scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-09-13 23:11:40 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 101

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

Browse source 
# Conflicts:
#	litellm/proxy/auth/auth_checks.py
#	litellm/proxy/hooks/model_max_budget_limiter.py
This commit is contained in:
ryan-crabbe-berri 2026-08-05 11:15:09 -07:00
commit a0eaf8f6a0
1884 changed files with 72273 additions and 46662 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

148
.circleci/config.yml
View file

@ -88,6 +88,59 @@ commands:
rm -f /tmp/uv-install.sh
echo 'export PATH="$HOME/.local/bin:$PATH"' >> "$BASH_ENV"
export PATH="$HOME/.local/bin:$PATH"
install_node:
description: "Install the Node.js version pinned in ui/litellm-dashboard/.nvmrc (24.19.0, which bundles npm 11.17.0) with checksum verification, and prepend it to PATH. Run this on any executor whose image does not already ship that version, or `npm ci` in ui/litellm-dashboard fails EBADENGINE against the engines floor. Installs into /opt/node rather than over /usr/local on purpose: cimg/python:*-browsers ships its own node there, and unpacking the tarball on top of it leaves npm 11.17 files merged with the image's npm 11.9 tree, which reports the new version and then exits 1 on `npm ci` with no error text at all. Requires checkout, which the .nvmrc drift check reads."
steps:
- run:
name: Install Node.js 24.19.0
command: |
NODE_VERSION="24.19.0"
NODE_TARBALL="node-v${NODE_VERSION}-linux-x64.tar.xz"
NODE_EXPECTED_SHA="14b342e71204f811bde6153be8e04b62aef63c236fef92b55f9c83154b409647"
NVMRC_VERSION="$(tr -d '[:space:]' < ui/litellm-dashboard/.nvmrc)"
if [ "$NVMRC_VERSION" != "$NODE_VERSION" ]; then
echo "install_node: ui/litellm-dashboard/.nvmrc pins ${NVMRC_VERSION} but this command pins ${NODE_VERSION}; update NODE_VERSION and NODE_EXPECTED_SHA together" >&2
exit 1
fi
curl -sSLf -o "/tmp/${NODE_TARBALL}" "https://nodejs.org/dist/v${NODE_VERSION}/${NODE_TARBALL}"
echo "${NODE_EXPECTED_SHA} /tmp/${NODE_TARBALL}" | sha256sum -c -
sudo mkdir -p /opt/node
sudo tar -xJf "/tmp/${NODE_TARBALL}" -C /opt/node --strip-components=1
rm -f "/tmp/${NODE_TARBALL}"
echo 'export PATH="/opt/node/bin:$PATH"' >> "$BASH_ENV"
export PATH="/opt/node/bin:$PATH"
node --version
npm --version
install_rust:
description: "Install pinned rustup (1.28.2) and Rust toolchain (1.97.1) with checksum verification. Adds ~/.cargo/bin to PATH. Run this before any `uv sync` or `uv build` of the workspace: the root package builds litellm-rust through maturin, and on an image without cargo maturin fetches an unpinned rustup and a floating toolchain by itself."
steps:
- run:
name: Install Rust (rustup 1.28.2, toolchain 1.97.1)
command: |
case "$(uname -m)" in
x86_64)
RUSTUP_TRIPLE=x86_64-unknown-linux-gnu
RUSTUP_SHA256=20a06e644b0d9bd2fbdbfd52d42540bdde820ea7df86e92e533c073da0cdd43c
;;
aarch64)
RUSTUP_TRIPLE=aarch64-unknown-linux-gnu
RUSTUP_SHA256=e3853c5a252fca15252d07cb23a1bdd9377a8c6f3efa01531109281ae47f841c
;;
*)
echo "install_rust: unsupported architecture $(uname -m)" >&2
exit 1
;;
esac
curl -sSLf -o /tmp/rustup-init \
"https://static.rust-lang.org/rustup/archive/1.28.2/${RUSTUP_TRIPLE}/rustup-init"
echo "${RUSTUP_SHA256} /tmp/rustup-init" | sha256sum -c -
chmod +x /tmp/rustup-init
/tmp/rustup-init -y --no-modify-path --profile minimal --default-toolchain 1.97.1
rm -f /tmp/rustup-init
echo 'export PATH="$HOME/.cargo/bin:$PATH"' >> "$BASH_ENV"
export PATH="$HOME/.cargo/bin:$PATH"
rustc --version
cargo --version
start_postgres:
description: "Start a postgres-db container on port 5432 and wait until it accepts connections."
parameters:
@ -163,6 +216,26 @@ commands:
done
echo "fake OpenAI endpoint did not become ready" >&2
exit 1
start_cost_center_service:
description: "Start the stand-in cost center validation service (tests/store_model_in_db_tests/cost_center_service.py) on host port 9414 and wait until healthy. The proxy's team-metadata validator (team_metadata_validator_e2e.py, impl 'http') reaches it via TEAM_METADATA_VALIDATION_SERVICE_URL=http://host.docker.internal:9414/validate. Run after uv deps are synced."
steps:
- run:
name: Start cost center validation service
background: true
command: |
uv run --no-sync python tests/store_model_in_db_tests/cost_center_service.py --host 0.0.0.0 --port 9414
- run:
name: Wait for cost center validation service
command: |
for i in $(seq 1 30); do
if curl -sf http://localhost:9414/health >/dev/null 2>&1; then
echo "cost center validation service is up"
exit 0
fi
sleep 1
done
echo "cost center validation service did not become ready" >&2
exit 1
setup_litellm_enterprise_pip:
steps:
- run:
@ -178,6 +251,7 @@ commands:
- checkout
- setup_google_dns
- install_uv
- install_rust
- restore_cache:
keys:
- v1-uv-cache-{{ checksum "uv.lock" }}
@ -292,6 +366,7 @@ jobs:
- checkout
- setup_google_dns
- install_uv
- install_rust
- run:
name: Build the wheel
environment:
@ -324,6 +399,7 @@ jobs:
keys:
- v1-uv-cache-{{ checksum "uv.lock" }}
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -397,6 +473,7 @@ jobs:
keys:
- v1-uv-cache-{{ checksum "uv.lock" }}
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -471,6 +548,7 @@ jobs:
keys:
- v1-uv-cache-{{ checksum "uv.lock" }}
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -522,6 +600,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -588,6 +667,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -628,6 +708,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -669,6 +750,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -702,6 +784,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- restore_cache:
keys:
- v1-uv-cache-{{ checksum "uv.lock" }}
@ -752,6 +835,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- restore_cache:
keys:
- v1-uv-cache-{{ checksum "uv.lock" }}
@ -803,6 +887,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -836,6 +921,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- restore_cache:
keys:
- v1-uv-cache-{{ checksum "uv.lock" }}
@ -882,6 +968,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -928,6 +1015,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -970,6 +1058,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -1016,6 +1105,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -1063,6 +1153,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- restore_cache:
keys:
- v1-uv-cache-{{ checksum "uv.lock" }}
@ -1103,6 +1194,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -1148,6 +1240,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -1192,6 +1285,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -1224,6 +1318,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -1267,6 +1362,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -1311,6 +1407,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -1355,6 +1452,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -1386,6 +1484,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -1432,6 +1531,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -1477,6 +1577,7 @@ jobs:
keys:
- v1-uv-cache-{{ checksum "uv.lock" }}
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -1527,6 +1628,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -1551,6 +1653,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -1577,6 +1680,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -1678,6 +1782,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -1773,6 +1878,7 @@ jobs:
at: ~/project
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -1861,6 +1967,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -1944,6 +2051,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -2076,6 +2184,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -2162,6 +2271,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -2258,12 +2368,14 @@ jobs:
- 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
- start_postgres
- start_fake_openai_endpoint
- start_cost_center_service
- attach_workspace:
at: ~/project
- run:
@ -2283,11 +2395,13 @@ jobs:
-e STORE_MODEL_IN_DB="True" \
-e LITELLM_MASTER_KEY="sk-1234" \
-e FAKE_OPENAI_API_BASE=http://host.docker.internal:8190 \
-e TEAM_METADATA_VALIDATION_SERVICE_URL=http://host.docker.internal:9414/validate \
-e LITELLM_LICENSE=$LITELLM_LICENSE \
-e LITELLM_LOG=ERROR \
--add-host host.docker.internal:host-gateway \
--name my-app \
-v $(pwd)/litellm/proxy/example_config_yaml/store_model_db_config.yaml:/app/config.yaml \
-v $(pwd)/litellm/proxy/example_config_yaml/team_metadata_validator_e2e.py:/app/team_metadata_validator_e2e.py \
litellm-docker-database:ci \
--config /app/config.yaml \
--port 4000
@ -2333,6 +2447,7 @@ jobs:
- setup_google_dns
# Remove Docker CLI installation since it's already available in machine executor
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -2414,6 +2529,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -2499,20 +2615,9 @@ jobs:
bundle exec rspec
no_output_timeout: 30m
# Install Node.js directly from nodejs.org with SHA256 verification,
# instead of piping NodeSource's setup_18.x apt-repo installer into
# instead of piping NodeSource's setup_24.x apt-repo installer into
# sudo bash (which runs a mutable upstream script unattended).
- run:
name: Install Node.js 18.20.8
command: |
NODE_VERSION="18.20.8"
NODE_TARBALL="node-v${NODE_VERSION}-linux-x64.tar.xz"
NODE_EXPECTED_SHA="5467ee62d6af1411d46b6a10e3fb5cacc92734dbcef465fea14e7b90993001c9"
curl -sSLf -o "/tmp/${NODE_TARBALL}" "https://nodejs.org/dist/v${NODE_VERSION}/${NODE_TARBALL}"
echo "${NODE_EXPECTED_SHA} /tmp/${NODE_TARBALL}" | sha256sum -c -
sudo tar -xJf "/tmp/${NODE_TARBALL}" -C /usr/local --strip-components=1
rm -f "/tmp/${NODE_TARBALL}"
node --version
npm --version
- install_node
- run:
name: Install Node.js test dependencies
@ -2553,6 +2658,7 @@ jobs:
- skip_if_unrelated_changes
- setup_google_dns
- install_uv
- install_rust
- run:
name: Install Dependencies
command: |
@ -2640,7 +2746,7 @@ jobs:
ui_build:
docker:
- image: cimg/node:20.19@sha256:35e64883e8d21bc345b0a7b04c35ee46442c127607ed1d8d7d37d8a1ed76db81
- image: cimg/node:24.19@sha256:8966565f07189a67d64d6808a2b127f31dafae566508e3547f55640e1070bfad
auth:
username: ${DOCKERHUB_USERNAME}
password: ${DOCKERHUB_PASSWORD}
@ -2684,7 +2790,7 @@ jobs:
ui_unit_tests:
docker:
- image: cimg/node:20.19@sha256:35e64883e8d21bc345b0a7b04c35ee46442c127607ed1d8d7d37d8a1ed76db81
- image: cimg/node:24.19@sha256:8966565f07189a67d64d6808a2b127f31dafae566508e3547f55640e1070bfad
auth:
username: ${DOCKERHUB_USERNAME}
password: ${DOCKERHUB_PASSWORD}
@ -2742,7 +2848,9 @@ jobs:
- skip_if_unrelated_changes:
category: client
- setup_google_dns
- install_node
- install_uv
- install_rust
- restore_cache:
keys:
- v1-uv-cache-{{ checksum "uv.lock" }}
@ -2757,7 +2865,7 @@ jobs:
- ~/.cache/uv
- restore_cache:
keys:
- ui-e2e-node-deps-v3-{{ checksum "ui/litellm-dashboard/package-lock.json" }}-{{ checksum "tests/e2e/ui/package-lock.json" }}
- ui-e2e-node-deps-v4-{{ checksum "ui/litellm-dashboard/package-lock.json" }}-{{ checksum "tests/e2e/ui/package-lock.json" }}
- run:
name: Install Node dependencies and Playwright
# The cimg/python:3.12-browsers image already ships the Chromium system
@ -2772,7 +2880,7 @@ jobs:
npm ci
npx playwright install chromium
- save_cache:
key: ui-e2e-node-deps-v3-{{ checksum "ui/litellm-dashboard/package-lock.json" }}-{{ checksum "tests/e2e/ui/package-lock.json" }}
key: ui-e2e-node-deps-v4-{{ checksum "ui/litellm-dashboard/package-lock.json" }}-{{ checksum "tests/e2e/ui/package-lock.json" }}
paths:
- ui/litellm-dashboard/node_modules
- tests/e2e/ui/node_modules
@ -2884,7 +2992,9 @@ jobs:
- skip_if_unrelated_changes:
category: client
- setup_google_dns
- install_node
- install_uv
- install_rust
- restore_cache:
keys:
- v1-uv-cache-{{ checksum "uv.lock" }}
@ -2899,7 +3009,7 @@ jobs:
- ~/.cache/uv
- restore_cache:
keys:
- ui-e2e-node-deps-v3-{{ checksum "ui/litellm-dashboard/package-lock.json" }}-{{ checksum "tests/e2e/ui/package-lock.json" }}
- ui-e2e-node-deps-v4-{{ checksum "ui/litellm-dashboard/package-lock.json" }}-{{ checksum "tests/e2e/ui/package-lock.json" }}
- run:
name: Install Node dependencies and Playwright
command: |
@ -2909,7 +3019,7 @@ jobs:
npm ci
npx playwright install chromium
- save_cache:
key: ui-e2e-node-deps-v3-{{ checksum "ui/litellm-dashboard/package-lock.json" }}-{{ checksum "tests/e2e/ui/package-lock.json" }}
key: ui-e2e-node-deps-v4-{{ checksum "ui/litellm-dashboard/package-lock.json" }}-{{ checksum "tests/e2e/ui/package-lock.json" }}
paths:
- ui/litellm-dashboard/node_modules
- tests/e2e/ui/node_modules

46
.flake8
View file

@ -1,46 +0,0 @@
[flake8]
ignore =
# The following ignores can be removed when formatting using black
W191,W291,W292,W293,W391,W504
E101,E111,E114,E116,E117,E121,E122,E123,E124,E125,E126,E127,E128,E129,E131,
E201,E202,E221,E222,E225,E226,E231,E241,E251,E252,E261,E265,E271,E272,E275,
E301,E302,E303,E305,E306,
# line break before binary operator
W503,
# inline comment should start with '# '
E262,
# too many leading '#' for block comment
E266,
# multiple imports on one line
E401,
# module level import not at top of file
E402,
# Line too long (82 > 79 characters)
E501,
# comparison to None should be 'if cond is None:'
E711,
# comparison to True should be 'if cond is True:' or 'if cond:'
E712,
# do not compare types, for exact checks use `is` / `is not`, for instance checks use `isinstance()`
E721,
# do not use bare 'except'
E722,
# x is imported but unused
F401,
# 'from . import *' used; unable to detect undefined names
F403,
# x may be undefined, or defined from star imports:
F405,
# f-string is missing placeholders
F541,
# dictionary key '' repeated with different values
F601,
# redefinition of unused x from line 123
F811,
# undefined name x
F821,
# local variable x is assigned to but never used
F841,
# https://black.readthedocs.io/en/stable/guides/using_black_with_other_tools.html#flake8
extend-ignore = E203

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

@ -154,6 +154,19 @@ jobs:
merge-multiple: true
- name: Upload to Codecov
id: codecov-upload
continue-on-error: true
uses: codecov/codecov-action@75cd11691c0faa626561e295848008c8a7dddffe # v5.5.4
with:
use_oidc: true
directory: coverage-reports
root_dir: ${{ github.workspace }}
flags: ${{ inputs.artifact-name }}
fail_ci_if_error: false
- name: Upload to Codecov (retry)
if: steps.codecov-upload.outcome == 'failure'
continue-on-error: true
uses: codecov/codecov-action@75cd11691c0faa626561e295848008c8a7dddffe # v5.5.4
with:
use_oidc: true

2
.github/workflows/check-ui-api-types.yml vendored
View file

@ -56,7 +56,7 @@ jobs:
- name: Set up Node.js
uses: actions/setup-node@a0853c24544627f65ddf259abe73b1d18a591444 # v5.0.0
with:
node-version: "20"
node-version-file: ui/litellm-dashboard/.nvmrc
cache: "npm"
cache-dependency-path: ui/litellm-dashboard/package-lock.json

3
.github/workflows/test-linting.yml vendored
View file

@ -104,9 +104,8 @@ jobs:
- name: Check basedpyright budget (delta vs base)
env:
BASE_SHA: ${{ github.event.pull_request.base.sha }}
NODE_OPTIONS: --max-old-space-size=12288
run: |
(uv run --no-sync basedpyright --outputjson || true) | uv run --no-sync python scripts/type_check_gate.py --base "$BASE_SHA"
uv run --no-sync python scripts/type_check_gate.py --base "$BASE_SHA"
- name: Check tests/e2e basedpyright (zero errors)
env:

2
.github/workflows/test-litellm-ui-build.yml vendored
View file

@ -27,7 +27,7 @@ jobs:
- name: Setup Node.js
uses: actions/setup-node@a0853c24544627f65ddf259abe73b1d18a591444 # v5.0.0
with:
node-version: "20"
node-version-file: ui/litellm-dashboard/.nvmrc
cache: "npm"
cache-dependency-path: ui/litellm-dashboard/package-lock.json

2
.github/workflows/test-litellm-ui-lint.yml vendored
View file

@ -61,7 +61,7 @@ jobs:
if: steps.changed.outputs.has_files == 'true'
uses: actions/setup-node@a0853c24544627f65ddf259abe73b1d18a591444 # v5.0.0
with:
node-version: "20"
node-version-file: ui/litellm-dashboard/.nvmrc
cache: "npm"
cache-dependency-path: ui/litellm-dashboard/package-lock.json

2
.github/workflows/test-litellm-ui-unit.yml vendored
View file

@ -35,7 +35,7 @@ jobs:
- name: Setup Node.js
uses: actions/setup-node@a0853c24544627f65ddf259abe73b1d18a591444 # v5.0.0
with:
node-version: "20"
node-version-file: ui/litellm-dashboard/.nvmrc
cache: "npm"
cache-dependency-path: ui/litellm-dashboard/package-lock.json

9
CLAUDE.md
View file

@ -29,7 +29,7 @@ Never use `pytest` commands or the like as "Screenshots / Proof of Fix". We pref
If you ever make public-facing PR descriptions, comments, issues, commit messages, etc., always follow these guidelines to sound less AI-y:
- don't use emojis
- don't use "—". Instead, reach for ";", ".", etc.
- 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
- 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
- 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 "."
@ -39,15 +39,11 @@ Don't hesitate to use values in .env to get needed API keys and other secrets, a
Python max line length is 120, not 88
On a fresh worktree or clone, run `make bootstrap` before anything else. It provisions everything tests, `make pre-commit`, and a local proxy need
Run tests before you commit. Also, run `make pre-commit` right before each commit, which generates types (as needed) and formats/lints your code. Any errors found must be fixed. It only runs when there are staged frontend and/or backend changes and calculates violations, generates types, etc. based on the worktree, so stage what you need or stash/delete unwanted files in litellm/ or ui/ (where backend and frontend lint run, respectively) before running it. If it fails because dashboard api types are stale, it already regenerated them for you. You just need to stage the schema.d.ts, re-run `make pre-commit` to confirm it passes, and commit
When you fix violations gated by `ruff-strict-budget.json`, `type-discipline-budget.json`, or `basedpyright-code-budget.json`, run `make lint-budget-update` and commit the lowered limits so the ceilings ratchet down instead of leaving stale headroom. It measures the working tree, so it must contain exactly the fixes you're committing
If you're trying to create a new function that relies on untyped stuff, instead of adding more Any's and pushing `reportAny` / `reportExplicitAny` closer to their basedpyright ceilings, just validate it in the caller with Pydantic (a model or `TypeAdapter` that returns the typed thing or raises will do) and then pass the now typed variable in
If you get an LIT001 or LIT002 fail, refactor the code to follow functional programming best practices rather than introducing mutable data structures. For example, build values in one shot with comprehensions or generators wrapped in `tuple()` / `frozenset()` instead of seeding an empty `list`/`dict`/`set` and mutating it over time. Ideally `# mutable-ok` is never used; reach for it only as a genuine last resort when an immutable rewrite is truly impossible, and always pair it with a real reason
If you get an LIT001 or LIT002 fail, refactor the code to follow functional programming best practices rather than introducing mutable data structures. For example, build values in one shot with comprehensions or generators wrapped in `tuple()` / `frozenset()` instead of seeding an empty `list`/`dict`/`set` and mutating it over time. Ideally, `# mutable-ok` is never used; reach for it only as a genuine last resort when an immutable rewrite is truly impossible, and always pair it with a real reason
Every lint or type suppression must name the exact rule inside brackets and carry a reason comment, e.g. `# pyright: ignore[reportArgumentType] # stubs lack async overload` or `# noqa: TID251 # <reason>`. `# type: ignore` is banned (LIT009): pyrightconfig.json sets `enableTypeIgnoreComments` to false, so it silently does nothing
@ -75,6 +71,7 @@ Follow these coding conventions for new/updated code (a three-line fix in a lega
- Never-nester: early returns over deep nesting
- Don't throw; model failures as values (One function (e.g., raise_public) maps error union to existing public exception contracts via exhaustive match + assert_never)
- No mutation; don't reassign variables, global or local. Instead of mutable lists and dicts, prefer tuples, frozen dataclasses (with slots=True), etc.
- Annotate every variable with `: Final` (LIT010). Unpacking and walrus targets cannot carry the annotation, so they are implicitly final. Don't rebind them. Never rebind or mutate function parameters (LIT011); `self`/`cls` attribute stores are the exception. If rebinding or in-place mutation is truly unavoidable, suppress with `# rebind-ok: <reason>` explaining why
- Use dependency injection
- Fully typed; no `Any` or coarse types like `dict[str, Any]` or just `dict`. Every function parameter must be strongly typed
- Use tagged unions + match

2
Dockerfile
View file

@ -7,7 +7,7 @@ ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
# Pinned by digest like the other base images; bump explicitly on Node upgrades.
ARG UI_BUILD_IMAGE=node:20.18-alpine3.20@sha256:3488b10bf958af7125a176419d2d8a9937d895bf124012aae811651988d2ffe6
ARG UI_BUILD_IMAGE=node:24.19-alpine3.24@sha256:d32cdf619f63fe0471182d08996dd516c6275bb5fd31ae06e55a570bd9e1ad43
FROM $UV_IMAGE AS uvbin

8
Makefile
View file

@ -75,7 +75,7 @@ install-dev:
bootstrap:
$(UV) sync --inexact --frozen --extra proxy --group proxy-dev --group e2e-dev
$(UV_RUN) python scripts/prisma_generate_if_needed.py
cd ui/litellm-dashboard && npm install --no-audit --no-fund
cd ui/litellm-dashboard && ../../scripts/with_dashboard_node.sh npm install --no-audit --no-fund
@main_root=$$(git worktree list --porcelain | head -1 | sed 's/^worktree //'); \
if [ "$$main_root" != "$$(git rev-parse --show-toplevel)" ] && [ -f "$$main_root/.env" ] && [ ! -f .env ]; then \
cp "$$main_root/.env" .env && echo "bootstrap: copied .env from $$main_root"; \
@ -176,10 +176,8 @@ lint-ruff-FULL-dev: install-dev
if [ -n "$$files" ]; then echo "$$files" | xargs $(UV_RUN) ruff check; \
else echo "No changed .py files to check."; fi
lint-basedpyright lint-basedpyright-budget-update: export NODE_OPTIONS := --max-old-space-size=12288
lint-basedpyright: $(LINT_DEP_INSTALL) $(LINT_DEP_BASE)
($(UV_RUN) basedpyright --outputjson || true) | $(UV_RUN) python scripts/type_check_gate.py --base origin/litellm_internal_staging
$(UV_RUN) python scripts/type_check_gate.py --base origin/litellm_internal_staging
lint-e2e-basedpyright: $(LINT_E2E_DEP_INSTALL)
$(UV_RUN) basedpyright tests/e2e
@ -192,7 +190,7 @@ lint-type-discipline: $(LINT_DEP_INSTALL) $(LINT_DEP_BASE)
# --update lowers each limit by what this branch fixed since its branch point, so
# it needs the base ref fetched to resolve the merge-base.
lint-basedpyright-budget-update: install-dev lint-fetch-base
($(UV_RUN) basedpyright --outputjson || true) | $(UV_RUN) python scripts/type_check_gate.py --update
$(UV_RUN) python scripts/type_check_gate.py --update
lint-format: format-check

1
backend/routes/allowlist.py
View file

@ -44,6 +44,7 @@ BACKEND_PATH_PREFIXES: tuple[str, ...] = (
"/router/",
"/router_settings",
"/adaptive_router/",
"/auto_router/",
"/fallback",
"/fallbacks",
"/cache_settings",

50
basedpyright-code-budget.json
View file

@ -1,6 +1,6 @@
{
"reportAny": {
"limit": 29813
"limit": 29809
},
"reportArgumentType": {
"limit": 2645
@ -15,40 +15,40 @@
"limit": 123
},
"reportConstantRedefinition": {
"limit": 59
"limit": 40
},
"reportDeprecated": {
"limit": 325
"limit": 215
},
"reportDuplicateImport": {
"limit": 42
"limit": 24
},
"reportExplicitAny": {
"limit": 9473
},
"reportFunctionMemberAccess": {
"limit": 11
"limit": 7
},
"reportGeneralTypeIssues": {
"limit": 227
"limit": 157
},
"reportIncompatibleMethodOverride": {
"limit": 77
"limit": 56
},
"reportIncompatibleVariableOverride": {
"limit": 12
"limit": 8
},
"reportInconsistentOverload": {
"limit": 18
"limit": 12
},
"reportIndexIssue": {
"limit": 37
"limit": 35
},
"reportInvalidTypeForm": {
"limit": 35
},
"reportInvalidTypeVarUse": {
"limit": 5
"limit": 2
},
"reportMatchNotExhaustive": {
"limit": 0
@ -57,10 +57,10 @@
"limit": 5855
},
"reportMissingTypeArgument": {
"limit": 15852
"limit": 15849
},
"reportMissingTypeStubs": {
"limit": 41
"limit": 40
},
"reportOperatorIssue": {
"limit": 0
@ -81,13 +81,13 @@
"limit": 0
},
"reportPossiblyUnboundVariable": {
"limit": 77
"limit": 56
},
"reportPrivateUsage": {
"limit": 2437
"limit": 1825
},
"reportRedeclaration": {
"limit": 12
"limit": 8
},
"reportReturnType": {
"limit": 219
@ -99,7 +99,7 @@
"limit": 0
},
"reportUnknownArgumentType": {
"limit": 45324
"limit": 45262
},
"reportUnknownLambdaType": {
"limit": 113
@ -114,16 +114,16 @@
"limit": 31978
},
"reportUnnecessaryCast": {
"limit": 177
"limit": 124
},
"reportUnnecessaryComparison": {
"limit": 1021
"limit": 703
},
"reportUnnecessaryContains": {
"limit": 7
"limit": 5
},
"reportUnnecessaryIsInstance": {
"limit": 1204
"limit": 866
},
"reportUntypedBaseClass": {
"limit": 165
@ -132,15 +132,15 @@
"limit": 33
},
"reportUnusedClass": {
"limit": 33
"limit": 23
},
"reportUnusedFunction": {
"limit": 204
"limit": 139
},
"reportUnusedImport": {
"limit": 1003
"limit": 588
},
"reportUnusedVariable": {
"limit": 1297
"limit": 147
}
}

2
docker/Dockerfile.database
View file

@ -7,7 +7,7 @@ ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17
ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b17a5ee8bc23b532600a44d705acef2409e0933c1251b45f
ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
# Pinned by digest like the other base images; bump explicitly on Node upgrades.
ARG UI_BUILD_IMAGE=node:20.18-alpine3.20@sha256:3488b10bf958af7125a176419d2d8a9937d895bf124012aae811651988d2ffe6
ARG UI_BUILD_IMAGE=node:24.19-alpine3.24@sha256:d32cdf619f63fe0471182d08996dd516c6275bb5fd31ae06e55a570bd9e1ad43
FROM $UV_IMAGE AS uvbin

2
docker/Dockerfile.non_root
View file

@ -6,7 +6,7 @@ ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:42df77a9974d6ec8b
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.
ARG UI_BUILD_IMAGE=node:20.18-alpine3.20@sha256:3488b10bf958af7125a176419d2d8a9937d895bf124012aae811651988d2ffe6
ARG UI_BUILD_IMAGE=node:24.19-alpine3.24@sha256:d32cdf619f63fe0471182d08996dd516c6275bb5fd31ae06e55a570bd9e1ad43
FROM $UV_IMAGE AS uvbin

5
docker/build_admin_ui.sh
View file

@ -54,8 +54,9 @@ else
fi || { echo "nvm checksum verification failed"; exit 1; }
bash "$NVM_SCRIPT"
source ~/.nvm/nvm.sh
nvm install v18.17.0
nvm use v18.17.0
NODE_VERSION="$(cat ui/litellm-dashboard/.nvmrc)"
nvm install "v${NODE_VERSION}"
nvm use "v${NODE_VERSION}"
# cd in to /ui/litellm-dashboard

0
enterprise/litellm_enterprise/enterprise_callbacks/__init__.py Normal file
View file

0
enterprise/litellm_enterprise/enterprise_callbacks/pagerduty/__init__.py Normal file
View file

0
enterprise/litellm_enterprise/enterprise_callbacks/send_emails/__init__.py Normal file
View file

0
enterprise/litellm_enterprise/integrations/__init__.py Normal file
View file

0
enterprise/litellm_enterprise/litellm_core_utils/__init__.py Normal file
View file

0
enterprise/litellm_enterprise/proxy/hooks/__init__.py Normal file
View file

6
enterprise/litellm_enterprise/proxy/hooks/managed_files.py
View file

@ -382,7 +382,11 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
"flat_model_file_ids": {"hasSome": model_object_ids},
}
)
return [OpenAIFileObject.model_validate(file_object.file_object) for file_object in file_ids]
return [
OpenAIFileObject.model_validate(file_object.file_object)
for file_object in file_ids
if file_object.file_object is not None
]
async def check_managed_file_id_access(
self, data: Dict, user_api_key_dict: UserAPIKeyAuth

0
enterprise/litellm_enterprise/proxy/vector_stores/__init__.py Normal file
View file

0
enterprise/litellm_enterprise/py.typed Normal file
View file

0
enterprise/litellm_enterprise/types/__init__.py Normal file
View file

0
enterprise/litellm_enterprise/types/enterprise_callbacks/__init__.py Normal file
View file

0
enterprise/litellm_enterprise/types/proxy/__init__.py Normal file
View file

2
helm/litellm-helm/README.md
View file

@ -39,7 +39,7 @@ If `db.useStackgresOperator` is used (not yet implemented):
| `livenessProbe.*` | Liveness probe settings for the LiteLLM container (`path`, `periodSeconds`, `timeoutSeconds`, thresholds, and initial delay). | See `values.yaml` |
| `readinessProbe.*` | Readiness probe settings for the LiteLLM container (`path`, `periodSeconds`, `timeoutSeconds`, thresholds, and initial delay). | See `values.yaml` |
| `startupProbe.*` | Startup probe settings for the LiteLLM container (`path`, `periodSeconds`, `timeoutSeconds`, thresholds, and initial delay). | See `values.yaml` |
| `resources.*` | CPU/memory requests and limits for the LiteLLM container. | `{}` |
| `resources.*` | CPU/memory requests and limits for the LiteLLM container. Unset by default; production deployments should set 1 CPU and 4Gi of memory per worker. | `{}` |
| `service.loadBalancerClass` | Optional LoadBalancer implementation class (only used when `service.type` is `LoadBalancer`) | `""` |
| `ingress.labels` | Additional labels for the Ingress resource | `{}` |
| `ingress.*` | See [values.yaml](./values.yaml) for example settings | N/A |

27
helm/litellm-helm/values.yaml
View file

@ -181,16 +181,19 @@ proxy_config:
resources:
{}
# We usually recommend not to specify default resources and to leave this as a conscious
# choice for the user. This also increases chances charts run on environments with little
# resources, such as Minikube. If you do want to specify resources, uncomment the following
# lines, adjust them as necessary, and remove the curly braces after 'resources:'.
# limits:
# cpu: 100m
# memory: 128Mi
# Unset by default so the chart installs on small clusters such as Minikube, and so an
# upgrade never leaves a running pod Pending. Production deployments should set these.
# A proxy at DB-connected steady state needs about 1 CPU and 4Gi of memory per worker;
# sizing below that gets the pod OOMKilled once traffic and DB connections ramp up.
# Scale both figures with --num_workers, then uncomment the lines below and remove the
# curly braces after 'resources:'. See "Recommended Machine Specifications" in
# https://docs.litellm.ai/docs/proxy/prod.
# requests:
# cpu: 100m
# memory: 128Mi
# cpu: "1"
# memory: 4Gi
# limits:
# cpu: "1"
# memory: 4Gi
autoscaling:
enabled: false
@ -432,9 +435,9 @@ migrationJob:
annotations: {}
ttlSecondsAfterFinished: 120
resources: {}
# requests:
# cpu: 100m
# memory: 100Mi
# Unset by default. This job runs the database migration and exits, so it does not
# need the steady-state headroom the proxy does; size it from your own migration
# runs rather than from the proxy figures above.
extraContainers: []
extraInitContainers: []

3
litellm-proxy-extras/litellm_proxy_extras/migrations/20260729000000_add_reload_tracking_to_litellm_config/migration.sql Normal file
View file

@ -0,0 +1,3 @@
-- AlterTable
ALTER TABLE "LiteLLM_Config" ADD COLUMN IF NOT EXISTS "last_run_at" TIMESTAMP(3),
ADD COLUMN IF NOT EXISTS "reload_revision" BIGINT NOT NULL DEFAULT 0;

17
litellm-proxy-extras/litellm_proxy_extras/migrations/20260731000000_add_autorouter_savings_spend/migration.sql Normal file
View file

@ -0,0 +1,17 @@
-- AlterTable
ALTER TABLE "LiteLLM_DailyUserSpend" ADD COLUMN IF NOT EXISTS "autorouter_savings_spend" DOUBLE PRECISION NOT NULL DEFAULT 0.0;
-- AlterTable
ALTER TABLE "LiteLLM_DailyOrganizationSpend" ADD COLUMN IF NOT EXISTS "autorouter_savings_spend" DOUBLE PRECISION NOT NULL DEFAULT 0.0;
-- AlterTable
ALTER TABLE "LiteLLM_DailyEndUserSpend" ADD COLUMN IF NOT EXISTS "autorouter_savings_spend" DOUBLE PRECISION NOT NULL DEFAULT 0.0;
-- AlterTable
ALTER TABLE "LiteLLM_DailyAgentSpend" ADD COLUMN IF NOT EXISTS "autorouter_savings_spend" DOUBLE PRECISION NOT NULL DEFAULT 0.0;
-- AlterTable
ALTER TABLE "LiteLLM_DailyTeamSpend" ADD COLUMN IF NOT EXISTS "autorouter_savings_spend" DOUBLE PRECISION NOT NULL DEFAULT 0.0;
-- AlterTable
ALTER TABLE "LiteLLM_DailyTagSpend" ADD COLUMN IF NOT EXISTS "autorouter_savings_spend" DOUBLE PRECISION NOT NULL DEFAULT 0.0;

177
litellm-proxy-extras/litellm_proxy_extras/prisma_toolchain.py Normal file
View file

@ -0,0 +1,177 @@
"""Prepare the Node toolchain the Prisma CLI needs, separately from migrations.
The Prisma CLI is a Node program. The first invocation inside a fresh
container installs a private Node runtime and npm-installs the CLI itself,
which can take minutes on a cold or slow machine. Sharing one timeout between
that one-time bootstrap and the migration commands makes a slow bootstrap
indistinguishable from a slow migration, so the bootstrap gets killed long
before it can finish.
A killed bootstrap does not correct itself. The installer leaves its cache
directory behind, and Prisma decides whether to install by testing that
directory for existence alone, so every later attempt skips the install and
then fails on a Node binary that was never written. Deleting a cache directory
that exists without a Node binary is what turns a killed bootstrap back into a
recoverable one.
Both budgets are overridable so an operator can widen them without a release:
``LITELLM_PRISMA_BOOTSTRAP_TIMEOUT`` for the toolchain install and
``LITELLM_PRISMA_COMMAND_TIMEOUT`` for every individual Prisma command.
"""
import math
import os
import shutil
import subprocess
from dataclasses import dataclass
from pathlib import Path
from typing import Optional
from litellm_proxy_extras._logging import logger
try:
from prisma import config as prisma_config
except ImportError:
prisma_config = None
PRISMA_COMMAND_TIMEOUT_ENV_VAR = "LITELLM_PRISMA_COMMAND_TIMEOUT"
PRISMA_BOOTSTRAP_TIMEOUT_ENV_VAR = "LITELLM_PRISMA_BOOTSTRAP_TIMEOUT"
NODEENV_CACHE_DIR_ENV_VAR = "PRISMA_NODEENV_CACHE_DIR"
DEFAULT_PRISMA_COMMAND_TIMEOUT = 60.0
DEFAULT_PRISMA_BOOTSTRAP_TIMEOUT = 600.0
BOOTSTRAP_ARG = "--version"
@dataclass(frozen=True)
class ToolchainBootstrap:
"""Outcome of preparing the Prisma toolchain."""
healed_incomplete_cache: bool
ready: bool
def _timeout_from_env(env_var: str, default: float) -> float:
raw = os.getenv(env_var)
if raw is None:
return default
try:
seconds = float(raw)
except ValueError:
logger.warning(
"%s=%r is not a number, falling back to %ss", env_var, raw, default
)
return default
if not math.isfinite(seconds) or seconds <= 0:
logger.warning(
"%s=%r is not a finite positive number, falling back to %ss",
env_var,
raw,
default,
)
return default
return seconds
def prisma_command_timeout() -> float:
"""Seconds any single Prisma command may run for."""
return _timeout_from_env(
PRISMA_COMMAND_TIMEOUT_ENV_VAR, DEFAULT_PRISMA_COMMAND_TIMEOUT
)
def prisma_bootstrap_timeout() -> float:
"""Seconds the one-time Node toolchain install may run for."""
return _timeout_from_env(
PRISMA_BOOTSTRAP_TIMEOUT_ENV_VAR, DEFAULT_PRISMA_BOOTSTRAP_TIMEOUT
)
def nodeenv_cache_dir() -> Optional[Path]:
"""Where Prisma installs its private Node runtime, or None if unknowable."""
override = os.getenv(NODEENV_CACHE_DIR_ENV_VAR)
if override:
return Path(override).absolute()
if prisma_config is not None:
try:
return Path(prisma_config.nodeenv_cache_dir).absolute()
except (OSError, ValueError) as e:
logger.warning("Could not read the Prisma nodeenv cache dir: %s", e)
try:
return Path.home() / ".cache" / "prisma-python" / "nodeenv"
except RuntimeError:
logger.warning(
"No resolvable home directory, cannot locate the Prisma nodeenv cache"
)
return None
def node_binary_path(cache_dir: Path) -> Path:
"""Path the Node binary occupies once the toolchain is fully installed."""
if os.name == "nt":
return cache_dir / "Scripts" / "node.exe"
return cache_dir / "bin" / "node"
def heal_incomplete_nodeenv_cache() -> bool:
"""Delete a nodeenv cache directory left without a Node binary.
Returns True when a half-installed toolchain was removed, so the next
Prisma invocation reinstalls it instead of failing on a missing binary.
"""
cache_dir = nodeenv_cache_dir()
if cache_dir is None or not cache_dir.is_dir():
return False
if node_binary_path(cache_dir).exists():
return False
logger.warning(
"Node toolchain at %s has no %s, so a previous install was interrupted. "
"Removing it so it can be reinstalled.",
cache_dir,
node_binary_path(cache_dir).name,
)
try:
shutil.rmtree(cache_dir)
except OSError as e:
logger.warning("Could not remove %s: %s", cache_dir, e)
return False
return True
def ensure_prisma_toolchain(
prisma_command: str, prisma_env: dict[str, str]
) -> ToolchainBootstrap:
"""Install whatever the Prisma CLI needs to run, under its own timeout.
Never raises. A toolchain that cannot be prepared is reported so the
caller can go on and let the real Prisma command produce the real error.
"""
healed = heal_incomplete_nodeenv_cache()
timeout = prisma_bootstrap_timeout()
logger.info("Preparing the Prisma CLI toolchain (timeout %ss)", timeout)
try:
subprocess.run(
[prisma_command, BOOTSTRAP_ARG],
timeout=timeout,
check=True,
capture_output=True,
text=True,
env=prisma_env,
)
except subprocess.TimeoutExpired:
logger.warning(
"Preparing the Prisma CLI toolchain timed out after %ss. Raise %s "
"if this machine needs longer to install it.",
timeout,
PRISMA_BOOTSTRAP_TIMEOUT_ENV_VAR,
)
return ToolchainBootstrap(healed_incomplete_cache=healed, ready=False)
except subprocess.CalledProcessError as e:
logger.warning("Preparing the Prisma CLI toolchain failed: %s", e.stderr)
return ToolchainBootstrap(healed_incomplete_cache=healed, ready=False)
except OSError as e:
logger.warning("Could not run the Prisma CLI: %s", e)
return ToolchainBootstrap(healed_incomplete_cache=healed, ready=False)
logger.info("Prisma CLI toolchain ready")
return ToolchainBootstrap(healed_incomplete_cache=healed, ready=True)

0
litellm-proxy-extras/litellm_proxy_extras/py.typed Normal file
View file

3
litellm-proxy-extras/litellm_proxy_extras/replica_identity.py
View file

@ -16,6 +16,7 @@ import tempfile
from pathlib import Path
from litellm_proxy_extras._logging import logger
from litellm_proxy_extras.prisma_toolchain import prisma_command_timeout
REPLICA_IDENTITY_FULL_ENV_VAR = "LITELLM_SET_REPLICA_IDENTITY_FULL"
@ -75,7 +76,7 @@ def apply_replica_identity_full(
"--schema",
schema_path,
],
timeout=60,
timeout=prisma_command_timeout(),
check=True,
capture_output=True,
text=True,

8
litellm-proxy-extras/litellm_proxy_extras/schema.prisma
View file

@ -601,6 +601,8 @@ model LiteLLM_TagTable {
model LiteLLM_Config {
param_name String @id
param_value Json?
last_run_at DateTime?
reload_revision BigInt @default(0)
}
// View spend, model, api_key per request
@ -748,6 +750,7 @@ model LiteLLM_DailyUserSpend {
compression_saved_tokens BigInt @default(0)
compression_savings_spend Float @default(0.0)
prompt_caching_savings_spend Float @default(0.0)
autorouter_savings_spend Float @default(0.0)
spend Float @default(0.0)
api_requests BigInt @default(0)
successful_requests BigInt @default(0)
@ -782,6 +785,7 @@ model LiteLLM_DailyOrganizationSpend {
compression_saved_tokens BigInt @default(0)
compression_savings_spend Float @default(0.0)
prompt_caching_savings_spend Float @default(0.0)
autorouter_savings_spend Float @default(0.0)
spend Float @default(0.0)
api_requests BigInt @default(0)
successful_requests BigInt @default(0)
@ -816,6 +820,7 @@ model LiteLLM_DailyEndUserSpend {
compression_saved_tokens BigInt @default(0)
compression_savings_spend Float @default(0.0)
prompt_caching_savings_spend Float @default(0.0)
autorouter_savings_spend Float @default(0.0)
spend Float @default(0.0)
api_requests BigInt @default(0)
successful_requests BigInt @default(0)
@ -849,6 +854,7 @@ model LiteLLM_DailyAgentSpend {
compression_saved_tokens BigInt @default(0)
compression_savings_spend Float @default(0.0)
prompt_caching_savings_spend Float @default(0.0)
autorouter_savings_spend Float @default(0.0)
spend Float @default(0.0)
api_requests BigInt @default(0)
successful_requests BigInt @default(0)
@ -882,6 +888,7 @@ model LiteLLM_DailyTeamSpend {
compression_saved_tokens BigInt @default(0)
compression_savings_spend Float @default(0.0)
prompt_caching_savings_spend Float @default(0.0)
autorouter_savings_spend Float @default(0.0)
spend Float @default(0.0)
api_requests BigInt @default(0)
successful_requests BigInt @default(0)
@ -917,6 +924,7 @@ model LiteLLM_DailyTagSpend {
compression_saved_tokens BigInt @default(0)
compression_savings_spend Float @default(0.0)
prompt_caching_savings_spend Float @default(0.0)
autorouter_savings_spend Float @default(0.0)
spend Float @default(0.0)
api_requests BigInt @default(0)
successful_requests BigInt @default(0)

33
litellm-proxy-extras/litellm_proxy_extras/utils.py
View file

@ -14,6 +14,10 @@ from litellm_proxy_extras.replica_identity import (
REPLICA_IDENTITY_FULL_ENV_VAR,
apply_replica_identity_full,
)
from litellm_proxy_extras.prisma_toolchain import (
ensure_prisma_toolchain,
prisma_command_timeout,
)
def str_to_bool(value: Optional[str]) -> bool:
@ -142,7 +146,7 @@ class ProxyExtrasDBManager:
],
stdout=open(migration_file, "w"),
check=True,
timeout=30,
timeout=prisma_command_timeout(),
env=prisma_env,
)
@ -157,7 +161,7 @@ class ProxyExtrasDBManager:
"0_init",
],
check=True,
timeout=30,
timeout=prisma_command_timeout(),
env=prisma_env,
)
@ -193,7 +197,7 @@ class ProxyExtrasDBManager:
"--rolled-back",
migration_name,
],
timeout=60,
timeout=prisma_command_timeout(),
check=True,
capture_output=True,
env=prisma_env,
@ -205,7 +209,7 @@ class ProxyExtrasDBManager:
prisma_env = _get_prisma_env()
subprocess.run(
[_get_prisma_command(), "migrate", "resolve", "--applied", migration_name],
timeout=60,
timeout=prisma_command_timeout(),
check=True,
capture_output=True,
env=prisma_env,
@ -303,7 +307,7 @@ class ProxyExtrasDBManager:
"--script",
],
check=True,
timeout=60,
timeout=prisma_command_timeout(),
stdout=f,
env=_get_prisma_env(),
)
@ -335,7 +339,7 @@ class ProxyExtrasDBManager:
"--schema",
schema_path,
],
timeout=60,
timeout=prisma_command_timeout(),
check=True,
capture_output=True,
text=True,
@ -364,7 +368,7 @@ class ProxyExtrasDBManager:
"--schema",
schema_path,
],
timeout=60,
timeout=prisma_command_timeout(),
check=True,
capture_output=True,
text=True,
@ -393,7 +397,7 @@ class ProxyExtrasDBManager:
"--applied",
migration_name,
],
timeout=60,
timeout=prisma_command_timeout(),
check=True,
capture_output=True,
text=True,
@ -530,7 +534,7 @@ class ProxyExtrasDBManager:
try:
subprocess.run(
[_get_prisma_command(), "db", "push", "--accept-data-loss"],
timeout=60,
timeout=prisma_command_timeout(),
check=True,
env=_get_prisma_env(),
)
@ -555,7 +559,7 @@ class ProxyExtrasDBManager:
try:
result = subprocess.run(
[_get_prisma_command(), "migrate", "deploy"],
timeout=60,
timeout=prisma_command_timeout(),
check=True,
capture_output=True,
text=True,
@ -731,6 +735,9 @@ class ProxyExtrasDBManager:
Returns:
bool: True if setup was successful, False otherwise
"""
ensure_prisma_toolchain(
prisma_command=_get_prisma_command(), prisma_env=_get_prisma_env()
)
migrated = ProxyExtrasDBManager._run_migrations(
use_migrate=use_migrate, use_v2_resolver=use_v2_resolver
)
@ -757,7 +764,7 @@ class ProxyExtrasDBManager:
# Set migrations directory for Prisma
result = subprocess.run(
[_get_prisma_command(), "migrate", "deploy"],
timeout=60,
timeout=prisma_command_timeout(),
check=True,
capture_output=True,
text=True,
@ -840,7 +847,7 @@ class ProxyExtrasDBManager:
"--rolled-back",
failed_migration,
],
timeout=60,
timeout=prisma_command_timeout(),
check=True,
capture_output=True,
text=True,
@ -968,7 +975,7 @@ class ProxyExtrasDBManager:
# Use prisma db push with increased timeout
subprocess.run(
[_get_prisma_command(), "db", "push", "--accept-data-loss"],
timeout=60,
timeout=prisma_command_timeout(),
check=True,
)
return True

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

@ -1,6 +1,6 @@
[project]
name = "litellm-proxy-extras"
version = "0.4.81"
version = "0.4.83"
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 = "0.4.81"
version = "0.4.83"
version_files = [
"pyproject.toml:^version",
"../pyproject.toml:litellm-proxy-extras==",

78
litellm/__init__.py
View file

@ -27,18 +27,19 @@ if os.getenv("LITELLM_MODE", "DEV") == "DEV":
_dotenv.load_dotenv(override=_dev_env_hot_reload_enabled())
from typing import (
Callable,
List,
Optional,
Dict,
Union,
Any,
Literal,
Callable,
Dict,
Final,
get_args,
TYPE_CHECKING,
Tuple,
List,
Literal,
Optional,
overload,
Tuple,
Type,
TYPE_CHECKING,
Union,
)
from litellm.types.integrations.datadog import DatadogInitParams
from litellm.types.integrations.newrelic import NewRelicInitParams
@ -264,6 +265,7 @@ 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
@ -680,12 +682,12 @@ def is_bedrock_pricing_only_model(key: str) -> bool:
bool: True if the key matches the Bedrock pattern, False otherwise.
"""
# Regex to match 'bedrock/<region>/<model>'
bedrock_pattern = re.compile(r"^bedrock/[a-zA-Z0-9_-]+/.+$")
bedrock_pattern: Final = re.compile(r"^bedrock/[a-zA-Z0-9_-]+/.+$")
if "month-commitment" in key:
return True
is_match = bedrock_pattern.match(key)
is_match: Final = bedrock_pattern.match(key)
return is_match is not None
@ -703,7 +705,7 @@ def is_openai_finetune_model(key: str) -> bool:
def add_known_models(model_cost_map: Optional[Dict] = None):
_map = model_cost_map if model_cost_map is not None else model_cost
_map: Final = model_cost_map if model_cost_map is not None else model_cost
for key, value in _map.items():
if value.get("litellm_provider") == "openai" and not is_openai_finetune_model(key):
open_ai_chat_completion_models.add(key)
@ -1267,8 +1269,8 @@ from .llms.xai.common_utils import XAIModelInfo
from litellm.types.utils import LlmProviders
## Lazy loading this is not straightforward, will leave it here for now.
from .main import * # type: ignore
from .compression import compress # type: ignore[no-redef]
from .main import *
from .compression import compress
# Skills API
from .skills.main import (
@ -1339,7 +1341,7 @@ from .assistants.main import *
from .batches.main import *
from .images.main import *
from .videos.main import *
from .batch_completion.main import * # type: ignore
from .batch_completion.main import *
from .rerank_api.main import *
from .llms.anthropic.experimental_pass_through.messages.handler import *
from .responses.main import *
@ -2052,7 +2054,7 @@ if TYPE_CHECKING:
supports_reasoning: Callable[..., bool]
acreate: Callable[..., Any]
get_max_tokens: Callable[..., int]
get_model_info: Callable[..., _ModelInfoType] # type: ignore[no-redef]
get_model_info: Callable[..., _ModelInfoType]
register_prompt_template: Callable[..., None]
validate_environment: Callable[..., dict]
check_valid_key: Callable[..., bool]
@ -2139,18 +2141,18 @@ def __getattr__(name: str) -> Any:
# Use cached registry from _lazy_imports instead of importing tuples every time
from ._lazy_imports import _get_lazy_import_registry
registry = _get_lazy_import_registry()
registry: Final = _get_lazy_import_registry()
# Check if name is in registry and call the cached handler function
if name in registry:
handler_func = registry[name]
handler_func: Final = registry[name]
return handler_func(name)
# Lazy load encoding from main.py to avoid heavy tiktoken import
if name == "encoding":
from ._lazy_imports import _get_litellm_globals
from ._lazy_imports import get_litellm_globals
_globals = _get_litellm_globals()
_globals = get_litellm_globals()
# Check if already cached
if "encoding" not in _globals:
from .main import encoding as _encoding
@ -2160,9 +2162,9 @@ def __getattr__(name: str) -> Any:
# Lazy load bedrock_tool_name_mappings instance
if name == "bedrock_tool_name_mappings":
from ._lazy_imports import _get_litellm_globals
from ._lazy_imports import get_litellm_globals
_globals = _get_litellm_globals()
_globals = get_litellm_globals()
# Check if already cached
if "bedrock_tool_name_mappings" not in _globals:
from .llms.bedrock.chat.invoke_handler import (
@ -2174,9 +2176,9 @@ def __getattr__(name: str) -> Any:
# Lazy load AzureOpenAIError exception class
if name == "AzureOpenAIError":
from ._lazy_imports import _get_litellm_globals
from ._lazy_imports import get_litellm_globals
_globals = _get_litellm_globals()
_globals = get_litellm_globals()
# Check if already cached
if "AzureOpenAIError" not in _globals:
from .llms.azure.common_utils import AzureOpenAIError as _AzureOpenAIError
@ -2186,9 +2188,9 @@ def __getattr__(name: str) -> Any:
# Lazy load openaiOSeriesConfig instance
if name == "openaiOSeriesConfig":
from ._lazy_imports import _get_litellm_globals
from ._lazy_imports import get_litellm_globals
_globals = _get_litellm_globals()
_globals = get_litellm_globals()
if "openaiOSeriesConfig" not in _globals:
# Import the config class and instantiate it
config_class = __getattr__("OpenAIOSeriesConfig")
@ -2196,7 +2198,7 @@ def __getattr__(name: str) -> Any:
return _globals["openaiOSeriesConfig"]
# Lazy load other config instances
_config_instances = {
_config_instances: Final = {
"openAIGPTConfig": "OpenAIGPTConfig",
"openAIGPTAudioConfig": "OpenAIGPTAudioConfig",
"openAIGPT5Config": "OpenAIGPT5Config",
@ -2204,9 +2206,9 @@ def __getattr__(name: str) -> Any:
"nvidiaNimEmbeddingConfig": "NvidiaNimEmbeddingConfig",
}
if name in _config_instances:
from ._lazy_imports import _get_litellm_globals
from ._lazy_imports import get_litellm_globals
_globals = _get_litellm_globals()
_globals = get_litellm_globals()
if name not in _globals:
# Import the config class and instantiate it
config_class = __getattr__(_config_instances[name])
@ -2219,9 +2221,9 @@ def __getattr__(name: str) -> Any:
# Lazy load provider_list
if name == "provider_list":
from ._lazy_imports import _get_litellm_globals
from ._lazy_imports import get_litellm_globals
_globals = _get_litellm_globals()
_globals = get_litellm_globals()
# Check if already cached
if "provider_list" not in _globals:
# LlmProviders is eagerly imported above, so we can import it directly
@ -2232,33 +2234,33 @@ def __getattr__(name: str) -> Any:
# Lazy load priority_reservation_settings instance
if name == "priority_reservation_settings":
from ._lazy_imports import _get_litellm_globals
from ._lazy_imports import get_litellm_globals
_globals = _get_litellm_globals()
_globals = get_litellm_globals()
# Check if already cached
if "priority_reservation_settings" not in _globals:
# Import the class and instantiate it
PriorityReservationSettings = __getattr__("PriorityReservationSettings")
PriorityReservationSettings: Final = __getattr__("PriorityReservationSettings")
_globals["priority_reservation_settings"] = PriorityReservationSettings()
return _globals["priority_reservation_settings"]
# Lazy load logging_callback_manager instance
if name == "logging_callback_manager":
from ._lazy_imports import _get_litellm_globals
from ._lazy_imports import get_litellm_globals
_globals = _get_litellm_globals()
_globals = get_litellm_globals()
# Check if already cached
if "logging_callback_manager" not in _globals:
# Import the class and instantiate it
LoggingCallbackManager = __getattr__("LoggingCallbackManager")
LoggingCallbackManager: Final = __getattr__("LoggingCallbackManager")
_globals["logging_callback_manager"] = LoggingCallbackManager()
return _globals["logging_callback_manager"]
# Lazy load _service_logger module
if name == "_service_logger":
from ._lazy_imports import _get_litellm_globals
from ._lazy_imports import get_litellm_globals
_globals = _get_litellm_globals()
_globals = get_litellm_globals()
# Check if already cached
if "_service_logger" not in _globals:
# Import the module lazily

3
litellm/_internal_context.py
View file

@ -7,7 +7,8 @@ asyncio task and cannot be injected via HTTP request bodies.
"""
from contextvars import ContextVar
from typing import Final
# When True, suppresses async logging and billing for internal sub-calls
# (e.g., emulated file-search steps that make nested LLM calls).
is_internal_call: ContextVar[bool] = ContextVar("is_internal_call", default=False)
is_internal_call: Final[ContextVar[bool]] = ContextVar("is_internal_call", default=False)

32
litellm/_lazy_imports.py
View file

@ -18,7 +18,7 @@ until they're actually needed.
import importlib
import sys
from collections.abc import Callable
from typing import Any, cast
from typing import Any, Final, cast
# 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
@ -54,7 +54,7 @@ from ._lazy_imports_registry import (
)
def _get_litellm_globals() -> dict:
def get_litellm_globals() -> dict:
"""
Get the globals dictionary of the litellm module.
@ -233,7 +233,7 @@ def _generic_lazy_import(name: str, import_map: dict[str, tuple[str, str]], cate
raise AttributeError(f"{category} lazy import: unknown attribute {name!r}")
# Step 2: Get the cache (where we store imported things)
_globals = _get_litellm_globals()
_globals: Final = get_litellm_globals()
# Step 3: If we've already imported it, just return the cached version
if name in _globals:
@ -255,7 +255,7 @@ 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 = getattr(module, attr_name)
value: Final = getattr(module, attr_name)
# Step 7: Cache it so we don't have to import again next time
_globals[name] = value
@ -332,14 +332,14 @@ def _lazy_import_utils_module(name: str) -> Any:
Handler for utils module lazy imports.
This uses a custom implementation because utils module needs to use
_get_utils_globals() instead of _get_litellm_globals() for caching.
_get_utils_globals() instead of get_litellm_globals() for caching.
"""
# Check if this attribute exists in our map
if name not in _UTILS_MODULE_IMPORT_MAP:
raise AttributeError(f"Utils module lazy import: unknown attribute {name!r}")
# Get the cache (where we store imported things) - use utils globals
_globals = _get_utils_globals()
_globals: Final = _get_utils_globals()
# If we've already imported it, just return the cached version
if name in _globals:
@ -355,7 +355,7 @@ def _lazy_import_utils_module(name: str) -> Any:
module = importlib.import_module(module_path)
# Get the actual attribute from the module
value = getattr(module, attr_name)
value: Final = getattr(module, attr_name)
# Cache it so we don't have to import again next time
_globals[name] = value
@ -379,15 +379,15 @@ def _lazy_import_llm_client_cache(name: str) -> Any:
- "in_memory_llm_clients_cache" is a singleton instance of that class
So we need custom logic to handle both cases.
"""
_globals = _get_litellm_globals()
_globals: Final = get_litellm_globals()
# If already cached, return it
if name in _globals:
return _globals[name]
# Import the class
module = importlib.import_module("litellm.caching.llm_caching_handler")
LLMClientCache = getattr(module, "LLMClientCache")
module: Final = importlib.import_module("litellm.caching.llm_caching_handler")
LLMClientCache: Final = getattr(module, "LLMClientCache")
# If they want the class itself, return it
if name == "LLMClientCache":
@ -396,7 +396,7 @@ def _lazy_import_llm_client_cache(name: str) -> Any:
# If they want the singleton instance, create it (only once)
if name == "in_memory_llm_clients_cache":
instance = LLMClientCache()
instance: Final = LLMClientCache()
_globals["in_memory_llm_clients_cache"] = instance
return instance
@ -412,7 +412,7 @@ def _lazy_import_http_handlers(name: str) -> Any:
- They need configuration (timeout, etc.) from the module globals
- They use factory functions instead of direct instantiation
"""
_globals = _get_litellm_globals()
_globals: Final = get_litellm_globals()
if name == "module_level_aclient":
# Create an async HTTP client using the factory function
@ -420,11 +420,11 @@ def _lazy_import_http_handlers(name: str) -> Any:
# Get timeout from module config (if set)
timeout = _globals.get("request_timeout")
params = {"timeout": timeout, "client_alias": "module level aclient"}
params: Final = {"timeout": timeout, "client_alias": "module level aclient"}
# Create the client instance
provider_id = cast(Any, "litellm_module_level_client")
async_client = get_async_httpx_client(
provider_id: Final = cast(Any, "litellm_module_level_client")
async_client: Final = get_async_httpx_client(
llm_provider=provider_id,
params=params,
)
@ -438,7 +438,7 @@ def _lazy_import_http_handlers(name: str) -> Any:
from litellm.llms.custom_httpx.http_handler import HTTPHandler
timeout = _globals.get("request_timeout")
sync_client = HTTPHandler(timeout=timeout)
sync_client: Final = HTTPHandler(timeout=timeout)
# Cache it
_globals["module_level_client"] = sync_client

