")
(dict "path" "/litellm-asset-prefix" "pathType" "Prefix")
(dict "path" "/_next" "pathType" "Prefix")
(dict "path" "/ui" "pathType" "Prefix")
(dict "path" "/*.txt" "pathType" "ImplementationSpecific")
-}}
{{/*
Gateway data-plane prefixes — must mirror gateway/routes/allowlist.py.
Versioned paths are listed explicitly to avoid routing management routes
@ -39,6 +74,21 @@
routes at startup -> 404. So /test is rendered as a standalone Exact path
and /test/* falls through to the backend catch-all.
*/}}
{{/*
Every "<path>|<pathType>" this template renders on its own. An
ingress.extraPaths entry that repeats one of these is rejected: duplicates
in a single rule are resolved by position or by controller-specific tie
breaking, so the operator entry could take over a built-in route (an entry
at "/" Prefix would swallow the whole backend management API) instead of
adding to it.
*/}}
{{- $builtinPathKeys := list "/test|Exact" "/|Prefix" -}}
{{- range $uiPaths }}
{{- $builtinPathKeys = append $builtinPathKeys (printf "%s|%s" .path .pathType) }}
{{- end }}
{{- range $gatewayPrefixes }}
{{- $builtinPathKeys = append $builtinPathKeys (printf "%s|Prefix" .) }}
{{- end }}
apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
@ -64,65 +114,15 @@ spec:
http:
paths:
# --- UI (Next.js static export) ---
- path: /
pathType: Exact
backend:
service:
name: {{ $uiName }}
port:
number: {{ $uiPort }}
- path: /favicon.ico
pathType: Exact
backend:
service:
name: {{ $uiName }}
port:
number: {{ $uiPort }}
- path: /litellm-asset-prefix
pathType: Prefix
backend:
service:
name: {{ $uiName }}
port:
number: {{ $uiPort }}
- path: /_next
pathType: Prefix
backend:
service:
name: {{ $uiName }}
port:
number: {{ $uiPort }}
# /ui/* is where the Next.js SPA serves its login + dashboard
# routes (e.g. /ui/login). Without this, /ui/* falls into the
# catch-all → backend → 404.
- path: /ui
pathType: Prefix
backend:
service:
name: {{ $uiName }}
port:
number: {{ $uiPort }}
# Next.js App Router (output: "export", basePath: "") emits the
# RSC/flight payload for every route as a ROOT-level <route>.txt
# (/index.txt, /teams.txt, /__next._tree.txt, ...). The client
# router fetches these on every soft navigation / prefetch as
# <route>.txt?_rsc=<hash> (the query string is irrelevant to path
# matching). They are not under /ui, /_next, or
# /litellm-asset-prefix, so without this rule they fall to the
# backend catch-all → 404 → client-side navigation never settles
# and the login flow spins in an infinite redirect loop
# (/ ⇄ /ui/login). ui/nginx.conf already serves *.txt from the
# export; this rule only routes the request to it. Needs an
# ingress controller whose ImplementationSpecific path is a
# wildcard pattern (AWS ALB: `*` = 0+ chars); this chart targets
# the AWS Load Balancer Controller.
- path: /*.txt
pathType: ImplementationSpecific
{{- range $uiPaths }}
- path: {{ .path }}
pathType: {{ .pathType }}
backend:
service:
name: {{ $uiName }}
port:
number: {{ $uiPort }}
{{- end }}
# --- Gateway data plane ---
# Exact /test only (see the $gatewayPrefixes comment above);
# /test/* MCP management endpoints fall to the backend catch-all.
@ -142,6 +142,46 @@ spec:
port:
number: {{ $gatewayPort }}
{{- end }}
{{- /*
--- Operator-supplied extra paths (ingress.extraPaths) ---
Rendered after every built-in path so an entry can never take
precedence over a default, and before the backend catch-all.
Position only decides the match on controllers that honour manifest
order: the AWS Load Balancer Controller this chart targets sorts
Exact paths first and Prefix paths longest-first, but keeps
ImplementationSpecific paths in manifest order, which is what the
/*.txt rule above already depends on.
*/}}
{{- range $idx, $extra := .Values.ingress.extraPaths }}
{{- if not (kindIs "map" $extra) }}
{{- fail (printf "ingress.extraPaths[%d]: each entry must be a mapping with a 'path' key" $idx) }}
{{- end }}
{{- if not $extra.path }}
{{- fail (printf "ingress.extraPaths[%d]: 'path' is required" $idx) }}
{{- end }}
{{- $service := $extra.service | default "gateway" }}
{{- $target := get $extraPathBackends $service }}
{{- if not $target }}
{{- fail (printf "ingress.extraPaths[%d] (path %s): unknown service %q, expected one of backend, gateway, ui" $idx $extra.path $service) }}
{{- end }}
{{- $pathType := $extra.pathType | default "Prefix" }}
{{- if not (has $pathType (list "Prefix" "Exact" "ImplementationSpecific")) }}
{{- fail (printf "ingress.extraPaths[%d] (path %s): unknown pathType %q, expected one of Exact, ImplementationSpecific, Prefix" $idx $extra.path $pathType) }}
{{- end }}
{{- if eq $extra.path "/" }}
{{- fail (printf "ingress.extraPaths[%d]: path / is already routed in both directions, Exact to ui and Prefix to backend, so no pathType leaves a request for an entry here to capture" $idx) }}
{{- end }}
{{- if has (printf "%s|%s" $extra.path $pathType) $builtinPathKeys }}
{{- fail (printf "ingress.extraPaths[%d]: path %s with pathType %s is already routed by this chart, and a duplicate would take it over rather than add to it" $idx $extra.path $pathType) }}
{{- end }}
- path: {{ $extra.path | quote }}
pathType: {{ $pathType }}
backend:
service:
name: {{ $target.name }}
port:
number: {{ $target.port }}
{{- end }}
# --- Catch-all → backend (management API: /key/*, /user/*, /team/*, ...) ---
- path: /
pathType: Prefix

12
helm/litellm/templates/migrations-job.yaml
View file

@ -7,6 +7,8 @@
#
# Running this pre-upgrade closes the window where new application pods would
# otherwise serve traffic against the previous release's unmigrated schema.
# Argo CD users can swap the Helm hook for a PreSync hook through
# `migrationJob.hooks`, which re-runs the Job on every sync.
apiVersion: batch/v1
kind: Job
metadata:
@ -14,10 +16,18 @@ metadata:
labels:
{{- include "litellm.commonLabels" . | nindent 4 }}
app.kubernetes.io/component: migrations
{{- if or .Values.migrationJob.hooks.helm.enabled .Values.migrationJob.hooks.argocd.enabled }}
annotations:
{{- if .Values.migrationJob.hooks.helm.enabled }}
helm.sh/hook: pre-install,pre-upgrade
helm.sh/hook-delete-policy: before-hook-creation
helm.sh/hook-weight: "0"
helm.sh/hook-weight: {{ .Values.migrationJob.hooks.helm.weight | default "0" | quote }}
{{- end }}
{{- if .Values.migrationJob.hooks.argocd.enabled }}
argocd.argoproj.io/hook: PreSync
argocd.argoproj.io/hook-delete-policy: BeforeHookCreation
{{- end }}
{{- end }}
spec:
backoffLimit: {{ .Values.migrationJob.backoffLimit }}
ttlSecondsAfterFinished: {{ .Values.migrationJob.ttlSecondsAfterFinished }}

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

@ -7,6 +7,10 @@ metadata:
{{- include "litellm.commonLabels" . | nindent 4 }}
app.kubernetes.io/component: ui
spec:
{{- with .Values.ui.strategy }}
strategy:
{{- toYaml . | nindent 4 }}
{{- end }}
selector:
matchLabels:
{{- include "litellm.ui.selectorLabels" . | nindent 6 }}

317
helm/litellm/tests/ingress_extra_paths_tests.yaml Normal file
View file

@ -0,0 +1,317 @@
suite: test ingress.extraPaths
templates:
- ingress.yaml
values:
- ./values/required.yaml
tests:
- it: renders nothing extra between the built-in gateway prefixes and the backend catch-all when unset
set:
ingress.enabled: true
asserts:
- equal:
path: spec.rules[0].http.paths[-1]
value:
path: /
pathType: Prefix
backend:
service:
name: RELEASE-NAME-litellm-backend
port:
number: 4001
- equal:
path: spec.rules[0].http.paths[-2]
value:
path: /metrics
pathType: Prefix
backend:
service:
name: RELEASE-NAME-litellm-gateway
port:
number: 4000
- it: routes an extra path to the gateway by default, immediately before the backend catch-all
set:
ingress.enabled: true
ingress.extraPaths:
- path: /watsonx
asserts:
- equal:
path: spec.rules[0].http.paths[-2]
value:
path: /watsonx
pathType: Prefix
backend:
service:
name: RELEASE-NAME-litellm-gateway
port:
number: 4000
- equal:
path: spec.rules[0].http.paths[-1]
value:
path: /
pathType: Prefix
backend:
service:
name: RELEASE-NAME-litellm-backend
port:
number: 4001
- it: keeps every built-in path when extra paths are supplied
set:
ingress.enabled: true
ingress.extraPaths:
- path: /watsonx
asserts:
- contains:
path: spec.rules[0].http.paths
content:
path: /
pathType: Exact
backend:
service:
name: RELEASE-NAME-litellm-ui
port:
number: 3000
- contains:
path: spec.rules[0].http.paths
content:
path: /ui
pathType: Prefix
backend:
service:
name: RELEASE-NAME-litellm-ui
port:
number: 3000
- contains:
path: spec.rules[0].http.paths
content:
path: /test
pathType: Exact
backend:
service:
name: RELEASE-NAME-litellm-gateway
port:
number: 4000
- contains:
path: spec.rules[0].http.paths
content:
path: /v1/chat
pathType: Prefix
backend:
service:
name: RELEASE-NAME-litellm-gateway
port:
number: 4000
- contains:
path: spec.rules[0].http.paths
content:
path: /vertex_ai
pathType: Prefix
backend:
service:
name: RELEASE-NAME-litellm-gateway
port:
number: 4000
- it: renders every entry in order and honours the service and pathType selectors
set:
ingress.enabled: true
ingress.extraPaths:
- path: /watsonx
service: gateway
- path: /my-passthrough
pathType: Exact
service: backend
- path: /brand.txt
pathType: ImplementationSpecific
service: ui
asserts:
- equal:
path: spec.rules[0].http.paths[-4]
value:
path: /watsonx
pathType: Prefix
backend:
service:
name: RELEASE-NAME-litellm-gateway
port:
number: 4000
- equal:
path: spec.rules[0].http.paths[-3]
value:
path: /my-passthrough
pathType: Exact
backend:
service:
name: RELEASE-NAME-litellm-backend
port:
number: 4001
- equal:
path: spec.rules[0].http.paths[-2]
value:
path: /brand.txt
pathType: ImplementationSpecific
backend:
service:
name: RELEASE-NAME-litellm-ui
port:
number: 3000
- it: addresses the component services by their configured ports
set:
ingress.enabled: true
gateway.service.port: 8000
backend.service.port: 8001
ui.service.port: 8080
ingress.extraPaths:
- path: /watsonx
- path: /my-passthrough
service: backend
- path: /brand.txt
service: ui
asserts:
- equal:
path: spec.rules[0].http.paths[-4].backend.service.port.number
value: 8000
- equal:
path: spec.rules[0].http.paths[-3].backend.service.port.number
value: 8001
- equal:
path: spec.rules[0].http.paths[-2].backend.service.port.number
value: 8080
- it: rejects an entry naming a service the chart does not deploy
set:
ingress.enabled: true
ingress.extraPaths:
- path: /watsonx
service: proxy
asserts:
- failedTemplate:
errorMessage: 'ingress.extraPaths[0] (path /watsonx): unknown service "proxy", expected one of backend, gateway, ui'
- it: rejects an entry whose pathType is not a kubernetes pathType
set:
ingress.enabled: true
ingress.extraPaths:
- path: /watsonx
pathType: prefix
asserts:
- failedTemplate:
errorMessage: 'ingress.extraPaths[0] (path /watsonx): unknown pathType "prefix", expected one of Exact, ImplementationSpecific, Prefix'
- it: rejects an entry with no path
set:
ingress.enabled: true
ingress.extraPaths:
- service: gateway
asserts:
- failedTemplate:
errorMessage: "ingress.extraPaths[0]: 'path' is required"
- it: rejects a root entry that would take over the backend catch-all
set:
ingress.enabled: true
ingress.extraPaths:
- path: /
service: gateway
asserts:
- failedTemplate:
errorMessage: "ingress.extraPaths[0]: path / is already routed in both directions, Exact to ui and Prefix to backend, so no pathType leaves a request for an entry here to capture"
- it: rejects a root entry that would take over the UI root
set:
ingress.enabled: true
ingress.extraPaths:
- path: /
pathType: Exact
service: gateway
asserts:
- failedTemplate:
errorMessage: "ingress.extraPaths[0]: path / is already routed in both directions, Exact to ui and Prefix to backend, so no pathType leaves a request for an entry here to capture"
# A root ImplementationSpecific entry duplicates no built-in pair, so the
# duplicate check alone would admit it. It is still dead: the built-in
# Exact / sorts ahead of it on the AWS Load Balancer Controller and claims
# the only request its pattern matches, so it renders and never routes.
- it: rejects a root entry that would render but never match
set:
ingress.enabled: true
ingress.extraPaths:
- path: /
pathType: ImplementationSpecific
service: gateway
asserts:
- failedTemplate:
errorMessage: "ingress.extraPaths[0]: path / is already routed in both directions, Exact to ui and Prefix to backend, so no pathType leaves a request for an entry here to capture"
- it: rejects an entry that would take over a UI prefix
set:
ingress.enabled: true
ingress.extraPaths:
- path: /ui
service: gateway
asserts:
- failedTemplate:
errorMessage: "ingress.extraPaths[0]: path /ui with pathType Prefix is already routed by this chart, and a duplicate would take it over rather than add to it"
- it: rejects an entry that would take over the UI RSC payload rule
set:
ingress.enabled: true
ingress.extraPaths:
- path: /*.txt
pathType: ImplementationSpecific
service: backend
asserts:
- failedTemplate:
errorMessage: "ingress.extraPaths[0]: path /*.txt with pathType ImplementationSpecific is already routed by this chart, and a duplicate would take it over rather than add to it"
- it: rejects an entry that would take over a gateway data-plane prefix
set:
ingress.enabled: true
ingress.extraPaths:
- path: /v1/chat
service: backend
asserts:
- failedTemplate:
errorMessage: "ingress.extraPaths[0]: path /v1/chat with pathType Prefix is already routed by this chart, and a duplicate would take it over rather than add to it"
- it: rejects an entry that would take over the exact /test route
set:
ingress.enabled: true
ingress.extraPaths:
- path: /test
pathType: Exact
service: backend
asserts:
- failedTemplate:
errorMessage: "ingress.extraPaths[0]: path /test with pathType Exact is already routed by this chart, and a duplicate would take it over rather than add to it"
- it: allows a built-in path under a different pathType, which is a distinct rule
set:
ingress.enabled: true
ingress.extraPaths:
- path: /ui
pathType: Exact
service: ui
asserts:
- equal:
path: spec.rules[0].http.paths[-2]
value:
path: /ui
pathType: Exact
backend:
service:
name: RELEASE-NAME-litellm-ui
port:
number: 3000
- it: rejects a bare string entry instead of failing on template internals
set:
ingress.enabled: true
ingress.extraPaths:
- /watsonx
asserts:
- failedTemplate:
errorMessage: "ingress.extraPaths[0]: each entry must be a mapping with a 'path' key"

63
helm/litellm/tests/migration_job_hooks_tests.yaml Normal file
View file

@ -0,0 +1,63 @@
suite: test migrations Job hook annotations
templates:
- migrations-job.yaml
values:
- ./values/required.yaml
tests:
- it: runs as a Helm pre-install / pre-upgrade hook by default
asserts:
- equal:
path: metadata.annotations["helm.sh/hook"]
value: pre-install,pre-upgrade
- equal:
path: metadata.annotations["helm.sh/hook-delete-policy"]
value: before-hook-creation
- equal:
path: metadata.annotations["helm.sh/hook-weight"]
value: "0"
- notExists:
path: metadata.annotations["argocd.argoproj.io/hook"]
- it: adds the Argo CD PreSync hook when asked
set:
migrationJob.hooks.argocd.enabled: true
asserts:
- equal:
path: metadata.annotations["argocd.argoproj.io/hook"]
value: PreSync
- equal:
path: metadata.annotations["argocd.argoproj.io/hook-delete-policy"]
value: BeforeHookCreation
- it: drops the Helm hook so Argo CD owns the Job
set:
migrationJob.hooks.argocd.enabled: true
migrationJob.hooks.helm.enabled: false
asserts:
- equal:
path: metadata.annotations["argocd.argoproj.io/hook"]
value: PreSync
- notExists:
path: metadata.annotations["helm.sh/hook"]
- notExists:
path: metadata.annotations["helm.sh/hook-delete-policy"]
- notExists:
path: metadata.annotations["helm.sh/hook-weight"]
- it: renders an ordinary Job when both hooks are disabled
set:
migrationJob.hooks.helm.enabled: false
asserts:
- notExists:
path: metadata.annotations
- equal:
path: kind
value: Job
- it: honours a custom Helm hook weight
set:
migrationJob.hooks.helm.weight: "-5"
asserts:
- equal:
path: metadata.annotations["helm.sh/hook-weight"]
value: "-5"

66
helm/litellm/tests/rollout_strategy_tests.yaml Normal file
View file

@ -0,0 +1,66 @@
suite: test rolling update strategy on the component deployments
templates:
- gateway/deployment.yaml
- gateway/configmap.yaml
- backend/deployment.yaml
- ui/deployment.yaml
values:
- ./values/required.yaml
tests:
- it: leaves the strategy to Kubernetes defaults when unset
asserts:
- notExists:
path: spec.strategy
- it: renders the configured strategy on each deployment
set:
gateway.strategy:
type: RollingUpdate
rollingUpdate:
maxUnavailable: 0
maxSurge: 1
backend.strategy:
type: RollingUpdate
rollingUpdate:
maxUnavailable: "25%"
maxSurge: 2
ui.strategy:
type: Recreate
asserts:
- equal:
path: spec.strategy
value:
type: RollingUpdate
rollingUpdate:
maxUnavailable: 0
maxSurge: 1
template: gateway/deployment.yaml
- equal:
path: spec.strategy
value:
type: RollingUpdate
rollingUpdate:
maxUnavailable: 25%
maxSurge: 2
template: backend/deployment.yaml
- equal:
path: spec.strategy
value:
type: Recreate
template: ui/deployment.yaml
- it: keeps a component on the cluster default when only another one sets a strategy
set:
gateway.strategy:
type: Recreate
asserts:
- equal:
path: spec.strategy.type
value: Recreate
template: gateway/deployment.yaml
- notExists:
path: spec.strategy
template: backend/deployment.yaml
- notExists:
path: spec.strategy
template: ui/deployment.yaml

58
helm/litellm/values.yaml
View file

@ -13,6 +13,27 @@ ingress:
annotations: {}
host: "" # optional; if set, becomes the rule's host
tls: []
# Extra HTTP paths appended to the ingress rule. Additive: every built-in
# UI / gateway / backend path is still rendered, these entries are placed
# after them and before the backend catch-all, and an entry that repeats a
# path the chart already routes is rejected at render time rather than
# silently taking it over.
#
# The chart's built-in gateway prefix list is a snapshot of the data-plane
# surface at release time. Use extraPaths for passthrough routes it does not
# cover: a provider prefix added upstream after this chart version, or a
# custom general_settings.pass_through_endpoints route.
#
# path required; the HTTP path to route
# service which component serves it: gateway (default), backend, or ui
# pathType Prefix (default), Exact, or ImplementationSpecific
#
# The target component only answers paths its own route allowlist keeps, so
# a path here still has to be one that component serves.
extraPaths: []
# - path: /watsonx
# pathType: Prefix
# service: gateway
# Per-component ServiceAccounts for gateway, backend, and ui.
#
@ -54,6 +75,22 @@ serviceAccounts:
# generate` — the migration engine doesn't need the generated client.
migrationJob:
enabled: true
# Which controller is responsible for running the Job.
#
# `helm.enabled` renders the Helm pre-install / pre-upgrade hook, so the Job
# runs whenever `helm upgrade` sees a change to apply. `argocd.enabled`
# renders an Argo CD PreSync hook instead, which runs the Job on every sync
# even when the rendered manifests are unchanged: the way to re-run
# migrations on demand from a GitOps pipeline. Turning the Helm hook off
# while the Argo CD hook is on leaves the Job out of Helm's own upgrade
# path, which is what Argo CD users want since Argo, not Helm, applies the
# manifests.
hooks:
helm:
enabled: true
weight: "0"
argocd:
enabled: false
backoffLimit: 4
ttlSecondsAfterFinished: 120
# Wall-clock budget for the whole Job, shared across every `backoffLimit`
@ -236,6 +273,15 @@ gateway:
initialDelaySeconds: 5
periodSeconds: 10
timeoutSeconds: 10
# Rolling update tuning for the gateway Deployment. Empty by default, so
# Kubernetes applies its own RollingUpdate defaults (25% maxSurge /
# 25% maxUnavailable). Example, for a surge-only rollout behind a load
# balancer that must never lose capacity:
# type: RollingUpdate
# rollingUpdate:
# maxUnavailable: 0
# maxSurge: 1
strategy: {}
# 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
@ -348,6 +394,8 @@ backend:
initialDelaySeconds: 5
periodSeconds: 10
timeoutSeconds: 10
# Same shape as gateway.strategy.
strategy: {}
# Optional startupProbe; same shape as gateway.startupProbe. Empty by default.
startupProbe: {}
hpa:
@ -412,6 +460,8 @@ ui:
httpGet: { path: /, port: http }
initialDelaySeconds: 2
periodSeconds: 10
# Same shape as gateway.strategy.
strategy: {}
# Optional startupProbe; same shape as gateway.startupProbe. Empty by default.
startupProbe: {}
hpa:
@ -428,9 +478,11 @@ ui:
maxUnavailable: ""
podAnnotations: {}
# Same shape as the gateway blocks of the same name. The nginx runtime
# writes its pid, cache, and proxy temp files under the image's root
# filesystem, so `securityContext.readOnlyRootFilesystem: true` here needs
# emptyDir volumes mounted over those paths.
# writes its pid, cache, and proxy temp files under /tmp, so it boots as
# any (arbitrary, non-root) uid; `securityContext.readOnlyRootFilesystem:
# true` here needs an emptyDir volume mounted over /tmp. Images before
# the /tmp move instead need emptyDirs over /var/cache/nginx and /run to
# run as a non-root uid at all.
podLabels: {}
podSecurityContext: {}
securityContext: {}

12
litellm-proxy-extras/litellm_proxy_extras/migrations/20260804162853_add_budget_window_spend_table/migration.sql Normal file
View file

@ -0,0 +1,12 @@
CREATE TABLE IF NOT EXISTS "LiteLLM_BudgetWindowSpend" (
"entity_type" TEXT NOT NULL,
"entity_id" TEXT NOT NULL,
"window_duration" TEXT NOT NULL,
"window_start" TIMESTAMP(3) NOT NULL,
"spend" DOUBLE PRECISION NOT NULL DEFAULT 0.0,
"created_at" TIMESTAMP(3) NOT NULL DEFAULT CURRENT_TIMESTAMP,
"updated_at" TIMESTAMP(3) NOT NULL,
CONSTRAINT "LiteLLM_BudgetWindowSpend_pkey" PRIMARY KEY ("entity_type","entity_id","window_duration")
);

4
litellm-proxy-extras/litellm_proxy_extras/migrations/20260819000000_backfill_spend_log_timestamps/migration.sql
View file

@ -1,4 +0,0 @@
UPDATE "LiteLLM_SpendLogs"
SET "created_at" = "endTime",
"updated_at" = "endTime"
WHERE "created_at" > "endTime" + interval '1 hour';

18
litellm-proxy-extras/litellm_proxy_extras/migrations/20260824000000_add_gateway_injected_caching_savings_spend/migration.sql Normal file
View file

@ -0,0 +1,18 @@
-- AlterTable
ALTER TABLE "LiteLLM_DailyUserSpend" ADD COLUMN IF NOT EXISTS "gateway_injected_caching_savings_spend" DOUBLE PRECISION NOT NULL DEFAULT 0.0;
-- AlterTable
ALTER TABLE "LiteLLM_DailyOrganizationSpend" ADD COLUMN IF NOT EXISTS "gateway_injected_caching_savings_spend" DOUBLE PRECISION NOT NULL DEFAULT 0.0;
-- AlterTable
ALTER TABLE "LiteLLM_DailyEndUserSpend" ADD COLUMN IF NOT EXISTS "gateway_injected_caching_savings_spend" DOUBLE PRECISION NOT NULL DEFAULT 0.0;
-- AlterTable
ALTER TABLE "LiteLLM_DailyAgentSpend" ADD COLUMN IF NOT EXISTS "gateway_injected_caching_savings_spend" DOUBLE PRECISION NOT NULL DEFAULT 0.0;
-- AlterTable
ALTER TABLE "LiteLLM_DailyTeamSpend" ADD COLUMN IF NOT EXISTS "gateway_injected_caching_savings_spend" DOUBLE PRECISION NOT NULL DEFAULT 0.0;
-- AlterTable
ALTER TABLE "LiteLLM_DailyTagSpend" ADD COLUMN IF NOT EXISTS "gateway_injected_caching_savings_spend" DOUBLE PRECISION NOT NULL DEFAULT 0.0;

22
litellm-proxy-extras/litellm_proxy_extras/migrations/20260828000000_shadow_eval_cost_comparison/migration.sql Normal file
View file

@ -0,0 +1,22 @@
-- AlterTable
ALTER TABLE "LiteLLM_ShadowEvalAttempt" ADD COLUMN IF NOT EXISTS "real_cost" DOUBLE PRECISION;
-- AlterTable
ALTER TABLE "LiteLLM_ShadowEvalAttempt" ADD COLUMN IF NOT EXISTS "real_classifier_cost" DOUBLE PRECISION NOT NULL DEFAULT 0;
-- AlterTable
ALTER TABLE "LiteLLM_ShadowEvalAttempt" ADD COLUMN IF NOT EXISTS "shadow_classifier_cost" DOUBLE PRECISION NOT NULL DEFAULT 0;
-- AlterTable
ALTER TABLE "LiteLLM_ShadowEvalAttempt" ADD COLUMN IF NOT EXISTS "real_cache_hit" BOOLEAN NOT NULL DEFAULT false;
-- CreateTable
CREATE TABLE IF NOT EXISTS "LiteLLM_ShadowEvalFunnel" (
"job_id" TEXT NOT NULL,
"not_sampled" INTEGER NOT NULL DEFAULT 0,
"unjudgeable" INTEGER NOT NULL DEFAULT 0,
"shed" INTEGER NOT NULL DEFAULT 0,
"withheld" INTEGER NOT NULL DEFAULT 0,
CONSTRAINT "LiteLLM_ShadowEvalFunnel_pkey" PRIMARY KEY ("job_id")
);

