Children spawned as [sys.executable, -c, ...] put the working directory first on sys.path, so under 'make test-rust-extension' a source checkout shadows the installed wheel and the child imports a litellm with no compiled extension. A shared helper spawns them with -I and asserts the child resolved the same litellm.__file__ as the parent, and a new TQ009 rule flags un-isolated sys.executable spawns.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Add an aws_textract OCR provider on the Rust route, with no Python path. The
detect-document-text model returns plain lines and analyze-document renders
layout and tables as markdown. Both use Textract's synchronous API, so a
multi-page PDF or TIFF is rejected with an error that names the single-page
limit. A call with no region fails instead of falling back to Bedrock's default
SigV4 covers the request body, and host hooks can rewrite that body before it
is sent. litellm-http now has OutboundRequest, which serializes the body once,
shows those bytes to a RequestSigner and is the only thing a route can send.
Chat, audio transcription and OCR build it after their hooks ran, so a callback
that redacts the body still produces a valid Bedrock or Textract signature
ChatCompletionsAuth and AudioTranscriptionAuth are replaced by
litellm_auth::RequestAuth, and one helper in core turns it into a signed or
unsigned request. Audio transcription now signs only the AWS header set and
rejects a forwarded header that SigV4 computes, the same as chat
The OCR catalog routes aws_textract as Rust required, and the dispatch context
reads the provider from the model prefix so a provider scoped rule can match
The OCR route called back into Python's get_secret_str for every env
fallback. With no secret manager configured that is os.environ behind a GIL
hop, and with one configured it blocked a tokio worker on vault I/O and also
sent the Azure and GCP identity variables, which Python reads with os.getenv,
to the vault. The other Rust routes already read the process environment.
Read the process environment here too. When litellm would read secrets from
a secret manager, decline the Rust route so the Python route serves the call
with the vault-backed keys
Python resolves the Vertex project and location as call params, then the
litellm.vertex_project / litellm.vertex_location globals, then env, and
Azure AD token refresh from litellm.enable_azure_ad_token_refresh alone.
Native OCR skipped the globals, so a config.yaml litellm_settings value
silently fell through to the credential's project and us-central1, and a
managed identity setup without an API key failed. The bridge now reads
them through a provider_defaults settings group into OcrSettings, and
VertexConfig / AzureAuthInputs slot them in at Python's precedence.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Python reads every provider credential fallback (MISTRAL_API_KEY,
AZURE_AI_API_KEY, AZURE_DOCUMENT_INTELLIGENCE_ENDPOINT, Azure AD and
Vertex env, ...) through get_secret_str, which consults the configured
key_management_system before os.environ. Native OCR read std::env
directly, so a key held only in the vault went missing and a stale env
copy silently won. OcrClient now carries an injected secret Lookup that
the connection exposes to providers and auth crates; the bridge backs it
with settings.secret -> get_secret_str, pure Rust keeps the process env.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
litellm-http now builds the rustls config itself, so one route-neutral place covers roots, the client certificate, ALPN, ssl_ecdh_curve and ssl_security_level. A curve picks the single key exchange group. A cipher string restricts the TLS 1.2 suites it names, and entries rustls cannot express, such as @SECLEVEL=1, are logged once and skipped.
user_url_validation and user_url_allowed_hosts are applied by the media fetcher. Document downloads honor the environment proxy whenever provider calls do, keeping the per-hop address check, and stay on the pinned resolver when no proxy applies.
AIOHTTP_SO_KEEPALIVE, AIOHTTP_TCP_KEEPIDLE, AIOHTTP_TCP_KEEPINTVL, AIOHTTP_TCP_KEEPCNT and AIOHTTP_KEEPALIVE_TIMEOUT map onto the client. A client= argument and a live SSLContext are ignored
litellm.user_url_validation and litellm.user_url_allowed_hosts are only implemented by the Python document fetcher, so an allowlisted internal document was rejected by the Rust route's network policy. The bridge now declines when either is changed from its default
Python's aiohttp transport reads HTTP(S)_PROXY on every request unless disable_aiohttp_trust_env is set, so the Rust clients now do the same instead of requiring aiohttp_trust_env. Transport error messages include reqwest's source chain, so a rejected certificate or refused connection is no longer reported as just 'error sending request'
Honor environment proxies whenever Python would use httpx (sync calls, HTTP/2, aiohttp disabled), apply the per-call ssl_verify argument, ignore empty or missing SSL env values the way http_handler.py does, expire pooled clients after an hour so rotated certificates reload, keep the client certificate off media downloads, and decline instead of raising when a litellm global has an unexpected type
Drop the core ocr() facade so VertexAuth and the http pool stay out of litellm-core's
public API, move the http Error enum to error.rs, and inject the media DNS resolver into
HttpClientPool instead of a per-call builder hook the cache key ignored.