54
litellm/_lazy_imports_registry.py
View file

@ -5,21 +5,23 @@ This module contains all the name tuples and import maps used by the lazy import
Separated from the handler functions for better organization.
"""
from typing import Final
# Cost calculator names that support lazy loading via _lazy_import_cost_calculator
COST_CALCULATOR_NAMES = (
COST_CALCULATOR_NAMES: Final = (
"completion_cost",
"cost_per_token",
"response_cost_calculator",
)
# Litellm logging names that support lazy loading via _lazy_import_litellm_logging
LITELLM_LOGGING_NAMES = (
LITELLM_LOGGING_NAMES: Final = (
"Logging",
"modify_integration",
)
# Utils names that support lazy loading via _lazy_import_utils
UTILS_NAMES = (
UTILS_NAMES: Final = (
"exception_type",
"get_optional_params",
"get_response_string",
@ -66,20 +68,20 @@ UTILS_NAMES = (
)
# Token counter names that support lazy loading via _lazy_import_token_counter
TOKEN_COUNTER_NAMES = ("get_modified_max_tokens",)
TOKEN_COUNTER_NAMES: Final = ("get_modified_max_tokens",)
# LLM client cache names that support lazy loading via _lazy_import_llm_client_cache
LLM_CLIENT_CACHE_NAMES = (
LLM_CLIENT_CACHE_NAMES: Final = (
"LLMClientCache",
"in_memory_llm_clients_cache",
)
# Bedrock type names that support lazy loading via _lazy_import_bedrock_types
BEDROCK_TYPES_NAMES = ("COHERE_EMBEDDING_INPUT_TYPES",)
BEDROCK_TYPES_NAMES: Final = ("COHERE_EMBEDDING_INPUT_TYPES",)
# Common types from litellm.types.utils that support lazy loading via
# _lazy_import_types_utils
TYPES_UTILS_NAMES = (
TYPES_UTILS_NAMES: Final = (
"ImageObject",
"BudgetConfig",
"all_litellm_params",
@ -92,7 +94,7 @@ TYPES_UTILS_NAMES = (
)
# Caching / cache classes that support lazy loading via _lazy_import_caching
CACHING_NAMES = (
CACHING_NAMES: Final = (
"Cache",
"DualCache",
"RedisCache",
@ -100,20 +102,20 @@ CACHING_NAMES = (
)
# HTTP handler names that support lazy loading via _lazy_import_http_handlers
HTTP_HANDLER_NAMES = (
HTTP_HANDLER_NAMES: Final = (
"module_level_aclient",
"module_level_client",
)
# Dotprompt integration names that support lazy loading via _lazy_import_dotprompt
DOTPROMPT_NAMES = (
DOTPROMPT_NAMES: Final = (
"global_prompt_manager",
"global_prompt_directory",
"set_global_prompt_directory",
)
# LLM config classes that support lazy loading via _lazy_import_llm_configs
LLM_CONFIG_NAMES = (
LLM_CONFIG_NAMES: Final = (
"AmazonConverseConfig",
"OpenAILikeChatConfig",
"GaladrielChatConfig",
@ -328,7 +330,7 @@ LLM_CONFIG_NAMES = (
)
# Types that support lazy loading via _lazy_import_types
TYPES_NAMES = (
TYPES_NAMES: Final = (
"GuardrailItem",
"DefaultTeamSSOParams",
"LiteLLM_UpperboundKeyGenerateParams",
@ -344,14 +346,14 @@ TYPES_NAMES = (
)
# LLM provider logic names that support lazy loading via _lazy_import_llm_provider_logic
LLM_PROVIDER_LOGIC_NAMES = (
LLM_PROVIDER_LOGIC_NAMES: Final = (
"get_llm_provider",
"remove_index_from_tool_calls",
)
# Utils module names that support lazy loading via _lazy_import_utils_module
# These are attributes accessed from litellm.utils module
UTILS_MODULE_NAMES = (
UTILS_MODULE_NAMES: Final = (
"encoding",
"BaseVectorStore",
"CredentialAccessor",
@ -423,7 +425,7 @@ UTILS_MODULE_NAMES = (
)
# Import maps for registry pattern - reduces repetition
_UTILS_IMPORT_MAP = {
_UTILS_IMPORT_MAP: Final = {
"exception_type": (".utils", "exception_type"),
"get_optional_params": (".utils", "get_optional_params"),
"get_response_string": (".utils", "get_response_string"),
@ -478,13 +480,13 @@ _UTILS_IMPORT_MAP = {
),
}
_COST_CALCULATOR_IMPORT_MAP = {
_COST_CALCULATOR_IMPORT_MAP: Final = {
"completion_cost": (".cost_calculator", "completion_cost"),
"cost_per_token": (".cost_calculator", "cost_per_token"),
"response_cost_calculator": (".cost_calculator", "response_cost_calculator"),
}
_TYPES_UTILS_IMPORT_MAP = {
_TYPES_UTILS_IMPORT_MAP: Final = {
"ImageObject": (".types.utils", "ImageObject"),
"BudgetConfig": (".types.utils", "BudgetConfig"),
"all_litellm_params": (".types.utils", "all_litellm_params"),
@ -496,28 +498,28 @@ _TYPES_UTILS_IMPORT_MAP = {
"GenericStreamingChunk": (".types.utils", "GenericStreamingChunk"),
}
_TOKEN_COUNTER_IMPORT_MAP = {
_TOKEN_COUNTER_IMPORT_MAP: Final = {
"get_modified_max_tokens": (
"litellm.litellm_core_utils.token_counter",
"get_modified_max_tokens",
),
}
_BEDROCK_TYPES_IMPORT_MAP = {
_BEDROCK_TYPES_IMPORT_MAP: Final = {
"COHERE_EMBEDDING_INPUT_TYPES": (
"litellm.types.llms.bedrock",
"COHERE_EMBEDDING_INPUT_TYPES",
),
}
_CACHING_IMPORT_MAP = {
_CACHING_IMPORT_MAP: Final = {
"Cache": ("litellm.caching.caching", "Cache"),
"DualCache": ("litellm.caching.caching", "DualCache"),
"RedisCache": ("litellm.caching.caching", "RedisCache"),
"InMemoryCache": ("litellm.caching.caching", "InMemoryCache"),
}
_LITELLM_LOGGING_IMPORT_MAP = {
_LITELLM_LOGGING_IMPORT_MAP: Final = {
"Logging": ("litellm.litellm_core_utils.litellm_logging", "Logging"),
"modify_integration": (
"litellm.litellm_core_utils.litellm_logging",
@ -525,7 +527,7 @@ _LITELLM_LOGGING_IMPORT_MAP = {
),
}
_DOTPROMPT_IMPORT_MAP = {
_DOTPROMPT_IMPORT_MAP: Final = {
"global_prompt_manager": (
"litellm.integrations.dotprompt",
"global_prompt_manager",
@ -540,7 +542,7 @@ _DOTPROMPT_IMPORT_MAP = {
),
}
_TYPES_IMPORT_MAP = {
_TYPES_IMPORT_MAP: Final = {
"GuardrailItem": ("litellm.types.guardrails", "GuardrailItem"),
"DefaultTeamSSOParams": (
"litellm.types.proxy.management_endpoints.ui_sso",
@ -569,7 +571,7 @@ _TYPES_IMPORT_MAP = {
),
}
_LLM_PROVIDER_LOGIC_IMPORT_MAP = {
_LLM_PROVIDER_LOGIC_IMPORT_MAP: Final = {
"get_llm_provider": (
"litellm.litellm_core_utils.get_llm_provider_logic",
"get_llm_provider",
@ -580,7 +582,7 @@ _LLM_PROVIDER_LOGIC_IMPORT_MAP = {
),
}
_LLM_CONFIGS_IMPORT_MAP = {
_LLM_CONFIGS_IMPORT_MAP: Final = {
"AmazonConverseConfig": (
".llms.bedrock.chat.converse_transformation",
"AmazonConverseConfig",
@ -1215,7 +1217,7 @@ _LLM_CONFIGS_IMPORT_MAP = {
}
# Import map for utils module lazy imports
_UTILS_MODULE_IMPORT_MAP = {
_UTILS_MODULE_IMPORT_MAP: Final = {
"encoding": ("litellm.main", "encoding"),
"BaseVectorStore": (
"litellm.integrations.vector_store_integrations.base_vector_store",

58
litellm/_logging.py
View file

@ -4,7 +4,7 @@ import os
import sys
from datetime import datetime
from logging import Formatter
from typing import Any
from typing import Any, Final
from litellm.litellm_core_utils.safe_json_dumps import safe_dumps
from litellm.litellm_core_utils.safe_json_loads import safe_json_loads
@ -17,7 +17,7 @@ if set_verbose is True:
"`litellm.set_verbose` is deprecated. Please set `os.environ['LITELLM_LOG'] = 'DEBUG'` for debug logs."
)
_ENABLE_SECRET_REDACTION = os.getenv("LITELLM_DISABLE_REDACT_SECRETS", "").lower() != "true"
_ENABLE_SECRET_REDACTION: Final = os.getenv("LITELLM_DISABLE_REDACT_SECRETS", "").lower() != "true"
def _redact_string(value: str) -> str:
@ -74,14 +74,14 @@ class SecretRedactionFilter(logging.Filter):
return True
_secret_filter = SecretRedactionFilter()
_secret_filter: Final = SecretRedactionFilter()
json_logs = bool(os.getenv("JSON_LOGS", False))
# Create a handler for the logger (you may need to adapt this based on your needs)
log_level = os.getenv("LITELLM_LOG", "DEBUG")
numeric_level: str = getattr(logging, log_level.upper())
handler = logging.StreamHandler()
log_level: Final = os.getenv("LITELLM_LOG", "DEBUG")
numeric_level: Final[str] = getattr(logging, log_level.upper())
handler: Final = logging.StreamHandler()
handler.setLevel(numeric_level)
handler.addFilter(_secret_filter)
@ -94,10 +94,10 @@ def _try_parse_json_message(message: str) -> dict[str, Any] | None:
"""
if not message or not isinstance(message, str):
return None
msg_stripped = message.strip()
msg_stripped: Final = message.strip()
if not (msg_stripped.startswith("{") or msg_stripped.startswith("[")):
return None
parsed = safe_json_loads(message, default=None)
parsed: Final = safe_json_loads(message, default=None)
if parsed is None or not isinstance(parsed, dict):
return None
return parsed
@ -144,7 +144,7 @@ def _get_standard_record_attrs() -> frozenset:
return frozenset(logging.LogRecord("", 0, "", 0, "", (), None).__dict__.keys())
_STANDARD_RECORD_ATTRS = _get_standard_record_attrs()
_STANDARD_RECORD_ATTRS: Final = _get_standard_record_attrs()
class JsonFormatter(Formatter):
@ -153,12 +153,12 @@ class JsonFormatter(Formatter):
def formatTime(self, record, datefmt=None):
# Use datetime to format the timestamp in ISO 8601 format
dt = datetime.fromtimestamp(record.created)
dt: Final = datetime.fromtimestamp(record.created)
return dt.isoformat()
def format(self, record):
message_str = record.getMessage()
json_record: dict[str, Any] = {
message_str: Final = record.getMessage()
json_record: Final[dict[str, Any]] = {
"message": message_str,
"level": record.levelname,
"timestamp": self.formatTime(record),
@ -193,13 +193,13 @@ class JsonFormatter(Formatter):
# Function to set up exception handlers for JSON logging
def _setup_json_exception_handlers(formatter):
# Create a handler with JSON formatting for exceptions
error_handler = logging.StreamHandler()
error_handler: Final = logging.StreamHandler()
error_handler.setFormatter(formatter)
error_handler.addFilter(_secret_filter)
# Setup excepthook for uncaught exceptions
def json_excepthook(exc_type, exc_value, exc_traceback):
record = logging.LogRecord(
record: Final = logging.LogRecord(
name="LiteLLM",
level=logging.ERROR,
pathname="",
@ -217,10 +217,10 @@ def _setup_json_exception_handlers(formatter):
import asyncio
def async_json_exception_handler(loop, context):
exception = context.get("exception")
exception: Final = context.get("exception")
if exception:
exc_type = type(exception)
record = logging.LogRecord(
exc_type: Final = type(exception)
record: Final = logging.LogRecord(
name="LiteLLM",
level=logging.ERROR,
pathname="",
@ -243,7 +243,7 @@ if json_logs:
handler.setFormatter(JsonFormatter())
_setup_json_exception_handlers(JsonFormatter())
else:
formatter = logging.Formatter(
formatter: Final = logging.Formatter(
"\033[92m%(asctime)s - %(name)s:%(levelname)s\033[0m: %(filename)s:%(lineno)s - %(message)s",
datefmt="%H:%M:%S",
)
@ -263,20 +263,20 @@ verbose_logger.addHandler(handler)
def _suppress_loggers():
"""Suppress noisy loggers at INFO level"""
# Suppress httpx request logging at INFO level
httpx_logger = logging.getLogger("httpx")
httpx_logger: Final = logging.getLogger("httpx")
httpx_logger.setLevel(logging.WARNING)
# Suppress APScheduler logging at INFO level
apscheduler_executors_logger = logging.getLogger("apscheduler.executors.default")
apscheduler_executors_logger: Final = logging.getLogger("apscheduler.executors.default")
apscheduler_executors_logger.setLevel(logging.WARNING)
apscheduler_scheduler_logger = logging.getLogger("apscheduler.scheduler")
apscheduler_scheduler_logger: Final = logging.getLogger("apscheduler.scheduler")
apscheduler_scheduler_logger.setLevel(logging.WARNING)
# Call the suppression function
_suppress_loggers()
ALL_LOGGERS = [
ALL_LOGGERS: Final = [
logging.getLogger(),
verbose_logger,
verbose_router_logger,
@ -293,11 +293,11 @@ def _get_loggers_to_initialize():
"""
import litellm
loggers = list(ALL_LOGGERS)
loggers: Final = list(ALL_LOGGERS)
# Add langfuse logger if langfuse is being used as a callback
langfuse_callbacks = {"langfuse", "langfuse_otel"}
all_callbacks = set(litellm.success_callback + litellm.failure_callback)
langfuse_callbacks: Final = {"langfuse", "langfuse_otel"}
all_callbacks: Final = set(litellm.success_callback + litellm.failure_callback)
if langfuse_callbacks & all_callbacks:
loggers.append(logging.getLogger("langfuse"))
@ -325,12 +325,12 @@ def _get_uvicorn_json_log_config():
This ensures that uvicorn's access logs, error logs, and all application logs
are formatted as JSON when json_logs is enabled.
"""
json_formatter_class = "litellm._logging.JsonFormatter"
json_formatter_class: Final = "litellm._logging.JsonFormatter"
# Use the module-level log_level variable for consistency
uvicorn_log_level = log_level.upper()
uvicorn_log_level: Final = log_level.upper()
log_config = {
log_config: Final = {
"version": 1,
"disable_existing_loggers": False,
"formatters": {
@ -384,7 +384,7 @@ def _turn_on_json():
- Adds a JSON formatter to all loggers
"""
handler = logging.StreamHandler()
handler: Final = logging.StreamHandler()
handler.setFormatter(JsonFormatter())
_initialize_loggers_with_handler(handler)
# Set up exception handlers

167
litellm/_redis.py
View file

@ -13,9 +13,10 @@ 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 typing import Final
import redis # type: ignore
import redis.asyncio as async_redis # type: ignore
import redis
import redis.asyncio as async_redis
from litellm import get_secret, get_secret_str
from litellm._redis_credential_provider import (
@ -32,20 +33,20 @@ from litellm.litellm_core_utils.sensitive_data_masker import SensitiveDataMasker
from ._logging import verbose_logger
AZURE_REDIS_SCOPE = "https://redis.azure.com/.default"
AZURE_REDIS_SCOPE: Final = "https://redis.azure.com/.default"
def _get_redis_kwargs():
arg_spec = inspect.getfullargspec(redis.Redis)
arg_spec: Final = inspect.getfullargspec(redis.Redis)
# Only allow primitive arguments
exclude_args = {
exclude_args: Final = {
"self",
"connection_pool",
"retry",
}
include_args = {
include_args: Final = {
"url",
"redis_connect_func",
"gcp_service_account",
@ -56,7 +57,7 @@ def _get_redis_kwargs():
"azure_client_secret",
}
available_args = {x for x in arg_spec.args if x not in exclude_args} | include_args
available_args: Final = {x for x in arg_spec.args if x not in exclude_args} | include_args
return available_args
@ -92,9 +93,9 @@ def _get_redis_url_kwargs(client: type | None = None) -> tuple[str, ...]:
"""
if client is None:
client = redis.Redis
connection_cls = async_redis.Connection if client is async_redis.Redis else redis.Connection
connection_cls: Final = async_redis.Connection if client is async_redis.Redis else redis.Connection
exclude_args = frozenset(
exclude_args: Final = frozenset(
{
"self",
"connection_pool",
@ -103,7 +104,7 @@ def _get_redis_url_kwargs(client: type | None = None) -> tuple[str, ...]:
)
# Only allow primitive arguments
include_args = ("url", "max_connections")
include_args: Final = ("url", "max_connections")
return tuple(x for x in _init_arg_names(connection_cls) if x not in exclude_args) + include_args
@ -111,10 +112,10 @@ def _get_redis_url_kwargs(client: type | None = None) -> tuple[str, ...]:
def _get_redis_cluster_kwargs(client=None):
if client is None:
client = redis.Redis.from_url
arg_spec = inspect.getfullargspec(redis.RedisCluster)
arg_spec: Final = inspect.getfullargspec(redis.RedisCluster)
# Only allow primitive arguments
exclude_args = {"self", "connection_pool", "retry", "host", "port", "startup_nodes"}
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 |= {
@ -142,17 +143,17 @@ def _get_redis_cluster_kwargs(client=None):
def _get_redis_env_kwarg_mapping():
PREFIX = "REDIS_"
PREFIX: Final = "REDIS_"
return {f"{PREFIX}{x.upper()}": x for x in _get_redis_kwargs()}
def _redis_kwargs_from_environment():
mapping = _get_redis_env_kwarg_mapping()
mapping: Final = _get_redis_env_kwarg_mapping()
return_dict = {}
return_dict: Final = {}
for k, v in mapping.items():
value = get_secret(k, default_value=None) # type: ignore
value = get_secret(k, default_value=None)
if value is not None:
return_dict[v] = value
return return_dict
@ -183,7 +184,7 @@ def create_gcp_iam_redis_connect_func(
self._parser.on_connect(self)
auth_args = (_generate_gcp_iam_access_token(service_account),)
auth_args: Final = (_generate_gcp_iam_access_token(service_account),)
self.send_command("AUTH", *auth_args, check_health=False)
try:
@ -224,9 +225,9 @@ def _build_azure_credential(
"azure-identity is required for Azure AD Redis authentication. Install it with: pip install azure-identity"
)
_client_id = azure_client_id or os.environ.get("AZURE_CLIENT_ID")
_tenant_id = azure_tenant_id or os.environ.get("AZURE_TENANT_ID")
_client_secret = azure_client_secret or os.environ.get("AZURE_CLIENT_SECRET")
_client_id: Final = azure_client_id or os.environ.get("AZURE_CLIENT_ID")
_tenant_id: Final = azure_tenant_id or os.environ.get("AZURE_TENANT_ID")
_client_secret: Final = azure_client_secret or os.environ.get("AZURE_CLIENT_SECRET")
if _client_id and _tenant_id and _client_secret:
return ClientSecretCredential(
@ -253,12 +254,12 @@ def _generate_azure_ad_redis_token(
(``AzureADCredentialProvider``) keep the credential alive across
connections so the Azure SDK's internal cache + silent refresh apply.
"""
credential = _build_azure_credential(
credential: Final = _build_azure_credential(
azure_client_id=azure_client_id,
azure_tenant_id=azure_tenant_id,
azure_client_secret=azure_client_secret,
)
token = credential.get_token(AZURE_REDIS_SCOPE)
token: Final = credential.get_token(AZURE_REDIS_SCOPE)
return token.token
@ -274,7 +275,7 @@ def create_azure_ad_redis_connect_func(
closure) and reused across connections — the Azure SDK handles token caching
and silent renewal internally. Only ``get_token`` is called per connection.
"""
credential = _build_azure_credential(
credential: Final = _build_azure_credential(
azure_client_id=azure_client_id,
azure_tenant_id=azure_tenant_id,
azure_client_secret=azure_client_secret,
@ -290,11 +291,11 @@ def create_azure_ad_redis_connect_func(
self._parser.on_connect(self)
access_token = credential.get_token(AZURE_REDIS_SCOPE).token
access_token: Final = credential.get_token(AZURE_REDIS_SCOPE).token
# Only include username when explicitly set — sending AUTH "" <token>
# is invalid for most ACL-configured Azure Redis instances.
username = os.environ.get("REDIS_USERNAME", "")
username: Final = os.environ.get("REDIS_USERNAME", "")
if username:
auth_args = (username, access_token)
else:
@ -316,7 +317,7 @@ def create_azure_ad_redis_connect_func(
# AzureADCredentialProvider for refresh-aware token retrieval. The raw
# client_id/tenant_id/secret are intentionally NOT exposed here — the
# credential closure already holds them.
ad_connect._azure_credential = credential # type: ignore[attr-defined]
ad_connect._azure_credential = credential
return ad_connect
@ -350,26 +351,26 @@ def _get_redis_client_logic(**env_overrides):
for k, v in env_overrides.items():
if isinstance(v, str) and v.startswith("os.environ/"):
v = v.replace("os.environ/", "")
value = get_secret(v) # type: ignore
value = get_secret(v)
env_overrides[k] = value
environment_kwargs = _redis_kwargs_from_environment()
environment_kwargs: Final = _redis_kwargs_from_environment()
# An explicitly configured connection target outranks REDIS_URL from the
# environment. Without this, the url branch below strips the caller's
# host/port/password and silently connects to whatever REDIS_URL names.
caller_named_a_target = any(
caller_named_a_target: Final = any(
env_overrides.get(key) is not None for key in ("host", "startup_nodes", "sentinel_nodes")
)
if caller_named_a_target and env_overrides.get("url") is None:
environment_kwargs.pop("url", None)
redis_kwargs = {
redis_kwargs: Final = {
**environment_kwargs,
**env_overrides,
}
_startup_nodes: str | list | None = redis_kwargs.get("startup_nodes", None) or get_secret( # type: ignore
_startup_nodes: Final[str | list | None] = redis_kwargs.get("startup_nodes", None) or get_secret(
"REDIS_CLUSTER_NODES"
)
@ -380,21 +381,21 @@ def _get_redis_client_logic(**env_overrides):
elif _startup_nodes is None:
redis_kwargs.pop("startup_nodes", None)
_sentinel_nodes: str | list | None = redis_kwargs.get("sentinel_nodes", None) or get_secret( # type: ignore
_sentinel_nodes: Final[str | list | None] = redis_kwargs.get("sentinel_nodes", None) or get_secret(
"REDIS_SENTINEL_NODES"
)
if _sentinel_nodes is not None and isinstance(_sentinel_nodes, str):
redis_kwargs["sentinel_nodes"] = json.loads(_sentinel_nodes)
_sentinel_password: str | None = redis_kwargs.get("sentinel_password", None) or get_secret_str(
_sentinel_password: Final[str | None] = redis_kwargs.get("sentinel_password", None) or get_secret_str(
"REDIS_SENTINEL_PASSWORD"
)
if _sentinel_password is not None:
redis_kwargs["sentinel_password"] = _sentinel_password
_service_name: str | None = redis_kwargs.get("service_name", None) or get_secret( # type: ignore
_service_name: Final[str | None] = redis_kwargs.get("service_name", None) or get_secret(
"REDIS_SERVICE_NAME"
)
@ -402,8 +403,8 @@ def _get_redis_client_logic(**env_overrides):
redis_kwargs["service_name"] = _service_name
# Handle GCP IAM authentication
_gcp_service_account = redis_kwargs.get("gcp_service_account") or get_secret_str("REDIS_GCP_SERVICE_ACCOUNT")
_gcp_ssl_ca_certs = redis_kwargs.get("gcp_ssl_ca_certs") or get_secret_str("REDIS_GCP_SSL_CA_CERTS")
_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.")
@ -411,7 +412,7 @@ def _get_redis_client_logic(**env_overrides):
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 # type: ignore[attr-defined]
redis_kwargs["redis_connect_func"]._gcp_service_account = _gcp_service_account
# Remove GCP-specific kwargs that shouldn't be passed to Redis client
redis_kwargs.pop("gcp_service_account", None)
@ -422,9 +423,9 @@ def _get_redis_client_logic(**env_overrides):
redis_kwargs["ssl_ca_certs"] = _gcp_ssl_ca_certs
# Handle Azure AD authentication (after GCP IAM block)
_azure_redis_ad_token = redis_kwargs.get("azure_redis_ad_token") or get_secret("REDIS_AZURE_AD_TOKEN")
_azure_redis_ad_token: Final = redis_kwargs.get("azure_redis_ad_token") or get_secret("REDIS_AZURE_AD_TOKEN")
_azure_ad_enabled = _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(
@ -433,9 +434,9 @@ def _get_redis_client_logic(**env_overrides):
)
if _azure_ad_enabled and _gcp_service_account is None:
_azure_client_id = redis_kwargs.get("azure_client_id") or get_secret_str("AZURE_CLIENT_ID")
_azure_tenant_id = redis_kwargs.get("azure_tenant_id") or get_secret_str("AZURE_TENANT_ID")
_azure_client_secret = redis_kwargs.get("azure_client_secret") or get_secret_str("AZURE_CLIENT_SECRET")
_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(
@ -448,7 +449,7 @@ def _get_redis_client_logic(**env_overrides):
# `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 # type: ignore[attr-defined]
redis_kwargs["redis_connect_func"]._azure_redis_ad_token = True
# Always remove Azure-specific kwargs that shouldn't be passed to Redis client
redis_kwargs.pop("azure_redis_ad_token", None)
@ -480,7 +481,7 @@ def _get_redis_client_logic(**env_overrides):
def init_redis_cluster(redis_kwargs) -> redis.RedisCluster:
_redis_cluster_nodes_in_env: str | None = get_secret("REDIS_CLUSTER_NODES") # type: ignore
_redis_cluster_nodes_in_env: Final[str | None] = get_secret("REDIS_CLUSTER_NODES")
if _redis_cluster_nodes_in_env is not None:
try:
redis_kwargs["startup_nodes"] = json.loads(_redis_cluster_nodes_in_env)
@ -492,24 +493,24 @@ def init_redis_cluster(redis_kwargs) -> redis.RedisCluster:
verbose_logger.debug("init_redis_cluster: startup nodes are being initialized.")
from redis.cluster import ClusterNode
args = _get_redis_cluster_kwargs()
cluster_kwargs = {}
args: Final = _get_redis_cluster_kwargs()
cluster_kwargs: Final = {}
for arg in redis_kwargs:
if arg in args:
cluster_kwargs[arg] = redis_kwargs[arg]
new_startup_nodes: list[ClusterNode] = []
new_startup_nodes: Final[list[ClusterNode]] = []
for item in redis_kwargs["startup_nodes"]:
new_startup_nodes.append(ClusterNode(**item))
cluster_kwargs.pop("startup_nodes", None)
return redis.RedisCluster(startup_nodes=new_startup_nodes, **cluster_kwargs) # type: ignore
return redis.RedisCluster(startup_nodes=new_startup_nodes, **cluster_kwargs)
def _get_redis_sentinel_connection_kwargs(redis_kwargs: dict) -> dict:
connection_kwargs = {}
args = _get_redis_kwargs()
connection_kwargs: Final = {}
args: Final = _get_redis_kwargs()
for arg in redis_kwargs:
if arg in args:
connection_kwargs[arg] = redis_kwargs[arg]
@ -518,12 +519,12 @@ def _get_redis_sentinel_connection_kwargs(redis_kwargs: dict) -> dict:
def _init_redis_sentinel(redis_kwargs) -> redis.Redis:
sentinel_nodes = redis_kwargs.get("sentinel_nodes")
sentinel_password = redis_kwargs.get("sentinel_password")
service_name = redis_kwargs.get("service_name")
connection_kwargs = _get_redis_sentinel_connection_kwargs(redis_kwargs)
sentinel_nodes: Final = redis_kwargs.get("sentinel_nodes")
sentinel_password: Final = redis_kwargs.get("sentinel_password")
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 = dict(connection_kwargs)
sentinel_kwargs: Final = dict(connection_kwargs)
sentinel_kwargs["password"] = sentinel_password
if not sentinel_nodes or not service_name:
@ -532,7 +533,7 @@ def _init_redis_sentinel(redis_kwargs) -> redis.Redis:
verbose_logger.debug("init_redis_sentinel: sentinel nodes are being initialized.")
# Set up the Sentinel client
sentinel = redis.Sentinel(
sentinel: Final = redis.Sentinel(
sentinel_nodes,
sentinel_kwargs=sentinel_kwargs,
)
@ -543,12 +544,12 @@ def _init_redis_sentinel(redis_kwargs) -> redis.Redis:
def _init_async_redis_sentinel(redis_kwargs) -> async_redis.Redis:
sentinel_nodes = redis_kwargs.get("sentinel_nodes")
sentinel_password = redis_kwargs.get("sentinel_password")
service_name = redis_kwargs.get("service_name")
connection_kwargs = _get_redis_sentinel_connection_kwargs(redis_kwargs)
sentinel_nodes: Final = redis_kwargs.get("sentinel_nodes")
sentinel_password: Final = redis_kwargs.get("sentinel_password")
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 = dict(connection_kwargs)
sentinel_kwargs: Final = dict(connection_kwargs)
sentinel_kwargs["password"] = sentinel_password
if not sentinel_nodes or not service_name:
@ -557,7 +558,7 @@ def _init_async_redis_sentinel(redis_kwargs) -> async_redis.Redis:
verbose_logger.debug("init_redis_sentinel: sentinel nodes are being initialized.")
# Set up the Sentinel client
sentinel = async_redis.Sentinel(
sentinel: Final = async_redis.Sentinel(
sentinel_nodes,
sentinel_kwargs=sentinel_kwargs,
)
@ -568,14 +569,14 @@ def _init_async_redis_sentinel(redis_kwargs) -> async_redis.Redis:
def get_redis_client(**env_overrides):
redis_kwargs = _get_redis_client_logic(**env_overrides)
redis_kwargs: Final = _get_redis_client_logic(**env_overrides)
if "startup_nodes" in redis_kwargs:
return init_redis_cluster(redis_kwargs)
if "url" in redis_kwargs and redis_kwargs["url"] is not None:
args = _get_redis_url_kwargs()
url_kwargs = {}
args: Final = _get_redis_url_kwargs()
url_kwargs: Final = {}
for arg in redis_kwargs:
if arg in args:
url_kwargs[arg] = redis_kwargs[arg]
@ -593,13 +594,13 @@ def get_redis_async_client(
connection_pool: async_redis.BlockingConnectionPool | None = None,
**env_overrides,
) -> async_redis.Redis | async_redis.RedisCluster:
redis_kwargs = _get_redis_client_logic(**env_overrides)
redis_kwargs: Final = _get_redis_client_logic(**env_overrides)
if "startup_nodes" in redis_kwargs:
from redis.cluster import ClusterNode
args = _get_redis_cluster_kwargs()
cluster_kwargs = {}
cluster_kwargs: Final = {}
for arg in redis_kwargs:
if arg in args:
cluster_kwargs[arg] = redis_kwargs[arg]
@ -621,7 +622,7 @@ def get_redis_async_client(
username=os.environ.get("REDIS_USERNAME") or None,
)
new_startup_nodes: list[ClusterNode] = []
new_startup_nodes: Final[list[ClusterNode]] = []
for item in redis_kwargs["startup_nodes"]:
new_startup_nodes.append(ClusterNode(**item))
@ -635,9 +636,9 @@ def get_redis_async_client(
cluster_kwargs.setdefault("socket_keepalive", True)
# Create async RedisCluster with IAM token as password if available
cluster_client = async_redis.RedisCluster(
cluster_client: Final = async_redis.RedisCluster(
startup_nodes=new_startup_nodes,
**cluster_kwargs, # type: ignore
**cluster_kwargs,
)
return cluster_client
@ -646,12 +647,14 @@ def get_redis_async_client(
if connection_pool is not None:
return async_redis.Redis(connection_pool=connection_pool)
args = _get_redis_url_kwargs(client=async_redis.Redis)
url_kwargs = {}
url_kwargs: Final = {}
for arg in redis_kwargs:
if arg in args:
url_kwargs[arg] = redis_kwargs[arg]
else:
verbose_logger.debug(f"REDIS: ignoring argument: {arg}. Not an allowed async_redis.Redis.from_url arg.")
verbose_logger.debug(
"REDIS: ignoring argument: %s. Not an allowed async_redis.Redis.from_url arg.", arg
)
return async_redis.Redis.from_url(**url_kwargs)
# Check for Redis Sentinel
@ -684,15 +687,15 @@ def get_redis_async_client(
def get_redis_connection_pool(
**env_overrides,
) -> async_redis.BlockingConnectionPool | None:
redis_kwargs = _get_redis_client_logic(**env_overrides)
redis_kwargs: Final = _get_redis_client_logic(**env_overrides)
verbose_logger.debug("get_redis_connection_pool: redis_kwargs", redis_kwargs)
if "startup_nodes" in redis_kwargs:
return None
if "url" in redis_kwargs and redis_kwargs["url"] is not None:
allowed_args = _get_redis_url_kwargs(client=async_redis.Redis)
pool_kwargs = {k: v for k, v in redis_kwargs.items() if k in allowed_args and k != "max_connections"}
allowed_args: Final = _get_redis_url_kwargs(client=async_redis.Redis)
pool_kwargs: Final = {k: v for k, v in redis_kwargs.items() if k in allowed_args and k != "max_connections"}
pool_kwargs["timeout"] = REDIS_CONNECTION_POOL_TIMEOUT
pool_kwargs["url"] = redis_kwargs["url"]
if "max_connections" in redis_kwargs:
@ -708,7 +711,7 @@ def get_redis_connection_pool(
# Wrap GCP / Azure AD auth in a CredentialProvider so pool-managed
# connections re-fetch tokens via the SDK's internal cache + silent refresh
# rather than reusing a single token captured at pool creation.
redis_connect_func = redis_kwargs.pop("redis_connect_func", None)
redis_connect_func: Final = redis_kwargs.pop("redis_connect_func", None)
if redis_connect_func and hasattr(redis_connect_func, "_azure_credential"):
redis_kwargs["credential_provider"] = AzureADCredentialProvider(
redis_connect_func._azure_credential,
@ -735,7 +738,7 @@ def _pretty_print_redis_config(redis_kwargs: dict) -> None:
if not verbose_logger.isEnabledFor(logging.DEBUG):
return
console = Console()
console: Final = Console()
# Initialize the sensitive data masker
masker = SensitiveDataMasker()
@ -744,10 +747,10 @@ def _pretty_print_redis_config(redis_kwargs: dict) -> None:
masked_redis_kwargs = masker.mask_dict(redis_kwargs)
# Create main panel title
title = Text("Redis Configuration", style="bold blue")
title: Final = Text("Redis Configuration", style="bold blue")
# Create configuration table
config_table = Table(
config_table: Final = Table(
title="🔧 Redis Connection Parameters",
show_header=True,
header_style="bold magenta",
@ -784,7 +787,7 @@ def _pretty_print_redis_config(redis_kwargs: dict) -> None:
connection_type = "Redis (URL-based)"
# Create connection type info
info_table = Table(
info_table: Final = Table(
title="📊 Connection Info",
show_header=True,
header_style="bold green",
@ -805,6 +808,6 @@ def _pretty_print_redis_config(redis_kwargs: dict) -> None:
# Fallback to simple logging if rich is not available
masker = SensitiveDataMasker()
masked_redis_kwargs = masker.mask_dict(redis_kwargs)
verbose_logger.info(f"Redis configuration: {masked_redis_kwargs}")
verbose_logger.info("Redis configuration: %s", masked_redis_kwargs)
except Exception as e:
verbose_logger.error(f"Error pretty printing Redis configuration: {e}")
verbose_logger.error("Error pretty printing Redis configuration: %s", e)

26
litellm/_redis_credential_provider.py
View file

@ -1,21 +1,21 @@
import asyncio
import threading
import time
from typing import Any
from typing import Any, Final
from redis.credentials import CredentialProvider # type: ignore[attr-defined]
from redis.credentials import CredentialProvider
# Azure AD scope for Redis Cache for Azure.
AZURE_REDIS_SCOPE = "https://redis.azure.com/.default"
AZURE_REDIS_SCOPE: Final = "https://redis.azure.com/.default"
# GCP IAM tokens are valid for 1 hour. Cache for 55 minutes to refresh before expiry.
_GCP_IAM_TOKEN_TTL_SECONDS = 3300
_GCP_IAM_TOKEN_TTL_SECONDS: Final = 3300
# Module-level cache shared across all GCPIAMCredentialProvider instances for the
# same service account, so multiple Redis connections on the same pod share one token.
# Keyed by service_account → (token, expiry_monotonic_timestamp).
_token_cache: dict[str, tuple[str, float]] = {}
_token_cache_lock = threading.Lock()
_token_cache: Final[dict[str, tuple[str, float]]] = {}
_token_cache_lock: Final = threading.Lock()
def _generate_gcp_iam_access_token(service_account: str) -> str:
@ -36,12 +36,12 @@ def _generate_gcp_iam_access_token(service_account: str) -> str:
"Install it with: pip install google-cloud-iam"
)
client = iam_credentials_v1.IAMCredentialsClient()
request = iam_credentials_v1.GenerateAccessTokenRequest(
client: Final = iam_credentials_v1.IAMCredentialsClient()
request: Final = iam_credentials_v1.GenerateAccessTokenRequest(
name=service_account,
scope=["https://www.googleapis.com/auth/cloud-platform"],
)
response = client.generate_access_token(request=request)
response: Final = client.generate_access_token(request=request)
return str(response.access_token)
@ -96,11 +96,11 @@ class GCPIAMCredentialProvider(CredentialProvider):
self._gcp_service_account = gcp_service_account
def get_credentials(self) -> tuple[str]:
token = _get_cached_gcp_iam_token(self._gcp_service_account)
token: Final = _get_cached_gcp_iam_token(self._gcp_service_account)
return (token,)
async def get_credentials_async(self) -> tuple[str]:
token = await asyncio.to_thread(_get_cached_gcp_iam_token, self._gcp_service_account)
token: Final = await asyncio.to_thread(_get_cached_gcp_iam_token, self._gcp_service_account)
return (token,)
@ -120,13 +120,13 @@ class AzureADCredentialProvider(CredentialProvider):
self._username = username
def get_credentials(self) -> tuple[str] | tuple[str, str]:
token = self._credential.get_token(AZURE_REDIS_SCOPE).token
token: Final = self._credential.get_token(AZURE_REDIS_SCOPE).token
if self._username:
return (self._username, token)
return (token,)
async def get_credentials_async(self) -> tuple[str] | tuple[str, str]:
token_obj = await asyncio.to_thread(self._credential.get_token, AZURE_REDIS_SCOPE)
token_obj: Final = await asyncio.to_thread(self._credential.get_token, AZURE_REDIS_SCOPE)
if self._username:
return (self._username, token_obj.token)
return (token_obj.token,)

14
litellm/_service_logger.py
View file

@ -1,6 +1,6 @@
import asyncio
from datetime import datetime, timedelta
from typing import TYPE_CHECKING, Any, Union
from typing import TYPE_CHECKING, Any, Final, Union
import litellm
from litellm._logging import verbose_logger
@ -67,7 +67,7 @@ class ServiceLogging(CustomLogger):
whether the callback is the logger instance itself or the ``"otel"`` string
(which routes to the proxy's registered ``open_telemetry_logger``).
"""
otel_v2_cls = _get_otel_v2_class()
otel_v2_cls: Final = _get_otel_v2_class()
def _is_otel_logger(obj: Any) -> bool:
if isinstance(obj, OpenTelemetry):
@ -101,7 +101,7 @@ class ServiceLogging(CustomLogger):
try:
# Try to get the current event loop
loop = asyncio.get_event_loop()
loop: Final = asyncio.get_event_loop()
# Check if the loop is running
if loop.is_running():
# If we're in a running loop, create a task
@ -163,7 +163,7 @@ class ServiceLogging(CustomLogger):
if self.mock_testing:
self.mock_testing_async_success_hook += 1
payload = ServiceLoggerPayload(
payload: Final = ServiceLoggerPayload(
is_error=False,
error=None,
service=service,
@ -178,7 +178,7 @@ class ServiceLogging(CustomLogger):
# (the V2 logger self-registers its instance even when the string is
# present, unlike V1). Without this guard each such reference emits its own
# span, so a single DB call shows up as duplicate ``postgres ...`` spans.
emitted_otel_logger_ids: set = set()
emitted_otel_logger_ids: Final[set] = set()
for callback in litellm.service_callback:
if callback == "prometheus_system":
await self.init_prometheus_services_logger_if_none()
@ -267,7 +267,7 @@ class ServiceLogging(CustomLogger):
elif isinstance(error, str):
error_message = error
payload = ServiceLoggerPayload(
payload: Final = ServiceLoggerPayload(
is_error=True,
error=error_message,
service=service,
@ -278,7 +278,7 @@ class ServiceLogging(CustomLogger):
# Dedupe OTel loggers per event — see ``async_service_success_hook`` for why
# the same logger can be referenced twice in ``service_callback``.
emitted_otel_logger_ids: set = set()
emitted_otel_logger_ids: Final[set] = set()
for callback in litellm.service_callback:
if callback == "prometheus_system":
await self.init_prometheus_services_logger_if_none()

2
litellm/_uuid.py
View file

@ -4,7 +4,7 @@ Internal unified UUID helper.
Always uses fastuuid for performance.
"""
import fastuuid as _uuid # type: ignore
import fastuuid as _uuid
# Expose a module-like alias so callers can use: uuid.uuid4()
uuid = _uuid

32
litellm/a2a_protocol/card_resolver.py
View file

@ -4,7 +4,7 @@ Custom A2A Card Resolver for LiteLLM.
Extends the A2A SDK's card resolver to support multiple well-known paths.
"""
from typing import TYPE_CHECKING, Any
from typing import TYPE_CHECKING, Any, Final
from litellm._logging import verbose_logger
from litellm.constants import LOCALHOST_URL_PATTERNS
@ -18,8 +18,8 @@ AGENT_CARD_WELL_KNOWN_PATH: str = "/.well-known/agent-card.json"
PREV_AGENT_CARD_WELL_KNOWN_PATH: str = "/.well-known/agent.json"
try:
from a2a.client import A2ACardResolver as _A2ACardResolver # type: ignore[no-redef]
from a2a.utils.constants import ( # type: ignore[no-redef]
from a2a.client import A2ACardResolver as _A2ACardResolver
from a2a.utils.constants import (
AGENT_CARD_WELL_KNOWN_PATH,
PREV_AGENT_CARD_WELL_KNOWN_PATH,
)
@ -43,18 +43,18 @@ def is_localhost_or_internal_url(url: str | None) -> bool:
if not url:
return False
url_lower = url.lower()
url_lower: Final = url.lower()
return any(pattern in url_lower for pattern in LOCALHOST_URL_PATTERNS)
def get_agent_card_url(agent_card: "AgentCard") -> str | None:
"""Return the agent endpoint URL from the resolved SDK card."""
url = getattr(agent_card, "url", None)
url: Final = getattr(agent_card, "url", None)
if url:
return url
interfaces = getattr(agent_card, "supported_interfaces", None)
interfaces: Final = getattr(agent_card, "supported_interfaces", None)
if interfaces:
return getattr(interfaces[0], "url", None)
return None
@ -62,11 +62,11 @@ def get_agent_card_url(agent_card: "AgentCard") -> str | None:
def set_agent_card_url(agent_card: "AgentCard", url: str) -> None:
"""Set the agent endpoint URL on the resolved SDK card."""
normalized = url.rstrip("/") + "/"
normalized: Final = url.rstrip("/") + "/"
if hasattr(agent_card, "url"):
agent_card.url = normalized
interfaces = getattr(agent_card, "supported_interfaces", None)
interfaces: Final = getattr(agent_card, "supported_interfaces", None)
if interfaces:
interfaces[0].url = normalized
@ -86,23 +86,23 @@ def fix_agent_card_url(agent_card: "AgentCard", base_url: str) -> "AgentCard":
Returns:
The agent card with the URL fixed if necessary
"""
card_url = getattr(agent_card, "url", None)
card_url: Final = getattr(agent_card, "url", None)
if card_url and is_localhost_or_internal_url(card_url):
# Normalize base_url to ensure it ends with /
fixed_url = base_url.rstrip("/") + "/"
fixed_url: Final = base_url.rstrip("/") + "/"
agent_card.url = fixed_url
interfaces = getattr(agent_card, "supported_interfaces", None)
interfaces: Final = getattr(agent_card, "supported_interfaces", None)
if interfaces:
interface_url = getattr(interfaces[0], "url", None)
interface_url: Final = getattr(interfaces[0], "url", None)
if interface_url and is_localhost_or_internal_url(interface_url):
interfaces[0].url = base_url.rstrip("/") + "/"
return agent_card
class LiteLLMA2ACardResolver(_A2ACardResolver): # type: ignore[misc]
class LiteLLMA2ACardResolver(_A2ACardResolver):
"""
Custom A2A card resolver that supports multiple well-known paths.
@ -140,7 +140,7 @@ class LiteLLMA2ACardResolver(_A2ACardResolver): # type: ignore[misc]
)
# Try both well-known paths
paths = [
paths: Final = [
AGENT_CARD_WELL_KNOWN_PATH,
PREV_AGENT_CARD_WELL_KNOWN_PATH,
]
@ -148,13 +148,13 @@ class LiteLLMA2ACardResolver(_A2ACardResolver): # type: ignore[misc]
last_error = None
for path in paths:
try:
verbose_logger.debug(f"Attempting to fetch agent card from {self.base_url}{path}")
verbose_logger.debug("Attempting to fetch agent card from %s%s", self.base_url, path)
return await super().get_agent_card(
relative_card_path=path,
http_kwargs=http_kwargs,
)
except Exception as e:
verbose_logger.debug(f"Failed to fetch agent card from {self.base_url}{path}: {e}")
verbose_logger.debug("Failed to fetch agent card from %s%s: %s", self.base_url, path, e)
last_error = e
continue

6
litellm/a2a_protocol/client.py
View file

@ -5,7 +5,7 @@ Provides a class-based interface for A2A agent invocation.
"""
from collections.abc import AsyncIterator
from typing import TYPE_CHECKING
from typing import TYPE_CHECKING, Final
from litellm.types.agents import LiteLLMSendMessageResponse
@ -92,7 +92,7 @@ class A2AClient:
"""Send a message to the A2A agent."""
from litellm.a2a_protocol.main import asend_message
a2a_client = await self._get_client()
a2a_client: Final = await self._get_client()
return await asend_message(a2a_client=a2a_client, request=request)
async def send_message_streaming(
@ -101,6 +101,6 @@ class A2AClient:
"""Send a streaming message to the A2A agent."""
from litellm.a2a_protocol.main import asend_message_streaming
a2a_client = await self._get_client()
a2a_client: Final = await self._get_client()
async for chunk in asend_message_streaming(a2a_client=a2a_client, request=request):
yield chunk

22
litellm/a2a_protocol/cost_calculator.py
View file

@ -5,7 +5,7 @@ Supports dynamic cost parameters that allow platform owners
to define custom costs per agent query or per token.
"""
from typing import TYPE_CHECKING, Any
from typing import TYPE_CHECKING, Any, Final
if TYPE_CHECKING:
from litellm.litellm_core_utils.litellm_logging import (
@ -42,23 +42,23 @@ class A2ACostCalculator:
if litellm_logging_obj is None:
return 0.0
model_call_details = litellm_logging_obj.model_call_details
model_call_details: Final = litellm_logging_obj.model_call_details
# Check if user set a custom response cost (backward compatibility)
response_cost = model_call_details.get("response_cost", None)
response_cost: Final = model_call_details.get("response_cost", None)
if response_cost is not None:
return float(response_cost)
# Get litellm_params for cost parameters
litellm_params = model_call_details.get("litellm_params", {}) or {}
litellm_params: Final = model_call_details.get("litellm_params", {}) or {}
# Check for cost_per_query (fixed cost per query)
if litellm_params.get("cost_per_query") is not None:
return float(litellm_params["cost_per_query"])
# Check for token-based pricing
input_cost_per_token = litellm_params.get("input_cost_per_token")
output_cost_per_token = litellm_params.get("output_cost_per_token")
input_cost_per_token: Final = litellm_params.get("input_cost_per_token")
output_cost_per_token: Final = litellm_params.get("output_cost_per_token")
if input_cost_per_token is not None or output_cost_per_token is not None:
return A2ACostCalculator._calculate_token_based_cost(
@ -88,16 +88,16 @@ class A2ACostCalculator:
float: The calculated cost
"""
# Get usage from model_call_details
usage = model_call_details.get("usage")
usage: Final = model_call_details.get("usage")
if usage is None:
return 0.0
# Get token counts
prompt_tokens = getattr(usage, "prompt_tokens", 0) or 0
completion_tokens = getattr(usage, "completion_tokens", 0) or 0
prompt_tokens: Final = getattr(usage, "prompt_tokens", 0) or 0
completion_tokens: Final = getattr(usage, "completion_tokens", 0) or 0
# Calculate costs
input_cost = prompt_tokens * (float(input_cost_per_token) if input_cost_per_token else 0.0)
output_cost = completion_tokens * (float(output_cost_per_token) if output_cost_per_token else 0.0)
input_cost: Final = prompt_tokens * (float(input_cost_per_token) if input_cost_per_token else 0.0)
output_cost: Final = completion_tokens * (float(output_cost_per_token) if output_cost_per_token else 0.0)
return input_cost + output_cost

34
litellm/a2a_protocol/exception_mapping_utils.py
View file

@ -4,7 +4,7 @@ A2A Protocol Exception Mapping Utils.
Maps A2A SDK exceptions to LiteLLM A2A exception types.
"""
from typing import TYPE_CHECKING, Any
from typing import TYPE_CHECKING, Any, Final
from litellm._logging import verbose_logger
from litellm.a2a_protocol.card_resolver import (
@ -29,9 +29,9 @@ try:
A2A_SDK_AVAILABLE = True
except ImportError:
A2A_SDK_AVAILABLE = False
Client = None # type: ignore[misc, assignment]
ClientConfig = None # type: ignore[misc, assignment]
create_client = None # type: ignore[misc, assignment]
Client = None
ClientConfig = None
create_client = None
class A2AExceptionCheckers:
@ -53,7 +53,7 @@ class A2AExceptionCheckers:
if not isinstance(error_str, str):
return False
error_str_lower = error_str.lower()
error_str_lower: Final = error_str.lower()
return any(pattern in error_str_lower for pattern in CONNECTION_ERROR_PATTERNS)
@staticmethod
@ -83,8 +83,8 @@ class A2AExceptionCheckers:
if not isinstance(error_str, str):
return False
error_str_lower = error_str.lower()
agent_card_patterns = [
error_str_lower: Final = error_str.lower()
agent_card_patterns: Final = [
"agent card",
"agent-card",
".well-known",
@ -118,7 +118,7 @@ def map_a2a_exception(
A2AAgentCardError: If the error is related to agent card issues
A2AError: For other A2A-related errors
"""
error_str = str(original_exception)
error_str: Final = str(original_exception)
# Check for localhost URL connection error (special case - retryable)
if (
@ -190,11 +190,13 @@ async def handle_a2a_localhost_retry(
"rewrite, so the upstream URL cannot be corrected."
)
request_type = "streaming " if is_streaming else ""
request_type: Final = "streaming " if is_streaming else ""
verbose_logger.warning(
f"A2A {request_type}request to '{error.localhost_url}' failed: {error.original_error}. "
f"Agent card contains localhost/internal URL. "
f"Retrying with base_url '{error.base_url}'."
"A2A %srequest to '%s' failed: %s. Agent card contains localhost/internal URL. Retrying with base_url '%s'.",
request_type,
error.localhost_url,
error.original_error,
error.base_url,
)
# Fix the agent card URL
@ -203,20 +205,20 @@ async def handle_a2a_localhost_retry(
# Reuse the httpx client LiteLLM attached at creation. It carries this agent's
# trace-id and auth headers, so a fresh client would drop them. Only clients built
# by ``create_a2a_client`` have it; an externally-supplied client cannot be retried.
httpx_client = getattr(a2a_client, "_litellm_httpx_client", None)
httpx_client: Final = getattr(a2a_client, "_litellm_httpx_client", None)
if httpx_client is None:
raise RuntimeError(
"Cannot retry A2A localhost URL fix: the client was not created by "
"create_a2a_client, so no LiteLLM httpx client is attached."
)
new_client = await create_client( # pyright: ignore[reportOptionalCall]
new_client: Final = await create_client( # pyright: ignore[reportOptionalCall]
agent_card,
client_config=ClientConfig( # pyright: ignore[reportOptionalCall]
httpx_client=httpx_client,
streaming=is_streaming,
),
)
new_client._litellm_httpx_client = httpx_client # type: ignore[attr-defined]
new_client._litellm_agent_card = agent_card # type: ignore[attr-defined]
new_client._litellm_httpx_client = httpx_client
new_client._litellm_agent_card = agent_card
return new_client

62
litellm/a2a_protocol/litellm_completion_bridge/handler.py
View file

@ -11,7 +11,7 @@ A2A Streaming Events (in order):
"""
from collections.abc import AsyncIterator
from typing import Any
from typing import Any, Final
import litellm
from litellm._logging import verbose_logger
@ -25,10 +25,10 @@ from litellm.interactions.agents.utils import merge_agent_headers
# litellm_params key carrying the authenticated principal (hashed virtual key) so
# A2A provider configs can scope provider-side state (e.g. LangFlow session memory)
# per key instead of trusting the client-supplied A2A contextId.
A2A_USER_API_KEY_HASH_PARAM = "litellm_a2a_user_api_key_hash"
A2A_USER_API_KEY_HASH_PARAM: Final = "litellm_a2a_user_api_key_hash"
# Agent metadata fields stored in litellm_params that are not valid litellm.acompletion() kwargs
_AGENT_ONLY_PARAMS = frozenset(
_AGENT_ONLY_PARAMS: Final = frozenset(
{
"is_public",
"agent_name",
@ -70,13 +70,13 @@ class A2ACompletionBridgeHandler:
"""
custom_llm_provider = litellm_params.get("custom_llm_provider")
if not _skip_a2a_provider_routing:
a2a_provider_config = A2AProviderConfigManager.get_provider_config(
a2a_provider_config: Final = A2AProviderConfigManager.get_provider_config(
custom_llm_provider=custom_llm_provider,
model=litellm_params.get("model"),
)
if a2a_provider_config is not None:
verbose_logger.info(f"A2A: Using provider config for {custom_llm_provider}")
verbose_logger.info("A2A: Using provider config for %s", custom_llm_provider)
return await a2a_provider_config.handle_non_streaming(
request_id=request_id,
@ -87,14 +87,14 @@ class A2ACompletionBridgeHandler:
)
# Extract message from params
message = params.get("message", {})
message: Final = params.get("message", {})
# Transform A2A message to OpenAI format
openai_messages = A2ACompletionBridgeTransformation.a2a_message_to_openai_messages(message)
openai_messages: Final = A2ACompletionBridgeTransformation.a2a_message_to_openai_messages(message)
# Get completion params
custom_llm_provider = litellm_params.get("custom_llm_provider")
model = litellm_params.get("model", "agent")
model: Final = litellm_params.get("model", "agent")
# Build full model string if provider specified
# Skip prepending if model already starts with the provider prefix
@ -103,17 +103,17 @@ class A2ACompletionBridgeHandler:
else:
full_model = model
verbose_logger.info(f"A2A completion bridge: model={full_model}, api_base={api_base}")
verbose_logger.info("A2A completion bridge: model=%s, api_base=%s", full_model, api_base)
# Build completion params dict
completion_params: dict[str, Any] = {
completion_params: Final[dict[str, Any]] = {
"model": full_model,
"messages": openai_messages,
"api_base": api_base,
"stream": False,
}
# Add litellm_params (contains api_key, client_id, client_secret, tenant_id, etc.)
litellm_params_to_add = {
litellm_params_to_add: Final = {
k: v
for k, v in litellm_params.items()
if k not in ("model", "custom_llm_provider") and k not in _AGENT_ONLY_PARAMS
@ -135,15 +135,15 @@ class A2ACompletionBridgeHandler:
)
# Call litellm.acompletion
response = await litellm.acompletion(**completion_params)
response: Final = await litellm.acompletion(**completion_params)
# Transform response to A2A format
a2a_response = A2ACompletionBridgeTransformation.openai_response_to_a2a_response(
a2a_response: Final = A2ACompletionBridgeTransformation.openai_response_to_a2a_response(
response=response,
request_id=request_id,
)
verbose_logger.info(f"A2A completion bridge completed: request_id={request_id}")
verbose_logger.info("A2A completion bridge completed: request_id=%s", request_id)
return a2a_response
@ -179,13 +179,13 @@ class A2ACompletionBridgeHandler:
"""
custom_llm_provider = litellm_params.get("custom_llm_provider")
if not _skip_a2a_provider_routing:
a2a_provider_config = A2AProviderConfigManager.get_provider_config(
a2a_provider_config: Final = A2AProviderConfigManager.get_provider_config(
custom_llm_provider=custom_llm_provider,
model=litellm_params.get("model"),
)
if a2a_provider_config is not None:
verbose_logger.info(f"A2A: Using provider config for {custom_llm_provider} (streaming)")
verbose_logger.info("A2A: Using provider config for %s (streaming)", custom_llm_provider)
async for chunk in a2a_provider_config.handle_streaming(
request_id=request_id,
@ -199,20 +199,20 @@ class A2ACompletionBridgeHandler:
return
# Extract message from params
message = params.get("message", {})
message: Final = params.get("message", {})
# Create streaming context
ctx = A2AStreamingContext(
ctx: Final = A2AStreamingContext(
request_id=request_id,
input_message=message,
)
# Transform A2A message to OpenAI format
openai_messages = A2ACompletionBridgeTransformation.a2a_message_to_openai_messages(message)
openai_messages: Final = A2ACompletionBridgeTransformation.a2a_message_to_openai_messages(message)
# Get completion params
custom_llm_provider = litellm_params.get("custom_llm_provider")
model = litellm_params.get("model", "agent")
model: Final = litellm_params.get("model", "agent")
# Build full model string if provider specified
# Skip prepending if model already starts with the provider prefix
@ -221,17 +221,17 @@ class A2ACompletionBridgeHandler:
else:
full_model = model
verbose_logger.info(f"A2A completion bridge streaming: model={full_model}, api_base={api_base}")
verbose_logger.info("A2A completion bridge streaming: model=%s, api_base=%s", full_model, api_base)
# Build completion params dict
completion_params: dict[str, Any] = {
completion_params: Final[dict[str, Any]] = {
"model": full_model,
"messages": openai_messages,
"api_base": api_base,
"stream": True,
}
# Add litellm_params (contains api_key, client_id, client_secret, tenant_id, etc.)
litellm_params_to_add = {
litellm_params_to_add: Final = {
k: v
for k, v in litellm_params.items()
if k not in ("model", "custom_llm_provider") and k not in _AGENT_ONLY_PARAMS
@ -253,11 +253,11 @@ class A2ACompletionBridgeHandler:
)
# 1. Emit initial task event (kind: "task", status: "submitted")
task_event = A2ACompletionBridgeTransformation.create_task_event(ctx)
task_event: Final = A2ACompletionBridgeTransformation.create_task_event(ctx)
yield task_event
# 2. Emit status update (kind: "status-update", status: "working")
working_event = A2ACompletionBridgeTransformation.create_status_update_event(
working_event: Final = A2ACompletionBridgeTransformation.create_status_update_event(
ctx=ctx,
state="working",
final=False,
@ -266,12 +266,12 @@ class A2ACompletionBridgeHandler:
yield working_event
# Call litellm.acompletion with streaming
response = await litellm.acompletion(**completion_params)
response: Final = await litellm.acompletion(**completion_params)
# 3. Accumulate content and emit artifact update
accumulated_text = ""
chunk_count = 0
async for chunk in response: # type: ignore[union-attr]
async for chunk in response:
chunk_count += 1
# Extract delta content
@ -286,21 +286,23 @@ class A2ACompletionBridgeHandler:
# Emit artifact update with accumulated content
if accumulated_text:
artifact_event = A2ACompletionBridgeTransformation.create_artifact_update_event(
artifact_event: Final = A2ACompletionBridgeTransformation.create_artifact_update_event(
ctx=ctx,
text=accumulated_text,
)
yield artifact_event
# 4. Emit final status update (kind: "status-update", status: "completed", final: true)
completed_event = A2ACompletionBridgeTransformation.create_status_update_event(
completed_event: Final = A2ACompletionBridgeTransformation.create_status_update_event(
ctx=ctx,
state="completed",
final=True,
)
yield completed_event
verbose_logger.info(f"A2A completion bridge streaming completed: request_id={request_id}, chunks={chunk_count}")
verbose_logger.info(
"A2A completion bridge streaming completed: request_id=%s, chunks=%s", request_id, chunk_count
)
# Convenience functions that delegate to the class methods