20
litellm-proxy-extras/litellm_proxy_extras/migrations/20260829000000_add_model_access_group_budget_table/migration.sql Normal file
View file

@ -0,0 +1,20 @@
-- CreateTable
CREATE TABLE IF NOT EXISTS "LiteLLM_ModelAccessGroupBudgetTable" (
"access_group_name" TEXT NOT NULL,
"spend" DOUBLE PRECISION NOT NULL DEFAULT 0.0,
"budget_id" TEXT,
"created_at" TIMESTAMP(3) NOT NULL DEFAULT CURRENT_TIMESTAMP,
"created_by" TEXT,
"updated_at" TIMESTAMP(3) NOT NULL DEFAULT CURRENT_TIMESTAMP,
"updated_by" TEXT,
CONSTRAINT "LiteLLM_ModelAccessGroupBudgetTable_pkey" PRIMARY KEY ("access_group_name")
);
-- AddForeignKey
DO $$
BEGIN
IF NOT EXISTS (SELECT 1 FROM pg_constraint WHERE conname = 'LiteLLM_ModelAccessGroupBudgetTable_budget_id_fkey') THEN
ALTER TABLE "LiteLLM_ModelAccessGroupBudgetTable" ADD CONSTRAINT "LiteLLM_ModelAccessGroupBudgetTable_budget_id_fkey" FOREIGN KEY ("budget_id") REFERENCES "LiteLLM_BudgetTable"("budget_id") ON DELETE SET NULL ON UPDATE CASCADE;
END IF;
END $$;

21
litellm-proxy-extras/litellm_proxy_extras/migrations/20260831000000_shadow_eval_typed_targets/migration.sql Normal file
View file

@ -0,0 +1,21 @@
DO $$
BEGIN
IF EXISTS (
SELECT 1 FROM information_schema.columns
WHERE table_name = 'LiteLLM_ShadowEvalJob' AND column_name = 'api_key_id'
) THEN
ALTER TABLE "LiteLLM_ShadowEvalJob" RENAME COLUMN "api_key_id" TO "target_id";
END IF;
END $$;
ALTER TABLE "LiteLLM_ShadowEvalJob" ADD COLUMN IF NOT EXISTS "target_type" TEXT NOT NULL DEFAULT 'key';
DROP INDEX IF EXISTS "LiteLLM_ShadowEvalJob_one_active_per_key_direction";
CREATE UNIQUE INDEX IF NOT EXISTS "LiteLLM_ShadowEvalJob_one_active_per_target_direction"
ON "LiteLLM_ShadowEvalJob"("target_type", "target_id", "direction") WHERE "stopped_at" IS NULL;
DROP INDEX IF EXISTS "LiteLLM_ShadowEvalJob_api_key_id_idx";
CREATE INDEX IF NOT EXISTS "LiteLLM_ShadowEvalJob_target_type_target_id_idx"
ON "LiteLLM_ShadowEvalJob"("target_type", "target_id");

3
litellm-proxy-extras/litellm_proxy_extras/migrations/20260901000000_shadow_eval_multi_router/migration.sql Normal file
View file

@ -0,0 +1,3 @@
ALTER TABLE "LiteLLM_ShadowEvalJob" ADD COLUMN IF NOT EXISTS "router_names" TEXT[] NOT NULL DEFAULT ARRAY[]::TEXT[];
ALTER TABLE "LiteLLM_ShadowEvalAttempt" ADD COLUMN IF NOT EXISTS "router_name" TEXT;

59
litellm-proxy-extras/litellm_proxy_extras/schema.prisma
View file

@ -29,6 +29,7 @@ model LiteLLM_BudgetTable {
keys LiteLLM_VerificationToken[] // multiple keys can have the same budget
end_users LiteLLM_EndUserTable[] // multiple end-users can have the same budget
tags LiteLLM_TagTable[] // multiple tags can have the same budget
model_access_groups LiteLLM_ModelAccessGroupBudgetTable[] // multiple model access groups can have the same budget
team_membership LiteLLM_TeamMembership[] // budgets of Users within a Team
organization_membership LiteLLM_OrganizationMembership[] // budgets of Users within a Organization
}
@ -585,6 +586,20 @@ model LiteLLM_EndUserTable {
blocked Boolean @default(false)
}
// Budget and shared spend for a model access group. The groups themselves are not rows anywhere:
// they are free-text strings in LiteLLM_ProxyModelTable.model_info.access_groups, so a row here
// exists only once someone gives that group a budget.
model LiteLLM_ModelAccessGroupBudgetTable {
access_group_name String @id
spend Float @default(0.0)
budget_id String?
litellm_budget_table LiteLLM_BudgetTable? @relation(fields: [budget_id], references: [budget_id])
created_at DateTime @default(now()) @map("created_at")
created_by String?
updated_at DateTime @default(now()) @updatedAt
updated_by String?
}
// Track tags with budgets and spend
model LiteLLM_TagTable {
tag_name String @id
@ -649,6 +664,18 @@ model LiteLLM_SpendLogs {
@@index([session_id])
}
model LiteLLM_BudgetWindowSpend {
entity_type String
entity_id String
window_duration String
window_start DateTime
spend Float @default(0.0)
created_at DateTime @default(now())
updated_at DateTime @updatedAt
@@id([entity_type, entity_id, window_duration])
}
// View spend, model, api_key per request
model LiteLLM_ErrorLogs {
request_id String @id @default(uuid())
@ -754,6 +781,7 @@ model LiteLLM_DailyUserSpend {
compression_saved_tokens BigInt @default(0)
compression_savings_spend Float @default(0.0)
prompt_caching_savings_spend Float @default(0.0)
gateway_injected_caching_savings_spend Float @default(0.0)
autorouter_savings_spend Float @default(0.0)
spend Float @default(0.0)
api_requests BigInt @default(0)
@ -789,6 +817,7 @@ model LiteLLM_DailyOrganizationSpend {
compression_saved_tokens BigInt @default(0)
compression_savings_spend Float @default(0.0)
prompt_caching_savings_spend Float @default(0.0)
gateway_injected_caching_savings_spend Float @default(0.0)
autorouter_savings_spend Float @default(0.0)
spend Float @default(0.0)
api_requests BigInt @default(0)
@ -824,6 +853,7 @@ model LiteLLM_DailyEndUserSpend {
compression_saved_tokens BigInt @default(0)
compression_savings_spend Float @default(0.0)
prompt_caching_savings_spend Float @default(0.0)
gateway_injected_caching_savings_spend Float @default(0.0)
autorouter_savings_spend Float @default(0.0)
spend Float @default(0.0)
api_requests BigInt @default(0)
@ -858,6 +888,7 @@ model LiteLLM_DailyAgentSpend {
compression_saved_tokens BigInt @default(0)
compression_savings_spend Float @default(0.0)
prompt_caching_savings_spend Float @default(0.0)
gateway_injected_caching_savings_spend Float @default(0.0)
autorouter_savings_spend Float @default(0.0)
spend Float @default(0.0)
api_requests BigInt @default(0)
@ -892,6 +923,7 @@ model LiteLLM_DailyTeamSpend {
compression_saved_tokens BigInt @default(0)
compression_savings_spend Float @default(0.0)
prompt_caching_savings_spend Float @default(0.0)
gateway_injected_caching_savings_spend Float @default(0.0)
autorouter_savings_spend Float @default(0.0)
spend Float @default(0.0)
api_requests BigInt @default(0)
@ -929,6 +961,7 @@ model LiteLLM_DailyTagSpend {
compression_saved_tokens BigInt @default(0)
compression_savings_spend Float @default(0.0)
prompt_caching_savings_spend Float @default(0.0)
gateway_injected_caching_savings_spend Float @default(0.0)
autorouter_savings_spend Float @default(0.0)
spend Float @default(0.0)
api_requests BigInt @default(0)
@ -1496,14 +1529,16 @@ model LiteLLM_AutoRouterSession {
model LiteLLM_ShadowEvalJob {
id String @id @default(cuid())
group_id String // legs of one job share this; the API's job id
api_key_id String // hashed virtual key whose traffic this leg shadows
router_name String // the auto-router under evaluation, in either direction
target_type String @default("key") // key | team | user
target_id String // hashed virtual key, team_id, or user_id whose traffic this leg shadows
router_name String // first (often only) auto-router under evaluation; router_names is the full set
router_names String[] @default([]) // all routers this job runs as shadow arms; empty on legacy rows, whose set is (router_name)
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-count ceiling: the whole budget on pre-max_budget jobs, the error-loop valve otherwise
max_budget Float? // per-key USD cap on the eval's own shadow + judge spend; null on jobs from before spend budgets
max_budget Float? // per-target USD cap on the eval's own shadow + judge spend; null on jobs from before spend budgets
created_at DateTime @default(now())
created_by String?
ends_at DateTime
@ -1511,7 +1546,7 @@ model LiteLLM_ShadowEvalJob {
stopped_by String? // operator who stopped it early; null when it ended on its own
@@index([group_id])
@@index([api_key_id])
@@index([target_type, target_id])
@@index([created_at])
}
@ -1521,18 +1556,34 @@ model LiteLLM_ShadowEvalAttempt {
job_id String
request_id String // the judged real request
outcome String // real | shadow | tie | error
router_name String? // the arm this verdict scores; NULL on legacy rows, meaning the job's own router
tier String? // router's tier for the prompt, when classified
real_model String?
shadow_model String?
confidence Float?
judge_cost Float @default(0)
shadow_cost Float @default(0)
real_cost Float? // NULL = row predates cost measurement; comparisons read only measured rows
real_classifier_cost Float @default(0)
shadow_classifier_cost Float @default(0)
real_cache_hit Boolean @default(false)
error String?
created_at DateTime @default(now())
@@index([job_id])
}
// Per-leg sampling funnel counters the attempt rows cannot derive: requests an
// admitting job saw but did not judge. attempted = the leg's attempt rows; the
// leg's eligible traffic = not_sampled + unjudgeable + shed + withheld + attempted.
model LiteLLM_ShadowEvalFunnel {
job_id String @id
not_sampled Int @default(0)
unjudgeable Int @default(0)
shed Int @default(0)
withheld Int @default(0)
}
// ---------------------------------------------------------------------------
// Workflow Run Tracking
//

148
litellm-proxy-extras/litellm_proxy_extras/utils.py
View file

@ -40,6 +40,65 @@ def _get_prisma_env() -> dict:
_MIGRATION_TS_RE = re.compile(r"^(\d{14})_")
_SPEND_LOGS_ALTER_RE = re.compile(r'^ALTER\s+TABLE\s+"LiteLLM_SpendLogs"\s', re.IGNORECASE)
_SPEND_LOGS_ARTIFACT_DROP_RE = re.compile(
r'^DROP\s+TABLE\s+"LiteLLM_SpendLogs_[^"]*"', re.IGNORECASE
)
_SPEND_LOGS_PK_CLAUSE_RE = re.compile(
r'^(?:DROP\s+CONSTRAINT\s+"[^"]*_pkey"'
r'|ADD\s+(?:CONSTRAINT\s+"[^"]*"\s+)?PRIMARY\s+KEY\s*\([^)]*\))$',
re.IGNORECASE,
)
PARTITIONED_SPEND_LOGS_PUSH_ERROR = (
"LiteLLM_SpendLogs is a partitioned table (see db_scripts/partition_spend_logs.sql), "
"so its primary key must include the partition key (\"startTime\"). `prisma db push` "
"reconciles the database against schema.prisma, which declares the unpartitioned "
"primary key (\"request_id\"), and Postgres rejects that rewrite with: unique "
"constraint on partitioned table must include all partitioning columns. Start the "
"proxy without --use_prisma_db_push so it uses `prisma migrate deploy`, which only "
"applies shipped migrations and leaves the partitioned primary key alone."
)
def _without_sql_comments(statement: str) -> str:
return "\n".join(
line
for line in statement.splitlines()
if line.strip() and not line.strip().startswith("--")
).strip()
def _without_spend_logs_pk_clauses(statement: str) -> Optional[str]:
prefix_match = _SPEND_LOGS_ALTER_RE.match(statement)
if not prefix_match:
return statement
kept = tuple(
clause.strip()
for clause in statement[prefix_match.end():].split(",\n")
if not _SPEND_LOGS_PK_CLAUSE_RE.match(clause.strip())
)
if not kept:
return None
return statement[: prefix_match.end()] + ",\n".join(kept)
def filter_partitioned_spend_logs_diff(diff_sql: str) -> str:
"""Drop statements from a `prisma migrate diff` script that fight the
SpendLogs partitioning runbook (db_scripts/partition_spend_logs.sql): the
primary-key rewrite on "LiteLLM_SpendLogs", which Postgres rejects on a
partitioned table, and drops of runbook artifacts such as
"LiteLLM_SpendLogs_legacy"."""
kept = tuple(
filtered
for statement in diff_sql.split(";")
for bare in (_without_sql_comments(statement),)
if bare and not _SPEND_LOGS_ARTIFACT_DROP_RE.match(bare)
for filtered in (_without_spend_logs_pk_clauses(bare),)
if filtered is not None
)
return "".join(f"{statement};\n\n" for statement in kept)
def _migration_timestamp(name: str) -> int:
"""Extract the leading `YYYYMMDDHHMMSS` timestamp from a migration name.
@ -355,7 +414,24 @@ class ProxyExtrasDBManager:
return
logger.info(f"Migration diff created at {diff_sql_path}")
if ProxyExtrasDBManager.spend_logs_is_partitioned():
filtered_sql = filter_partitioned_spend_logs_diff(
diff_sql_path.read_text()
)
diff_sql_path.write_text(filtered_sql)
logger.info(
"LiteLLM_SpendLogs is partitioned; removed its primary-key "
"rewrite and partitioning artifacts from the drift script"
)
if not filtered_sql.strip():
logger.info("Drift script is empty after filtering; nothing to apply")
if not mark_all_applied:
return
ProxyExtrasDBManager._mark_migrations_applied(migrations_dir)
return
# 2. Run prisma db execute to apply the migration
applied_ok = False
try:
logger.info("Running prisma db execute to apply the migration diff...")
result = subprocess.run(
@ -376,6 +452,7 @@ class ProxyExtrasDBManager:
)
logger.info(f"prisma db execute stdout: {result.stdout}")
logger.info("✅ Migration diff applied successfully")
applied_ok = True
except subprocess.CalledProcessError as e:
logger.warning(f"Failed to apply migration diff: {e.stderr}")
except subprocess.TimeoutExpired:
@ -384,6 +461,16 @@ class ProxyExtrasDBManager:
# 3. Mark all migrations as applied
if not mark_all_applied:
return
if not applied_ok:
logger.warning(
"Drift script failed to apply; NOT marking migrations as "
"applied so a later migration run can retry them"
)
return
ProxyExtrasDBManager._mark_migrations_applied(migrations_dir)
@staticmethod
def _mark_migrations_applied(migrations_dir: str) -> None:
migration_names = ProxyExtrasDBManager._get_migration_names(migrations_dir)
logger.info(f"Resolving {len(migration_names)} migrations")
for migration_name in migration_names:
@ -410,6 +497,62 @@ class ProxyExtrasDBManager:
f"Failed to resolve migration {migration_name}: {e.stderr}"
)
@staticmethod
def spend_logs_is_partitioned() -> bool:
"""True when the connected database's LiteLLM_SpendLogs is a
partitioned table in Prisma's target schema (the `schema` URL param,
falling back to Prisma's default target, public), i.e. the operator
ran db_scripts/partition_spend_logs.sql. Returns False when psycopg is
unavailable or the database cannot be reached, preserving the
pre-existing behavior in those cases."""
database_url = os.getenv("DATABASE_URL")
if not database_url:
return False
try:
import psycopg
except ImportError:
logger.warning(
"psycopg is not installed; skipping the LiteLLM_SpendLogs "
"partition check. If this table is partitioned (see "
"db_scripts/partition_spend_logs.sql), schema reconciliation "
"will try to rewrite its primary key and fail. Install the "
"litellm[extra_proxy] extra, which now includes psycopg."
)
return False
cleaned_url = ProxyExtrasDBManager._strip_prisma_query_params(database_url)
try:
with psycopg.connect(
cleaned_url, connect_timeout=10, autocommit=True
) as conn:
row = conn.execute(
"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 = 'LiteLLM_SpendLogs' "
" AND n.nspname = %s",
(
ProxyExtrasDBManager._prisma_schema_param(database_url)
or "public",
),
).fetchone()
except (psycopg.OperationalError, psycopg.DatabaseError):
return False
return row is not None
@staticmethod
def _prisma_schema_param(url: str) -> Optional[str]:
"""The `schema` query param Prisma uses to pick its target schema,
or None when the URL does not set one."""
from urllib.parse import urlparse, parse_qsl
return next(
(v for k, v in parse_qsl(urlparse(url).query) if k == "schema"),
None,
)
@staticmethod
def _strip_prisma_query_params(url: str) -> str:
"""Remove Prisma-specific query params (connection_limit, pool_timeout,
@ -528,7 +671,8 @@ class ProxyExtrasDBManager:
migrations_dir = ProxyExtrasDBManager._get_prisma_dir()
if not use_migrate:
# Preserve `prisma db push` path unchanged.
if ProxyExtrasDBManager.spend_logs_is_partitioned():
raise RuntimeError(PARTITIONED_SPEND_LOGS_PUSH_ERROR)
original_dir = os.getcwd()
os.chdir(migrations_dir)
try:
@ -972,6 +1116,8 @@ class ProxyExtrasDBManager:
)
raise
else:
if ProxyExtrasDBManager.spend_logs_is_partitioned():
raise RuntimeError(PARTITIONED_SPEND_LOGS_PUSH_ERROR)
# Use prisma db push with increased timeout
subprocess.run(
[_get_prisma_command(), "db", "push", "--accept-data-loss"],

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

@ -1,6 +1,6 @@
[project]
name = "litellm-proxy-extras"
version = "0.4.88"
version = "0.4.92"
description = "Additional files for the LiteLLM Proxy. Reduces the size of the main litellm package."
readme = "README.md"
requires-python = ">=3.9"
@ -26,7 +26,7 @@ required-version = ">=0.10.9"
module-root = ""
[tool.commitizen]
version = "0.4.88"
version = "0.4.92"
version_files = [
"pyproject.toml:^version",
"../pyproject.toml:litellm-proxy-extras==",

340
litellm-rust/Cargo.lock generated
View file

@ -2,6 +2,36 @@
# It is not intended for manual editing.
version = 4
[[package]]
name = "aho-corasick"
version = "1.1.5"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "c982642fa9e8606056828ee9a8505737230110bb1099153c79efe865c59d12ba"
dependencies = [
"memchr",
]
[[package]]
name = "alloca"
version = "0.4.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "e5a7d05ea6aea7e9e64d25b9156ba2fee3fdd659e34e41063cd2fc7cd020d7f4"
dependencies = [
"cc",
]
[[package]]
name = "anes"
version = "0.1.6"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "4b46cbb362ab8752921c97e041f5e366ee6297bd428a31275b9fcf1e380f7299"
[[package]]
name = "anstyle"
version = "1.0.14"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "940b3a0ca603d1eade50a4846a2afffd5ef57a9feac2c0e2ec2e14f9ead76000"
[[package]]
name = "arc-swap"
version = "1.9.2"
@ -506,6 +536,12 @@ dependencies = [
"either",
]
[[package]]
name = "cast"
version = "0.3.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "37b2a672a2cb129a2e41c10b1224bb368f9f37a2b16b612598138befd7b37eb5"
[[package]]
name = "cc"
version = "1.3.0"
@ -541,6 +577,58 @@ dependencies = [
"rand_core 0.10.1",
]
[[package]]
name = "ciborium"
version = "0.2.2"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "42e69ffd6f0917f5c029256a24d0161db17cea3997d185db0d35926308770f0e"
dependencies = [
"ciborium-io",
"ciborium-ll",
"serde",
]
[[package]]
name = "ciborium-io"
version = "0.2.2"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "05afea1e0a06c9be33d539b876f1ce3692f4afea2cb41f740e7743225ed1c757"
[[package]]
name = "ciborium-ll"
version = "0.2.2"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "57663b653d948a338bfb3eeba9bb2fd5fcfaecb9e199e87e1eda4d9e8b240fd9"
dependencies = [
"ciborium-io",
"half",
]
[[package]]
name = "clap"
version = "4.6.6"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "473c7e07f409a8d772161724aa8db6a765a2532a70f9667eeb7b49d3d02fbdca"
dependencies = [
"clap_builder",
]
[[package]]
name = "clap_builder"
version = "4.6.6"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "7b48fea5a88e9ae728a2dcbedbfc0e730f7d60da42e1cb049a83c9fb8b789889"
dependencies = [
"anstyle",
"clap_lex",
]
[[package]]
name = "clap_lex"
version = "1.1.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "c8d4a3bb8b1e0c1050499d1815f5ab16d04f0959b233085fb31653fbfc9d98f9"
[[package]]
name = "cmake"
version = "0.1.58"
@ -596,6 +684,72 @@ dependencies = [
"libc",
]
[[package]]
name = "criterion"
version = "0.8.2"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "950046b2aa2492f9a536f5f4f9a3de7b9e2476e575e05bd6c333371add4d98f3"
dependencies = [
"alloca",
"anes",
"cast",
"ciborium",
"clap",
"criterion-plot",
"itertools",
"num-traits",
"oorandom",
"page_size",
"plotters",
"rayon",
"regex",
"serde",
"serde_json",
"tinytemplate",
"walkdir",
]
[[package]]
name = "criterion-plot"
version = "0.8.2"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "d8d80a2f4f5b554395e47b5d8305bc3d27813bacb73493eb1001e8f76dae29ea"
dependencies = [
"cast",
"itertools",
]
[[package]]
name = "crossbeam-deque"
version = "0.8.7"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "5181e0de7b61eb03a81e347d6dd8797bae9da5146707b51077e2d71a54ec0ceb"
dependencies = [
"crossbeam-epoch",
"crossbeam-utils",
]
[[package]]
name = "crossbeam-epoch"
version = "0.9.20"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "2d6914041f254d6e9176c01941b21115dcfb7089e55135a35411081bd106ef3f"
dependencies = [
"crossbeam-utils",
]
[[package]]
name = "crossbeam-utils"
version = "0.8.22"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "61803da095bee82a81bb1a452ecc25d3b2f1416d1897eb86430c6159ef717c17"
[[package]]
name = "crunchy"
version = "0.2.4"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "460fbee9c2c2f33933d720630a6a0bac33ba7053db5344fac858d4b8952d77d5"
[[package]]
name = "crypto-common"
version = "0.1.7"
@ -856,6 +1010,17 @@ dependencies = [
"tracing",
]
[[package]]
name = "half"
version = "2.7.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "6ea2d84b969582b4b1864a92dc5d27cd2b77b622a8d79306834f1be5ba20d84b"
dependencies = [
"cfg-if",
"crunchy",
"zerocopy",
]
[[package]]
name = "hashbrown"
version = "0.17.1"
@ -1179,6 +1344,15 @@ version = "2.12.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "d98f6fed1fde3f8c21bc40a1abb88dd75e67924f9cffc3ef95607bad8017f8e2"
[[package]]
name = "itertools"
version = "0.13.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "413ee7dfc52ee1a4949ceeb7dbc8a33f2d6c088194d9f922fb8318faf1f01186"
dependencies = [
"either",
]
[[package]]
name = "itoa"
version = "1.0.18"
@ -1255,10 +1429,13 @@ dependencies = [
name = "litellm-python-bridge"
version = "0.1.0"
dependencies = [
"criterion",
"litellm-ai-gateway",
"litellm-core",
"pyo3",
"pyo3-async-runtimes",
"pythonize",
"serde",
"serde_json",
"tokio",
]
@ -1340,6 +1517,12 @@ version = "1.21.4"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "9f7c3e4beb33f85d45ae3e3a1792185706c8e16d043238c593331cc7cd313b50"
[[package]]
name = "oorandom"
version = "11.1.5"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "d6790f58c7ff633d8771f42965289203411a5e5c68388703c06e14f24770b41e"
[[package]]
name = "openssl-probe"
version = "0.2.1"
@ -1352,6 +1535,16 @@ version = "0.5.2"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "1a80800c0488c3a21695ea981a54918fbb37abf04f4d0720c453632255e2ff0e"
[[package]]
name = "page_size"
version = "0.6.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "30d5b2194ed13191c1999ae0704b7839fb18384fa22e49b57eeaa97d79ce40da"
dependencies = [
"libc",
"winapi",
]
[[package]]
name = "percent-encoding"
version = "2.3.2"
@ -1376,6 +1569,34 @@ version = "0.3.33"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "19f132c84eca552bf34cab8ec81f1c1dcc229b811638f9d283dceabe58c5569e"
[[package]]
name = "plotters"
version = "0.3.7"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "5aeb6f403d7a4911efb1e33402027fc44f29b5bf6def3effcc22d7bb75f2b747"
dependencies = [
"num-traits",
"plotters-backend",
"plotters-svg",
"wasm-bindgen",
"web-sys",
]
[[package]]
name = "plotters-backend"
version = "0.3.7"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "df42e13c12958a16b3f7f4386b9ab1f3e7933914ecea48da7139435263a4172a"
[[package]]
name = "plotters-svg"
version = "0.3.7"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "51bae2ac328883f7acdfea3d66a7c35751187f870bc81f94563733a154d7a670"
dependencies = [
"plotters-backend",
]
[[package]]
name = "portable-atomic"
version = "1.14.0"
@ -1486,6 +1707,16 @@ dependencies = [
"syn 2.0.119",
]
[[package]]
name = "pythonize"
version = "0.29.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "6ec376e1216e0c929a74964ce2020012a1a39f32d80e78aa688721219ea7fb89"
dependencies = [
"pyo3",
"serde",
]
[[package]]
name = "quinn"
version = "0.11.11"
@ -1613,12 +1844,61 @@ dependencies = [
"rand_core 0.10.1",
]
[[package]]
name = "rayon"
version = "1.12.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "fb39b166781f92d482534ef4b4b1b2568f42613b53e5b6c160e24cfbfa30926d"
dependencies = [
"either",
"rayon-core",
]
[[package]]
name = "rayon-core"
version = "1.13.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "22e18b0f0062d30d4230b2e85ff77fdfe4326feb054b9783a3460d8435c8ab91"
dependencies = [
"crossbeam-deque",
"crossbeam-utils",
]
[[package]]
name = "regex"
version = "1.13.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "f020237b6c8eed93db2e2cb53c00c60a8e1bc73da7d073199a1180401450218d"
dependencies = [
"aho-corasick",
"memchr",
"regex-automata",
"regex-syntax",
]
[[package]]
name = "regex-automata"
version = "0.4.18"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "ad8553b9b26413251cbf30e620595c7a41b3887f03da04579c0e6b0d6a06b4b2"
dependencies = [
"aho-corasick",
"memchr",
"regex-syntax",
]
[[package]]
name = "regex-lite"
version = "0.1.9"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "cab834c73d247e67f4fae452806d17d3c7501756d98c8808d7c9c7aa7d18f973"
[[package]]
name = "regex-syntax"
version = "0.8.11"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "d6f6ff9a378485b298a5286656da665ba74413d36db0979633275d2e708145d4"
[[package]]
name = "reqwest"
version = "0.12.28"
@ -1774,6 +2054,15 @@ version = "1.0.23"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "9774ba4a74de5f7b1c1451ed6cd5285a32eddb5cccb8cc655a4e50009e06477f"
[[package]]
name = "same-file"
version = "1.0.6"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "93fc1dc3aaa9bfed95e02e6eadabb4baf7e3078b0bd1b4d7b6b0b68378900502"
dependencies = [
"winapi-util",
]
[[package]]
name = "schannel"
version = "0.1.29"
@ -2099,6 +2388,16 @@ dependencies = [
"zerovec",
]
[[package]]
name = "tinytemplate"
version = "1.2.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "be4d6b5f19ff7664e8c98d03e2139cb510db9b0a60b55f8e8709b689d939b6bc"
dependencies = [
"serde",
"serde_json",
]
[[package]]
name = "tinyvec"
version = "1.12.0"
@ -2363,6 +2662,16 @@ version = "0.8.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "5c3082ca00d5a5ef149bb8b555a72ae84c9c59f7250f013ac822ac2e49b19c64"
[[package]]
name = "walkdir"
version = "2.5.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "29790946404f91d9c5d06f9874efddea1dc06c5efe94541a7d6863108e3a5e4b"
dependencies = [
"same-file",
"winapi-util",
]
[[package]]
name = "want"
version = "0.3.1"
@ -2475,6 +2784,37 @@ dependencies = [
"rustls-pki-types",
]
[[package]]
name = "winapi"
version = "0.3.9"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "5c839a674fcd7a98952e593242ea400abe93992746761e38641405d28b00f419"
dependencies = [
"winapi-i686-pc-windows-gnu",
"winapi-x86_64-pc-windows-gnu",
]
[[package]]
name = "winapi-i686-pc-windows-gnu"
version = "0.4.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "ac3b87c63620426dd9b991e5ce0329eff545bccbbb34f3be09ff6fb6ab51b7b6"
[[package]]
name = "winapi-util"
version = "0.1.11"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "c2a7b1c03c876122aa43f3020e6c3c3ee5c05081c9a00739faf7503aeba10d22"
dependencies = [
"windows-sys 0.61.2",
]
[[package]]
name = "winapi-x86_64-pc-windows-gnu"
version = "0.4.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "712e227841d057c1ee1cd2fb22fa7e5a5461ae8e48fa2ca79ec42cfc1931183f"
[[package]]
name = "windows-link"
version = "0.2.1"