The bridge now reads litellm.* HTTP settings only through litellm/rust_bridge/settings.py,
pinned by python_settings.json, while env overrides stay in Rust. This adds the Python
default User-Agent, parses string ssl_verify globals like get_ssl_verify, drops per-call
ssl_verify that Python OCR never honored, and removes the unused request_timeout.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every legacy callback call from callbacks-legacy now goes through one typed
Python shim, litellm.rust_bridge.legacy_callbacks, the only Python module
the crate reaches. Before, the crate called Logging methods, litellm.utils
hooks, the logging worker, the executor and several litellm globals
directly, and its tests retyped those signatures by hand, so an outdated
fake could accept a call the real code rejects. python_contract.json lists
each shim function's parameters: a Python test pins it to the real
signatures and a Rust test pins it to the Rust enum.
The lifecycle contract changes to match the Python @client wrapper:
- the driver emits CallEvent::Started before begin, so every host sees one
start time
- RequestContext carries the route-resolved api_key, so legacy pre_call and
post_call receive it, and post_call's additional_args match the Python OCR
path
- Passthrough and its re-aliasing are gone
- async deployment hooks always run, and the "no callbacks" shortcut that
skipped the logging payload is removed, as in the Python path
The OCR api_key is a SecretValue from the wire request onward, so Debug
output upstream of the callback contract cannot leak it.
host-python's RouteHost now classifies native failures once through
classify, and host ops return HostOpError. The OCR route host keeps main's
public errors by sending both through the existing Python map_failure.
The httpx.Response built for a Rust upstream failure had no request attached,
so constructing openai.AuthenticationError raised RuntimeError inside the
exception mapper and every bad-key OCR call surfaced as APIConnectionError 500
instead of AuthenticationError 401 (the Python path already returned 401)
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Extracted from #41733 without the router loop, the cache machine layer, streaming, or the
error, timeout and route-pruning work that moved to #41745
litellm-callbacks holds the contract a native call and its host share: Machine, HostOp,
CallEvent, the in-process run loop, and Passthrough, which is built only by comparing the
caller's inputs with the body the route sends, so a route can never mark a key it rewrote.
litellm-host-python (formerly python-interop) owns the CPython driver and the Execution
handle, and litellm-callbacks-legacy is the @client wrapper as the native call sees it:
function_setup, the deployment hooks, pre_call and post_call, the success and failure fan-out
and the deferred proxy release. OCR is the one route on it, and the old core and bridge
lifecycles are gone
The passthrough rule is the structural fix for the bug #41719 patched in core and #41716
reworks: an inlined remote document no longer counts as the caller's value, so the legacy
adapter never hands the caller's URL back into the body. core/tests/ocr/passthrough.rs pins
it for every route and document source, including that unchanged values stay passthrough,
and callbacks-legacy/tests/payload.rs pins the adapter side with a real pre_call callback
Python OCR integration tests that only exercised core behavior now live as Rust tests, so
tests/test_litellm_rust keeps the cases that need the full Python stack
Each route now has litellm/rust_bridge/<route>/{entrypoints,callbacks}.py and a
public dispatch module (litellm/chat_completions/dispatch.py,
litellm/responses/dispatch.py, litellm/messages/dispatch.py) that binds the
public call to the legacy Python signature, builds a frozen request, and asks
the runtime to pick Rust or Python from the catalog. The legacy implementations
stay in litellm/main.py, litellm/responses/main.py and the anthropic messages
handler, and litellm/__init__.py re-exports the dispatch names over them the
same way it already does for ocr
The per-handler shims in rust_bridge/chat_completions/native.py and
rust_bridge/messages/native.py are removed along with their call sites in the
anthropic and bedrock chat handlers and the http handler. The exception
mapping that every callbacks module repeated moves to rust_bridge/failures.py
and the signature binding helpers to rust_bridge/public_call.py
Move each route's bridge module under litellm/rust_bridge/<route>/ so a folder
means a Rust implementation exists while the catalog row says whether it is
used. OCR now keeps the Python implementation in litellm/ocr/main.py and the
Rust selection in litellm/ocr/rust.py, removing litellm/ocr/legacy.py
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Parse the switch with pydantic TypeAdapter(bool) so 1/true/yes/on and 0/false/no/off all work, and treat an unparseable value as unset instead of off. PYTHON_ONLY and RUST_REQUIRED still ignore both switches
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Replace the per-route enablement helpers (rust_enabled, rust_ocr_enabled, RUST_CHAT_COMPLETIONS_PROVIDERS, FallbackMode) with a single rule table in litellm/rust_bridge/catalog.py that maps a Context(route, provider, model, delivery) to one of four rollout tiers, and a pure decide() that turns tier plus process/env switches into a Decision. runtime.run/arun own the only fallback path: Python for PYTHON, native then Python on missing binding or admission decline for RUST_WITH_FALLBACK, raise for RUST_REQUIRED. OCR is the first route on the shared runtime; chat completions, Anthropic messages, and Responses websocket policy checks now read the catalog.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
num_retries_per_request has always capped the retries of one request with its fallback hops included. #40930 started reading the per-hop attempted_retries counter instead, and every fallback hop restarts that counter at zero, so a request could spend a fresh retry budget on each hop and the legacy fallback cap test started seeing the hop run.