38
litellm/a2a_protocol/litellm_completion_bridge/transformation.py
View file

@ -18,7 +18,7 @@ A2A Streaming Events:
"""
from datetime import datetime, timezone
from typing import Any
from typing import Any, Final
from uuid import uuid4
from litellm._logging import verbose_logger
@ -48,7 +48,7 @@ class A2ACompletionBridgeTransformation:
@staticmethod
def _extract_text_from_a2a_parts(parts: list[dict[str, Any]]) -> str:
"""Extract text from A2A parts (with or without explicit ``kind``)."""
content_parts: list[str] = []
content_parts: Final[list[str]] = []
for part in parts:
if not isinstance(part, dict):
continue
@ -71,10 +71,10 @@ class A2ACompletionBridgeTransformation:
Forwarded once on the LangGraph run payload (``metadata``), not duplicated on
each input message — see ``apply_forward_metadata_to_completion_params``.
"""
merged: dict[str, Any] = {}
merged: Final[dict[str, Any]] = {}
if params and isinstance(params.get("metadata"), dict):
merged.update(params["metadata"])
message_metadata = a2a_message.get("metadata")
message_metadata: Final = a2a_message.get("metadata")
if isinstance(message_metadata, dict):
merged.update(message_metadata)
return merged or None
@ -90,7 +90,7 @@ class A2ACompletionBridgeTransformation:
Uses ``extra_body`` so we do not collide with LiteLLM's spend-log ``metadata`` kwarg.
"""
forward_metadata = A2ACompletionBridgeTransformation.get_forward_metadata(
forward_metadata: Final = A2ACompletionBridgeTransformation.get_forward_metadata(
a2a_message=a2a_message,
params=params,
)
@ -103,13 +103,13 @@ class A2ACompletionBridgeTransformation:
# 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 = extra_body.get("metadata")
existing_dict: dict[str, Any] = existing_metadata if isinstance(existing_metadata, dict) else {}
merged_metadata: dict[str, Any] = {**forward_metadata, **existing_dict}
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
verbose_logger.debug(f"A2A -> completion forward metadata keys={list(forward_metadata.keys())}")
verbose_logger.debug("A2A -> completion forward metadata keys=%s", list(forward_metadata.keys()))
@staticmethod
def a2a_message_to_openai_messages(
@ -124,7 +124,7 @@ class A2ACompletionBridgeTransformation:
Returns:
List of OpenAI-format messages
"""
role = a2a_message.get("role", "user")
role: Final = a2a_message.get("role", "user")
parts = a2a_message.get("parts", [])
# Map A2A roles to OpenAI roles
@ -139,13 +139,15 @@ class A2ACompletionBridgeTransformation:
if not isinstance(parts, list):
parts = []
content = A2ACompletionBridgeTransformation._extract_text_from_a2a_parts(parts)
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: dict[str, Any] = {"role": openai_role, "content": content}
openai_message: Final[dict[str, Any]] = {"role": openai_role, "content": content}
verbose_logger.debug(f"A2A -> OpenAI transform: role={role} -> {openai_role}, content_length={len(content)}")
verbose_logger.debug(
"A2A -> OpenAI transform: role=%s -> %s, content_length=%s", role, openai_role, len(content)
)
return [openai_message]
@ -167,12 +169,12 @@ class A2ACompletionBridgeTransformation:
# Extract content from response
content = ""
if hasattr(response, "choices") and response.choices:
choice = response.choices[0]
choice: Final = response.choices[0]
if hasattr(choice, "message") and choice.message:
content = choice.message.content or ""
# Build A2A message
a2a_message = {
a2a_message: Final = {
"kind": "message",
"role": "agent",
"parts": [{"kind": "text", "text": content}],
@ -180,13 +182,13 @@ class A2ACompletionBridgeTransformation:
}
# Build A2A response
a2a_response = {
a2a_response: Final = {
"jsonrpc": "2.0",
"id": request_id,
"result": a2a_message,
}
verbose_logger.debug(f"OpenAI -> A2A transform: content_length={len(content)}")
verbose_logger.debug("OpenAI -> A2A transform: content_length=%s", len(content))
return a2a_response
@ -243,7 +245,7 @@ class A2ACompletionBridgeTransformation:
final: Whether this is the final event
message_text: Optional message text for 'working' status
"""
status: dict[str, Any] = {
status: Final[dict[str, Any]] = {
"state": state,
"timestamp": A2ACompletionBridgeTransformation._get_timestamp(),
}

133
litellm/a2a_protocol/main.py
View file

@ -13,12 +13,7 @@ import asyncio
import datetime
import uuid
from collections.abc import AsyncIterator, Coroutine
from typing import (
TYPE_CHECKING,
Any,
Optional,
cast,
)
from typing import TYPE_CHECKING, Any, Final, Optional, cast
import litellm
from litellm._logging import verbose_logger, verbose_proxy_logger
@ -64,9 +59,9 @@ try:
A2A_SDK_AVAILABLE = True
except ImportError:
Client = None # type: ignore[misc, assignment]
ClientConfig = None # type: ignore[misc, assignment]
create_client = None # type: ignore[misc, assignment]
Client = None
ClientConfig = None
create_client = None
# Import our custom card resolver that supports multiple well-known paths
from litellm.a2a_protocol.card_resolver import (
@ -80,7 +75,7 @@ from litellm.a2a_protocol.exception_mapping_utils import (
from litellm.a2a_protocol.exceptions import A2ALocalhostURLError
# Use our custom resolver instead of the default A2A SDK resolver
A2ACardResolver = LiteLLMA2ACardResolver
A2ACardResolver: Final = LiteLLMA2ACardResolver
def _set_usage_on_logging_obj(
@ -96,9 +91,9 @@ def _set_usage_on_logging_obj(
prompt_tokens: Number of input tokens
completion_tokens: Number of output tokens
"""
litellm_logging_obj = kwargs.get("litellm_logging_obj")
litellm_logging_obj: Final = kwargs.get("litellm_logging_obj")
if litellm_logging_obj is not None:
usage = litellm.Usage(
usage: Final = litellm.Usage(
prompt_tokens=prompt_tokens,
completion_tokens=completion_tokens,
total_tokens=prompt_tokens + completion_tokens,
@ -120,13 +115,13 @@ def _set_agent_id_on_logging_obj(
if agent_id is None:
return
litellm_logging_obj = kwargs.get("litellm_logging_obj")
litellm_logging_obj: Final = kwargs.get("litellm_logging_obj")
if litellm_logging_obj is not None:
# Set agent_id directly on model_call_details (same pattern as custom_llm_provider)
litellm_logging_obj.model_call_details["agent_id"] = agent_id
_A2A_COST_PARAM_KEYS = ("cost_per_query", "input_cost_per_token", "output_cost_per_token")
_A2A_COST_PARAM_KEYS: Final = ("cost_per_query", "input_cost_per_token", "output_cost_per_token")
def _set_litellm_params_on_logging_obj(
@ -141,7 +136,7 @@ def _set_litellm_params_on_logging_obj(
litellm_params already carries metadata / proxy_server_request / user-key
context, so merge the pricing keys in rather than replacing the dict.
"""
logging_obj = kwargs.get("litellm_logging_obj")
logging_obj: Final = kwargs.get("litellm_logging_obj")
if logging_obj is None:
return
@ -149,7 +144,7 @@ def _set_litellm_params_on_logging_obj(
if not cost_params:
return
existing = logging_obj.model_call_details.get("litellm_params") or {}
existing: Final = logging_obj.model_call_details.get("litellm_params") or {}
logging_obj.model_call_details["litellm_params"] = {**existing, **cost_params}
@ -162,17 +157,17 @@ def _get_a2a_model_info(a2a_client: Any, kwargs: dict[str, Any]) -> str:
"""
agent_name = "unknown"
agent_card = _get_a2a_client_agent_card(a2a_client)
agent_card: Final = _get_a2a_client_agent_card(a2a_client)
if agent_card is not None:
agent_name = getattr(agent_card, "name", "unknown") or "unknown"
# Build model string
model = f"a2a_agent/{agent_name}"
custom_llm_provider = "a2a_agent"
model: Final = f"a2a_agent/{agent_name}"
custom_llm_provider: Final = "a2a_agent"
# Set on litellm_logging_obj if available (for standard logging payload)
litellm_logging_obj = kwargs.get("litellm_logging_obj")
litellm_logging_obj: Final = kwargs.get("litellm_logging_obj")
if litellm_logging_obj is not None:
litellm_logging_obj.model = model
litellm_logging_obj.custom_llm_provider = custom_llm_provider
@ -204,7 +199,7 @@ async def _send_message_via_completion_bridge(
Requires request; api_base is optional for providers that derive endpoint from model.
"""
verbose_logger.info(f"A2A using completion bridge: provider={custom_llm_provider}, api_base={api_base}")
verbose_logger.info("A2A using completion bridge: provider=%s, api_base=%s", custom_llm_provider, api_base)
from litellm.a2a_protocol.litellm_completion_bridge.handler import (
A2ACompletionBridgeHandler,
@ -212,7 +207,7 @@ async def _send_message_via_completion_bridge(
params = request.params.model_dump(mode="json") if hasattr(request.params, "model_dump") else dict(request.params)
response_dict = await A2ACompletionBridgeHandler.handle_non_streaming(
response_dict: Final = await A2ACompletionBridgeHandler.handle_non_streaming(
request_id=str(request.id),
params=params,
litellm_params=litellm_params,
@ -230,18 +225,18 @@ async def _send_message(a2a_client: "A2AClientType", request: "SendMessageReques
"The 'a2a' package is required for A2A agent invocation. Install it with: pip install a2a-sdk"
)
pb_request = _a2a_conversions.to_core_send_message_request(request)
pb_request: Final = _a2a_conversions.to_core_send_message_request(request)
last_event = None
async for event in a2a_client.send_message(pb_request):
last_event = event
if last_event is None:
raise RuntimeError("A2A send_message failed: no response received from agent.")
stream_compat = _a2a_conversions.to_compat_stream_response(
stream_compat: Final = _a2a_conversions.to_compat_stream_response(
last_event,
request_id=request.id,
)
result = stream_compat.result
result: Final = stream_compat.result
if not isinstance(result, (Message, Task)):
raise RuntimeError(
"A2A send_message failed: non-streaming message/send expects the "
@ -305,7 +300,7 @@ async def _stream_messages(
"The 'a2a' package is required for A2A agent invocation. Install it with: pip install a2a-sdk"
)
pb_request = _a2a_conversions.to_core_send_message_request(request)
pb_request: Final = _a2a_conversions.to_core_send_message_request(request)
async for event in a2a_client.send_message(pb_request):
compat_chunk = _a2a_conversions.to_compat_stream_response(
event,
@ -425,9 +420,9 @@ async def asend_message(
```
"""
litellm_params = litellm_params or {}
logging_obj = kwargs.get("litellm_logging_obj")
logging_obj: Final = kwargs.get("litellm_logging_obj")
trace_id = getattr(logging_obj, "litellm_trace_id", None) if logging_obj else None
custom_llm_provider = litellm_params.get("custom_llm_provider")
custom_llm_provider: Final = litellm_params.get("custom_llm_provider")
# Route through completion bridge if custom_llm_provider is set
if custom_llm_provider:
@ -450,7 +445,7 @@ async def asend_message(
if api_base is None:
raise ValueError("Either a2a_client or api_base is required for standard A2A flow")
trace_id = trace_id or str(uuid.uuid4())
extra_headers: dict[str, str] = {"X-LiteLLM-Trace-Id": trace_id}
extra_headers: Final[dict[str, str]] = {"X-LiteLLM-Trace-Id": trace_id}
if agent_id:
extra_headers["X-LiteLLM-Agent-Id"] = agent_id
# Overlay agent-level headers (agent headers take precedence over LiteLLM internal ones)
@ -461,15 +456,15 @@ async def asend_message(
# Type assertion: a2a_client is guaranteed to be non-None here
assert a2a_client is not None
agent_name = _get_a2a_model_info(a2a_client, kwargs)
agent_name: Final = _get_a2a_model_info(a2a_client, kwargs)
verbose_logger.info(f"A2A send_message request_id={request.id}, agent={agent_name}")
verbose_logger.info("A2A send_message request_id=%s, agent=%s", request.id, agent_name)
# Get agent card URL for localhost retry logic
agent_card = _get_a2a_client_agent_card(a2a_client)
card_url = get_agent_card_url(agent_card) if agent_card else None
agent_card: Final = _get_a2a_client_agent_card(a2a_client)
card_url: Final = get_agent_card_url(agent_card) if agent_card else None
a2a_response = await _execute_a2a_send_with_retry(
a2a_response: Final = await _execute_a2a_send_with_retry(
a2a_client=a2a_client,
request=request,
agent_card=agent_card,
@ -478,13 +473,13 @@ async def asend_message(
agent_name=agent_name,
)
verbose_logger.info(f"A2A send_message completed, request_id={request.id}")
verbose_logger.info("A2A send_message completed, request_id=%s", request.id)
# Wrap in LiteLLM response type for _hidden_params support
response = LiteLLMSendMessageResponse.from_a2a_response(a2a_response, request_id=str(request.id))
response: Final = LiteLLMSendMessageResponse.from_a2a_response(a2a_response, request_id=str(request.id))
# Calculate token usage from request and response
response_dict = a2a_response.model_dump(mode="json", exclude_none=True)
response_dict: Final = a2a_response.model_dump(mode="json", exclude_none=True)
(
prompt_tokens,
completion_tokens,
@ -549,10 +544,10 @@ def _build_streaming_logging_obj(
proxy_server_request: dict[str, Any] | None,
) -> Logging:
"""Build logging object for streaming A2A requests."""
start_time = datetime.datetime.now()
model = f"a2a_agent/{agent_name}"
start_time: Final = datetime.datetime.now()
model: Final = f"a2a_agent/{agent_name}"
logging_obj = Logging(
logging_obj: Final = Logging(
model=model,
messages=[{"role": "user", "content": "streaming-request"}],
stream=False,
@ -569,7 +564,7 @@ def _build_streaming_logging_obj(
if agent_id:
logging_obj.model_call_details["agent_id"] = agent_id
_litellm_params = litellm_params.copy() if litellm_params else {}
_litellm_params: Final = litellm_params.copy() if litellm_params else {}
if metadata:
_litellm_params["metadata"] = metadata
if proxy_server_request:
@ -632,7 +627,7 @@ async def asend_message_streaming(
```
"""
litellm_params = litellm_params or {}
custom_llm_provider = litellm_params.get("custom_llm_provider")
custom_llm_provider: Final = litellm_params.get("custom_llm_provider")
# Route through completion bridge if custom_llm_provider is set
if custom_llm_provider:
@ -640,14 +635,14 @@ async def asend_message_streaming(
raise ValueError("request is required for completion bridge")
# api_base is optional for providers that derive endpoint from model (e.g., bedrock/agentcore)
verbose_logger.info(f"A2A streaming using completion bridge: provider={custom_llm_provider}")
verbose_logger.info("A2A streaming using completion bridge: provider=%s", custom_llm_provider)
from litellm.a2a_protocol.litellm_completion_bridge.handler import (
A2ACompletionBridgeHandler,
)
# Extract params from request
params = (
params: Final = (
request.params.model_dump(mode="json") if hasattr(request.params, "model_dump") else dict(request.params)
)
@ -664,15 +659,15 @@ async def asend_message_streaming(
if request is None:
raise ValueError("request is required")
_raw_logging_obj = kwargs.get("litellm_logging_obj")
_raw_logging_obj: Final = kwargs.get("litellm_logging_obj")
logging_obj: Logging | None = _raw_logging_obj if isinstance(_raw_logging_obj, Logging) else None
if a2a_client is None:
if api_base is None:
raise ValueError("Either a2a_client or api_base is required for standard A2A flow")
logging_trace_id = getattr(logging_obj, "litellm_trace_id", None) if logging_obj else None
trace_id = logging_trace_id or (str(request.id) if request.id else str(uuid.uuid4()))
extra_headers: dict[str, str] = {"X-LiteLLM-Trace-Id": trace_id}
logging_trace_id: Final = getattr(logging_obj, "litellm_trace_id", None) if logging_obj else None
trace_id: Final = logging_trace_id or (str(request.id) if request.id else str(uuid.uuid4()))
extra_headers: Final[dict[str, str]] = {"X-LiteLLM-Trace-Id": trace_id}
if agent_id:
extra_headers["X-LiteLLM-Agent-Id"] = agent_id
if agent_extra_headers:
@ -685,7 +680,7 @@ async def asend_message_streaming(
assert a2a_client is not None
agent_name = _get_a2a_model_info(a2a_client, kwargs)
agent_name: Final = _get_a2a_model_info(a2a_client, kwargs)
if logging_obj is None:
logging_obj = _build_streaming_logging_obj(
@ -697,12 +692,12 @@ async def asend_message_streaming(
proxy_server_request=proxy_server_request,
)
verbose_logger.info(f"A2A send_message_streaming request_id={request.id}, agent={agent_name}")
verbose_logger.info("A2A send_message_streaming request_id=%s, agent=%s", request.id, agent_name)
agent_card = _get_a2a_client_agent_card(a2a_client)
card_url = get_agent_card_url(agent_card) if agent_card else None
agent_card: Final = _get_a2a_client_agent_card(a2a_client)
card_url: Final = get_agent_card_url(agent_card) if agent_card else None
stream = _execute_a2a_stream_with_retry(
stream: Final = _execute_a2a_stream_with_retry(
a2a_client=a2a_client,
request=request,
agent_card=agent_card,
@ -759,7 +754,7 @@ async def create_a2a_client(
"The 'a2a' package is required for A2A agent invocation. Install it with: pip install a2a-sdk"
)
verbose_logger.info(f"Creating A2A client for {base_url}")
verbose_logger.info("Creating A2A client for %s", base_url)
# Use get_async_httpx_client with per-agent params so that different agents
# (with different extra_headers) get separate cached clients. The params
@ -769,21 +764,21 @@ async def create_a2a_client(
# Only pass params that AsyncHTTPHandler.__init__ accepts (e.g. timeout).
# Use "disable_aiohttp_transport" key for cache-key-only data (it's
# filtered out before reaching the constructor).
_client_params: dict = {"timeout": timeout}
_client_params: Final[dict] = {"timeout": timeout}
if extra_headers:
# Encode headers into a cache-key-only param so each unique header
# set produces a distinct cache key.
_client_params["disable_aiohttp_transport"] = str(sorted(extra_headers.items()))
_async_handler = get_async_httpx_client(
_async_handler: Final = get_async_httpx_client(
llm_provider=httpxSpecialProvider.A2AProvider,
params=_client_params,
)
httpx_client = _async_handler.client
httpx_client: Final = _async_handler.client
if extra_headers:
httpx_client.headers.update(extra_headers)
verbose_proxy_logger.debug(f"A2A client created with extra_headers={list(extra_headers.keys())}")
verbose_proxy_logger.debug("A2A client created with extra_headers=%s", list(extra_headers.keys()))
a2a_client = await create_client( # pyright: ignore[reportOptionalCall]
a2a_client: Final = await create_client( # pyright: ignore[reportOptionalCall]
base_url,
client_config=ClientConfig( # pyright: ignore[reportOptionalCall]
httpx_client=httpx_client,
@ -793,12 +788,12 @@ async def create_a2a_client(
# Stash LiteLLM-owned handles on the client so the localhost-retry path can reuse
# the configured httpx client (with this agent's trace-id/auth headers) without
# excavating a2a-sdk private internals.
a2a_client._litellm_httpx_client = httpx_client # type: ignore[attr-defined]
agent_card = getattr(a2a_client, "_card", None)
a2a_client._litellm_httpx_client = httpx_client
agent_card: Final = getattr(a2a_client, "_card", None)
if agent_card is not None:
a2a_client._litellm_agent_card = agent_card # type: ignore[attr-defined]
a2a_client._litellm_agent_card = agent_card
verbose_logger.info(f"A2A client created for {base_url}")
verbose_logger.info("A2A client created for %s", base_url)
return a2a_client
@ -824,20 +819,20 @@ async def aget_agent_card(
"The 'a2a' package is required for A2A agent invocation. Install it with: pip install a2a-sdk"
)
verbose_logger.info(f"Fetching agent card from {base_url}")
verbose_logger.info("Fetching agent card from %s", base_url)
# Use LiteLLM's cached httpx client
http_handler = get_async_httpx_client(
http_handler: Final = get_async_httpx_client(
llm_provider=httpxSpecialProvider.A2A,
params={"timeout": timeout},
)
httpx_client = http_handler.client
httpx_client: Final = http_handler.client
resolver = A2ACardResolver(
resolver: Final = A2ACardResolver(
httpx_client=httpx_client,
base_url=base_url,
)
agent_card = await resolver.get_agent_card()
agent_card: Final = await resolver.get_agent_card()
verbose_logger.info(f"Fetched agent card: {agent_card.name if hasattr(agent_card, 'name') else 'unknown'}")
verbose_logger.info("Fetched agent card: %s", agent_card.name if hasattr(agent_card, "name") else "unknown")
return agent_card

6
litellm/a2a_protocol/providers/bedrock_agentcore/config.py
View file

@ -3,7 +3,7 @@ Bedrock AgentCore A2A provider configuration.
"""
from collections.abc import AsyncIterator
from typing import Any
from typing import Any, Final
from litellm.a2a_protocol.providers.base import BaseA2AProviderConfig
from litellm.a2a_protocol.providers.bedrock_agentcore.handler import (
@ -28,7 +28,7 @@ class BedrockAgentCoreA2AConfig(BaseA2AProviderConfig):
**kwargs,
) -> dict[str, Any]:
"""Handle non-streaming request to AgentCore A2A agent."""
litellm_params = kwargs.get("litellm_params")
litellm_params: Final = kwargs.get("litellm_params")
if not litellm_params:
raise ValueError(
"litellm_params is required for BedrockAgentCoreA2AConfig (must contain model with AgentCore ARN)"
@ -48,7 +48,7 @@ class BedrockAgentCoreA2AConfig(BaseA2AProviderConfig):
**kwargs,
) -> AsyncIterator[dict[str, Any]]:
"""Handle streaming request to AgentCore A2A agent."""
litellm_params = kwargs.get("litellm_params")
litellm_params: Final = kwargs.get("litellm_params")
if not litellm_params:
raise ValueError(
"litellm_params is required for BedrockAgentCoreA2AConfig (must contain model with AgentCore ARN)"

24
litellm/a2a_protocol/providers/bedrock_agentcore/handler.py
View file

@ -7,7 +7,7 @@ completion bridge that would otherwise strip the envelope.
import json
from collections.abc import AsyncIterator
from typing import Any, cast
from typing import Any, Final, cast
from litellm._logging import verbose_logger
from litellm.a2a_protocol.providers.bedrock_agentcore.transformation import (
@ -53,21 +53,21 @@ class BedrockAgentCoreA2AHandler:
agent_extra_headers=agent_extra_headers,
)
verbose_logger.info(f"BedrockAgentCore A2A: Sending non-streaming request to {url}")
verbose_logger.info("BedrockAgentCore A2A: Sending non-streaming request to %s", url)
client = get_async_httpx_client(
client: Final = get_async_httpx_client(
llm_provider=cast(Any, httpxSpecialProvider.A2AProvider),
)
response = await client.post(
response: Final = await client.post(
url,
headers=headers,
data=body,
)
response.raise_for_status()
response_data = response.json()
response_data: Final = response.json()
if "error" in response_data:
verbose_logger.warning(f"BedrockAgentCore A2A: Agent returned error: {response_data['error']}")
verbose_logger.warning("BedrockAgentCore A2A: Agent returned error: %s", response_data["error"])
return response_data
@ -100,12 +100,12 @@ class BedrockAgentCoreA2AHandler:
agent_extra_headers=agent_extra_headers,
)
verbose_logger.info(f"BedrockAgentCore A2A: Sending streaming request to {url}")
verbose_logger.info("BedrockAgentCore A2A: Sending streaming request to %s", url)
client = get_async_httpx_client(
client: Final = get_async_httpx_client(
llm_provider=cast(Any, httpxSpecialProvider.A2AProvider),
)
response = await client.post(
response: Final = await client.post(
url,
headers=headers,
data=body,
@ -114,15 +114,15 @@ class BedrockAgentCoreA2AHandler:
response.raise_for_status()
# Check content type — AgentCore may return JSON instead of SSE
content_type = response.headers.get("content-type", "").lower()
content_type: Final = response.headers.get("content-type", "").lower()
if "application/json" in content_type:
# Single JSON response fallback (not SSE)
verbose_logger.debug(
"BedrockAgentCore A2A streaming: received JSON instead of SSE, yielding as single event"
)
response_body = await response.aread()
response_data = json.loads(response_body)
response_body: Final = await response.aread()
response_data: Final = json.loads(response_body)
yield response_data
else:
# SSE stream — parse data: lines

30
litellm/a2a_protocol/providers/bedrock_agentcore/transformation.py
View file

@ -7,7 +7,7 @@ and signs requests via AmazonAgentCoreConfig (SigV4 or JWT).
import json
from collections.abc import AsyncIterator, Mapping
from typing import Any
from typing import Any, Final
from litellm._logging import verbose_logger
from litellm.llms.bedrock.chat.agentcore.transformation import AmazonAgentCoreConfig
@ -23,13 +23,13 @@ from litellm.llms.bedrock.chat.agentcore.transformation import AmazonAgentCoreCo
# ``runtimeSessionId`` / ``runtimeUserId`` in the agent's ``litellm_params``;
# ``authorization`` is set by the AgentCore signer (JWT or SigV4); ``host`` and
# the ``x-amz-*`` family are owned by SigV4 itself.
_RESERVED_EXACT_HEADERS = frozenset(
_RESERVED_EXACT_HEADERS: Final = frozenset(
{
"authorization",
"host",
}
)
_RESERVED_PREFIX_HEADERS: tuple[str, ...] = (
_RESERVED_PREFIX_HEADERS: Final[tuple[str, ...]] = (
"x-amzn-bedrock-agentcore-runtime-",
"x-amz-",
)
@ -47,8 +47,8 @@ def _filter_reserved_headers(
if not agent_extra_headers:
return None
filtered: dict[str, str] = {}
dropped: list = []
filtered: Final[dict[str, str]] = {}
dropped: Final[list] = []
for k, v in agent_extra_headers.items():
k_lower = k.lower()
if k_lower in _RESERVED_EXACT_HEADERS or any(k_lower.startswith(prefix) for prefix in _RESERVED_PREFIX_HEADERS):
@ -107,19 +107,19 @@ class BedrockAgentCoreA2ATransformation:
"""
# Extract model and strip the "bedrock/" prefix
# "bedrock/agentcore/arn:aws:..." → "agentcore/arn:aws:..."
model = litellm_params.get("model", "")
model: Final = litellm_params.get("model", "")
if model.startswith("bedrock/"):
agentcore_model = model[len("bedrock/") :]
else:
agentcore_model = model
# Build optional_params from litellm_params (everything except model and custom_llm_provider)
optional_params = {k: v for k, v in litellm_params.items() if k not in ("model", "custom_llm_provider")}
optional_params: Final = {k: v for k, v in litellm_params.items() if k not in ("model", "custom_llm_provider")}
agentcore_config = AmazonAgentCoreConfig()
agentcore_config: Final = AmazonAgentCoreConfig()
# Derive URL from ARN
url = agentcore_config.get_complete_url(
url: Final = agentcore_config.get_complete_url(
api_base=optional_params.get("api_base"),
api_key=optional_params.get("api_key"),
model=agentcore_model,
@ -129,7 +129,7 @@ class BedrockAgentCoreA2ATransformation:
)
# Construct JSON-RPC 2.0 envelope
json_rpc_body = {
json_rpc_body: Final = {
"jsonrpc": "2.0",
"method": method,
"id": request_id,
@ -138,17 +138,17 @@ class BedrockAgentCoreA2ATransformation:
# Set required AgentCore session headers (normally set by transform_request,
# which we skip because it also builds {"prompt": "..."})
headers: dict = {}
session_id = agentcore_config._get_runtime_session_id(optional_params)
headers: Final[dict] = {}
session_id: Final = agentcore_config._get_runtime_session_id(optional_params)
headers["X-Amzn-Bedrock-AgentCore-Runtime-Session-Id"] = session_id
runtime_user_id = agentcore_config._get_runtime_user_id(optional_params)
runtime_user_id: Final = agentcore_config._get_runtime_user_id(optional_params)
if runtime_user_id:
headers["X-Amzn-Bedrock-AgentCore-Runtime-User-Id"] = runtime_user_id
# Merge per-request agent headers before signing so SigV4 covers them.
# Reserved headers are stripped first to prevent client-controlled values
# from spoofing the AgentCore runtime identity / SigV4 metadata.
safe_extra_headers = _filter_reserved_headers(agent_extra_headers)
safe_extra_headers: Final = _filter_reserved_headers(agent_extra_headers)
if safe_extra_headers:
headers.update(safe_extra_headers)
@ -195,5 +195,5 @@ class BedrockAgentCoreA2ATransformation:
event = json.loads(data_str)
yield event
except json.JSONDecodeError:
verbose_logger.debug(f"BedrockAgentCore A2A: Skipping non-JSON SSE line: {data_str[:100]}")
verbose_logger.debug("BedrockAgentCore A2A: Skipping non-JSON SSE line: %s", data_str[:100])
continue

10
litellm/a2a_protocol/providers/pydantic_ai_agents/handler.py
View file

@ -6,7 +6,7 @@ This handler provides fake streaming by converting non-streaming responses into
"""
from collections.abc import AsyncIterator
from typing import Any
from typing import Any, Final
from litellm._logging import verbose_logger
from litellm.a2a_protocol.providers.pydantic_ai_agents.transformation import (
@ -47,10 +47,10 @@ class PydanticAIHandler:
"""
if api_base is None:
raise ValueError("api_base is required for Pydantic AI agents")
verbose_logger.info(f"Pydantic AI: Routing to Pydantic AI agent at {api_base}")
verbose_logger.info("Pydantic AI: Routing to Pydantic AI agent at %s", api_base)
# Send request directly to Pydantic AI agent
response_data = await PydanticAITransformation.send_non_streaming_request(
response_data: Final = await PydanticAITransformation.send_non_streaming_request(
api_base=api_base,
request_id=request_id,
params=params,
@ -92,10 +92,10 @@ class PydanticAIHandler:
"""
if api_base is None:
raise ValueError("api_base is required for Pydantic AI agents")
verbose_logger.info(f"Pydantic AI: Faking streaming for Pydantic AI agent at {api_base}")
verbose_logger.info("Pydantic AI: Faking streaming for Pydantic AI agent at %s", api_base)
# Get raw task response first (not the transformed A2A format)
raw_response = await PydanticAITransformation.send_and_get_raw_response(
raw_response: Final = await PydanticAITransformation.send_and_get_raw_response(
api_base=api_base,
request_id=request_id,
params=params,

58
litellm/a2a_protocol/providers/pydantic_ai_agents/transformation.py
View file

@ -7,7 +7,7 @@ This module provides fake streaming by converting non-streaming responses into s
import asyncio
from collections.abc import AsyncIterator
from typing import Any, cast
from typing import Any, Final, cast
from uuid import uuid4
from litellm._logging import verbose_logger
@ -118,7 +118,7 @@ class PydanticAITransformation:
status = result.get("status", {})
state = status.get("state", "")
verbose_logger.debug(f"Pydantic AI: Poll attempt {attempt + 1}/{max_attempts}, state={state}")
verbose_logger.debug("Pydantic AI: Poll attempt %s/%s, state=%s", attempt + 1, max_attempts, state)
if state == "completed":
return poll_data
@ -163,7 +163,7 @@ class PydanticAITransformation:
params_dict["message"]["kind"] = "message"
# Build A2A JSON-RPC request using message/send method for FastA2A compatibility
a2a_request = {
a2a_request: Final = {
"jsonrpc": "2.0",
"id": request_id,
"method": "message/send",
@ -171,16 +171,16 @@ class PydanticAITransformation:
}
# FastA2A uses root endpoint (/) not /messages
endpoint = api_base.rstrip("/")
endpoint: Final = api_base.rstrip("/")
verbose_logger.info(f"Pydantic AI: Sending non-streaming request to {endpoint}")
verbose_logger.info("Pydantic AI: Sending non-streaming request to %s", endpoint)
# Send request to Pydantic AI agent using shared async HTTP client
client = get_async_httpx_client(
client: Final = get_async_httpx_client(
llm_provider=cast(Any, "pydantic_ai_agent"),
params={"timeout": timeout},
)
response = await client.post(
response: Final = await client.post(
endpoint,
json=a2a_request,
headers={
@ -192,15 +192,15 @@ class PydanticAITransformation:
response_data = response.json()
# Check if task is already completed
result = response_data.get("result", {})
status = result.get("status", {})
state = status.get("state", "")
result: Final = response_data.get("result", {})
status: Final = result.get("status", {})
state: Final = status.get("state", "")
if state != "completed":
# Need to poll for completion
task_id = result.get("id")
task_id: Final = result.get("id")
if task_id:
verbose_logger.info(f"Pydantic AI: Task {task_id} submitted, polling for completion...")
verbose_logger.info("Pydantic AI: Task %s submitted, polling for completion...", task_id)
response_data = await PydanticAITransformation._poll_for_completion(
client=client,
endpoint=endpoint,
@ -209,7 +209,7 @@ class PydanticAITransformation:
agent_extra_headers=agent_extra_headers,
)
verbose_logger.info(f"Pydantic AI: Received completed response for request_id={request_id}")
verbose_logger.info("Pydantic AI: Received completed response for request_id=%s", request_id)
return response_data
@ -235,7 +235,7 @@ class PydanticAITransformation:
Standard A2A non-streaming response format with message
"""
# Get raw task response
raw_response = await PydanticAITransformation._send_and_poll_raw(
raw_response: Final = await PydanticAITransformation._send_and_poll_raw(
api_base=api_base,
request_id=request_id,
params=params,
@ -313,7 +313,7 @@ class PydanticAITransformation:
full_text, message_id, parts = PydanticAITransformation._extract_response_text(response_data)
# Build standard A2A message
a2a_message = {
a2a_message: Final = {
"kind": "message",
"role": "agent",
"parts": parts if parts else [{"kind": "text", "text": full_text}],
@ -342,10 +342,10 @@ class PydanticAITransformation:
Returns:
Tuple of (full_text, message_id, parts)
"""
result = response_data.get("result", {})
result: Final = response_data.get("result", {})
# Try to extract from artifacts first (preferred for results)
artifacts = result.get("artifacts", [])
artifacts: Final = result.get("artifacts", [])
if artifacts:
for artifact in artifacts:
parts = artifact.get("parts", [])
@ -356,7 +356,7 @@ class PydanticAITransformation:
return text, str(uuid4()), parts
# Fall back to history - get the last agent message
history = result.get("history", [])
history: Final = result.get("history", [])
for msg in reversed(history):
if msg.get("role") == "agent":
parts = msg.get("parts", [])
@ -369,7 +369,7 @@ class PydanticAITransformation:
return full_text, message_id, parts
# Fall back to message field (original format)
message = result.get("message", {})
message: Final = result.get("message", {})
if message:
parts = message.get("parts", [])
message_id = message.get("messageId", str(uuid4()))
@ -410,8 +410,8 @@ class PydanticAITransformation:
full_text, message_id, parts = PydanticAITransformation._extract_response_text(response_data)
# Extract input message from raw response for history
result = response_data.get("result", {})
history = result.get("history", [])
result: Final = response_data.get("result", {})
history: Final = result.get("history", [])
input_message = {}
for msg in history:
if msg.get("role") == "user":
@ -419,14 +419,14 @@ class PydanticAITransformation:
break
# Generate IDs for streaming events
task_id = str(uuid4())
context_id = str(uuid4())
artifact_id = str(uuid4())
input_message_id = input_message.get("messageId", str(uuid4()))
task_id: Final = str(uuid4())
context_id: Final = str(uuid4())
artifact_id: Final = str(uuid4())
input_message_id: Final = input_message.get("messageId", str(uuid4()))
# 1. Emit initial task event (kind: "task", status: "submitted")
# Format matches A2ACompletionBridgeTransformation.create_task_event
task_event = {
task_event: Final = {
"jsonrpc": "2.0",
"id": request_id,
"result": {
@ -452,7 +452,7 @@ class PydanticAITransformation:
# 2. Emit status update (kind: "status-update", status: "working")
# Format matches A2ACompletionBridgeTransformation.create_status_update_event
working_event = {
working_event: Final = {
"jsonrpc": "2.0",
"id": request_id,
"result": {
@ -503,7 +503,7 @@ class PydanticAITransformation:
await asyncio.sleep(delay_ms / 1000.0)
# 4. Emit final status update (kind: "status-update", status: "completed", final: true)
completed_event = {
completed_event: Final = {
"jsonrpc": "2.0",
"id": request_id,
"result": {
@ -518,4 +518,4 @@ class PydanticAITransformation:
}
yield completed_event
verbose_logger.info(f"Pydantic AI: Fake streaming completed for request_id={request_id}")
verbose_logger.info("Pydantic AI: Fake streaming completed for request_id=%s", request_id)

6
litellm/a2a_protocol/providers/watsonx_orchestrate/config.py
View file

@ -3,7 +3,7 @@ A2A provider configuration for IBM watsonx Orchestrate (WXO).
"""
from collections.abc import AsyncIterator
from typing import Any
from typing import Any, Final
from litellm.a2a_protocol.providers.base import BaseA2AProviderConfig
from litellm.a2a_protocol.providers.watsonx_orchestrate.handler import (
@ -22,7 +22,7 @@ class WatsonxOrchestrateA2AConfig(BaseA2AProviderConfig):
**kwargs: Any,
) -> dict[str, Any]:
"""Handle a non-streaming A2A request via WXO runs API."""
litellm_params = kwargs.get("litellm_params")
litellm_params: Final = kwargs.get("litellm_params")
if not litellm_params:
raise ValueError(
"litellm_params is required for WatsonxOrchestrateA2AConfig "
@ -42,7 +42,7 @@ class WatsonxOrchestrateA2AConfig(BaseA2AProviderConfig):
**kwargs: Any,
) -> AsyncIterator[dict[str, Any]]:
"""Handle a streaming A2A request via WXO streaming runs API."""
litellm_params = kwargs.get("litellm_params")
litellm_params: Final = kwargs.get("litellm_params")
if not litellm_params:
raise ValueError(
"litellm_params is required for WatsonxOrchestrateA2AConfig "

88
litellm/a2a_protocol/providers/watsonx_orchestrate/handler.py
View file

@ -7,7 +7,7 @@ import hashlib
import json
import time
from collections.abc import AsyncIterator
from typing import Any, NamedTuple, cast
from typing import Any, Final, NamedTuple, cast
import httpx
@ -21,11 +21,11 @@ from litellm.llms.custom_httpx.http_handler import (
)
from litellm.types.llms.custom_http import httpxSpecialProvider
_IBM_CLOUD_IAM_URL = "https://iam.cloud.ibm.com/identity/token"
_POLL_INTERVAL_S = 2.0
_MAX_POLL_ATTEMPTS = 90
_TOKEN_CACHE_TTL_BUFFER_S = 60
_token_cache: dict[str, tuple[str, float]] = {}
_IBM_CLOUD_IAM_URL: Final = "https://iam.cloud.ibm.com/identity/token"
_POLL_INTERVAL_S: Final = 2.0
_MAX_POLL_ATTEMPTS: Final = 90
_TOKEN_CACHE_TTL_BUFFER_S: Final = 60
_token_cache: Final[dict[str, tuple[str, float]]] = {}
class WXORequestParams(NamedTuple):
@ -53,14 +53,14 @@ class WatsonxOrchestrateHandler:
api_key: str,
username: str | None,
) -> str:
material = f"{auth_mode}:{cp4d_host}:{username or ''}:{api_key}"
material: Final = f"{auth_mode}:{cp4d_host}:{username or ''}:{api_key}"
return hashlib.sha256(material.encode()).hexdigest()
@staticmethod
def _cp4d_token_ttl_seconds(expiration: Any, now_wall: float | None = None) -> int:
# CP4D returns expiration as absolute Unix epoch seconds, not a duration.
expires_at = int(expiration)
wall = now_wall if now_wall is not None else time.time()
expires_at: Final = int(expiration)
wall: Final = now_wall if now_wall is not None else time.time()
return max(expires_at - int(wall), 0)
@staticmethod
@ -71,9 +71,9 @@ class WatsonxOrchestrateHandler:
username: str | None = None,
client: AsyncHTTPHandler | None = None,
) -> str:
cache_key = WatsonxOrchestrateHandler._token_cache_key(auth_mode, cp4d_host, api_key, username)
now = time.monotonic()
cached = _token_cache.get(cache_key)
cache_key: Final = WatsonxOrchestrateHandler._token_cache_key(auth_mode, cp4d_host, api_key, username)
now: Final = time.monotonic()
cached: Final = _token_cache.get(cache_key)
if cached and cached[1] > now:
return cached[0]
@ -96,7 +96,7 @@ class WatsonxOrchestrateHandler:
else:
if not username:
raise ValueError("'username' is required in litellm_params when auth_mode='cp4d'")
token_url = f"{cp4d_host.rstrip('/')}/icp4d-api/v1/authorize"
token_url: Final = f"{cp4d_host.rstrip('/')}/icp4d-api/v1/authorize"
response = await client.post(
token_url,
json={"username": username, "api_key": api_key},
@ -105,13 +105,13 @@ class WatsonxOrchestrateHandler:
response.raise_for_status()
payload = response.json()
token = str(payload["token"])
expiration = payload.get("expiration")
expiration: Final = payload.get("expiration")
if expiration is None:
ttl_s = 3600
else:
ttl_s = WatsonxOrchestrateHandler._cp4d_token_ttl_seconds(expiration)
expires_at = now + max(ttl_s - _TOKEN_CACHE_TTL_BUFFER_S, 0)
expires_at: Final = now + max(ttl_s - _TOKEN_CACHE_TTL_BUFFER_S, 0)
_token_cache[cache_key] = (token, expires_at)
for stale_key, (_, stale_expires_at) in list(_token_cache.items()):
if stale_expires_at <= now:
@ -127,7 +127,7 @@ class WatsonxOrchestrateHandler:
max_attempts: int = _MAX_POLL_ATTEMPTS,
interval_s: float = _POLL_INTERVAL_S,
) -> dict[str, Any]:
url = f"{base_url}/v1/orchestrate/runs/{run_id}"
url: Final = f"{base_url}/v1/orchestrate/runs/{run_id}"
for attempt in range(max_attempts):
await asyncio.sleep(interval_s)
@ -135,7 +135,7 @@ class WatsonxOrchestrateHandler:
response.raise_for_status()
result: dict[str, Any] = response.json()
status = result.get("status", "")
verbose_logger.debug(f"WXO: Poll {attempt + 1}/{max_attempts} run='{run_id}' status='{status}'")
verbose_logger.debug("WXO: Poll %s/%s run='%s' status='%s'", attempt + 1, max_attempts, run_id, status)
if status in WatsonxOrchestrateTransformation.TERMINAL_STATES:
return result
@ -152,7 +152,7 @@ class WatsonxOrchestrateHandler:
) -> dict[str, Any]:
status = run_data.get("status", "")
if status not in WatsonxOrchestrateTransformation.TERMINAL_STATES:
run_id = run_data.get("run_id") or run_data.get("id") or ""
run_id: Final = run_data.get("run_id") or run_data.get("id") or ""
if not run_id:
raise ValueError(f"WXO: No run_id in response: {run_data}")
run_data = await WatsonxOrchestrateHandler._poll_run(
@ -188,10 +188,10 @@ class WatsonxOrchestrateHandler:
@staticmethod
def _extract_litellm_params(litellm_params: dict[str, Any]) -> WXORequestParams:
cp4d_host = litellm_params.get("cp4d_host") or ""
instance_id = litellm_params.get("instance_id") or ""
wxo_agent_id = litellm_params.get("wxo_agent_id") or ""
api_key = litellm_params.get("api_key") or ""
cp4d_host: Final = litellm_params.get("cp4d_host") or ""
instance_id: Final = litellm_params.get("instance_id") or ""
wxo_agent_id: Final = litellm_params.get("wxo_agent_id") or ""
api_key: Final = litellm_params.get("api_key") or ""
if not cp4d_host:
raise ValueError("'cp4d_host' is required in litellm_params for WXO agents")
@ -218,29 +218,29 @@ class WatsonxOrchestrateHandler:
params: dict[str, Any],
litellm_params: dict[str, Any],
) -> dict[str, Any]:
wxo = WatsonxOrchestrateHandler._extract_litellm_params(litellm_params)
wxo: Final = WatsonxOrchestrateHandler._extract_litellm_params(litellm_params)
client = WatsonxOrchestrateHandler._http_client(timeout=90.0)
token = await WatsonxOrchestrateHandler._get_bearer_token(
client: Final = WatsonxOrchestrateHandler._http_client(timeout=90.0)
token: Final = await WatsonxOrchestrateHandler._get_bearer_token(
cp4d_host=wxo.cp4d_host,
auth_mode=wxo.auth_mode,
api_key=wxo.api_key,
username=wxo.username,
client=client,
)
base_url = WatsonxOrchestrateTransformation.get_api_base_url(wxo.cp4d_host, wxo.instance_id)
auth_headers = {
base_url: Final = WatsonxOrchestrateTransformation.get_api_base_url(wxo.cp4d_host, wxo.instance_id)
auth_headers: Final = {
"Authorization": f"Bearer {token}",
"Content-Type": "application/json",
"Accept": "application/json",
}
text = WatsonxOrchestrateTransformation.extract_text_from_a2a_params(params)
body = WatsonxOrchestrateTransformation.build_wxo_run_body(
text: Final = WatsonxOrchestrateTransformation.extract_text_from_a2a_params(params)
body: Final = WatsonxOrchestrateTransformation.build_wxo_run_body(
wxo_agent_id=wxo.wxo_agent_id, text=text, thread_id=wxo.thread_id
)
run_response = await client.post(
run_response: Final = await client.post(
f"{base_url}/v1/orchestrate/runs",
json=body,
headers=auth_headers,
@ -255,7 +255,7 @@ class WatsonxOrchestrateHandler:
client=client,
)
response_text = WatsonxOrchestrateTransformation.extract_text_from_wxo_result(run_data)
response_text: Final = WatsonxOrchestrateTransformation.extract_text_from_wxo_result(run_data)
return WatsonxOrchestrateTransformation.build_a2a_message_response(request_id=request_id, text=response_text)
@staticmethod
@ -266,29 +266,29 @@ class WatsonxOrchestrateHandler:
chunk_size: int = 50,
delay_ms: int = 10,
) -> AsyncIterator[dict[str, Any]]:
wxo = WatsonxOrchestrateHandler._extract_litellm_params(litellm_params)
wxo: Final = WatsonxOrchestrateHandler._extract_litellm_params(litellm_params)
client = WatsonxOrchestrateHandler._http_client(timeout=120.0)
token = await WatsonxOrchestrateHandler._get_bearer_token(
client: Final = WatsonxOrchestrateHandler._http_client(timeout=120.0)
token: Final = await WatsonxOrchestrateHandler._get_bearer_token(
cp4d_host=wxo.cp4d_host,
auth_mode=wxo.auth_mode,
api_key=wxo.api_key,
username=wxo.username,
client=client,
)
base_url = WatsonxOrchestrateTransformation.get_api_base_url(wxo.cp4d_host, wxo.instance_id)
auth_headers = {
base_url: Final = WatsonxOrchestrateTransformation.get_api_base_url(wxo.cp4d_host, wxo.instance_id)
auth_headers: Final = {
"Authorization": f"Bearer {token}",
"Content-Type": "application/json",
"Accept": "text/event-stream, application/json",
}
text = WatsonxOrchestrateTransformation.extract_text_from_a2a_params(params)
body = WatsonxOrchestrateTransformation.build_wxo_run_body(
text: Final = WatsonxOrchestrateTransformation.extract_text_from_a2a_params(params)
body: Final = WatsonxOrchestrateTransformation.build_wxo_run_body(
wxo_agent_id=wxo.wxo_agent_id, text=text, thread_id=wxo.thread_id
)
try:
response = await client.post(
response: Final = await client.post(
f"{base_url}/v1/orchestrate/runs/stream",
json=body,
headers=auth_headers,
@ -297,8 +297,8 @@ class WatsonxOrchestrateHandler:
response.raise_for_status()
except httpx.TransportError as exc:
verbose_logger.warning(
f"WXO: Streaming request failed before a run was submitted "
f"({exc!r}), falling back to non-streaming + fake streaming",
"WXO: Streaming request failed before a run was submitted (%r), falling back to non-streaming + fake streaming",
exc,
exc_info=True,
)
result = await WatsonxOrchestrateHandler.handle_non_streaming(
@ -306,7 +306,7 @@ class WatsonxOrchestrateHandler:
params=params,
litellm_params=litellm_params,
)
response_text = WatsonxOrchestrateTransformation.extract_text_from_a2a_message_response(result)
response_text: Final = WatsonxOrchestrateTransformation.extract_text_from_a2a_message_response(result)
async for chunk in WatsonxOrchestrateTransformation.fake_streaming_from_text(
text=response_text,
request_id=request_id,
@ -316,9 +316,9 @@ class WatsonxOrchestrateHandler:
yield chunk
return
content_type = response.headers.get("content-type", "").lower()
content_type: Final = response.headers.get("content-type", "").lower()
if "text/event-stream" not in content_type:
response_body = await response.aread()
response_body: Final = await response.aread()
result = json.loads(response_body)
result = await WatsonxOrchestrateHandler._get_successful_run_data(
run_data=result,

26
litellm/a2a_protocol/providers/watsonx_orchestrate/transformation.py
View file

@ -9,7 +9,7 @@ WXO uses a REST API (not A2A/JSON-RPC) with an async-poll execution model:
import asyncio
from collections.abc import AsyncIterator
from typing import Any
from typing import Any, Final
from uuid import uuid4
from litellm._logging import verbose_logger
@ -35,9 +35,9 @@ class WatsonxOrchestrateTransformation:
A2A format: params.message.parts[*] where part.kind == "text"
"""
message = params.get("message", {})
parts = message.get("parts", [])
texts = []
message: Final = params.get("message", {})
parts: Final = message.get("parts", [])
texts: Final = []
for part in parts:
if not isinstance(part, dict):
continue
@ -53,7 +53,7 @@ class WatsonxOrchestrateTransformation:
thread_id: str | None = None,
) -> dict[str, Any]:
"""Build the WXO POST /v1/orchestrate/runs request body."""
body: dict[str, Any] = {
body: Final[dict[str, Any]] = {
"agent_id": wxo_agent_id,
"message": {
"role": "user",
@ -96,7 +96,7 @@ class WatsonxOrchestrateTransformation:
pass
# Tertiary: results as a raw string
results = result.get("results")
results: Final = result.get("results")
if results and isinstance(results, str):
return results
@ -104,11 +104,11 @@ class WatsonxOrchestrateTransformation:
@staticmethod
def extract_text_from_a2a_message_response(a2a_response: dict[str, Any]) -> str:
result = a2a_response.get("result")
result: Final = a2a_response.get("result")
if not isinstance(result, dict):
verbose_logger.warning("WXO: A2A response missing result object")
return ""
parts = result.get("parts")
parts: Final = result.get("parts")
if not isinstance(parts, list):
verbose_logger.warning("WXO: A2A result has no parts list")
return ""
@ -150,9 +150,9 @@ class WatsonxOrchestrateTransformation:
3. artifact-update chunks
4. status-update (kind="status-update", state="completed", final=True)
"""
task_id = str(uuid4())
context_id = str(uuid4())
artifact_id = str(uuid4())
task_id: Final = str(uuid4())
context_id: Final = str(uuid4())
artifact_id: Final = str(uuid4())
# 1. Task submitted
yield {
@ -181,7 +181,7 @@ class WatsonxOrchestrateTransformation:
await asyncio.sleep(delay_ms / 1000.0)
# 3. Artifact chunks (always emit at least one chunk, even for empty text)
text_to_chunk = text or ""
text_to_chunk: Final = text or ""
for i in range(0, max(len(text_to_chunk), 1), chunk_size):
chunk_text = text_to_chunk[i : i + chunk_size]
is_last = (i + chunk_size) >= max(len(text_to_chunk), 1)
@ -214,4 +214,4 @@ class WatsonxOrchestrateTransformation:
},
}
verbose_logger.debug(f"WXO: Fake streaming completed for request_id={request_id}")
verbose_logger.debug("WXO: Fake streaming completed for request_id=%s", request_id)

48
litellm/a2a_protocol/streaming_iterator.py
View file

@ -5,7 +5,7 @@ A2A Streaming Iterator with token tracking and logging support.
import asyncio
from collections.abc import AsyncIterator
from datetime import datetime
from typing import TYPE_CHECKING, Any
from typing import TYPE_CHECKING, Any, Final
import litellm
from litellm._logging import verbose_logger
@ -47,7 +47,7 @@ class A2AStreamingIterator:
async def __anext__(self) -> "SendStreamingMessageResponse":
try:
chunk = await self.stream.__anext__()
chunk: Final = await self.stream.__anext__()
# Store chunk
self.chunks.append(chunk)
@ -71,8 +71,8 @@ class A2AStreamingIterator:
def _collect_text_from_chunk(self, chunk: Any) -> None:
"""Extract text from a streaming chunk and add to collected parts."""
try:
chunk_dict = chunk.model_dump(mode="json", exclude_none=True) if hasattr(chunk, "model_dump") else {}
text = A2ARequestUtils.extract_text_from_response(chunk_dict)
chunk_dict: Final = chunk.model_dump(mode="json", exclude_none=True) if hasattr(chunk, "model_dump") else {}
text: Final = A2ARequestUtils.extract_text_from_response(chunk_dict)
if text:
self.collected_text_parts.append(text)
except Exception:
@ -81,10 +81,10 @@ class A2AStreamingIterator:
def _is_completed_chunk(self, chunk: Any) -> bool:
"""Check if chunk indicates stream completion."""
try:
chunk_dict = chunk.model_dump(mode="json", exclude_none=True) if hasattr(chunk, "model_dump") else {}
result = chunk_dict.get("result", {})
chunk_dict: Final = chunk.model_dump(mode="json", exclude_none=True) if hasattr(chunk, "model_dump") else {}
result: Final = chunk_dict.get("result", {})
if isinstance(result, dict):
status = result.get("status", {})
status: Final = result.get("status", {})
if isinstance(status, dict):
return status.get("state") == "completed"
except Exception:
@ -94,21 +94,21 @@ class A2AStreamingIterator:
async def _handle_stream_complete(self) -> None:
"""Handle logging and token counting when stream completes."""
try:
end_time = datetime.now()
end_time: Final = datetime.now()
# Calculate tokens from collected text
input_message = A2ARequestUtils.get_input_message_from_request(self.request)
input_text = A2ARequestUtils.extract_text_from_message(input_message)
prompt_tokens = A2ARequestUtils.count_tokens(input_text)
input_message: Final = A2ARequestUtils.get_input_message_from_request(self.request)
input_text: Final = A2ARequestUtils.extract_text_from_message(input_message)
prompt_tokens: Final = A2ARequestUtils.count_tokens(input_text)
# Use the last (most complete) text from chunks
output_text = self.collected_text_parts[-1] if self.collected_text_parts else ""
completion_tokens = A2ARequestUtils.count_tokens(output_text)
output_text: Final = self.collected_text_parts[-1] if self.collected_text_parts else ""
completion_tokens: Final = A2ARequestUtils.count_tokens(output_text)
total_tokens = prompt_tokens + completion_tokens
total_tokens: Final = prompt_tokens + completion_tokens
# Create usage object
usage = litellm.Usage(
usage: Final = litellm.Usage(
prompt_tokens=prompt_tokens,
completion_tokens=completion_tokens,
total_tokens=total_tokens,
@ -120,11 +120,11 @@ class A2AStreamingIterator:
self.logging_obj.model_call_details["stream"] = False
# Calculate cost using A2ACostCalculator
response_cost = A2ACostCalculator.calculate_a2a_cost(self.logging_obj)
response_cost: Final = A2ACostCalculator.calculate_a2a_cost(self.logging_obj)
self.logging_obj.model_call_details["response_cost"] = response_cost
# Build result for logging
result = self._build_logging_result(usage)
result: Final = self._build_logging_result(usage)
# Call success handlers - they will build standard_logging_object
asyncio.create_task(
@ -138,17 +138,19 @@ class A2AStreamingIterator:
)
verbose_logger.info(
f"A2A streaming completed: prompt_tokens={prompt_tokens}, "
f"completion_tokens={completion_tokens}, total_tokens={total_tokens}, "
f"response_cost={response_cost}"
"A2A streaming completed: prompt_tokens=%s, completion_tokens=%s, total_tokens=%s, response_cost=%s",
prompt_tokens,
completion_tokens,
total_tokens,
response_cost,
)
except Exception as e:
verbose_logger.debug(f"Error in A2A streaming completion handler: {e}")
verbose_logger.debug("Error in A2A streaming completion handler: %s", e)
def _build_logging_result(self, usage: litellm.Usage) -> dict[str, Any]:
"""Build a result dict for logging."""
result: dict[str, Any] = {
result: Final[dict[str, Any]] = {
"id": getattr(self.request, "id", "unknown"),
"jsonrpc": "2.0",
"usage": (usage.model_dump() if hasattr(usage, "model_dump") else dict(usage)),
@ -157,7 +159,7 @@ class A2AStreamingIterator:
# Add final chunk result if available
if self.final_chunk:
try:
chunk_dict = self.final_chunk.model_dump(mode="json", exclude_none=True)
chunk_dict: Final = self.final_chunk.model_dump(mode="json", exclude_none=True)
result["result"] = chunk_dict.get("result", {})
except Exception:
pass