10
litellm-rust/Cargo.toml
View file

@ -18,6 +18,7 @@ litellm-ai-gateway = { path = "crates/ai-gateway", default-features = false }
axum = "0.7"
pyo3 = "0.29.0"
pyo3-async-runtimes = { version = "0.29.0", features = ["tokio-runtime"] }
pythonize = "0.29.0"
rand = "0.8"
reqwest = { version = "0.12", default-features = false, features = ["blocking", "json", "rustls-tls", "http2", "stream"] }
serde = { version = "1.0", features = ["derive"] }
@ -29,3 +30,12 @@ tokio = { version = "1", features = ["rt-multi-thread", "macros", "time", "net"]
tokio-tungstenite = { version = "0.24", default-features = false, features = ["connect", "rustls-tls-native-roots"] }
futures-util = { version = "0.3", default-features = false, features = ["sink", "std"] }
base64 = "0.22"
[profile.release]
opt-level = 3
lto = "thin"
codegen-units = 1
panic = "unwind"
debug = false
incremental = false
strip = "symbols"

16
litellm-rust/crates/python-bridge/Cargo.toml
View file

@ -9,10 +9,24 @@ repository.workspace = true
name = "_native"
crate-type = ["cdylib"]
[features]
default = ["abi3"]
abi3 = ["pyo3/abi3-py310"]
extension-module = ["pyo3/extension-module"]
[dependencies]
litellm-core = { workspace = true, features = ["bedrock-auth"] }
litellm-ai-gateway = { workspace = true, default-features = false }
pyo3 = { workspace = true, features = ["extension-module"] }
pyo3.workspace = true
pyo3-async-runtimes.workspace = true
pythonize.workspace = true
serde.workspace = true
serde_json.workspace = true
tokio.workspace = true
[dev-dependencies]
criterion = "0.8.2"
[[bench]]
name = "serialization"
harness = false

103
litellm-rust/crates/python-bridge/benches/serialization.rs Normal file
View file

@ -0,0 +1,103 @@
use std::hint::black_box;
use std::time::Duration;
use criterion::{BenchmarkId, Criterion, criterion_group, criterion_main};
use pyo3::prelude::*;
use pyo3::types::PyDict;
use serde_json::{Value, json};
const PAYLOAD_SIZES: &[(&str, usize)] = &[
("1_KiB", 1024),
("64_KiB", 64 * 1024),
("1_MiB", 1024 * 1024),
("4_MiB", 4 * 1024 * 1024),
("16_MiB", 16 * 1024 * 1024),
];
fn former_json_roundtrip_from_py(py: Python<'_>, value: &Bound<'_, PyAny>) -> Value {
let json = py.import("json").expect("Python json module should import");
let encoded: String = json
.call_method1("dumps", (value,))
.expect("payload should serialize")
.extract()
.expect("json.dumps should return a string");
serde_json::from_str(&encoded).expect("serialized JSON should parse")
}
fn pythonize_from_py(value: &Bound<'_, PyAny>) -> Value {
pythonize::depythonize(value).expect("payload should depythonize")
}
fn former_json_roundtrip_to_py(py: Python<'_>, value: &Value) -> Py<PyAny> {
let json = py.import("json").expect("Python json module should import");
let encoded = serde_json::to_string(value).expect("response should serialize");
json.call_method1("loads", (encoded,))
.expect("serialized response should parse in Python")
.unbind()
}
fn pythonize_to_py(py: Python<'_>, value: &Value) -> Py<PyAny> {
pythonize::pythonize(py, value)
.expect("response should pythonize")
.unbind()
}
fn serialization(c: &mut Criterion) {
Python::initialize();
Python::attach(|py| {
for &(label, payload_bytes) in PAYLOAD_SIZES {
let data_uri = format!("data:image/png;base64,{}", "A".repeat(payload_bytes));
let document = PyDict::new(py);
document
.set_item("type", "image_url")
.expect("document type should be set");
document
.set_item("image_url", &data_uri)
.expect("document URL should be set");
let response = json!({
"pages": [{
"index": 0,
"markdown": "OCR text",
"images": [{"image_base64": data_uri}],
}],
"model": "mistral-ocr-latest",
"document_annotation": null,
"usage_info": {"pages_processed": 1},
"object": "ocr",
});
c.bench_with_input(
BenchmarkId::new("python_to_rust_json", label),
&document,
|b, document| {
b.iter(|| former_json_roundtrip_from_py(py, black_box(document.as_any())))
},
);
c.bench_with_input(
BenchmarkId::new("python_to_rust_pythonize", label),
&document,
|b, document| b.iter(|| pythonize_from_py(black_box(document.as_any()))),
);
c.bench_with_input(
BenchmarkId::new("rust_to_python_json", label),
&response,
|b, response| b.iter(|| former_json_roundtrip_to_py(py, black_box(response))),
);
c.bench_with_input(
BenchmarkId::new("rust_to_python_pythonize", label),
&response,
|b, response| b.iter(|| pythonize_to_py(py, black_box(response))),
);
}
});
}
criterion_group! {
name = benches;
config = Criterion::default()
.sample_size(20)
.warm_up_time(Duration::from_secs(1))
.measurement_time(Duration::from_secs(4));
targets = serialization
}
criterion_main!(benches);

48
litellm-rust/crates/python-bridge/src/lib.rs
View file

@ -19,6 +19,9 @@ use pyo3::types::{PyAny, PyDict};
use serde_json::{Map, Value};
mod gil;
mod marshal;
use marshal::{from_py, to_py};
pyo3::create_exception!(
_native,
@ -41,35 +44,18 @@ type MarshaledOcrInputs = (
Option<Duration>,
);
fn py_to_json(py: Python<'_>, value: &Bound<'_, PyAny>) -> PyResult<Value> {
let json = py.import("json")?;
let encoded: String = json.call_method1("dumps", (value,))?.extract()?;
serde_json::from_str(&encoded).map_err(|err| PyValueError::new_err(err.to_string()))
}
fn json_to_py(py: Python<'_>, value: Value) -> PyResult<Py<PyAny>> {
let json = py.import("json")?;
let encoded =
serde_json::to_string(&value).map_err(|err| PyValueError::new_err(err.to_string()))?;
Ok(json.call_method1("loads", (encoded,))?.unbind())
}
fn messages_response_to_py(
py: Python<'_>,
response: AnthropicMessagesResponse,
) -> PyResult<Py<PyAny>> {
let value =
serde_json::to_value(response).map_err(|err| PyValueError::new_err(err.to_string()))?;
json_to_py(py, value)
to_py(py, &response)
}
fn chat_completions_response_to_py(
py: Python<'_>,
response: ChatCompletionsResponse,
) -> PyResult<Py<PyAny>> {
let value =
serde_json::to_value(response).map_err(|err| PyValueError::new_err(err.to_string()))?;
json_to_py(py, value)
to_py(py, &response)
}
fn core_error_to_pyerr(err: CoreError) -> PyErr {
@ -116,7 +102,7 @@ fn optional_object_to_map(
value: Option<Py<PyAny>>,
) -> PyResult<Map<String, Value>> {
match value {
Some(value) => match py_to_json(py, value.bind(py))? {
Some(value) => match from_py(value.bind(py))? {
Value::Object(map) => Ok(map),
_ => Err(PyValueError::new_err(format!("{name} must be a dict"))),
},
@ -139,7 +125,7 @@ fn marshal_headers(
headers: Option<Py<PyAny>>,
) -> PyResult<HashMap<String, String>> {
let value = match headers {
Some(headers) => py_to_json(py, headers.bind(py))?,
Some(headers) => from_py(headers.bind(py))?,
None => Value::Object(Map::new()),
};
let Value::Object(headers) = value else {
@ -211,7 +197,7 @@ fn marshal_inputs(
optional_params: Option<Py<PyAny>>,
timeout_seconds: Option<f64>,
) -> PyResult<MarshaledOcrInputs> {
let document = py_to_json(py, document.bind(py))?;
let document = from_py(document.bind(py))?;
let extra_headers = match extra_headers {
Some(headers) => Some(optional_object_to_map(py, "extra_headers", Some(headers))?),
None => None,
@ -262,7 +248,7 @@ fn ocr(
});
match result {
Ok(value) => json_to_py(py, value),
Ok(value) => to_py(py, &value),
Err(err) => Err(core_error_to_pyerr(err)),
}
}
@ -307,7 +293,7 @@ fn aocr(
.await
.map_err(core_error_to_pyerr)?;
Python::attach(|py| json_to_py(py, value))
Python::attach(|py| to_py(py, &value))
})
}
@ -325,7 +311,7 @@ fn transcription(
optional_params: Option<Py<PyAny>>,
timeout_seconds: Option<f64>,
) -> PyResult<Py<PyAny>> {
let audio = py_to_json(py, audio.bind(py))?;
let audio = from_py(audio.bind(py))?;
let extra_headers = match extra_headers {
Some(headers) => Some(optional_object_to_map(py, "extra_headers", Some(headers))?),
None => None,
@ -351,7 +337,7 @@ fn transcription(
))
});
match result {
Ok(value) => json_to_py(py, value),
Ok(value) => to_py(py, &value),
Err(err) => Err(core_error_to_pyerr(err)),
}
}
@ -370,7 +356,7 @@ fn atranscription(
optional_params: Option<Py<PyAny>>,
timeout_seconds: Option<f64>,
) -> PyResult<Bound<'_, PyAny>> {
let audio = py_to_json(py, audio.bind(py))?;
let audio = from_py(audio.bind(py))?;
let extra_headers = match extra_headers {
Some(headers) => Some(optional_object_to_map(py, "extra_headers", Some(headers))?),
None => None,
@ -394,7 +380,7 @@ fn atranscription(
})
.await
.map_err(core_error_to_pyerr)?;
Python::attach(|py| json_to_py(py, value))
Python::attach(|py| to_py(py, &value))
})
}
@ -406,7 +392,7 @@ fn marshal_messages_inputs(
extra_headers: Option<Py<PyAny>>,
timeout_seconds: Option<f64>,
) -> PyResult<MarshaledMessagesInputs> {
let body = py_to_json(py, body.bind(py))?;
let body: Value = from_py(body.bind(py))?;
if !body.is_object() {
return Err(PyValueError::new_err("body must be a dict"));
}
@ -498,7 +484,7 @@ fn marshal_chat_completions_inputs(
extra_headers: Option<Py<PyAny>>,
timeout_seconds: Option<f64>,
) -> PyResult<MarshaledChatCompletionsInputs> {
let messages = py_to_json(py, messages.bind(py))?;
let messages: Value = from_py(messages.bind(py))?;
if !messages.is_array() {
return Err(PyValueError::new_err("messages must be a list"));
}
@ -527,7 +513,7 @@ fn chat_completions_decline(
optional_params: Option<Py<PyAny>>,
custom_llm_provider: Option<String>,
) -> PyResult<Option<String>> {
let messages = py_to_json(py, messages.bind(py))?;
let messages = from_py(messages.bind(py))?;
let optional_params = optional_object_to_map(py, "optional_params", optional_params)?;
Ok(chat_completions_decline_reason(
&model,

20
litellm-rust/crates/python-bridge/src/marshal.rs Normal file
View file

@ -0,0 +1,20 @@
use pyo3::exceptions::PyValueError;
use pyo3::prelude::*;
use serde::Serialize;
use serde::de::DeserializeOwned;
pub fn from_py<T>(value: &Bound<'_, PyAny>) -> PyResult<T>
where
T: DeserializeOwned,
{
pythonize::depythonize(value).map_err(|error| PyValueError::new_err(error.to_string()))
}
pub fn to_py<T>(py: Python<'_>, value: &T) -> PyResult<Py<PyAny>>
where
T: Serialize + ?Sized,
{
pythonize::pythonize(py, value)
.map(Bound::unbind)
.map_err(|error| PyValueError::new_err(error.to_string()))
}

52
litellm-rust/crates/python-bridge/tests/marshal_boundary.rs Normal file
View file

@ -0,0 +1,52 @@
use std::fs;
use std::path::{Path, PathBuf};
const DISALLOWED_OUTSIDE_MARSHAL: &[&str] = &[
"py.import(\"json\")",
"pythonize::",
"serde_json::to_string",
"serde_json::from_str",
];
fn source_root() -> PathBuf {
Path::new(env!("CARGO_MANIFEST_DIR")).join("src")
}
fn rust_sources(directory: &Path) -> Vec<PathBuf> {
fs::read_dir(directory)
.expect("bridge source directory should be readable")
.map(|entry| {
entry
.expect("bridge source entry should be readable")
.path()
})
.flat_map(|path| {
if path.is_dir() {
rust_sources(&path)
} else if path.extension().is_some_and(|extension| extension == "rs") {
vec![path]
} else {
Vec::new()
}
})
.collect()
}
#[test]
fn serialization_is_centralized_in_marshal_module() {
let root = source_root();
for path in rust_sources(&root) {
if path == root.join("marshal.rs") {
continue;
}
let source = fs::read_to_string(&path).expect("bridge source should be readable");
for disallowed in DISALLOWED_OUTSIDE_MARSHAL {
assert!(
!source.contains(disallowed),
"{} bypasses the typed marshal module with `{disallowed}`",
path.display()
);
}
}
}

46
litellm/__init__.py
View file

@ -7,6 +7,9 @@ warnings.filterwarnings("ignore", message=".*conflict with protected namespace.*
# Suppress Pydantic 2.11+ deprecation warning about accessing model_fields on instances
# This warning can accumulate during streaming and cause memory leaks
warnings.filterwarnings("ignore", message=".*Accessing the.*attribute on the instance is deprecated.*")
# ReadOnly on TypedDict fields is repo-wide static discipline (LIT012); pydantic warns it
# cannot enforce it at runtime, which floods proxy boot once such a type is schema-walked
warnings.filterwarnings("ignore", message=".*`ReadOnly` qualifier.*")
### INIT VARIABLES #########################
import threading
import os
@ -199,6 +202,7 @@ standard_logging_payload_excluded_fields: Optional[List[str]] = (
None # Fields to exclude from StandardLoggingPayload before callbacks receive it
)
log_raw_request_response: bool = False
log_client_error_tracebacks: bool = False
request_correlation_in_logs: bool = False
redact_messages_in_exceptions: Optional[bool] = False
redact_user_api_key_info: Optional[bool] = False
@ -273,7 +277,6 @@ databricks_key: Optional[str] = None
openai_like_key: Optional[str] = None
azure_key: Optional[str] = None
anthropic_key: Optional[str] = None
autorouter_savings_baseline_model: Optional[str] = None
replicate_key: Optional[str] = None
bytez_key: Optional[str] = None
gdc_key: Optional[str] = None
@ -444,6 +447,7 @@ max_ui_session_budget: Optional[float] = (
1.0 # USD budget for each dashboard login session (playground, test connection)
)
internal_user_budget_duration: Optional[str] = None
budget_rollover: bool = False # carry spend beyond max_budget into the next window instead of zeroing it
tag_budget_config: Optional[Dict[str, "BudgetConfig"]] = None
max_end_user_budget: Optional[float] = None
max_end_user_budget_id: Optional[str] = None
@ -463,6 +467,11 @@ prometheus_metrics_config: Optional[List] = None
prometheus_exclude_metrics: Optional[List[str]] = None
prometheus_exclude_labels: Optional[List[str]] = None
prometheus_emit_stream_label: bool = False
prometheus_deployment_and_latency_caller_identity: Literal[
"api_key_alias",
"user_email",
"both",
] = "api_key_alias"
# Opt-in: emit `rate_limit_category` and `rate_limit_type` labels on
# `litellm_proxy_failed_requests_metric`. Off by default to preserve the
# pre-unification label set so existing dashboards / recording rules keyed on
@ -480,6 +489,7 @@ public_mcp_servers: Optional[List[str]] = None
public_mcp_hub_strict_whitelist: bool = True
public_model_groups: Optional[List[str]] = None
public_agent_groups: Optional[List[str]] = None
agent_search_embedding_model: Optional[str] = None
# Supports both old format (Dict[str, str]) and new format (Dict[str, Dict[str, Any]])
# New format: { "displayName": { "url": "...", "index": 0 } }
# Old format: { "displayName": "url" } (for backward compatibility)
@ -649,6 +659,8 @@ aiml_models: Set = set()
deepgram_models: Set = set()
elevenlabs_models: Set = set()
dashscope_models: Set = set()
qwencloud_models: Set = set()
qwen_ai_platform_models: Set = set()
moonshot_models: Set = set()
publicai_models: Set = set()
darkbloom_models: Set = set()
@ -899,6 +911,10 @@ def _populate_provider_model_sets(model_cost_map: Dict) -> None:
heroku_models.add(key)
elif value.get("litellm_provider") == "dashscope":
dashscope_models.add(key)
elif value.get("litellm_provider") == "qwencloud":
qwencloud_models.add(key)
elif value.get("litellm_provider") == "qwen_ai_platform":
qwen_ai_platform_models.add(key)
elif value.get("litellm_provider") == "modelscope":
modelscope_models.add(key)
elif value.get("litellm_provider") == "moonshot":
@ -1062,6 +1078,8 @@ model_list = list(
| deepgram_models
| elevenlabs_models
| dashscope_models
| qwencloud_models
| qwen_ai_platform_models
| moonshot_models
| publicai_models
| darkbloom_models
@ -1168,6 +1186,8 @@ def _build_models_by_provider() -> dict:
"elevenlabs": elevenlabs_models,
"heroku": heroku_models,
"dashscope": dashscope_models,
"qwencloud": qwencloud_models,
"qwen_ai_platform": qwen_ai_platform_models,
"modelscope": modelscope_models,
"moonshot": moonshot_models,
"publicai": publicai_models,
@ -1628,6 +1648,9 @@ if TYPE_CHECKING:
AmazonMantleMessagesConfig as AmazonMantleMessagesConfig,
)
from .llms.together_ai.chat import TogetherAIConfig as TogetherAIConfig
from .llms.together_ai.chat.transformation import (
TogetherAIChatConfig as TogetherAIChatConfig,
)
from .llms.nlp_cloud.chat.handler import NLPCloudConfig as NLPCloudConfig
from .llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
VertexGeminiConfig as VertexGeminiConfig,
@ -1801,6 +1824,9 @@ if TYPE_CHECKING:
from .llms.gemini.interactions.transformation import (
GoogleAIStudioInteractionsConfig as GoogleAIStudioInteractionsConfig,
)
from .llms.vertex_ai.interactions.transformation import (
VertexAIInteractionsConfig as VertexAIInteractionsConfig,
)
from .llms.openai.chat.o_series_transformation import (
OpenAIOSeriesConfig as OpenAIOSeriesConfig,
OpenAIOSeriesConfig as OpenAIO1Config,
@ -1998,6 +2024,24 @@ if TYPE_CHECKING:
from .llms.dashscope.rerank.transformation import (
DashScopeRerankConfig as DashScopeRerankConfig,
)
from .llms.dashscope.qwencloud import (
QwenCloudChatConfig as QwenCloudChatConfig,
)
from .llms.dashscope.qwencloud import (
QwenCloudEmbeddingConfig as QwenCloudEmbeddingConfig,
)
from .llms.dashscope.qwencloud import (
QwenCloudRerankConfig as QwenCloudRerankConfig,
)
from .llms.dashscope.qwen_ai_platform import (
QwenAIPlatformChatConfig as QwenAIPlatformChatConfig,
)
from .llms.dashscope.qwen_ai_platform import (
QwenAIPlatformEmbeddingConfig as QwenAIPlatformEmbeddingConfig,
)
from .llms.dashscope.qwen_ai_platform import (
QwenAIPlatformRerankConfig as QwenAIPlatformRerankConfig,
)
from .llms.modelscope.chat.transformation import (
ModelScopeChatConfig as ModelScopeChatConfig,
)