Router.log_retry now also keeps request_retry_count on the request metadata, incremented on every retry and fallback hop and never truncated the way previous_models is, and max_retries_per_request_hit reads that count. The flat retry records, the litellm_metadata coverage and caps above four from #40930 stay as they are, and the legacy test goes back to its previous_models == 0 assertion.
Router.log_retry used to copy the failed attempt's kwargs and metadata into
metadata.previous_models. Nothing downstream read those copies, but they carried
client credentials into spend logs and grew the payload on every retry. Each
attempt now leaves a flat record (model group, deployment id, exception type and
string, attempt number), which drops RETRY_BREADCRUMB_EXCLUDED_KWARGS and the
per-retry credential masking.
num_retries_per_request was enforced from len(previous_models), which only
looked at the metadata bucket and never exceeded four records. The sync and
async client wrappers and the Rust lifecycle guard now read attempted_retries
from whichever metadata bucket the call carries.
Resolves LIT-7505
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(rust): count tiktoken cl100k_base admission tokens in Rust
The Rust admission token counter only had the Anthropic tokenizer, so every
other model (OpenAI gpt-4 family, Azure, Gemini, Bedrock non-Claude, Mistral)
tokenized with tiktoken on the Python inference worker.
Add an exact cl100k_base counter to litellm-token-counter: the vendored rank
file (base64 token / rank lines, the bytes Python's tiktoken uses) is parsed
into a byte-level BPE model and the cl100k split pattern is a handwritten
scanner over the shared Unicode classes, so no regex engine runs per request.
Both tokenizers share the message, tool and reply-priming accounting.
The PyO3 TokenCounter gains a from_cl100k_ranks constructor; Python reads the
rank file and passes it in, the way claude_json_str already works. The bridge
selects the counter through the same predicates litellm.token_counter uses
(huggingface_tokenizer_kind, openai_tokenizer_encoding), declines o200k_base,
downloaded HuggingFace and custom tokenizers to Python, and budget reservation
counts once per distinct tokenizer a request names.
The legacy gpt-3.5-turbo-0301 message accounting (4 per message, -1 per name)
stays in Python: the selector declines it through the predicate token_counter
itself uses.
* feat(rust): count tiktoken o200k_base admission tokens in Rust (#40794)
Add a handwritten o200k_base split scanner and TokenCounter::from_o200k_ranks
next to the cl100k_base counter, sharing MergeRanks and the request
accounting. The Python bridge selects it when openai_tokenizer_encoding
names o200k_base, so gpt-4o, gpt-4.1, gpt-5, o1/o3/o4 and chatgpt-4o
requests stop tokenizing on the Python worker under LITELLM_RUST=true
Co-authored-by: yassin <yassin@berri.ai>
---------
Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(rust_bridge): count budget-check input tokens in Rust on all LLM routes
Rust counts input tokens from the raw JSON body with the GIL released inside the existing budget reservation, covering every LLM route the auth dependency guards. It only fires for models on the Anthropic tokenizer when a budget is set, and Python counts whenever Rust is off, missing, or declines a body shape.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* perf(rust): count byte-level BPE tokens without the GPT-2 split regex (#40594)
The oniguruma run of the ByteLevel pre-tokenizer regex is about 90% of
encode_fast on a 100k token body (100 ms of the ~110 ms Rust admission
count in the gateway pod). A hand-written scanner that yields the same
pieces, then feeds the model directly, counts the same text in 10 ms.
It only engages for tokenizers with the Anthropic shape (optional NFKC,
ByteLevel without prefix space, no post-processor) and falls back to the
full encoder when the text contains an added token. Parity with
encode_fast is tested on random texts, the pieces are compared with the
real pre-tokenizer, and the \p{L}/\p{N}/\s tables are checked against
oniguruma for every code point.
NFKC runs through unicode-normalization-alignments, the crate and
Unicode tables NormalizedString::nfkc already uses, so the fast path
normalizes exactly what the full encoder would. Using the newer
unicode-normalization crate changed the count for 171 code points that
gained compatibility decompositions after Unicode 9 (U+32FF, U+A7F1..).
The fast normalizer is compared with the tokenizer's for every scalar
value and on random texts.
The scanner is built without mutable state: byte_char and mapped_len replace the const table builders and the reusable mapped buffer, and iter::successors replaces the stateful piece iterator. byte_chars_match_the_byte_level_alphabet checks the byte mapping against ByteLevel for every scalar value.
Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(rust_bridge): bound concurrent token-count encodes and share the Anthropic tokenizer predicate
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Adds a chat_completions route module to litellm-core, mirroring the messages
route, plus Anthropic Messages and Bedrock Converse provider configs. The
per-model `rust: true` opt-in now covers /chat/completions for both providers.
The core accepts an allowlisted subset (text conversations, non-streaming) and
returns CoreError::Unsupported for anything else, so tool calls, multimodal
content and streaming fall back to the Python path transparently.
Resolves LIT-5698