20
litellm/a2a_protocol/utils.py
View file

@ -2,7 +2,7 @@
Utility functions for A2A protocol.
"""
from typing import TYPE_CHECKING, Any
from typing import TYPE_CHECKING, Any, Final
import litellm
from litellm._logging import verbose_logger
@ -34,7 +34,7 @@ class A2ARequestUtils:
else:
parts = getattr(message, "parts", []) or []
text_parts: list[str] = []
text_parts: Final[list[str]] = []
for part in parts:
if isinstance(part, dict):
if part.get("kind") == "text":
@ -56,7 +56,7 @@ class A2ARequestUtils:
Returns:
Text from response message parts
"""
result = response_dict.get("result", {})
result: Final = response_dict.get("result", {})
if not isinstance(result, dict):
return ""
@ -66,7 +66,7 @@ class A2ARequestUtils:
if result.get("kind") == "message":
return A2ARequestUtils.extract_text_from_message(result)
message = result.get("message", {})
message: Final = result.get("message", {})
return A2ARequestUtils.extract_text_from_message(message)
@staticmethod
@ -82,7 +82,7 @@ class A2ARequestUtils:
Returns:
The message object/dict or None
"""
params = getattr(request, "params", None)
params: Final = getattr(request, "params", None)
if params is None:
return None
return getattr(params, "message", None)
@ -128,14 +128,14 @@ class A2ARequestUtils:
input_message = A2ARequestUtils.get_input_message_from_request(request)
if input_message is not None and hasattr(input_message, "model_dump"):
input_message = input_message.model_dump(mode="json")
input_text = A2ARequestUtils.extract_text_from_message(input_message)
prompt_tokens = A2ARequestUtils.count_tokens(input_text)
input_text: Final = A2ARequestUtils.extract_text_from_message(input_message)
prompt_tokens: Final = A2ARequestUtils.count_tokens(input_text)
# Count output tokens
output_text = A2ARequestUtils.extract_text_from_response(response_dict)
completion_tokens = A2ARequestUtils.count_tokens(output_text)
output_text: Final = A2ARequestUtils.extract_text_from_response(response_dict)
completion_tokens: Final = A2ARequestUtils.count_tokens(output_text)
total_tokens = prompt_tokens + completion_tokens
total_tokens: Final = prompt_tokens + completion_tokens
return prompt_tokens, completion_tokens, total_tokens

51
litellm/anthropic_beta_headers_manager.py
View file

@ -25,6 +25,7 @@ Environment Variables:
import json
import os
from importlib.resources import files
from typing import Final
import httpx
@ -46,12 +47,12 @@ class GetAnthropicBetaHeadersConfig:
def load_local_beta_headers_config() -> dict:
"""Load the local backup beta headers config bundled with the package."""
try:
content = json.loads(
content: Final = json.loads(
files("litellm").joinpath("anthropic_beta_headers_config.json").read_text(encoding="utf-8")
)
return content
except Exception as e:
verbose_logger.error(f"Failed to load local beta headers config: {e}")
verbose_logger.error("Failed to load local beta headers config: %s", e)
# Return empty config as fallback
return {
"anthropic": {},
@ -79,14 +80,14 @@ class GetAnthropicBetaHeadersConfig:
return False
# Check for at least one provider key
provider_keys = [
provider_keys: Final = [
"anthropic",
"azure_ai",
"bedrock",
"bedrock_converse",
"vertex_ai",
]
has_provider = any(key in fetched_config for key in provider_keys)
has_provider: Final = any(key in fetched_config for key in provider_keys)
if not has_provider:
verbose_logger.warning(
@ -113,7 +114,7 @@ class GetAnthropicBetaHeadersConfig:
Returns the parsed JSON dict. Raises on network/parse errors
(caller is expected to handle).
"""
response = httpx.get(url, timeout=timeout)
response: Final = httpx.get(url, timeout=timeout)
response.raise_for_status()
return response.json()
@ -138,7 +139,7 @@ def get_beta_headers_config(url: str) -> dict:
return GetAnthropicBetaHeadersConfig.load_local_beta_headers_config()
try:
content = GetAnthropicBetaHeadersConfig.fetch_remote_beta_headers_config(url)
content: Final = GetAnthropicBetaHeadersConfig.fetch_remote_beta_headers_config(url)
except Exception as e:
verbose_logger.warning(
"LiteLLM: Failed to fetch remote beta headers config from %s: %s. Falling back to local backup.",
@ -206,8 +207,8 @@ def get_provider_name(provider: str) -> str:
Returns:
Canonical provider name
"""
config = _load_beta_headers_config()
aliases = config.get("provider_aliases", {})
config: Final = _load_beta_headers_config()
aliases: Final = config.get("provider_aliases", {})
return aliases.get(provider, provider)
@ -233,20 +234,22 @@ def filter_and_transform_beta_headers(
if not beta_headers:
return []
config = _load_beta_headers_config()
config: Final = _load_beta_headers_config()
provider = get_provider_name(provider)
# Get the header mapping for this provider
provider_mapping = config.get(provider, {})
provider_mapping: Final = config.get(provider, {})
filtered_headers: set[str] = set()
filtered_headers: Final[set[str]] = set()
for header in beta_headers:
header = header.strip()
# Check if header is in the mapping
if header not in provider_mapping:
verbose_logger.debug(f"Dropping unknown beta header '{header}' for provider '{provider}' (not in mapping)")
verbose_logger.debug(
"Dropping unknown beta header '%s' for provider '%s' (not in mapping)", header, provider
)
continue
# Get the mapped header value
@ -254,7 +257,7 @@ def filter_and_transform_beta_headers(
# Skip if header is unsupported (null value)
if mapped_header is None:
verbose_logger.debug(f"Dropping unsupported beta header '{header}' for provider '{provider}'")
verbose_logger.debug("Dropping unsupported beta header '%s' for provider '%s'", header, provider)
continue
# Add the mapped header
@ -277,9 +280,9 @@ def is_beta_header_supported(
Returns:
True if the header is in the mapping with a non-null value, False otherwise
"""
config = _load_beta_headers_config()
config: Final = _load_beta_headers_config()
provider = get_provider_name(provider)
provider_mapping = config.get(provider, {})
provider_mapping: Final = config.get(provider, {})
# Header is supported if it's in the mapping and has a non-null value
return beta_header in provider_mapping and provider_mapping[beta_header] is not None
@ -301,11 +304,11 @@ def get_provider_beta_header(
Returns:
The provider-specific header name if supported, or None if unsupported/unknown
"""
config = _load_beta_headers_config()
config: Final = _load_beta_headers_config()
provider = get_provider_name(provider)
# Get the header mapping for this provider
provider_mapping = config.get(provider, {})
provider_mapping: Final = config.get(provider, {})
# Check if header is in the mapping
if anthropic_beta_header not in provider_mapping:
@ -330,15 +333,15 @@ def update_headers_with_filtered_beta(
Returns:
Updated headers dict
"""
existing_beta = headers.get("anthropic-beta")
existing_beta: Final = headers.get("anthropic-beta")
if not existing_beta:
return headers
# Parse existing beta headers
beta_values = [b.strip() for b in existing_beta.split(",") if b.strip()]
beta_values: Final = [b.strip() for b in existing_beta.split(",") if b.strip()]
# Filter and transform based on provider
filtered_beta_values = filter_and_transform_beta_headers(
filtered_beta_values: Final = filter_and_transform_beta_headers(
beta_headers=beta_values,
provider=provider,
)
@ -372,11 +375,11 @@ def update_request_with_filtered_beta(
"""
headers = update_headers_with_filtered_beta(headers=headers, provider=provider)
existing_body_betas = request_data.get("anthropic_beta")
existing_body_betas: Final = request_data.get("anthropic_beta")
if not existing_body_betas:
return headers, request_data
filtered_body_betas = filter_and_transform_beta_headers(
filtered_body_betas: Final = filter_and_transform_beta_headers(
beta_headers=existing_body_betas,
provider=provider,
)
@ -399,9 +402,9 @@ def get_unsupported_headers(provider: str) -> list[str]:
Returns:
List of unsupported Anthropic beta header names
"""
config = _load_beta_headers_config()
config: Final = _load_beta_headers_config()
provider = get_provider_name(provider)
provider_mapping = config.get(provider, {})
provider_mapping: Final = config.get(provider, {})
# Return headers with null values
return [header for header, value in provider_mapping.items() if value is None]

12
litellm/anthropic_interface/exceptions/exception_mapping_utils.py
View file

@ -4,13 +4,15 @@ Utilities for mapping exceptions to Anthropic error format.
Similar to litellm/litellm_core_utils/exception_mapping_utils.py but for Anthropic response format.
"""
from typing import Final
from litellm.litellm_core_utils.safe_json_loads import safe_json_loads
from .exceptions import AnthropicErrorResponse, AnthropicErrorType
# HTTP status code -> Anthropic error type
# Source: https://docs.anthropic.com/en/api/errors
ANTHROPIC_ERROR_TYPE_MAP: dict[int, AnthropicErrorType] = {
ANTHROPIC_ERROR_TYPE_MAP: Final[dict[int, AnthropicErrorType]] = {
400: "invalid_request_error",
401: "authentication_error",
403: "permission_error",
@ -50,9 +52,9 @@ class AnthropicExceptionMapping:
"request_id": "req_..."
}
"""
error_type = AnthropicExceptionMapping.get_error_type(status_code)
error_type: Final = AnthropicExceptionMapping.get_error_type(status_code)
response: AnthropicErrorResponse = {
response: Final[AnthropicErrorResponse] = {
"type": "error",
"error": {
"type": error_type,
@ -76,7 +78,7 @@ class AnthropicExceptionMapping:
- Generic: {"message": "..."}
- Plain strings
"""
parsed = safe_json_loads(raw_message)
parsed: Final = safe_json_loads(raw_message)
if isinstance(parsed, dict):
# Bedrock format
if "detail" in parsed and isinstance(parsed["detail"], dict):
@ -151,7 +153,7 @@ class AnthropicExceptionMapping:
# Optionally add request_id if provided and not present
if request_id and "request_id" not in parsed:
parsed["request_id"] = request_id
return parsed # type: ignore
return parsed
# Extract message - use parsed dict if available, otherwise raw string
if parsed is not None:

358
litellm/assistants/main.py
View file

@ -5,7 +5,7 @@ import contextvars
import os
from collections.abc import Coroutine, Iterable
from functools import partial
from typing import Any, Literal
from typing import Any, Final, Literal
import httpx
from openai import AsyncOpenAI, OpenAI
@ -29,8 +29,8 @@ from ..types.router import *
from .utils import get_optional_params_add_message
####### ENVIRONMENT VARIABLES ###################
openai_assistants_api = OpenAIAssistantsAPI()
azure_assistants_api = AzureAssistantsAPI()
openai_assistants_api: Final = OpenAIAssistantsAPI()
azure_assistants_api: Final = AzureAssistantsAPI()
### ASSISTANTS ###
@ -40,28 +40,26 @@ async def aget_assistants(
client: AsyncOpenAI | None = None,
**kwargs,
) -> AsyncCursorPage[Assistant]:
loop = asyncio.get_event_loop()
loop: Final = asyncio.get_event_loop()
### PASS ARGS TO GET ASSISTANTS ###
kwargs["aget_assistants"] = True
try:
# Use a partial function to pass your keyword arguments
func = partial(get_assistants, custom_llm_provider, client, **kwargs)
func: Final = partial(get_assistants, custom_llm_provider, client, **kwargs)
# Add the context to the function
ctx = contextvars.copy_context()
func_with_context = partial(ctx.run, func)
ctx: Final = contextvars.copy_context()
func_with_context: Final = partial(ctx.run, func)
_, custom_llm_provider, _, _ = get_llm_provider( # type: ignore
model="", custom_llm_provider=custom_llm_provider
) # type: ignore
_, custom_llm_provider, _, _ = get_llm_provider(model="", custom_llm_provider=custom_llm_provider)
# Await normally
init_response = await loop.run_in_executor(None, func_with_context)
init_response: Final = await loop.run_in_executor(None, func_with_context)
if asyncio.iscoroutine(init_response):
response = await init_response
else:
response = init_response
return response # type: ignore
return response
except Exception as e:
raise exception_type(
model="",
@ -80,11 +78,11 @@ def get_assistants(
api_version: str | None = None,
**kwargs,
) -> SyncCursorPage[Assistant]:
aget_assistants: bool | None = kwargs.pop("aget_assistants", None)
aget_assistants: Final[bool | None] = kwargs.pop("aget_assistants", None)
if aget_assistants is not None and not isinstance(aget_assistants, bool):
raise Exception("Invalid value passed in for aget_assistants. Only bool or None allowed")
optional_params = GenericLiteLLMParams(api_key=api_key, api_base=api_base, api_version=api_version, **kwargs)
litellm_params_dict = get_litellm_params(**kwargs)
optional_params: Final = GenericLiteLLMParams(api_key=api_key, api_base=api_base, api_version=api_version, **kwargs)
litellm_params_dict: Final = get_litellm_params(**kwargs)
### TIMEOUT LOGIC ###
timeout = optional_params.timeout or kwargs.get("request_timeout", 600) or 600
@ -95,10 +93,10 @@ def get_assistants(
and isinstance(timeout, httpx.Timeout)
and supports_httpx_timeout(custom_llm_provider) is False
):
read_timeout = timeout.read or 600
read_timeout: Final = timeout.read or 600
timeout = read_timeout # default 10 min timeout
elif timeout is not None and not isinstance(timeout, httpx.Timeout):
timeout = float(timeout) # type: ignore
timeout = float(timeout)
elif timeout is None:
timeout = 600.0
@ -111,7 +109,7 @@ def get_assistants(
or os.getenv("OPENAI_API_BASE")
or "https://api.openai.com/v1"
)
organization = (
organization: Final = (
optional_params.organization
or litellm.organization
or os.getenv("OPENAI_ORGANIZATION", None)
@ -132,12 +130,12 @@ def get_assistants(
max_retries=optional_params.max_retries,
organization=organization,
client=client,
aget_assistants=aget_assistants, # type: ignore
) # type: ignore
aget_assistants=aget_assistants,
)
elif custom_llm_provider == "azure":
api_base = optional_params.api_base or litellm.api_base or get_secret("AZURE_API_BASE") # type: ignore
api_base = optional_params.api_base or litellm.api_base or get_secret("AZURE_API_BASE")
api_version = optional_params.api_version or litellm.api_version or get_secret("AZURE_API_VERSION") # type: ignore
api_version = optional_params.api_version or litellm.api_version or get_secret("AZURE_API_VERSION")
api_key = (
optional_params.api_key
@ -145,14 +143,14 @@ def get_assistants(
or litellm.azure_key
or get_secret("AZURE_OPENAI_API_KEY")
or get_secret("AZURE_API_KEY")
) # type: ignore
)
extra_body = optional_params.get("extra_body", {})
extra_body: Final = optional_params.get("extra_body", {})
azure_ad_token: str | None = None
if extra_body is not None:
azure_ad_token = extra_body.pop("azure_ad_token", None)
else:
azure_ad_token = get_secret("AZURE_AD_TOKEN") # type: ignore
azure_ad_token = get_secret("AZURE_AD_TOKEN")
response = azure_assistants_api.get_assistants(
api_base=api_base,
@ -162,7 +160,7 @@ def get_assistants(
timeout=timeout,
max_retries=optional_params.max_retries,
client=client,
aget_assistants=aget_assistants, # type: ignore
aget_assistants=aget_assistants,
litellm_params=litellm_params_dict,
)
else:
@ -173,7 +171,7 @@ def get_assistants(
response=httpx.Response(
status_code=400,
content="Unsupported provider",
request=httpx.Request(method="create_thread", url="https://github.com/BerriAI/litellm"), # type: ignore
request=httpx.Request(method="create_thread", url="https://github.com/BerriAI/litellm"),
),
)
@ -185,7 +183,7 @@ def get_assistants(
response=httpx.Response(
status_code=400,
content="Unsupported provider",
request=httpx.Request(method="create_thread", url="https://github.com/BerriAI/litellm"), # type: ignore
request=httpx.Request(method="create_thread", url="https://github.com/BerriAI/litellm"),
),
)
@ -197,30 +195,28 @@ async def acreate_assistants(
client: AsyncOpenAI | None = None,
**kwargs,
) -> Assistant:
loop = asyncio.get_event_loop()
loop: Final = asyncio.get_event_loop()
### PASS ARGS TO GET ASSISTANTS ###
kwargs["async_create_assistants"] = True
model = kwargs.pop("model", None)
model: Final = kwargs.pop("model", None)
try:
kwargs["client"] = client
# Use a partial function to pass your keyword arguments
func = partial(create_assistants, custom_llm_provider, model, **kwargs)
func: Final = partial(create_assistants, custom_llm_provider, model, **kwargs)
# Add the context to the function
ctx = contextvars.copy_context()
func_with_context = partial(ctx.run, func)
ctx: Final = contextvars.copy_context()
func_with_context: Final = partial(ctx.run, func)
_, custom_llm_provider, _, _ = get_llm_provider( # type: ignore
model=model, custom_llm_provider=custom_llm_provider
) # type: ignore
_, custom_llm_provider, _, _ = get_llm_provider(model=model, custom_llm_provider=custom_llm_provider)
# Await normally
init_response = await loop.run_in_executor(None, func_with_context)
init_response: Final = await loop.run_in_executor(None, func_with_context)
if asyncio.iscoroutine(init_response):
response = await init_response
else:
response = init_response
return response # type: ignore
return response
except Exception as e:
raise exception_type(
model=model,
@ -249,11 +245,11 @@ def create_assistants(
api_version: str | None = None,
**kwargs,
) -> Assistant | Coroutine[Any, Any, Assistant]:
async_create_assistants: bool | None = kwargs.pop("async_create_assistants", None)
async_create_assistants: Final[bool | None] = kwargs.pop("async_create_assistants", None)
if async_create_assistants is not None and not isinstance(async_create_assistants, bool):
raise ValueError("Invalid value passed in for async_create_assistants. Only bool or None allowed")
optional_params = GenericLiteLLMParams(api_key=api_key, api_base=api_base, api_version=api_version, **kwargs)
litellm_params_dict = get_litellm_params(**kwargs)
optional_params: Final = GenericLiteLLMParams(api_key=api_key, api_base=api_base, api_version=api_version, **kwargs)
litellm_params_dict: Final = get_litellm_params(**kwargs)
### TIMEOUT LOGIC ###
timeout = optional_params.timeout or kwargs.get("request_timeout", 600) or 600
@ -264,10 +260,10 @@ def create_assistants(
and isinstance(timeout, httpx.Timeout)
and supports_httpx_timeout(custom_llm_provider) is False
):
read_timeout = timeout.read or 600
read_timeout: Final = timeout.read or 600
timeout = read_timeout # default 10 min timeout
elif timeout is not None and not isinstance(timeout, httpx.Timeout):
timeout = float(timeout) # type: ignore
timeout = float(timeout)
elif timeout is None:
timeout = 600.0
@ -296,7 +292,7 @@ def create_assistants(
or os.getenv("OPENAI_API_BASE")
or "https://api.openai.com/v1"
)
organization = (
organization: Final = (
optional_params.organization
or litellm.organization
or os.getenv("OPENAI_ORGANIZATION", None)
@ -318,12 +314,12 @@ def create_assistants(
organization=organization,
create_assistant_data=create_assistant_data,
client=client,
async_create_assistants=async_create_assistants, # type: ignore
) # type: ignore
async_create_assistants=async_create_assistants,
)
elif custom_llm_provider == "azure":
api_base = optional_params.api_base or litellm.api_base or get_secret("AZURE_API_BASE") # type: ignore
api_base = optional_params.api_base or litellm.api_base or get_secret("AZURE_API_BASE")
api_version = optional_params.api_version or litellm.api_version or get_secret("AZURE_API_VERSION") # type: ignore
api_version = optional_params.api_version or litellm.api_version or get_secret("AZURE_API_VERSION")
api_key = (
optional_params.api_key
@ -331,14 +327,14 @@ def create_assistants(
or litellm.azure_key
or get_secret("AZURE_OPENAI_API_KEY")
or get_secret("AZURE_API_KEY")
) # type: ignore
)
extra_body = optional_params.get("extra_body", {})
extra_body: Final = optional_params.get("extra_body", {})
azure_ad_token: str | None = None
if extra_body is not None:
azure_ad_token = extra_body.pop("azure_ad_token", None)
else:
azure_ad_token = get_secret("AZURE_AD_TOKEN") # type: ignore
azure_ad_token = get_secret("AZURE_AD_TOKEN")
if isinstance(client, OpenAI):
client = None # only pass client if it's AzureOpenAI
@ -363,7 +359,7 @@ def create_assistants(
response=httpx.Response(
status_code=400,
content="Unsupported provider",
request=httpx.Request(method="create_thread", url="https://github.com/BerriAI/litellm"), # type: ignore
request=httpx.Request(method="create_thread", url="https://github.com/BerriAI/litellm"),
),
)
if response is None:
@ -380,29 +376,27 @@ async def adelete_assistant(
client: AsyncOpenAI | None = None,
**kwargs,
) -> AssistantDeleted:
loop = asyncio.get_event_loop()
loop: Final = asyncio.get_event_loop()
### PASS ARGS TO GET ASSISTANTS ###
kwargs["async_delete_assistants"] = True
try:
kwargs["client"] = client
# Use a partial function to pass your keyword arguments
func = partial(delete_assistant, custom_llm_provider, **kwargs)
func: Final = partial(delete_assistant, custom_llm_provider, **kwargs)
# Add the context to the function
ctx = contextvars.copy_context()
func_with_context = partial(ctx.run, func)
ctx: Final = contextvars.copy_context()
func_with_context: Final = partial(ctx.run, func)
_, custom_llm_provider, _, _ = get_llm_provider( # type: ignore
model="", custom_llm_provider=custom_llm_provider
) # type: ignore
_, custom_llm_provider, _, _ = get_llm_provider(model="", custom_llm_provider=custom_llm_provider)
# Await normally
init_response = await loop.run_in_executor(None, func_with_context)
init_response: Final = await loop.run_in_executor(None, func_with_context)
if asyncio.iscoroutine(init_response):
response = await init_response
else:
response = init_response
return response # type: ignore
return response
except Exception as e:
raise exception_type(
model="",
@ -422,11 +416,11 @@ def delete_assistant(
api_version: str | None = None,
**kwargs,
) -> AssistantDeleted | Coroutine[Any, Any, AssistantDeleted]:
optional_params = GenericLiteLLMParams(api_key=api_key, api_base=api_base, api_version=api_version, **kwargs)
optional_params: Final = GenericLiteLLMParams(api_key=api_key, api_base=api_base, api_version=api_version, **kwargs)
litellm_params_dict = get_litellm_params(**kwargs)
litellm_params_dict: Final = get_litellm_params(**kwargs)
async_delete_assistants: bool | None = kwargs.pop("async_delete_assistants", None)
async_delete_assistants: Final[bool | None] = kwargs.pop("async_delete_assistants", None)
if async_delete_assistants is not None and not isinstance(async_delete_assistants, bool):
raise ValueError("Invalid value passed in for async_delete_assistants. Only bool or None allowed")
@ -439,10 +433,10 @@ def delete_assistant(
and isinstance(timeout, httpx.Timeout)
and supports_httpx_timeout(custom_llm_provider) is False
):
read_timeout = timeout.read or 600
read_timeout: Final = timeout.read or 600
timeout = read_timeout # default 10 min timeout
elif timeout is not None and not isinstance(timeout, httpx.Timeout):
timeout = float(timeout) # type: ignore
timeout = float(timeout)
elif timeout is None:
timeout = 600.0
@ -455,7 +449,7 @@ def delete_assistant(
or os.getenv("OPENAI_API_BASE")
or "https://api.openai.com/v1"
)
organization = (
organization: Final = (
optional_params.organization or litellm.organization or os.getenv("OPENAI_ORGANIZATION", None) or None
)
# set API KEY
@ -472,9 +466,9 @@ def delete_assistant(
async_delete_assistants=async_delete_assistants,
)
elif custom_llm_provider == "azure":
api_base = optional_params.api_base or litellm.api_base or get_secret("AZURE_API_BASE") # type: ignore
api_base = optional_params.api_base or litellm.api_base or get_secret("AZURE_API_BASE")
api_version = optional_params.api_version or litellm.api_version or get_secret("AZURE_API_VERSION") # type: ignore
api_version = optional_params.api_version or litellm.api_version or get_secret("AZURE_API_VERSION")
api_key = (
optional_params.api_key
@ -482,14 +476,14 @@ def delete_assistant(
or litellm.azure_key
or get_secret("AZURE_OPENAI_API_KEY")
or get_secret("AZURE_API_KEY")
) # type: ignore
)
extra_body = optional_params.get("extra_body", {})
extra_body: Final = optional_params.get("extra_body", {})
azure_ad_token: str | None = None
if extra_body is not None:
azure_ad_token = extra_body.pop("azure_ad_token", None)
else:
azure_ad_token = get_secret("AZURE_AD_TOKEN") # type: ignore
azure_ad_token = get_secret("AZURE_AD_TOKEN")
if isinstance(client, OpenAI):
client = None # only pass client if it's AzureOpenAI
@ -530,28 +524,26 @@ def delete_assistant(
async def acreate_thread(custom_llm_provider: Literal["openai", "azure"], **kwargs) -> Thread:
loop = asyncio.get_event_loop()
loop: Final = asyncio.get_event_loop()
### PASS ARGS TO GET ASSISTANTS ###
kwargs["acreate_thread"] = True
try:
# Use a partial function to pass your keyword arguments
func = partial(create_thread, custom_llm_provider, **kwargs)
func: Final = partial(create_thread, custom_llm_provider, **kwargs)
# Add the context to the function
ctx = contextvars.copy_context()
func_with_context = partial(ctx.run, func)
ctx: Final = contextvars.copy_context()
func_with_context: Final = partial(ctx.run, func)
_, custom_llm_provider, _, _ = get_llm_provider( # type: ignore
model="", custom_llm_provider=custom_llm_provider
) # type: ignore
_, custom_llm_provider, _, _ = get_llm_provider(model="", custom_llm_provider=custom_llm_provider)
# Await normally
init_response = await loop.run_in_executor(None, func_with_context)
init_response: Final = await loop.run_in_executor(None, func_with_context)
if asyncio.iscoroutine(init_response):
response = await init_response
else:
response = init_response
return response # type: ignore
return response
except Exception as e:
raise exception_type(
model="",
@ -592,9 +584,9 @@ def create_thread(
)
```
"""
acreate_thread = kwargs.get("acreate_thread", None)
optional_params = GenericLiteLLMParams(**kwargs)
litellm_params_dict = get_litellm_params(**kwargs)
acreate_thread: Final = kwargs.get("acreate_thread", None)
optional_params: Final = GenericLiteLLMParams(**kwargs)
litellm_params_dict: Final = get_litellm_params(**kwargs)
### TIMEOUT LOGIC ###
timeout = optional_params.timeout or kwargs.get("request_timeout", 600) or 600
@ -605,10 +597,10 @@ def create_thread(
and isinstance(timeout, httpx.Timeout)
and supports_httpx_timeout(custom_llm_provider) is False
):
read_timeout = timeout.read or 600
read_timeout: Final = timeout.read or 600
timeout = read_timeout # default 10 min timeout
elif timeout is not None and not isinstance(timeout, httpx.Timeout):
timeout = float(timeout) # type: ignore
timeout = float(timeout)
elif timeout is None:
timeout = 600.0
@ -624,7 +616,7 @@ def create_thread(
or os.getenv("OPENAI_API_BASE")
or "https://api.openai.com/v1"
)
organization = (
organization: Final = (
optional_params.organization
or litellm.organization
or os.getenv("OPENAI_ORGANIZATION", None)
@ -649,7 +641,7 @@ def create_thread(
acreate_thread=acreate_thread,
)
elif custom_llm_provider == "azure":
api_base = optional_params.api_base or litellm.api_base or get_secret("AZURE_API_BASE") # type: ignore
api_base = optional_params.api_base or litellm.api_base or get_secret("AZURE_API_BASE")
api_key = (
optional_params.api_key
@ -657,16 +649,16 @@ def create_thread(
or litellm.azure_key
or get_secret("AZURE_OPENAI_API_KEY")
or get_secret("AZURE_API_KEY")
) # type: ignore
)
api_version: str | None = optional_params.api_version or litellm.api_version or get_secret("AZURE_API_VERSION") # type: ignore
api_version: str | None = optional_params.api_version or litellm.api_version or get_secret("AZURE_API_VERSION")
extra_body = optional_params.get("extra_body", {})
extra_body: Final = optional_params.get("extra_body", {})
azure_ad_token: str | None = None
if extra_body is not None:
azure_ad_token = extra_body.pop("azure_ad_token", None)
else:
azure_ad_token = get_secret("AZURE_AD_TOKEN") # type: ignore
azure_ad_token = get_secret("AZURE_AD_TOKEN")
if isinstance(client, OpenAI):
client = None # only pass client if it's AzureOpenAI
@ -692,10 +684,10 @@ def create_thread(
response=httpx.Response(
status_code=400,
content="Unsupported provider",
request=httpx.Request(method="create_thread", url="https://github.com/BerriAI/litellm"), # type: ignore
request=httpx.Request(method="create_thread", url="https://github.com/BerriAI/litellm"),
),
)
return response # type: ignore
return response
async def aget_thread(
@ -704,28 +696,26 @@ async def aget_thread(
client: AsyncOpenAI | None = None,
**kwargs,
) -> Thread:
loop = asyncio.get_event_loop()
loop: Final = asyncio.get_event_loop()
### PASS ARGS TO GET ASSISTANTS ###
kwargs["aget_thread"] = True
try:
# Use a partial function to pass your keyword arguments
func = partial(get_thread, custom_llm_provider, thread_id, client, **kwargs)
func: Final = partial(get_thread, custom_llm_provider, thread_id, client, **kwargs)
# Add the context to the function
ctx = contextvars.copy_context()
func_with_context = partial(ctx.run, func)
ctx: Final = contextvars.copy_context()
func_with_context: Final = partial(ctx.run, func)
_, custom_llm_provider, _, _ = get_llm_provider( # type: ignore
model="", custom_llm_provider=custom_llm_provider
) # type: ignore
_, custom_llm_provider, _, _ = get_llm_provider(model="", custom_llm_provider=custom_llm_provider)
# Await normally
init_response = await loop.run_in_executor(None, func_with_context)
init_response: Final = await loop.run_in_executor(None, func_with_context)
if asyncio.iscoroutine(init_response):
response = await init_response
else:
response = init_response
return response # type: ignore
return response
except Exception as e:
raise exception_type(
model="",
@ -743,9 +733,9 @@ def get_thread(
**kwargs,
) -> Thread:
"""Get the thread object, given a thread_id"""
aget_thread = kwargs.pop("aget_thread", None)
optional_params = GenericLiteLLMParams(**kwargs)
litellm_params_dict = get_litellm_params(**kwargs)
aget_thread: Final = kwargs.pop("aget_thread", None)
optional_params: Final = GenericLiteLLMParams(**kwargs)
litellm_params_dict: Final = get_litellm_params(**kwargs)
### TIMEOUT LOGIC ###
timeout = optional_params.timeout or kwargs.get("request_timeout", 600) or 600
# set timeout for 10 minutes by default
@ -755,10 +745,10 @@ def get_thread(
and isinstance(timeout, httpx.Timeout)
and supports_httpx_timeout(custom_llm_provider) is False
):
read_timeout = timeout.read or 600
read_timeout: Final = timeout.read or 600
timeout = read_timeout # default 10 min timeout
elif timeout is not None and not isinstance(timeout, httpx.Timeout):
timeout = float(timeout) # type: ignore
timeout = float(timeout)
elif timeout is None:
timeout = 600.0
api_base: str | None = None
@ -772,7 +762,7 @@ def get_thread(
or os.getenv("OPENAI_API_BASE")
or "https://api.openai.com/v1"
)
organization = (
organization: Final = (
optional_params.organization
or litellm.organization
or os.getenv("OPENAI_ORGANIZATION", None)
@ -797,9 +787,9 @@ def get_thread(
aget_thread=aget_thread,
)
elif custom_llm_provider == "azure":
api_base = optional_params.api_base or litellm.api_base or get_secret("AZURE_API_BASE") # type: ignore
api_base = optional_params.api_base or litellm.api_base or get_secret("AZURE_API_BASE")
api_version: str | None = optional_params.api_version or litellm.api_version or get_secret("AZURE_API_VERSION") # type: ignore
api_version: str | None = optional_params.api_version or litellm.api_version or get_secret("AZURE_API_VERSION")
api_key = (
optional_params.api_key
@ -807,14 +797,14 @@ def get_thread(
or litellm.azure_key
or get_secret("AZURE_OPENAI_API_KEY")
or get_secret("AZURE_API_KEY")
) # type: ignore
)
extra_body = optional_params.get("extra_body", {})
extra_body: Final = optional_params.get("extra_body", {})
azure_ad_token: str | None = None
if extra_body is not None:
azure_ad_token = extra_body.pop("azure_ad_token", None)
else:
azure_ad_token = get_secret("AZURE_AD_TOKEN") # type: ignore
azure_ad_token = get_secret("AZURE_AD_TOKEN")
if isinstance(client, OpenAI):
client = None # only pass client if it's AzureOpenAI
@ -839,10 +829,10 @@ def get_thread(
response=httpx.Response(
status_code=400,
content="Unsupported provider",
request=httpx.Request(method="create_thread", url="https://github.com/BerriAI/litellm"), # type: ignore
request=httpx.Request(method="create_thread", url="https://github.com/BerriAI/litellm"),
),
)
return response # type: ignore
return response
### MESSAGES ###
@ -858,12 +848,12 @@ async def a_add_message(
client=None,
**kwargs,
) -> OpenAIMessage:
loop = asyncio.get_event_loop()
loop: Final = asyncio.get_event_loop()
### PASS ARGS TO GET ASSISTANTS ###
kwargs["a_add_message"] = True
try:
# Use a partial function to pass your keyword arguments
func = partial(
func: Final = partial(
add_message,
custom_llm_provider,
thread_id,
@ -876,21 +866,19 @@ async def a_add_message(
)
# Add the context to the function
ctx = contextvars.copy_context()
func_with_context = partial(ctx.run, func)
ctx: Final = contextvars.copy_context()
func_with_context: Final = partial(ctx.run, func)
_, custom_llm_provider, _, _ = get_llm_provider( # type: ignore
model="", custom_llm_provider=custom_llm_provider
) # type: ignore
_, custom_llm_provider, _, _ = get_llm_provider(model="", custom_llm_provider=custom_llm_provider)
# Await normally
init_response = await loop.run_in_executor(None, func_with_context)
init_response: Final = await loop.run_in_executor(None, func_with_context)
if asyncio.iscoroutine(init_response):
response = await init_response
else:
# Call the synchronous function using run_in_executor
response = init_response
return response # type: ignore
return response
except Exception as e:
raise exception_type(
model="",
@ -912,12 +900,12 @@ def add_message(
**kwargs,
) -> OpenAIMessage:
### COMMON OBJECTS ###
a_add_message = kwargs.pop("a_add_message", None)
_message_data = MessageData(role=role, content=content, attachments=attachments, metadata=metadata)
litellm_params_dict = get_litellm_params(**kwargs)
optional_params = GenericLiteLLMParams(**kwargs)
a_add_message: Final = kwargs.pop("a_add_message", None)
_message_data: Final = MessageData(role=role, content=content, attachments=attachments, metadata=metadata)
litellm_params_dict: Final = get_litellm_params(**kwargs)
optional_params: Final = GenericLiteLLMParams(**kwargs)
message_data = get_optional_params_add_message(
message_data: Final = get_optional_params_add_message(
role=_message_data["role"],
content=_message_data["content"],
attachments=_message_data["attachments"],
@ -934,10 +922,10 @@ def add_message(
and isinstance(timeout, httpx.Timeout)
and supports_httpx_timeout(custom_llm_provider) is False
):
read_timeout = timeout.read or 600
read_timeout: Final = timeout.read or 600
timeout = read_timeout # default 10 min timeout
elif timeout is not None and not isinstance(timeout, httpx.Timeout):
timeout = float(timeout) # type: ignore
timeout = float(timeout)
elif timeout is None:
timeout = 600.0
api_key: str | None = None
@ -951,7 +939,7 @@ def add_message(
or os.getenv("OPENAI_API_BASE")
or "https://api.openai.com/v1"
)
organization = (
organization: Final = (
optional_params.organization
or litellm.organization
or os.getenv("OPENAI_ORGANIZATION", None)
@ -976,9 +964,9 @@ def add_message(
a_add_message=a_add_message,
)
elif custom_llm_provider == "azure":
api_base = optional_params.api_base or litellm.api_base or get_secret("AZURE_API_BASE") # type: ignore
api_base = optional_params.api_base or litellm.api_base or get_secret("AZURE_API_BASE")
api_version: str | None = optional_params.api_version or litellm.api_version or get_secret("AZURE_API_VERSION") # type: ignore
api_version: str | None = optional_params.api_version or litellm.api_version or get_secret("AZURE_API_VERSION")
api_key = (
optional_params.api_key
@ -986,14 +974,14 @@ def add_message(
or litellm.azure_key
or get_secret("AZURE_OPENAI_API_KEY")
or get_secret("AZURE_API_KEY")
) # type: ignore
)
extra_body = optional_params.get("extra_body", {})
extra_body: Final = optional_params.get("extra_body", {})
azure_ad_token: str | None = None
if extra_body is not None:
azure_ad_token = extra_body.pop("azure_ad_token", None)
else:
azure_ad_token = get_secret("AZURE_AD_TOKEN") # type: ignore
azure_ad_token = get_secret("AZURE_AD_TOKEN")
response = azure_assistants_api.add_message(
thread_id=thread_id,
@ -1016,11 +1004,11 @@ def add_message(
response=httpx.Response(
status_code=400,
content="Unsupported provider",
request=httpx.Request(method="create_thread", url="https://github.com/BerriAI/litellm"), # type: ignore
request=httpx.Request(method="create_thread", url="https://github.com/BerriAI/litellm"),
),
)
return response # type: ignore
return response
async def aget_messages(
@ -1029,12 +1017,12 @@ async def aget_messages(
client: AsyncOpenAI | None = None,
**kwargs,
) -> AsyncCursorPage[OpenAIMessage]:
loop = asyncio.get_event_loop()
loop: Final = asyncio.get_event_loop()
### PASS ARGS TO GET ASSISTANTS ###
kwargs["aget_messages"] = True
try:
# Use a partial function to pass your keyword arguments
func = partial(
func: Final = partial(
get_messages,
custom_llm_provider,
thread_id,
@ -1043,21 +1031,19 @@ async def aget_messages(
)
# Add the context to the function
ctx = contextvars.copy_context()
func_with_context = partial(ctx.run, func)
ctx: Final = contextvars.copy_context()
func_with_context: Final = partial(ctx.run, func)
_, custom_llm_provider, _, _ = get_llm_provider( # type: ignore
model="", custom_llm_provider=custom_llm_provider
) # type: ignore
_, custom_llm_provider, _, _ = get_llm_provider(model="", custom_llm_provider=custom_llm_provider)
# Await normally
init_response = await loop.run_in_executor(None, func_with_context)
init_response: Final = await loop.run_in_executor(None, func_with_context)
if asyncio.iscoroutine(init_response):
response = await init_response
else:
# Call the synchronous function using run_in_executor
response = init_response
return response # type: ignore
return response
except Exception as e:
raise exception_type(
model="",
@ -1074,9 +1060,9 @@ def get_messages(
client: Any | None = None,
**kwargs,
) -> SyncCursorPage[OpenAIMessage]:
aget_messages = kwargs.pop("aget_messages", None)
optional_params = GenericLiteLLMParams(**kwargs)
litellm_params_dict = get_litellm_params(**kwargs)
aget_messages: Final = kwargs.pop("aget_messages", None)
optional_params: Final = GenericLiteLLMParams(**kwargs)
litellm_params_dict: Final = get_litellm_params(**kwargs)
### TIMEOUT LOGIC ###
timeout = optional_params.timeout or kwargs.get("request_timeout", 600) or 600
@ -1087,10 +1073,10 @@ def get_messages(
and isinstance(timeout, httpx.Timeout)
and supports_httpx_timeout(custom_llm_provider) is False
):
read_timeout = timeout.read or 600
read_timeout: Final = timeout.read or 600
timeout = read_timeout # default 10 min timeout
elif timeout is not None and not isinstance(timeout, httpx.Timeout):
timeout = float(timeout) # type: ignore
timeout = float(timeout)
elif timeout is None:
timeout = 600.0
@ -1105,7 +1091,7 @@ def get_messages(
or os.getenv("OPENAI_API_BASE")
or "https://api.openai.com/v1"
)
organization = (
organization: Final = (
optional_params.organization
or litellm.organization
or os.getenv("OPENAI_ORGANIZATION", None)
@ -1129,9 +1115,9 @@ def get_messages(
aget_messages=aget_messages,
)
elif custom_llm_provider == "azure":
api_base = optional_params.api_base or litellm.api_base or get_secret("AZURE_API_BASE") # type: ignore
api_base = optional_params.api_base or litellm.api_base or get_secret("AZURE_API_BASE")
api_version: str | None = optional_params.api_version or litellm.api_version or get_secret("AZURE_API_VERSION") # type: ignore
api_version: str | None = optional_params.api_version or litellm.api_version or get_secret("AZURE_API_VERSION")
api_key = (
optional_params.api_key
@ -1139,14 +1125,14 @@ def get_messages(
or litellm.azure_key
or get_secret("AZURE_OPENAI_API_KEY")
or get_secret("AZURE_API_KEY")
) # type: ignore
)
extra_body = optional_params.get("extra_body", {})
extra_body: Final = optional_params.get("extra_body", {})
azure_ad_token: str | None = None
if extra_body is not None:
azure_ad_token = extra_body.pop("azure_ad_token", None)
else:
azure_ad_token = get_secret("AZURE_AD_TOKEN") # type: ignore
azure_ad_token = get_secret("AZURE_AD_TOKEN")
response = azure_assistants_api.get_messages(
thread_id=thread_id,
@ -1168,11 +1154,11 @@ def get_messages(
response=httpx.Response(
status_code=400,
content="Unsupported provider",
request=httpx.Request(method="create_thread", url="https://github.com/BerriAI/litellm"), # type: ignore
request=httpx.Request(method="create_thread", url="https://github.com/BerriAI/litellm"),
),
)
return response # type: ignore
return response
### RUNS ###
@ -1182,7 +1168,7 @@ def arun_thread_stream(
**kwargs,
) -> AsyncAssistantStreamManager[AsyncAssistantEventHandler]:
kwargs["arun_thread"] = True
return run_thread(stream=True, event_handler=event_handler, **kwargs) # type: ignore
return run_thread(stream=True, event_handler=event_handler, **kwargs)
async def arun_thread(
@ -1198,12 +1184,12 @@ async def arun_thread(
client: Any | None = None,
**kwargs,
) -> Run:
loop = asyncio.get_event_loop()
loop: Final = asyncio.get_event_loop()
### PASS ARGS TO GET ASSISTANTS ###
kwargs["arun_thread"] = True
try:
# Use a partial function to pass your keyword arguments
func = partial(
func: Final = partial(
run_thread,
custom_llm_provider,
thread_id,
@ -1219,21 +1205,19 @@ async def arun_thread(
)
# Add the context to the function
ctx = contextvars.copy_context()
func_with_context = partial(ctx.run, func)
ctx: Final = contextvars.copy_context()
func_with_context: Final = partial(ctx.run, func)
_, custom_llm_provider, _, _ = get_llm_provider( # type: ignore
model="", custom_llm_provider=custom_llm_provider
) # type: ignore
_, custom_llm_provider, _, _ = get_llm_provider(model="", custom_llm_provider=custom_llm_provider)
# Await normally
init_response = await loop.run_in_executor(None, func_with_context)
init_response: Final = await loop.run_in_executor(None, func_with_context)
if asyncio.iscoroutine(init_response):
response = await init_response
else:
# Call the synchronous function using run_in_executor
response = init_response
return response # type: ignore
return response
except Exception as e:
raise exception_type(
model="",
@ -1249,7 +1233,7 @@ def run_thread_stream(
event_handler: AssistantEventHandler | None = None,
**kwargs,
) -> AssistantStreamManager[AssistantEventHandler]:
return run_thread(stream=True, event_handler=event_handler, **kwargs) # type: ignore
return run_thread(stream=True, event_handler=event_handler, **kwargs)
def run_thread(
@ -1267,9 +1251,9 @@ def run_thread(
**kwargs,
) -> Run:
"""Run a given thread + assistant."""
arun_thread = kwargs.pop("arun_thread", None)
optional_params = GenericLiteLLMParams(**kwargs)
litellm_params_dict = get_litellm_params(**kwargs)
arun_thread: Final = kwargs.pop("arun_thread", None)
optional_params: Final = GenericLiteLLMParams(**kwargs)
litellm_params_dict: Final = get_litellm_params(**kwargs)
### TIMEOUT LOGIC ###
timeout = optional_params.timeout or kwargs.get("request_timeout", 600) or 600
@ -1280,10 +1264,10 @@ def run_thread(
and isinstance(timeout, httpx.Timeout)
and supports_httpx_timeout(custom_llm_provider) is False
):
read_timeout = timeout.read or 600
read_timeout: Final = timeout.read or 600
timeout = read_timeout # default 10 min timeout
elif timeout is not None and not isinstance(timeout, httpx.Timeout):
timeout = float(timeout) # type: ignore
timeout = float(timeout)
elif timeout is None:
timeout = 600.0
@ -1296,7 +1280,7 @@ def run_thread(
or os.getenv("OPENAI_API_BASE")
or "https://api.openai.com/v1"
)
organization = (
organization: Final = (
optional_params.organization
or litellm.organization
or os.getenv("OPENAI_ORGANIZATION", None)
@ -1329,9 +1313,9 @@ def run_thread(
event_handler=event_handler,
)
elif custom_llm_provider == "azure":
api_base = optional_params.api_base or litellm.api_base or get_secret("AZURE_API_BASE") # type: ignore
api_base = optional_params.api_base or litellm.api_base or get_secret("AZURE_API_BASE")
api_version = optional_params.api_version or litellm.api_version or get_secret("AZURE_API_VERSION") # type: ignore
api_version = optional_params.api_version or litellm.api_version or get_secret("AZURE_API_VERSION")
api_key = (
optional_params.api_key
@ -1339,14 +1323,14 @@ def run_thread(
or litellm.azure_key
or get_secret("AZURE_OPENAI_API_KEY")
or get_secret("AZURE_API_KEY")
) # type: ignore
)
extra_body = optional_params.get("extra_body", {})
extra_body: Final = optional_params.get("extra_body", {})
azure_ad_token = None
if extra_body is not None:
azure_ad_token = extra_body.pop("azure_ad_token", None)
else:
azure_ad_token = get_secret("AZURE_AD_TOKEN") # type: ignore
azure_ad_token = get_secret("AZURE_AD_TOKEN")
response = azure_assistants_api.run_thread(
thread_id=thread_id,
@ -1366,7 +1350,7 @@ def run_thread(
client=client,
arun_thread=arun_thread,
litellm_params=litellm_params_dict,
) # type: ignore
)
else:
raise litellm.exceptions.BadRequestError(
message=f"LiteLLM doesn't support {custom_llm_provider} for 'run_thread'. Only 'openai' is supported.",
@ -1375,7 +1359,7 @@ def run_thread(
response=httpx.Response(
status_code=400,
content="Unsupported provider",
request=httpx.Request(method="create_thread", url="https://github.com/BerriAI/litellm"), # type: ignore
request=httpx.Request(method="create_thread", url="https://github.com/BerriAI/litellm"),
),
)
return response # type: ignore
return response

20
litellm/assistants/utils.py
View file

@ -1,3 +1,5 @@
from typing import Final
import litellm
from ..exceptions import UnsupportedParamsError
@ -17,13 +19,13 @@ def get_optional_params_add_message(
Reference - https://learn.microsoft.com/en-us/azure/ai-services/openai/assistants-reference-messages?tabs=python#create-message
"""
passed_params = locals()
passed_params: Final = locals()
custom_llm_provider = passed_params.pop("custom_llm_provider")
special_params = passed_params.pop("kwargs")
special_params: Final = passed_params.pop("kwargs")
for k, v in special_params.items():
passed_params[k] = v
default_params = {
default_params: Final = {
"role": None,
"content": None,
"attachments": None,
@ -36,7 +38,7 @@ def get_optional_params_add_message(
## raise exception if non-default value passed for non-openai/azure embedding calls
def _check_valid_arg(supported_params):
if len(non_default_params.keys()) > 0:
keys = list(non_default_params.keys())
keys: Final = list(non_default_params.keys())
for k in keys:
if litellm.drop_params is True and k not in supported_params: # drop the unsupported non-default values
non_default_params.pop(k, None)
@ -50,7 +52,7 @@ def get_optional_params_add_message(
if custom_llm_provider == "openai":
optional_params = non_default_params
elif custom_llm_provider == "azure":
supported_params = litellm.AzureOpenAIAssistantsAPIConfig().get_supported_openai_create_message_params()
supported_params: Final = litellm.AzureOpenAIAssistantsAPIConfig().get_supported_openai_create_message_params()
_check_valid_arg(supported_params=supported_params)
optional_params = litellm.AzureOpenAIAssistantsAPIConfig().map_openai_params_create_message_params(
non_default_params=non_default_params, optional_params=optional_params
@ -72,13 +74,13 @@ def get_optional_params_image_gen(
**kwargs,
):
# retrieve all parameters passed to the function
passed_params = locals()
passed_params: Final = locals()
custom_llm_provider = passed_params.pop("custom_llm_provider")
special_params = passed_params.pop("kwargs")
special_params: Final = passed_params.pop("kwargs")
for k, v in special_params.items():
passed_params[k] = v
default_params = {
default_params: Final = {
"n": None,
"quality": None,
"response_format": None,
@ -93,7 +95,7 @@ def get_optional_params_image_gen(
## raise exception if non-default value passed for non-openai/azure embedding calls
def _check_valid_arg(supported_params):
if len(non_default_params.keys()) > 0:
keys = list(non_default_params.keys())
keys: Final = list(non_default_params.keys())
for k in keys:
if litellm.drop_params is True and k not in supported_params: # drop the unsupported non-default values
non_default_params.pop(k, None)

19
litellm/batch_completion/main.py
View file

@ -1,4 +1,5 @@
from concurrent.futures import FIRST_COMPLETED, ThreadPoolExecutor, wait
from typing import Final
import litellm
from litellm._logging import print_verbose
@ -55,17 +56,17 @@ def batch_completion(
Returns:
list: A list of completion results.
"""
args = locals()
args: Final = locals()
batch_messages = messages
completions = []
batch_messages: Final = messages
completions: Final = []
model = model
custom_llm_provider = None
if model.split("/", 1)[0] in litellm.provider_list:
custom_llm_provider = model.split("/", 1)[0]
model = model.split("/", 1)[1]
if custom_llm_provider == "vllm":
optional_params = get_optional_params(
optional_params: Final = get_optional_params(
functions=functions,
function_call=function_call,
temperature=temperature,
@ -145,7 +146,7 @@ def batch_completion_models(*args, **kwargs):
if "model" in kwargs:
kwargs.pop("model")
if "models" in kwargs:
models = kwargs["models"]
models: Final = kwargs["models"]
kwargs.pop("models")
futures = {}
with ThreadPoolExecutor(max_workers=len(models)) as executor:
@ -156,10 +157,10 @@ def batch_completion_models(*args, **kwargs):
if future.result() is not None:
return future.result()
elif "deployments" in kwargs:
deployments = kwargs["deployments"]
deployments: Final = kwargs["deployments"]
kwargs.pop("deployments")
kwargs.pop("model_list")
nested_kwargs = kwargs.pop("kwargs", {})
nested_kwargs: Final = kwargs.pop("kwargs", {})
futures = {}
with ThreadPoolExecutor(max_workers=len(deployments)) as executor:
for deployment in deployments:
@ -238,10 +239,10 @@ def batch_completion_models_all_responses(*args, **kwargs):
if len(models) == 0:
return []
responses = []
responses: Final = []
with concurrent.futures.ThreadPoolExecutor(max_workers=len(models)) as executor:
futures = [executor.submit(litellm.completion, *args, model=model, **kwargs) for model in models]
futures: Final = [executor.submit(litellm.completion, *args, model=model, **kwargs) for model in models]
for future in futures:
try:

56
litellm/batches/batch_utils.py
View file

@ -1,7 +1,7 @@
import json
from collections.abc import Iterable, Iterator
from dataclasses import dataclass
from typing import Any, Literal
from typing import Any, Final, Literal
import litellm
from litellm._logging import verbose_logger
@ -141,9 +141,9 @@ def _aggregate_batch_cost_usage_models(
) -> 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 = tuple(_iter_successful_output_line_stats(entries, custom_llm_provider, model_name, model_info))
line_stats: Final = tuple(_iter_successful_output_line_stats(entries, custom_llm_provider, model_name, model_info))
cache_token_params = {
cache_token_params: Final = {
key: tokens
for key, tokens in (
("cache_read_input_tokens", sum(stats.cache_read_tokens for stats in line_stats)),
@ -151,14 +151,14 @@ def _aggregate_batch_cost_usage_models(
)
if tokens > 0
}
batch_usage = Usage(
batch_usage: Final = Usage(
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),
**cache_token_params,
)
batch_models = [model_name] if model_name else [stats.model for stats in line_stats if stats.model]
total_cost = sum((stats.cost for stats in line_stats), 0.0)
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
@ -184,7 +184,7 @@ def calculate_vertex_ai_batch_cost_and_usage(
total_tokens = 0
prompt_tokens = 0
completion_tokens = 0
actual_model_name = model_name or "gemini-2.0-flash-001"
actual_model_name: Final = model_name or "gemini-2.0-flash-001"
for response in vertex_ai_batch_responses:
response_body = response.get("response")
@ -254,27 +254,27 @@ async def _fetch_batch_output_file_content(
raise ValueError("Output file id is None cannot retrieve file content")
file_id = batch.output_file_id
is_base64_unified_file_id = _is_base64_encoded_unified_file_id(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(f"Extracted LLM output file ID from unified file ID: {file_id}")
verbose_logger.debug("Extracted LLM output file ID from unified file ID: %s", file_id)
except (IndexError, AttributeError) as e:
verbose_logger.error(
f"Failed to extract LLM output file ID from unified file ID: {batch.output_file_id}, error: {e}"
"Failed to extract LLM output file ID from unified file ID: %s, error: %s", batch.output_file_id, e
)
# Build kwargs for afile_content with credentials from litellm_params
file_content_kwargs = {
file_content_kwargs: Final = {
"file_id": file_id,
"custom_llm_provider": custom_llm_provider,
}
# Extract and add credentials for file access
credentials = _extract_file_access_credentials(litellm_params)
credentials: Final = _extract_file_access_credentials(litellm_params)
file_content_kwargs.update(credentials)
_file_content = await afile_content(**file_content_kwargs) # type: ignore[reportArgumentType]
_file_content: Final = await afile_content(**file_content_kwargs)
return _file_content.content
@ -291,11 +291,11 @@ def _extract_file_access_credentials(litellm_params: dict | None) -> dict:
Returns:
Dictionary containing only the credentials needed for file access
"""
credentials = {}
credentials: Final = {}
if litellm_params:
# List of credential keys that should be passed to file operations
credential_keys = [
credential_keys: Final = [
"api_key",
"api_base",
"api_version",
@ -355,7 +355,7 @@ def _iter_batch_input_entries(file_content: bytes) -> Iterator[dict]:
# A batch request's input tokens scale roughly with its serialized size, so this
# is a conservative per-row fallback when the token counter cannot measure a row.
_BATCH_TOKEN_ESTIMATE_BYTES_PER_TOKEN = 4
_BATCH_TOKEN_ESTIMATE_BYTES_PER_TOKEN: Final = 4
def _estimate_batch_entry_tokens(raw_line: bytes) -> int:
@ -370,18 +370,18 @@ def _count_entry_tokens(
model_name: str | None = None,
) -> int:
"""Token-count a single batch input entry's body (chat / text / embedding)."""
body = entry.get("body", {}) or {}
model = body.get("model", model_name or "")
body: Final = entry.get("body", {}) or {}
model: Final = body.get("model", model_name or "")
messages = body.get("messages")
messages: Final = body.get("messages")
if messages:
return token_counter(model=model, messages=messages)
prompt = body.get("prompt")
prompt: Final = body.get("prompt")
if prompt:
return _count_prompt_or_input_tokens(model=model, value=prompt)
input_data = body.get("input")
input_data: Final = body.get("input")
if input_data:
return _count_prompt_or_input_tokens(model=model, value=input_data)
@ -432,8 +432,12 @@ def _get_batch_job_usage_from_response_body(response_body: dict, custom_llm_prov
usage_object=response_body.get("usage", None) or {},
reasoning_content=None,
)
_usage_dict = response_body.get("usage", None) or {}
usage: Usage = Usage(**_usage_dict)
from litellm.responses.utils import ResponseAPILoggingUtils
_usage_dict: Final = response_body.get("usage", None) or {}
if ResponseAPILoggingUtils._is_response_api_usage(_usage_dict):
return ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(_usage_dict)
usage: Final[Usage] = Usage(**_usage_dict)
return usage
@ -455,8 +459,8 @@ def _get_response_from_batch_job_output_file(batch_job_output_file: dict, custom
return _get_anthropic_result_from_batch_results_line(batch_job_output_file).get("message", None) or {}
if custom_llm_provider == "bedrock":
return batch_job_output_file.get("modelOutput", None) or {}
_response: dict = batch_job_output_file.get("response", None) or {}
_response_body = _response.get("body", None) or {}
_response: Final[dict] = batch_job_output_file.get("response", None) or {}
_response_body: Final = _response.get("body", None) or {}
return _response_body
@ -472,5 +476,5 @@ def _batch_response_was_successful(batch_job_output_file: dict, custom_llm_provi
return _get_anthropic_result_from_batch_results_line(batch_job_output_file).get("type") == "succeeded"
if custom_llm_provider == "bedrock":
return batch_job_output_file.get("modelOutput") is not None and batch_job_output_file.get("error") is None
_response: dict = batch_job_output_file.get("response", None) or {}
_response: Final[dict] = batch_job_output_file.get("response", None) or {}
return _response.get("status_code", None) == 200

202
litellm/batches/main.py
View file

@ -15,7 +15,7 @@ import contextvars
import os
from collections.abc import Coroutine
from functools import partial
from typing import Any, Literal, cast
from typing import Any, Final, Literal, cast
import httpx
from openai.types.batch import BatchRequestCounts
@ -54,10 +54,10 @@ from litellm.utils import (
)
####### ENVIRONMENT VARIABLES ###################
openai_batches_instance = OpenAIBatchesAPI()
azure_batches_instance = AzureBatchesAPI()
vertex_ai_batches_instance = VertexAIBatchPrediction(gcs_bucket_name="")
anthropic_batches_instance = AnthropicBatchesHandler()
openai_batches_instance: Final = OpenAIBatchesAPI()
azure_batches_instance: Final = AzureBatchesAPI()
vertex_ai_batches_instance: Final = VertexAIBatchPrediction(gcs_bucket_name="")
anthropic_batches_instance: Final = AnthropicBatchesHandler()
base_llm_http_handler = BaseLLMHTTPHandler()
#################################################
@ -80,13 +80,13 @@ def _resolve_timeout(
Returns:
Resolved timeout as float
"""
timeout = optional_params.timeout or kwargs.get("request_timeout", default_timeout) or default_timeout
timeout: Final = optional_params.timeout or kwargs.get("request_timeout", default_timeout) or default_timeout
# Handle httpx.Timeout objects
if isinstance(timeout, httpx.Timeout):
if supports_httpx_timeout(custom_llm_provider) is False:
# Extract read timeout for providers that don't support httpx.Timeout
read_timeout = timeout.read or default_timeout
read_timeout: Final = timeout.read or default_timeout
return float(read_timeout)
else:
# For providers that support httpx.Timeout, we still need to return a float
@ -104,7 +104,7 @@ def _resolve_timeout(
@client
async def acreate_batch(
completion_window: Literal["24h"],
endpoint: Literal["/v1/chat/completions", "/v1/embeddings", "/v1/completions"],
endpoint: Literal["/v1/chat/completions", "/v1/embeddings", "/v1/completions", "/v1/responses"],
input_file_id: str,
custom_llm_provider: Literal["openai", "azure", "vertex_ai", "bedrock", "hosted_vllm"] = "openai",
metadata: dict[str, str] | None = None,
@ -119,11 +119,11 @@ async def acreate_batch(
LiteLLM Equivalent of POST: https://api.openai.com/v1/batches
"""
try:
loop = asyncio.get_event_loop()
loop: Final = asyncio.get_event_loop()
kwargs["acreate_batch"] = True
# Use a partial function to pass your keyword arguments
func = partial(
func: Final = partial(
create_batch,
completion_window,
endpoint,
@ -137,9 +137,9 @@ async def acreate_batch(
)
# Add the context to the function
ctx = contextvars.copy_context()
func_with_context = partial(ctx.run, func)
init_response = await loop.run_in_executor(None, func_with_context)
ctx: Final = contextvars.copy_context()
func_with_context: Final = partial(ctx.run, func)
init_response: Final = await loop.run_in_executor(None, func_with_context)
if asyncio.iscoroutine(init_response):
response = await init_response
@ -154,7 +154,7 @@ async def acreate_batch(
@client
def create_batch(
completion_window: Literal["24h"],
endpoint: Literal["/v1/chat/completions", "/v1/embeddings", "/v1/completions"],
endpoint: Literal["/v1/chat/completions", "/v1/embeddings", "/v1/completions", "/v1/responses"],
input_file_id: str,
custom_llm_provider: Literal["openai", "azure", "vertex_ai", "bedrock", "hosted_vllm"] = "openai",
metadata: dict[str, str] | None = None,
@ -169,10 +169,10 @@ def create_batch(
LiteLLM Equivalent of POST: https://api.openai.com/v1/batches
"""
try:
optional_params = GenericLiteLLMParams(**kwargs)
litellm_call_id = kwargs.get("litellm_call_id", None)
proxy_server_request = kwargs.get("proxy_server_request", None)
model_info = kwargs.get("model_info", None)
optional_params: Final = GenericLiteLLMParams(**kwargs)
litellm_call_id: Final = kwargs.get("litellm_call_id", None)
proxy_server_request: Final = kwargs.get("proxy_server_request", None)
model_info: Final = kwargs.get("model_info", None)
model: str | None = kwargs.get("model", None)
try:
if model is not None:
@ -182,14 +182,14 @@ def create_batch(
)
except Exception as e:
verbose_logger.exception(
f"litellm.batches.main.py::create_batch() - Error inferring custom_llm_provider - {e}"
"litellm.batches.main.py::create_batch() - Error inferring custom_llm_provider - %s", e
)
_is_async = kwargs.pop("acreate_batch", False) is True
litellm_params = dict(GenericLiteLLMParams(**kwargs))
litellm_logging_obj: LiteLLMLoggingObj = cast(LiteLLMLoggingObj, kwargs.get("litellm_logging_obj", None))
_is_async: Final = kwargs.pop("acreate_batch", False) is True
litellm_params: Final = dict(GenericLiteLLMParams(**kwargs))
litellm_logging_obj: Final[LiteLLMLoggingObj] = cast(LiteLLMLoggingObj, kwargs.get("litellm_logging_obj", None))
### TIMEOUT LOGIC ###
timeout = _resolve_timeout(optional_params, kwargs, custom_llm_provider)
timeout: Final = _resolve_timeout(optional_params, kwargs, custom_llm_provider)
litellm_logging_obj.update_from_kwargs(
kwargs=kwargs,
model=model,
@ -206,7 +206,7 @@ def create_batch(
custom_llm_provider=custom_llm_provider,
)
_create_batch_request = CreateBatchRequest(
_create_batch_request: Final = CreateBatchRequest(
completion_window=completion_window,
endpoint=endpoint,
input_file_id=input_file_id,
@ -248,7 +248,7 @@ def create_batch(
or os.getenv("OPENAI_API_BASE")
or "https://api.openai.com/v1"
)
organization = (
organization: Final = (
optional_params.organization
or litellm.organization
or os.getenv("OPENAI_ORGANIZATION", None)
@ -287,7 +287,7 @@ def create_batch(
if extra_body is not None:
extra_body.pop("azure_ad_token", None)
else:
get_secret_str("AZURE_AD_TOKEN") # type: ignore
get_secret_str("AZURE_AD_TOKEN")
response = azure_batches_instance.create_batch(
_is_async=_is_async,
@ -301,13 +301,13 @@ def create_batch(
)
elif custom_llm_provider == "vertex_ai":
api_base = optional_params.api_base or ""
vertex_ai_project = (
vertex_ai_project: Final = (
optional_params.vertex_project or litellm.vertex_project or get_secret_str("VERTEXAI_PROJECT")
)
vertex_ai_location = (
vertex_ai_location: Final = (
optional_params.vertex_location or litellm.vertex_location or get_secret_str("VERTEXAI_LOCATION")
)
vertex_credentials = optional_params.vertex_credentials or get_secret_str("VERTEXAI_CREDENTIALS")
vertex_credentials: Final = optional_params.vertex_credentials or get_secret_str("VERTEXAI_CREDENTIALS")
response = vertex_ai_batches_instance.create_batch(
_is_async=_is_async,
@ -327,7 +327,7 @@ def create_batch(
response=httpx.Response(
status_code=400,
content="Unsupported provider",
request=httpx.Request(method="create_batch", url="https://github.com/BerriAI/litellm"), # type: ignore
request=httpx.Request(method="create_batch", url="https://github.com/BerriAI/litellm"),
),
)
return response
@ -350,11 +350,11 @@ async def aretrieve_batch(
LiteLLM Equivalent of GET https://api.openai.com/v1/batches/{batch_id}
"""
try:
loop = asyncio.get_event_loop()
loop: Final = asyncio.get_event_loop()
kwargs["aretrieve_batch"] = True
# Use a partial function to pass your keyword arguments
func = partial(
func: Final = partial(
retrieve_batch,
batch_id,
custom_llm_provider,
@ -364,13 +364,13 @@ async def aretrieve_batch(
**kwargs,
)
# Add the context to the function
ctx = contextvars.copy_context()
func_with_context = partial(ctx.run, func)
init_response = await loop.run_in_executor(None, func_with_context)
ctx: Final = contextvars.copy_context()
func_with_context: Final = partial(ctx.run, func)
init_response: Final = await loop.run_in_executor(None, func_with_context)
if asyncio.iscoroutine(init_response):
response = await init_response
else:
response = init_response # type: ignore
response = init_response
return response
except Exception as e:
@ -397,7 +397,7 @@ def _handle_retrieve_batch_providers_without_provider_config(
or os.getenv("OPENAI_API_BASE")
or "https://api.openai.com/v1"
)
organization = (
organization: Final = (
optional_params.organization
or litellm.organization
or os.getenv("OPENAI_ORGANIZATION", None)
@ -422,7 +422,7 @@ def _handle_retrieve_batch_providers_without_provider_config(
)
elif custom_llm_provider == "azure":
api_base = optional_params.api_base or litellm.api_base or get_secret_str("AZURE_API_BASE")
api_version = optional_params.api_version or litellm.api_version or get_secret_str("AZURE_API_VERSION")
api_version: Final = optional_params.api_version or litellm.api_version or get_secret_str("AZURE_API_VERSION")
api_key = (
optional_params.api_key
@ -432,11 +432,11 @@ def _handle_retrieve_batch_providers_without_provider_config(
or get_secret_str("AZURE_API_KEY")
)
extra_body = optional_params.get("extra_body", {})
extra_body: Final = optional_params.get("extra_body", {})
if extra_body is not None:
extra_body.pop("azure_ad_token", None)
else:
get_secret_str("AZURE_AD_TOKEN") # type: ignore
get_secret_str("AZURE_AD_TOKEN")
response = azure_batches_instance.retrieve_batch(
_is_async=_is_async,
@ -450,13 +450,13 @@ def _handle_retrieve_batch_providers_without_provider_config(
)
elif custom_llm_provider == "vertex_ai":
api_base = optional_params.api_base or ""
vertex_ai_project = (
vertex_ai_project: Final = (
optional_params.vertex_project or litellm.vertex_project or get_secret_str("VERTEXAI_PROJECT")
)
vertex_ai_location = (
vertex_ai_location: Final = (
optional_params.vertex_location or litellm.vertex_location or get_secret_str("VERTEXAI_LOCATION")
)
vertex_credentials = optional_params.vertex_credentials or get_secret_str("VERTEXAI_CREDENTIALS")
vertex_credentials: Final = optional_params.vertex_credentials or get_secret_str("VERTEXAI_CREDENTIALS")
response = vertex_ai_batches_instance.retrieve_batch(
_is_async=_is_async,
@ -498,7 +498,7 @@ def _handle_retrieve_batch_providers_without_provider_config(
response=httpx.Response(
status_code=400,
content="Unsupported provider",
request=httpx.Request(method="retrieve_batch", url="https://github.com/BerriAI/litellm"), # type: ignore
request=httpx.Request(method="retrieve_batch", url="https://github.com/BerriAI/litellm"),
),
)
return response
@ -519,11 +519,11 @@ def retrieve_batch(
LiteLLM Equivalent of GET https://api.openai.com/v1/batches/{batch_id}
"""
try:
optional_params = GenericLiteLLMParams(**kwargs)
litellm_logging_obj: LiteLLMLoggingObj | None = kwargs.get("litellm_logging_obj", None)
optional_params: Final = GenericLiteLLMParams(**kwargs)
litellm_logging_obj: Final[LiteLLMLoggingObj | None] = kwargs.get("litellm_logging_obj", None)
### TIMEOUT LOGIC ###
timeout = optional_params.timeout or kwargs.get("request_timeout", 600) or 600
litellm_params = get_litellm_params(
litellm_params: Final = get_litellm_params(
custom_llm_provider=custom_llm_provider,
**kwargs,
)
@ -542,21 +542,21 @@ def retrieve_batch(
and isinstance(timeout, httpx.Timeout)
and supports_httpx_timeout(custom_llm_provider) is False
):
read_timeout = timeout.read or 600
read_timeout: Final = timeout.read or 600
timeout = read_timeout # default 10 min timeout
elif timeout is not None and not isinstance(timeout, httpx.Timeout):
timeout = float(timeout) # type: ignore
timeout = float(timeout)
elif timeout is None:
timeout = 600.0
_retrieve_batch_request = RetrieveBatchRequest(
_retrieve_batch_request: Final = RetrieveBatchRequest(
batch_id=batch_id,
extra_headers=extra_headers,
extra_body=extra_body,
)
_is_async = kwargs.pop("aretrieve_batch", False) is True
client = kwargs.get("client", None)
_is_async: Final = kwargs.pop("aretrieve_batch", False) is True
client: Final = kwargs.get("client", None)
# Bedrock has two distinct ARN families that need different APIs:
# * async-invoke ARNs (Twelve Labs Marengo embeddings) -> bedrock-runtime data plane
@ -568,7 +568,7 @@ def retrieve_batch(
if batch_id.startswith("arn:aws") and ":bedrock:" in batch_id:
if ":async-invoke/" in batch_id:
# Remove aws_region_name from kwargs to avoid duplicate parameter
async_kwargs = kwargs.copy()
async_kwargs: Final = kwargs.copy()
async_kwargs.pop("aws_region_name", None)
return BedrockBatchesHandler._handle_async_invoke_status(
@ -578,7 +578,7 @@ def retrieve_batch(
**async_kwargs,
)
if ":model-invocation-job/" in batch_id:
mij_kwargs = kwargs.copy()
mij_kwargs: Final = kwargs.copy()
mij_kwargs.pop("aws_region_name", None)
return BedrockBatchesHandler._handle_model_invocation_job_status(
@ -589,7 +589,7 @@ def retrieve_batch(
)
# Try to use provider config first (for providers like bedrock)
model: str | None = kwargs.get("model", None)
model: Final[str | None] = kwargs.get("model", None)
if model is not None:
provider_config = ProviderConfigManager.get_provider_batches_config(
model=model,
@ -599,7 +599,7 @@ def retrieve_batch(
provider_config = None
if provider_config is not None:
response = base_llm_http_handler.retrieve_batch(
response: Final = base_llm_http_handler.retrieve_batch(
batch_id=batch_id,
provider_config=provider_config,
litellm_params=litellm_params,
@ -656,11 +656,11 @@ async def alist_batches(
"""
try:
loop = asyncio.get_event_loop()
loop: Final = asyncio.get_event_loop()
kwargs["alist_batches"] = True
# Use a partial function to pass your keyword arguments
func = partial(
func: Final = partial(
list_batches,
after,
limit,
@ -671,13 +671,13 @@ async def alist_batches(
)
# Add the context to the function
ctx = contextvars.copy_context()
func_with_context = partial(ctx.run, func)
init_response = await loop.run_in_executor(None, func_with_context)
ctx: Final = contextvars.copy_context()
func_with_context: Final = partial(ctx.run, func)
init_response: Final = await loop.run_in_executor(None, func_with_context)
if asyncio.iscoroutine(init_response):
response = await init_response
else:
response = init_response # type: ignore
response = init_response
return response
except Exception as e:
@ -700,8 +700,8 @@ def list_batches(
"""
try:
# set API KEY
optional_params = GenericLiteLLMParams(**kwargs)
litellm_params = get_litellm_params(
optional_params: Final = GenericLiteLLMParams(**kwargs)
litellm_params: Final = get_litellm_params(
custom_llm_provider=custom_llm_provider,
**kwargs,
)
@ -720,14 +720,14 @@ def list_batches(
and isinstance(timeout, httpx.Timeout)
and supports_httpx_timeout(custom_llm_provider) is False
):
read_timeout = timeout.read or 600
read_timeout: Final = timeout.read or 600
timeout = read_timeout # default 10 min timeout
elif timeout is not None and not isinstance(timeout, httpx.Timeout):
timeout = float(timeout) # type: ignore
timeout = float(timeout)
elif timeout is None:
timeout = 600.0
_is_async = kwargs.pop("alist_batches", False) is True
_is_async: Final = kwargs.pop("alist_batches", False) is True
if custom_llm_provider in OPENAI_COMPATIBLE_BATCH_AND_FILES_PROVIDERS:
# for deepinfra/perplexity/anyscale/groq we check in get_llm_provider and pass in the api base from there
api_base = (
@ -737,7 +737,7 @@ def list_batches(
or os.getenv("OPENAI_API_BASE")
or "https://api.openai.com/v1"
)
organization = (
organization: Final = (
optional_params.organization
or litellm.organization
or os.getenv("OPENAI_ORGANIZATION", None)
@ -755,7 +755,7 @@ def list_batches(
max_retries=optional_params.max_retries,
)
elif custom_llm_provider == "azure":
api_base = optional_params.api_base or litellm.api_base or get_secret_str("AZURE_API_BASE") # type: ignore
api_base = optional_params.api_base or litellm.api_base or get_secret_str("AZURE_API_BASE")
api_version = optional_params.api_version or litellm.api_version or get_secret_str("AZURE_API_VERSION")
api_key = (
@ -770,7 +770,7 @@ def list_batches(
if extra_body is not None:
extra_body.pop("azure_ad_token", None)
else:
get_secret_str("AZURE_AD_TOKEN") # type: ignore
get_secret_str("AZURE_AD_TOKEN")
response = azure_batches_instance.list_batches(
_is_async=_is_async,
@ -783,13 +783,13 @@ def list_batches(
)
elif custom_llm_provider == "vertex_ai":
api_base = optional_params.api_base or ""
vertex_ai_project = (
vertex_ai_project: Final = (
optional_params.vertex_project or litellm.vertex_project or get_secret_str("VERTEXAI_PROJECT")
)
vertex_ai_location = (
vertex_ai_location: Final = (
optional_params.vertex_location or litellm.vertex_location or get_secret_str("VERTEXAI_LOCATION")
)
vertex_credentials = optional_params.vertex_credentials or get_secret_str("VERTEXAI_CREDENTIALS")
vertex_credentials: Final = optional_params.vertex_credentials or get_secret_str("VERTEXAI_CREDENTIALS")
response = vertex_ai_batches_instance.list_batches(
_is_async=_is_async,
@ -813,7 +813,7 @@ def list_batches(
response=httpx.Response(
status_code=400,
content="Unsupported provider",
request=httpx.Request(method="create_thread", url="https://github.com/BerriAI/litellm"), # type: ignore
request=httpx.Request(method="create_thread", url="https://github.com/BerriAI/litellm"),
),
)
return response
@ -836,14 +836,14 @@ async def acancel_batch(
LiteLLM Equivalent of POST https://api.openai.com/v1/batches/{batch_id}/cancel
"""
try:
loop = asyncio.get_event_loop()
loop: Final = asyncio.get_event_loop()
kwargs["acancel_batch"] = True
# Preserve model parameter - only pop from kwargs if it exists there
# (to avoid passing it twice), otherwise keep the function parameter value
model = kwargs.pop("model", None) or model
# Use a partial function to pass your keyword arguments
func = partial(
func: Final = partial(
cancel_batch,
batch_id,
model,
@ -854,9 +854,9 @@ async def acancel_batch(
**kwargs,
)
# Add the context to the function
ctx = contextvars.copy_context()
func_with_context = partial(ctx.run, func)
init_response = await loop.run_in_executor(None, func_with_context)
ctx: Final = contextvars.copy_context()
func_with_context: Final = partial(ctx.run, func)
init_response: Final = await loop.run_in_executor(None, func_with_context)
if asyncio.iscoroutine(init_response):
response = await init_response
else:
@ -890,10 +890,10 @@ def cancel_batch(
)
except Exception as e:
verbose_logger.exception(
f"litellm.batches.main.py::cancel_batch() - Error inferring custom_llm_provider - {e}"
"litellm.batches.main.py::cancel_batch() - Error inferring custom_llm_provider - %s", e
)
optional_params = GenericLiteLLMParams(**kwargs)
litellm_params = get_litellm_params(
optional_params: Final = GenericLiteLLMParams(**kwargs)
litellm_params: Final = get_litellm_params(
custom_llm_provider=custom_llm_provider,
**kwargs,
)
@ -906,20 +906,20 @@ def cancel_batch(
and isinstance(timeout, httpx.Timeout)
and supports_httpx_timeout(custom_llm_provider) is False
):
read_timeout = timeout.read or 600
read_timeout: Final = timeout.read or 600
timeout = read_timeout # default 10 min timeout
elif timeout is not None and not isinstance(timeout, httpx.Timeout):
timeout = float(timeout) # type: ignore
timeout = float(timeout)
elif timeout is None:
timeout = 600.0
_cancel_batch_request = CancelBatchRequest(
_cancel_batch_request: Final = CancelBatchRequest(
batch_id=batch_id,
extra_headers=extra_headers,
extra_body=extra_body,
)
_is_async = kwargs.pop("acancel_batch", False) is True
_is_async: Final = kwargs.pop("acancel_batch", False) is True
api_base: str | None = None
if custom_llm_provider in OPENAI_COMPATIBLE_BATCH_AND_FILES_PROVIDERS:
api_base = (
@ -929,7 +929,7 @@ def cancel_batch(
or os.getenv("OPENAI_API_BASE")
or "https://api.openai.com/v1"
)
organization = (
organization: Final = (
optional_params.organization or litellm.organization or os.getenv("OPENAI_ORGANIZATION", None) or None
)
api_key = optional_params.api_key or litellm.api_key or litellm.openai_key or os.getenv("OPENAI_API_KEY")
@ -959,7 +959,7 @@ def cancel_batch(
if extra_body is not None:
extra_body.pop("azure_ad_token", None)
else:
get_secret_str("AZURE_AD_TOKEN") # type: ignore
get_secret_str("AZURE_AD_TOKEN")
response = azure_batches_instance.cancel_batch(
_is_async=_is_async,
@ -973,13 +973,13 @@ def cancel_batch(
)
elif custom_llm_provider == "vertex_ai":
api_base = optional_params.api_base or None
vertex_ai_project = (
vertex_ai_project: Final = (
optional_params.vertex_project or litellm.vertex_project or get_secret_str("VERTEXAI_PROJECT")
)
vertex_ai_location = (
vertex_ai_location: Final = (
optional_params.vertex_location or litellm.vertex_location or get_secret_str("VERTEXAI_LOCATION")
)
vertex_credentials = optional_params.vertex_credentials or get_secret_str("VERTEXAI_CREDENTIALS")
vertex_credentials: Final = optional_params.vertex_credentials or get_secret_str("VERTEXAI_CREDENTIALS")
response = vertex_ai_batches_instance.cancel_batch(
_is_async=_is_async,
@ -999,7 +999,7 @@ def cancel_batch(
response=httpx.Response(
status_code=400,
content="Unsupported provider",
request=httpx.Request(method="cancel_batch", url="https://github.com/BerriAI/litellm"), # type: ignore
request=httpx.Request(method="cancel_batch", url="https://github.com/BerriAI/litellm"),
),
)
return response
@ -1025,10 +1025,10 @@ def _handle_async_invoke_status(batch_id: str, aws_region_name: str, logging_obj
async def _async_get_status():
# Create embedding handler instance
embedding_handler = BedrockEmbedding()
embedding_handler: Final = BedrockEmbedding()
# Get the status of the async invoke job
status_response = await embedding_handler._get_async_invoke_status(
status_response: Final = await embedding_handler._get_async_invoke_status(
invocation_arn=batch_id,
aws_region_name=aws_region_name,
logging_obj=logging_obj,
@ -1040,16 +1040,16 @@ def _handle_async_invoke_status(batch_id: str, aws_region_name: str, logging_obj
from litellm.types.utils import LiteLLMBatch
# Normalize status to lowercase (AWS returns 'Completed', 'Failed', etc.)
aws_status_raw = status_response.get("status", "")
aws_status_lower = aws_status_raw.lower()
aws_status_raw: Final = status_response.get("status", "")
aws_status_lower: Final = aws_status_raw.lower()
# Map AWS status values to LiteLLM expected values
status_mapping: dict[str, BatchJobStatus] = {
status_mapping: Final[dict[str, BatchJobStatus]] = {
"completed": "completed",
"failed": "failed",
"inprogress": "in_progress",
"in_progress": "in_progress",
}
normalized_status: BatchJobStatus = status_mapping.get(
normalized_status: Final[BatchJobStatus] = status_mapping.get(
aws_status_lower, "failed"
) # Default to "failed" if unknown status
@ -1073,7 +1073,7 @@ def _handle_async_invoke_status(batch_id: str, aws_region_name: str, logging_obj
_,
_,
) = BedrockBatchesConfig()._parse_timestamps_and_status(status_response, aws_status_raw)
result = LiteLLMBatch(
result: Final = LiteLLMBatch(
id=status_response["invocationArn"],
object="batch",
status=normalized_status,
@ -1105,7 +1105,7 @@ def _handle_async_invoke_status(batch_id: str, aws_region_name: str, logging_obj
import concurrent.futures
def run_in_thread():
new_loop = asyncio.new_event_loop()
new_loop: Final = asyncio.new_event_loop()
asyncio.set_event_loop(new_loop)
try:
return new_loop.run_until_complete(_async_get_status())
@ -1113,5 +1113,5 @@ def _handle_async_invoke_status(batch_id: str, aws_region_name: str, logging_obj
new_loop.close()
with concurrent.futures.ThreadPoolExecutor() as executor:
future = executor.submit(run_in_thread)
future: Final = executor.submit(run_in_thread)
return future.result()

26
litellm/budget_manager.py
View file

@ -11,7 +11,7 @@ import json
import os
import threading
import time
from typing import Literal
from typing import Final, Literal
import litellm
from litellm.constants import (
@ -60,8 +60,8 @@ class BudgetManager:
self.print_verbose(f"user dict from local: {self.user_dict}")
elif self.client_type == "hosted":
# Load the user_dict from hosted db
url = self.api_base + "/get_budget"
data = {"project_name": self.project_name}
url: Final = self.api_base + "/get_budget"