92
litellm/_lazy_imports.py
View file

@ -17,8 +17,11 @@ until they're actually needed.
import importlib
import sys
from collections.abc import Callable
from typing import Any, Final, cast
from collections.abc import Callable, Mapping
from types import ModuleType
from typing import TYPE_CHECKING, Any, Final, cast
from typing_extensions import ReadOnly, TypedDict
# Import all the data structures that define what can be lazy-loaded
# These are just lists of names and maps of where to find them
@ -53,8 +56,12 @@ from ._lazy_imports_registry import (
UTILS_NAMES,
)
if TYPE_CHECKING:
import httpx
from tiktoken import Encoding
def get_litellm_globals() -> dict:
def get_litellm_globals() -> dict[str, object]:
"""
Get the globals dictionary of the litellm module.
@ -64,7 +71,7 @@ def get_litellm_globals() -> dict:
return sys.modules["litellm"].__dict__
def _get_utils_globals() -> dict:
def _get_utils_globals() -> dict[str, object]:
"""
Get the globals dictionary of the utils module.
@ -74,14 +81,19 @@ def _get_utils_globals() -> dict:
return sys.modules["litellm.utils"].__dict__
def _get_module_level_client_timeout(litellm_globals: Mapping[str, Any]) -> "float | httpx.Timeout | None":
"""Read the configured `litellm.request_timeout` used for the module level http clients."""
return litellm_globals.get("request_timeout")
# These are special lazy loaders for things that are used internally
# They're separate from the main lazy import system because they have specific use cases
# Lazy loader for default encoding - avoids importing heavy tiktoken library at startup
_default_encoding: Any | None = None
_default_encoding: "Encoding | None" = None
def _get_default_encoding() -> Any:
def _get_default_encoding() -> "Encoding":
"""
Lazily load and cache the default OpenAI encoding.
@ -100,10 +112,10 @@ def _get_default_encoding() -> Any:
# Lazy loader for get_modified_max_tokens to avoid importing token_counter at module import time
_get_modified_max_tokens_func: Any | None = None
_get_modified_max_tokens_func: "Callable[..., int | None] | None" = None
def _get_modified_max_tokens() -> Any:
def _get_modified_max_tokens() -> "Callable[..., int | None]":
"""
Lazily load and cache the get_modified_max_tokens function.
@ -124,10 +136,10 @@ def _get_modified_max_tokens() -> Any:
# Lazy loader for token_counter to avoid importing token_counter module at module import time
_token_counter_new_func: Any | None = None
_token_counter_new_func: "Callable[..., int] | None" = None
def _get_token_counter_new() -> Any:
def _get_token_counter_new() -> "Callable[..., int]":
"""
Lazily load and cache the token_counter function (aliased as token_counter_new).
@ -154,10 +166,10 @@ def _get_token_counter_new() -> Any:
# This registry maps attribute names (like "ModelResponse") to handler functions
# It's built once the first time someone accesses a lazy-loaded attribute
# Example: {"ModelResponse": _lazy_import_utils, "Cache": _lazy_import_caching, ...}
_LAZY_IMPORT_REGISTRY: dict[str, Callable[[str], Any]] | None = None
_LAZY_IMPORT_REGISTRY: dict[str, Callable[[str], object]] | None = None
def _get_lazy_import_registry() -> dict[str, Callable[[str], Any]]:
def _get_lazy_import_registry() -> dict[str, Callable[[str], object]]:
"""
Build the registry that maps attribute names to their handler functions.
@ -206,7 +218,18 @@ def _get_lazy_import_registry() -> dict[str, Callable[[str], Any]]:
return _LAZY_IMPORT_REGISTRY
def _generic_lazy_import(name: str, import_map: dict[str, tuple[str, str]], category: str) -> Any:
class _AttributeView(TypedDict):
"""Holds one module attribute so the lazily fetched value is read back as ``object``."""
value: ReadOnly[object]
def _module_attribute(module: ModuleType, attr_name: str) -> object:
attribute: Final[_AttributeView] = {"value": getattr(module, attr_name)}
return attribute["value"]
def _generic_lazy_import(name: str, import_map: dict[str, tuple[str, str]], category: str) -> object:
"""
Generic function that handles lazy importing for most attributes.
@ -255,7 +278,7 @@ def _generic_lazy_import(name: str, import_map: dict[str, tuple[str, str]], cate
# Step 6: Get the actual attribute from the module
# Example: getattr(utils_module, "ModelResponse") returns the ModelResponse class
value: Final = getattr(module, attr_name)
value: Final = _module_attribute(module, attr_name)
# Step 7: Cache it so we don't have to import again next time
_globals[name] = value
@ -272,62 +295,62 @@ def _generic_lazy_import(name: str, import_map: dict[str, tuple[str, str]], cate
# The registry (above) maps attribute names to these handler functions.
def _lazy_import_utils(name: str) -> Any:
def _lazy_import_utils(name: str) -> object:
"""Handler for utils module attributes (ModelResponse, token_counter, etc.)"""
return _generic_lazy_import(name, _UTILS_IMPORT_MAP, "Utils")
def _lazy_import_cost_calculator(name: str) -> Any:
def _lazy_import_cost_calculator(name: str) -> object:
"""Handler for cost calculator functions (completion_cost, cost_per_token, etc.)"""
return _generic_lazy_import(name, _COST_CALCULATOR_IMPORT_MAP, "Cost calculator")
def _lazy_import_token_counter(name: str) -> Any:
def _lazy_import_token_counter(name: str) -> object:
"""Handler for token counter utilities"""
return _generic_lazy_import(name, _TOKEN_COUNTER_IMPORT_MAP, "Token counter")
def _lazy_import_bedrock_types(name: str) -> Any:
def _lazy_import_bedrock_types(name: str) -> object:
"""Handler for Bedrock type aliases"""
return _generic_lazy_import(name, _BEDROCK_TYPES_IMPORT_MAP, "Bedrock types")
def _lazy_import_types_utils(name: str) -> Any:
def _lazy_import_types_utils(name: str) -> object:
"""Handler for types from litellm.types.utils (BudgetConfig, ImageObject, etc.)"""
return _generic_lazy_import(name, _TYPES_UTILS_IMPORT_MAP, "Types utils")
def _lazy_import_caching(name: str) -> Any:
def _lazy_import_caching(name: str) -> object:
"""Handler for caching classes (Cache, DualCache, RedisCache, etc.)"""
return _generic_lazy_import(name, _CACHING_IMPORT_MAP, "Caching")
def _lazy_import_dotprompt(name: str) -> Any:
def _lazy_import_dotprompt(name: str) -> object:
"""Handler for dotprompt integration globals"""
return _generic_lazy_import(name, _DOTPROMPT_IMPORT_MAP, "Dotprompt")
def _lazy_import_types(name: str) -> Any:
def _lazy_import_types(name: str) -> object:
"""Handler for type classes (GuardrailItem, etc.)"""
return _generic_lazy_import(name, _TYPES_IMPORT_MAP, "Types")
def _lazy_import_llm_configs(name: str) -> Any:
def _lazy_import_llm_configs(name: str) -> object:
"""Handler for LLM config classes (AnthropicConfig, OpenAILikeChatConfig, etc.)"""
return _generic_lazy_import(name, _LLM_CONFIGS_IMPORT_MAP, "LLM config")
def _lazy_import_litellm_logging(name: str) -> Any:
def _lazy_import_litellm_logging(name: str) -> object:
"""Handler for litellm_logging module (Logging, modify_integration)"""
return _generic_lazy_import(name, _LITELLM_LOGGING_IMPORT_MAP, "Litellm logging")
def _lazy_import_llm_provider_logic(name: str) -> Any:
def _lazy_import_llm_provider_logic(name: str) -> object:
"""Handler for LLM provider logic functions (get_llm_provider, etc.)"""
return _generic_lazy_import(name, _LLM_PROVIDER_LOGIC_IMPORT_MAP, "LLM provider logic")
def _lazy_import_utils_module(name: str) -> Any:
def _lazy_import_utils_module(name: str) -> object:
"""
Handler for utils module lazy imports.
@ -355,7 +378,7 @@ def _lazy_import_utils_module(name: str) -> Any:
module = importlib.import_module(module_path)
# Get the actual attribute from the module
value: Final = getattr(module, attr_name)
value: Final = _module_attribute(module, attr_name)
# Cache it so we don't have to import again next time
_globals[name] = value
@ -370,7 +393,7 @@ def _lazy_import_utils_module(name: str) -> Any:
# These handlers have custom logic that doesn't fit the generic pattern
def _lazy_import_llm_client_cache(name: str) -> Any:
def _lazy_import_llm_client_cache(name: str) -> object:
"""
Handler for LLM client cache - has special logic for singleton instance.
@ -386,8 +409,7 @@ def _lazy_import_llm_client_cache(name: str) -> Any:
return _globals[name]
# Import the class
module: Final = importlib.import_module("litellm.caching.llm_caching_handler")
LLMClientCache: Final = getattr(module, "LLMClientCache")
from litellm.caching.llm_caching_handler import LLMClientCache
# If they want the class itself, return it
if name == "LLMClientCache":
@ -403,7 +425,7 @@ def _lazy_import_llm_client_cache(name: str) -> Any:
raise AttributeError(f"LLM client cache lazy import: unknown attribute {name!r}")
def _lazy_import_http_handlers(name: str) -> Any:
def _lazy_import_http_handlers(name: str) -> object:
"""
Handler for HTTP clients - has special logic for creating client instances.
@ -419,8 +441,8 @@ def _lazy_import_http_handlers(name: str) -> Any:
from litellm.llms.custom_httpx.http_handler import get_async_httpx_client
# Get timeout from module config (if set)
timeout = _globals.get("request_timeout")
params: Final = {"timeout": timeout, "client_alias": "module level aclient"}
async_timeout: Final = _get_module_level_client_timeout(_globals)
params: Final = {"timeout": async_timeout, "client_alias": "module level aclient"}
# Create the client instance
provider_id: Final = cast(Any, "litellm_module_level_client")
@ -437,8 +459,8 @@ def _lazy_import_http_handlers(name: str) -> Any:
# Create a sync HTTP client
from litellm.llms.custom_httpx.http_handler import HTTPHandler
timeout = _globals.get("request_timeout")
sync_client: Final = HTTPHandler(timeout=timeout)
sync_timeout: Final = _get_module_level_client_timeout(_globals)
sync_client: Final = HTTPHandler(timeout=sync_timeout)
# Cache it
_globals["module_level_client"] = sync_client

20
litellm/_lazy_imports_registry.py
View file

@ -177,6 +177,7 @@ LLM_CONFIG_NAMES: Final = (
"AmazonAnthropicClaudeMessagesConfig",
"AmazonMantleMessagesConfig",
"TogetherAIConfig",
"TogetherAIChatConfig",
"NLPCloudConfig",
"VertexGeminiConfig",
"GoogleAIStudioGeminiConfig",
@ -242,6 +243,7 @@ LLM_CONFIG_NAMES: Final = (
"OpenRouterResponsesAPIConfig",
"BedrockMantleResponsesAPIConfig",
"GoogleAIStudioInteractionsConfig",
"VertexAIInteractionsConfig",
"OpenAIOSeriesConfig",
"AnthropicSkillsConfig",
"BaseSkillsAPIConfig",
@ -308,6 +310,8 @@ LLM_CONFIG_NAMES: Final = (
"GigaChatConfig",
"GigaChatEmbeddingConfig",
"DashScopeChatConfig",
"QwenCloudChatConfig",
"QwenAIPlatformChatConfig",
"ModelScopeChatConfig",
"MoonshotChatConfig",
"DockerModelRunnerChatConfig",
@ -740,6 +744,10 @@ _LLM_CONFIGS_IMPORT_MAP: Final = {
"AmazonMantleMessagesConfig",
),
"TogetherAIConfig": (".llms.together_ai.chat", "TogetherAIConfig"),
"TogetherAIChatConfig": (
".llms.together_ai.chat.transformation",
"TogetherAIChatConfig",
),
"NLPCloudConfig": (".llms.nlp_cloud.chat.handler", "NLPCloudConfig"),
"VertexGeminiConfig": (
".llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini",
@ -977,6 +985,10 @@ _LLM_CONFIGS_IMPORT_MAP: Final = {
".llms.gemini.interactions.transformation",
"GoogleAIStudioInteractionsConfig",
),
"VertexAIInteractionsConfig": (
".llms.vertex_ai.interactions.transformation",
"VertexAIInteractionsConfig",
),
"OpenAIOSeriesConfig": (
".llms.openai.chat.o_series_transformation",
"OpenAIOSeriesConfig",
@ -1162,6 +1174,14 @@ _LLM_CONFIGS_IMPORT_MAP: Final = {
".llms.dashscope.chat.transformation",
"DashScopeChatConfig",
),
"QwenCloudChatConfig": (
".llms.dashscope.qwencloud",
"QwenCloudChatConfig",
),
"QwenAIPlatformChatConfig": (
".llms.dashscope.qwen_ai_platform",
"QwenAIPlatformChatConfig",
),
"GDCGeminiConfig": (
".llms.gdc.chat.transformation",
"GDCGeminiConfig",

99
litellm/_logging.py
View file

@ -5,7 +5,7 @@ import os
import sys
from datetime import datetime
from logging import Formatter
from typing import Any, Final
from typing import Any, Final, TextIO
import litellm
from litellm.constants import (
@ -88,6 +88,24 @@ def redact_secrets(value: str) -> str:
return _redact_string(value)
def _substituted_color_message(record: logging.LogRecord) -> str | None:
"""Render a record's ``color_message`` against its args, or None if absent.
uvicorn's colorized formatter re-renders `color_message` against
record.args at emit time (see uvicorn.logging.ColourizedFormatter) instead
of using the already-formatted record.msg, so it has to be substituted
before args are cleared or it is later formatted with no args and prints
the raw "%s://%s:%d" placeholders instead of the URL.
"""
color_message: Final = record.__dict__.get("color_message")
if not isinstance(color_message, str) or not record.args:
return None
try:
return color_message % record.args
except TypeError:
return color_message
class SecretRedactionFilter(logging.Filter):
"""Scrubs known secret/credential patterns from log records."""
@ -97,6 +115,12 @@ class SecretRedactionFilter(logging.Filter):
if not _ENABLE_SECRET_REDACTION:
return True
# Runs before args are cleared, and before the extra-field loop below
# that redacts the substituted result.
substituted_color_message: Final = _substituted_color_message(record)
if substituted_color_message is not None:
record.color_message = substituted_color_message # rebind-ok: a Filter scrubs records in place
try:
record.msg = _redact_string(record.getMessage())
record.args = None
@ -210,11 +234,69 @@ class CorrelationContextFilter(logging.Filter):
_correlation_filter: Final = CorrelationContextFilter()
json_logs = bool(os.getenv("JSON_LOGS", False))
_LOG_FORMAT_PREFIX: Final = "%(asctime)s - %(name)s:%(levelname)s"
_LOG_FORMAT_SUFFIX: Final = ": %(filename)s:%(lineno)s - %(message)s"
_PLAIN_LOG_FORMAT: Final = _LOG_FORMAT_PREFIX + _LOG_FORMAT_SUFFIX
_COLOR_LOG_FORMAT: Final = f"\033[92m{_LOG_FORMAT_PREFIX}\033[0m{_LOG_FORMAT_SUFFIX}"
def _stream_is_tty(stream: TextIO | None) -> bool:
"""True when the stream is an open interactive terminal; never raises.
A stream can be None (pythonw/embedded interpreters), lack isatty entirely
(GUI log-redirect shims), or be closed; import must survive all three.
"""
try:
return stream is not None and stream.isatty()
except (AttributeError, ValueError):
return False
def _plain_log_format(stdout: TextIO | None, stderr: TextIO | None) -> str:
"""The plain-text log format, colorized only when both streams are an interactive terminal.
Honors the NO_COLOR convention from no-color.org: color is disabled when
NO_COLOR is present with a non-empty value.
"""
if os.environ.get("NO_COLOR"):
return _PLAIN_LOG_FORMAT
return _COLOR_LOG_FORMAT if _stream_is_tty(stdout) and _stream_is_tty(stderr) else _PLAIN_LOG_FORMAT
class LevelRoutingStreamHandler(logging.StreamHandler):
"""Writes records below WARNING and invalid-key warnings to stdout, others to stderr.
Collectors that derive severity from the stream report every stderr line as an error.
Invalid-key warnings route to stdout so LITELLM_LOG=ERROR can suppress them.
"""
def emit(self, record: logging.LogRecord) -> None:
is_stdout_record: Final = record.levelno < logging.WARNING or (
record.levelno == logging.WARNING and record.name == verbose_proxy_stdout_logger.name
)
preferred: Final = sys.stdout if is_stdout_record else sys.stderr
if preferred is None or getattr(preferred, "closed", False):
self.stream = sys.stderr # rebind-ok: fall back to the pre-fix stream rather than raising per record
else:
self.stream = preferred # rebind-ok: StreamHandler.emit writes self.stream under the handler lock
super().emit(record)
def _parse_json_logs_env(value: str | None) -> bool:
"""Strict opt-in parse for the JSON_LOGS env var: only "true" (any case) enables JSON logs.
Matches the reader in litellm-proxy-extras/_logging.py. The previous
bool(os.getenv(...)) treated any non-empty value, including "false" and "0",
as enabled.
"""
return (value or "").lower() == "true"
json_logs: Final = _parse_json_logs_env(os.getenv("JSON_LOGS"))
# Create a handler for the logger (you may need to adapt this based on your needs)
log_level: Final = os.getenv("LITELLM_LOG", "DEBUG")
numeric_level: Final[str] = getattr(logging, log_level.upper())
handler: Final = logging.StreamHandler()
handler: Final = LevelRoutingStreamHandler()
handler.setLevel(numeric_level)
handler.addFilter(_secret_filter)
handler.addFilter(_correlation_filter)
@ -423,13 +505,16 @@ if json_logs:
_setup_json_exception_handlers(JsonFormatter())
else:
formatter: Final = CorrelationPlainFormatter(
"\033[92m%(asctime)s - %(name)s:%(levelname)s\033[0m: %(filename)s:%(lineno)s - %(message)s",
_plain_log_format(sys.stdout, sys.stderr),
datefmt="%H:%M:%S",
)
handler.setFormatter(formatter)
verbose_proxy_logger = logging.getLogger("LiteLLM Proxy")
# Malformed virtual key rejections log through this child; LevelRoutingStreamHandler
# writes its WARNING records to stdout. It has no handler or level of its own.
verbose_proxy_stdout_logger: Final = verbose_proxy_logger.getChild("stdout")
verbose_router_logger = logging.getLogger("LiteLLM Router")
verbose_logger = logging.getLogger("LiteLLM")
@ -442,6 +527,7 @@ verbose_logger.addHandler(handler)
# handlers (JSON mode, uvicorn log config, a host app's root handler).
verbose_router_logger.addFilter(_stdout_truncation_filter)
verbose_proxy_logger.addFilter(_stdout_truncation_filter)
verbose_proxy_stdout_logger.addFilter(_stdout_truncation_filter)
verbose_logger.addFilter(_stdout_truncation_filter)
@ -604,7 +690,8 @@ def _turn_on_json():
- Adds a JSON formatter to all loggers
"""
handler: Final = logging.StreamHandler()
handler: Final = LevelRoutingStreamHandler()
handler.setLevel(numeric_level)
handler.setFormatter(JsonFormatter())
_initialize_loggers_with_handler(handler)
# Set up exception handlers
@ -622,12 +709,14 @@ def _disable_debugging():
verbose_logger.disabled = True
verbose_router_logger.disabled = True
verbose_proxy_logger.disabled = True
verbose_proxy_stdout_logger.disabled = True
def _enable_debugging():
verbose_logger.disabled = False
verbose_router_logger.disabled = False
verbose_proxy_logger.disabled = False
verbose_proxy_stdout_logger.disabled = False
def print_verbose(print_statement):

265
litellm/_redis.py
View file

@ -12,8 +12,10 @@ import json
# s/o [@Frank Colson](https://www.linkedin.com/in/frank-colson-422b9b183/) for this redis implementation
import os
from collections.abc import Callable
from collections.abc import Callable, Mapping
from types import MappingProxyType
from typing import Final
from urllib.parse import urlsplit, urlunsplit
import redis
import redis.asyncio as async_redis
@ -37,9 +39,25 @@ from ._logging import verbose_logger
AZURE_REDIS_SCOPE: Final = "https://redis.azure.com/.default"
def _get_redis_kwargs():
arg_spec: Final = inspect.getfullargspec(redis.Redis)
def _unwrapped_init_args(cls: type) -> frozenset[str]:
"""Every parameter on a single class's own ``__init__``, decorator-unwrapped.
Unlike ``_init_arg_names`` below, this does not walk the MRO: ``redis.Redis``
and ``redis.RedisCluster`` (sync and async) each declare every real
constructor parameter directly on their own ``__init__``, so MRO-walking is
unnecessary — and it actively breaks the several tests here that mock the
class with ``patch(..., autospec=True)``, since ``inspect.getmro`` needs a
real ``__mro__`` that an autospec'd stand-in for a class does not provide.
Still unwraps first: redis-py >= 7.4 decorates these ``__init__``s with
``@deprecated_args`` too, which the same class of bug as ``_init_arg_names``
would otherwise silently empty this allowlist through (see its docstring).
"""
spec: Final = inspect.getfullargspec(inspect.unwrap(cls.__init__))
return frozenset(spec.args + spec.kwonlyargs)
def _get_redis_kwargs():
# Only allow primitive arguments
exclude_args: Final = {
"self",
@ -50,6 +68,7 @@ def _get_redis_kwargs():
include_args: Final = {
"url",
"redis_connect_func",
"credential_provider",
"gcp_service_account",
"gcp_ssl_ca_certs",
"azure_redis_ad_token",
@ -58,7 +77,7 @@ def _get_redis_kwargs():
"azure_client_secret",
}
available_args: Final = {x for x in arg_spec.args if x not in exclude_args} | include_args
available_args: Final = {x for x in _unwrapped_init_args(redis.Redis) if x not in exclude_args} | include_args
return available_args
@ -118,15 +137,23 @@ def _get_redis_url_kwargs(client: type | None = None) -> tuple[str, ...]:
return tuple(x for x in _init_arg_names(connection_cls) if x not in exclude_args) + include_args
def _get_redis_cluster_kwargs(client=None):
def _get_redis_cluster_kwargs(client: type | None = None):
"""Config kwargs the target cluster client's constructor actually accepts.
Defaults to the sync ``redis.RedisCluster``, but the async cluster client
(``redis.asyncio.cluster.RedisCluster``) declares connection settings such as
``decode_responses`` on its own constructor, where the sync class takes them
through ``**kwargs`` and so never names them in its signature. Introspecting
only the sync class regardless of which client is actually built silently
drops those for every async cluster caller.
"""
if client is None:
client = redis.Redis.from_url
arg_spec: Final = inspect.getfullargspec(redis.RedisCluster)
client = redis.RedisCluster
# Only allow primitive arguments
exclude_args: Final = {"self", "connection_pool", "retry", "host", "port", "startup_nodes"}
available_args = {x for x in arg_spec.args if x not in exclude_args}
available_args = {x for x in _unwrapped_init_args(client) if x not in exclude_args}
available_args |= {
"password",
"username",
@ -155,7 +182,81 @@ def _get_redis_cluster_kwargs(client=None):
def _get_redis_env_kwarg_mapping():
PREFIX: Final = "REDIS_"
return {f"{PREFIX}{x.upper()}": x for x in _get_redis_kwargs()}
exclude_from_environment: Final = frozenset({"credential_provider"})
return {f"{PREFIX}{x.upper()}": x for x in _get_redis_kwargs() if x not in exclude_from_environment}
def _str_to_bool(value: str) -> bool:
return value.lower() in ("true", "1", "yes")
def _coerce_redis_kwargs_types(
redis_kwargs: Mapping[str, object],
client: type | tuple[type, ...] = redis.Redis,
) -> dict[str, object]: # mutable-ok: a caller mutates the returned kwargs before constructing its client
"""Coerces string values to the numeric/boolean type ``client``'s constructor
declares for that parameter. ``client`` may be a tuple of client classes; a
parameter's type is taken from the first signature that declares it, which
lets cluster callers coerce cluster-only kwargs such as
``cluster_error_retry_attempts`` alongside the shared connection kwargs.
Environment variables are always strings, and Helm ``--set`` stringifies values
too, so a config value like ``health_check_interval`` or ``socket_timeout``
can arrive as ``"30"``/``"5.5"`` rather than a real number. redis-py's own
connection-health-check arithmetic (``loop.time() + self.health_check_interval``)
then raises ``TypeError`` on every Redis operation instead of connecting.
``max_connections``, ``socket_timeout``, and ``socket_connect_timeout`` use an
explicit target type rather than the parameter's own signature default: redis-py
8.x changed the timeout defaults from ``None`` to int ``5``, so inferring the
type from the default would make a fractional ``"5.5"`` fail ``int()`` and get
silently dropped on 8.x while working on older versions. ``socket_keepalive``
is explicit too: its signature default is ``None``, which carries no type to
infer from, and leaving it a string makes ``"false"`` truthy.
"""
signatures: Final = tuple(inspect.signature(c) for c in (client if isinstance(client, tuple) else (client,)))
explicit_param_types: Final = MappingProxyType(
{
"max_connections": int,
"socket_timeout": float,
"socket_connect_timeout": float,
"socket_keepalive": bool,
}
)
result: Final = dict(redis_kwargs) # mutable-ok: per-key try/except coercion below needs to drop individual keys
for key, value in redis_kwargs.items():
if not isinstance(value, str):
continue
param = next((sig.parameters[key] for sig in signatures if key in sig.parameters), None)
if param is None:
continue
explicit_type = explicit_param_types.get(key)
if explicit_type is bool:
result[key] = _str_to_bool(value)
continue
if explicit_type is not None:
try:
result[key] = explicit_type(value)
except (ValueError, TypeError):
del result[key]
continue
default: object = param.default # pyright: ignore[reportAny] # inspect.Parameter.default is stubbed as Any
if default is inspect.Parameter.empty:
continue
# bool must be checked before int, since bool subclasses int
if isinstance(default, bool):
result[key] = _str_to_bool(value)
elif isinstance(default, int):
try:
result[key] = int(value)
except (ValueError, TypeError):
del result[key]
elif isinstance(default, float):
try:
result[key] = float(value)
except (ValueError, TypeError):
del result[key]
return result
def _redis_kwargs_from_environment():
@ -353,6 +454,12 @@ def get_redis_url_from_environment():
return f"{redis_protocol}://{auth_part}{os.environ['REDIS_HOST']}:{os.environ['REDIS_PORT']}"
def _url_without_userinfo(url: str) -> str:
parts: Final = urlsplit(url)
netloc: Final = parts.netloc.rsplit("@", 1)[-1]
return urlunsplit((parts.scheme, netloc, parts.path, parts.query, parts.fragment))
def _get_redis_client_logic(**env_overrides):
"""
Common functionality across sync + async redis client implementations
@ -410,54 +517,58 @@ def _get_redis_client_logic(**env_overrides):
if _service_name is not None:
redis_kwargs["service_name"] = _service_name
# Handle GCP IAM authentication
_gcp_service_account: Final = redis_kwargs.get("gcp_service_account") or get_secret_str("REDIS_GCP_SERVICE_ACCOUNT")
_gcp_ssl_ca_certs: Final = redis_kwargs.get("gcp_ssl_ca_certs") or get_secret_str("REDIS_GCP_SSL_CA_CERTS")
if _gcp_service_account is not None:
verbose_logger.debug("Setting up GCP IAM authentication for Redis with service account.")
redis_kwargs["redis_connect_func"] = create_gcp_iam_redis_connect_func(
service_account=_gcp_service_account, ssl_ca_certs=_gcp_ssl_ca_certs
if redis_kwargs.get("credential_provider") is None:
# Handle GCP IAM authentication
_gcp_service_account: Final = redis_kwargs.get("gcp_service_account") or get_secret_str(
"REDIS_GCP_SERVICE_ACCOUNT"
)
# Store GCP service account in redis_connect_func for async cluster access
redis_kwargs["redis_connect_func"]._gcp_service_account = _gcp_service_account
_gcp_ssl_ca_certs: Final = redis_kwargs.get("gcp_ssl_ca_certs") or get_secret_str("REDIS_GCP_SSL_CA_CERTS")
# Remove GCP-specific kwargs that shouldn't be passed to Redis client
redis_kwargs.pop("gcp_service_account", None)
redis_kwargs.pop("gcp_ssl_ca_certs", None)
if _gcp_service_account is not None:
verbose_logger.debug("Setting up GCP IAM authentication for Redis with service account.")
redis_kwargs["redis_connect_func"] = create_gcp_iam_redis_connect_func(
service_account=_gcp_service_account, ssl_ca_certs=_gcp_ssl_ca_certs
)
# Store GCP service account in redis_connect_func for async cluster access
redis_kwargs["redis_connect_func"]._gcp_service_account = _gcp_service_account
# Only enable SSL if explicitly requested AND SSL CA certs are provided
if _gcp_ssl_ca_certs and redis_kwargs.get("ssl", False):
redis_kwargs["ssl_ca_certs"] = _gcp_ssl_ca_certs
# Only enable SSL if explicitly requested AND SSL CA certs are provided
if _gcp_ssl_ca_certs and redis_kwargs.get("ssl", False):
redis_kwargs["ssl_ca_certs"] = _gcp_ssl_ca_certs
# Handle Azure AD authentication (after GCP IAM block)
_azure_redis_ad_token: Final = redis_kwargs.get("azure_redis_ad_token") or get_secret("REDIS_AZURE_AD_TOKEN")
# Handle Azure AD authentication (after GCP IAM block)
_azure_redis_ad_token: Final = redis_kwargs.get("azure_redis_ad_token") or get_secret("REDIS_AZURE_AD_TOKEN")
_azure_ad_enabled: Final = _azure_redis_ad_token is not None and str(_azure_redis_ad_token).lower() == "true"
_azure_ad_enabled: Final = _azure_redis_ad_token is not None and str(_azure_redis_ad_token).lower() == "true"
if _azure_ad_enabled and _gcp_service_account is not None:
verbose_logger.warning(
"Both GCP IAM (gcp_service_account) and Azure AD (azure_redis_ad_token) are configured for Redis. "
"Using GCP IAM. Remove one to avoid misconfiguration."
)
if _azure_ad_enabled and _gcp_service_account is not None:
verbose_logger.warning(
"Both GCP IAM (gcp_service_account) and Azure AD (azure_redis_ad_token) are configured for Redis. "
"Using GCP IAM. Remove one to avoid misconfiguration."
)
if _azure_ad_enabled and _gcp_service_account is None:
_azure_client_id: Final = redis_kwargs.get("azure_client_id") or get_secret_str("AZURE_CLIENT_ID")
_azure_tenant_id: Final = redis_kwargs.get("azure_tenant_id") or get_secret_str("AZURE_TENANT_ID")
_azure_client_secret: Final = redis_kwargs.get("azure_client_secret") or get_secret_str("AZURE_CLIENT_SECRET")
if _azure_ad_enabled and _gcp_service_account is None:
_azure_client_id: Final = redis_kwargs.get("azure_client_id") or get_secret_str("AZURE_CLIENT_ID")
_azure_tenant_id: Final = redis_kwargs.get("azure_tenant_id") or get_secret_str("AZURE_TENANT_ID")
_azure_client_secret: Final = redis_kwargs.get("azure_client_secret") or get_secret_str(
"AZURE_CLIENT_SECRET"
)
verbose_logger.debug("Setting up Azure AD authentication for Redis.")
redis_kwargs["redis_connect_func"] = create_azure_ad_redis_connect_func(
azure_client_id=_azure_client_id,
azure_tenant_id=_azure_tenant_id,
azure_client_secret=_azure_client_secret,
)
# Marker for async paths to detect Azure AD auth. The live credential
# object is attached separately as `_azure_credential` by
# `create_azure_ad_redis_connect_func`; the raw client_id/tenant_id/secret
# are intentionally NOT exposed on the function to avoid leaking
# credentials via inspection or logging.
redis_kwargs["redis_connect_func"]._azure_redis_ad_token = True
verbose_logger.debug("Setting up Azure AD authentication for Redis.")
redis_kwargs["redis_connect_func"] = create_azure_ad_redis_connect_func(
azure_client_id=_azure_client_id,
azure_tenant_id=_azure_tenant_id,
azure_client_secret=_azure_client_secret,
)
# Marker for async paths to detect Azure AD auth. The live credential
# object is attached separately as `_azure_credential` by
# `create_azure_ad_redis_connect_func`; the raw client_id/tenant_id/secret
# are intentionally NOT exposed on the function to avoid leaking
# credentials via inspection or logging.
redis_kwargs["redis_connect_func"]._azure_redis_ad_token = True
redis_kwargs.pop("gcp_service_account", None)
redis_kwargs.pop("gcp_ssl_ca_certs", None)
# Always remove Azure-specific kwargs that shouldn't be passed to Redis client
redis_kwargs.pop("azure_redis_ad_token", None)
@ -465,6 +576,13 @@ def _get_redis_client_logic(**env_overrides):
redis_kwargs.pop("azure_tenant_id", None)
redis_kwargs.pop("azure_client_secret", None)
if redis_kwargs.get("credential_provider") is not None:
redis_kwargs.pop("redis_connect_func", None)
redis_kwargs.pop("username", None)
redis_kwargs.pop("password", None)
if redis_kwargs.get("url") is not None:
redis_kwargs["url"] = _url_without_userinfo(redis_kwargs["url"])
if "url" in redis_kwargs and redis_kwargs["url"] is not None:
# Only strip host/port/db/password when not routing to a cluster.
# When startup_nodes is also present the cluster path takes priority and
@ -485,7 +603,12 @@ def _get_redis_client_logic(**env_overrides):
raise ValueError("Either 'host' or 'url' must be specified for redis.")
# litellm.print_verbose(f"redis_kwargs: {redis_kwargs}")
return redis_kwargs
coercion_client: Final = (
(redis.Redis, redis.RedisCluster, async_redis.RedisCluster)
if redis_kwargs.get("startup_nodes")
else redis.Redis
)
return _coerce_redis_kwargs_types(redis_kwargs, client=coercion_client)
def init_redis_cluster(redis_kwargs) -> redis.RedisCluster:
@ -532,8 +655,7 @@ def _init_redis_sentinel(redis_kwargs) -> redis.Redis:
service_name: Final = redis_kwargs.get("service_name")
connection_kwargs: Final = _get_redis_sentinel_connection_kwargs(redis_kwargs)
connection_kwargs.setdefault("socket_timeout", REDIS_SOCKET_TIMEOUT)
sentinel_kwargs: Final = dict(connection_kwargs)
sentinel_kwargs["password"] = sentinel_password
sentinel_kwargs: Final = _sentinel_auth_kwargs(connection_kwargs, sentinel_password)
if not sentinel_nodes or not service_name:
raise ValueError("Both 'sentinel_nodes' and 'service_name' are required for Redis Sentinel.")
@ -605,7 +727,12 @@ def _async_credential_provider(redis_connect_func: object | None) -> CredentialP
def _async_auth_kwargs(redis_kwargs: dict) -> dict:
"""Swaps a connect func an async path cannot run for the equivalent credential provider,
which supersedes any static username or password redis-py would otherwise reject it with."""
credential_provider: Final = _async_credential_provider(redis_kwargs.get("redis_connect_func"))
explicit_provider: Final = redis_kwargs.get("credential_provider")
credential_provider: Final = (
explicit_provider
if explicit_provider is not None
else _async_credential_provider(redis_kwargs.get("redis_connect_func"))
)
if credential_provider is None:
return redis_kwargs
@ -633,7 +760,9 @@ def get_redis_client(**env_overrides):
if "sentinel_nodes" in redis_kwargs and "service_name" in redis_kwargs:
return _init_redis_sentinel(redis_kwargs)
return redis.Redis(**redis_kwargs)
return redis.Redis( # pyright: ignore[reportCallIssue] # object-valued kwargs match no overload statically
**redis_kwargs, # pyright: ignore[reportArgumentType] # allow-listed and coerced against this signature
)
def get_redis_async_client(
@ -645,7 +774,7 @@ def get_redis_async_client(
if "startup_nodes" in redis_kwargs:
from redis.cluster import ClusterNode
args = _get_redis_cluster_kwargs()
args = _get_redis_cluster_kwargs(async_redis.RedisCluster)
cluster_kwargs: Final = {}
for arg in redis_kwargs:
if arg in args:
@ -665,8 +794,16 @@ def get_redis_async_client(
cluster_kwargs.setdefault("health_check_interval", REDIS_CLUSTER_HEALTH_CHECK_INTERVAL)
cluster_kwargs.setdefault("socket_keepalive", True)
# A single node's client-side timeout must reset only that node's connections,
# not tear down the whole cluster client for every concurrent caller.
from litellm.caching.redis_cluster_node_isolation import (
get_litellm_async_redis_cluster_class,
)
async_redis_cluster_class: Final = get_litellm_async_redis_cluster_class()
# Create async RedisCluster with IAM token as password if available
cluster_client: Final = async_redis.RedisCluster(
cluster_client: Final = async_redis_cluster_class(
startup_nodes=new_startup_nodes,
**cluster_kwargs,
)
@ -730,8 +867,20 @@ def get_redis_connection_pool(
return async_redis.BlockingConnectionPool(timeout=REDIS_CONNECTION_POOL_TIMEOUT, **redis_kwargs)
def _redis_kwargs_for_logging(redis_kwargs: Mapping[str, object]) -> Mapping[str, object]:
return {
key: "<credential provider>"
if key == "credential_provider" and value is not None
else "<redis connect function>"
if key == "redis_connect_func" and value is not None
else value
for key, value in redis_kwargs.items()
}
def _pretty_print_redis_config(redis_kwargs: dict) -> None:
"""Pretty print the Redis configuration using rich with sensitive data masking"""
redis_kwargs_for_logging: Final = _redis_kwargs_for_logging(redis_kwargs)
try:
import logging
@ -749,7 +898,7 @@ def _pretty_print_redis_config(redis_kwargs: dict) -> None:
masker = SensitiveDataMasker()
# Mask sensitive data in redis_kwargs
masked_redis_kwargs = masker.mask_dict(redis_kwargs)
masked_redis_kwargs = masker.mask_dict(redis_kwargs_for_logging)
# Create main panel title
title: Final = Text("Redis Configuration", style="bold blue")
@ -812,7 +961,7 @@ def _pretty_print_redis_config(redis_kwargs: dict) -> None:
except ImportError:
# Fallback to simple logging if rich is not available
masker = SensitiveDataMasker()
masked_redis_kwargs = masker.mask_dict(redis_kwargs)
masked_redis_kwargs = masker.mask_dict(redis_kwargs_for_logging)
verbose_logger.info("Redis configuration: %s", masked_redis_kwargs)
except Exception as e:
verbose_logger.error("Error pretty printing Redis configuration: %s", e)

