Decode *managed_id*. Returns ``None`` for anything that is not a passthrough managed ID — raw OpenAI IDs, unified-endpoint IDs, garbage,

Flush the logging queue. Waits until every enqueued task has completed. ``queue.join()`` blocks on the queue's unfinished-task counte

Must return the base completion call args

Use this to create a new INTERNAL user with a budget. Internal Users can access LiteLLM Admin UI to make keys, request access to models.

Common utility function to process audio files for audio transcription APIs. Handles various input types: - File paths (str, os.PathLike

Count the number of tokens in a given text using a specified model. Args: model (str): The name of the model to use for tokenization. De

Transform LiteLLM data to CBF format, dropping records with zero successful_requests or invalid CZRNs.

Ensure mock_prisma.db.litellm_config exists with async methods (the conftest only stubs ``litellm_configtable`` — this is a different table).

Normalize the endpoint URL for a specific OpenTelemetry signal type. The OTLP exporters expect endpoints to use signal-specific path

Attempt to acquire the lock for a specific cron job using Redis. Uses the SET command with NX and EX options to ensure atomicity.

Arize is used mainly for LLM I/O tracing, sending router+caching metrics adds bloat to arize logs

Create streaming response, checking if the first chunk is an error. If the first chunk is an error, return a standard JSON error response.

Check if a record exists.

Find multiple records matching the criteria.

Get all custom loggers that are instances of the given class type

Raises Exception if Virtual Key is not allowed to call the route

Disconnect and reconnect the Prisma client with a new database URL. Kills the old engine subprocess directly (SIGTERM → SIGKILL) rather than

Build a request-like object whose same-origin is ``base_url``. ``get_request_base_url`` defers to ``request.base_url`` unless the caller is a

[Non-Blocking] Safely get the request headers. Caches the result on request.state to avoid re-creating dict(request.headers) per call. W

Key - the key in cache Value - float - the value you want to increment by Refresh_ttl - bool - if True, resets the Redis TT

- handles scenario where content is list and not string - content list is just text, and no images Motivation: mistral api + azure ai do

Generate config class dynamically from JSON configuration

Encode a file/batch ID with model routing information. Format: <prefix><base64(litellm:<original_id>;model,<model_name>)> The result pre

Returns the deployment based on routing strategy

Get the cache key for the given arguments. Args: **kwargs: kwargs to litellm.completion() or embedding() Return

Get the base URL for Vertex AI API calls. - ``global`` uses the global control plane host. - Multi-region geographies (e.g. ``us``, ``eu

Calls async_post_call_response_headers_hook on all CustomLogger callbacks. Merges all returned header dicts (later callbacks override

Prints the given `print_statement` to the console if `litellm.set_verbose` is True. Also logs the `print_statement` at the debug level using

Release the lock if the current pod holds it. Uses an atomic Lua compare-and-delete to prevent TOCTOU races where a stale own

Factory mirroring user_api_key_auth: SERVER span + url.path + http.route.

Call PANW Prisma AIRS API to scan content or a tool_event.

Get the OpenAI compatible provider info for Baseten

Update the litellm params for health check. - gets a short `messages` param for health check - adds a bounded `max_tokens` when the depl

Auto-apply ``pytest.mark.vcr`` to collected items. Skip semantics (in priority order): 1. ``vcr_disabled()`` — global env-var off-switch (``