data: Final = {"project_name": self.project_name}
response = litellm.module_level_client.post(url, headers=self.headers, json=data)
response = response.json()
if response["status"] == "error":
@ -100,11 +100,11 @@ class BudgetManager:
return self.user_dict[user]
def projected_cost(self, model: str, messages: list, user: str):
text = "".join(message["content"] for message in messages)
prompt_tokens = litellm.token_counter(model=model, text=text)
text: Final = "".join(message["content"] for message in messages)
prompt_tokens: Final = litellm.token_counter(model=model, text=text)
prompt_cost, _ = litellm.cost_per_token(model=model, prompt_tokens=prompt_tokens, completion_tokens=0)
current_cost = self.user_dict[user].get("current_cost", 0)
projected_cost = prompt_cost + current_cost
current_cost: Final = self.user_dict[user].get("current_cost", 0)
projected_cost: Final = prompt_cost + current_cost
return projected_cost
def get_total_budget(self, user: str):
@ -178,11 +178,11 @@ class BudgetManager:
def reset_on_duration(self, user: str):
# Get current and creation time
last_updated_at = self.user_dict[user]["last_updated_at"]
current_time = time.time()
last_updated_at: Final = self.user_dict[user]["last_updated_at"]
current_time: Final = time.time()
# Convert duration from days to seconds
duration_in_seconds = self.user_dict[user]["duration"] * HOURS_IN_A_DAY * 60 * 60
duration_in_seconds: Final = self.user_dict[user]["duration"] * HOURS_IN_A_DAY * 60 * 60
# Check if duration has elapsed
if current_time - last_updated_at >= duration_in_seconds:
@ -197,7 +197,7 @@ class BudgetManager:
self.reset_on_duration(user)
def _save_data_thread(self):
thread = threading.Thread(target=self.save_data) # [Non-Blocking]: saves data without blocking execution
thread: Final = threading.Thread(target=self.save_data) # [Non-Blocking]: saves data without blocking execution
thread.start()
def save_data(self):
@ -209,8 +209,8 @@ class BudgetManager:
json.dump(self.user_dict, json_file, indent=4) # Indent for pretty formatting
return {"status": "success"}
elif self.client_type == "hosted":
url = self.api_base + "/set_budget"
data = {"project_name": self.project_name, "user_dict": self.user_dict}
url: Final = self.api_base + "/set_budget"
data: Final = {"project_name": self.project_name, "user_dict": self.user_dict}
response = litellm.module_level_client.post(url, headers=self.headers, json=data)
response = response.json()
return response

6
litellm/caching/_embedding_router.py
View file

@ -12,7 +12,7 @@ This module is dependency-injected: callers pass the proxy ``llm_router`` and
from __future__ import annotations
from typing import TYPE_CHECKING, Any
from typing import TYPE_CHECKING, Any, Final
if TYPE_CHECKING:
from litellm.router import Router
@ -26,7 +26,7 @@ def resolve_embedding_router(
"""Return ``llm_router`` iff it serves ``embedding_model`` as a deployment."""
if llm_router is None:
return None
router_model_names: list[str] = (
router_model_names: Final[list[str]] = (
[m["model_name"] for m in llm_model_list if "model_name" in m] if llm_model_list is not None else []
)
if embedding_model in router_model_names:
@ -38,6 +38,6 @@ def build_router_embedding_metadata(
request_metadata: dict[str, Any] | None,
) -> dict[str, Any]:
"""Forward the caller's full metadata, flagged as a semantic-cache embedding."""
metadata: dict[str, Any] = dict(request_metadata or {})
metadata: Final[dict[str, Any]] = dict(request_metadata or {})
metadata["semantic-cache-embedding"] = True
return metadata

4
litellm/caching/_internal_lru_cache.py
View file

@ -1,6 +1,6 @@
from collections.abc import Callable
from functools import lru_cache
from typing import TypeVar
from typing import Final, TypeVar
T = TypeVar("T")
@ -21,7 +21,7 @@ def lru_cache_wrapper(
return ("error", e)
def wrapped(*args, **kwargs):
result = wrapper(*args, **kwargs)
result: Final = wrapper(*args, **kwargs)
if result[0] == "error":
raise result[1]
return result[1]

31
litellm/caching/azure_blob_cache.py
View file

@ -11,6 +11,7 @@ Has 4 methods:
import asyncio
import json
from contextlib import suppress
from typing import Final
from litellm._logging import print_verbose, verbose_logger
@ -41,7 +42,7 @@ class AzureBlobCache(BaseCache):
def set_cache(self, key, value, **kwargs) -> None:
print_verbose(f"LiteLLM SET Cache - Azure Blob. Key={key}. Value={value}")
serialized_value = json.dumps(value)
serialized_value: Final = json.dumps(value)
try:
self.container_client.upload_blob(key, serialized_value)
except Exception as e:
@ -50,7 +51,7 @@ class AzureBlobCache(BaseCache):
async def async_set_cache(self, key, value, **kwargs) -> None:
print_verbose(f"LiteLLM SET Cache - Azure Blob. Key={key}. Value={value}")
serialized_value = json.dumps(value)
serialized_value: Final = json.dumps(value)
try:
await self.async_container_client.upload_blob(key, serialized_value, overwrite=True)
except Exception as e:
@ -62,12 +63,15 @@ class AzureBlobCache(BaseCache):
try:
print_verbose(f"Get Azure Blob Cache: key: {key}")
as_bytes = self.container_client.download_blob(key).readall()
as_str = as_bytes.decode("utf-8")
cached_response = json.loads(as_str)
as_bytes: Final = self.container_client.download_blob(key).readall()
as_str: Final = as_bytes.decode("utf-8")
cached_response: Final = json.loads(as_str)
verbose_logger.debug(
f"Got Azure Blob Cache: key: {key}, cached_response {cached_response}. Type Response {type(cached_response)}"
"Got Azure Blob Cache: key: %s, cached_response %s. Type Response %s",
key,
cached_response,
type(cached_response),
)
return cached_response
@ -79,12 +83,15 @@ class AzureBlobCache(BaseCache):
try:
print_verbose(f"Get Azure Blob Cache: key: {key}")
blob = await self.async_container_client.download_blob(key)
as_bytes = await blob.readall()
as_str = as_bytes.decode("utf-8")
cached_response = json.loads(as_str)
blob: Final = await self.async_container_client.download_blob(key)
as_bytes: Final = await blob.readall()
as_str: Final = as_bytes.decode("utf-8")
cached_response: Final = json.loads(as_str)
verbose_logger.debug(
f"Got Azure Blob Cache: key: {key}, cached_response {cached_response}. Type Response {type(cached_response)}"
"Got Azure Blob Cache: key: %s, cached_response %s. Type Response %s",
key,
cached_response,
type(cached_response),
)
return cached_response
except ResourceNotFoundError:
@ -99,7 +106,7 @@ class AzureBlobCache(BaseCache):
await self.async_container_client.close()
async def async_set_cache_pipeline(self, cache_list, **kwargs) -> None:
tasks = []
tasks: Final = []
for val in cache_list:
tasks.append(self.async_set_cache(val[0], val[1], **kwargs))
await asyncio.gather(*tasks)

4
litellm/caching/base_cache.py
View file

@ -9,7 +9,7 @@ Has 4 methods:
"""
from abc import ABC, abstractmethod
from typing import TYPE_CHECKING, Any, Union
from typing import TYPE_CHECKING, Any, Final, Union
if TYPE_CHECKING:
from opentelemetry.trace import Span as _Span
@ -24,7 +24,7 @@ class BaseCache(ABC):
self.default_ttl = default_ttl
def get_ttl(self, **kwargs) -> int | None:
kwargs_ttl: int | None = kwargs.get("ttl")
kwargs_ttl: Final[int | None] = kwargs.get("ttl")
if kwargs_ttl is not None:
try:
return int(kwargs_ttl)

122
litellm/caching/caching.py
View file

@ -13,7 +13,7 @@ import json
import time
import traceback
from enum import Enum
from typing import Any
from typing import Any, Final
from pydantic import BaseModel
@ -169,13 +169,13 @@ class Cache:
if type == LiteLLMCacheType.REDIS:
# Check REDIS_CLUSTER_NODES env var if no explicit startup nodes
if not redis_startup_nodes:
_env_cluster_nodes = litellm.get_secret("REDIS_CLUSTER_NODES")
_env_cluster_nodes: Final = litellm.get_secret("REDIS_CLUSTER_NODES")
if _env_cluster_nodes is not None and isinstance(_env_cluster_nodes, str):
redis_startup_nodes = json.loads(_env_cluster_nodes)
if redis_startup_nodes:
# Only pass GCP parameters if they are provided
cluster_kwargs = {
cluster_kwargs: Final = {
"host": host,
"port": port,
"password": password,
@ -312,9 +312,9 @@ class Cache:
)
def _get_semantic_cache_tenant_scope(self, kwargs: dict) -> str:
metadata: dict = kwargs.get("metadata") or {}
litellm_params: dict = kwargs.get("litellm_params") or {}
metadata_in_litellm_params: dict = litellm_params.get("metadata") or {}
metadata: Final[dict] = kwargs.get("metadata") or {}
litellm_params: Final[dict] = kwargs.get("litellm_params") or {}
metadata_in_litellm_params: Final[dict] = litellm_params.get("metadata") or {}
scope = ""
for field in self._SEMANTIC_CACHE_TENANT_SCOPE_FIELDS:
@ -338,15 +338,15 @@ class Cache:
cache_key = ""
# verbose_logger.debug("\nGetting Cache key. Kwargs: %s", kwargs)
preset_cache_key = self._get_preset_cache_key_from_kwargs(**kwargs)
preset_cache_key: Final = self._get_preset_cache_key_from_kwargs(**kwargs)
if preset_cache_key is not None:
verbose_logger.debug("\nReturning preset cache key: %s", preset_cache_key)
return preset_cache_key
combined_kwargs = ModelParamHelper._get_all_llm_api_params()
litellm_param_kwargs = all_litellm_params
is_semantic_cache = self._is_semantic_cache()
scope_excluded_params = self._SEMANTIC_CACHE_SCOPE_EXCLUDED_PARAMS if is_semantic_cache else frozenset()
combined_kwargs: Final = ModelParamHelper._get_all_llm_api_params()
litellm_param_kwargs: Final = all_litellm_params
is_semantic_cache: Final = self._is_semantic_cache()
scope_excluded_params: Final = self._SEMANTIC_CACHE_SCOPE_EXCLUDED_PARAMS if is_semantic_cache else frozenset()
for param in kwargs:
if param in scope_excluded_params:
continue
@ -373,7 +373,7 @@ class Cache:
)
# Remove preset_cache_key from kwargs to avoid "got multiple values" TypeError
# when kwargs already contains preset_cache_key from upstream callers
kwargs_for_preset = {k: v for k, v in kwargs.items() if k != "preset_cache_key"}
kwargs_for_preset: Final = {k: v for k, v in kwargs.items() if k != "preset_cache_key"}
self._set_preset_cache_key_in_kwargs(preset_cache_key=hashed_cache_key, **kwargs_for_preset)
return hashed_cache_key
@ -399,15 +399,15 @@ class Cache:
2. Else if a model_group is set, then return the model_group as the model. This is used for all requests sent through the litellm.Router()
3. Else use the `model` passed in kwargs
"""
metadata: dict = kwargs.get("metadata", {}) or {}
litellm_params: dict = kwargs.get("litellm_params", {}) or {}
metadata_in_litellm_params: dict = litellm_params.get("metadata", {}) or {}
model_group: str | None = metadata.get("model_group") or metadata_in_litellm_params.get("model_group")
caching_group = self._get_caching_group(metadata, model_group)
metadata: Final[dict] = kwargs.get("metadata", {}) or {}
litellm_params: Final[dict] = kwargs.get("litellm_params", {}) or {}
metadata_in_litellm_params: Final[dict] = litellm_params.get("metadata", {}) or {}
model_group: Final[str | None] = metadata.get("model_group") or metadata_in_litellm_params.get("model_group")
caching_group: Final = self._get_caching_group(metadata, model_group)
return caching_group or model_group or kwargs["model"]
def _get_caching_group(self, metadata: dict, model_group: str | None) -> str | None:
caching_groups: list | None = metadata.get("caching_groups", [])
caching_groups: Final[list | None] = metadata.get("caching_groups", [])
if caching_groups:
for group in caching_groups:
if model_group in group:
@ -418,9 +418,9 @@ class Cache:
"""
Handles getting the value for the 'file' param from kwargs. Used for `transcription` requests
"""
file = kwargs.get("file")
metadata = kwargs.get("metadata", {})
litellm_params = kwargs.get("litellm_params", {})
file: Final = kwargs.get("file")
metadata: Final = kwargs.get("metadata", {})
litellm_params: Final = kwargs.get("litellm_params", {})
return (
metadata.get("file_checksum")
or getattr(file, "name", None)
@ -467,9 +467,9 @@ class Cache:
Returns:
str: The hashed cache key.
"""
hash_object = hashlib.sha256(cache_key.encode())
hash_object: Final = hashlib.sha256(cache_key.encode())
# Hexadecimal representation of the hash
hash_hex = hash_object.hexdigest()
hash_hex: Final = hash_object.hexdigest()
verbose_logger.debug("Hashed cache key (SHA-256): %s", hash_hex)
return hash_hex
@ -484,16 +484,16 @@ class Cache:
Returns:
str: The final hashed cache key with the redis namespace.
"""
dynamic_cache_control: DynamicCacheControl = kwargs.get("cache", {})
metadata = kwargs.get("metadata") or {}
namespace = dynamic_cache_control.get("namespace") or metadata.get("redis_namespace") or self.namespace
dynamic_cache_control: Final[DynamicCacheControl] = kwargs.get("cache", {})
metadata: Final = kwargs.get("metadata") or {}
namespace: Final = dynamic_cache_control.get("namespace") or metadata.get("redis_namespace") or self.namespace
if namespace:
hash_hex = f"{namespace}:{hash_hex}"
verbose_logger.debug("Final hashed key: %s", hash_hex)
return hash_hex
def generate_streaming_content(self, content):
chunk_size = 5 # Adjust the chunk size as needed
chunk_size: Final = 5 # Adjust the chunk size as needed
for i in range(0, len(content), chunk_size):
yield {
"choices": [
@ -517,11 +517,11 @@ class Cache:
"""
# Check if a timestamp was stored with the cached response
if cached_result is not None and isinstance(cached_result, dict) and "timestamp" in cached_result:
timestamp = cached_result["timestamp"]
current_time = time.time()
timestamp: Final = cached_result["timestamp"]
current_time: Final = time.time()
# Calculate age of the cached response
response_age = current_time - timestamp
response_age: Final = current_time - timestamp
# Check if the cached response is older than the max-age
if max_age is not None and response_age > max_age:
@ -534,22 +534,20 @@ class Cache:
if isinstance(cached_response, dict):
pass
else:
cached_response = json.loads(
cached_response # type: ignore
) # Convert string to dictionary
cached_response = json.loads(cached_response) # Convert string to dictionary
except Exception:
cached_response = ast.literal_eval(cached_response) # type: ignore
cached_response = ast.literal_eval(cached_response)
return cached_response
return cached_result
@staticmethod
def _get_safe_cache_lookup_kwargs(kwargs: dict[str, Any]) -> dict[str, Any]:
cache_lookup_kwargs: dict[str, Any] = {}
cache_lookup_kwargs: Final[dict[str, Any]] = {}
for prompt_kwarg in ("messages", "input"):
if prompt_kwarg in kwargs:
cache_lookup_kwargs[prompt_kwarg] = kwargs[prompt_kwarg]
metadata = kwargs.get("metadata")
metadata: Final = kwargs.get("metadata")
if isinstance(metadata, dict):
cache_lookup_kwargs["metadata"] = dict(metadata)
@ -559,8 +557,8 @@ class Cache:
def _update_metadata_from_cache_lookup_kwargs(
original_kwargs: dict[str, Any], cache_lookup_kwargs: dict[str, Any]
) -> None:
original_metadata = original_kwargs.get("metadata")
cache_lookup_metadata = cache_lookup_kwargs.get("metadata")
original_metadata: Final = original_kwargs.get("metadata")
cache_lookup_metadata: Final = cache_lookup_kwargs.get("metadata")
if not isinstance(original_metadata, dict) or not isinstance(cache_lookup_metadata, dict):
return
@ -586,9 +584,9 @@ class Cache:
else:
cache_key = self.get_cache_key(**kwargs)
if cache_key is not None:
cache_control_args: DynamicCacheControl = kwargs.get("cache", {})
cache_control_args: Final[DynamicCacheControl] = kwargs.get("cache", {})
max_age = cache_control_args.get("s-maxage") or cache_control_args.get("s-max-age") or float("inf")
cache_lookup_kwargs = self._get_safe_cache_lookup_kwargs(kwargs)
cache_lookup_kwargs: Final = self._get_safe_cache_lookup_kwargs(kwargs)
if dynamic_cache_object is not None:
cached_result = dynamic_cache_object.get_cache(cache_key, **cache_lookup_kwargs)
else:
@ -618,8 +616,8 @@ class Cache:
else:
cache_key = self.get_cache_key(**kwargs)
if cache_key is not None:
cache_control_args = kwargs.get("cache", {})
max_age = cache_control_args.get("s-max-age", cache_control_args.get("s-maxage", float("inf")))
cache_control_args: Final = kwargs.get("cache", {})
max_age: Final = cache_control_args.get("s-max-age", cache_control_args.get("s-maxage", float("inf")))
if dynamic_cache_object is not None:
cached_result = await dynamic_cache_object.async_get_cache(cache_key, **kwargs)
else:
@ -646,13 +644,13 @@ class Cache:
if self.ttl is not None:
kwargs["ttl"] = self.ttl
## Get Cache-Controls ##
_cache_kwargs = kwargs.get("cache", None)
_cache_kwargs: Final = kwargs.get("cache", None)
if isinstance(_cache_kwargs, dict):
for k, v in _cache_kwargs.items():
if k == "ttl":
kwargs["ttl"] = v
cached_data = {"timestamp": time.time(), "response": result}
cached_data: Final = {"timestamp": time.time(), "response": result}
return cache_key, cached_data, kwargs
else:
raise Exception("cache key is None")
@ -676,7 +674,7 @@ class Cache:
cache_key, cached_data, kwargs = self._add_cache_logic(result=result, **kwargs)
self.cache.set_cache(cache_key, cached_data, **kwargs)
except Exception as e:
verbose_logger.exception(f"LiteLLM Cache: Excepton add_cache: {e}")
verbose_logger.exception("LiteLLM Cache: Excepton add_cache: %s", e)
async def async_add_cache(self, result, dynamic_cache_object: BaseCache | None = None, **kwargs):
"""
@ -695,7 +693,7 @@ class Cache:
else:
await self.cache.async_set_cache(cache_key, cached_data, **kwargs)
except Exception as e:
verbose_logger.exception(f"LiteLLM Cache: Excepton add_cache: {e}")
verbose_logger.exception("LiteLLM Cache: Excepton add_cache: %s", e)
def _convert_to_cached_embedding(
self,
@ -756,7 +754,7 @@ class Cache:
if result.usage is None or result.usage.prompt_tokens_details is None:
return None
details = result.usage.prompt_tokens_details
details: Final = result.usage.prompt_tokens_details
if hasattr(details, "model_dump"):
details_dict = details.model_dump(exclude_none=True)
elif isinstance(details, dict):
@ -767,12 +765,12 @@ class Cache:
if not details_dict:
return None
num_items = len(result.data)
num_items: Final = len(result.data)
if num_items <= 1:
return details_dict
# Distribute integer/float fields evenly across items
per_item: dict = {}
per_item: Final[dict] = {}
for key, value in details_dict.items():
if isinstance(value, int):
quotient, remainder = divmod(value, num_items)
@ -798,8 +796,8 @@ class Cache:
if result.usage is None or result.usage.prompt_tokens is None:
return None
total = result.usage.prompt_tokens
num_items = len(result.data)
total: Final = result.usage.prompt_tokens
num_items: Final = len(result.data)
if num_items <= 1:
return total
@ -813,23 +811,23 @@ class Cache:
kwargs: dict,
idx_in_result_data: int = 0,
) -> tuple[str, dict, dict]:
preset_cache_key = self.get_cache_key(**{**kwargs, "input": input})
preset_cache_key: Final = self.get_cache_key(**{**kwargs, "input": input})
kwargs["cache_key"] = preset_cache_key
embedding_response = result.data[idx_in_result_data]
embedding_response: Final = result.data[idx_in_result_data]
# Extract per-item prompt_tokens + details from response usage
prompt_tokens = self._get_per_item_prompt_tokens(
prompt_tokens: Final = self._get_per_item_prompt_tokens(
result=result,
idx_in_result_data=idx_in_result_data,
)
prompt_tokens_details = self._get_per_item_prompt_tokens_details(
prompt_tokens_details: Final = self._get_per_item_prompt_tokens_details(
result=result,
idx_in_result_data=idx_in_result_data,
)
# Always convert to properly typed CachedEmbedding
model_name = result.model
embedding_dict: CachedEmbedding = self._convert_to_cached_embedding(
model_name: Final = result.model
embedding_dict: Final[CachedEmbedding] = self._convert_to_cached_embedding(
embedding_response,
model_name,
prompt_tokens=prompt_tokens,
@ -856,7 +854,7 @@ class Cache:
if self.ttl is not None:
kwargs["ttl"] = self.ttl
cache_list = []
cache_list: Final = []
if isinstance(kwargs["input"], list):
for idx, i in enumerate(kwargs["input"]):
(
@ -874,7 +872,7 @@ class Cache:
else:
await self.cache.async_set_cache_pipeline(cache_list=cache_list, **kwargs)
except Exception as e:
verbose_logger.exception(f"LiteLLM Cache: Excepton add_cache: {e}")
verbose_logger.exception("LiteLLM Cache: Excepton add_cache: %s", e)
def should_use_cache(self, **kwargs):
"""
@ -887,7 +885,7 @@ class Cache:
return True
# when mode == default_off -> Cache is opt in only
_cache = kwargs.get("cache", None)
_cache: Final = kwargs.get("cache", None)
verbose_logger.debug("should_use_cache: kwargs: %s; _cache: %s", kwargs, _cache)
if _cache and isinstance(_cache, dict):
if _cache.get("use-cache", False) is True:
@ -899,13 +897,13 @@ class Cache:
await self.cache.batch_cache_write(cache_key, cached_data, **kwargs)
async def ping(self):
cache_ping = getattr(self.cache, "ping")
cache_ping: Final = getattr(self.cache, "ping")
if cache_ping:
return await cache_ping()
return None
async def delete_cache_keys(self, keys):
cache_delete_cache_keys = getattr(self.cache, "delete_cache_keys")
cache_delete_cache_keys: Final = getattr(self.cache, "delete_cache_keys")
if cache_delete_cache_keys:
return await cache_delete_cache_keys(keys)
return None

92
litellm/caching/caching_handler.py
View file

@ -19,11 +19,7 @@ import datetime
import inspect
import time
from collections.abc import AsyncGenerator, Callable, Generator
from typing import (
TYPE_CHECKING,
Any,
Optional,
)
from typing import TYPE_CHECKING, Any, Final, Optional
from pydantic import BaseModel
@ -76,7 +72,7 @@ class CachingHandlerResponse(BaseModel):
embedding_all_elements_cache_hit: bool = False # this is set to True when all elements in the list have a cache hit in the embedding cache, if true return the final_embedding_cached_response no need to make an API call
in_memory_cache_obj = InMemoryCache()
in_memory_cache_obj: Final = InMemoryCache()
def _drop_logging_obj_from_kwargs(request_kwargs: dict[str, object]) -> dict[str, object]:
@ -96,10 +92,10 @@ def _drop_logging_obj_from_kwargs(request_kwargs: dict[str, object]) -> dict[str
def _is_chat_completion_cached_dict(cached_result: dict) -> bool:
cached_id = cached_result.get("id")
cached_id: Final = cached_result.get("id")
if isinstance(cached_id, str) and cached_id.startswith("chatcmpl"):
return True
obj = cached_result.get("object")
obj: Final = cached_result.get("object")
if isinstance(obj, str):
return obj.startswith("chat.completion")
return "choices" in cached_result
@ -184,10 +180,10 @@ class LLMCachingHandler:
#########################################################
# Init cache timing metrics
#########################################################
cache_check_start_time = time.perf_counter()
cache_check_start_time: Final = time.perf_counter()
cache_check_end_time: float | None = None
#########################################################
parent_otel_span = _get_parent_otel_span_from_kwargs(kwargs)
parent_otel_span: Final = _get_parent_otel_span_from_kwargs(kwargs)
kwargs["parent_otel_span"] = parent_otel_span
if litellm.cache is not None and self._is_call_type_supported_by_cache(original_function=original_function):
@ -201,15 +197,15 @@ class LLMCachingHandler:
if cached_result is not None and not isinstance(cached_result, list):
verbose_logger.debug("Cache Hit!")
cache_hit = True
end_time = datetime.datetime.now()
cache_hit: Final = True
end_time: Final = datetime.datetime.now()
model, custom_llm_provider, _, _ = litellm.get_llm_provider(
model=model,
custom_llm_provider=kwargs.get("custom_llm_provider", None),
api_base=kwargs.get("api_base", None),
api_key=kwargs.get("api_key", None),
)
cache_duration_ms = (cache_check_end_time - cache_check_start_time) * 1000
cache_duration_ms: Final = (cache_check_end_time - cache_check_start_time) * 1000
self._update_litellm_logging_obj_environment(
logging_obj=logging_obj,
model=model,
@ -240,13 +236,13 @@ class LLMCachingHandler:
end_time=end_time,
cache_hit=cache_hit,
)
cache_key = (
cache_key: Final = (
self.preset_cache_key
or self.request_kwargs.get("cache_key")
or litellm.cache.get_cache_key(**self.request_kwargs)
)
if hasattr(cached_result, "_hidden_params"):
cached_result._hidden_params["cache_key"] = cache_key # type: ignore
cached_result._hidden_params["cache_key"] = cache_key
return CachingHandlerResponse(cached_result=cached_result)
elif (
call_type == CallTypes.aembedding.value
@ -271,7 +267,7 @@ class LLMCachingHandler:
embedding_all_elements_cache_hit=embedding_all_elements_cache_hit,
)
verbose_logger.debug(f"CACHE RESULT: {cached_result}")
verbose_logger.debug("CACHE RESULT: %s", cached_result)
return CachingHandlerResponse(
cached_result=cached_result,
final_embedding_cached_response=final_embedding_cached_response,
@ -295,7 +291,7 @@ class LLMCachingHandler:
if litellm.cache is not None and self._is_call_type_supported_by_cache(original_function=original_function):
args = args or ()
# Now that we confirmed caching will happen, prepare kwargs
new_kwargs = kwargs.copy()
new_kwargs: Final = kwargs.copy()
new_kwargs.update(
convert_args_to_kwargs(
self.original_function,
@ -326,8 +322,8 @@ class LLMCachingHandler:
)
# LOG SUCCESS
cache_hit = True
end_time = datetime.datetime.now()
cache_hit: Final = True
end_time: Final = datetime.datetime.now()
(
model,
custom_llm_provider,
@ -354,13 +350,13 @@ class LLMCachingHandler:
end_time=end_time,
cache_hit=cache_hit,
)
cache_key = (
cache_key: Final = (
self.preset_cache_key
or self.request_kwargs.get("cache_key")
or litellm.cache.get_cache_key(**self.request_kwargs)
)
if hasattr(cached_result, "_hidden_params"):
cached_result._hidden_params["cache_key"] = cache_key # type: ignore
cached_result._hidden_params["cache_key"] = cache_key
return CachingHandlerResponse(cached_result=cached_result)
return CachingHandlerResponse(cached_result=cached_result)
@ -420,9 +416,9 @@ class LLMCachingHandler:
"""
embedding_all_elements_cache_hit: bool = False
remaining_list = []
non_null_list = []
kwargs_input_as_list = self.handle_kwargs_input_list_or_str(kwargs)
remaining_list: Final = []
non_null_list: Final = []
kwargs_input_as_list: Final = self.handle_kwargs_input_list_or_str(kwargs)
for idx, cr in enumerate(cached_result):
if cr is None:
remaining_list.append(kwargs_input_as_list[idx])
@ -479,7 +475,7 @@ class LLMCachingHandler:
prompt_tokens_details = PromptTokensDetailsWrapper(**aggregated_details)
except Exception:
prompt_tokens_details = None
usage = Usage(
usage: Final = Usage(
prompt_tokens=prompt_tokens,
completion_tokens=0,
total_tokens=prompt_tokens,
@ -488,9 +484,9 @@ class LLMCachingHandler:
final_embedding_cached_response.usage = usage
if len(remaining_list) == 0:
# LOG SUCCESS
cache_hit = True
cache_hit: Final = True
embedding_all_elements_cache_hit = True
end_time = datetime.datetime.now()
end_time: Final = datetime.datetime.now()
(
model,
custom_llm_provider,
@ -546,10 +542,10 @@ class LLMCachingHandler:
if details2 is None:
return details1
dict1 = details1.model_dump(exclude_none=True) if hasattr(details1, "model_dump") else {}
dict2 = details2.model_dump(exclude_none=True) if hasattr(details2, "model_dump") else {}
dict1: Final = details1.model_dump(exclude_none=True) if hasattr(details1, "model_dump") else {}
dict2: Final = details2.model_dump(exclude_none=True) if hasattr(details2, "model_dump") else {}
merged: dict = {}
merged: Final[dict] = {}
for key in set(dict1.keys()) | set(dict2.keys()):
v1 = dict1.get(key, 0)
v2 = dict2.get(key, 0)
@ -607,7 +603,7 @@ class LLMCachingHandler:
return embedding_response
idx = 0
final_data_list = []
final_data_list: Final = []
for item in _caching_handler_response.final_embedding_cached_response.data:
if item is None and embedding_response.data is not None:
final_data_list.append(embedding_response.data[idx])
@ -690,7 +686,7 @@ class LLMCachingHandler:
if litellm.cache is None:
return None
new_kwargs = kwargs.copy()
new_kwargs: Final = kwargs.copy()
new_kwargs.update(
convert_args_to_kwargs(
self.original_function,
@ -708,7 +704,7 @@ class LLMCachingHandler:
new_kwargs["input"] = [new_kwargs["input"]]
elif not isinstance(new_kwargs["input"], list):
raise ValueError("input must be a string or a list")
tasks = []
tasks: Final = []
for idx, i in enumerate(new_kwargs["input"]):
preset_cache_key = litellm.cache.get_cache_key(**{**new_kwargs, "input": i})
tasks.append(
@ -724,8 +720,8 @@ class LLMCachingHandler:
if all(result is None for result in cached_result):
cached_result = None
else:
request_kwargs = new_kwargs.copy()
request_cache_key = request_kwargs.pop("cache_key", None)
request_kwargs: Final = new_kwargs.copy()
request_cache_key: Final = request_kwargs.pop("cache_key", None)
if litellm.cache._supports_async() is True:
## check if dual cache is supported ##
self.preset_cache_key = request_cache_key or litellm.cache.get_cache_key(**request_kwargs)
@ -828,7 +824,7 @@ class LLMCachingHandler:
elif (call_type == CallTypes.atranscription.value or call_type == CallTypes.transcription.value) and isinstance(
cached_result, dict
):
hidden_params = {
hidden_params: Final = {
"model": "whisper-1",
"custom_llm_provider": custom_llm_provider,
"cache_hit": True,
@ -840,10 +836,10 @@ class LLMCachingHandler:
hidden_params=hidden_params,
)
elif (call_type == "aresponses" or call_type == "responses") and isinstance(cached_result, dict):
use_chat_completion_cache = _is_chat_completion_cached_dict(cached_result)
use_chat_completion_cache: Final = _is_chat_completion_cached_dict(cached_result)
if use_chat_completion_cache:
if kwargs.get("stream", False) is True:
bridge_call_type = (
bridge_call_type: Final = (
CallTypes.acompletion.value if call_type == "aresponses" else CallTypes.completion.value
)
cached_result = self._convert_cached_stream_response(
@ -862,7 +858,7 @@ class LLMCachingHandler:
CachedResponsesAPIStreamingIterator,
)
response_obj = ResponsesAPIResponse(**cached_result)
response_obj: Final = ResponsesAPIResponse(**cached_result)
if (
hasattr(response_obj, "_hidden_params")
and response_obj._hidden_params is not None
@ -957,14 +953,14 @@ class LLMCachingHandler:
if litellm.cache is None:
return
new_kwargs = kwargs.copy()
new_kwargs: Final = kwargs.copy()
new_kwargs.update(
convert_args_to_kwargs(
original_function,
args,
)
)
parent_otel_span = _get_parent_otel_span_from_kwargs(new_kwargs)
parent_otel_span: Final = _get_parent_otel_span_from_kwargs(new_kwargs)
new_kwargs["parent_otel_span"] = parent_otel_span
# [OPTIONAL] ADD TO CACHE
if self._should_store_result_in_cache(original_function=original_function, kwargs=new_kwargs):
@ -1006,7 +1002,7 @@ class LLMCachingHandler:
Sync internal method to add the result to the cache
"""
new_kwargs = kwargs.copy()
new_kwargs: Final = kwargs.copy()
new_kwargs.update(
convert_args_to_kwargs(
self.original_function,
@ -1067,7 +1063,7 @@ class LLMCachingHandler:
"""
complete_streaming_response: ModelResponse | TextCompletionResponse | None = (
complete_streaming_response: Final[ModelResponse | TextCompletionResponse | None] = (
_assemble_complete_response_from_streaming_chunks(
result=processed_chunk,
start_time=self.start_time,
@ -1089,7 +1085,7 @@ class LLMCachingHandler:
"""
Sync internal method to add the streaming response to the cache
"""
complete_streaming_response: ModelResponse | TextCompletionResponse | None = (
complete_streaming_response: Final[ModelResponse | TextCompletionResponse | None] = (
_assemble_complete_response_from_streaming_chunks(
result=processed_chunk,
start_time=self.start_time,
@ -1133,7 +1129,7 @@ class LLMCachingHandler:
Returns:
None
"""
litellm_params = {
litellm_params: Final = {
"logger_fn": kwargs.get("logger_fn", None),
"acompletion": is_async,
"api_base": kwargs.get("api_base", ""),
@ -1173,13 +1169,13 @@ def convert_args_to_kwargs(
args: tuple[Any, ...] | None = None,
) -> dict[str, Any]:
# Get the signature of the original function
signature = inspect.signature(original_function)
signature: Final = inspect.signature(original_function)
# Get parameter names in the order they appear in the original function
param_names = list(signature.parameters.keys())
param_names: Final = list(signature.parameters.keys())
# Create a mapping of positional arguments to parameter names
args_to_kwargs = {}
args_to_kwargs: Final = {}
if args:
for index, arg in enumerate(args):
if index < len(param_names):

16
litellm/caching/disk_cache.py
View file

@ -1,5 +1,5 @@
import json
from typing import TYPE_CHECKING, Any, Union
from typing import TYPE_CHECKING, Any, Final, Union
from .base_cache import BaseCache
@ -41,17 +41,17 @@ class DiskCache(BaseCache):
self.set_cache(key=cache_key, value=cache_value)
def get_cache(self, key, **kwargs):
original_cached_response = self.disk_cache.get(key)
original_cached_response: Final = self.disk_cache.get(key)
if original_cached_response:
try:
cached_response = json.loads(original_cached_response) # type: ignore
cached_response = json.loads(original_cached_response)
except Exception:
cached_response = original_cached_response
return cached_response
return None
def batch_get_cache(self, keys: list, **kwargs):
return_val = []
return_val: Final = []
for k in keys:
val = self.get_cache(key=k, **kwargs)
return_val.append(val)
@ -59,9 +59,9 @@ class DiskCache(BaseCache):
def increment_cache(self, key, value: int, **kwargs) -> int:
with self.disk_cache.transact():
cached_value = self.get_cache(key=key)
init_value = cached_value if isinstance(cached_value, int) else 0
new_value = init_value + value
cached_value: Final = self.get_cache(key=key)
init_value: Final = cached_value if isinstance(cached_value, int) else 0
new_value: Final = init_value + value
self.set_cache(key, new_value, **kwargs)
return new_value
@ -69,7 +69,7 @@ class DiskCache(BaseCache):
return self.get_cache(key=key, **kwargs)
async def async_batch_get_cache(self, keys: list, **kwargs):
return_val = []
return_val: Final = []
for k in keys:
val = self.get_cache(key=k, **kwargs)
return_val.append(val)

34
litellm/caching/dual_cache.py
View file

@ -13,7 +13,7 @@ import time
import traceback
from concurrent.futures import ThreadPoolExecutor
from threading import Lock
from typing import TYPE_CHECKING, Any, Union
from typing import TYPE_CHECKING, Any, Final, Union
if TYPE_CHECKING:
from litellm.types.caching import RedisPipelineIncrementOperation
@ -147,7 +147,7 @@ class DualCache(BaseCache):
return result
except Exception as e:
verbose_logger.error(f"LiteLLM Cache: Excepton async add_cache: {e}")
verbose_logger.error("LiteLLM Cache: Excepton async add_cache: %s", e)
raise e
def get_cache(
@ -161,14 +161,14 @@ class DualCache(BaseCache):
try:
result = None
if self.in_memory_cache is not None:
in_memory_result = self.in_memory_cache.get_cache(key, **kwargs)
in_memory_result: Final = self.in_memory_cache.get_cache(key, **kwargs)
if in_memory_result is not None:
result = in_memory_result
if result is None and self.redis_cache is not None and local_only is False:
# If not found in in-memory cache, try fetching from Redis
redis_result = self.redis_cache.get_cache(key, parent_otel_span=parent_otel_span)
redis_result: Final = self.redis_cache.get_cache(key, parent_otel_span=parent_otel_span)
if redis_result is not None:
# Update in-memory cache with the value from Redis
@ -188,12 +188,12 @@ class DualCache(BaseCache):
local_only: bool = False,
**kwargs,
):
received_args = locals()
received_args: Final = locals()
received_args.pop("self")
def run_in_new_loop():
"""Run the coroutine in a new event loop within this thread."""
new_loop = asyncio.new_event_loop()
new_loop: Final = asyncio.new_event_loop()
try:
asyncio.set_event_loop(new_loop)
return new_loop.run_until_complete(self.async_batch_get_cache(**received_args))
@ -207,7 +207,7 @@ class DualCache(BaseCache):
# If we're already in an event loop, run in a separate thread
# to avoid nested event loop issues
with ThreadPoolExecutor(max_workers=1) as executor:
future = executor.submit(run_in_new_loop)
future: Final = executor.submit(run_in_new_loop)
return future.result()
except RuntimeError:
@ -226,7 +226,7 @@ class DualCache(BaseCache):
print_verbose(f"async get cache: cache key: {key}; local_only: {local_only}")
result = None
if self.in_memory_cache is not None:
in_memory_result = await self.in_memory_cache.async_get_cache(key, **kwargs)
in_memory_result: Final = await self.in_memory_cache.async_get_cache(key, **kwargs)
print_verbose(f"in_memory_result: {in_memory_result}")
if in_memory_result is not None:
@ -234,7 +234,7 @@ class DualCache(BaseCache):
if result is None and self.redis_cache is not None and local_only is False:
# If not found in in-memory cache, try fetching from Redis
redis_result = await self.redis_cache.async_get_cache(key, parent_otel_span=parent_otel_span)
redis_result: Final = await self.redis_cache.async_get_cache(key, parent_otel_span=parent_otel_span)
if redis_result is not None:
# Update in-memory cache with the value from Redis
@ -257,8 +257,8 @@ class DualCache(BaseCache):
Atomically choose keys to fetch from Redis and reserve their access time.
This prevents check-then-act races under concurrent async callers.
"""
sublist_keys: list[str] = []
previous_access_times: dict[str, float | None] = {}
sublist_keys: Final[list[str]] = []
previous_access_times: Final[dict[str, float | None]] = {}
with self._last_redis_batch_access_time_lock:
for key, value in zip(keys, result):
@ -293,7 +293,7 @@ class DualCache(BaseCache):
try:
result = [None] * len(keys)
if self.in_memory_cache is not None:
in_memory_result = await self.in_memory_cache.async_batch_get_cache(keys, **kwargs)
in_memory_result: Final = await self.in_memory_cache.async_batch_get_cache(keys, **kwargs)
if in_memory_result is not None:
result = in_memory_result
@ -303,14 +303,14 @@ class DualCache(BaseCache):
- for the none values in the result
- check the redis cache
"""
current_time = time.time()
current_time: Final = time.time()
sublist_keys, previous_access_times = self._reserve_redis_batch_keys(current_time, keys, result)
# Only hit Redis if enough time has passed since last access.
if len(sublist_keys) > 0:
try:
# If not found in in-memory cache, try fetching from Redis
redis_result = await self.redis_cache.async_batch_get_cache(
redis_result: Final = await self.redis_cache.async_batch_get_cache(
sublist_keys, parent_otel_span=parent_otel_span
)
except Exception:
@ -323,7 +323,7 @@ class DualCache(BaseCache):
return result
# Pre-compute key-to-index mapping for O(1) lookup
key_to_index = {key: i for i, key in enumerate(keys)}
key_to_index: Final = {key: i for i, key in enumerate(keys)}
# Update both result and in-memory cache in a single loop
for key, value in redis_result.items():
@ -347,7 +347,7 @@ class DualCache(BaseCache):
if self.redis_cache is not None and local_only is False:
await self.redis_cache.async_set_cache(key, value, **kwargs)
except Exception as e:
verbose_logger.exception(f"LiteLLM Cache: Excepton async add_cache: {e}")
verbose_logger.exception("LiteLLM Cache: Excepton async add_cache: %s", e)
# async_batch_set_cache
async def async_set_cache_pipeline(self, cache_list: list, local_only: bool = False, **kwargs):
@ -366,7 +366,7 @@ class DualCache(BaseCache):
cache_list=cache_list, ttl=kwargs.pop("ttl", None), **kwargs
)
except Exception as e:
verbose_logger.exception(f"LiteLLM Cache: Excepton async add_cache: {e}")
verbose_logger.exception("LiteLLM Cache: Excepton async add_cache: %s", e)
async def async_increment_cache(
self,

276
litellm/caching/evicted_client_closer.py
View file

@ -1,276 +0,0 @@
"""
Deferred close of HTTP/SDK clients that the LLM client cache has evicted.
Eviction only drops the cache's reference to a client. Every OpenAI/Azure SDK
client is a reference cycle (each resource namespace holds the client back), so
an evicted client and its pooled TCP connections survive until a generational
collection runs, which under load is thousands of requests later.
Closing at eviction time is not an option: a request that was handed the client
just before it was evicted is still using it, and closing it underneath that
request raises ``RuntimeError: Cannot send a request, as the client has been
closed.``
So an evicted client is closed once two conditions hold. A grace window must
have passed since its eviction, which covers a request that holds the client
but is momentarily not on the wire, and the client must report no connection in
flight. The second condition is what keeps the first honest: a request may run
for ``litellm.request_timeout`` seconds, 6000 by default, and a streaming
response is bounded only by how long the upstream keeps sending, so no deadline
on its own can promise that a request has finished.
Only clients litellm itself created are closed; a client the caller supplied is
left alone because litellm does not own its lifecycle.
A client that closes synchronously is closed from wherever the cache is next
used. One whose close is a coroutine needs the event loop it was evicted on, so
it waits for a call from that loop rather than having work scheduled onto a loop
it does not belong to. Queued clients are therefore bucketed by what it takes to
close them, and each bucket is ordered by deadline, so a reap walks the entries
that are due rather than the whole queue.
The queue holds its clients weakly, so waiting out a grace window never keeps
alive anything the collector would have reclaimed first.
"""
import asyncio
import inspect
import threading
import time
import weakref
from collections import deque
from collections.abc import Awaitable, Callable, Iterator
from dataclasses import dataclass, replace
from litellm.constants import (
EVICTED_LLM_CLIENT_CLOSE_GRACE_SECONDS,
EVICTED_LLM_CLIENT_CLOSE_MAX_PENDING,
)
_CLOSABLE_ANYWHERE = "closable-anywhere"
_CLOSABLE_ON_ANY_LOOP = "closable-on-any-loop"
_BucketKey = str | int
@dataclass(frozen=True, slots=True)
class _PendingClose:
"""A queued close.
The client is held weakly, so queueing one never keeps alive anything the
collector would otherwise have reclaimed first.
``needs_loop`` is set for a client whose close is a coroutine; those can only
be closed from the event loop they were evicted on, recorded in ``loop_id``.
A client that closes synchronously carries neither constraint.
"""
client_ref: "weakref.ref[object]"
loop_id: int | None
needs_loop: bool
close_after: float
def _bucket_key(pending: _PendingClose) -> _BucketKey:
"""Which reaps can close this entry: any at all, any running a loop, or one loop's."""
if not pending.needs_loop:
return _CLOSABLE_ANYWHERE
if pending.loop_id is None:
return _CLOSABLE_ON_ANY_LOOP
return pending.loop_id
def _running_loop_id() -> int | None:
try:
return id(asyncio.get_running_loop())
except RuntimeError:
return None
def _close_function(client: object) -> Callable[[], object] | None:
close_fn: Callable[[], object] | None = getattr(client, "aclose", None) or getattr(client, "close", None)
return close_fn
def _transport_of(client: object) -> object:
"""The httpx transport behind an SDK wrapper, a litellm handler, or a bare client."""
for holder in (getattr(client, "_client", None), getattr(client, "client", None), client):
transport: object = getattr(holder, "_transport", None)
if transport is not None:
return transport
return None
def _connection_is_idle(connection: object) -> bool:
"""A pooled connection is idle unless it is servicing a request."""
is_idle: object = getattr(connection, "is_idle", None)
return bool(is_idle()) if callable(is_idle) else True
def _pool_has_busy_connection(transport: object) -> bool | None:
"""Whether the httpcore pool behind the transport is servicing a request.
``None`` when there is no such pool, so the caller can ask the other backend.
"""
pooled: object = getattr(getattr(transport, "_pool", None), "connections", None)
if not isinstance(pooled, (list, tuple)):
return None
return any(
not _connection_is_idle(connection) # pyright: ignore[reportUnknownArgumentType] # untyped pool list
for connection in pooled # pyright: ignore[reportUnknownVariableType] # untyped pool list
)
def _has_connection_in_flight(client: object) -> bool:
"""Whether the client is servicing a request right now.
Both connection backends litellm uses already account for the connections
they have handed out, so this reads the client's own lease accounting rather
than inferring it from elapsed time: httpcore reports a non-idle connection
for the whole of a response including a stream, and aiohttp holds the
connection in ``_acquired`` over the same span.
A client that cannot answer is reported as idle, which leaves the grace
window as the only guard, exactly as it was before this check existed.
"""
try:
transport = _transport_of(client)
pooled_busy = _pool_has_busy_connection(transport)
if pooled_busy is not None:
return pooled_busy
session: object = getattr(transport, "client", None)
return bool(getattr(getattr(session, "connector", None), "_acquired", None))
except Exception: # noqa: BLE001 - a client that cannot report its state is treated as idle
return False
async def _close_quietly(closing: Awaitable[object]) -> None:
try:
await closing
except Exception: # noqa: BLE001 - a discarded client's close must never surface to callers
pass
class EvictedClientCloser:
"""Closes evicted, litellm-owned clients once they are idle and out of grace."""
def __init__(
self,
grace_seconds: float = EVICTED_LLM_CLIENT_CLOSE_GRACE_SECONDS,
max_pending: int = EVICTED_LLM_CLIENT_CLOSE_MAX_PENDING,
clock: Callable[[], float] = time.monotonic,
) -> None:
self._grace_seconds = grace_seconds
self._max_pending = max_pending
self._clock = clock
self._owned: weakref.WeakSet[object] = weakref.WeakSet()
self._buckets: dict[_BucketKey, deque[_PendingClose]] = {} # mutable-ok: deadline-ordered queues
self._pending_count = 0
self._queue_lock = threading.Lock() # the cache is reachable from every worker thread's loop
self._close_tasks: set[asyncio.Task[None]] = set() # mutable-ok: strong refs to running closes
def mark_owned(self, client: object) -> None:
"""Record that litellm created this client, so it may be closed on eviction."""
try:
self._owned.add(client)
except TypeError:
pass # values that cannot be weak-referenced are never litellm clients
def _is_owned(self, client: object) -> bool:
try:
return client in self._owned
except TypeError:
return False # unhashable values are never litellm clients
def schedule(self, client: object) -> None:
"""Queue an evicted client for closing once it is idle and out of grace.
Past ``max_pending`` the client is left to the collector instead, so a
workload that churns the cache cannot grow this queue without bound.
Every queued entry comes due within one grace window, so the capacity it
occupies is returned within that window rather than held.
"""
if client is None or not self._is_owned(client):
return
close_fn = _close_function(client)
if close_fn is None:
return
if self._pending_count >= self._max_pending:
return
self._enqueue(
_PendingClose(
client_ref=weakref.ref(client),
loop_id=_running_loop_id(),
needs_loop=inspect.iscoroutinefunction(close_fn),
close_after=self._clock() + self._grace_seconds,
)
)
def reap(self) -> None:
"""Close every queued client that is due, idle, and closable from here.
Called from the cache's read path, so the empty-queue exit comes first and
the work done past it is proportional to what is due, not to the queue.
"""
if not self._pending_count:
return
now = self._clock()
for pending in self._take_due(_running_loop_id(), now):
client = pending.client_ref()
if client is None:
continue
if _has_connection_in_flight(client):
self._enqueue(replace(pending, close_after=now + self._grace_seconds))
continue
self._close(client)
@property
def pending_count(self) -> int:
return self._pending_count
def _enqueue(self, pending: _PendingClose) -> None:
"""Append to the entry's bucket, dropping any dead entries it queues behind.
Deadlines only ever move forward, so appending keeps each bucket ordered
by deadline, and entries whose client the collector already took sit at
the front rather than having to be searched for.
"""
with self._queue_lock:
bucket = self._buckets.setdefault(_bucket_key(pending), deque()) # mutable-ok: FIFO by design
while bucket and bucket[0].client_ref() is None:
bucket.popleft()
self._pending_count -= 1
bucket.append(pending)
self._pending_count += 1
def _take_due(self, loop_id: int | None, now: float) -> tuple[_PendingClose, ...]:
buckets = (_CLOSABLE_ANYWHERE,) if loop_id is None else (_CLOSABLE_ANYWHERE, _CLOSABLE_ON_ANY_LOOP, loop_id)
with self._queue_lock:
return tuple(pending for key in buckets for pending in self._drain_locked(key, now))
def _drain_locked(self, key: _BucketKey, now: float) -> Iterator[_PendingClose]:
bucket = self._buckets.get(key)
if bucket is None:
return
while bucket and bucket[0].close_after <= now:
self._pending_count -= 1
yield bucket.popleft()
if not bucket:
del self._buckets[key]
def _close(self, client: object) -> None:
close_fn = _close_function(client)
if close_fn is None:
return
try:
closing = close_fn()
except Exception: # noqa: BLE001 - a discarded client's close must never surface to callers
return
if not inspect.isawaitable(closing):
return
task = asyncio.get_running_loop().create_task(_close_quietly(closing))
self._close_tasks.add(task)
task.add_done_callback(self._close_tasks.discard)
default_evicted_client_closer = EvictedClientCloser()

48
litellm/caching/gcs_cache.py
View file

@ -4,6 +4,7 @@ Supports syncing responses to Google Cloud Storage Buckets using HTTP requests.
import asyncio
import json
from typing import Final
from urllib.parse import quote
from litellm._logging import print_verbose, verbose_logger
@ -33,7 +34,7 @@ class GCSCache(BaseCache):
self.sync_client = _get_httpx_client()
def _construct_headers(self) -> dict:
base = GCSBucketBase(bucket_name=self.bucket_name)
base: Final = GCSBucketBase(bucket_name=self.bucket_name)
base.path_service_account_json = self.path_service_account
base.BUCKET_NAME = self.bucket_name
return base.sync_construct_request_headers()
@ -41,55 +42,58 @@ class GCSCache(BaseCache):
def set_cache(self, key, value, **kwargs):
try:
print_verbose(f"LiteLLM SET Cache - GCS. Key={key}. Value={value}")
headers = self._construct_headers()
object_name = self.key_prefix + key
bucket_name = self.bucket_name
headers: Final = self._construct_headers()
object_name: Final = self.key_prefix + key
bucket_name: Final = self.bucket_name
url = f"https://storage.googleapis.com/upload/storage/v1/b/{bucket_name}/o?uploadType=media&name={quote(object_name, safe='')}"
data = json.dumps(value)
data: Final = json.dumps(value)
self.sync_client.post(url=url, data=data, headers=headers)
except Exception as e:
print_verbose(f"GCS Caching: set_cache() - Got exception from GCS: {e}")
async def async_set_cache(self, key, value, **kwargs):
try:
headers = self._construct_headers()
object_name = self.key_prefix + key
bucket_name = self.bucket_name
headers: Final = self._construct_headers()
object_name: Final = self.key_prefix + key
bucket_name: Final = self.bucket_name
url = f"https://storage.googleapis.com/upload/storage/v1/b/{bucket_name}/o?uploadType=media&name={quote(object_name, safe='')}"
data = json.dumps(value)
data: Final = json.dumps(value)
await self.async_client.post(url=url, data=data, headers=headers)
except Exception as e:
print_verbose(f"GCS Caching: async_set_cache() - Got exception from GCS: {e}")
def get_cache(self, key, **kwargs):
try:
headers = self._construct_headers()
object_name = self.key_prefix + key
bucket_name = self.bucket_name
headers: Final = self._construct_headers()
object_name: Final = self.key_prefix + key
bucket_name: Final = self.bucket_name
url = f"https://storage.googleapis.com/storage/v1/b/{bucket_name}/o/{quote(object_name, safe='')}?alt=media"
response = self.sync_client.get(url=url, headers=headers)
response: Final = self.sync_client.get(url=url, headers=headers)
if response.status_code == 200:
cached_response = json.loads(response.text)
cached_response: Final = json.loads(response.text)
verbose_logger.debug(
f"Got GCS Cache: key: {key}, cached_response {cached_response}. Type Response {type(cached_response)}"
"Got GCS Cache: key: %s, cached_response %s. Type Response %s",
key,
cached_response,
type(cached_response),
)
return cached_response
return None
except Exception as e:
verbose_logger.error(f"GCS Caching: get_cache() - Got exception from GCS: {e}")
verbose_logger.error("GCS Caching: get_cache() - Got exception from GCS: %s", e)
async def async_get_cache(self, key, **kwargs):
try:
headers = self._construct_headers()
object_name = self.key_prefix + key
bucket_name = self.bucket_name
headers: Final = self._construct_headers()
object_name: Final = self.key_prefix + key
bucket_name: Final = self.bucket_name
url = f"https://storage.googleapis.com/storage/v1/b/{bucket_name}/o/{quote(object_name, safe='')}?alt=media"
response = await self.async_client.get(url=url, headers=headers)
response: Final = await self.async_client.get(url=url, headers=headers)
if response.status_code == 200:
return json.loads(response.text)
return None
except Exception as e:
verbose_logger.error(f"GCS Caching: async_get_cache() - Got exception from GCS: {e}")
verbose_logger.error("GCS Caching: async_get_cache() - Got exception from GCS: %s", e)
def flush_cache(self):
pass
@ -98,7 +102,7 @@ class GCSCache(BaseCache):
pass
async def async_set_cache_pipeline(self, cache_list, **kwargs):
tasks = []
tasks: Final = []
for val in cache_list:
tasks.append(self.async_set_cache(val[0], val[1], **kwargs))
await asyncio.gather(*tasks)

22
litellm/caching/in_memory_cache.py
View file

@ -13,7 +13,7 @@ import json
import sys
import threading
import time
from typing import TYPE_CHECKING, Any
from typing import TYPE_CHECKING, Any, Final
if TYPE_CHECKING:
from litellm.types.caching import RedisPipelineIncrementOperation
@ -67,7 +67,7 @@ class InMemoryCache(BaseCache):
# Handle special types without full conversion when possible
if hasattr(value, "__sizeof__"): # Use __sizeof__ if available
size = value.__sizeof__() / 1024
size: Final = value.__sizeof__() / 1024
return size <= self.max_size_per_item
# Fallback for complex types
@ -111,7 +111,7 @@ class InMemoryCache(BaseCache):
- 3. the size of in-memory cache is bounded
"""
current_time = time.time()
current_time: Final = time.time()
# Step 1: Remove expired or outdated items
while self.expiration_heap:
@ -144,7 +144,7 @@ class InMemoryCache(BaseCache):
"""
Check if ttl is set for a key
"""
ttl_time = self.ttl_dict.get(key)
ttl_time: Final = self.ttl_dict.get(key)
if ttl_time is None or float(ttl_time) < time.time(): # if ttl is not set, allow override
return True
else:
@ -186,7 +186,7 @@ class InMemoryCache(BaseCache):
Add value to set
"""
# get the value
init_value = self.get_cache(key=key) or set()
init_value: Final = self.get_cache(key=key) or set()
for val in value:
init_value.add(val)
self.set_cache(key, init_value, ttl=ttl)
@ -207,7 +207,7 @@ class InMemoryCache(BaseCache):
if key in self.cache_dict:
if self.evict_element_if_expired(key):
return None
original_cached_response = self.cache_dict[key]
original_cached_response: Final = self.cache_dict[key]
try:
cached_response = json.loads(original_cached_response)
except Exception:
@ -216,7 +216,7 @@ class InMemoryCache(BaseCache):
return None
def batch_get_cache(self, keys: list, **kwargs):
return_val = []
return_val: Final = []
for k in keys:
val = self.get_cache(key=k, **kwargs)
return_val.append(val)
@ -225,7 +225,7 @@ class InMemoryCache(BaseCache):
def increment_cache(self, key, value: float, **kwargs) -> float:
with self._increment_lock:
# keep read-modify-write atomic
init_value = self.get_cache(key=key) or 0
init_value: Final = self.get_cache(key=key) or 0
value = init_value + value
self.set_cache(key, value, **kwargs)
return value
@ -234,7 +234,7 @@ class InMemoryCache(BaseCache):
return self.get_cache(key=key, **kwargs)
async def async_batch_get_cache(self, keys: list, **kwargs):
return_val = []
return_val: Final = []
for k in keys:
val = self.get_cache(key=k, **kwargs)
return_val.append(val)
@ -246,7 +246,7 @@ class InMemoryCache(BaseCache):
async def async_increment_pipeline(
self, increment_list: list["RedisPipelineIncrementOperation"], **kwargs
) -> list[float] | None:
results = []
results: Final = []
for increment in increment_list:
result = await self.async_increment(increment["key"], increment["increment_value"], **kwargs)
results.append(result)
@ -274,5 +274,5 @@ class InMemoryCache(BaseCache):
Get the oldest n keys in the cache
"""
# sorted ttl dict by ttl
sorted_ttl_dict = sorted(self.ttl_dict.items(), key=lambda x: x[1])
sorted_ttl_dict: Final = sorted(self.ttl_dict.items(), key=lambda x: x[1])
return [key for key, _ in sorted_ttl_dict[:n]]

50
litellm/caching/llm_caching_handler.py
View file

@ -3,73 +3,45 @@ Add the event loop to the cache key, to prevent event loop closed errors.
"""
import asyncio
from typing import Final
from .evicted_client_closer import EvictedClientCloser, default_evicted_client_closer
from .in_memory_cache import InMemoryCache
class LLMClientCache(InMemoryCache):
"""Cache for LLM HTTP clients (OpenAI, Azure, httpx, etc.).
An evicted client is never closed on the spot: a request handed the client
just before eviction is still using it, and closing it there raises
``RuntimeError: Cannot send a request, as the client has been closed.``
IMPORTANT: This cache intentionally does NOT close clients on eviction.
Evicted clients may still be in use by in-flight requests. Closing them
eagerly causes ``RuntimeError: Cannot send a request, as the client has
been closed.`` errors in production after the TTL (1 hour) expires.
Nor can eviction be left to rely on garbage collection. The SDK clients are
reference cycles, so an evicted client and its open TCP connections survive
until a generational collection runs. Instead a client litellm created is
handed to ``EvictedClientCloser``, which closes it once a grace window has
passed. Clients the caller supplied are left untouched.
Clients that are no longer referenced will be garbage-collected normally.
For explicit shutdown cleanup, use ``close_litellm_async_clients()``.
"""
def __init__(
self,
max_size_in_memory: int | None = 200,
default_ttl: int | None = 600,
max_size_per_item: int | None = 1024,
evicted_client_closer: EvictedClientCloser | None = None,
):
super().__init__(
max_size_in_memory=max_size_in_memory,
default_ttl=default_ttl,
max_size_per_item=max_size_per_item,
)
self.evicted_client_closer = evicted_client_closer or default_evicted_client_closer
def _remove_key(self, key: str) -> None:
evicted: object = self.cache_dict.get(key)
super()._remove_key(key)
self.evicted_client_closer.schedule(evicted)
self.evicted_client_closer.reap()
def update_cache_key_with_event_loop(self, key):
"""
Add the event loop to the cache key, to prevent event loop closed errors.
If none, use the key as is.
"""
try:
event_loop = asyncio.get_running_loop()
stringified_event_loop = str(id(event_loop))
event_loop: Final = asyncio.get_running_loop()
stringified_event_loop: Final = str(id(event_loop))
return f"{key}-{stringified_event_loop}"
except RuntimeError: # handle no current running event loop
return key
def set_cache(self, key: str, value: object, litellm_owned_client: bool = False, **kwargs):
"""``litellm_owned_client`` marks a client litellm built, so it may be closed once evicted."""
if litellm_owned_client:
self.evicted_client_closer.mark_owned(value)
def set_cache(self, key, value, **kwargs):
key = self.update_cache_key_with_event_loop(key)
return super().set_cache(key, value, **kwargs)
async def async_set_cache(self, key: str, value: object, litellm_owned_client: bool = False, **kwargs):
if litellm_owned_client:
self.evicted_client_closer.mark_owned(value)
async def async_set_cache(self, key, value, **kwargs):
key = self.update_cache_key_with_event_loop(key)
return await super().async_set_cache(key, value, **kwargs)
def get_cache(self, key, **kwargs):
key = self.update_cache_key_with_event_loop(key)
self.evicted_client_closer.reap()
return super().get_cache(key, **kwargs)

80
litellm/caching/qdrant_semantic_cache.py
View file

@ -12,7 +12,7 @@ import ast
import asyncio
import json
import os
from typing import Any, cast
from typing import Any, Final, cast
import litellm
from litellm._logging import print_verbose
@ -88,7 +88,7 @@ class QdrantSemanticCache(BaseCache):
if quantization_config is None:
print_verbose("Quantization config is not provided. Default binary quantization will be used.")
collection_exists = self.sync_client.get(
collection_exists: Final = self.sync_client.get(
url=f"{self.qdrant_api_base}/collections/{self.collection_name}/exists",
headers=self.headers,
)
@ -124,7 +124,7 @@ class QdrantSemanticCache(BaseCache):
else:
raise Exception("Quantization config must be one of 'scalar', 'binary' or 'product'")
new_collection_status = self.sync_client.put(
new_collection_status: Final = self.sync_client.put(
url=f"{self.qdrant_api_base}/collections/{self.collection_name}",
json={
"vectors": {"size": self.vector_size, "distance": "Cosine"},
@ -167,7 +167,7 @@ class QdrantSemanticCache(BaseCache):
def _ensure_cache_key_payload_index(self) -> None:
try:
response = self.sync_client.put(
response: Final = self.sync_client.put(
url=f"{self.qdrant_api_base}/collections/{self.collection_name}/index",
headers=self.headers,
json={
@ -185,7 +185,7 @@ class QdrantSemanticCache(BaseCache):
# payload field. Reassigning them to a caller's key would risk
# cross-scope hits, so they're treated as misses and re-populated on
# the next set_cache.
cached_key = payload.get(self.CACHE_KEY_FIELD_NAME)
cached_key: Final = payload.get(self.CACHE_KEY_FIELD_NAME)
return cached_key is not None and str(cached_key) == str(key)
def _get_embedding(self, prompt: str, metadata: dict[str, Any] | None = None) -> EmbeddingResponse:
@ -196,7 +196,7 @@ class QdrantSemanticCache(BaseCache):
llm_model_list = None
llm_router = None
router = resolve_embedding_router(self.embedding_model, llm_router, llm_model_list)
router: Final = resolve_embedding_router(self.embedding_model, llm_router, llm_model_list)
if router is not None:
return router.embedding(
model=self.embedding_model,
@ -217,7 +217,7 @@ class QdrantSemanticCache(BaseCache):
llm_model_list = None
llm_router = None
router = resolve_embedding_router(self.embedding_model, llm_router, llm_model_list)
router: Final = resolve_embedding_router(self.embedding_model, llm_router, llm_model_list)
if router is not None:
return await router.aembedding(
model=self.embedding_model,
@ -237,22 +237,22 @@ class QdrantSemanticCache(BaseCache):
from litellm._uuid import uuid
# get the prompt
messages = kwargs["messages"]
prompt = get_str_from_messages(messages)
messages: Final = kwargs["messages"]
prompt: Final = get_str_from_messages(messages)
# create an embedding for prompt
embedding_response = cast(
embedding_response: Final = cast(
EmbeddingResponse,
self._get_embedding(prompt, metadata=kwargs.get("metadata")),
)
# get the embedding
embedding = embedding_response["data"][0]["embedding"]
embedding: Final = embedding_response["data"][0]["embedding"]
value = str(value)
assert isinstance(value, str)
data = {
data: Final = {
"points": [
{
"id": str(uuid.uuid4()),
@ -275,19 +275,19 @@ class QdrantSemanticCache(BaseCache):
print_verbose(f"sync qdrant semantic-cache get_cache, kwargs: {kwargs}")
# get the messages
messages = kwargs["messages"]
prompt = get_str_from_messages(messages)
messages: Final = kwargs["messages"]
prompt: Final = get_str_from_messages(messages)
# convert to embedding
embedding_response = cast(
embedding_response: Final = cast(
EmbeddingResponse,
self._get_embedding(prompt, metadata=kwargs.get("metadata")),
)
# get the embedding
embedding = embedding_response["data"][0]["embedding"]
embedding: Final = embedding_response["data"][0]["embedding"]
data = {
data: Final = {
"vector": embedding,
"params": {
"quantization": {
@ -301,12 +301,12 @@ class QdrantSemanticCache(BaseCache):
}
self._add_cache_key_filter_to_search_data(data=data, key=key)
search_response = self.sync_client.post(
search_response: Final = self.sync_client.post(
url=f"{self.qdrant_api_base}/collections/{self.collection_name}/points/search",
headers=self.headers,
json=data,
)
results = search_response.json()["result"]
results: Final = search_response.json()["result"]
if results is None:
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
@ -316,14 +316,14 @@ class QdrantSemanticCache(BaseCache):
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
return None
similarity = results[0]["score"]
payload = results[0]["payload"]
similarity: Final = results[0]["score"]
payload: Final = results[0]["payload"]
if not self._payload_matches_cache_key(payload=payload, key=key):
print_verbose("Qdrant semantic-cache hit did not match cache key scope")
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
return None
cached_prompt = payload["text"]
cached_prompt: Final = payload["text"]
# check similarity, if more than self.similarity_threshold, return results
print_verbose(
@ -335,7 +335,7 @@ class QdrantSemanticCache(BaseCache):
if similarity >= self.similarity_threshold:
# cache hit !
cached_value = payload["response"]
cached_value: Final = payload["response"]
print_verbose(
f"got a cache hit, similarity: {similarity}, Current prompt: {prompt}, cached_prompt: {cached_prompt}"
)
@ -350,17 +350,17 @@ class QdrantSemanticCache(BaseCache):
print_verbose(f"async qdrant semantic-cache set_cache, kwargs: {kwargs}")
# get the prompt
messages = kwargs["messages"]
prompt = get_str_from_messages(messages)
embedding_response = await self._get_async_embedding(prompt, metadata=kwargs.get("metadata"))
messages: Final = kwargs["messages"]
prompt: Final = get_str_from_messages(messages)
embedding_response: Final = await self._get_async_embedding(prompt, metadata=kwargs.get("metadata"))
# get the embedding
embedding = embedding_response["data"][0]["embedding"]
embedding: Final = embedding_response["data"][0]["embedding"]
value = str(value)
assert isinstance(value, str)
data = {
data: Final = {
"points": [
{
"id": str(uuid.uuid4()),
@ -384,15 +384,15 @@ class QdrantSemanticCache(BaseCache):
print_verbose(f"async qdrant semantic-cache get_cache, kwargs: {kwargs}")
# get the messages
messages = kwargs["messages"]
prompt = get_str_from_messages(messages)
messages: Final = kwargs["messages"]
prompt: Final = get_str_from_messages(messages)
embedding_response = await self._get_async_embedding(prompt, metadata=kwargs.get("metadata"))
embedding_response: Final = await self._get_async_embedding(prompt, metadata=kwargs.get("metadata"))
# get the embedding
embedding = embedding_response["data"][0]["embedding"]
embedding: Final = embedding_response["data"][0]["embedding"]
data = {
data: Final = {
"vector": embedding,
"params": {
"quantization": {
@ -406,13 +406,13 @@ class QdrantSemanticCache(BaseCache):
}
self._add_cache_key_filter_to_search_data(data=data, key=key)
search_response = await self.async_client.post(
search_response: Final = await self.async_client.post(
url=f"{self.qdrant_api_base}/collections/{self.collection_name}/points/search",
headers=self.headers,
json=data,
)
results = search_response.json()["result"]
results: Final = search_response.json()["result"]
if results is None:
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
@ -422,14 +422,14 @@ class QdrantSemanticCache(BaseCache):
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
return None
similarity = results[0]["score"]
payload = results[0]["payload"]
similarity: Final = results[0]["score"]
payload: Final = results[0]["payload"]
if not self._payload_matches_cache_key(payload=payload, key=key):
print_verbose("Qdrant semantic-cache hit did not match cache key scope")
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
return None
cached_prompt = payload["text"]
cached_prompt: Final = payload["text"]
# check similarity, if more than self.similarity_threshold, return results
print_verbose(
@ -441,7 +441,7 @@ class QdrantSemanticCache(BaseCache):
if similarity >= self.similarity_threshold:
# cache hit !
cached_value = payload["response"]
cached_value: Final = payload["response"]
print_verbose(
f"got a cache hit, similarity: {similarity}, Current prompt: {prompt}, cached_prompt: {cached_prompt}"
)
@ -454,7 +454,7 @@ class QdrantSemanticCache(BaseCache):
return self.collection_info
async def async_set_cache_pipeline(self, cache_list, **kwargs):
tasks = []
tasks: Final = []
for val in cache_list:
tasks.append(self.async_set_cache(val[0], val[1], **kwargs))
await asyncio.gather(*tasks)

277
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, TypeVar, Union, cast
from typing import TYPE_CHECKING, Any, Final, TypeVar, Union, cast
import litellm
from litellm._logging import print_verbose, verbose_logger
@ -69,18 +69,18 @@ def _get_call_stack_info(num_frames: int = 2) -> str:
A string with format "current_function <- caller_function [<- grandparent_function]"
"""
try:
current_frame = inspect.currentframe()
current_frame: Final = inspect.currentframe()
if current_frame is None:
return "unknown"
# Skip this function and the immediate caller (which sets call_type)
f_back = current_frame.f_back
f_back: Final = current_frame.f_back
if f_back is None:
return "unknown"
frame = f_back.f_back
if frame is None:
return "unknown"
function_names = []
function_names: Final = []
for _ in range(num_frames):
if frame is None:
@ -172,7 +172,7 @@ class RedisCircuitBreaker:
_RedisCallResult = TypeVar("_RedisCallResult")
_swallowed_redis_failures: ContextVar[int] = ContextVar("litellm_swallowed_redis_failures", default=0)
_swallowed_redis_failures: Final[ContextVar[int]] = ContextVar("litellm_swallowed_redis_failures", default=0)
@functools.lru_cache(maxsize=1)
@ -230,9 +230,9 @@ async def _run_under_circuit_breaker(
"""
if breaker.is_open():
raise Exception(f"Redis circuit breaker is open — skipping {name}")
swallowed_before = _swallowed_redis_failures.get()
swallowed_before: Final = _swallowed_redis_failures.get()
try:
result = await call()
result: Final = await call()
except Exception as e:
if _is_redis_health_failure(e):
breaker.record_failure()
@ -242,7 +242,7 @@ async def _run_under_circuit_breaker(
return result
def _redis_circuit_breaker_guard(method): # type: ignore
def _redis_circuit_breaker_guard(method):
"""
Decorator for RedisCache async methods.
Checks the circuit breaker before each call; records success/failure after.
@ -256,7 +256,7 @@ def _redis_circuit_breaker_guard(method): # type: ignore
"""
@functools.wraps(method)
async def wrapper(self, *args, **kwargs): # type: ignore
async def wrapper(self, *args, **kwargs):
return await _run_under_circuit_breaker(
self._circuit_breaker, method.__name__, lambda: method(self, *args, **kwargs)
)
@ -282,7 +282,7 @@ class RedisCache(BaseCache):
from .._redis import get_redis_client, get_redis_connection_pool
redis_kwargs = {}
redis_kwargs: Final = {}
if host is not None:
redis_kwargs["host"] = host
if port is not None:
@ -319,7 +319,7 @@ class RedisCache(BaseCache):
self.redis_version = "Unknown"
try:
if not coroutine_checker.is_async_callable(self.redis_client):
self.redis_version = self.redis_client.info()["redis_version"] # type: ignore
self.redis_version = self.redis_client.info()["redis_version"]
except Exception:
pass
@ -346,7 +346,8 @@ class RedisCache(BaseCache):
verbose_logger.debug("Ignoring async redis ping. No running event loop.")
else:
verbose_logger.error(
f"Error connecting to Async Redis client - {e}",
"Error connecting to Async Redis client - %s",
e,
extra={"error": str(e)},
)
self._handle_async_ping_error(e)
@ -354,7 +355,7 @@ class RedisCache(BaseCache):
# SYNC HEALTH PING
try:
if hasattr(self.redis_client, "ping"):
self.redis_client.ping() # type: ignore
self.redis_client.ping()
except Exception as e:
verbose_logger.error("Error connecting to Sync Redis client", extra={"error": str(e)})
self._handle_sync_ping_error(e)
@ -362,9 +363,9 @@ class RedisCache(BaseCache):
def _handle_async_ping_error(self, e: Exception):
"""Handle async ping error with service failure hook."""
try:
loop = asyncio.get_running_loop()
start_time = time.time()
end_time = start_time
loop: Final = asyncio.get_running_loop()
start_time: Final = time.time()
end_time: Final = start_time
loop.create_task(
self.service_logger_obj.async_service_failure_hook(
service=ServiceTypes.REDIS,
@ -379,9 +380,9 @@ class RedisCache(BaseCache):
def _handle_sync_ping_error(self, e: Exception):
"""Handle sync ping error with service failure hook."""
try:
loop = asyncio.get_running_loop()
start_time = time.time()
end_time = start_time
loop: Final = asyncio.get_running_loop()
start_time: Final = time.time()
end_time: Final = start_time
loop.create_task(
self.service_logger_obj.async_service_failure_hook(
service=ServiceTypes.REDIS,
@ -400,9 +401,9 @@ class RedisCache(BaseCache):
"""
# Create a stable representation of redis_kwargs for hashing
# Sort keys to ensure consistent hash regardless of parameter order
sorted_kwargs = sorted(self.redis_kwargs.items())
kwargs_str = json.dumps(sorted_kwargs, sort_keys=True)
kwargs_hash = hashlib.sha256(kwargs_str.encode()).hexdigest()[:16]
sorted_kwargs: Final = sorted(self.redis_kwargs.items())
kwargs_str: Final = json.dumps(sorted_kwargs, sort_keys=True)
kwargs_hash: Final = hashlib.sha256(kwargs_str.encode()).hexdigest()[:16]
return f"async-redis-client-{kwargs_hash}"
def init_async_client(
@ -412,8 +413,8 @@ class RedisCache(BaseCache):
from .._redis import get_redis_async_client, get_redis_connection_pool
cache_key = self._get_async_client_cache_key()
cached_client = in_memory_llm_clients_cache.get_cache(key=cache_key)
cache_key: Final = self._get_async_client_cache_key()
cached_client: Final = in_memory_llm_clients_cache.get_cache(key=cache_key)
if cached_client is not None:
redis_async_client = cast(async_redis_client | async_redis_cluster_client, cached_client)
else:
@ -422,7 +423,7 @@ class RedisCache(BaseCache):
redis_async_client = get_redis_async_client(connection_pool=self.async_redis_conn_pool, **self.redis_kwargs)
in_memory_llm_clients_cache.set_cache(key=cache_key, value=redis_async_client)
self.redis_async_client = redis_async_client # type: ignore
self.redis_async_client = redis_async_client
return redis_async_client
def check_and_fix_namespace(self, key: str) -> str:
@ -430,7 +431,7 @@ class RedisCache(BaseCache):
Make sure each key starts with the given namespace
"""
if key is None:
return key # type: ignore[return-value]
return key
if self.namespace is not None and not key.startswith(self.namespace):
key = self.namespace + ":" + key
@ -453,7 +454,7 @@ class RedisCache(BaseCache):
return DEFAULT_REDIS_MAJOR_VERSION
try:
version_str = str(self.redis_version).strip()
version_str: Final = str(self.redis_version).strip()
# Handle cases where there's no dot (e.g., "7" or 7)
if "." in version_str:
major_version = int(version_str.split(".")[0])
@ -466,14 +467,14 @@ class RedisCache(BaseCache):
return DEFAULT_REDIS_MAJOR_VERSION
def set_cache(self, key, value, **kwargs):
ttl = self.get_ttl(**kwargs)
ttl: Final = self.get_ttl(**kwargs)
print_verbose(f"Set Redis Cache: key: {key}\nValue {value}\nttl={ttl}, redis_version={self.redis_version}")
key = self.check_and_fix_namespace(key=key)
try:
start_time = time.time()
start_time: Final = time.time()
self.redis_client.set(name=key, value=str(value), ex=ttl)
end_time = time.time()
_duration = end_time - start_time
end_time: Final = time.time()
_duration: Final = end_time - start_time
self.service_logger_obj.service_success_hook(
service=ServiceTypes.REDIS,
duration=_duration,
@ -486,13 +487,13 @@ class RedisCache(BaseCache):
print_verbose(f"litellm.caching.caching: set() - Got exception from REDIS : {e}")
def increment_cache(self, key, value: int, ttl: float | None = None, **kwargs) -> int:
_redis_client = self.redis_client
_redis_client: Final = self.redis_client
start_time = time.time()
set_ttl = self.get_ttl(ttl=ttl)
set_ttl: Final = self.get_ttl(ttl=ttl)
key = self.check_and_fix_namespace(key=key)
try:
start_time = time.time()
result: int = _redis_client.incr(name=key, amount=value) # type: ignore
result: Final[int] = _redis_client.incr(name=key, amount=value)
end_time = time.time()
_duration = end_time - start_time
self.service_logger_obj.service_success_hook(
@ -506,7 +507,7 @@ class RedisCache(BaseCache):
if set_ttl is not None:
# check if key already has ttl, if not -> set ttl
start_time = time.time()
current_ttl = _redis_client.ttl(key)
current_ttl: Final = _redis_client.ttl(key)
end_time = time.time()
_duration = end_time - start_time
self.service_logger_obj.service_success_hook(
@ -519,7 +520,7 @@ class RedisCache(BaseCache):
if current_ttl == -1:
# Key has no expiration
start_time = time.time()
_redis_client.expire(key, set_ttl) # type: ignore
_redis_client.expire(key, set_ttl)
end_time = time.time()
_duration = end_time - start_time
self.service_logger_obj.service_success_hook(
@ -543,10 +544,10 @@ class RedisCache(BaseCache):
@_redis_circuit_breaker_guard
async def async_scan_iter(self, pattern: str, count: int = 100) -> list:
start_time = time.time()
start_time: Final = time.time()
try:
keys = []
_redis_client = self.init_async_client()
keys: Final = []
_redis_client: Final = self.init_async_client()
if not hasattr(_redis_client, "scan_iter"):
verbose_logger.debug(
"Redis client does not support scan_iter, potentially using Redis Cluster. Returning empty list."
@ -554,7 +555,7 @@ class RedisCache(BaseCache):
return []
pattern = self.check_and_fix_namespace(key=pattern)
async for key in _redis_client.scan_iter(match=pattern + "*", count=count): # type: ignore
async for key in _redis_client.scan_iter(match=pattern + "*", count=count):
keys.append(key)
if len(keys) >= count:
break
@ -619,7 +620,7 @@ class RedisCache(BaseCache):
# different key prefixes never share an executor; in_memory_llm_clients_cache
# then adds the running loop, completing the per-(client, namespace, loop)
# scoping.
script_cache_key = (
script_cache_key: Final = (
f"redis-registered-script-{self._get_async_client_cache_key()}-"
f"{self.namespace}-{hashlib.sha256(script.encode()).hexdigest()[:16]}"
)
@ -645,21 +646,21 @@ class RedisCache(BaseCache):
Kept separate from async_register_script so each loop caches its own
executor; see that method for why the binding must be per loop.
"""
_redis_client: Any = self.init_async_client()
_redis_client: Final[Any] = self.init_async_client()
if hasattr(_redis_client, "register_script"):
registered_script = _redis_client.register_script(script)
registered_script: Final = _redis_client.register_script(script)
async def standalone_executor(keys: Sequence[str], args: Sequence[Any], client: Any = None) -> Any:
namespaced_keys = tuple(self.check_and_fix_namespace(key=key) for key in keys)
namespaced_keys: Final = tuple(self.check_and_fix_namespace(key=key) for key in keys)
return await registered_script(keys=namespaced_keys, args=args, client=client)
return standalone_executor
if hasattr(_redis_client, "script_load"):
script_sha = _redis_client.script_load(script)
script_sha: Final = _redis_client.script_load(script)
async def cluster_executor(keys: Sequence[str], args: Sequence[Any], client: Any = None) -> Any:
namespaced_keys = tuple(self.check_and_fix_namespace(key=key) for key in keys)
namespaced_keys: Final = tuple(self.check_and_fix_namespace(key=key) for key in keys)
return await _redis_client.evalsha(script_sha, len(namespaced_keys), *namespaced_keys, *args)
return cluster_executor
@ -677,9 +678,9 @@ class RedisCache(BaseCache):
)
return None
start_time = time.time()
start_time: Final = time.time()
try:
_redis_client: Redis = self.init_async_client() # type: ignore
_redis_client: Final[Redis] = self.init_async_client()
except Exception as e:
end_time = time.time()
_duration = end_time - start_time
@ -703,14 +704,14 @@ class RedisCache(BaseCache):
raise e
key = self.check_and_fix_namespace(key=key)
ttl = self.get_ttl(**kwargs)
nx = kwargs.get("nx", False)
ttl: Final = self.get_ttl(**kwargs)
nx: Final = kwargs.get("nx", False)
print_verbose(f"Set ASYNC Redis Cache: key: {key}\nValue {value}\nttl={ttl}")
try:
if not hasattr(_redis_client, "set"):
raise Exception("Redis client cannot set cache. Attribute not found.")
result = await _redis_client.set(
result: Final = await _redis_client.set(
name=key,
value=json.dumps(value),
nx=nx,
@ -772,13 +773,13 @@ class RedisCache(BaseCache):
_td: timedelta | None = None
if ttl is not None:
_td = timedelta(seconds=ttl)
pipe.set( # type: ignore
pipe.set(
name=cache_key,
value=json_cache_value,
ex=_td,
)
# Execute the pipeline and return the results.
results = await pipe.execute()
results: Final = await pipe.execute()
return results
@_redis_circuit_breaker_guard
@ -790,14 +791,14 @@ class RedisCache(BaseCache):
if len(cache_list) == 0:
return
_redis_client = self.init_async_client()
start_time = time.time()
_redis_client: Final = self.init_async_client()
start_time: Final = time.time()
print_verbose(f"Set Async Redis Cache: key list: {cache_list}\nttl={ttl}, redis_version={self.redis_version}")
cache_value: Any = None
cache_value: Final[Any] = None
try:
async with _redis_client.pipeline(transaction=False) as pipe:
results = await self._pipeline_helper(pipe, cache_list, ttl)
results: Final = await self._pipeline_helper(pipe, cache_list, ttl)
print_verbose(f"pipeline results: {results}")
# Optionally, you could process 'results' to make sure that all set operations were successful.
@ -848,9 +849,9 @@ class RedisCache(BaseCache):
"""Helper function for async_set_cache_sadd. Separated for testing."""
ttl = self.get_ttl(ttl=ttl)
try:
await redis_client.sadd(key, *value) # type: ignore
await redis_client.sadd(key, *value)
if ttl is not None:
_td = timedelta(seconds=ttl)
_td: Final = timedelta(seconds=ttl)
await redis_client.expire(key, _td)
except Exception:
raise
@ -859,9 +860,9 @@ class RedisCache(BaseCache):
async def async_set_cache_sadd(self, key, value: list, ttl: float | None, **kwargs):
from redis.asyncio import Redis
start_time = time.time()
start_time: Final = time.time()
try:
_redis_client: Redis = self.init_async_client() # type: ignore
_redis_client: Final[Redis] = self.init_async_client()
except Exception as e:
end_time = time.time()
_duration = end_time - start_time
@ -944,17 +945,17 @@ class RedisCache(BaseCache):
) -> float:
from redis.asyncio import Redis
_redis_client: Redis = self.init_async_client() # type: ignore
start_time = time.time()
_used_ttl = self.get_ttl(ttl=ttl)
_redis_client: Final[Redis] = self.init_async_client()
start_time: Final = time.time()
_used_ttl: Final = self.get_ttl(ttl=ttl)
key = self.check_and_fix_namespace(key=key)
try:
result = await _redis_client.incrbyfloat(name=key, amount=value)
result: Final = await _redis_client.incrbyfloat(name=key, amount=value)
if _used_ttl is not None:
if refresh_ttl:
await _redis_client.expire(key, _used_ttl)
else:
current_ttl = await _redis_client.ttl(key)
current_ttl: Final = await _redis_client.ttl(key)
if current_ttl == -1:
await _redis_client.expire(key, _used_ttl)
@ -1011,10 +1012,10 @@ class RedisCache(BaseCache):
GET/compare/SET runs in a single Lua call, so it is also atomic across
racing callers and pods. Returns the resulting value.
"""
_redis_client = self.init_async_client()
_used_ttl = self.get_ttl(ttl=ttl)
_redis_client: Final = self.init_async_client()
_used_ttl: Final = self.get_ttl(ttl=ttl)
key = self.check_and_fix_namespace(key=key)
lua = (
lua: Final = (
"local cur = redis.call('GET', KEYS[1]) "
"if cur == false or tonumber(cur) < tonumber(ARGV[1]) then "
"redis.call('SET', KEYS[1], ARGV[1]) "
@ -1055,10 +1056,10 @@ class RedisCache(BaseCache):
try:
key = self.check_and_fix_namespace(key=key)
print_verbose(f"Get Redis Cache: key: {key}")
start_time = time.time()
cached_response = self.redis_client.get(key)
end_time = time.time()
_duration = end_time - start_time
start_time: Final = time.time()
cached_response: Final = self.redis_client.get(key)
end_time: Final = time.time()
_duration: Final = end_time - start_time
self.service_logger_obj.service_success_hook(
service=ServiceTypes.REDIS,
duration=_duration,
@ -1079,7 +1080,7 @@ class RedisCache(BaseCache):
We use a wrapper so RedisCluster can override this method
"""
return self.redis_client.mget(keys=keys) # type: ignore
return self.redis_client.mget(keys=keys)
async def _async_run_redis_mget_operation(self, keys: list[str]) -> list[Any]:
"""
@ -1087,8 +1088,8 @@ class RedisCache(BaseCache):
We use a wrapper so RedisCluster can override this method
"""
async_redis_client = self.init_async_client()
return await async_redis_client.mget(keys=keys) # type: ignore
async_redis_client: Final = self.init_async_client()
return await async_redis_client.mget(keys=keys)
def batch_get_cache(
self,
@ -1106,17 +1107,17 @@ class RedisCache(BaseCache):
dict: A dictionary mapping keys to their cached values
"""
key_value_dict = {}
_key_list = [key for key in key_list if key is not None]
_key_list: Final = [key for key in key_list if key is not None]
try:
_keys = []
_keys: Final = []
for cache_key in _key_list:
cache_key = self.check_and_fix_namespace(key=cache_key or "")
_keys.append(cache_key)
start_time = time.time()
results: list = self._run_redis_mget_operation(keys=_keys)
end_time = time.time()
_duration = end_time - start_time
start_time: Final = time.time()
results: Final[list] = self._run_redis_mget_operation(keys=_keys)
end_time: Final = time.time()
_duration: Final = end_time - start_time
self.service_logger_obj.service_success_hook(
service=ServiceTypes.REDIS,
duration=_duration,
@ -1130,7 +1131,7 @@ class RedisCache(BaseCache):
# 'results' is a list of values corresponding to the order of keys in '_key_list'.
key_value_dict = dict(zip(_key_list, results))
decoded_results = {}
decoded_results: Final = {}
for k, v in key_value_dict.items():
if isinstance(k, bytes):
k = k.decode("utf-8")
@ -1139,22 +1140,22 @@ class RedisCache(BaseCache):
return decoded_results
except Exception as e:
verbose_logger.error(f"Error occurred in batch get cache - {e}")
verbose_logger.error("Error occurred in batch get cache - %s", e)
return key_value_dict
@_redis_circuit_breaker_guard
async def async_get_cache(self, key, parent_otel_span: Span | None = None, **kwargs):
from redis.asyncio import Redis
_redis_client: Redis = self.init_async_client() # type: ignore
_redis_client: Final[Redis] = self.init_async_client()
key = self.check_and_fix_namespace(key=key)
start_time = time.time()
start_time: Final = time.time()
try:
print_verbose(f"Get Async Redis Cache: key: {key}")
cached_response = await _redis_client.get(key)
cached_response: Final = await _redis_client.get(key)
print_verbose(f"Got Async Redis Cache: key: {key}, cached_response {cached_response}")
response = self._get_cache_logic(cached_response=cached_response)
response: Final = self._get_cache_logic(cached_response=cached_response)
end_time = time.time()
_duration = end_time - start_time
@ -1208,14 +1209,14 @@ class RedisCache(BaseCache):
"""
# typed as Any, redis python lib has incomplete type stubs for RedisCluster and does not include `mget`
key_value_dict = {}
start_time = time.time()
_key_list = [key for key in key_list if key is not None]
start_time: Final = time.time()
_key_list: Final = [key for key in key_list if key is not None]
try:
_keys = []
_keys: Final = []
for cache_key in _key_list:
cache_key = self.check_and_fix_namespace(key=cache_key)
_keys.append(cache_key)
results = await self._async_run_redis_mget_operation(keys=_keys)
results: Final = await self._async_run_redis_mget_operation(keys=_keys)
## LOGGING ##
end_time = time.time()
_duration = end_time - start_time
@ -1234,7 +1235,7 @@ class RedisCache(BaseCache):
# 'results' is a list of values corresponding to the order of keys in 'key_list'.
key_value_dict = dict(zip(_key_list, results))
decoded_results = {}
decoded_results: Final = {}
for k, v in key_value_dict.items():
if isinstance(k, bytes):
k = k.decode("utf-8")
@ -1257,7 +1258,7 @@ class RedisCache(BaseCache):
parent_otel_span=parent_otel_span,
)
)
verbose_logger.error(f"Error occurred in async batch get cache - {e}")
verbose_logger.error("Error occurred in async batch get cache - %s", e)
_record_swallowed_redis_failure(self._circuit_breaker, e)
return key_value_dict
@ -1266,9 +1267,9 @@ class RedisCache(BaseCache):
Tests if the sync redis client is correctly setup.
"""
print_verbose("Pinging Sync Redis Cache")
start_time = time.time()
start_time: Final = time.time()
try:
response: bool = self.redis_client.ping() # type: ignore
response: Final[bool] = self.redis_client.ping()
print_verbose(f"Redis Cache PING: {response}")
## LOGGING ##
end_time = time.time()
@ -1292,16 +1293,16 @@ class RedisCache(BaseCache):
error=e,
call_type=f"sync_ping <- {_get_call_stack_info()}",
)
verbose_logger.error(f"LiteLLM Redis Cache PING: - Got exception from REDIS : {e}")
verbose_logger.error("LiteLLM Redis Cache PING: - Got exception from REDIS : %s", e)
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: Any = self.init_async_client()
start_time = time.time()
_redis_client: Final[Any] = self.init_async_client()
start_time: Final = time.time()
print_verbose("Pinging Async Redis Cache")
try:
response = await _redis_client.ping()
response: Final = await _redis_client.ping()
## LOGGING ##
end_time = time.time()
_duration = end_time - start_time
@ -1326,23 +1327,23 @@ class RedisCache(BaseCache):
call_type=f"async_ping <- {_get_call_stack_info()}",
)
)
verbose_logger.error(f"LiteLLM Redis Cache PING: - Got exception from REDIS : {e}")
verbose_logger.error("LiteLLM Redis Cache PING: - Got exception from REDIS : %s", e)
raise e
@_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: Any = self.init_async_client()
_redis_client: Final[Any] = self.init_async_client()
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)
def client_list(self) -> list:
client_list: list = self.redis_client.client_list() # type: ignore
client_list: Final[list] = self.redis_client.client_list()
return client_list
def info(self):
info = self.redis_client.info()
info: Final = self.redis_client.info()
return info
def flush_cache(self):
@ -1372,13 +1373,13 @@ class RedisCache(BaseCache):
import redis.asyncio as redis_async
# Create a fresh Redis client with current settings
redis_client = redis_async.Redis(**self.redis_kwargs)
redis_client: Final = redis_async.Redis(**self.redis_kwargs)
# Test the connection
ping_result = await redis_client.ping() # type: ignore[misc]
ping_result: Final = await redis_client.ping()
# Close the connection
await redis_client.aclose() # type: ignore[attr-defined]
await redis_client.aclose()
if ping_result:
return {
@ -1388,7 +1389,7 @@ class RedisCache(BaseCache):
else:
return {"status": "failed", "message": "Redis ping returned False"}
except Exception as e:
verbose_logger.error(f"Redis connection test failed: {e}")
verbose_logger.error("Redis connection test failed: %s", e)
return {
"status": "failed",
"message": f"Redis connection failed: {e}",
@ -1398,7 +1399,7 @@ 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: Any = self.init_async_client()
_redis_client: Final[Any] = self.init_async_client()
key = self.check_and_fix_namespace(key=key)
# keys is str
return await _redis_client.delete(key)
@ -1424,9 +1425,9 @@ class RedisCache(BaseCache):
_td = timedelta(seconds=increment_op["ttl"])
pipe.expire(cache_key, _td)
# Execute the pipeline and return results
results = await pipe.execute()
results: Final = await pipe.execute()
# only return float values
verbose_logger.debug(f"Increment ASYNC Redis Cache PIPELINE: results: {results}")
verbose_logger.debug("Increment ASYNC Redis Cache PIPELINE: results: %s", results)
return [r for r in results if isinstance(r, float)]
@_redis_circuit_breaker_guard
@ -1447,14 +1448,14 @@ class RedisCache(BaseCache):
from redis.asyncio import Redis
_redis_client: Redis = self.init_async_client() # type: ignore
start_time = time.time()
_redis_client: Final[Redis] = self.init_async_client()
start_time: Final = time.time()
print_verbose(f"Increment Async Redis Cache Pipeline: increment list: {increment_list}")
try:
async with _redis_client.pipeline(transaction=False) as pipe:
results = await self._pipeline_increment_helper(pipe, increment_list)
results: Final = await self._pipeline_increment_helper(pipe, increment_list)
## LOGGING ##
end_time = time.time()
@ -1506,14 +1507,14 @@ class RedisCache(BaseCache):
"""
try:
# typed as Any, redis python lib has incomplete type stubs for RedisCluster and does not include `ttl`
_redis_client: Any = self.init_async_client()
_redis_client: Final[Any] = self.init_async_client()
key = self.check_and_fix_namespace(key=key)
ttl = await _redis_client.ttl(key)
ttl: Final = await _redis_client.ttl(key)
if ttl <= -1: # -1 means the key does not exist, -2 key does not exist
return None
return ttl
except Exception as e:
verbose_logger.debug(f"Redis TTL Error: {e}")
verbose_logger.debug("Redis TTL Error: %s", e)
_record_swallowed_redis_failure(self._circuit_breaker, e)
return None
@ -1536,11 +1537,11 @@ class RedisCache(BaseCache):
Returns:
int: The length of the list after the push operation
"""
_redis_client: Any = self.init_async_client()
_redis_client: Final[Any] = self.init_async_client()
key = self.check_and_fix_namespace(key=key)
start_time = time.time()
start_time: Final = time.time()
try:
response = await _redis_client.rpush(key, *values)
response: Final = await _redis_client.rpush(key, *values)
## LOGGING ##
end_time = time.time()
_duration = end_time - start_time
@ -1565,7 +1566,7 @@ class RedisCache(BaseCache):
call_type=f"async_rpush <- {_get_call_stack_info()}",
)
)
verbose_logger.error(f"LiteLLM Redis Cache RPUSH: - Got exception from REDIS : {e}")
verbose_logger.error("LiteLLM Redis Cache RPUSH: - Got exception from REDIS : %s", e)
raise e
async def _pipeline_rpush_helper(
@ -1577,7 +1578,7 @@ class RedisCache(BaseCache):
for rpush_op in rpush_list:
key = self.check_and_fix_namespace(key=rpush_op["key"])
pipe.rpush(key, *rpush_op["values"])
results = await pipe.execute()
results: Final = await pipe.execute()
# Preserve positional correspondence — raise on per-command errors
for r in results:
if isinstance(r, Exception):
@ -1603,12 +1604,12 @@ class RedisCache(BaseCache):
if len(rpush_list) == 0:
return []
_redis_client: Any = self.init_async_client()
start_time = time.time()
_redis_client: Final[Any] = self.init_async_client()
start_time: Final = time.time()
try:
async with _redis_client.pipeline(transaction=False) as pipe:
results = await self._pipeline_rpush_helper(pipe, rpush_list)
results: Final = await self._pipeline_rpush_helper(pipe, rpush_list)
## LOGGING ##
end_time = time.time()
@ -1640,7 +1641,7 @@ class RedisCache(BaseCache):
raise e
async def handle_lpop_count_for_older_redis_versions(self, pipe: pipeline, key: str, count: int) -> list[bytes]:
result: list[bytes] = []
result: Final[list[bytes]] = []
for _ in range(count):
pipe.lpop(key)
results = await pipe.execute()
@ -1660,12 +1661,12 @@ class RedisCache(BaseCache):
parent_otel_span: Span | None = None,
**kwargs,
) -> Any | list[Any]:
_redis_client: Any = self.init_async_client()
_redis_client: Final[Any] = self.init_async_client()
key = self.check_and_fix_namespace(key=key)
start_time = time.time()
start_time: Final = time.time()
print_verbose(f"LPOP from Redis list: key: {key}, count: {count}")
try:
major_version = self._parse_redis_major_version()
major_version: Final = self._parse_redis_major_version()
if count is not None and major_version < 7:
# For Redis < 7.0, use pipeline to execute multiple LPOP commands
@ -1711,7 +1712,7 @@ class RedisCache(BaseCache):
call_type=f"async_lpop <- {_get_call_stack_info()}",
)
)
verbose_logger.error(f"LiteLLM Redis Cache LPOP: - Got exception from REDIS : {e}")
verbose_logger.error("LiteLLM Redis Cache LPOP: - Got exception from REDIS : %s", e)
raise e
async def _pipeline_lpop_helper(
@ -1724,7 +1725,7 @@ class RedisCache(BaseCache):
For Redis >= 7, queues one LPOP(key, count) per operation.
For Redis < 7, queues `count` individual LPOP(key) commands per operation.
"""
major_version = self._parse_redis_major_version()
major_version: Final = self._parse_redis_major_version()
if major_version >= 7:
for lpop_op in lpop_list:
@ -1734,14 +1735,14 @@ class RedisCache(BaseCache):
else:
# For Redis < 7, LPOP doesn't support count param.
# Issue `count` individual LPOP commands per key, all in one pipeline.
counts: list[int] = []
counts: Final[list[int]] = []
for lpop_op in lpop_list:
key = self.check_and_fix_namespace(key=lpop_op["key"])
count = lpop_op["count"] or 1
counts.append(count)
for _ in range(count):
pipe.lpop(key)
flat_results = await pipe.execute()
flat_results: Final = await pipe.execute()
# Re-group the flat results back into per-key lists
raw_results = []
@ -1757,7 +1758,7 @@ class RedisCache(BaseCache):
raise r
# Decode bytes -> str for each result set
decoded_results: list[list[str] | None] = []
decoded_results: Final[list[list[str] | None]] = []
for r in raw_results:
if r is None:
decoded_results.append(None)
@ -1768,7 +1769,7 @@ class RedisCache(BaseCache):
or None
)
except Exception:
decoded_results.append(r) # type: ignore
decoded_results.append(r)
else:
decoded_results.append(None)
return decoded_results
@ -1792,12 +1793,12 @@ class RedisCache(BaseCache):
if len(lpop_list) == 0:
return []
_redis_client: Any = self.init_async_client()
start_time = time.time()
_redis_client: Final[Any] = self.init_async_client()
start_time: Final = time.time()
try:
async with _redis_client.pipeline(transaction=False) as pipe:
results = await self._pipeline_lpop_helper(pipe, lpop_list)
results: Final = await self._pipeline_lpop_helper(pipe, lpop_list)
## LOGGING ##
end_time = time.time()

26
litellm/caching/redis_cluster_cache.py
View file

@ -5,7 +5,7 @@ Key differences:
- RedisClient NEEDs to be re-used across requests, adds 3000ms latency if it's re-created
"""
from typing import TYPE_CHECKING, Any, Union
from typing import TYPE_CHECKING, Any, Final, Union
from litellm.caching.redis_cache import RedisCache
@ -37,7 +37,7 @@ class RedisClusterCache(RedisCache):
if self.redis_async_redis_cluster_client:
return self.redis_async_redis_cluster_client
_redis_client = get_redis_async_client(connection_pool=self.async_redis_conn_pool, **self.redis_kwargs)
_redis_client: Final = get_redis_async_client(connection_pool=self.async_redis_conn_pool, **self.redis_kwargs)
if isinstance(_redis_client, RedisCluster):
self.redis_async_redis_cluster_client = _redis_client
@ -47,14 +47,14 @@ class RedisClusterCache(RedisCache):
"""
Overrides `_run_redis_mget_operation` in redis_cache.py
"""
return self.redis_client.mget_nonatomic(keys=keys) # type: ignore
return self.redis_client.mget_nonatomic(keys=keys)
async def _async_run_redis_mget_operation(self, keys: list[str]) -> list[Any]:
"""
Overrides `_async_run_redis_mget_operation` in redis_cache.py
"""
async_redis_cluster_client = self.init_async_client()
return await async_redis_cluster_client.mget_nonatomic(keys=keys) # type: ignore
async_redis_cluster_client: Final = self.init_async_client()
return await async_redis_cluster_client.mget_nonatomic(keys=keys)
async def test_connection(self) -> dict:
"""
@ -68,24 +68,24 @@ class RedisClusterCache(RedisCache):
from redis.cluster import ClusterNode
# Create ClusterNode objects from startup_nodes
cluster_kwargs = self.redis_kwargs.copy()
startup_nodes = cluster_kwargs.pop("startup_nodes", [])
cluster_kwargs: Final = self.redis_kwargs.copy()
startup_nodes: Final = cluster_kwargs.pop("startup_nodes", [])
new_startup_nodes: list[ClusterNode] = []
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 = redis_async.RedisCluster(
redis_client: Final = redis_async.RedisCluster(
startup_nodes=new_startup_nodes,
**cluster_kwargs, # type: ignore
**cluster_kwargs,
)
# Test the connection
ping_result = await redis_client.ping() # type: ignore[attr-defined, misc]
ping_result: Final = await redis_client.ping()
# Close the connection
await redis_client.aclose() # type: ignore[attr-defined]
await redis_client.aclose()
if ping_result:
return {
@ -100,7 +100,7 @@ class RedisClusterCache(RedisCache):
except Exception as e:
from litellm._logging import verbose_logger
verbose_logger.error(f"Redis Cluster connection test failed: {e}")
verbose_logger.error("Redis Cluster connection test failed: %s", e)
return {
"status": "failed",
"message": f"Redis Cluster connection failed: {e}",

92
litellm/caching/redis_semantic_cache.py
View file

@ -13,7 +13,7 @@ import ast
import asyncio
import json
import os
from typing import Any, cast
from typing import Any, Final, cast
import litellm
from litellm._logging import print_verbose, verbose_logger
@ -95,7 +95,7 @@ class RedisSemanticCache(BaseCache):
password = password or os.environ["REDIS_PASSWORD"]
except KeyError as e:
# Raise a more informative exception if any of the required keys are missing
missing_var = e.args[0]
missing_var: Final = e.args[0]
raise ValueError(
f"Missing required Redis configuration: {missing_var}. Provide {missing_var} or redis_url."
) from e
@ -126,11 +126,11 @@ class RedisSemanticCache(BaseCache):
# CustomTextVectorizer probes its embedding dimension at construction by
# embedding "dimension test", so the first cache request issues one extra
# billable embedding on top of the request's own.
from redisvl.extensions.llmcache import SemanticCache # type: ignore[import-not-found, import-untyped]
from redisvl.utils.vectorize import CustomTextVectorizer # type: ignore[import-not-found, import-untyped]
from redisvl.extensions.llmcache import SemanticCache
from redisvl.utils.vectorize import CustomTextVectorizer
try:
cache_vectorizer = CustomTextVectorizer(self._get_embedding)
cache_vectorizer: Final = CustomTextVectorizer(self._get_embedding)
return self._init_semantic_cache(
semantic_cache_cls=SemanticCache,
index_name=self._index_name,
@ -138,7 +138,7 @@ class RedisSemanticCache(BaseCache):
cache_vectorizer=cache_vectorizer,
)
except Exception as e:
verbose_logger.error(f"Redis semantic-cache index build failed: {e}")
verbose_logger.error("Redis semantic-cache index build failed: %s", e)
raise
@classmethod
@ -156,7 +156,7 @@ class RedisSemanticCache(BaseCache):
cache_vectorizer: Any,
) -> Any:
def _is_schema_mismatch(exc: ValueError) -> bool:
error_message = str(exc).lower()
error_message: Final = str(exc).lower()
return any(phrase in error_message for phrase in ("schema does not match", "index schema"))
try:
@ -172,7 +172,7 @@ class RedisSemanticCache(BaseCache):
if not _is_schema_mismatch(exc):
raise
isolated_index_name = f"{index_name}_isolated"
isolated_index_name: Final = f"{index_name}_isolated"
print_verbose(
"Redis semantic-cache existing index schema is not isolated; "
f"using isolated index - {isolated_index_name}"
@ -207,7 +207,7 @@ class RedisSemanticCache(BaseCache):
return {self.CACHE_KEY_FIELD_NAME: str(key)}
def _get_cache_key_filter_expression(self, key: str) -> Any:
from redisvl.query.filter import Tag # type: ignore[import-not-found, import-untyped]
from redisvl.query.filter import Tag
return Tag(self.CACHE_KEY_FIELD_NAME) == str(key)
@ -239,16 +239,16 @@ class RedisSemanticCache(BaseCache):
"""
Extract a semantic-cache prompt from chat or Responses API request kwargs.
"""
messages = kwargs.get("messages")
messages: Final = kwargs.get("messages")
if messages:
return get_str_from_messages(messages)
if "input" not in kwargs:
return None
prompt_parts: list[str] = []
prompt_parts: Final[list[str]] = []
cls._collect_responses_input_text(kwargs.get("input"), prompt_parts)
prompt = "\n".join(prompt_parts).strip()
prompt: Final = "\n".join(prompt_parts).strip()
return prompt or None
@classmethod
@ -258,7 +258,7 @@ class RedisSemanticCache(BaseCache):
return
if isinstance(value, str):
stripped_value = value.strip()
stripped_value: Final = value.strip()
if stripped_value:
prompt_parts.append(stripped_value)
return
@ -298,10 +298,10 @@ class RedisSemanticCache(BaseCache):
@staticmethod
def _coerce_response_input_value(value: Any) -> Any:
model_dump = getattr(value, "model_dump", None)
model_dump: Final = getattr(value, "model_dump", None)
if callable(model_dump):
return model_dump()
dict_method = getattr(value, "dict", None)
dict_method: Final = getattr(value, "dict", None)
if callable(dict_method):
return dict_method()
return value
@ -318,7 +318,7 @@ class RedisSemanticCache(BaseCache):
llm_model_list = None
llm_router = None
router = resolve_embedding_router(self.embedding_model, llm_router, llm_model_list)
router: Final = resolve_embedding_router(self.embedding_model, llm_router, llm_model_list)
if router is not None:
embedding_response = cast(
EmbeddingResponse,
@ -383,22 +383,22 @@ class RedisSemanticCache(BaseCache):
value_str: str | None = None
try:
prompt = self._get_prompt_from_kwargs(**kwargs)
prompt: Final = self._get_prompt_from_kwargs(**kwargs)
if prompt is None:
print_verbose("No prompt provided for semantic caching")
return
value_str = str(value)
prompt_embedding = self._get_embedding(prompt, metadata=kwargs.get("metadata"))
prompt_embedding: Final = self._get_embedding(prompt, metadata=kwargs.get("metadata"))
store_kwargs: dict[str, Any] = {
store_kwargs: Final[dict[str, Any]] = {
"vector": prompt_embedding,
"filters": self._get_cache_filters(key),
}
# Get TTL and store in Redis semantic cache
ttl = self._get_ttl(**kwargs)
ttl: Final = self._get_ttl(**kwargs)
if ttl is not None:
store_kwargs["ttl"] = int(ttl)
self.llmcache.store(prompt, value_str, **store_kwargs)
@ -419,7 +419,7 @@ class RedisSemanticCache(BaseCache):
print_verbose(f"Redis semantic-cache get_cache, kwargs: {kwargs}")
try:
prompt = self._get_prompt_from_kwargs(**kwargs)
prompt: Final = self._get_prompt_from_kwargs(**kwargs)
if prompt is None:
print_verbose("No prompt provided for semantic cache lookup")
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
@ -427,13 +427,13 @@ class RedisSemanticCache(BaseCache):
# Check the cache for semantically similar prompts in this exact
# LiteLLM cache-key scope.
prompt_embedding = self._get_embedding(prompt, metadata=kwargs.get("metadata"))
check_kwargs: dict[str, Any] = {
prompt_embedding: Final = self._get_embedding(prompt, metadata=kwargs.get("metadata"))
check_kwargs: Final[dict[str, Any]] = {
"prompt": prompt,
"vector": prompt_embedding,
"filter_expression": self._get_cache_key_filter_expression(key),
}
results = self.llmcache.check(**check_kwargs)
results: Final = self.llmcache.check(**check_kwargs)
# Return None if no similar prompts found
if not results:
@ -441,20 +441,20 @@ class RedisSemanticCache(BaseCache):
return None
# Process the best matching result
cache_hit = results[0]
cache_hit: Final = results[0]
if not self._cache_hit_matches_key(cache_hit=cache_hit, key=key):
print_verbose("Redis semantic-cache hit did not match cache key scope")
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
return None
vector_distance = float(cache_hit["vector_distance"])
vector_distance: Final = float(cache_hit["vector_distance"])
# Convert vector distance back to similarity score
# For cosine distance: 0 = most similar, 2 = least similar
# While similarity: 1 = most similar, 0 = least similar
similarity = 1 - vector_distance
similarity: Final = 1 - vector_distance
cached_prompt = cache_hit["prompt"]
cached_response = cache_hit["response"]
cached_prompt: Final = cache_hit["prompt"]
cached_response: Final = cache_hit["response"]
# update kwargs["metadata"] with similarity, don't rewrite the original metadata
kwargs.setdefault("metadata", {})["semantic-similarity"] = similarity
@ -488,7 +488,7 @@ class RedisSemanticCache(BaseCache):
llm_model_list = None
llm_router = None
router = resolve_embedding_router(self.embedding_model, llm_router, llm_model_list)
router: Final = resolve_embedding_router(self.embedding_model, llm_router, llm_model_list)
try:
if router is not None:
embedding_response = await router.aembedding(
@ -521,23 +521,23 @@ class RedisSemanticCache(BaseCache):
print_verbose(f"Async Redis semantic-cache set_cache, kwargs: {kwargs}")
try:
prompt = self._get_prompt_from_kwargs(**kwargs)
prompt: Final = self._get_prompt_from_kwargs(**kwargs)
if prompt is None:
print_verbose("No prompt provided for semantic caching")
return
value_str = str(value)
value_str: Final = str(value)
# Generate embedding for the value (response) to cache
prompt_embedding = await self._get_async_embedding(prompt, metadata=kwargs.get("metadata"))
prompt_embedding: Final = await self._get_async_embedding(prompt, metadata=kwargs.get("metadata"))
store_kwargs: dict[str, Any] = {
store_kwargs: Final[dict[str, Any]] = {
"vector": prompt_embedding,
"filters": self._get_cache_filters(key),
}
# Get TTL and store in Redis semantic cache
ttl = self._get_ttl(**kwargs)
ttl: Final = self._get_ttl(**kwargs)
if ttl is not None:
store_kwargs["ttl"] = ttl
await self.llmcache.astore(
@ -562,43 +562,43 @@ class RedisSemanticCache(BaseCache):
print_verbose(f"Async Redis semantic-cache get_cache, kwargs: {kwargs}")
try:
prompt = self._get_prompt_from_kwargs(**kwargs)
prompt: Final = self._get_prompt_from_kwargs(**kwargs)
if prompt is None:
print_verbose("No prompt provided for semantic cache lookup")
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
return None
# Generate embedding for the prompt
prompt_embedding = await self._get_async_embedding(prompt, metadata=kwargs.get("metadata"))
prompt_embedding: Final = await self._get_async_embedding(prompt, metadata=kwargs.get("metadata"))
# Check the cache for semantically similar prompts in this exact
# LiteLLM cache-key scope.
check_kwargs: dict[str, Any] = {
check_kwargs: Final[dict[str, Any]] = {
"prompt": prompt,
"vector": prompt_embedding,
"filter_expression": self._get_cache_key_filter_expression(key),
}
results = await self.llmcache.acheck(**check_kwargs)
results: Final = await self.llmcache.acheck(**check_kwargs)
# handle results / cache hit
if not results:
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
return None
cache_hit = results[0]
cache_hit: Final = results[0]
if not self._cache_hit_matches_key(cache_hit=cache_hit, key=key):
print_verbose("Redis semantic-cache hit did not match cache key scope")
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
return None
vector_distance = float(cache_hit["vector_distance"])
vector_distance: Final = float(cache_hit["vector_distance"])
# Convert vector distance back to similarity
# For cosine distance: 0 = most similar, 2 = least similar
# While similarity: 1 = most similar, 0 = least similar
similarity = 1 - vector_distance
similarity: Final = 1 - vector_distance
cached_prompt = cache_hit["prompt"]
cached_response = cache_hit["response"]
cached_prompt: Final = cache_hit["prompt"]
cached_response: Final = cache_hit["response"]
# update kwargs["metadata"] with similarity, don't rewrite the original metadata
kwargs.setdefault("metadata", {})["semantic-similarity"] = similarity
@ -622,7 +622,7 @@ class RedisSemanticCache(BaseCache):
Returns:
Dict[str, Any]: Information about the Redis index
"""
aindex = await self.llmcache._get_async_index()
aindex: Final = await self.llmcache._get_async_index()
return await aindex.info()
async def async_set_cache_pipeline(self, cache_list: list[tuple[str, Any]], **kwargs) -> None:
@ -634,7 +634,7 @@ class RedisSemanticCache(BaseCache):
**kwargs: Additional arguments
"""
try:
tasks = []
tasks: Final = []
for val in cache_list:
tasks.append(self.async_set_cache(val[0], val[1], **kwargs))
await asyncio.gather(*tasks)

42
litellm/caching/s3_cache.py
View file

@ -13,6 +13,7 @@ import asyncio
import json
from datetime import datetime, timedelta, timezone
from functools import partial
from typing import Final
from litellm._logging import print_verbose, verbose_logger
@ -62,16 +63,16 @@ class S3Cache(BaseCache):
def set_cache(self, key, value, **kwargs):
try:
print_verbose(f"LiteLLM SET Cache - S3. Key={key}. Value={value}")
ttl = kwargs.get("ttl", None)
ttl: Final = kwargs.get("ttl", None)
# Convert value to JSON before storing in S3
serialized_value = json.dumps(value)
serialized_value: Final = json.dumps(value)
key = self._to_s3_key(key)
if ttl is not None:
cache_control = f"immutable, max-age={ttl}, s-maxage={ttl}"
# Calculate expiration time
expiration_time = datetime.now(timezone.utc) + timedelta(seconds=ttl)
expiration_time: Final = datetime.now(timezone.utc) + timedelta(seconds=ttl)
# Upload the data to S3 with the calculated expiration time
self.s3_client.put_object(
Bucket=self.bucket_name,
@ -104,12 +105,12 @@ class S3Cache(BaseCache):
Compatible with Python 3.8+.
"""
try:
verbose_logger.debug(f"Set ASYNC S3 Cache: Key={key}. Value={value}")
loop = asyncio.get_event_loop()
func = partial(self.set_cache, key, value, **kwargs)
verbose_logger.debug("Set ASYNC S3 Cache: Key=%s. Value=%s", key, value)
loop: Final = asyncio.get_event_loop()
func: Final = partial(self.set_cache, key, value, **kwargs)
await loop.run_in_executor(None, func)
except Exception as e:
verbose_logger.error(f"S3 Caching: async_set_cache() - Got exception from S3: {e}")
verbose_logger.error("S3 Caching: async_set_cache() - Got exception from S3: %s", e)
def get_cache(self, key, **kwargs):
import botocore
@ -123,8 +124,8 @@ class S3Cache(BaseCache):
if cached_response is not None:
if "Expires" in cached_response:
expires_time = cached_response["Expires"]
current_time = datetime.now(expires_time.tzinfo)
expires_time: Final = cached_response["Expires"]
current_time: Final = datetime.now(expires_time.tzinfo)
if current_time > expires_time:
return None
@ -138,17 +139,20 @@ class S3Cache(BaseCache):
if not isinstance(cached_response, dict):
cached_response = dict(cached_response)
verbose_logger.debug(
f"Got S3 Cache: key: {key}, cached_response {cached_response}. Type Response {type(cached_response)}"
"Got S3 Cache: key: %s, cached_response %s. Type Response %s",
key,
cached_response,
type(cached_response),
)
return cached_response
except botocore.exceptions.ClientError as e: # type: ignore
except botocore.exceptions.ClientError as e:
if e.response["Error"]["Code"] == "NoSuchKey":
verbose_logger.debug(f"S3 Cache: The specified key '{key}' does not exist in the S3 bucket.")
verbose_logger.debug("S3 Cache: The specified key '%s' does not exist in the S3 bucket.", key)
return None
except Exception as e:
verbose_logger.error(f"S3 Caching: get_cache() - Got exception from S3: {e}")
verbose_logger.error("S3 Caching: get_cache() - Got exception from S3: %s", e)
async def async_get_cache(self, key, **kwargs):
"""
@ -156,13 +160,13 @@ class S3Cache(BaseCache):
Compatible with Python 3.8+.
"""
try:
verbose_logger.debug(f"Get ASYNC S3 Cache: key: {key}")
loop = asyncio.get_event_loop()
func = partial(self.get_cache, key, **kwargs)
result = await loop.run_in_executor(None, func)
verbose_logger.debug("Get ASYNC S3 Cache: key: %s", key)
loop: Final = asyncio.get_event_loop()
func: Final = partial(self.get_cache, key, **kwargs)
result: Final = await loop.run_in_executor(None, func)
return result
except Exception as e:
verbose_logger.error(f"S3 Caching: async_get_cache() - Got exception from S3: {e}")
verbose_logger.error("S3 Caching: async_get_cache() - Got exception from S3: %s", e)
return None
def flush_cache(self):
@ -172,7 +176,7 @@ class S3Cache(BaseCache):
pass
async def async_set_cache_pipeline(self, cache_list, **kwargs):
tasks = []
tasks: Final = []
for val in cache_list:
tasks.append(self.async_set_cache(val[0], val[1], **kwargs))
await asyncio.gather(*tasks)

56
litellm/caching/valkey_semantic_cache.py
View file

@ -19,7 +19,7 @@ import hashlib
import os
import struct
from dataclasses import dataclass
from typing import Any
from typing import Any, Final
from redis import Redis
from redis.asyncio import Redis as AsyncRedis
@ -85,12 +85,8 @@ class ValkeySemanticCache(RedisSemanticCache):
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) # type: ignore[arg-type]
)
self.async_client = (
async_client if async_client is not None else AsyncRedis.from_url(resolved_url) # type: ignore[arg-type]
)
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}")
@ -106,8 +102,8 @@ class ValkeySemanticCache(RedisSemanticCache):
"(or VALKEY_HOST/VALKEY_PORT), or pass redis_url."
)
credentials = f":{password}@" if password else ""
scheme = "rediss" if ssl else "redis"
credentials: Final = f":{password}@" if password else ""
scheme: Final = "rediss" if ssl else "redis"
return f"{scheme}://{credentials}{host}:{port}"
@classmethod
@ -154,7 +150,7 @@ class ValkeySemanticCache(RedisSemanticCache):
return None
def _assert_dim_matches(self, info: dict, dim: int) -> None:
existing_dim = self._extract_index_dim(info)
existing_dim: Final = self._extract_index_dim(info)
if existing_dim is not None and existing_dim != dim:
raise ValueError(
f"Valkey semantic-cache index '{self.index_name}' already exists with "
@ -186,7 +182,7 @@ class ValkeySemanticCache(RedisSemanticCache):
except Exception as exc:
if not self._is_index_exists_error(exc):
raise
info = await self.async_client.ft(self.index_name).info()
info: Final = await self.async_client.ft(self.index_name).info()
self._assert_dim_matches(info, dim)
self._index_dim = dim
@ -202,8 +198,8 @@ class ValkeySemanticCache(RedisSemanticCache):
}
def _knn_query(self, key: str) -> Query:
scope = self._scope_tag(key)
query_string = (
scope: Final = self._scope_tag(key)
query_string: Final = (
f"(@{self.CACHE_KEY_FIELD_NAME}:{{{scope}}})"
f"=>[KNN 1 @{self.EMBEDDING_FIELD_NAME} $vec AS {self.DISTANCE_FIELD_NAME}]"
)
@ -211,10 +207,10 @@ class ValkeySemanticCache(RedisSemanticCache):
@classmethod
def _first_hit(cls, search_result: Any) -> _ValkeyCacheHit | None:
docs = getattr(search_result, "docs", [])
docs: Final = getattr(search_result, "docs", [])
if not docs:
return None
doc = docs[0]
doc: Final = docs[0]
return _ValkeyCacheHit(
response=str(getattr(doc, cls.RESPONSE_FIELD_NAME)),
distance=float(getattr(doc, cls.DISTANCE_FIELD_NAME)),
@ -225,7 +221,7 @@ class ValkeySemanticCache(RedisSemanticCache):
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
return None
similarity = 1 - hit.distance
similarity: Final = 1 - hit.distance
kwargs.setdefault("metadata", {})["semantic-similarity"] = similarity
if similarity < self.similarity_threshold:
@ -235,17 +231,17 @@ class ValkeySemanticCache(RedisSemanticCache):
def set_cache(self, key: str, value: Any, **kwargs: Any) -> None:
print_verbose(f"Valkey semantic-cache set_cache, kwargs: {kwargs}")
try:
prompt = self._get_prompt_from_kwargs(**kwargs)
prompt: Final = self._get_prompt_from_kwargs(**kwargs)
if prompt is None:
print_verbose("No prompt provided for semantic caching")
return
embedding = self._get_embedding(prompt)
embedding: Final = self._get_embedding(prompt)
self._ensure_index_sync(len(embedding))
doc_key = self._doc_key(key)
doc_key: Final = self._doc_key(key)
self.sync_client.hset(doc_key, mapping=self._doc_mapping(key, prompt, str(value), embedding))
ttl = self._get_ttl(**kwargs)
ttl: Final = self._get_ttl(**kwargs)
if ttl is not None:
self.sync_client.expire(doc_key, ttl)
except Exception as e:
@ -254,15 +250,15 @@ class ValkeySemanticCache(RedisSemanticCache):
def get_cache(self, key: str, **kwargs: Any) -> Any:
print_verbose(f"Valkey semantic-cache get_cache, kwargs: {kwargs}")
try:
prompt = self._get_prompt_from_kwargs(**kwargs)
prompt: Final = self._get_prompt_from_kwargs(**kwargs)
if prompt is None:
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
return None
embedding = self._get_embedding(prompt)
embedding: Final = self._get_embedding(prompt)
self._ensure_index_sync(len(embedding))
search_result = self.sync_client.ft(self.index_name).search(
search_result: Final = self.sync_client.ft(self.index_name).search(
self._knn_query(key),
query_params={"vec": self._embedding_to_bytes(embedding)},
)
@ -274,17 +270,17 @@ class ValkeySemanticCache(RedisSemanticCache):
async def async_set_cache(self, key: str, value: Any, **kwargs: Any) -> None:
print_verbose(f"Async Valkey semantic-cache set_cache, kwargs: {kwargs}")
try:
prompt = self._get_prompt_from_kwargs(**kwargs)
prompt: Final = self._get_prompt_from_kwargs(**kwargs)
if prompt is None:
print_verbose("No prompt provided for semantic caching")
return
embedding = await self._get_async_embedding(prompt, metadata=kwargs.get("metadata"))
embedding: Final = await self._get_async_embedding(prompt, metadata=kwargs.get("metadata"))
await self._ensure_index_async(len(embedding))
doc_key = self._doc_key(key)
doc_key: Final = self._doc_key(key)
await self.async_client.hset(doc_key, mapping=self._doc_mapping(key, prompt, str(value), embedding))
ttl = self._get_ttl(**kwargs)
ttl: Final = self._get_ttl(**kwargs)
if ttl is not None:
await self.async_client.expire(doc_key, ttl)
except Exception as e:
@ -293,15 +289,15 @@ class ValkeySemanticCache(RedisSemanticCache):
async def async_get_cache(self, key: str, **kwargs: Any) -> Any:
print_verbose(f"Async Valkey semantic-cache get_cache, kwargs: {kwargs}")
try:
prompt = self._get_prompt_from_kwargs(**kwargs)
prompt: Final = self._get_prompt_from_kwargs(**kwargs)
if prompt is None:
kwargs.setdefault("metadata", {})["semantic-similarity"] = 0.0
return None
embedding = await self._get_async_embedding(prompt, metadata=kwargs.get("metadata"))
embedding: Final = await self._get_async_embedding(prompt, metadata=kwargs.get("metadata"))
await self._ensure_index_async(len(embedding))
search_result = await self.async_client.ft(self.index_name).search(
search_result: Final = await self.async_client.ft(self.index_name).search(
self._knn_query(key),
query_params={"vec": self._embedding_to_bytes(embedding)},
)

102
litellm/completion_extras/litellm_responses_transformation/handler.py
View file

@ -3,7 +3,7 @@ Handler for transforming /chat/completions api requests to litellm.responses req
"""
from collections.abc import Coroutine
from typing import TYPE_CHECKING, Any, Union
from typing import TYPE_CHECKING, Any, Final, Union
from typing_extensions import TypedDict
@ -60,7 +60,7 @@ class ResponsesToCompletionBridgeHandler:
raise ValueError("Unexpected responses stream payload")
if hidden_params:
existing = getattr(response, "_hidden_params", None)
existing: Final = getattr(response, "_hidden_params", None)
if not isinstance(existing, dict) or not existing:
setattr(response, "_hidden_params", dict(hidden_params))
else:
@ -72,13 +72,13 @@ class ResponsesToCompletionBridgeHandler:
for _ in stream_iter:
pass
completed = getattr(stream_iter, "completed_response", None)
response_obj = getattr(completed, "response", None) if completed else None
completed: Final = getattr(stream_iter, "completed_response", None)
response_obj: Final = getattr(completed, "response", None) if completed else None
if response_obj is None:
raise ValueError("Stream ended without a completed response")
hidden_params = getattr(stream_iter, "_hidden_params", None)
response = self._coerce_response_object(response_obj, hidden_params)
hidden_params: Final = getattr(stream_iter, "_hidden_params", None)
response: Final = self._coerce_response_object(response_obj, hidden_params)
if not isinstance(response, ResponsesAPIResponse):
raise ValueError("Stream completed response is invalid")
return response
@ -87,13 +87,13 @@ class ResponsesToCompletionBridgeHandler:
async for _ in stream_iter:
pass
completed = getattr(stream_iter, "completed_response", None)
response_obj = getattr(completed, "response", None) if completed else None
completed: Final = getattr(stream_iter, "completed_response", None)
response_obj: Final = getattr(completed, "response", None) if completed else None
if response_obj is None:
raise ValueError("Stream ended without a completed response")
hidden_params = getattr(stream_iter, "_hidden_params", None)
response = self._coerce_response_object(response_obj, hidden_params)
hidden_params: Final = getattr(stream_iter, "_hidden_params", None)
response: Final = self._coerce_response_object(response_obj, hidden_params)
if not isinstance(response, ResponsesAPIResponse):
raise ValueError("Stream completed response is invalid")
return response
@ -102,35 +102,35 @@ class ResponsesToCompletionBridgeHandler:
from litellm import LiteLLMLoggingObj
from litellm.types.utils import ModelResponse
model = kwargs.get("model")
model: Final = kwargs.get("model")
if model is None or not isinstance(model, str):
raise ValueError("model is required")
custom_llm_provider = kwargs.get("custom_llm_provider")
custom_llm_provider: Final = kwargs.get("custom_llm_provider")
if custom_llm_provider is None or not isinstance(custom_llm_provider, str):
raise ValueError("custom_llm_provider is required")
messages = kwargs.get("messages")
messages: Final = kwargs.get("messages")
if messages is None or not isinstance(messages, list):
raise ValueError("messages is required")
optional_params = kwargs.get("optional_params")
optional_params: Final = kwargs.get("optional_params")
if optional_params is None or not isinstance(optional_params, dict):
raise ValueError("optional_params is required")
litellm_params = kwargs.get("litellm_params")
litellm_params: Final = kwargs.get("litellm_params")
if litellm_params is None or not isinstance(litellm_params, dict):
raise ValueError("litellm_params is required")
headers = kwargs.get("headers")
headers: Final = kwargs.get("headers")
if headers is None or not isinstance(headers, dict):
raise ValueError("headers is required")
model_response = kwargs.get("model_response")
model_response: Final = kwargs.get("model_response")
if model_response is None or not isinstance(model_response, ModelResponse):
raise ValueError("model_response is required")
logging_obj = kwargs.get("logging_obj")
logging_obj: Final = kwargs.get("logging_obj")
if logging_obj is None or not isinstance(logging_obj, LiteLLMLoggingObj):
raise ValueError("logging_obj is required")
@ -158,19 +158,19 @@ class ResponsesToCompletionBridgeHandler:
from litellm import responses
from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper
validated_kwargs = self.validate_input_kwargs(kwargs)
model = validated_kwargs["model"]
messages = validated_kwargs["messages"]
validated_kwargs: Final = self.validate_input_kwargs(kwargs)
model: Final = validated_kwargs["model"]
messages: Final = validated_kwargs["messages"]
optional_params = validated_kwargs["optional_params"]
litellm_params = validated_kwargs["litellm_params"]
headers = validated_kwargs["headers"]
model_response = validated_kwargs["model_response"]
logging_obj = validated_kwargs["logging_obj"]
custom_llm_provider = validated_kwargs["custom_llm_provider"]
litellm_params: Final = validated_kwargs["litellm_params"]
headers: Final = validated_kwargs["headers"]
model_response: Final = validated_kwargs["model_response"]
logging_obj: Final = validated_kwargs["logging_obj"]
custom_llm_provider: Final = validated_kwargs["custom_llm_provider"]
if kwargs.get("stream") is True and "stream" not in optional_params:
optional_params = {**optional_params, "stream": True}
request_data = self.transformation_handler.transform_request(
request_data: Final = self.transformation_handler.transform_request(
model=model,
messages=messages,
optional_params=optional_params,
@ -188,13 +188,13 @@ class ResponsesToCompletionBridgeHandler:
# than adding an explicit kwarg) avoids the duplicate-keyword
# TypeError that would otherwise fire on the real bridge path.
request_data["custom_llm_provider"] = custom_llm_provider
result = responses(
result: Final = responses(
**request_data,
)
from litellm.types.utils import ModelResponse
stream = self._resolve_stream_flag(optional_params, litellm_params)
stream: Final = self._resolve_stream_flag(optional_params, litellm_params)
if isinstance(result, ResponsesAPIResponse):
return self.transformation_handler.transform_response(
model=model,
@ -220,7 +220,7 @@ class ResponsesToCompletionBridgeHandler:
json_mode=kwargs.get("json_mode"),
)
elif not stream:
responses_api_response = self._collect_response_from_stream(result)
responses_api_response: Final = self._collect_response_from_stream(result)
return self.transformation_handler.transform_response(
model=model,
raw_response=responses_api_response,
@ -237,12 +237,12 @@ class ResponsesToCompletionBridgeHandler:
else:
if self._is_preformatted_cached_chat_stream(result):
return self._apply_post_stream_processing(result, model, custom_llm_provider)
completion_stream = self.transformation_handler.get_model_response_iterator(
streaming_response=result, # type: ignore
completion_stream: Final = self.transformation_handler.get_model_response_iterator(
streaming_response=result,
sync_stream=True,
json_mode=kwargs.get("json_mode"),
)
streamwrapper = CustomStreamWrapper(
streamwrapper: Final = CustomStreamWrapper(
completion_stream=completion_stream,
model=model,
custom_llm_provider=custom_llm_provider,
@ -254,20 +254,20 @@ class ResponsesToCompletionBridgeHandler:
from litellm import aresponses
from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper
validated_kwargs = self.validate_input_kwargs(kwargs)
model = validated_kwargs["model"]
messages = validated_kwargs["messages"]
validated_kwargs: Final = self.validate_input_kwargs(kwargs)
model: Final = validated_kwargs["model"]
messages: Final = validated_kwargs["messages"]
optional_params = validated_kwargs["optional_params"]
litellm_params = validated_kwargs["litellm_params"]
headers = validated_kwargs["headers"]
model_response = validated_kwargs["model_response"]
logging_obj = validated_kwargs["logging_obj"]
custom_llm_provider = validated_kwargs["custom_llm_provider"]
litellm_params: Final = validated_kwargs["litellm_params"]
headers: Final = validated_kwargs["headers"]
model_response: Final = validated_kwargs["model_response"]
logging_obj: Final = validated_kwargs["logging_obj"]
custom_llm_provider: Final = validated_kwargs["custom_llm_provider"]
if kwargs.get("stream") is True and "stream" not in optional_params:
optional_params = {**optional_params, "stream": True}
try:
request_data = self.transformation_handler.transform_request(
request_data: Final = self.transformation_handler.transform_request(
model=model,
messages=messages,
optional_params=optional_params,
@ -285,14 +285,14 @@ class ResponsesToCompletionBridgeHandler:
# keyword TypeError when `sanitized_litellm_params` already
# carries `custom_llm_provider`.
request_data["custom_llm_provider"] = custom_llm_provider
result = await aresponses(
result: Final = await aresponses(
**request_data,
aresponses=True,
)
from litellm.types.utils import ModelResponse
stream = self._resolve_stream_flag(optional_params, litellm_params)
stream: Final = self._resolve_stream_flag(optional_params, litellm_params)
if isinstance(result, ResponsesAPIResponse):
return self.transformation_handler.transform_response(
model=model,
@ -318,7 +318,7 @@ class ResponsesToCompletionBridgeHandler:
json_mode=kwargs.get("json_mode"),
)
elif not stream:
responses_api_response = await self._collect_response_from_stream_async(result)
responses_api_response: Final = await self._collect_response_from_stream_async(result)
return self.transformation_handler.transform_response(
model=model,
raw_response=responses_api_response,
@ -335,12 +335,12 @@ class ResponsesToCompletionBridgeHandler:
else:
if self._is_preformatted_cached_chat_stream(result):
return self._apply_post_stream_processing(result, model, custom_llm_provider)
completion_stream = self.transformation_handler.get_model_response_iterator(
streaming_response=result, # type: ignore
completion_stream: Final = self.transformation_handler.get_model_response_iterator(
streaming_response=result,
sync_stream=False,
json_mode=kwargs.get("json_mode"),
)
streamwrapper = CustomStreamWrapper(
streamwrapper: Final = CustomStreamWrapper(
completion_stream=completion_stream,
model=model,
custom_llm_provider=custom_llm_provider,
@ -359,7 +359,7 @@ class ResponsesToCompletionBridgeHandler:
from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper
from litellm.llms.base_llm.base_model_iterator import MockResponseIterator
streamwrapper = CustomStreamWrapper(
streamwrapper: Final = CustomStreamWrapper(
completion_stream=MockResponseIterator(model_response=response, json_mode=json_mode),
model=model,
custom_llm_provider=custom_llm_provider,
@ -378,7 +378,7 @@ class ResponsesToCompletionBridgeHandler:
from litellm.utils import ProviderConfigManager
try:
provider_config = ProviderConfigManager.get_provider_chat_config(
provider_config: Final = ProviderConfigManager.get_provider_chat_config(
model=model, provider=LlmProviders(custom_llm_provider)
)
except (ValueError, KeyError):
@ -389,4 +389,4 @@ class ResponsesToCompletionBridgeHandler:
return stream
responses_api_bridge = ResponsesToCompletionBridgeHandler()
responses_api_bridge: Final = ResponsesToCompletionBridgeHandler()

477
litellm/completion_extras/litellm_responses_transformation/transformation.py
View file

@ -4,15 +4,17 @@ Handler for transforming /chat/completions api requests to litellm.responses req
import json
import os
from collections.abc import AsyncIterator, Callable, Iterable, Iterator
from typing import (
TYPE_CHECKING,
Any,
Literal,
Union,
cast,
)
from collections.abc import AsyncIterator, Callable, Iterable, Iterator, Mapping
from typing import TYPE_CHECKING, Any, Final, Literal, Union, cast
from openai.types.responses.custom_tool_param import CustomToolParam
from openai.types.responses.response_input_param import (
FunctionCallOutput,
ResponseCustomToolCallOutputParam,
ResponseCustomToolCallParam,
)
from openai.types.responses.tool_choice_custom_param import ToolChoiceCustomParam
from openai.types.responses.tool_choice_function_param import ToolChoiceFunctionParam
from openai.types.responses.tool_param import FunctionToolParam
from pydantic import BaseModel
@ -32,6 +34,8 @@ from litellm.responses.utils import normalize_responses_api_stream_options
from litellm.types.llms.openai import (
ChatCompletionAnnotation,
ChatCompletionReasoningItem,
ChatCompletionToolCallChunk,
ChatCompletionToolCallFunctionChunk,
ChatCompletionToolParamFunctionChunk,
Reasoning,
ResponsesAPIOptionalRequestParams,
@ -59,9 +63,9 @@ def _get_reasoning_items(
msg: "AllMessageValues",
) -> list[ChatCompletionReasoningItem]:
"""Extract reasoning_items from a message dict with proper typing."""
items = msg.get("reasoning_items") # type: ignore[union-attr]
items: Final = msg.get("reasoning_items")
if items:
return items # type: ignore[return-value]
return items
return []
@ -74,7 +78,7 @@ def _build_reasoning_item(
Handles both pydantic objects (attribute access) and plain dicts.
"""
summary: list[dict[str, Any]] = []
summary: Final[list[dict[str, Any]]] = []
for s in summary_raw or []:
if isinstance(s, dict):
summary.append({"type": s.get("type", "summary_text"), "text": s.get("text", "")})
@ -93,11 +97,55 @@ def _build_reasoning_item(
}
class _ChatToolCallDict(ChatCompletionToolCallChunk, total=False):
provider_specific_fields: Mapping[str, Any]
def _tool_call_dict_from_output_item(item: Mapping[str, Any], index: int) -> _ChatToolCallDict:
"""Convert a ``function_call`` or ``custom_tool_call`` output item dict to a chat
completions tool_call dict. Custom (grammar/freeform) tool calls carry their raw
string payload in ``input`` rather than ``arguments``; both map to
``function.arguments`` so chat clients (e.g. Cursor agent mode) receive them like
any other tool call. The single conversion rule shared by the non-streaming
accumulator and the streaming ``output_item.added`` branch."""
from litellm.responses.litellm_completion_transformation.transformation import (
LiteLLMCompletionResponsesConfig,
)
is_custom: Final = item.get("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)
tool_call_dict: Final = _ChatToolCallDict(
id=LiteLLMCompletionResponsesConfig._tool_call_id_from_responses_item(item.get("id"), item.get("call_id")),
type="function",
function=function_chunk,
index=index,
)
raw_provider_fields: Final = 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__"):
provider_specific_fields = vars(raw_provider_fields)
else:
provider_specific_fields = None
if provider_specific_fields:
tool_call_dict["provider_specific_fields"] = provider_specific_fields
function_chunk["provider_specific_fields"] = provider_specific_fields
return tool_call_dict
def _flat_responses_tool_choice(choice_type: str, name: str) -> ToolChoiceFunctionParam | ToolChoiceCustomParam:
if choice_type == "custom":
return ToolChoiceCustomParam(type="custom", name=name)
return ToolChoiceFunctionParam(type="function", name=name)
def _reasoning_item_to_response_input(
r_item: ChatCompletionReasoningItem | dict[str, Any],
) -> dict[str, Any]:
"""Convert a stored ChatCompletionReasoningItem back to a Responses API input item."""
r_input: dict[str, Any] = {
r_input: Final[dict[str, Any]] = {
"type": "reasoning",
"id": r_item.get("id") or f"rs_{id(r_item)}",
# summary is always required by the Responses API, even when empty
@ -117,17 +165,20 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
pass
def _normalize_tool_choice_for_responses_api(self, tool_choice: Any) -> Any:
"""Chat tool_choice uses function.name; Responses API expects top-level name."""
if not isinstance(tool_choice, dict) or tool_choice.get("type") != "function":
"""Chat tool_choice nests the name under function/custom; Responses API expects top-level name."""
if not isinstance(tool_choice, dict):
return tool_choice
choice_type: Final = tool_choice.get("type")
if choice_type not in ("function", "custom"):
return tool_choice
if isinstance(tool_choice.get("name"), str) and tool_choice.get("name"):
# Return only Responses shape so stray chat ``function`` key is not sent upstream.
return {"type": "function", "name": tool_choice["name"]}
fn = tool_choice.get("function")
if isinstance(fn, dict):
fn_name = fn.get("name")
if isinstance(fn_name, str) and fn_name:
return {"type": "function", "name": fn_name}
# Return only Responses shape so stray chat ``function``/``custom`` keys are not sent upstream.
return _flat_responses_tool_choice(choice_type, tool_choice["name"])
nested: Final = tool_choice.get(choice_type)
if isinstance(nested, dict):
nested_name: Final = nested.get("name")
if isinstance(nested_name, str) and nested_name:
return _flat_responses_tool_choice(choice_type, nested_name)
return tool_choice
def _handle_raw_dict_response_item(self, item: dict[str, Any], index: int) -> tuple[Any | None, int]:
@ -143,7 +194,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
"""
from litellm.types.utils import Choices, Message
item_type = item.get("type")
item_type: Final = item.get("type")
# Ignore reasoning items for now
if item_type == "reasoning":
@ -151,7 +202,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
# Handle message items with output_text content
if item_type == "message":
content_list = item.get("content", [])
content_list: Final = item.get("content", [])
for content_item in content_list:
if isinstance(content_item, dict):
content_type = content_item.get("type")
@ -169,36 +220,8 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
choice = Choices(message=msg, finish_reason="stop", index=index)
return choice, index + 1
# Handle function_call items (e.g., from GPT-5 Codex format)
if item_type == "function_call":
# Extract provider_specific_fields if present and pass through as-is
provider_specific_fields = item.get("provider_specific_fields")
if provider_specific_fields and not isinstance(provider_specific_fields, dict):
provider_specific_fields = (
dict(provider_specific_fields) if hasattr(provider_specific_fields, "__dict__") else {}
)
tool_call_dict = {
"id": item.get("call_id") or item.get("id", ""),
"function": {
"name": item.get("name", ""),
"arguments": item.get("arguments", ""),
},
"type": "function",
}