38
litellm/a2a_protocol/card_resolver.py
View file

@ -4,6 +4,7 @@ Custom A2A Card Resolver for LiteLLM.
Extends the A2A SDK's card resolver to support multiple well-known paths.
"""
from types import MappingProxyType
from typing import TYPE_CHECKING, Any, Final
from litellm._logging import verbose_logger
@ -48,6 +49,43 @@ def is_localhost_or_internal_url(url: str | None) -> bool:
return any(pattern in url_lower for pattern in LOCALHOST_URL_PATTERNS)
_CANONICAL_PROTOCOL_BINDINGS: Final = MappingProxyType(
{
"jsonrpc": "JSONRPC",
"http+json": "HTTP+JSON",
"grpc": "GRPC",
}
)
_LEGACY_PROTOCOL_VERSION: Final = "0.3"
def normalize_agent_card_interfaces(agent_card: "AgentCard") -> "AgentCard":
"""
Canonicalize the supported interfaces of spec-adjacent agent cards.
Some A2A servers (e.g. LangGraph Platform) serve agent cards with lowercase
bindings like "jsonrpc", but a2a-sdk's ClientFactory matches bindings
case-sensitively against its uppercase TransportProtocol constants and fails
with "no compatible transports found." for spec-adjacent casings.
The same servers also speak the A2A 0.3 JSON dialect ("kind"-discriminated
payloads) while declaring protocolVersion "1.0", which a2a-sdk's strict v1
proto parsing rejects. A mis-cased binding fingerprints such a server, so its
declared version is downgraded to 0.3 to route the SDK's ClientFactory onto
its v0.3 compat transport, which speaks that dialect.
"""
normalized: Final = type(agent_card)()
normalized.CopyFrom(agent_card)
for interface in normalized.supported_interfaces:
canonical: str | None = _CANONICAL_PROTOCOL_BINDINGS.get(interface.protocol_binding.lower())
if canonical is None or canonical == interface.protocol_binding:
continue
interface.protocol_binding = canonical
interface.protocol_version = _LEGACY_PROTOCOL_VERSION
return normalized
def get_agent_card_url(agent_card: "AgentCard") -> str | None:
"""Return the agent endpoint URL from the resolved SDK card."""
url: Final = getattr(agent_card, "url", None)

132
litellm/a2a_protocol/litellm_completion_bridge/transformation.py
View file

@ -17,11 +17,27 @@ A2A Streaming Events:
- Artifact update (kind: "artifact-update") - Content/artifact delivery
"""
from collections.abc import Mapping, MutableMapping, Sequence
from datetime import datetime, timezone
from typing import Any, Final
from typing import TYPE_CHECKING, Final
from uuid import uuid4
from pydantic import JsonValue, TypeAdapter, ValidationError
from litellm._logging import verbose_logger
from litellm.types.utils import ModelResponse
if TYPE_CHECKING:
from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper
_STR_KEY_MAPPING_ADAPTER: Final = TypeAdapter(Mapping[str, object])
def _as_object_mapping(value: object) -> Mapping[str, object]:
try:
return _STR_KEY_MAPPING_ADAPTER.validate_python(value)
except ValidationError:
return {}
class A2AStreamingContext:
@ -30,7 +46,7 @@ class A2AStreamingContext:
Tracks task_id, context_id, and message accumulation.
"""
def __init__(self, request_id: str, input_message: dict[str, Any]):
def __init__(self, request_id: str, input_message: Mapping[str, JsonValue]):
self.request_id = request_id
self.task_id = str(uuid4())
self.context_id = str(uuid4())
@ -46,44 +62,46 @@ class A2ACompletionBridgeTransformation:
"""
@staticmethod
def _extract_text_from_a2a_parts(parts: list[dict[str, Any]]) -> str:
def _text_from_a2a_part(part: JsonValue) -> str | None:
if not isinstance(part, dict):
return None
text: Final = part.get("text")
if text is None:
return None
if part.get("kind") not in (None, "", "text"):
return None
return str(text)
@staticmethod
def _extract_text_from_a2a_parts(parts: Sequence[JsonValue]) -> str:
"""Extract text from A2A parts (with or without explicit ``kind``)."""
content_parts: Final[list[str]] = []
for part in parts:
if not isinstance(part, dict):
continue
kind = part.get("kind")
text = part.get("text")
if text is None:
continue
if kind in (None, "", "text"):
content_parts.append(str(text))
return "\n".join(content_parts)
extracted: Final = (A2ACompletionBridgeTransformation._text_from_a2a_part(part) for part in parts)
return "\n".join(text for text in extracted if text is not None)
@staticmethod
def get_forward_metadata(
a2a_message: dict[str, Any],
params: dict[str, Any] | None = None,
) -> dict[str, Any] | None:
a2a_message: Mapping[str, JsonValue],
params: Mapping[str, JsonValue] | None = None,
) -> Mapping[str, JsonValue] | None:
"""
Merge A2A metadata from MessageSendParams and the message for downstream providers.
Forwarded once on the LangGraph run payload (``metadata``), not duplicated on
each input message — see ``apply_forward_metadata_to_completion_params``.
"""
merged: Final[dict[str, Any]] = {}
if params and isinstance(params.get("metadata"), dict):
merged.update(params["metadata"])
params_metadata: Final = params.get("metadata") if params else None
message_metadata: Final = a2a_message.get("metadata")
if isinstance(message_metadata, dict):
merged.update(message_metadata)
merged: Final[dict[str, JsonValue]] = {
**(params_metadata if isinstance(params_metadata, dict) else {}),
**(message_metadata if isinstance(message_metadata, dict) else {}),
}
return merged or None
@staticmethod
def apply_forward_metadata_to_completion_params(
completion_params: dict[str, Any],
a2a_message: dict[str, Any],
params: dict[str, Any] | None = None,
completion_params: MutableMapping[str, object],
a2a_message: Mapping[str, JsonValue],
params: Mapping[str, JsonValue] | None = None,
) -> None:
"""
Attach A2A metadata to completion kwargs for provider bridges (e.g. LangGraph).
@ -97,24 +115,20 @@ class A2ACompletionBridgeTransformation:
if not forward_metadata:
return
extra_body = completion_params.get("extra_body")
if not isinstance(extra_body, dict):
extra_body = {}
extra_body: Final = _as_object_mapping(completion_params.get("extra_body"))
# Layer client-supplied A2A metadata under any agent-owner-configured
# ``extra_body.metadata`` so the configured keys remain authoritative
# and an A2A caller cannot overwrite server-set run metadata.
existing_metadata: Final = extra_body.get("metadata")
existing_dict: Final[dict[str, Any]] = existing_metadata if isinstance(existing_metadata, dict) else {}
merged_metadata: Final[dict[str, Any]] = {**forward_metadata, **existing_dict}
extra_body = {**extra_body, "metadata": merged_metadata}
completion_params["extra_body"] = extra_body
existing_dict: Final = _as_object_mapping(extra_body.get("metadata"))
merged_metadata: Final[dict[str, object]] = {**forward_metadata, **existing_dict}
completion_params["extra_body"] = {**extra_body, "metadata": merged_metadata}
verbose_logger.debug("A2A -> completion forward metadata keys=%s", list(forward_metadata.keys()))
@staticmethod
def a2a_message_to_openai_messages(
a2a_message: dict[str, Any],
) -> list[dict[str, Any]]:
a2a_message: Mapping[str, JsonValue],
) -> list[dict[str, object]]:
"""
Transform an A2A message to OpenAI message format.
@ -125,25 +139,19 @@ class A2ACompletionBridgeTransformation:
List of OpenAI-format messages
"""
role: Final = a2a_message.get("role", "user")
parts = a2a_message.get("parts", [])
raw_parts: Final = a2a_message.get("parts", [])
# Map A2A roles to OpenAI roles
openai_role = role
if role == "user":
openai_role = "user"
elif role == "assistant":
openai_role = "assistant"
elif role == "system":
openai_role = "system"
if not isinstance(parts, list):
parts = []
openai_role: Final = (
"user" if role == "user" else "assistant" if role == "assistant" else "system" if role == "system" else role
)
parts: Final = raw_parts if isinstance(raw_parts, list) else []
content: Final = A2ACompletionBridgeTransformation._extract_text_from_a2a_parts(parts)
# Do not attach A2A message.metadata here — the completion bridge forwards it
# once at run level via extra_body.metadata (LangGraph POST /runs/wait shape).
openai_message: Final[dict[str, Any]] = {"role": openai_role, "content": content}
openai_message: Final[dict[str, object]] = {"role": openai_role, "content": content}
verbose_logger.debug(
"A2A -> OpenAI transform: role=%s -> %s, content_length=%s", role, openai_role, len(content)
@ -151,11 +159,20 @@ class A2ACompletionBridgeTransformation:
return [openai_message]
@staticmethod
def _extract_response_content(response: "ModelResponse | CustomStreamWrapper") -> str:
if not isinstance(response, ModelResponse) or not response.choices:
return ""
choice: Final = response.choices[0]
if not choice.message:
return ""
return choice.message.content or ""
@staticmethod
def openai_response_to_a2a_response(
response: Any,
response: "ModelResponse | CustomStreamWrapper",
request_id: str | None = None,
) -> dict[str, Any]:
) -> dict[str, object]:
"""
Transform a LiteLLM ModelResponse to A2A SendMessageResponse format.
@ -166,12 +183,7 @@ class A2ACompletionBridgeTransformation:
Returns:
A2A SendMessageResponse dict
"""
# Extract content from response
content = ""
if hasattr(response, "choices") and response.choices:
choice: Final = response.choices[0]
if hasattr(choice, "message") and choice.message:
content = choice.message.content or ""
content: Final = A2ACompletionBridgeTransformation._extract_response_content(response)
# Build A2A message
a2a_message: Final = {
@ -182,7 +194,7 @@ class A2ACompletionBridgeTransformation:
}
# Build A2A response
a2a_response: Final = {
a2a_response: Final[dict[str, object]] = {
"jsonrpc": "2.0",
"id": request_id,
"result": a2a_message,
@ -200,7 +212,7 @@ class A2ACompletionBridgeTransformation:
@staticmethod
def create_task_event(
ctx: A2AStreamingContext,
) -> dict[str, Any]:
) -> dict[str, object]:
"""
Create the initial task event with status 'submitted'.
@ -235,7 +247,7 @@ class A2ACompletionBridgeTransformation:
state: str,
final: bool = False,
message_text: str | None = None,
) -> dict[str, Any]:
) -> dict[str, object]:
"""
Create a status update event.
@ -245,7 +257,7 @@ class A2ACompletionBridgeTransformation:
final: Whether this is the final event
message_text: Optional message text for 'working' status
"""
status: Final[dict[str, Any]] = {
status: Final[dict[str, object]] = {
"state": state,
"timestamp": A2ACompletionBridgeTransformation._get_timestamp(),
}
@ -277,7 +289,7 @@ class A2ACompletionBridgeTransformation:
def create_artifact_update_event(
ctx: A2AStreamingContext,
text: str,
) -> dict[str, Any]:
) -> dict[str, object]:
"""
Create an artifact update event with content.

41
litellm/a2a_protocol/main.py
View file

@ -73,6 +73,7 @@ except ImportError:
from litellm.a2a_protocol.card_resolver import (
LiteLLMA2ACardResolver,
get_agent_card_url,
normalize_agent_card_interfaces,
)
from litellm.a2a_protocol.exception_mapping_utils import (
handle_a2a_localhost_retry,
@ -85,7 +86,7 @@ A2ACardResolver: Final = LiteLLMA2ACardResolver
def _set_usage_on_logging_obj(
kwargs: dict[str, Any],
kwargs: Mapping[str, object],
prompt_tokens: int,
completion_tokens: int,
) -> None:
@ -98,7 +99,7 @@ def _set_usage_on_logging_obj(
completion_tokens: Number of output tokens
"""
litellm_logging_obj: Final = kwargs.get("litellm_logging_obj")
if litellm_logging_obj is not None:
if isinstance(litellm_logging_obj, Logging):
usage: Final = litellm.Usage(
prompt_tokens=prompt_tokens,
completion_tokens=completion_tokens,
@ -108,7 +109,7 @@ def _set_usage_on_logging_obj(
def _set_agent_id_on_logging_obj(
kwargs: dict[str, Any],
kwargs: Mapping[str, object],
agent_id: str | None,
) -> None:
"""
@ -122,7 +123,7 @@ def _set_agent_id_on_logging_obj(
return
litellm_logging_obj: Final = kwargs.get("litellm_logging_obj")
if litellm_logging_obj is not None:
if isinstance(litellm_logging_obj, Logging):
# Set agent_id directly on model_call_details (same pattern as custom_llm_provider)
litellm_logging_obj.model_call_details["agent_id"] = agent_id
@ -131,7 +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],
kwargs: Mapping[str, object],
litellm_params: Mapping[str, object],
) -> None:
"""
@ -143,18 +144,22 @@ def _set_litellm_params_on_logging_obj(
context, so merge the pricing keys in rather than replacing the dict.
"""
logging_obj: Final = kwargs.get("litellm_logging_obj")
if logging_obj is None:
if not isinstance(logging_obj, Logging):
return
cost_params = {key: litellm_params[key] for key in _A2A_COST_PARAM_KEYS if litellm_params.get(key) is not None}
cost_params: Final = {
key: litellm_params[key] for key in _A2A_COST_PARAM_KEYS if litellm_params.get(key) is not None
}
if not cost_params:
return
existing: Final = logging_obj.model_call_details.get("litellm_params") or {}
logging_obj.model_call_details["litellm_params"] = {**existing, **cost_params}
logging_obj.model_call_details["litellm_params"] = {
**(logging_obj.model_call_details.get("litellm_params") or {}),
**cost_params,
}
def _get_a2a_model_info(a2a_client: "A2AClientType", kwargs: dict[str, Any]) -> str:
def _get_a2a_model_info(a2a_client: "A2AClientType", kwargs: Mapping[str, object]) -> str:
"""
Extract agent info and set model/custom_llm_provider for cost tracking.
@ -174,7 +179,7 @@ def _get_a2a_model_info(a2a_client: "A2AClientType", kwargs: dict[str, Any]) ->
# Set on litellm_logging_obj if available (for standard logging payload)
litellm_logging_obj: Final = kwargs.get("litellm_logging_obj")
if litellm_logging_obj is not None:
if isinstance(litellm_logging_obj, Logging):
litellm_logging_obj.model = model
litellm_logging_obj.custom_llm_provider = custom_llm_provider
litellm_logging_obj.model_call_details["model"] = model
@ -497,7 +502,7 @@ 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[dict[str, object]] = a2a_response.model_dump(mode="json", exclude_none=True)
response_dict: Final[dict[str, object]] = a2a_response.root.model_dump(mode="json", exclude_none=True)
(
prompt_tokens,
completion_tokens,
@ -782,13 +787,17 @@ async def create_a2a_client(
if extra_headers:
verbose_proxy_logger.debug("A2A client created with extra_headers=%s", list(extra_headers.keys()))
resolver: Final = A2ACardResolver(httpx_client=httpx_client, base_url=base_url)
agent_card: Final = normalize_agent_card_interfaces(
await resolver.get_agent_card(http_kwargs={"headers": extra_headers} if extra_headers else None)
)
a2a_client: Final = await create_client( # pyright: ignore[reportOptionalCall]
base_url,
agent_card,
client_config=ClientConfig( # pyright: ignore[reportOptionalCall]
httpx_client=httpx_client,
streaming=streaming,
),
resolver_http_kwargs={"headers": extra_headers} if extra_headers else None,
)
# Stash LiteLLM-owned handles on the client so the localhost-retry path can reuse
# the configured httpx client and this agent's headers without excavating
@ -799,9 +808,7 @@ async def create_a2a_client(
if extra_headers
else None
)
agent_card: Final = getattr(a2a_client, "_card", None)
if agent_card is not None:
a2a_client._litellm_agent_card = agent_card
a2a_client._litellm_agent_card = agent_card
verbose_logger.info("A2A client created for %s", base_url)

269
litellm/batches/batch_utils.py
View file

@ -1,6 +1,8 @@
import json
from collections.abc import Iterable, Iterator, Mapping
from dataclasses import dataclass
from dataclasses import replace as dataclasses_replace
from enum import Enum
from typing import Any, Final, Literal
import litellm
@ -12,12 +14,23 @@ from litellm.types.utils import CallTypes, ModelInfo, Usage
from litellm.utils import token_counter
@dataclass(frozen=True, slots=True)
class BatchCostUsageResult:
"""Aggregate cost, usage, and per-line pass/fail counts for a completed batch."""
cost: float
usage: Usage
models: list[str]
successful_requests: int
failed_requests: int
async def calculate_batch_cost_and_usage(
file_content_dictionary: list[dict],
custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic"],
model_name: str | None = None,
model_info: ModelInfo | None = None,
) -> tuple[float, Usage, list[str]]:
) -> BatchCostUsageResult:
"""
Calculate the cost and usage of a batch.
@ -32,8 +45,7 @@ async def calculate_batch_cost_and_usage(
and model_name
and getattr(litellm, "disable_vertex_batch_output_transformation", False)
):
batch_cost, batch_usage = calculate_vertex_ai_batch_cost_and_usage(file_content_dictionary, model_name)
return batch_cost, batch_usage, [model_name]
return calculate_vertex_ai_batch_cost_and_usage(file_content_dictionary, model_name)
return _aggregate_batch_cost_usage_models(
entries=file_content_dictionary,
@ -49,7 +61,7 @@ async def _handle_completed_batch(
model_name: str | None = None,
litellm_params: dict | None = None,
model_info: ModelInfo | None = None,
) -> tuple[float, Usage, list[str]]:
) -> BatchCostUsageResult:
"""Fetch a completed batch's output file and aggregate its cost, usage, and
models in a single pass over the JSONL lines, so the parsed file content is
never materialized in memory.
@ -72,27 +84,49 @@ async def _handle_completed_batch(
# The generic retrieval helper keeps raising for callers that explicitly ask
# for a missing output file.
if batch.output_file_id is None:
return 0.0, Usage(prompt_tokens=0, completion_tokens=0, total_tokens=0), []
return BatchCostUsageResult(
cost=0.0,
usage=Usage(prompt_tokens=0, completion_tokens=0, total_tokens=0),
models=[], # mutable-ok: no output file means no model was ever priced; BatchCostUsageResult.models requires list[str]
successful_requests=0,
failed_requests=await count_error_file_failed_requests(
batch, custom_llm_provider=custom_llm_provider, litellm_params=litellm_params
),
)
file_content = await _fetch_batch_output_file_content(batch, custom_llm_provider, litellm_params=litellm_params)
if (
custom_llm_provider == "vertex_ai"
and model_name
and getattr(litellm, "disable_vertex_batch_output_transformation", False)
):
batch_cost, batch_usage = calculate_vertex_ai_batch_cost_and_usage(
_get_file_content_as_dictionary(file_content), model_name
)
return batch_cost, batch_usage, [model_name]
return _aggregate_batch_cost_usage_models(
entries=_iter_batch_output_entries(file_content),
custom_llm_provider=custom_llm_provider,
model_name=model_name,
model_info=model_info,
error_file_failed_requests: Final = await count_error_file_failed_requests(
batch, custom_llm_provider=custom_llm_provider, litellm_params=litellm_params
)
output_file_result: Final = (
calculate_vertex_ai_batch_cost_and_usage(_get_file_content_as_dictionary(file_content), model_name)
if (
custom_llm_provider == "vertex_ai"
and model_name
and getattr(litellm, "disable_vertex_batch_output_transformation", False)
)
else _aggregate_batch_cost_usage_models(
entries=_iter_batch_output_entries(file_content),
custom_llm_provider=custom_llm_provider,
model_name=model_name,
model_info=model_info,
)
)
if not error_file_failed_requests:
return output_file_result
return dataclasses_replace(
output_file_result, failed_requests=output_file_result.failed_requests + error_file_failed_requests
)
class _LineOutcome(Enum):
"""A batch output line that yielded no billable stats."""
PROVIDER_FAILED = "provider_failed"
UNCOSTABLE = "uncostable"
@dataclass(frozen=True, slots=True)
class _BatchOutputLineStats:
@ -102,19 +136,27 @@ class _BatchOutputLineStats:
total_tokens: int
cache_read_tokens: int
cache_creation_tokens: int
reasoning_tokens: int
model: str | None
def _iter_successful_output_line_stats(
def _classify_output_line_stats(
entries: Iterable[dict],
custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic", "bedrock"],
model_name: str | None,
model_info: ModelInfo | None,
) -> Iterator[_BatchOutputLineStats]:
) -> Iterator[_BatchOutputLineStats | _LineOutcome]:
"""Classify every output line in a single pass, so counting failures never needs
a second read of a potentially huge output file. A line the provider reported as
failed yields ``PROVIDER_FAILED``; a successful line litellm could not price
yields ``UNCOSTABLE`` and still counts as a successful request billed at $0, so
the counts stay reconcilable with the provider's own ``request_counts``."""
for entry in entries:
if not _batch_response_was_successful(entry, custom_llm_provider):
yield _LineOutcome.PROVIDER_FAILED
continue
stats = _safe_output_line_stats(entry, custom_llm_provider, model_name, model_info)
if stats is not None:
yield stats
yield stats if stats is not None else _LineOutcome.UNCOSTABLE
def _safe_output_line_stats(
@ -123,13 +165,11 @@ def _safe_output_line_stats(
model_name: str | None,
model_info: ModelInfo | None,
) -> _BatchOutputLineStats | None:
"""Return the stats for one batch output line, or None for a line that is
unsuccessful or cannot be costed, so a single bad line never aborts the
whole batch's cost accounting."""
"""Return the stats for one provider-successful batch output line, or None when
it cannot be costed, so a single bad line never aborts the whole batch's cost
accounting."""
custom_id: Final = entry.get("custom_id") if isinstance(entry, dict) else None
try:
if not _batch_response_was_successful(entry, custom_llm_provider):
return None
return _compute_output_line_stats(entry, custom_llm_provider, model_name, model_info)
except Exception as e: # noqa: BLE001 # any single line's costing failure must not abort the whole batch
verbose_logger.warning(
@ -152,6 +192,7 @@ def _compute_output_line_stats(
prompt_details: Final = parse_prompt_tokens_details(usage)
raw_model: Final = response_body.get("model")
response_model: Final = raw_model if isinstance(raw_model, str) and raw_model else None
completion_details: Final = usage.completion_tokens_details
return _BatchOutputLineStats(
cost=_output_line_cost(
response_body=response_body,
@ -166,6 +207,7 @@ def _compute_output_line_stats(
total_tokens=usage.total_tokens,
cache_read_tokens=prompt_details["cache_hit_tokens"],
cache_creation_tokens=prompt_details["cache_creation_tokens"],
reasoning_tokens=(completion_details.reasoning_tokens if completion_details else None) or 0,
model=response_model,
)
@ -203,10 +245,14 @@ def _aggregate_batch_cost_usage_models(
custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic", "bedrock"],
model_name: str | None = None,
model_info: ModelInfo | None = None,
) -> tuple[float, Usage, list[str]]:
"""Aggregate cost, usage, and models from batch output entries in a single
pass, holding one small stats record per line instead of the parsed file."""
line_stats: Final = tuple(_iter_successful_output_line_stats(entries, custom_llm_provider, model_name, model_info))
) -> BatchCostUsageResult:
"""Aggregate cost, usage, models, and pass/fail counts from batch output
entries in a single pass, holding one small stats record per line instead
of the parsed file."""
all_results: Final = tuple(_classify_output_line_stats(entries, custom_llm_provider, model_name, model_info))
line_stats: Final = tuple(result for result in all_results if isinstance(result, _BatchOutputLineStats))
failed_requests: Final = sum(1 for result in all_results if result is _LineOutcome.PROVIDER_FAILED)
successful_requests: Final = len(all_results) - failed_requests
cache_token_params: Final = {
key: tokens
@ -220,18 +266,32 @@ def _aggregate_batch_cost_usage_models(
total_tokens=sum(stats.total_tokens for stats in line_stats),
prompt_tokens=sum(stats.prompt_tokens for stats in line_stats),
completion_tokens=sum(stats.completion_tokens for stats in line_stats),
reasoning_tokens=sum(stats.reasoning_tokens for stats in line_stats),
**cache_token_params,
)
batch_models: Final = [model_name] if model_name else [stats.model for stats in line_stats if stats.model]
total_cost: Final = sum((stats.cost for stats in line_stats), 0.0)
verbose_logger.debug("batch output aggregate: cost=%s usage=%s models=%s", total_cost, batch_usage, batch_models)
return total_cost, batch_usage, batch_models
verbose_logger.debug(
"batch output aggregate: cost=%s usage=%s models=%s successful=%d failed=%d",
total_cost,
batch_usage,
batch_models,
successful_requests,
failed_requests,
)
return BatchCostUsageResult(
cost=total_cost,
usage=batch_usage,
models=batch_models,
successful_requests=successful_requests,
failed_requests=failed_requests,
)
def calculate_vertex_ai_batch_cost_and_usage(
vertex_ai_batch_responses: list[dict],
model_name: str | None = None,
) -> tuple[float, Usage]:
) -> BatchCostUsageResult:
"""
Calculate both cost and usage from raw Vertex AI batch responses.
@ -242,6 +302,10 @@ def calculate_vertex_ai_batch_cost_and_usage(
{"request": ..., "response": {"candidates": [...], "usageMetadata": {...}}}
usageMetadata contains promptTokenCount, candidatesTokenCount, totalTokenCount.
A row with no ``response`` is counted as failed - the same signal already
used to skip it from cost/usage aggregation, since Vertex batch prediction
output doesn't establish a distinct error shape in this (non-default) path.
"""
from litellm.cost_calculator import batch_cost_calculator
@ -249,12 +313,16 @@ def calculate_vertex_ai_batch_cost_and_usage(
total_tokens = 0
prompt_tokens = 0
completion_tokens = 0
successful_requests = 0 # rebind-ok: loop accumulator, matches total_cost/total_tokens above
failed_requests = 0 # rebind-ok: loop accumulator, matches total_cost/total_tokens above
actual_model_name: Final = model_name or "gemini-2.0-flash-001"
for response in vertex_ai_batch_responses:
response_body = response.get("response")
if response_body is None:
failed_requests += 1
continue
successful_requests += 1
usage_metadata = response_body.get("usageMetadata", {})
_prompt = usage_metadata.get("promptTokenCount", 0) or 0
@ -282,20 +350,84 @@ def calculate_vertex_ai_batch_cost_and_usage(
total_tokens += _total
verbose_logger.info(
"vertex_ai batch cost: cost=%s, prompt=%d, completion=%d, total=%d",
"vertex_ai batch cost: cost=%s, prompt=%d, completion=%d, total=%d, successful=%d, failed=%d",
total_cost,
prompt_tokens,
completion_tokens,
total_tokens,
successful_requests,
failed_requests,
)
return total_cost, Usage(
total_tokens=total_tokens,
prompt_tokens=prompt_tokens,
completion_tokens=completion_tokens,
return BatchCostUsageResult(
cost=total_cost,
usage=Usage(
total_tokens=total_tokens,
prompt_tokens=prompt_tokens,
completion_tokens=completion_tokens,
),
models=[actual_model_name],
successful_requests=successful_requests,
failed_requests=failed_requests,
)
def _provider_output_file_id(output_file_id: str) -> str:
"""
Resolve the file id the provider actually knows: unified ids yield their embedded
llm_output_file_id, model-encoded ids decode to the raw provider id, raw ids pass through.
"""
from litellm.proxy.openai_files_endpoints.common_utils import (
_is_base64_encoded_unified_file_id,
get_original_file_id,
)
unified_file_id: Final = _is_base64_encoded_unified_file_id(output_file_id)
if not unified_file_id:
return get_original_file_id(output_file_id)
try:
extracted: Final = unified_file_id.split("llm_output_file_id,")[1].split(";")[0]
except (IndexError, AttributeError) as e:
verbose_logger.error(
"Failed to extract LLM output file ID from unified file ID: %s, error: %s",
output_file_id,
e,
)
return output_file_id
verbose_logger.debug("Extracted LLM output file ID from unified file ID: %s", extracted)
return extracted
async def _fetch_batch_managed_file_content(
file_id: str,
custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic"] = "openai",
litellm_params: dict | None = None,
) -> bytes:
"""
Fetch a batch's output or error file and return its raw JSONL bytes.
Args:
file_id: The provider or unified (litellm-managed) file id to fetch
custom_llm_provider: The LLM provider
litellm_params: Optional litellm parameters containing credentials (api_key, api_base, etc.)
Required for Azure and other providers that need authentication
"""
from litellm.files.main import afile_content
# Build kwargs for afile_content with credentials from litellm_params
file_content_kwargs: Final = {
"file_id": _provider_output_file_id(file_id),
"custom_llm_provider": custom_llm_provider,
}
# Extract and add credentials for file access
credentials: Final = _extract_file_access_credentials(litellm_params)
file_content_kwargs.update(credentials)
_file_content: Final = await afile_content(**file_content_kwargs)
return _file_content.content
async def _fetch_batch_output_file_content(
batch: Batch,
custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic"] = "openai",
@ -310,37 +442,36 @@ async def _fetch_batch_output_file_content(
litellm_params: Optional litellm parameters containing credentials (api_key, api_base, etc.)
Required for Azure and other providers that need authentication
"""
from litellm.files.main import afile_content
from litellm.proxy.openai_files_endpoints.common_utils import (
_is_base64_encoded_unified_file_id,
)
if batch.output_file_id is None:
raise ValueError("Output file id is None cannot retrieve file content")
file_id = batch.output_file_id
is_base64_unified_file_id: Final = _is_base64_encoded_unified_file_id(file_id)
if is_base64_unified_file_id:
try:
file_id = is_base64_unified_file_id.split("llm_output_file_id,")[1].split(";")[0]
verbose_logger.debug("Extracted LLM output file ID from unified file ID: %s", file_id)
except (IndexError, AttributeError) as e:
verbose_logger.error(
"Failed to extract LLM output file ID from unified file ID: %s, error: %s", batch.output_file_id, e
)
return await _fetch_batch_managed_file_content(
batch.output_file_id, custom_llm_provider=custom_llm_provider, litellm_params=litellm_params
)
# Build kwargs for afile_content with credentials from litellm_params
file_content_kwargs: Final = {
"file_id": file_id,
"custom_llm_provider": custom_llm_provider,
}
# Extract and add credentials for file access
credentials: Final = _extract_file_access_credentials(litellm_params)
file_content_kwargs.update(credentials)
async def count_error_file_failed_requests(
batch: Batch,
custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic"],
litellm_params: dict | None,
) -> int:
"""Count failed requests reported only in the batch's separate error file.
_file_content: Final = await afile_content(**file_content_kwargs)
return _file_content.content
OpenAI-shaped batch providers write successful lines to ``output_file_id``
and per-request failures (e.g. a rejected param) to a distinct
``error_file_id`` - they never appear in the output file at all, so
counting failures from the output file alone silently undercounts them.
"""
if batch.error_file_id is None:
return 0
try:
error_file_content = await _fetch_batch_managed_file_content(
batch.error_file_id, custom_llm_provider=custom_llm_provider, litellm_params=litellm_params
)
except Exception as e: # noqa: BLE001 # a failed/missing error file must not abort cost tracking for the batch
verbose_logger.debug("Failed to fetch batch error file %s: %s", batch.error_file_id, e)
return 0
return sum(1 for _ in _iter_batch_input_lines(error_file_content))
def _extract_file_access_credentials(litellm_params: dict | None) -> dict:
@ -537,7 +668,7 @@ def _get_batch_job_usage_from_response_body(
return usage
def _get_anthropic_result_from_batch_results_line(batch_results_line: Mapping[str, Any]) -> dict:
def _get_anthropic_result_from_batch_results_line(batch_results_line: Mapping[str, Any]) -> Mapping[str, Any]:
"""
Get the ``result`` object from a line of an Anthropic message batch results JSONL file.
@ -549,7 +680,7 @@ def _get_anthropic_result_from_batch_results_line(batch_results_line: Mapping[st
def _get_response_from_batch_job_output_file(
batch_job_output_file: Mapping[str, Any], custom_llm_provider: str = "openai"
) -> Any:
) -> Mapping[str, Any]:
"""
Get the response from the batch job output file
"""

2
litellm/batches/main.py
View file

@ -390,7 +390,7 @@ def _handle_retrieve_batch_providers_without_provider_config(
custom_llm_provider: Literal[
"openai", "azure", "vertex_ai", "bedrock", "hosted_vllm", "litellm_proxy", "anthropic"
] = "openai",
logging_obj: Any | None = None,
logging_obj: LiteLLMLoggingObj | None = None,
):
api_base: str | None = None
if custom_llm_provider in OPENAI_COMPATIBLE_BATCH_AND_FILES_PROVIDERS:

9
litellm/caching/caching.py
View file

@ -12,6 +12,7 @@ import hashlib
import json
import time
import traceback
from collections.abc import Mapping
from enum import Enum
from typing import Any, Final
@ -506,7 +507,7 @@ class Cache:
def _get_cache_logic(
self,
cached_result: Any | None,
cached_result: object | None,
max_age: float | None,
):
"""
@ -538,8 +539,8 @@ class Cache:
return cached_result
@staticmethod
def _get_safe_cache_lookup_kwargs(kwargs: dict[str, Any]) -> dict[str, Any]:
cache_lookup_kwargs: Final[dict[str, Any]] = {}
def _get_safe_cache_lookup_kwargs(kwargs: Mapping[str, object]) -> dict[str, object]:
cache_lookup_kwargs: Final[dict[str, object]] = {}
for prompt_kwarg in ("messages", "input"):
if prompt_kwarg in kwargs:
cache_lookup_kwargs[prompt_kwarg] = kwargs[prompt_kwarg]
@ -552,7 +553,7 @@ class Cache:
@staticmethod
def _update_metadata_from_cache_lookup_kwargs(
original_kwargs: dict[str, Any], cache_lookup_kwargs: dict[str, Any]
original_kwargs: Mapping[str, object], cache_lookup_kwargs: Mapping[str, object]
) -> None:
original_metadata: Final = original_kwargs.get("metadata")
cache_lookup_metadata: Final = cache_lookup_kwargs.get("metadata")

43
litellm/caching/caching_handler.py
View file

@ -18,7 +18,7 @@ import asyncio
import datetime
import inspect
import time
from collections.abc import AsyncGenerator, AsyncIterator, Callable, Generator, Mapping
from collections.abc import AsyncGenerator, AsyncIterator, Awaitable, Callable, Generator, Mapping
from typing import TYPE_CHECKING, Any, Final, Optional, TypeVar
from pydantic import BaseModel
@ -27,6 +27,7 @@ import litellm
from litellm._logging import print_verbose, verbose_logger
from litellm.caching import InMemoryCache
from litellm.caching.caching import S3Cache
from litellm.constants import CACHE_WRITE_SHUTDOWN_FLUSH_TIMEOUT_SECONDS
from litellm.litellm_core_utils.llm_response_utils.response_metadata import (
update_response_metadata,
)
@ -124,6 +125,29 @@ def _prompt_tokens_details_as_mapping(details: "PromptTokensDetailsWrapper") ->
return details.model_dump(exclude_none=True) if hasattr(details, "model_dump") else {}
_PENDING_CACHE_WRITES: Final[set["asyncio.Task[None]"]] = set() # mutable-ok: strong refs to pending write tasks
async def _complete_cache_write_despite_cancellation(write_factory: Callable[[], Awaitable[None]]) -> None:
try:
await write_factory()
except asyncio.CancelledError:
try:
await asyncio.wait_for(write_factory(), timeout=CACHE_WRITE_SHUTDOWN_FLUSH_TIMEOUT_SECONDS)
except Exception as flush_error: # noqa: BLE001 # shutdown flush failures are logged, never raised
verbose_logger.warning(
"LiteLLM Cache: pending cache write failed during event loop shutdown: %s", flush_error
)
raise
def create_cache_write_task(write_factory: Callable[[], Awaitable[None]]) -> "asyncio.Task[None]":
task: Final = asyncio.create_task(_complete_cache_write_despite_cancellation(write_factory))
_PENDING_CACHE_WRITES.add(task)
task.add_done_callback(_PENDING_CACHE_WRITES.discard)
return task
def _request_cache_key(request_kwargs: Mapping[str, Any]) -> str | None:
"""Read the caller-supplied ``cache_key`` off the request kwargs."""
return request_kwargs.get("cache_key", None)
@ -983,6 +1007,7 @@ class LLMCachingHandler:
if litellm.cache is None:
return
cache: Final = litellm.cache
new_kwargs: Final = kwargs.copy()
new_kwargs.update(
@ -1004,24 +1029,24 @@ class LLMCachingHandler:
):
if (
isinstance(result, EmbeddingResponse)
and litellm.cache is not None
and not isinstance(litellm.cache.cache, S3Cache) # s3 doesn't support bulk writing. Exclude.
and not isinstance(cache.cache, S3Cache) # s3 doesn't support bulk writing. Exclude.
):
asyncio.create_task(
litellm.cache.async_add_cache_pipeline(
create_cache_write_task(
lambda: cache.async_add_cache_pipeline(
result, dynamic_cache_object=self.dual_cache, **new_kwargs
)
)
else:
asyncio.create_task(
litellm.cache.async_add_cache(
result.model_dump_json(),
result_json: Final = result.model_dump_json()
create_cache_write_task(
lambda: cache.async_add_cache(
result_json,
dynamic_cache_object=self.dual_cache,
**new_kwargs,
)
)
else:
asyncio.create_task(litellm.cache.async_add_cache(result, **new_kwargs))
create_cache_write_task(lambda: cache.async_add_cache(result, **new_kwargs))
def sync_set_cache(
self,

18
litellm/caching/qdrant_semantic_cache.py
View file

@ -12,7 +12,7 @@ import ast
import asyncio
import json
import os
from typing import TYPE_CHECKING, Any, Final, cast
from typing import TYPE_CHECKING, Any, Final, Protocol, cast
import litellm
from litellm._logging import print_verbose
@ -39,6 +39,12 @@ if TYPE_CHECKING:
from litellm.router import Router
class _QdrantCollectionDetailsResponse(Protocol):
"""The qdrant `/collections/{name}` response, whose body is kept as an opaque JSON object."""
def json(self) -> dict[str, object]: ...
class QdrantSemanticCache(BaseCache):
CACHE_KEY_FIELD_NAME = "litellm_cache_key"
embedding_max_input_tokens: int | None = None
@ -115,15 +121,15 @@ class QdrantSemanticCache(BaseCache):
raise ValueError(f"Error from qdrant checking if /collections exist {collection_exists.text}")
if collection_exists.json()["result"]["exists"]:
collection_details = self.sync_client.get(
collection_details: _QdrantCollectionDetailsResponse = self.sync_client.get(
url=f"{self.qdrant_api_base}/collections/{self.collection_name}",
headers=self.headers,
)
self.collection_info = collection_details.json()
self.collection_info: dict[str, object] = collection_details.json()
print_verbose(f"Collection already exists.\nCollection details:{self.collection_info}")
self._ensure_cache_key_payload_index()
else:
quantization_params: dict[str, Any]
quantization_params: dict[str, dict[str, object]]
if quantization_config is None or quantization_config == "binary":
quantization_params = {
"binary": {
@ -214,7 +220,7 @@ class QdrantSemanticCache(BaseCache):
resolve_embedding_max_input_tokens(self.embedding_max_input_tokens, self.embedding_model, router),
)
def _get_embedding(self, prompt: str, metadata: dict[str, Any] | None = None) -> EmbeddingResponse:
def _get_embedding(self, prompt: str, metadata: dict[str, object] | None = None) -> EmbeddingResponse:
"""Embed via the proxy Router when it serves the model, else direct."""
try:
from litellm.proxy.proxy_server import llm_model_list, llm_router
@ -241,7 +247,7 @@ class QdrantSemanticCache(BaseCache):
num_retries=0,
)
async def _get_async_embedding(self, prompt: str, metadata: dict[str, Any] | None = None) -> EmbeddingResponse:
async def _get_async_embedding(self, prompt: str, metadata: dict[str, object] | None = None) -> EmbeddingResponse:
try:
from litellm.proxy.proxy_server import llm_model_list, llm_router
except ImportError:

70
litellm/caching/redis_cache.py
View file

@ -18,7 +18,7 @@ import time
from collections.abc import Awaitable, Callable, Sequence
from contextvars import ContextVar
from datetime import timedelta
from typing import TYPE_CHECKING, Any, Final, TypeVar, cast
from typing import TYPE_CHECKING, Any, Final, Protocol, TypeVar, cast
import litellm
from litellm._logging import print_verbose, verbose_logger
@ -58,6 +58,26 @@ else:
Span = Any
class _AsyncRedisCommands(Protocol):
"""Async redis commands this cache issues.
redis-py's type stubs omit these methods on RedisCluster, so the union returned by
init_async_client() is untyped at every call site without this protocol.
"""
def ping(self) -> Awaitable[bool]: ...
def delete(self, *names: str) -> Awaitable[int]: ...
def ttl(self, name: str) -> Awaitable[int]: ...
def rpush(self, name: str, *values: str | bytes | float) -> Awaitable[int]: ...
def lpop(self, name: str, count: int | None = None) -> Awaitable[object]: ...
def pipeline(self, transaction: bool = True) -> "Pipeline[bytes]": ...
def _get_call_stack_info(num_frames: int = 2) -> str:
"""
Get the function names from the previous 1-2 functions in the call stack.
@ -175,6 +195,10 @@ _RedisCallResult = TypeVar("_RedisCallResult")
_swallowed_redis_failures: Final[ContextVar[int]] = ContextVar("litellm_swallowed_redis_failures", default=0)
def _opaque_kwarg_key(value: object) -> str:
return f"{type(value).__name__}-{id(value)}"
@functools.lru_cache(maxsize=1)
def _redis_health_error_types() -> tuple[type, ...]:
"""Exception types that mean the Redis backend itself is unhealthy.
@ -399,10 +423,9 @@ class RedisCache(BaseCache):
Generate a cache key for the async Redis client based on connection parameters.
This ensures different Redis configurations use different cached clients.
"""
# Create a stable representation of redis_kwargs for hashing
# Sort keys to ensure consistent hash regardless of parameter order
sorted_kwargs: Final = sorted(self.redis_kwargs.items())
kwargs_str: Final = json.dumps(sorted_kwargs, sort_keys=True)
kwargs_str: Final = json.dumps(sorted_kwargs, sort_keys=True, default=_opaque_kwarg_key)
kwargs_hash: Final = hashlib.sha256(kwargs_str.encode()).hexdigest()[:16]
return f"async-redis-client-{kwargs_hash}"
@ -426,13 +449,16 @@ class RedisCache(BaseCache):
self.redis_async_client = redis_async_client
return redis_async_client
def _async_commands(self) -> _AsyncRedisCommands:
return self.init_async_client()
def check_and_fix_namespace(self, key: str) -> str:
"""
Make sure each key starts with the given namespace
"""
if key is None:
return key
if self.namespace is not None and not key.startswith(self.namespace):
if self.namespace and not key.startswith(self.namespace + ":"):
key = self.namespace + ":" + key
return key
@ -1052,19 +1078,17 @@ class RedisCache(BaseCache):
await self.async_set_cache_pipeline(self.redis_batch_writing_buffer)
self.redis_batch_writing_buffer = []
def _get_cache_logic(self, cached_response: Any):
def _get_cache_logic(self, cached_response: bytes | str | None):
"""
Common 'get_cache_logic' across sync + async redis client implementations
"""
if cached_response is None:
return cached_response
# cached_response is in `b{} convert it to ModelResponse
cached_response = cached_response.decode("utf-8") # Convert bytes to string
return None
decoded: Final = cached_response.decode("utf-8") if isinstance(cached_response, bytes) else cached_response
try:
cached_response = json.loads(cached_response) # Convert string to dictionary
return json.loads(decoded)
except Exception:
cached_response = ast.literal_eval(cached_response)
return cached_response
return ast.literal_eval(decoded)
def get_cache(self, key, parent_otel_span: Span | None = None, **kwargs):
try:
@ -1311,8 +1335,7 @@ class RedisCache(BaseCache):
raise e
async def ping(self) -> bool:
# typed as Any, redis python lib has incomplete type stubs for RedisCluster and does not include `ping`
_redis_client: Final[Any] = self.init_async_client()
_redis_client: Final = self._async_commands()
start_time: Final = time.time()
print_verbose("Pinging Async Redis Cache")
try:
@ -1346,8 +1369,7 @@ class RedisCache(BaseCache):
@_redis_circuit_breaker_guard
async def delete_cache_keys(self, keys):
# typed as Any, redis python lib has incomplete type stubs for RedisCluster and does not include `delete`
_redis_client: Final[Any] = self.init_async_client()
_redis_client: Final = self._async_commands()
keys = [self.check_and_fix_namespace(key=key) for key in keys]
# keys is a list, unpack it so it gets passed as individual elements to delete
await _redis_client.delete(*keys)
@ -1384,10 +1406,10 @@ class RedisCache(BaseCache):
dict: {"status": "success" | "failed", "message": str, "error": Optional[str]}
"""
try:
import redis.asyncio as redis_async
from .._redis import get_redis_async_client
# Create a fresh Redis client with current settings
redis_client: Final = redis_async.Redis(**self.redis_kwargs)
redis_client: Final = get_redis_async_client(**self.redis_kwargs)
# Test the connection
ping_result: Final = await redis_client.ping()
@ -1412,8 +1434,7 @@ class RedisCache(BaseCache):
@_redis_circuit_breaker_guard
async def async_delete_cache(self, key: str):
# typed as Any, redis python lib has incomplete type stubs for RedisCluster and does not include `delete`
_redis_client: Final[Any] = self.init_async_client()
_redis_client: Final = self._async_commands()
key = self.check_and_fix_namespace(key=key)
# keys is str
return await _redis_client.delete(key)
@ -1520,8 +1541,7 @@ class RedisCache(BaseCache):
Redis ref: https://redis.io/docs/latest/commands/ttl/
"""
try:
# typed as Any, redis python lib has incomplete type stubs for RedisCluster and does not include `ttl`
_redis_client: Final[Any] = self.init_async_client()
_redis_client: Final = self._async_commands()
key = self.check_and_fix_namespace(key=key)
ttl: Final = await _redis_client.ttl(key)
if ttl <= -1: # -1 means the key does not exist, -2 key does not exist
@ -1551,7 +1571,7 @@ class RedisCache(BaseCache):
Returns:
int: The length of the list after the push operation
"""
_redis_client: Final[Any] = self.init_async_client()
_redis_client: Final = self._async_commands()
key = self.check_and_fix_namespace(key=key)
start_time: Final = time.time()
try:
@ -1618,7 +1638,7 @@ class RedisCache(BaseCache):
if len(rpush_list) == 0:
return []
_redis_client: Final[Any] = self.init_async_client()
_redis_client: Final = self._async_commands()
start_time: Final = time.time()
try:
@ -1675,7 +1695,7 @@ class RedisCache(BaseCache):
parent_otel_span: Span | None = None,
**kwargs,
) -> Any | list[Any]:
_redis_client: Final[Any] = self.init_async_client()
_redis_client: Final = self._async_commands()
key = self.check_and_fix_namespace(key=key)
start_time: Final = time.time()
print_verbose(f"LPOP from Redis list: key: {key}, count: {count}")
@ -1807,7 +1827,7 @@ class RedisCache(BaseCache):
if len(lpop_list) == 0:
return []
_redis_client: Final[Any] = self.init_async_client()
_redis_client: Final = self._async_commands()
start_time: Final = time.time()
try:

17
litellm/caching/redis_cluster_cache.py
View file

@ -64,22 +64,9 @@ class RedisClusterCache(RedisCache):
dict: {"status": "success" | "failed", "message": str, "error": Optional[str]}
"""
try:
import redis.asyncio as redis_async
from redis.cluster import ClusterNode
from .._redis import get_redis_async_client
# Create ClusterNode objects from startup_nodes
cluster_kwargs: Final = self.redis_kwargs.copy()
startup_nodes: Final = cluster_kwargs.pop("startup_nodes", [])
new_startup_nodes: Final[list[ClusterNode]] = []
for item in startup_nodes:
new_startup_nodes.append(ClusterNode(**item))
# Create a fresh Redis Cluster client with current settings
redis_client: Final = redis_async.RedisCluster(
startup_nodes=new_startup_nodes,
**cluster_kwargs,
)
redis_client: Final = get_redis_async_client(**self.redis_kwargs)
# Test the connection
ping_result: Final = await redis_client.ping()

201
litellm/caching/redis_cluster_node_isolation.py Normal file
View file

@ -0,0 +1,201 @@
"""Bounds the blast radius of a single node's transient connection error on the async
Redis Cluster client.
redis-py's ``RedisCluster._execute_command`` responds to a ``ConnectionError`` or
``TimeoutError`` on ANY one node by tearing down every node's connections and flipping
the client into "needs reinitialization", which forces every other concurrent caller
sharing this client through one reinit lock until the whole cluster topology is
re-walked. Under real proxy load, a client-side socket timeout on a single node is a
routine event (the event loop was too busy to read the response before ``socket_timeout``
elapsed) and does not mean the cluster's topology moved, so treating it as a full-cluster
event turns one slow node into a proxy-wide latency spike while Redis itself stays
healthy -- confirmed live: pausing one of three local cluster nodes made every concurrent
command against the other two, untouched nodes stall for the full pause duration too.
``get_litellm_async_redis_cluster_class`` returns a ``RedisCluster`` subclass that resets
only the node that actually failed (mirroring what a plain, non-cluster Redis client
already does when one of its pooled connections errors), leaving every other node's
connections untouched. Every other branch (MOVED, ASK, CLUSTERDOWN, slot-not-covered,
retry-exhaustion) is unchanged from upstream, since those already carry real evidence the
topology changed.
redis-py 8.x fixed this upstream with gentler machinery than this override's
``node.disconnect()`` (which also kills connections other coroutines are mid-operation
on, so one timeout cascades into a reconnect storm and, with TLS, a fresh handshake per
killed connection): it marks in-use connections for reconnect only after their current
operation completes, disconnects only the idle pooled ones, and defers reinitialization
to the outer retry loop. When the installed ``ClusterNode`` has that per-connection
recovery API, the factory returns the base ``RedisCluster`` unmodified.
"""
import asyncio
from typing import TYPE_CHECKING, Final, Protocol
from litellm._logging import verbose_logger
if TYPE_CHECKING:
from redis.asyncio.cluster import RedisCluster as _AsyncRedisClusterType
class _ClusterNodeAttrs(Protocol):
"""The subset of ``redis.asyncio.cluster.ClusterNode`` this override reads. redis-py
ships no resolvable stub for these members under the repo's current types-redis pin,
so a plain attribute access resolves every downstream use to ``Unknown`` under strict
mode; typing ``target_node`` as this Protocol at the one boundary keeps the override's
own logic fully typed without a banned ``typing.cast``."""
async def execute_command(
self,
*args: object,
**kwargs: object, # kwargs-ok: mirrors redis-py's own ClusterNode.execute_command signature, a raw command dispatch with no fixed keyword contract
) -> object: ...
async def disconnect(self) -> None: ...
class _NodesManagerAttrs(Protocol):
_moved_exception: object
def get_node_from_slot(
self, slot: int, read_from_replicas: bool, load_balancing_strategy: object
) -> _ClusterNodeAttrs: ...
class _ClusterAttrs(Protocol):
RedisClusterRequestTTL: int
reinitialize_counter: int
reinitialize_steps: int
read_from_replicas: bool
load_balancing_strategy: object
nodes_manager: _NodesManagerAttrs
def get_node(self, node_name: str) -> _ClusterNodeAttrs: ...
async def _determine_slot(self, *args: object) -> int: ...
async def aclose(self) -> None: ...
#: redis-py versions this override's copied ``_execute_command`` body has been verified
#: against. A version outside this set may have changed the method's structure in a way
#: this override can't see (Python won't error -- it'll just run our now-stale copy), so
#: construction logs a loud warning rather than silently trusting an unverified copy.
_VERIFIED_REDIS_VERSIONS: Final = frozenset({"5.3.1"})
def get_litellm_async_redis_cluster_class(
cluster_node_class: type | None = None,
) -> type["_AsyncRedisClusterType"]:
"""Returns the base ``RedisCluster`` when the installed redis-py already recovers a
node-level connection error per-connection (8.x+), else builds the ``RedisCluster``
subclass with the per-node isolation fix for older versions whose upstream branch
tears down the whole cluster client.
``cluster_node_class`` exists for dependency injection in tests; production callers
leave it unset and the installed ``ClusterNode`` is used.
Imported lazily because this module is reachable from a base ``import litellm`` while
redis is not a base dependency. Cheap to call repeatedly: the underlying redis
submodules are cached in ``sys.modules`` after the first import.
"""
import redis
from redis.asyncio.cluster import (
ClusterNode as _AsyncClusterNode, # pyright: ignore[reportUnknownVariableType] # redis-py ships no resolvable stub for this class under the repo's current (stale) types-redis pin
)
from redis.asyncio.cluster import (
RedisCluster as _BaseAsyncRedisCluster, # pyright: ignore[reportUnknownVariableType] # same stale-stub gap as the import above
)
from redis.cluster import get_node_name
from redis.commands import READ_COMMANDS
from redis.exceptions import (
AskError,
BusyLoadingError,
ClusterDownError,
ClusterError,
MaxConnectionsError,
MovedError,
SlotNotCoveredError,
TryAgainError,
)
from redis.exceptions import ConnectionError as _RedisConnectionError
from redis.exceptions import TimeoutError as _RedisTimeoutError
node_class: Final = cluster_node_class if cluster_node_class is not None else _AsyncClusterNode
if hasattr(node_class, "update_active_connections_for_reconnect"):
verbose_logger.debug(
"redis-py %s recovers a node-level connection error per-connection upstream; "
"using the base RedisCluster without litellm's node-isolation override.",
redis.__version__,
)
return _BaseAsyncRedisCluster
if redis.__version__ not in _VERIFIED_REDIS_VERSIONS:
verbose_logger.warning(
"redis-py %s is not in the set this cluster-teardown-storm fix was verified "
"against (%s). The per-node-isolation override may not match the installed library's "
"real _execute_command behavior.",
redis.__version__,
sorted(_VERIFIED_REDIS_VERSIONS),
)
class LiteLLMAsyncRedisCluster(
_BaseAsyncRedisCluster # pyright: ignore[reportUntypedBaseClass] # same stale-stub gap as the import above; the base class itself is unresolvable, not this subclass's own code
):
async def _execute_command(
self,
target_node: _ClusterNodeAttrs,
*args: object,
**kwargs: object, # kwargs-ok: overrides redis-py's own **kwargs signature; the keyword contract is defined by the Redis command being dispatched, not by this method
) -> object:
cluster: _ClusterAttrs = self
node = target_node
asking = moved = False
redirect_addr: str | None = None
ttl = cluster.RedisClusterRequestTTL
while ttl > 0:
ttl -= 1
try:
if asking:
assert redirect_addr is not None
node = cluster.get_node(node_name=redirect_addr)
await node.execute_command("ASKING")
asking = False
elif moved:
slot = await cluster._determine_slot(*args) # pyright: ignore[reportPrivateUsage] # mirrors upstream's own un-overridden branch, which makes this identical private call from the same subclass
node = cluster.nodes_manager.get_node_from_slot(
slot,
cluster.read_from_replicas and args[0] in READ_COMMANDS,
(cluster.load_balancing_strategy if args[0] in READ_COMMANDS else None),
)
moved = False
return await node.execute_command(*args, **kwargs)
except (BusyLoadingError, MaxConnectionsError):
raise
except (_RedisConnectionError, _RedisTimeoutError):
# Reset only the node that actually failed instead of the upstream
# default (`await self.aclose()`, a full-cluster teardown that forces
# every other concurrent caller through the shared reinit lock).
await node.disconnect()
raise
except (ClusterDownError, SlotNotCoveredError):
await cluster.aclose()
await asyncio.sleep(0.25)
raise
except MovedError as e:
cluster.reinitialize_counter += 1
if cluster.reinitialize_steps and cluster.reinitialize_counter % cluster.reinitialize_steps == 0:
await cluster.aclose()
cluster.reinitialize_counter = 0
else:
cluster.nodes_manager._moved_exception = e # pyright: ignore[reportPrivateUsage] # mirrors upstream's own un-overridden branch; redis-py exposes no public setter for this
moved = True
except AskError as e:
redirect_addr = get_node_name(host=e.host, port=e.port)
asking = True
except TryAgainError:
if ttl < cluster.RedisClusterRequestTTL / 2:
await asyncio.sleep(0.05)
raise ClusterError("TTL exhausted.")
return LiteLLMAsyncRedisCluster

85
litellm/caching/valkey_semantic_cache.py
View file

@ -17,6 +17,7 @@ RedisSemanticCache since those are backend agnostic.
import asyncio
import hashlib
import os
from collections.abc import Mapping, Sequence
from dataclasses import dataclass
from typing import Any, Final
@ -64,7 +65,7 @@ class ValkeySemanticCache(RedisSemanticCache):
async_client: AsyncRedis | None = None,
embedding_max_input_tokens: int | None = None,
embedding_timeout: float | None = None,
**kwargs: Any,
**kwargs: object,
):
if similarity_threshold is None:
raise ValueError("similarity_threshold must be provided, passed None")
@ -87,11 +88,13 @@ class ValkeySemanticCache(RedisSemanticCache):
self.key_prefix = f"{self.index_name}:"
self._index_dim: int | None = None
resolved_url = None
if sync_client is None or async_client is None:
resolved_url = redis_url or self._build_valkey_url(host, port, password, ssl)
self.sync_client = sync_client if sync_client is not None else Redis.from_url(resolved_url)
self.async_client = async_client if async_client is not None else AsyncRedis.from_url(resolved_url)
if sync_client is not None and async_client is not None:
self.sync_client = sync_client
self.async_client = async_client
else:
resolved_url: Final = redis_url or self._build_valkey_url(host, port, password, ssl)
self.sync_client = sync_client if sync_client is not None else Redis.from_url(resolved_url)
self.async_client = async_client if async_client is not None else AsyncRedis.from_url(resolved_url)
print_verbose(f"Valkey semantic-cache initializing index - {self.index_name}")
@ -118,7 +121,7 @@ class ValkeySemanticCache(RedisSemanticCache):
return hashlib.sha256(str(key).encode("utf-8")).hexdigest()
@staticmethod
def _embedding_to_bytes(embedding: list[float]) -> bytes:
def _embedding_to_bytes(embedding: Sequence[float]) -> bytes:
return pack_vector(embedding)
def _index_schema(self, dim: int) -> tuple[TagField, VectorField]:
@ -192,7 +195,9 @@ class ValkeySemanticCache(RedisSemanticCache):
def _doc_key(self, key: str) -> str:
return f"{self.key_prefix}{self._scope_tag(key)}:{uuid.uuid4()}"
def _doc_mapping(self, key: str, prompt: str, value_str: str, embedding: list[float]) -> dict:
def _doc_mapping(
self, key: str, prompt: str, value_str: str, embedding: Sequence[float]
) -> Mapping[str | bytes, str | bytes]:
return {
self.CACHE_KEY_FIELD_NAME: self._scope_tag(key),
self.PROMPT_FIELD_NAME: prompt,
@ -208,30 +213,49 @@ class ValkeySemanticCache(RedisSemanticCache):
)
return Query(query_string).return_fields(self.RESPONSE_FIELD_NAME, self.DISTANCE_FIELD_NAME).dialect(2)
async def _async_search(self, key: str, embedding: Sequence[float]) -> object:
"""Run the KNN query on the async client, stopping the untyped search surface here."""
return await self.async_client.ft(self.index_name).search(
self._knn_query(key),
query_params={"vec": self._embedding_to_bytes(embedding)}, # pyright: ignore[reportArgumentType] # redis stubs omit bytes; KNN vectors are raw bytes at runtime
)
@classmethod
def _first_hit(cls, search_result: Any) -> _ValkeyCacheHit | None:
docs: Final = getattr(search_result, "docs", [])
def _first_hit(cls, search_result: object) -> _ValkeyCacheHit | None:
docs: Final[Sequence[object]] = getattr(search_result, "docs", [])
if not docs:
return None
doc: Final = docs[0]
response_field: Final[object] = getattr(doc, cls.RESPONSE_FIELD_NAME)
distance_field: Final[str | bytes | float] = getattr(doc, cls.DISTANCE_FIELD_NAME)
return _ValkeyCacheHit(
response=str(getattr(doc, cls.RESPONSE_FIELD_NAME)),
distance=float(getattr(doc, cls.DISTANCE_FIELD_NAME)),
response=str(response_field),
distance=float(distance_field),
)
def _resolve_hit(self, hit: _ValkeyCacheHit | None, key: str, **kwargs: Any) -> Any:
@staticmethod
def _record_similarity(kwargs: dict[str, Any], similarity: float) -> None:
"""Stamp the semantic-similarity score onto the request metadata carried in ``kwargs``."""
kwargs.setdefault("metadata", {})["semantic-similarity"] = similarity
@staticmethod
def _embedding_metadata(kwargs: dict[str, Any]) -> dict[str, Any] | None:
"""The request metadata forwarded to the embedding call."""
return kwargs.get("metadata")
def _resolve_hit(self, hit: _ValkeyCacheHit | None, key: str, **kwargs: object) -> object:
if hit is None:
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
self._record_similarity(kwargs, 0.0)
return None
similarity: Final = 1 - hit.distance
kwargs.setdefault("metadata", {})["semantic-similarity"] = similarity
self._record_similarity(kwargs, similarity)
if similarity < self.similarity_threshold:
return None
return self._get_cache_logic(cached_response=hit.response)
def set_cache(self, key: str, value: Any, **kwargs: Any) -> None:
def set_cache(self, key: str, value: object, **kwargs: object) -> None:
print_verbose(f"Valkey semantic-cache set_cache, kwargs: {kwargs}")
try:
prompt: Final = self._get_prompt_from_kwargs(**kwargs)
@ -250,12 +274,12 @@ class ValkeySemanticCache(RedisSemanticCache):
except Exception as e:
print_verbose(f"Error in Valkey semantic-cache set_cache: {e}")
def get_cache(self, key: str, **kwargs: Any) -> Any:
def get_cache(self, key: str, **kwargs: object) -> object:
print_verbose(f"Valkey semantic-cache get_cache, kwargs: {kwargs}")
try:
prompt: Final = self._get_prompt_from_kwargs(**kwargs)
if prompt is None:
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
self._record_similarity(kwargs, 0.0)
return None
embedding: Final = self._get_embedding(prompt)
@ -263,14 +287,14 @@ class ValkeySemanticCache(RedisSemanticCache):
search_result: Final = self.sync_client.ft(self.index_name).search(
self._knn_query(key),
query_params={"vec": self._embedding_to_bytes(embedding)},
query_params={"vec": self._embedding_to_bytes(embedding)}, # pyright: ignore[reportArgumentType] # redis stubs omit bytes; KNN vectors are raw bytes at runtime
)
return self._resolve_hit(self._first_hit(search_result), key, **kwargs)
except Exception as e:
print_verbose(f"Error in Valkey semantic-cache get_cache: {e}")
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
self._record_similarity(kwargs, 0.0)
async def async_set_cache(self, key: str, value: Any, **kwargs: Any) -> None:
async def async_set_cache(self, key: str, value: object, **kwargs: object) -> None:
print_verbose(f"Async Valkey semantic-cache set_cache, kwargs: {kwargs}")
try:
prompt: Final = self._get_prompt_from_kwargs(**kwargs)
@ -278,7 +302,7 @@ class ValkeySemanticCache(RedisSemanticCache):
print_verbose("No prompt provided for semantic caching")
return
embedding: Final = await self._get_async_embedding(prompt, metadata=kwargs.get("metadata"))
embedding: Final = await self._get_async_embedding(prompt, metadata=self._embedding_metadata(kwargs))
await self._ensure_index_async(len(embedding))
doc_key: Final = self._doc_key(key)
@ -289,31 +313,28 @@ class ValkeySemanticCache(RedisSemanticCache):
except Exception as e:
print_verbose(f"Error in async Valkey semantic-cache set_cache: {e}")
async def async_get_cache(self, key: str, **kwargs: Any) -> Any:
async def async_get_cache(self, key: str, **kwargs: object) -> object:
print_verbose(f"Async Valkey semantic-cache get_cache, kwargs: {kwargs}")
try:
prompt: Final = self._get_prompt_from_kwargs(**kwargs)
if prompt is None:
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
self._record_similarity(kwargs, 0.0)
return None
embedding: Final = await self._get_async_embedding(prompt, metadata=kwargs.get("metadata"))
embedding: Final = await self._get_async_embedding(prompt, metadata=self._embedding_metadata(kwargs))
await self._ensure_index_async(len(embedding))
search_result: Final = await self.async_client.ft(self.index_name).search(
self._knn_query(key),
query_params={"vec": self._embedding_to_bytes(embedding)},
)
search_result: Final[object] = await self._async_search(key, embedding)
return self._resolve_hit(self._first_hit(search_result), key, **kwargs)
except Exception as e:
print_verbose(f"Error in async Valkey semantic-cache get_cache: {e}")
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
self._record_similarity(kwargs, 0.0)
async def async_set_cache_pipeline(self, cache_list: list[tuple[str, Any]], **kwargs: Any) -> None:
async def async_set_cache_pipeline(self, cache_list: list[tuple[str, object]], **kwargs: object) -> None:
try:
await asyncio.gather(*[self.async_set_cache(key, value, **kwargs) for key, value in cache_list])
except Exception as e:
print_verbose(f"Error in Valkey semantic-cache async_set_cache_pipeline: {e}")
async def _index_info(self) -> dict:
async def _index_info(self) -> Mapping[str, object]:
return await self.async_client.ft(self.index_name).info()

14
litellm/completion_extras/litellm_responses_transformation/handler.py
View file

@ -45,14 +45,14 @@ class ResponsesToCompletionBridgeHandler:
return bool(stream)
@staticmethod
def _is_preformatted_cached_chat_stream(result: Any) -> bool:
def _is_preformatted_cached_chat_stream(result: object) -> bool:
from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper
return isinstance(result, CustomStreamWrapper) and result.custom_llm_provider == "cached_response"
@staticmethod
def _coerce_response_object(
response_obj: Any,
response_obj: object,
hidden_params: dict | None,
) -> "ResponsesAPIResponse":
if isinstance(response_obj, ResponsesAPIResponse):
@ -78,8 +78,8 @@ class ResponsesToCompletionBridgeHandler:
for _ in stream_iter:
pass
completed: Final = getattr(stream_iter, "completed_response", None)
response_obj: Final = getattr(completed, "response", None) if completed else None
completed: Final[object] = getattr(stream_iter, "completed_response", None)
response_obj: Final[object] = getattr(completed, "response", None) if completed else None
if response_obj is None:
raise ValueError("Stream ended without a completed response")
@ -93,8 +93,8 @@ class ResponsesToCompletionBridgeHandler:
async for _ in stream_iter:
pass
completed: Final = getattr(stream_iter, "completed_response", None)
response_obj: Final = getattr(completed, "response", None) if completed else None
completed: Final[object] = getattr(stream_iter, "completed_response", None)
response_obj: Final[object] = getattr(completed, "response", None) if completed else None
if response_obj is None:
raise ValueError("Stream ended without a completed response")
@ -157,7 +157,7 @@ class ResponsesToCompletionBridgeHandler:
def completion(
self, *args, **kwargs
) -> Union[
Coroutine[Any, Any, Union["ModelResponse", "CustomStreamWrapper"]],
Coroutine[None, None, Union["ModelResponse", "CustomStreamWrapper"]],
"ModelResponse",
"CustomStreamWrapper",
]:

101
litellm/completion_extras/litellm_responses_transformation/transformation.py
View file

@ -5,7 +5,7 @@ Handler for transforming /chat/completions api requests to litellm.responses req
import json
import os
from collections.abc import AsyncIterator, Callable, Iterable, Iterator, Mapping, Sequence
from typing import TYPE_CHECKING, Any, Final, Literal, TypedDict, Union, cast
from typing import TYPE_CHECKING, Any, Final, Literal, TypedDict, Union, cast, get_args
from openai.types.responses.custom_tool_param import CustomToolParam
from openai.types.responses.response_input_param import (
@ -21,6 +21,9 @@ from pydantic import BaseModel
import litellm
from litellm import ModelResponse
from litellm._logging import verbose_logger
from litellm.litellm_core_utils.prompt_templates.common_utils import (
responses_reasoning_item_from_thinking_blocks,
)
from litellm.llms.base_llm.base_model_iterator import BaseModelResponseIterator
from litellm.llms.base_llm.bridges.completion_transformation import (
CompletionTransformationBridge,
@ -32,6 +35,7 @@ from litellm.responses.sse_output_recovery import (
)
from litellm.responses.utils import normalize_responses_api_stream_options
from litellm.types.llms.openai import (
REASONING_EFFORT,
ChatCompletionAnnotation,
ChatCompletionReasoningItem,
ChatCompletionToolCallChunk,
@ -55,9 +59,11 @@ if TYPE_CHECKING:
from litellm.types.llms.openai import (
ALL_RESPONSES_API_TOOL_PARAMS,
AllMessageValues,
ChatCompletionFileObject,
ChatCompletionImageObject,
ChatCompletionRedactedThinkingBlock,
ChatCompletionThinkingBlock,
ChatCompletionToolReferenceObject,
OpenAIMessageContentListBlock,
)
from litellm.types.utils import Choices
@ -85,6 +91,22 @@ def _get_reasoning_items(
return []
def _reasoning_input_items(msg: "AllMessageValues") -> list[dict[str, object]]: # mutable-ok: API message payload
"""Reasoning input items for an assistant message.
Stored reasoning items win because they carry an id the Responses API minted; thinking
blocks are the fallback for turns that arrived over another API surface.
"""
items: Final = _get_reasoning_items(msg)
stored: Final = [_reasoning_item_to_response_input(item) for item in items] # mutable-ok: API message payload
if stored:
return stored
raw_blocks: Final = msg.get("thinking_blocks") or ()
blocks: Final = cast("Iterable[ChatCompletionThinkingBlock]", raw_blocks) # cast-ok: untyped client json
from_thinking: Final = responses_reasoning_item_from_thinking_blocks(blocks)
return [] if from_thinking is None else [dict(from_thinking)] # mutable-ok: API message payload
def _build_reasoning_item(
item_id: str,
encrypted_content: str | None,
@ -155,6 +177,16 @@ def _map_incomplete_reason_to_finish_reason(incomplete_reason: str | None) -> Li
return "length"
def _input_file_from_file_value(file_value: object) -> dict[str, object]:
if not isinstance(file_value, dict):
return {"type": "input_file"}
file_dict: Final = cast("dict[str, object]", file_value) # cast-ok: runtime dict checked
return {
"type": "input_file",
**{key: file_dict[key] for key in ("file_id", "file_data", "filename") if key in file_dict},
}
def _incomplete_reason_from_response_payload(response_payload: object) -> str | None:
if not isinstance(response_payload, Mapping):
return None
@ -180,7 +212,8 @@ def _tool_call_dict_from_output_item(item: Mapping[str, Any], index: int) -> _Ch
LiteLLMCompletionResponsesConfig,
)
is_custom: Final = item.get("type") == "custom_tool_call"
item_type: Final[object] = item.get("type")
is_custom: Final = item_type == "custom_tool_call"
arguments: Final = (item.get("input") if is_custom else item.get("arguments")) or ""
name: Final = item.get("name") or ("custom_tool" if is_custom else "")
function_chunk: Final = ChatCompletionToolCallFunctionChunk(name=name, arguments=arguments)
@ -190,7 +223,7 @@ def _tool_call_dict_from_output_item(item: Mapping[str, Any], index: int) -> _Ch
function=function_chunk,
index=index,
)
raw_provider_fields: Final = item.get("provider_specific_fields")
raw_provider_fields: Final[object] = item.get("provider_specific_fields")
if isinstance(raw_provider_fields, dict):
provider_specific_fields = raw_provider_fields
elif raw_provider_fields and hasattr(raw_provider_fields, "__dict__"):
@ -372,8 +405,15 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
)
)
elif role == "assistant" and tool_calls and isinstance(tool_calls, list):
for r_item in _get_reasoning_items(msg):
input_items.append(_reasoning_item_to_response_input(r_item))
input_items.extend(_reasoning_input_items(msg))
if content:
input_items.append(
{ # mutable-ok: API message payload
"type": "message",
"role": "assistant",
"content": self._convert_content_to_responses_format(content, "assistant"),
}
)
for tool_call in tool_calls:
function = tool_call.get("function")
custom = tool_call.get("custom")
@ -400,15 +440,16 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
raise ValueError(f"tool call not supported: {tool_call}")
elif content is not None:
if role == "assistant":
for r_item in _get_reasoning_items(msg):
input_items.append(_reasoning_item_to_response_input(r_item))
input_items.extend(_reasoning_input_items(msg))
input_items.append(
{
{ # mutable-ok: API message payload
"type": "message",
"role": role,
"content": self._convert_content_to_responses_format(content, cast(str, role)),
}
)
elif role == "assistant":
input_items.extend(_reasoning_input_items(msg))
return input_items, instructions
@ -467,7 +508,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
def _merge_responses_api_request_into_request_data(
self,
request_data: dict[str, Any],
request_data: dict[str, object],
responses_api_request: "ResponsesAPIOptionalRequestParams",
instructions: str | None,
) -> None:
@ -929,7 +970,12 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
content: str
| list[object]
| Iterable[
Union["OpenAIMessageContentListBlock", "ChatCompletionThinkingBlock", "ChatCompletionRedactedThinkingBlock"]
Union[
"OpenAIMessageContentListBlock",
"ChatCompletionThinkingBlock",
"ChatCompletionRedactedThinkingBlock",
"ChatCompletionToolReferenceObject",
]
]
| None,
role: str,
@ -978,17 +1024,15 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
result.append(converted)
verbose_logger.debug("Chat provider: image -> %s", converted)
elif item_type == "file":
# Map Chat Completion file to Responses API input_file
# {"type": "file", "file": {"file_data": "...", "filename": "..."}}
# -> {"type": "input_file", "file_data": "...", "filename": "..."}
file_data = item.get("file", {})
converted = {"type": "input_file"}
if isinstance(file_data, dict):
for key in ["file_id", "file_data", "filename"]:
if key in file_data:
converted[key] = file_data[key]
converted = _input_file_from_file_value(
cast("ChatCompletionFileObject", item).get("file"), # cast-ok: type tag checked
)
result.append(converted)
verbose_logger.debug("Chat provider: file -> %s", converted)
elif item_type == "tool_reference":
verbose_logger.debug(
"Chat provider: tool_reference has no responses API equivalent; skipped"
)
elif item_type in [
"input_text",
"input_image",
@ -1086,22 +1130,11 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
litellm.reasoning_auto_summary or os.getenv("LITELLM_REASONING_AUTO_SUMMARY", "false").lower() == "true"
)
# If string is passed, map with optional summary based on flag/env var
if reasoning_effort == "none":
return Reasoning(effort="none", summary="detailed") if auto_summary_enabled else Reasoning(effort="none")
elif reasoning_effort == "high":
return Reasoning(effort="high", summary="detailed") if auto_summary_enabled else Reasoning(effort="high")
elif reasoning_effort == "xhigh":
return Reasoning(effort="xhigh", summary="detailed") if auto_summary_enabled else Reasoning(effort="xhigh")
elif reasoning_effort == "medium":
if reasoning_effort in get_args(REASONING_EFFORT):
return (
Reasoning(effort="medium", summary="detailed") if auto_summary_enabled else Reasoning(effort="medium")
)
elif reasoning_effort == "low":
return Reasoning(effort="low", summary="detailed") if auto_summary_enabled else Reasoning(effort="low")
elif reasoning_effort == "minimal":
return (
Reasoning(effort="minimal", summary="detailed") if auto_summary_enabled else Reasoning(effort="minimal")
Reasoning(effort=reasoning_effort, summary="detailed")
if auto_summary_enabled
else Reasoning(effort=reasoning_effort)
)
return None