# Pass through provider_specific_fields as-is if present
if provider_specific_fields:
tool_call_dict["provider_specific_fields"] = provider_specific_fields
# Also add to function's provider_specific_fields for consistency
tool_call_dict["function"]["provider_specific_fields"] = provider_specific_fields
msg = Message(
content=None,
tool_calls=[tool_call_dict],
)
choice = Choices(message=msg, finish_reason="tool_calls", index=index)
return choice, index + 1
# function_call / custom_tool_call dicts are intercepted and accumulated by
# _convert_response_output_to_choices before this callback is reached
# Unknown or unsupported type
return None, index
@ -206,8 +229,17 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
def convert_chat_completion_messages_to_responses_api(
self, messages: list["AllMessageValues"]
) -> tuple[list[Any], str | None]:
input_items: list[Any] = []
input_items: Final[list[Any]] = []
instructions: str | None = None
custom_tool_call_ids: Final = frozenset(
tool_call["id"]
for msg in messages
if msg.get("role") == "assistant" and isinstance(msg.get("tool_calls"), list)
for tool_call in msg.get("tool_calls") or ()
if isinstance(tool_call, dict)
and not tool_call.get("function")
and isinstance(tool_call.get("custom"), dict)
)
for msg in messages:
role = msg.get("role")
@ -229,8 +261,8 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
"type": "message",
"role": role,
"content": self._convert_content_to_responses_format(
content, # type: ignore[arg-type]
role, # type: ignore
content,
role,
),
}
)
@ -253,18 +285,28 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
else:
# Fallback: convert unexpected types to input_text
tool_output = [{"type": "input_text", "text": str(content)}]
input_items.append(
{
"type": "function_call_output",
"call_id": tool_call_id,
"output": tool_output,
}
)
if tool_call_id in custom_tool_call_ids:
input_items.append(
ResponseCustomToolCallOutputParam(
type="custom_tool_call_output",
call_id=tool_call_id,
output=content if isinstance(content, str) else tool_output,
)
)
else:
input_items.append(
FunctionCallOutput(
type="function_call_output",
call_id=tool_call_id,
output=tool_output,
)
)
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))
for tool_call in tool_calls:
function = tool_call.get("function")
custom = tool_call.get("custom")
if function:
input_tool_call: dict[str, Any] = {
"type": "function_call",
@ -275,6 +317,15 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
if "arguments" in function:
input_tool_call["arguments"] = function["arguments"]
input_items.append(input_tool_call)
elif isinstance(custom, dict):
input_items.append(
ResponseCustomToolCallParam(
type="custom_tool_call",
call_id=tool_call["id"],
name=custom.get("name", ""),
input=custom.get("input", ""),
)
)
else:
raise ValueError(f"tool call not supported: {tool_call}")
elif content is not None:
@ -285,7 +336,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
{
"type": "message",
"role": role,
"content": self._convert_content_to_responses_format(content, cast(str, role)), # type: ignore[arg-type]
"content": self._convert_content_to_responses_format(content, cast(str, role)),
}
)
@ -309,17 +360,15 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
elif key == "response_format":
text_format = self._transform_response_format_to_text_format(value)
if text_format:
responses_api_request["text"] = text_format # type: ignore
responses_api_request["text"] = text_format
elif key == "tool_choice":
responses_api_request["tool_choice"] = ( # type: ignore[assignment]
self._normalize_tool_choice_for_responses_api(value)
)
responses_api_request["tool_choice"] = self._normalize_tool_choice_for_responses_api(value)
elif key == "stream_options":
stream_options = normalize_responses_api_stream_options(value)
if stream_options is not None:
responses_api_request["stream_options"] = stream_options
elif key in ResponsesAPIOptionalRequestParams.__annotations__.keys():
responses_api_request[key] = value # type: ignore
responses_api_request[key] = value
elif key == "previous_response_id":
responses_api_request["previous_response_id"] = value
elif key == "reasoning_effort":
@ -329,13 +378,13 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
def _build_sanitized_litellm_params(self, litellm_params: dict) -> dict[str, Any]:
"""Build sanitized litellm_params with merged metadata."""
responses_optional_param_keys = set(ResponsesAPIOptionalRequestParams.__annotations__.keys())
sanitized: dict[str, Any] = {
responses_optional_param_keys: Final = set(ResponsesAPIOptionalRequestParams.__annotations__.keys())
sanitized: Final[dict[str, Any]] = {
key: value for key, value in litellm_params.items() if key not in responses_optional_param_keys
}
legacy_metadata = litellm_params.get("metadata")
existing_litellm_metadata = litellm_params.get("litellm_metadata")
merged_litellm_metadata: dict[str, Any] = {}
legacy_metadata: Final = litellm_params.get("metadata")
existing_litellm_metadata: Final = litellm_params.get("litellm_metadata")
merged_litellm_metadata: Final[dict[str, Any]] = {}
if isinstance(legacy_metadata, dict):
merged_litellm_metadata.update(legacy_metadata)
if isinstance(existing_litellm_metadata, dict):
@ -399,7 +448,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
optional_params = self._extract_extra_body_params(optional_params)
# Build responses API request using the reverse transformation logic
responses_api_request = ResponsesAPIOptionalRequestParams()
responses_api_request: Final = ResponsesAPIOptionalRequestParams()
# Set instructions if we found a system message
if instructions:
@ -407,30 +456,30 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
self._map_optional_params_to_responses_api_request(optional_params, responses_api_request)
stream = optional_params.get("stream") or litellm_params.get("stream", False)
verbose_logger.debug(f"Chat provider: Stream parameter: {stream}")
stream: Final = optional_params.get("stream") or litellm_params.get("stream", False)
verbose_logger.debug("Chat provider: Stream parameter: %s", stream)
# Ensure stream is properly set in the request
if stream:
responses_api_request["stream"] = True
# Handle session management if previous_response_id is provided
previous_response_id = optional_params.get("previous_response_id")
previous_response_id: Final = optional_params.get("previous_response_id")
if previous_response_id:
# Use the existing session handler for responses API
verbose_logger.debug(f"Chat provider: Warning ignoring previous response ID: {previous_response_id}")
verbose_logger.debug("Chat provider: Warning ignoring previous response ID: %s", previous_response_id)
# Convert back to responses API format for the actual request
api_model = model
api_model: Final = model
from litellm.types.utils import CallTypes
setattr(litellm_logging_obj, "call_type", CallTypes.responses.value)
sanitized_litellm_params = self._build_sanitized_litellm_params(litellm_params)
sanitized_litellm_params: Final = self._build_sanitized_litellm_params(litellm_params)
request_data = {
request_data: Final = {
"model": api_model,
"input": input_items,
"litellm_logging_obj": litellm_logging_obj,
@ -438,7 +487,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
"client": client,
}
verbose_logger.debug(f"Chat provider: Final request model={api_model}, input_items={len(input_items)}")
verbose_logger.debug("Chat provider: Final request model=%s, input_items=%s", api_model, len(input_items))
self._merge_responses_api_request_into_request_data(request_data, responses_api_request, instructions)
@ -473,18 +522,18 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
ResponseApplyPatchToolCall,
)
except ImportError:
ResponseApplyPatchToolCall = None # type: ignore[assignment,misc]
ResponseApplyPatchToolCall = None
from litellm.types.utils import Choices, Message
choices: list[Choices] = []
choices: Final[list[Choices]] = []
index = 0
reasoning_content: str | None = None
pending_reasoning_item: dict[str, Any] | None = None
# Collect all tool calls to put them in a single choice
# (Chat Completions API expects all tool calls in one message)
accumulated_tool_calls: list[dict[str, Any]] = []
accumulated_tool_calls: Final[list[dict[str, Any]]] = []
tool_call_index = 0
for item in output_items:
@ -555,11 +604,21 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
accumulated_tool_calls.append(tool_call_dict)
tool_call_index += 1
elif isinstance(item, dict) and handle_raw_dict_callback is not None:
# Handle raw dict responses (e.g., from GPT-5 Codex)
choice, index = handle_raw_dict_callback(item=item, index=index)
if choice is not None:
choices.append(choice)
elif isinstance(item, (dict, BaseModel)):
# Raw dict items (e.g., from GPT-5 Codex) and pydantic items matching no
# openai SDK class above: typed ResponseCustomToolCall and litellm's own
# GenericResponseOutputItem from the completion bridge both land here
raw_item = item if isinstance(item, dict) else item.model_dump()
if raw_item.get("type") in ("function_call", "custom_tool_call"):
# Tool calls accumulate into the single trailing tool_calls choice
# like the typed branches above; a choice per call would hide every
# call after choices[0] from chat clients
accumulated_tool_calls.append(_tool_call_dict_from_output_item(raw_item, tool_call_index))
tool_call_index += 1
elif handle_raw_dict_callback is not None:
choice, index = handle_raw_dict_callback(item=raw_item, index=index)
if choice is not None:
choices.append(choice)
else:
pass # don't fail request if item in list is not supported
@ -582,10 +641,10 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
@classmethod
def _extract_output_from_completed_event(cls, parsed_chunk: dict[str, Any]) -> list[dict[str, Any]] | None:
response_payload = parsed_chunk.get("response")
response_payload: Final = parsed_chunk.get("response")
if not isinstance(response_payload, dict):
return None
response_output = response_payload.get("output")
response_output: Final = response_payload.get("output")
if not isinstance(response_output, list) or len(response_output) == 0:
return None
return cast(list[dict[str, Any]], response_output)
@ -595,8 +654,8 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
if not raw_sse or not isinstance(raw_sse, str):
return []
recovered_output_items: dict[int, dict[str, Any]] = {}
recovered_text_only_items: dict[int, dict[str, Any]] = {}
recovered_output_items: Final[dict[int, dict[str, Any]]] = {}
recovered_text_only_items: Final[dict[int, dict[str, Any]]] = {}
for chunk in raw_sse.splitlines():
parsed_chunk = parse_sse_json_chunk(chunk)
@ -631,7 +690,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
# but text-only items at indices without a matching OUTPUT_ITEM_DONE
# must still be preserved (e.g. multi-output responses where some
# indices only emitted OUTPUT_TEXT_DONE).
merged_items: dict[int, dict[str, Any]] = {**recovered_text_only_items}
merged_items: Final[dict[int, dict[str, Any]]] = {**recovered_text_only_items}
merged_items.update(recovered_output_items)
if merged_items:
@ -641,8 +700,8 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
@classmethod
def _recover_output_items_from_logging(cls, logging_obj: "LiteLLMLoggingObj") -> list[dict[str, Any]]:
model_call_details = getattr(logging_obj, "model_call_details", {}) or {}
original_response = model_call_details.get("original_response")
model_call_details: Final = getattr(logging_obj, "model_call_details", {}) or {}
original_response: Final = model_call_details.get("original_response")
return cls._recover_output_items_from_raw_sse(original_response)
def transform_response(
@ -671,7 +730,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
output_items = raw_response.output
if len(output_items) == 0:
recovered_output_items = self._recover_output_items_from_logging(logging_obj)
recovered_output_items: Final = self._recover_output_items_from_logging(logging_obj)
if recovered_output_items:
output_items = cast(Any, recovered_output_items)
raw_response.output = cast(Any, recovered_output_items)
@ -681,7 +740,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
)
# Convert response output to choices using the static helper
choices = self._convert_response_output_to_choices(
choices: Final = self._convert_response_output_to_choices(
output_items=output_items,
handle_raw_dict_callback=self._handle_raw_dict_response_item,
)
@ -704,7 +763,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
# Preserve hidden params from the ResponsesAPIResponse, especially the headers
# which contain important provider information like x-request-id
raw_response_hidden_params = getattr(raw_response, "_hidden_params", {})
raw_response_hidden_params: Final = getattr(raw_response, "_hidden_params", {})
if raw_response_hidden_params:
if not hasattr(model_response, "_hidden_params") or model_response._hidden_params is None:
model_response._hidden_params = {}
@ -740,7 +799,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
) -> "ResponseInputImageParam":
from openai.types.responses import ResponseInputImageParam
content_image_url = content.get("image_url")
content_image_url: Final = content.get("image_url")
actual_image_url: str | None = None
detail: Literal["low", "high", "auto"] | None = None
@ -756,7 +815,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
if actual_image_url is None:
raise ValueError(f"Invalid image URL: {content_image_url}")
image_param = ResponseInputImageParam(image_url=actual_image_url, detail="auto", type="input_image")
image_param: Final = ResponseInputImageParam(image_url=actual_image_url, detail="auto", type="input_image")
if detail:
image_param["detail"] = detail
@ -776,29 +835,29 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
"""Convert chat completion content to responses API format"""
from litellm.types.llms.openai import ChatCompletionImageObject
verbose_logger.debug(f"Chat provider: Converting content to responses format - input type: {type(content)}")
verbose_logger.debug("Chat provider: Converting content to responses format - input type: %s", type(content))
if content is None:
return [self._convert_content_str_to_input_text("", role)]
elif isinstance(content, str):
result = [self._convert_content_str_to_input_text(content, role)]
verbose_logger.debug(f"Chat provider: String content -> {result}")
verbose_logger.debug("Chat provider: String content -> %s", result)
return result
elif isinstance(content, list):
result = []
for i, item in enumerate(content):
verbose_logger.debug(f"Chat provider: Processing content item {i}: {type(item)} = {item}")
verbose_logger.debug("Chat provider: Processing content item %s: %s = %s", i, type(item), item)
if isinstance(item, str):
converted = self._convert_content_str_to_input_text(item, role)
result.append(converted)
verbose_logger.debug(f"Chat provider: -> {converted}")
verbose_logger.debug("Chat provider: -> %s", converted)
elif isinstance(item, dict):
# Handle multimodal content
original_type = item.get("type")
if original_type == "text":
converted = self._convert_content_str_to_input_text(item.get("text", ""), role)
result.append(converted)
verbose_logger.debug(f"Chat provider: text -> {converted}")
verbose_logger.debug("Chat provider: text -> %s", converted)
elif original_type == "image_url":
# Map to responses API image format
converted = cast(
@ -808,14 +867,14 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
),
)
result.append(converted)
verbose_logger.debug(f"Chat provider: image_url -> {converted}")
verbose_logger.debug("Chat provider: image_url -> %s", converted)
else:
# Try to map other types to responses API format
item_type = original_type or "input_text"
if item_type == "image":
converted = {"type": "input_image", **item}
result.append(converted)
verbose_logger.debug(f"Chat provider: image -> {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": "..."}}
@ -827,7 +886,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
if key in file_data:
converted[key] = file_data[key]
result.append(converted)
verbose_logger.debug(f"Chat provider: file -> {converted}")
verbose_logger.debug("Chat provider: file -> %s", converted)
elif item_type in [
"input_text",
"input_image",
@ -839,22 +898,22 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
]:
# Already in responses API format
result.append(item)
verbose_logger.debug(f"Chat provider: passthrough -> {item}")
verbose_logger.debug("Chat provider: passthrough -> %s", item)
else:
# Default to input_text for unknown types
converted = self._convert_content_str_to_input_text(str(item.get("text", item)), role)
result.append(converted)
verbose_logger.debug(f"Chat provider: unknown({original_type}) -> {converted}")
verbose_logger.debug(f"Chat provider: Final converted content: {result}")
verbose_logger.debug("Chat provider: unknown(%s) -> %s", original_type, converted)
verbose_logger.debug("Chat provider: Final converted content: %s", result)
return result
else:
result = [self._convert_content_str_to_input_text(str(content), role)]
verbose_logger.debug(f"Chat provider: Other content type -> {result}")
verbose_logger.debug("Chat provider: Other content type -> %s", result)
return result
def _convert_tools_to_responses_format(self, tools: list[dict[str, Any]]) -> list["ALL_RESPONSES_API_TOOL_PARAMS"]:
"""Convert chat completion tools to responses API tools format"""
responses_tools: list[ALL_RESPONSES_API_TOOL_PARAMS] = []
responses_tools: Final[list[ALL_RESPONSES_API_TOOL_PARAMS]] = []
for tool in tools:
# convert function tool from chat completion to responses API format
if tool.get("type") == "function":
@ -868,8 +927,20 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
description=function_tool.get("description"),
)
)
elif tool.get("type") == "custom" and isinstance(tool.get("custom"), dict):
from litellm.litellm_core_utils.prompt_templates.common_utils import (
convert_custom_tool_format_to_responses_shape,
)
custom_payload = tool["custom"]
flat_custom = CustomToolParam(type="custom", name=custom_payload.get("name", ""))
if custom_payload.get("description") is not None:
flat_custom["description"] = custom_payload["description"]
if isinstance(custom_payload.get("format"), dict):
flat_custom["format"] = convert_custom_tool_format_to_responses_shape(custom_payload["format"])
responses_tools.append(flat_custom)
else:
responses_tools.append(tool) # type: ignore
responses_tools.append(tool)
return cast(list["ALL_RESPONSES_API_TOOL_PARAMS"], responses_tools)
@ -880,11 +951,11 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
unsupported params remain in extra_body.
"""
# Extract extra_body and separate supported params from unsupported ones
extra_body = optional_params.pop("extra_body", None) or {}
extra_body: Final = optional_params.pop("extra_body", None) or {}
if not extra_body:
return optional_params
supported_responses_api_params = set(ResponsesAPIOptionalRequestParams.__annotations__.keys())
supported_responses_api_params: Final = set(ResponsesAPIOptionalRequestParams.__annotations__.keys())
# Also include params we handle specially
supported_responses_api_params.update(
{
@ -894,7 +965,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
)
# Extract supported params from extra_body and merge into optional_params
extra_body_copy = extra_body.copy()
extra_body_copy: Final = extra_body.copy()
for key, value in extra_body_copy.items():
if key in supported_responses_api_params:
# Prefer extra_body value if it exists (may have more complete info like summary in reasoning_effort)
@ -905,21 +976,21 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
def _map_reasoning_effort(self, reasoning_effort: str | dict[str, Any]) -> Reasoning | None:
# If dict is passed, convert it directly to Reasoning object
if isinstance(reasoning_effort, dict):
return Reasoning(**reasoning_effort) # type: ignore[typeddict-item]
return Reasoning(**reasoning_effort)
# Check if auto-summary is enabled via flag or environment variable
# Priority: litellm.reasoning_auto_summary flag > LITELLM_REASONING_AUTO_SUMMARY env var
auto_summary_enabled = (
auto_summary_enabled: Final = (
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") # type: ignore
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") # type: ignore[typeddict-item]
return Reasoning(effort="xhigh", summary="detailed") if auto_summary_enabled else Reasoning(effort="xhigh")
elif reasoning_effort == "medium":
return (
Reasoning(effort="medium", summary="detailed") if auto_summary_enabled else Reasoning(effort="medium")
@ -953,7 +1024,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
tools = []
responses_api_request["tools"] = tools
web_search_tool: dict[str, Any] = {"type": "web_search"}
web_search_tool: Final[dict[str, Any]] = {"type": "web_search"}
if isinstance(web_search_options, dict):
web_search_tool.update(web_search_options)
@ -988,10 +1059,10 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
return None
if isinstance(response_format, dict):
format_type = response_format.get("type")
format_type: Final = response_format.get("type")
if format_type == "json_schema":
json_schema = response_format.get("json_schema", {})
json_schema: Final = response_format.get("json_schema", {})
return {
"format": {
"type": "json_schema",
@ -1020,7 +1091,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
if not annotations:
return None
result: list[ChatCompletionAnnotation] = []
result: Final[list[ChatCompletionAnnotation]] = []
for annotation in annotations:
try:
# Convert Pydantic models to dicts (handles both v1 and v2)
@ -1032,13 +1103,13 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
annotation_dict = annotation
else:
# Skip unsupported annotation types
verbose_logger.debug(f"Skipping unsupported annotation type: {type(annotation)}")
verbose_logger.debug("Skipping unsupported annotation type: %s", type(annotation))
continue
result.append(annotation_dict) # type: ignore
result.append(annotation_dict)
except Exception as e:
# Skip malformed annotations
verbose_logger.debug(f"Skipping malformed annotation: {annotation}, error: {e}")
verbose_logger.debug("Skipping malformed annotation: %s, error: %s", annotation, e)
continue
return result if result else None
@ -1048,7 +1119,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
if not status:
return "stop"
status_mapping = {
status_mapping: Final = {
"completed": "stop",
"incomplete": "length",
"failed": "stop",
@ -1062,6 +1133,7 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
def __init__(self, streaming_response, sync_stream: bool, json_mode: bool | None = False):
super().__init__(streaming_response, sync_stream, json_mode)
self._chat_completion_id: str | None = None
self._tool_call_index_map: dict[int, int] = {} # mutable-ok: per-stream accumulator state
def _handle_string_chunk(
self, str_line: Union[str, "BaseModel"]
@ -1074,21 +1146,41 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
if not str_line or str_line.startswith("event:"):
# ignore.
return GenericStreamingChunk(text="", tool_use=None, is_finished=False, finish_reason="", usage=None)
index = str_line.find("data:")
index: Final = str_line.find("data:")
if index != -1:
str_line = str_line[index + 5 :]
return self.chunk_parser(json.loads(str_line))
@staticmethod
def _sequential_tool_call_index(
tool_call_index_map: dict[int, int] | None, # mutable-ok: per-stream state, remapped in place
output_index: int,
) -> int:
"""Chat-completions tool_call indices must be 0-based and sequential, but
Responses API ``output_index`` counts every output item (reasoning,
message, ...), so the first tool call of a reasoning model arrives at
output_index >= 1 and strict SSE accumulators (e.g. Cursor agent mode)
misplace it. When a per-stream map is provided, remap each distinct
output_index to the next sequential slot; without a map (stateless
callers), fall back to the raw output_index."""
if tool_call_index_map is None:
return output_index
if output_index not in tool_call_index_map:
tool_call_index_map[output_index] = len(tool_call_index_map) # mutable-ok: per-stream accumulator state
return tool_call_index_map[output_index]
@staticmethod
def translate_responses_chunk_to_openai_stream(
parsed_chunk: dict | BaseModel,
tool_call_index_map: dict[int, int] | None = None, # mutable-ok: per-stream state, remapped in place
) -> "ModelResponseStream":
"""
Translate a Responses API streaming chunk to OpenAI chat completion streaming format.
Args:
parsed_chunk: Dict containing the Responses API event chunk
tool_call_index_map: Per-stream output_index -> sequential tool_call index map
Returns:
ModelResponseStream: OpenAI-formatted streaming chunk
@ -1122,11 +1214,11 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
):
return ModelResponseStream(**parsed_chunk)
verbose_logger.debug(f"Chat provider: Processing event type: {event_type}")
verbose_logger.debug("Chat provider: Processing event type: %s", event_type)
if event_type == "response.created":
# Initial response creation event
verbose_logger.debug(f"Chat provider: response.created -> {parsed_chunk}")
verbose_logger.debug("Chat provider: response.created -> %s", parsed_chunk)
return ModelResponseStream(
choices=[
StreamingChoices(
@ -1139,39 +1231,28 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
elif event_type == "response.output_item.added":
# New output item added
output_item = parsed_chunk.get("item", {})
if output_item.get("type") == "function_call":
# Extract provider_specific_fields if present
provider_specific_fields = output_item.get("provider_specific_fields")
if provider_specific_fields and not isinstance(provider_specific_fields, dict):
provider_specific_fields = (
dict(provider_specific_fields) if hasattr(provider_specific_fields, "__dict__") else {}
)
if output_item.get("type") in ("function_call", "custom_tool_call"):
converted: Final = _tool_call_dict_from_output_item(output_item, parsed_chunk.get("output_index", 0))
provider_specific_fields: Final = converted.get("provider_specific_fields")
function_chunk = ChatCompletionToolCallFunctionChunk(
name=output_item.get("name", None),
arguments=parsed_chunk.get("arguments", ""),
function_chunk: Final = ChatCompletionToolCallFunctionChunk(
name=converted["function"]["name"] or None,
arguments=converted["function"]["arguments"] or parsed_chunk.get("arguments") or "",
)
if provider_specific_fields:
function_chunk["provider_specific_fields"] = provider_specific_fields
from litellm.responses.litellm_completion_transformation.transformation import (
LiteLLMCompletionResponsesConfig,
tool_call_index = OpenAiResponsesToChatCompletionStreamIterator._sequential_tool_call_index(
tool_call_index_map, parsed_chunk.get("output_index", 0)
)
tool_call_index = parsed_chunk.get("output_index", 0)
tool_call_chunk = ChatCompletionToolCallChunk(
id=LiteLLMCompletionResponsesConfig._tool_call_id_from_responses_item(
output_item.get("id"), output_item.get("call_id")
),
tool_call_chunk: Final = ChatCompletionToolCallChunk(
id=converted["id"],
index=tool_call_index,
type="function",
function=function_chunk,
)
# Add provider_specific_fields if present
if provider_specific_fields:
tool_call_chunk.provider_specific_fields = provider_specific_fields # type: ignore
tool_call_chunk.provider_specific_fields = provider_specific_fields
return ModelResponseStream(
choices=[
@ -1182,10 +1263,15 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
)
]
)
elif event_type == "response.function_call_arguments.delta":
elif event_type in (
ResponsesAPIStreamEvents.FUNCTION_CALL_ARGUMENTS_DELTA,
ResponsesAPIStreamEvents.CUSTOM_TOOL_CALL_INPUT_DELTA,
):
content_part: str | None = parsed_chunk.get("delta", None)
if content_part:
tool_call_index = parsed_chunk.get("output_index", 0)
tool_call_index = OpenAiResponsesToChatCompletionStreamIterator._sequential_tool_call_index(
tool_call_index_map, parsed_chunk.get("output_index", 0)
)
return ModelResponseStream(
choices=[
StreamingChoices(
@ -1209,39 +1295,32 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
elif event_type == ResponsesAPIStreamEvents.OUTPUT_ITEM_DONE:
# New output item added
output_item = parsed_chunk.get("item", {})
if output_item.get("type") == "function_call":
# Extract provider_specific_fields if present
provider_specific_fields = output_item.get("provider_specific_fields")
if provider_specific_fields and not isinstance(provider_specific_fields, dict):
provider_specific_fields = (
dict(provider_specific_fields) if hasattr(provider_specific_fields, "__dict__") else {}
if output_item.get("type") in ("function_call", "custom_tool_call"):
if tool_call_index_map is None:
# Stateless callers (the responses guardrail handler extracting
# tool calls from a buffered output_item.done) get the complete
# tool call; per-stream callers already received it via
# output_item.added and the argument delta events
return ModelResponseStream(
choices=[ # mutable-ok: ModelResponseStream coerces only list choices
StreamingChoices(
index=0,
delta=Delta(
tool_calls=(
_tool_call_dict_from_output_item(
output_item, parsed_chunk.get("output_index", 0)
),
)
),
finish_reason=None,
)
]
)
function_chunk = ChatCompletionToolCallFunctionChunk(
name=output_item.get("name", None),
arguments="", # responses API sends everything again, we don't
)
# Add provider_specific_fields to function if present
if provider_specific_fields:
function_chunk["provider_specific_fields"] = provider_specific_fields
tool_call_index = parsed_chunk.get("output_index", 0)
tool_call_chunk = ChatCompletionToolCallChunk(
id=output_item.get("call_id"),
index=tool_call_index,
type="function",
function=function_chunk,
)
# Add provider_specific_fields if present
if provider_specific_fields:
tool_call_chunk.provider_specific_fields = provider_specific_fields # type: ignore
# Do NOT emit finish_reason here — response.completed handles the terminal
# finish_reason. Emitting "tool_calls" here would prematurely terminate
# the stream before subsequent tool calls arrive (same fix as #17246 for
# the message-type branch).
# the message-type branch). The item's fields were already streamed via
# output_item.added and the argument delta events.
return ModelResponseStream(
choices=[
StreamingChoices(
@ -1296,14 +1375,16 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
# Check if response contains function_call items in output
# to determine correct finish_reason
response_data = parsed_chunk.get("response", {})
output_items = response_data.get("output", []) if response_data else []
response_data: Final = parsed_chunk.get("response", {})
output_items: Final = response_data.get("output", []) if response_data else []
has_function_calls = any(
item.get("type") == "function_call" for item in output_items if isinstance(item, dict)
has_function_calls: Final = any(
item.get("type") in ("function_call", "custom_tool_call")
for item in output_items
if isinstance(item, dict)
)
finish_reason = "tool_calls" if has_function_calls else "stop"
finish_reason: Final = "tool_calls" if has_function_calls else "stop"
# Extract reasoning items with encrypted_content for round-tripping
completed_reasoning_items: list[dict[str, Any]] | None = None
@ -1319,7 +1400,7 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
summary_raw=item.get("summary"),
)
)
completed_reasoning_items_typed = cast(
completed_reasoning_items_typed: Final = cast(
list[ChatCompletionReasoningItem] | None,
completed_reasoning_items,
)
@ -1345,7 +1426,7 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
else:
pass
# For any unhandled event types, create a minimal valid chunk or skip
verbose_logger.debug(f"Chat provider: Unhandled event type '{event_type}', creating empty chunk")
verbose_logger.debug("Chat provider: Unhandled event type '%s', creating empty chunk", event_type)
# Return a minimal valid chunk for unknown events
return ModelResponseStream(
@ -1368,9 +1449,11 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
Returns:
ModelResponseStream: OpenAI-formatted streaming chunk
"""
verbose_logger.debug(f"Chat provider: transform_streaming_response called with chunk: {chunk}")
verbose_logger.debug("Chat provider: transform_streaming_response called with chunk: %s", chunk)
return self._with_stream_scoped_id(
OpenAiResponsesToChatCompletionStreamIterator.translate_responses_chunk_to_openai_stream(chunk)
OpenAiResponsesToChatCompletionStreamIterator.translate_responses_chunk_to_openai_stream(
chunk, tool_call_index_map=self._tool_call_index_map
)
)
def _with_stream_scoped_id(self, chunk: "ModelResponseStream") -> "ModelResponseStream":

72
litellm/compression/compress.py
View file

@ -4,7 +4,7 @@ scoring, message stubbing, and retrieval tool injection.
"""
from collections.abc import Mapping, Sequence
from typing import Any, cast
from typing import Any, Final, cast
from litellm.caching.dual_cache import DualCache
from litellm.compression.message_stubbing import (
@ -20,11 +20,11 @@ from litellm.types.utils import CallTypes
# CallTypes that produce Anthropic-shaped messages (structured content blocks).
# Everything else is treated as OpenAI chat-completions shape.
_ANTHROPIC_CALL_TYPES = frozenset({CallTypes.anthropic_messages.value})
_ANTHROPIC_CALL_TYPES: Final = frozenset({CallTypes.anthropic_messages.value})
# CallTypes that are valid targets for compression. Compression operates on
# message-shaped inputs, so we only accept call types whose payload is a list
# of role/content messages.
_SUPPORTED_CALL_TYPES = frozenset(
_SUPPORTED_CALL_TYPES: Final = frozenset(
{
CallTypes.completion.value,
CallTypes.acompletion.value,
@ -54,7 +54,7 @@ def _build_retrieval_tools(keys: list[str], call_type: str) -> list[dict]:
if not keys:
return []
openai_tools = [build_retrieval_tool(keys)]
openai_tools: Final = [build_retrieval_tool(keys)]
if not _is_anthropic_call_type(call_type):
return openai_tools
@ -77,8 +77,8 @@ def _content_to_text(content: Any) -> str:
Implemented iteratively (stack-based) to avoid unbounded recursion.
"""
parts: list[str] = []
stack: list[Any] = [content]
parts: Final[list[str]] = []
stack: Final[list[Any]] = [content]
while stack:
item = stack.pop()
if isinstance(item, str):
@ -111,9 +111,9 @@ def _normalize_messages_for_compression(
f"Unsupported call_type={call_type!r} for compression. Expected one of: {sorted(_SUPPORTED_CALL_TYPES)}."
)
original_messages: list[dict[str, Any]] = [dict(m) for m in messages]
original_messages: Final[list[dict[str, Any]]] = [dict(m) for m in messages]
normalized_messages: list[dict] = []
normalized_messages: Final[list[dict]] = []
for msg in original_messages:
normalized_messages.append(
{
@ -135,7 +135,7 @@ def _extract_last_user_message(messages: list[dict]) -> str:
def _extract_tool_use_ids(content: Any) -> list[str]:
if not isinstance(content, list):
return []
tool_use_ids: list[str] = []
tool_use_ids: Final[list[str]] = []
for part in content:
if not isinstance(part, dict):
continue
@ -150,7 +150,7 @@ def _extract_tool_use_ids(content: Any) -> list[str]:
def _extract_tool_result_ids(content: Any) -> set[str]:
if not isinstance(content, list):
return set()
tool_result_ids: set[str] = set()
tool_result_ids: Final[set[str]] = set()
for part in content:
if not isinstance(part, dict):
continue
@ -171,7 +171,7 @@ def _extract_anthropic_tool_exchange_spans(
Each assistant message containing `tool_use` must be immediately followed by a
user message containing matching `tool_result` blocks for all tool_use ids.
"""
spans: list[set[int]] = []
spans: Final[list[set[int]]] = []
i = 0
while i < len(messages):
current = messages[i]
@ -216,8 +216,8 @@ def get_protected_indices(messages: Sequence[Mapping[str, object]]) -> tuple[int
so compressing it replaces the live instruction with a marker. Compression
guardrails share this policy; see the Headroom guardrail.
"""
system_indices = tuple(index for index, msg in enumerate(messages) if msg.get("role", "") == "system")
last_user = tuple(index for index, msg in enumerate(messages) if msg.get("role", "") == "user")[-1:]
system_indices: Final = tuple(index for index, msg in enumerate(messages) if msg.get("role", "") == "system")
last_user: Final = tuple(index for index, msg in enumerate(messages) if msg.get("role", "") == "user")[-1:]
last_assistant = tuple(index for index, msg in enumerate(messages) if msg.get("role", "") == "assistant")[-1:]
return system_indices + last_user + last_assistant
@ -230,16 +230,16 @@ def _combine_scores(
"""Weighted average of BM25 and embedding scores, with min-max normalization."""
def _normalize(scores: list[float]) -> list[float]:
min_s = min(scores) if scores else 0.0
max_s = max(scores) if scores else 0.0
rng = max_s - min_s
min_s: Final = min(scores) if scores else 0.0
max_s: Final = max(scores) if scores else 0.0
rng: Final = max_s - min_s
if rng == 0:
return [0.0] * len(scores)
return [(s - min_s) / rng for s in scores]
norm_bm25 = _normalize(bm25_scores)
norm_emb = _normalize(emb_scores)
emb_weight = 1.0 - bm25_weight
norm_bm25: Final = _normalize(bm25_scores)
norm_emb: Final = _normalize(emb_scores)
emb_weight: Final = 1.0 - bm25_weight
return [bm25_weight * b + emb_weight * e for b, e in zip(norm_bm25, norm_emb)]
@ -253,7 +253,7 @@ def _select_kept_indices_for_budget(
initial_kept_indices: set[int],
tool_exchange_spans: list[set[int]],
) -> tuple[set[int], dict[int, dict]]:
kept_indices = set(initial_kept_indices)
kept_indices: Final = set(initial_kept_indices)
current_tokens = 0
for i in kept_indices:
current_tokens += token_counter(
@ -265,14 +265,14 @@ def _select_kept_indices_for_budget(
# A unit is either:
# 1) a single message index, or
# 2) an Anthropic tool-exchange span that must be kept/dropped atomically.
truncated_overrides: dict[int, dict] = {} # idx -> truncated message dict
span_id_by_index: dict[int, int] = {}
truncated_overrides: Final[dict[int, dict]] = {} # idx -> truncated message dict
span_id_by_index: Final[dict[int, int]] = {}
for span_id, span in enumerate(tool_exchange_spans):
for idx in span:
span_id_by_index[idx] = span_id
# Build single-message candidate units (non-span messages).
candidate_units: list[tuple[float, tuple[int, ...], bool]] = []
candidate_units: Final[list[tuple[float, tuple[int, ...], bool]]] = []
for idx in range(len(normalized_messages)):
if idx in span_id_by_index or idx in kept_indices:
continue
@ -323,7 +323,7 @@ def _select_kept_indices_for_budget(
def _get_dropped_tool_span_indices(kept_indices: set[int], tool_exchange_spans: list[set[int]]) -> set[int]:
dropped_tool_span_indices: set[int] = set()
dropped_tool_span_indices: Final[set[int]] = set()
for span in tool_exchange_spans:
if not any(idx in kept_indices for idx in span):
dropped_tool_span_indices.update(span)
@ -372,7 +372,7 @@ def compress(
A ``CompressedResult`` dict containing compressed messages, token
counts, a cache of original content, and the retrieval tool definition.
"""
call_type_str = _normalize_call_type(call_type)
call_type_str: Final = _normalize_call_type(call_type)
normalized_messages, original_messages = _normalize_messages_for_compression(
messages=messages,
call_type=call_type_str,
@ -381,7 +381,7 @@ def compress(
if compression_target is None:
compression_target = compression_trigger * 7 // 10
original_tokens = token_counter(
original_tokens: Final = token_counter(
model=model,
messages=cast(list[Any], original_messages),
)
@ -399,17 +399,17 @@ def compress(
)
# Extract query for relevance scoring
query = _extract_last_user_message(normalized_messages)
query: Final = _extract_last_user_message(normalized_messages)
# Score each message
bm25_scores = bm25_score_messages(query, normalized_messages)
bm25_scores: Final = bm25_score_messages(query, normalized_messages)
if embedding_model:
from litellm.compression.scoring.embedding_scorer import (
embedding_score_messages,
)
emb_scores = embedding_score_messages(
emb_scores: Final = embedding_score_messages(
query,
normalized_messages,
model=embedding_model,
@ -421,7 +421,7 @@ def compress(
combined_scores = bm25_scores
# Protected messages are never compressed
protected_indices = get_protected_indices(normalized_messages)
protected_indices: Final = get_protected_indices(normalized_messages)
kept_indices: set[int] = set(protected_indices)
tool_exchange_spans: list[set[int]] = []
@ -454,10 +454,10 @@ def compress(
)
# Build compressed messages and cache
compressed_messages: list[dict] = []
cache: dict[str, str] = {}
used_keys: set[str] = set()
dropped_tool_span_indices = _get_dropped_tool_span_indices(
compressed_messages: Final[list[dict]] = []
cache: Final[dict[str, str]] = {}
used_keys: Final[set[str]] = set()
dropped_tool_span_indices: Final = _get_dropped_tool_span_indices(
kept_indices=kept_indices, tool_exchange_spans=tool_exchange_spans
)
@ -474,9 +474,9 @@ def compress(
compressed_messages.append(stub_message(msg, key))
# Build retrieval tool in the target request schema
tools = _build_retrieval_tools(list(cache.keys()), call_type=call_type_str)
tools: Final = _build_retrieval_tools(list(cache.keys()), call_type=call_type_str)
compressed_tokens = token_counter(
compressed_tokens: Final = token_counter(
model=model,
messages=cast(list[Any], compressed_messages),
)

13
litellm/compression/content_detection.py
View file

@ -4,8 +4,9 @@ Auto-detect content type per message: code, JSON, or text.
import json
import re
from typing import Final
_CODE_KEYWORDS = re.compile(
_CODE_KEYWORDS: Final = re.compile(
r"\b(?:def |function |class |import |from |require\(|#include|fn |func |const |let |var |public |private |static )\b"
)
@ -16,7 +17,7 @@ def detect_content_type(content: str) -> str:
Returns one of: "code", "json", "text"
"""
stripped = content.strip()
stripped: Final = content.strip()
if not stripped:
return "text"
@ -30,10 +31,10 @@ def detect_content_type(content: str) -> str:
# Check code indicators
# Sample first 5000 chars for performance
sample = stripped[:5000]
keyword_matches = len(_CODE_KEYWORDS.findall(sample))
lines = sample.split("\n")
indented_lines = sum(1 for line in lines if line.startswith((" ", "\t")) and line.strip())
sample: Final = stripped[:5000]
keyword_matches: Final = len(_CODE_KEYWORDS.findall(sample))
lines: Final = sample.split("\n")
indented_lines: Final = sum(1 for line in lines if line.startswith((" ", "\t")) and line.strip())
# If we see multiple code keywords or significant indentation, it's likely code
if keyword_matches >= 3 or (indented_lines > len(lines) * 0.3 and len(lines) > 5):

25
litellm/compression/message_stubbing.py
View file

@ -3,11 +3,12 @@ Replace messages with compact stubs and extract human-readable keys.
"""
import re
from typing import Final
from litellm.compression.content_detection import detect_content_type
# Patterns for extracting file paths from content
_FILE_PATH_PATTERNS = [
_FILE_PATH_PATTERNS: Final = [
re.compile(r"^#\s*(\S+\.\w+)", re.MULTILINE), # # filename.py
re.compile(r"^//\s*(\S+\.\w+)", re.MULTILINE), # // filename.js
re.compile(r"^File:\s*(\S+)", re.MULTILINE), # File: path/to/file
@ -40,7 +41,7 @@ def extract_key(message: dict, fallback_index: int, used_keys: set[str]) -> str:
key = f"message_{fallback_index}"
# Handle duplicates
base_key = key
base_key: Final = key
counter = 2
while key in used_keys:
key = f"{base_key}_{counter}"
@ -61,10 +62,10 @@ def stub_message(message: dict, key: str) -> dict:
if isinstance(content, list):
content = " ".join(p.get("text", "") if isinstance(p, dict) else str(p) for p in content)
line_count = content.count("\n") + 1
content_type = detect_content_type(content)
line_count: Final = content.count("\n") + 1
content_type: Final = detect_content_type(content)
stub_content = (
stub_content: Final = (
f"[Compressed: {key} — {line_count} lines, {content_type}. "
f"Use litellm_content_retrieve tool to get full content.]"
)
@ -89,23 +90,23 @@ def truncate_message(message: dict, max_tokens: int) -> dict:
content = " ".join(p.get("text", "") if isinstance(p, dict) else str(p) for p in content)
# Rough conversion: 1 token ≈ 3 characters
target_chars = max(100, max_tokens * 3)
target_chars: Final = max(100, max_tokens * 3)
if len(content) <= target_chars:
return {**message, "content": content}
lines = content.split("\n")
lines: Final = content.split("\n")
# Estimate target line count from character budget
avg_line_len = max(1, len(content) // max(1, len(lines)))
target_lines = max(2, target_chars // avg_line_len)
avg_line_len: Final = max(1, len(content) // max(1, len(lines)))
target_lines: Final = max(2, target_chars // avg_line_len)
if len(lines) <= target_lines:
return {**message, "content": content}
first_count = (target_lines * 7) // 10
last_count = target_lines - first_count
truncated = (
first_count: Final = (target_lines * 7) // 10
last_count: Final = target_lines - first_count
truncated: Final = (
"\n".join(lines[:first_count]) + "\n...[truncated for context window]...\n" + "\n".join(lines[-last_count:])
)
return {**message, "content": truncated}

25
litellm/compression/scoring/bm25.py
View file

@ -7,6 +7,7 @@ No external dependencies — uses only stdlib.
import math
import re
from collections import Counter
from typing import Final
def _tokenize(text: str) -> list[str]:
@ -16,11 +17,11 @@ def _tokenize(text: str) -> list[str]:
def _extract_content(message: dict) -> str:
"""Extract text content from a message dict."""
content = message.get("content", "")
content: Final = message.get("content", "")
if isinstance(content, str):
return content
if isinstance(content, list):
parts = []
parts: Final = []
for part in content:
if isinstance(part, dict) and part.get("type") == "text":
parts.append(part.get("text", ""))
@ -48,32 +49,32 @@ def bm25_score_messages(
Returns:
List of float scores, one per message. Higher = more relevant.
"""
query_terms = _tokenize(query)
query_terms: Final = _tokenize(query)
if not query_terms:
return [0.0] * len(messages)
# Tokenize all documents
doc_tokens: list[list[str]] = []
doc_tokens: Final[list[list[str]]] = []
for msg in messages:
doc_tokens.append(_tokenize(_extract_content(msg)))
n = len(doc_tokens)
n: Final = len(doc_tokens)
if n == 0:
return []
# Average document length
doc_lengths = [len(dt) for dt in doc_tokens]
avgdl = sum(doc_lengths) / n if n > 0 else 1.0
doc_lengths: Final = [len(dt) for dt in doc_tokens]
avgdl: Final = sum(doc_lengths) / n if n > 0 else 1.0
# Document frequency for each term
df: dict[str, int] = {}
df: Final[dict[str, int]] = {}
for dt in doc_tokens:
seen = set(dt)
for term in seen:
df[term] = df.get(term, 0) + 1
# IDF for query terms
idf: dict[str, float] = {}
idf: Final[dict[str, float]] = {}
for term in set(query_terms):
term_df = df.get(term, 0)
# Standard BM25 IDF: log((N - df + 0.5) / (df + 0.5) + 1)
@ -83,9 +84,9 @@ def bm25_score_messages(
# document tokens that start with that term (min 4 chars match). This lets
# "cook" match "cooking" and "auth" match "authentication" without a full
# stemmer dependency.
def _expand_tf(query_term: str, tf_counts: Counter) -> int: # type: ignore[type-arg]
def _expand_tf(query_term: str, tf_counts: Counter) -> int:
"""Sum TF across all doc tokens that are prefixed by query_term."""
exact = tf_counts.get(query_term, 0)
exact: Final = tf_counts.get(query_term, 0)
if exact:
return exact
if len(query_term) < 4:
@ -93,7 +94,7 @@ def bm25_score_messages(
return sum(count for token, count in tf_counts.items() if token != query_term and token.startswith(query_term))
# Score each document
scores: list[float] = []
scores: Final[list[float]] = []
for i, dt in enumerate(doc_tokens):
if not dt:
scores.append(0.0)

26
litellm/compression/scoring/embedding_scorer.py
View file

@ -5,18 +5,18 @@ Computes cosine similarity between the query embedding and each message embeddin
"""
import math
from typing import Any
from typing import Any, Final
from litellm.caching.dual_cache import DualCache
def _extract_content(message: dict) -> str:
"""Extract text content from a message dict."""
content = message.get("content", "")
content: Final = message.get("content", "")
if isinstance(content, str):
return content
if isinstance(content, list):
parts = []
parts: Final = []
for part in content:
if isinstance(part, dict) and part.get("type") == "text":
parts.append(part.get("text", ""))
@ -30,15 +30,15 @@ def _truncate_text(text: str, max_chars: int = 30000) -> str:
"""Truncate long text, keeping first and last portions."""
if len(text) <= max_chars:
return text
half = max_chars // 2
half: Final = max_chars // 2
return text[:half] + "\n...\n" + text[-half:]
def _cosine_similarity(a: list[float], b: list[float]) -> float:
"""Compute cosine similarity between two vectors."""
dot = sum(x * y for x, y in zip(a, b))
norm_a = math.sqrt(sum(x * x for x in a))
norm_b = math.sqrt(sum(x * x for x in b))
dot: Final = sum(x * y for x, y in zip(a, b))
norm_a: Final = math.sqrt(sum(x * x for x in a))
norm_b: Final = math.sqrt(sum(x * x for x in b))
if norm_a == 0 or norm_b == 0:
return 0.0
return dot / (norm_a * norm_b)
@ -67,12 +67,12 @@ def embedding_score_messages(
"""
import litellm
texts = [_truncate_text(query)]
texts: Final = [_truncate_text(query)]
for msg in messages:
texts.append(_truncate_text(_extract_content(msg)))
# Filter out empty texts — replace with a placeholder to maintain indexing
processed_texts = [t if t.strip() else "empty" for t in texts]
processed_texts: Final = [t if t.strip() else "empty" for t in texts]
kwargs: dict[str, Any] = {
"model": model,
@ -82,13 +82,13 @@ def embedding_score_messages(
if embedding_model_params:
kwargs = {**kwargs, **embedding_model_params}
response = litellm.embedding(**kwargs)
response: Final = litellm.embedding(**kwargs)
# Extract embedding vectors
embeddings = [item["embedding"] for item in response.data]
embeddings: Final = [item["embedding"] for item in response.data]
query_embedding = embeddings[0]
scores: list[float] = []
query_embedding: Final = embeddings[0]
scores: Final[list[float]] = []
for i in range(1, len(embeddings)):
scores.append(_cosine_similarity(query_embedding, embeddings[i]))

769
litellm/constants.py
View file

File diff suppressed because it is too large Load diff

66
litellm/containers/endpoint_factory.py
View file

@ -11,7 +11,7 @@ import json
from collections.abc import Callable
from functools import partial
from pathlib import Path
from typing import Any, Literal
from typing import Any, Final, Literal
import litellm
from litellm.constants import request_timeout as DEFAULT_REQUEST_TIMEOUT
@ -28,7 +28,7 @@ from litellm.types.router import GenericLiteLLMParams
from litellm.utils import ProviderConfigManager, client
# Response type mapping
RESPONSE_TYPES: dict[str, type] = {
RESPONSE_TYPES: Final[dict[str, type]] = {
"ContainerFileListResponse": ContainerFileListResponse,
"ContainerFileObject": ContainerFileObject,
"DeleteContainerFileResponse": DeleteContainerFileResponse,
@ -37,7 +37,7 @@ RESPONSE_TYPES: dict[str, type] = {
def _load_endpoints_config() -> dict:
"""Load the endpoints configuration from JSON file."""
config_path = Path(__file__).parent / "endpoints.json"
config_path: Final = Path(__file__).parent / "endpoints.json"
with open(config_path) as f:
return json.load(f)
@ -48,9 +48,9 @@ def create_sync_endpoint_function(endpoint_config: dict) -> Callable:
Uses the generic container handler instead of individual handler methods.
"""
endpoint_name = endpoint_config["name"]
response_type = RESPONSE_TYPES.get(endpoint_config["response_type"])
path_params = endpoint_config.get("path_params", [])
endpoint_name: Final = endpoint_config["name"]
response_type: Final = RESPONSE_TYPES.get(endpoint_config["response_type"])
path_params: Final = endpoint_config.get("path_params", [])
@client
def endpoint_func(
@ -61,12 +61,12 @@ def create_sync_endpoint_function(endpoint_config: dict) -> Callable:
extra_body: dict[str, Any] | None = None,
**kwargs,
):
local_vars = locals()
local_vars: Final = locals()
try:
resolved_custom_llm_provider: str = custom_llm_provider
litellm_logging_obj: LiteLLMLoggingObj = kwargs.pop("litellm_logging_obj")
litellm_call_id: str | None = kwargs.get("litellm_call_id")
_is_async = kwargs.pop("async_call", False) is True
litellm_logging_obj: Final[LiteLLMLoggingObj] = kwargs.pop("litellm_logging_obj")
litellm_call_id: Final[str | None] = kwargs.get("litellm_call_id")
_is_async: Final = kwargs.pop("async_call", False) is True
# Check for mock response
mock_response = kwargs.get("mock_response")
@ -99,7 +99,7 @@ def create_sync_endpoint_function(endpoint_config: dict) -> Callable:
raise ValueError(f"Container provider config not found for: {resolved_custom_llm_provider}")
# Build optional params for logging
optional_params = {k: kwargs.get(k) for k in path_params if k in kwargs}
optional_params: Final = {k: kwargs.get(k) for k in path_params if k in kwargs}
# Pre-call logging
litellm_logging_obj.update_from_kwargs(
@ -150,12 +150,12 @@ def create_async_endpoint_function(
extra_body: dict[str, Any] | None = None,
**kwargs,
):
local_vars = locals()
local_vars: Final = locals()
try:
loop = asyncio.get_event_loop()
loop: Final = asyncio.get_event_loop()
kwargs["async_call"] = True
func = partial(
func: Final = partial(
sync_func,
timeout=timeout,
custom_llm_provider=custom_llm_provider,
@ -165,9 +165,9 @@ def create_async_endpoint_function(
**kwargs,
)
ctx = contextvars.copy_context()
func_with_context = partial(ctx.run, func)
init_response = await loop.run_in_executor(None, func_with_context)
ctx: Final = contextvars.copy_context()
func_with_context: Final = partial(ctx.run, func)
init_response: Final = await loop.run_in_executor(None, func_with_context)
if asyncio.iscoroutine(init_response):
response = await init_response
@ -193,8 +193,8 @@ def generate_container_endpoints() -> dict[str, Callable]:
Returns a dict mapping function names to their implementations.
"""
config = _load_endpoints_config()
endpoints = {}
config: Final = _load_endpoints_config()
endpoints: Final = {}
for endpoint_config in config["endpoints"]:
# Create sync function
@ -210,8 +210,8 @@ def generate_container_endpoints() -> dict[str, Callable]:
def get_all_endpoint_names() -> list[str]:
"""Get all endpoint names (sync and async) from config."""
config = _load_endpoints_config()
names = []
config: Final = _load_endpoints_config()
names: Final = []
for endpoint in config["endpoints"]:
names.append(endpoint["name"])
names.append(endpoint["async_name"])
@ -220,21 +220,21 @@ def get_all_endpoint_names() -> list[str]:
def get_async_endpoint_names() -> list[str]:
"""Get all async endpoint names for router registration."""
config = _load_endpoints_config()
config: Final = _load_endpoints_config()
return [endpoint["async_name"] for endpoint in config["endpoints"]]
# Generate endpoints on module load
_generated_endpoints = generate_container_endpoints()
_generated_endpoints: Final = generate_container_endpoints()
# Export generated functions dynamically
list_container_files = _generated_endpoints.get("list_container_files")
alist_container_files = _generated_endpoints.get("alist_container_files")
upload_container_file = _generated_endpoints.get("upload_container_file")
aupload_container_file = _generated_endpoints.get("aupload_container_file")
retrieve_container_file = _generated_endpoints.get("retrieve_container_file")
aretrieve_container_file = _generated_endpoints.get("aretrieve_container_file")
delete_container_file = _generated_endpoints.get("delete_container_file")
adelete_container_file = _generated_endpoints.get("adelete_container_file")
retrieve_container_file_content = _generated_endpoints.get("retrieve_container_file_content")
aretrieve_container_file_content = _generated_endpoints.get("aretrieve_container_file_content")
list_container_files: Final = _generated_endpoints.get("list_container_files")
alist_container_files: Final = _generated_endpoints.get("alist_container_files")
upload_container_file: Final = _generated_endpoints.get("upload_container_file")
aupload_container_file: Final = _generated_endpoints.get("aupload_container_file")
retrieve_container_file: Final = _generated_endpoints.get("retrieve_container_file")
aretrieve_container_file: Final = _generated_endpoints.get("aretrieve_container_file")
delete_container_file: Final = _generated_endpoints.get("delete_container_file")
adelete_container_file: Final = _generated_endpoints.get("adelete_container_file")
retrieve_container_file_content: Final = _generated_endpoints.get("retrieve_container_file_content")
aretrieve_container_file_content: Final = _generated_endpoints.get("aretrieve_container_file_content")

148
litellm/containers/main.py
View file

@ -3,7 +3,7 @@ import contextvars
import json
from collections.abc import Coroutine
from functools import partial
from typing import Any, Literal, overload
from typing import Any, Final, Literal, overload
import litellm
from litellm.constants import request_timeout as DEFAULT_REQUEST_TIMEOUT
@ -76,12 +76,12 @@ async def acreate_container(
Returns:
- `response` (ContainerObject): The created container object
"""
local_vars = locals()
local_vars: Final = locals()
try:
loop = asyncio.get_event_loop()
loop: Final = asyncio.get_event_loop()
kwargs["async_call"] = True
func = partial(
func: Final = partial(
create_container,
name=name,
expires_after=expires_after,
@ -94,9 +94,9 @@ async def acreate_container(
**kwargs,
)
ctx = contextvars.copy_context()
func_with_context = partial(ctx.run, func)
init_response = await loop.run_in_executor(None, func_with_context)
ctx: Final = contextvars.copy_context()
func_with_context: Final = partial(ctx.run, func)
init_response: Final = await loop.run_in_executor(None, func_with_context)
if asyncio.iscoroutine(init_response):
response = await init_response
@ -185,11 +185,11 @@ def create_container(
print(response)
```
"""
local_vars = locals()
local_vars: Final = locals()
try:
litellm_logging_obj: LiteLLMLoggingObj = kwargs.pop("litellm_logging_obj") # type: ignore
litellm_call_id: str | None = kwargs.get("litellm_call_id")
_is_async = kwargs.pop("async_call", False) is True
litellm_logging_obj: Final[LiteLLMLoggingObj] = kwargs.pop("litellm_logging_obj")
litellm_call_id: Final[str | None] = kwargs.get("litellm_call_id")
_is_async: Final = kwargs.pop("async_call", False) is True
# Check for mock response first
mock_response = kwargs.get("mock_response")
@ -197,12 +197,12 @@ def create_container(
if isinstance(mock_response, str):
mock_response = json.loads(mock_response)
response = ContainerObject(**mock_response)
response: Final = ContainerObject(**mock_response)
return response
# get llm provider logic
# Pass credential params explicitly since they're named args, not in kwargs
litellm_params = GenericLiteLLMParams(
litellm_params: Final = GenericLiteLLMParams(
api_key=api_key,
api_base=api_base,
api_version=api_version,
@ -218,12 +218,12 @@ def create_container(
local_vars.update(kwargs)
# Get ContainerCreateOptionalRequestParams with only valid parameters
container_create_optional_params: ContainerCreateOptionalRequestParams = (
container_create_optional_params: Final[ContainerCreateOptionalRequestParams] = (
ContainerRequestUtils.get_requested_container_create_optional_param(local_vars)
)
# Get optional parameters for the container API
container_create_request_params: dict = ContainerRequestUtils.get_optional_params_container_create(
container_create_request_params: Final[dict] = ContainerRequestUtils.get_optional_params_container_create(
container_provider_config=container_provider_config,
container_create_optional_params=container_create_optional_params,
)
@ -306,12 +306,12 @@ async def alist_containers(
Returns:
- `response` (ContainerListResponse): The list of containers
"""
local_vars = locals()
local_vars: Final = locals()
try:
loop = asyncio.get_event_loop()
loop: Final = asyncio.get_event_loop()
kwargs["async_call"] = True
func = partial(
func: Final = partial(
list_containers,
after=after,
limit=limit,
@ -324,9 +324,9 @@ async def alist_containers(
**kwargs,
)
ctx = contextvars.copy_context()
func_with_context = partial(ctx.run, func)
init_response = await loop.run_in_executor(None, func_with_context)
ctx: Final = contextvars.copy_context()
func_with_context: Final = partial(ctx.run, func)
init_response: Final = await loop.run_in_executor(None, func_with_context)
if asyncio.iscoroutine(init_response):
response = await init_response
@ -403,11 +403,11 @@ def list_containers(
Currently supports OpenAI
"""
local_vars = locals()
local_vars: Final = locals()
try:
litellm_logging_obj: LiteLLMLoggingObj = kwargs.pop("litellm_logging_obj") # type: ignore
litellm_call_id: str | None = kwargs.get("litellm_call_id")
_is_async = kwargs.pop("async_call", False) is True
litellm_logging_obj: Final[LiteLLMLoggingObj] = kwargs.pop("litellm_logging_obj")
litellm_call_id: Final[str | None] = kwargs.get("litellm_call_id")
_is_async: Final = kwargs.pop("async_call", False) is True
# Check for mock response first
mock_response = kwargs.get("mock_response")
@ -415,12 +415,12 @@ def list_containers(
if isinstance(mock_response, str):
mock_response = json.loads(mock_response)
response = ContainerListResponse(**mock_response)
response: Final = ContainerListResponse(**mock_response)
return response
# get llm provider logic
# Pass credential params explicitly since they're named args, not in kwargs
litellm_params = GenericLiteLLMParams(
litellm_params: Final = GenericLiteLLMParams(
api_key=api_key,
api_base=api_base,
api_version=api_version,
@ -435,7 +435,7 @@ def list_containers(
raise ValueError(f"Container provider config not found for provider: {custom_llm_provider}")
# Get container list request parameters
container_list_optional_params: ContainerListOptionalRequestParams = (
container_list_optional_params: Final[ContainerListOptionalRequestParams] = (
ContainerRequestUtils.get_requested_container_list_optional_param(local_vars)
)
@ -504,12 +504,12 @@ async def aretrieve_container(
Returns:
- `response` (ContainerObject): The container object
"""
local_vars = locals()
local_vars: Final = locals()
try:
loop = asyncio.get_event_loop()
loop: Final = asyncio.get_event_loop()
kwargs["async_call"] = True
func = partial(
func: Final = partial(
retrieve_container,
container_id=container_id,
timeout=timeout,
@ -520,9 +520,9 @@ async def aretrieve_container(
**kwargs,
)
ctx = contextvars.copy_context()
func_with_context = partial(ctx.run, func)
init_response = await loop.run_in_executor(None, func_with_context)
ctx: Final = contextvars.copy_context()
func_with_context: Final = partial(ctx.run, func)
init_response: Final = await loop.run_in_executor(None, func_with_context)
if asyncio.iscoroutine(init_response):
response = await init_response
@ -593,12 +593,12 @@ def retrieve_container(
Currently supports OpenAI
"""
local_vars = locals()
local_vars: Final = locals()
try:
resolved_custom_llm_provider: str = custom_llm_provider
litellm_logging_obj: LiteLLMLoggingObj = kwargs.pop("litellm_logging_obj") # type: ignore
litellm_call_id: str | None = kwargs.get("litellm_call_id")
_is_async = kwargs.pop("async_call", False) is True
litellm_logging_obj: Final[LiteLLMLoggingObj] = kwargs.pop("litellm_logging_obj")
litellm_call_id: Final[str | None] = kwargs.get("litellm_call_id")
_is_async: Final = kwargs.pop("async_call", False) is True
# Check for mock response first
mock_response = kwargs.get("mock_response")
@ -606,7 +606,7 @@ def retrieve_container(
if isinstance(mock_response, str):
mock_response = json.loads(mock_response)
response = ContainerObject(**mock_response)
response: Final = ContainerObject(**mock_response)
return response
# get llm provider logic
@ -625,7 +625,7 @@ def retrieve_container(
litellm_params=litellm_params,
)
# True when input was a LiteLLM-managed ID (any length); needed to re-encode output for routing affinity
was_encoded = original_container_id != container_id
was_encoded: Final = original_container_id != container_id
# get provider config
container_provider_config: BaseContainerConfig | None = ProviderConfigManager.get_provider_container_config(
@ -719,12 +719,12 @@ async def adelete_container(
Returns:
- `response` (DeleteContainerResult): The deletion result
"""
local_vars = locals()
local_vars: Final = locals()
try:
loop = asyncio.get_event_loop()
loop: Final = asyncio.get_event_loop()
kwargs["async_call"] = True
func = partial(
func: Final = partial(
delete_container,
container_id=container_id,
timeout=timeout,
@ -735,9 +735,9 @@ async def adelete_container(
**kwargs,
)
ctx = contextvars.copy_context()
func_with_context = partial(ctx.run, func)
init_response = await loop.run_in_executor(None, func_with_context)
ctx: Final = contextvars.copy_context()
func_with_context: Final = partial(ctx.run, func)
init_response: Final = await loop.run_in_executor(None, func_with_context)
if asyncio.iscoroutine(init_response):
response = await init_response
@ -808,12 +808,12 @@ def delete_container(
Currently supports OpenAI
"""
local_vars = locals()
local_vars: Final = locals()
try:
resolved_custom_llm_provider: str = custom_llm_provider
litellm_logging_obj: LiteLLMLoggingObj = kwargs.pop("litellm_logging_obj") # type: ignore
litellm_call_id: str | None = kwargs.get("litellm_call_id")
_is_async = kwargs.pop("async_call", False) is True
litellm_logging_obj: Final[LiteLLMLoggingObj] = kwargs.pop("litellm_logging_obj")
litellm_call_id: Final[str | None] = kwargs.get("litellm_call_id")
_is_async: Final = kwargs.pop("async_call", False) is True
# Check for mock response first
mock_response = kwargs.get("mock_response")
@ -821,7 +821,7 @@ def delete_container(
if isinstance(mock_response, str):
mock_response = json.loads(mock_response)
response = DeleteContainerResult(**mock_response)
response: Final = DeleteContainerResult(**mock_response)
return response
# get llm provider logic
@ -840,7 +840,7 @@ def delete_container(
litellm_params=litellm_params,
)
# True when input was a LiteLLM-managed ID (any length); needed to re-encode output for routing affinity
was_encoded = original_container_id != container_id
was_encoded: Final = original_container_id != container_id
# get provider config
container_provider_config: BaseContainerConfig | None = ProviderConfigManager.get_provider_container_config(
@ -938,12 +938,12 @@ async def alist_container_files(
Returns:
- `response` (ContainerFileListResponse): The list of container files
"""
local_vars = locals()
local_vars: Final = locals()
try:
loop = asyncio.get_event_loop()
loop: Final = asyncio.get_event_loop()
kwargs["async_call"] = True
func = partial(
func: Final = partial(
list_container_files,
container_id=container_id,
after=after,
@ -957,9 +957,9 @@ async def alist_container_files(
**kwargs,
)
ctx = contextvars.copy_context()
func_with_context = partial(ctx.run, func)
init_response = await loop.run_in_executor(None, func_with_context)
ctx: Final = contextvars.copy_context()
func_with_context: Final = partial(ctx.run, func)
init_response: Final = await loop.run_in_executor(None, func_with_context)
if asyncio.iscoroutine(init_response):
response = await init_response
@ -1037,12 +1037,12 @@ def list_container_files(
Currently supports OpenAI
"""
local_vars = locals()
local_vars: Final = locals()
try:
resolved_custom_llm_provider: str = custom_llm_provider
litellm_logging_obj: LiteLLMLoggingObj = kwargs.pop("litellm_logging_obj") # type: ignore
litellm_call_id: str | None = kwargs.get("litellm_call_id")
_is_async = kwargs.pop("async_call", False) is True
litellm_logging_obj: Final[LiteLLMLoggingObj] = kwargs.pop("litellm_logging_obj")
litellm_call_id: Final[str | None] = kwargs.get("litellm_call_id")
_is_async: Final = kwargs.pop("async_call", False) is True
# Check for mock response first
mock_response = kwargs.get("mock_response")
@ -1050,7 +1050,7 @@ def list_container_files(
if isinstance(mock_response, str):
mock_response = json.loads(mock_response)
response = ContainerFileListResponse(**mock_response)
response: Final = ContainerFileListResponse(**mock_response)
return response
# get llm provider logic
@ -1168,12 +1168,12 @@ async def aupload_container_file(
print(response)
```
"""
local_vars = locals()
local_vars: Final = locals()
try:
loop = asyncio.get_event_loop()
loop: Final = asyncio.get_event_loop()
kwargs["async_call"] = True
func = partial(
func: Final = partial(
upload_container_file,
container_id=container_id,
file=file,
@ -1185,9 +1185,9 @@ async def aupload_container_file(
**kwargs,
)
ctx = contextvars.copy_context()
func_with_context = partial(ctx.run, func)
init_response = await loop.run_in_executor(None, func_with_context)
ctx: Final = contextvars.copy_context()
func_with_context: Final = partial(ctx.run, func)
init_response: Final = await loop.run_in_executor(None, func_with_context)
if asyncio.iscoroutine(init_response):
response = await init_response
@ -1288,12 +1288,12 @@ def upload_container_file(
"""
from litellm.llms.custom_httpx.container_handler import generic_container_handler
local_vars = locals()
local_vars: Final = locals()
try:
resolved_custom_llm_provider: str = custom_llm_provider
litellm_logging_obj: LiteLLMLoggingObj = kwargs.pop("litellm_logging_obj") # type: ignore
litellm_call_id: str | None = kwargs.get("litellm_call_id")
_is_async = kwargs.pop("async_call", False) is True
litellm_logging_obj: Final[LiteLLMLoggingObj] = kwargs.pop("litellm_logging_obj")
litellm_call_id: Final[str | None] = kwargs.get("litellm_call_id")
_is_async: Final = kwargs.pop("async_call", False) is True
# Check for mock response first
mock_response = kwargs.get("mock_response")
@ -1301,7 +1301,7 @@ def upload_container_file(
if isinstance(mock_response, str):
mock_response = json.loads(mock_response)
response = ContainerFileObject(**mock_response)
response: Final = ContainerFileObject(**mock_response)
return response
# get llm provider logic

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