131
litellm/constants.py
View file

@ -13,6 +13,12 @@ DEFAULT_BATCH_SIZE: Final = int(os.getenv("DEFAULT_BATCH_SIZE", 512))
DEFAULT_FLUSH_INTERVAL_SECONDS: Final = int(os.getenv("DEFAULT_FLUSH_INTERVAL_SECONDS", 5))
DEFAULT_S3_FLUSH_INTERVAL_SECONDS: Final = int(os.getenv("DEFAULT_S3_FLUSH_INTERVAL_SECONDS", 10))
DEFAULT_S3_BATCH_SIZE: Final = int(os.getenv("DEFAULT_S3_BATCH_SIZE", 512))
# https://docs.aws.amazon.com/AmazonS3/latest/userguide/object-keys.html
MAX_S3_OBJECT_KEY_BYTES: Final = 1024
S3_BOUNDED_OBJECT_KEY_HEAD_BYTES: Final = 64
S3_PREFIX_DIGEST_CHARS: Final = 16
# s3 allows 2048 bytes of combined metadata headers, which Content-Disposition counts against
MAX_S3_OBJECT_DOWNLOAD_FILENAME_BYTES: Final = 1024
DEFAULT_SQS_FLUSH_INTERVAL_SECONDS: Final = int(os.getenv("DEFAULT_SQS_FLUSH_INTERVAL_SECONDS", 10))
DEFAULT_NUM_WORKERS_LITELLM_PROXY: Final = int(os.getenv("DEFAULT_NUM_WORKERS_LITELLM_PROXY", 1))
DYNAMIC_RATE_LIMIT_ERROR_THRESHOLD_PER_MINUTE = int(os.getenv("DYNAMIC_RATE_LIMIT_ERROR_THRESHOLD_PER_MINUTE", 1))
@ -35,6 +41,7 @@ DEFAULT_COOLDOWN_TIME_SECONDS: Final = int(os.getenv("DEFAULT_COOLDOWN_TIME_SECO
DEFAULT_REPLICATE_POLLING_RETRIES: Final = int(os.getenv("DEFAULT_REPLICATE_POLLING_RETRIES", 5))
DEFAULT_REPLICATE_POLLING_DELAY_SECONDS: Final = int(os.getenv("DEFAULT_REPLICATE_POLLING_DELAY_SECONDS", 1))
DEFAULT_IMAGE_TOKEN_COUNT: Final = int(os.getenv("DEFAULT_IMAGE_TOKEN_COUNT", 250))
HF_CONFIG_FETCH_TIMEOUT_SECONDS: Final = 10.0
# Maximum wall-clock seconds a streaming response is allowed to run.
# Streams exceeding this duration are terminated with a Timeout error.
@ -48,6 +55,9 @@ LITELLM_MAX_STREAMING_DURATION_SECONDS: Final = (
# Data URIs exceeding this are replaced with a size placeholder.
# Set to 0 to disable truncation.
MAX_BASE64_LENGTH_FOR_LOGGING: Final = int(os.getenv("MAX_BASE64_LENGTH_FOR_LOGGING", 64))
REDACTED_BY_LITELLM: Final = "redacted-by-litellm"
# in-memory stand-in handed to provider converters for redacted arguments; never stored
REDACTED_TOOL_CALL_ARGUMENTS_PLACEHOLDER: Final = "{}"
MAX_STRING_LENGTH_STDOUT_LOG: Final = get_env_int("MAX_STRING_LENGTH_STDOUT_LOG", 4096)
@ -126,6 +136,7 @@ MCP_CLIENT_TIMEOUT: Final = float(os.getenv("LITELLM_MCP_CLIENT_TIMEOUT", "60.0"
MCP_TOOL_LISTING_TIMEOUT: Final = float(os.getenv("LITELLM_MCP_TOOL_LISTING_TIMEOUT", "30.0"))
MCP_METADATA_TIMEOUT: Final = float(os.getenv("LITELLM_MCP_METADATA_TIMEOUT", "10.0"))
MCP_HEALTH_CHECK_TIMEOUT: Final = float(os.getenv("LITELLM_MCP_HEALTH_CHECK_TIMEOUT", "10.0"))
MCP_TOOL_LISTING_MAX_PAGES: Final = 1000
# Allowlist of commands permitted for MCP stdio transport.
# Prevents arbitrary command execution via /mcp-rest/test/* endpoints or server creation.
@ -146,6 +157,7 @@ LITELLM_UI_ALLOW_HEADERS: Final = [
"x-litellm-adaptive-router-model",
"x-litellm-applied-guardrails",
"x-litellm-guardrail-scan-id",
"x-litellm-cache-key",
]
# Gemini model-specific minimal thinking budget constants
@ -284,6 +296,7 @@ REDIS_DAILY_ORG_SPEND_UPDATE_BUFFER_KEY: Final = "litellm_daily_org_spend_update
REDIS_DAILY_END_USER_SPEND_UPDATE_BUFFER_KEY: Final = "litellm_daily_end_user_spend_update_buffer"
REDIS_DAILY_AGENT_SPEND_UPDATE_BUFFER_KEY: Final = "litellm_daily_agent_spend_update_buffer"
REDIS_DAILY_TAG_SPEND_UPDATE_BUFFER_KEY: Final = "litellm_daily_tag_spend_update_buffer"
REDIS_WINDOW_SPEND_UPDATE_BUFFER_KEY: Final = "litellm_window_spend_update_buffer"
MAX_REDIS_BUFFER_DEQUEUE_COUNT: Final = int(os.getenv("MAX_REDIS_BUFFER_DEQUEUE_COUNT", 100))
# Bounds asyncio.Queue() instances (log queues, spend update queues, etc.) to prevent unbounded memory growth
LITELLM_ASYNCIO_QUEUE_MAXSIZE: Final = int(os.getenv("LITELLM_ASYNCIO_QUEUE_MAXSIZE", 1000))
@ -292,6 +305,9 @@ GUARDRAIL_SCANNED_MESSAGES_CACHE_TTL_SECONDS: Final = int(
os.getenv("GUARDRAIL_SCANNED_MESSAGES_CACHE_TTL_SECONDS", 24 * 60 * 60)
)
BEDROCK_APPLY_GUARDRAIL_CHUNK_BUDGET_CHARS: Final = 25_000
DEFAULT_PRESIDIO_ANALYZE_CHUNK_SIZE_BYTES: Final = 500_000
PRESIDIO_ANALYZE_CHUNK_OVERLAP_CHARS: Final = 4096
PRESIDIO_ANALYZE_CHUNK_CONCURRENCY: Final = 8
# Aggregation threshold: default to 80% of the asyncio queue maxsize so the check can always trigger.
# Must be < LITELLM_ASYNCIO_QUEUE_MAXSIZE; if set higher the aggregation logic will never fire.
MAX_SIZE_IN_MEMORY_QUEUE: Final = int(os.getenv("MAX_SIZE_IN_MEMORY_QUEUE", int(LITELLM_ASYNCIO_QUEUE_MAXSIZE * 0.8)))
@ -377,6 +393,7 @@ AZURE_OPERATION_POLLING_TIMEOUT: Final = int(os.getenv("AZURE_OPERATION_POLLING_
AZURE_DOCUMENT_INTELLIGENCE_API_VERSION: Final = str(os.getenv("AZURE_DOCUMENT_INTELLIGENCE_API_VERSION", "2024-11-30"))
AZURE_DOCUMENT_INTELLIGENCE_DEFAULT_DPI: Final = int(os.getenv("AZURE_DOCUMENT_INTELLIGENCE_DEFAULT_DPI", 96))
REDIS_SOCKET_TIMEOUT: Final = float(os.getenv("REDIS_SOCKET_TIMEOUT", 0.1))
CACHE_WRITE_SHUTDOWN_FLUSH_TIMEOUT_SECONDS: Final[float] = 5.0
REDIS_CONNECTION_POOL_TIMEOUT: Final = int(os.getenv("REDIS_CONNECTION_POOL_TIMEOUT", 5))
REDIS_CIRCUIT_BREAKER_FAILURE_THRESHOLD: Final = int(os.getenv("REDIS_CIRCUIT_BREAKER_FAILURE_THRESHOLD", 5))
REDIS_CIRCUIT_BREAKER_RECOVERY_TIMEOUT: Final = int(os.getenv("REDIS_CIRCUIT_BREAKER_RECOVERY_TIMEOUT", 60))
@ -460,6 +477,8 @@ CONNECTION_ERROR_PATTERNS: Final[list[str]] = [
]
STREAM_SSE_DONE_STRING: Final[str] = "[DONE]"
STREAM_SSE_DATA_PREFIX: Final[str] = "data: "
STREAM_SSE_KEEPALIVE_PING_CHUNK: Final[str] = 'event: ping\ndata: {"type": "ping"}\n\n'
STREAM_SSE_KEEPALIVE_PING_BYTES: Final[bytes] = STREAM_SSE_KEEPALIVE_PING_CHUNK.encode("utf-8")
### SPEND TRACKING ###
DEFAULT_REPLICATE_GPU_PRICE_PER_SECOND: Final = float(
os.getenv("DEFAULT_REPLICATE_GPU_PRICE_PER_SECOND", 0.001400)
@ -472,6 +491,22 @@ FIREWORKS_AI_80_B: Final = int(os.getenv("FIREWORKS_AI_80_B", 80))
#### Logging callback constants ####
REDACTED_BY_LITELM_STRING: Final = "REDACTED_BY_LITELM"
MAX_LANGFUSE_INITIALIZED_CLIENTS: Final = int(os.getenv("MAX_LANGFUSE_INITIALIZED_CLIENTS", 50))
# Backpressure + lifetime bounds for the /v1/messages streaming relay (see
# BaseAnthropicMessagesStreamingIterator.async_sse_wrapper). The relay queue is
# bounded so a slow client throttles the upstream pump instead of letting it
# buffer the whole response in memory; the detached-drain cap bounds how many
# post-disconnect drains may run concurrently so client behavior can't create
# unbounded worker state.
ANTHROPIC_MESSAGES_STREAM_RELAY_QUEUE_MAXSIZE: Final = int(
os.getenv("ANTHROPIC_MESSAGES_STREAM_RELAY_QUEUE_MAXSIZE", "1024")
)
# Setting this to 0 disables detached draining entirely: every post-disconnect
# pump bills whatever partial output it has already collected and aborts the
# upstream stream immediately, instead of continuing to drain for the real
# terminal usage.
ANTHROPIC_MESSAGES_MAX_DETACHED_STREAM_DRAINS: Final = int(
os.getenv("ANTHROPIC_MESSAGES_MAX_DETACHED_STREAM_DRAINS", "100")
)
LOGGING_WORKER_CONCURRENCY: Final = int(os.getenv("LOGGING_WORKER_CONCURRENCY", 100)) # Must be above 0
LOGGING_WORKER_MAX_QUEUE_SIZE: Final = int(os.getenv("LOGGING_WORKER_MAX_QUEUE_SIZE", 50_000))
LOGGING_WORKER_MAX_TIME_PER_COROUTINE: Final = float(os.getenv("LOGGING_WORKER_MAX_TIME_PER_COROUTINE", 20.0))
@ -602,6 +637,8 @@ LITELLM_CHAT_PROVIDERS: Final = [
"nscale",
"nebius",
"dashscope",
"qwencloud",
"qwen_ai_platform",
"modelscope",
"moonshot",
"publicai",
@ -619,6 +656,15 @@ LITELLM_CHAT_PROVIDERS: Final = [
"amazon_nova",
]
# Resolving these providers runs an OAuth device flow (their provider info IS the login), so any
# metadata or capability lookup against them can block for minutes waiting on a human.
PROVIDERS_THAT_AUTHENTICATE_ON_PROVIDER_INFO: Final = frozenset(
{
"github_copilot",
"chatgpt",
}
)
LITELLM_EMBEDDING_PROVIDERS_SUPPORTING_INPUT_ARRAY_OF_TOKENS: Final = [
"openai",
"azure",
@ -749,6 +795,7 @@ openai_compatible_endpoints: Final[list] = [
"api.groq.com/openai/v1",
"https://integrate.api.nvidia.com/v1",
"api.deepseek.com/v1",
"api.together.ai/v1",
"api.together.xyz/v1",
"app.empower.dev/api/v1",
"https://api.friendli.ai/serverless/v1",
@ -761,6 +808,7 @@ openai_compatible_endpoints: Final[list] = [
"inference.api.nscale.com/v1",
"api.studio.nebius.ai/v1",
"https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
"https://dashscope.aliyuncs.com/compatible-mode/v1",
"https://api-inference.modelscope.cn/v1",
"https://api.moonshot.ai/v1",
"https://api.publicai.co/v1",
@ -783,6 +831,8 @@ openai_compatible_endpoints: Final[list] = [
"https://pinstripes.io/v1",
"https://api.meta.ai/v1",
"https://api.cognition.ai/v1",
"https://api.scx.ai/v1",
"https://gigachat.devices.sberbank.ru/api/v1",
]
@ -832,6 +882,8 @@ openai_compatible_providers: Final[list] = [
"nscale",
"nebius",
"dashscope",
"qwencloud",
"qwen_ai_platform",
"modelscope",
"moonshot",
"v0",
@ -851,6 +903,7 @@ openai_compatible_providers: Final[list] = [
"darkbloom",
"meta", # Meta Model API (Muse Spark) - JSON-configured provider
"cognition",
"scx-ai",
]
openai_text_completion_compatible_providers: Final[list] = [ # providers that support `/v1/completions`
"together_ai",
@ -861,6 +914,8 @@ openai_text_completion_compatible_providers: Final[list] = [ # providers that s
"featherless_ai",
"nebius",
"dashscope",
"qwencloud",
"qwen_ai_platform",
"modelscope",
"moonshot",
"publicai",
@ -1068,7 +1123,7 @@ nebius_models: Final[set] = set(
]
)
dashscope_models: Final[set] = set(
dashscope_models: Final[frozenset] = frozenset(
[
"qwen-turbo",
"qwen-plus",
@ -1083,6 +1138,10 @@ dashscope_models: Final[set] = set(
]
)
qwencloud_models: Final[frozenset] = frozenset(dashscope_models)
qwen_ai_platform_models: Final[frozenset] = frozenset(dashscope_models)
nebius_embedding_models: Final[set] = set(
[
"BAAI/bge-en-icl",
@ -1199,6 +1258,7 @@ BEDROCK_CONVERSE_MODELS: Final = [
"openai.gpt-oss-120b-1:0",
"anthropic.claude-haiku-4-5-20251001-v1:0",
"anthropic.claude-sonnet-4-5-20250929-v1:0",
"anthropic.claude-fable-5-1",
"anthropic.claude-fable-5",
"anthropic.claude-sonnet-5",
"anthropic.claude-opus-5",
@ -1386,6 +1446,12 @@ DEFAULT_SOFT_BUDGET: Final = float(
) # by default all litellm proxy keys have a soft budget of 50.0
# makes it clear this is a rate limit error for a litellm virtual key
RATE_LIMIT_ERROR_MESSAGE_FOR_VIRTUAL_KEY: Final = "LiteLLM Virtual Key user_api_key_hash"
# Prefix of the 401 raised when a submitted virtual key is not shaped like one.
INVALID_VIRTUAL_KEY_ERROR_MESSAGE: Final = "LiteLLM Virtual Key expected"
# Attribute stamped on that 401 at its raise site so log routing recognises it by
# provenance. Message text is caller-influenceable on other 401s, so it must not
# be used to classify.
INVALID_VIRTUAL_KEY_ERROR_MARKER: Final = "_litellm_invalid_virtual_key_error"
# Python garbage collection threshold configuration
# Format: "gen0,gen1,gen2" e.g., "1000,50,50"
@ -1463,6 +1529,12 @@ LITELLM_PROXY_MASTER_KEY_ALIAS: Final = "litellm_proxy_master_key"
# ``ProxyLogging._handle_logging_proxy_only_error``.
LITELLM_LOGGING_NO_UPSTREAM_LLM_CALL: Final = "litellm_no_upstream_llm_call"
# Key/team metadata fields naming the OTel Resource ``service.name``, highest
# precedence first. Shared between the OTel v2 tenant router (which reads them
# out of ``user_api_key_auth_metadata``) and proxy request setup (which re-applies
# the key's values after the team metadata merge so a key outranks its team).
OTEL_SERVICE_NAME_METADATA_KEYS: Final = ("otel_service_name_override", "otel_service_name")
# Key Rotation Constants
LITELLM_KEY_ROTATION_ENABLED: Final = os.getenv("LITELLM_KEY_ROTATION_ENABLED", "false")
LITELLM_KEY_ROTATION_CHECK_INTERVAL_SECONDS: Final = int(
@ -1518,6 +1590,7 @@ KEY_ROTATION_JOB_NAME: Final = "litellm_key_rotation_job"
EXPIRED_UI_SESSION_KEY_CLEANUP_JOB_NAME: Final = "litellm_expired_ui_session_key_cleanup_job"
WEEKLY_SPEND_REPORT_JOB_ID: Final = "weekly_spend_report_job"
MONTHLY_SPEND_REPORT_JOB_ID: Final = "monthly_spend_report_job"
USER_SPEND_ALERTS_JOB_ID: Final = "user_spend_alerts_job"
PROMETHEUS_FALLBACK_STATS_JOB_ID: Final = "prometheus_fallback_stats_job"
SLACK_DAILY_REPORT_LOCK_ID: Final = "slack_daily_report"
SLACK_MODEL_DEPRECATION_LOCK_ID: Final = "slack_model_deprecation_warning"
@ -1534,9 +1607,12 @@ 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))
SPEND_LOG_WRITE_BATCH_MAX_BYTES: Final = max(1, int(os.getenv("SPEND_LOG_WRITE_BATCH_MAX_BYTES", 2_000_000)))
SPEND_LOG_WRITE_BATCH_MAX_ROWS: Final = max(1, int(os.getenv("SPEND_LOG_WRITE_BATCH_MAX_ROWS", "100")))
SPEND_LOG_QUEUE_SIZE_THRESHOLD: Final = int(os.getenv("SPEND_LOG_QUEUE_SIZE_THRESHOLD", 100))
SPEND_LOG_QUEUE_MAX_BYTES: Final = max(1, int(os.getenv("SPEND_LOG_QUEUE_MAX_BYTES", "64000000")))
SPEND_LOG_QUEUE_POLL_INTERVAL: Final = float(os.getenv("SPEND_LOG_QUEUE_POLL_INTERVAL", 2.0))
RESPONSES_SESSION_LOOKUP_MAX_ATTEMPTS: Final = max(1, int(os.getenv("RESPONSES_SESSION_LOOKUP_MAX_ATTEMPTS", "3")))
RESPONSES_SESSION_LOOKUP_RETRY_INTERVAL: Final = float(os.getenv("RESPONSES_SESSION_LOOKUP_RETRY_INTERVAL", "0.2"))
SPEND_COUNTER_RESEED_LOCKS_MAX_SIZE: Final = int(os.getenv("SPEND_COUNTER_RESEED_LOCKS_MAX_SIZE", 10000))
DEFAULT_CRON_JOB_LOCK_TTL_SECONDS: Final = int(os.getenv("DEFAULT_CRON_JOB_LOCK_TTL_SECONDS", 60)) # 1 minute
PROXY_BUDGET_RESCHEDULER_MIN_TIME: Final = int(os.getenv("PROXY_BUDGET_RESCHEDULER_MIN_TIME", 597))
@ -1556,6 +1632,19 @@ STALE_OBJECT_CLEANUP_BATCH_SIZE: Final = max(1, int(os.getenv("STALE_OBJECT_CLEA
# installations with large numbers of stale managed objects).
_batch_polling_env: Final = os.getenv("PROXY_BATCH_POLLING_ENABLED", "true").lower()
PROXY_BATCH_POLLING_ENABLED: Final = _batch_polling_env == "true"
BACKGROUND_INTERACTION_COST_POLL_INITIAL_INTERVAL_SECONDS: Final = float(
os.getenv("BACKGROUND_INTERACTION_COST_POLL_INITIAL_INTERVAL_SECONDS", "5")
)
BACKGROUND_INTERACTION_COST_POLL_MAX_INTERVAL_SECONDS: Final = float(
os.getenv("BACKGROUND_INTERACTION_COST_POLL_MAX_INTERVAL_SECONDS", "60")
)
BACKGROUND_INTERACTION_COST_POLL_TIMEOUT_SECONDS: Final = float(
os.getenv("BACKGROUND_INTERACTION_COST_POLL_TIMEOUT_SECONDS", "3600")
)
_background_interaction_cost_polling_env: Final = os.getenv(
"BACKGROUND_INTERACTION_COST_POLLING_ENABLED", "true"
).lower()
BACKGROUND_INTERACTION_COST_POLLING_ENABLED: Final = _background_interaction_cost_polling_env == "true"
PROXY_BUDGET_RESCHEDULER_MAX_TIME: Final = int(os.getenv("PROXY_BUDGET_RESCHEDULER_MAX_TIME", 605))
PROXY_BATCH_WRITE_AT: Final = int(os.getenv("PROXY_BATCH_WRITE_AT", 10)) # in seconds, increased from 10
PROXY_CONFIG_RELOAD_INTERVAL_SECONDS: Final = get_env_int("PROXY_CONFIG_RELOAD_INTERVAL_SECONDS", 30)
@ -1620,6 +1709,7 @@ LITELLM_SETTINGS_SAFE_DB_OVERRIDES: Final = [
"enable_anthropic_prompt_caching",
"anthropic_prompt_caching_ttl",
"max_ui_session_budget",
"budget_rollover",
]
SPECIAL_LITELLM_AUTH_TOKEN: Final = ["ui-token"]
DEFAULT_MANAGEMENT_OBJECT_IN_MEMORY_CACHE_TTL = int(os.getenv("DEFAULT_MANAGEMENT_OBJECT_IN_MEMORY_CACHE_TTL", 60))
@ -1636,6 +1726,7 @@ DEFAULT_MCP_NAMESPACE_CSV_MAX_TOKENS: Final = 16
# Ceilings on the cached auth registries; larger tables fall back to per-row lookups
# instead of holding an unbounded id set in every worker.
TAG_REGISTRY_MAX_SIZE: Final = 5000
MODEL_ACCESS_GROUP_REGISTRY_MAX_SIZE: Final = 5000
END_USER_RESTRICTED_REGISTRY_MAX_SIZE: Final = 5000
# How long a failed registry load is remembered as "unusable", so a degraded Postgres
# is not re-scanned on every request on top of the per-id lookups it falls back to.
@ -1680,6 +1771,7 @@ SENTRY_DENYLIST: Final = [
"jwt_token",
"private_key",
"SLACK_WEBHOOK_URL",
"ALERTING_WEBHOOK_URL",
"webhook_url",
"LANGFUSE_SECRET_KEY",
# Email Configuration
@ -1786,6 +1878,43 @@ BROWSER_SECURITY_HEADERS: Final[frozenset[str]] = frozenset(
UNSAFE_PROXY_RESPONSE_HEADERS: Final[frozenset[str]] = HTTP_FRAMING_HEADERS | BROWSER_SECURITY_HEADERS
# A retrieved response replays the usage of the call that created it, so pricing these
# read/management routes like inference bills the same tokens twice.
NON_INFERENCE_CALL_TYPES: Final[frozenset[str]] = frozenset(
{
"get_responses",
"aget_responses",
"delete_responses",
"adelete_responses",
"cancel_responses",
"acancel_responses",
"list_input_items",
"alist_input_items",
"vector_store_create",
"avector_store_create",
"vector_store_retrieve",
"avector_store_retrieve",
"vector_store_list",
"avector_store_list",
"vector_store_update",
"avector_store_update",
"vector_store_delete",
"avector_store_delete",
"vector_store_file_create",
"avector_store_file_create",
"vector_store_file_list",
"avector_store_file_list",
"vector_store_file_retrieve",
"avector_store_file_retrieve",
"vector_store_file_content",
"avector_store_file_content",
"vector_store_file_update",
"avector_store_file_update",
"vector_store_file_delete",
"avector_store_file_delete",
}
)
# PTU reservation rollup writes rows to LiteLLM_DailyTeamSpend with this
# sentinel api_key so PTU flat cost stays distinguishable from real per-request
# spend under the table's composite unique constraint.

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