Fixes#16613
The issue was caused by two test files having the same module name
(test_transformation.py) in different directories, which caused pytest
to fail with an import file mismatch error.
Changes:
- Renamed tests/test_litellm/llms/xai/responses/test_transformation.py
to test_xai_responses_transformation.py
- Renamed tests/test_litellm/llms/openai_like/chat/test_transformation.py
to test_openai_like_chat_transformation.py
Both files now have unique, descriptive names that reflect their
specific test purposes and prevent module name collisions.
* add NewModelGroupRequest
* add endpoint for create_model_group
* fix model_access_group_management_router
* add UpdateModelGroupRequest, info and delete
* fix model management tag
* fix validate_models_exist
* fix get_all_access_groups_from_db
* test_create_duplicate_access_group_fails
* test fixes
* fix working create access groups
* fix access group management endpoints
* add is db model checks for model access groups
* feat(openai): Add support for reasoning_effort='none' in GPT-5.1
OpenAI's GPT-5.1 introduced a new reasoning effort parameter 'none'
which replaces the previous 'minimal' setting for faster, lower-latency
responses. This is now the default setting for GPT-5.1.
Changes:
- Updated REASONING_EFFORT type to include 'none' value
- Added GPT-5.1, GPT-5-mini, and GPT-5-nano to documentation
- Updated docs to reflect 'none' as GPT-5.1's default reasoning effort
- Added test to verify reasoning_effort='none' passes through correctly
Fixes#16633
* feat(responses): Add support for reasoning_effort='none' in Responses API transformation
* Refactor proxy embeddings to use shared processor
- allow ProxyBaseLLMRequestProcessing to accept the aembedding route so embeddings requests reuse the base pipeline hooks
- route embeddings requests through base_process_llm_request, sharing logging, hook execution, retries, and header handling with chat/responses
- tighten token array decoding logic by using router deployment lookups and the unified error handler
* Fix: Correctly process embedding requests with token arrays
The `test_embedding_input_array_of_tokens` test was failing due to a regression that caused embedding requests with token arrays to be processed incorrectly. This prevented the `aembedding` function from being called as expected.
This was caused by a combination of three distinct issues:
1. In `litellm/proxy/common_request_processing.py`, the `function_setup` utility was called with `aembedding` as the `original_function` for embedding routes. This has been corrected to `embedding` to ensure proper request setup.
2. In `litellm/proxy/proxy_server.py`, a `TypeError` occurred because the `get_deployment` method was called with the `model_name` keyword argument instead of the expected `model_id`. This has been corrected. Additionally, the check for token arrays was improved to validate that all elements in the input subarray are integers.
3. In `litellm/proxy/litellm_pre_call_utils.py`, the check for the `enforced_params` enterprise feature was too strict. It blocked valid requests even when the `enforced_params` list was empty. The condition has been adjusted to trigger the check only for non-empty lists.
Finally, the `test_embedding_input_array_of_tokens` assertion was updated to be more robust. The previous `assert_called_once_with` was overly strict, causing failures when unrelated internal parameters were added to the function call. The test now first asserts that `aembedding` is called and then separately verifies the `model` and `input` arguments. This makes the test more resilient to future changes without sacrificing its ability to catch regressions.
* test: align proxy embedding assertions
Update the embedding proxy test to match the new request pipeline: keep the data the proxy builds, expect the extra control kwargs, let the post-call hook return the actual response, and assert the normalized 'embeddings' hook type. This proves the refactor still forwards metadata and returns the mocked payload.
* Update proxy exception test
The proxy now forwards additional kwargs (request_timeout, litellm_call_id, litellm_logging_obj) to llm_router.aembedding. The test needs to accept these to match the real call signature and keep validating the error path instead of the kwargs list.
* testing: unsure of this change
I don't remember why I changed this, will revert and see if any tests fail since the manual test isn't failing without it.
* fix: remove unrelated change
This change was not related to the embeddings refactor and actually belonged to a different branch.
* fix: support Anthropic tool_use and tool_result in token counter
* refactor(token_counter): add dynamic field inference for Anthropic content blocks
* test: Add additional tests
* make format
* Fix lint error
* Fix mypy narrow type lint errors
Implements support for reasoning_effort="none" parameter for Gemini models,
providing significant cost savings (up to 96% cheaper) by disabling thinking
budget while maintaining response quality.
Changes:
- Added "supports_reasoning": true to gemini-2.0-flash-thinking-exp-01-21 in model config
- Implemented mapping for reasoning_effort="none" to thinkingConfig {thinkingBudget: 0, includeThoughts: false}
- Added unit test to verify the mapping works correctly
Performance impact:
- Without reasoning_effort: ~313 tokens
- With reasoning_effort="none": ~12 tokens (96% cheaper)
Closes#16420
Co-authored-by: Krish Dholakia <krrishdholakia@gmail.com>
Fixes#16533
Before this fix, non-ASCII characters (Japanese, Spanish, Chinese, etc.)
in function call arguments were being escaped as Unicode sequences.
Example:
- Before: "やあ" → "\u3084\u3042"
- After: "やあ" → "やあ" (preserved)
Changes:
- Add ensure_ascii=False to json.dumps() in _transform_parts()
- Add test for Japanese and Spanish Unicode character preservation
This is not a breaking change as both formats are equivalent in JSON.
The fix improves readability and aligns with OpenAI's behavior.
* fix(passthrough): use VLLM passthrough config for hosted VLLM provider instead of raising an error
* test(passthrough): add tests for hosted VLLM passthrough configuration and routing
* fix account linking for CLA
Fixes#16486
## Problem
Callbacks configured via litellm.success_callback (e.g., PostHog, LangSmith)
were not being invoked for litellm.acompletion() in short-lived scripts.
The callbacks worked correctly for synchronous completions but async
completions would queue callbacks that were lost when the script exited
before GLOBAL_LOGGING_WORKER could process them.
Root cause: asyncio.run() closes the event loop immediately after the
async function completes, preventing the background worker from processing
queued callbacks.
## Solution
Implemented a two-level atexit handler approach:
1. GLOBAL_LOGGING_WORKER atexit handler (logging_worker.py):
- Processes remaining callbacks from queue before exit
- Creates new event loop to run pending coroutines synchronously
- Applies time and iteration limits to prevent blocking shutdown
2. Integration-specific atexit handlers (posthog.py as example):
- Flushes internal queue to external service
- Uses synchronous HTTP client for reliable delivery
- Each integration needs its own handler due to varying sync APIs
## Changes
- litellm/litellm_core_utils/logging_worker.py:
- Added _flush_on_exit() method
- Registered atexit handler in __init__
- Processes up to MAX_ITERATIONS_TO_CLEAR_QUEUE events
- Time-limited to MAX_TIME_TO_CLEAR_QUEUE seconds
- litellm/integrations/posthog.py:
- Added _flush_on_exit() method
- Registered atexit handler in __init__
- Groups events by credentials for batch sending
- Uses sync_client for reliable HTTP delivery
- tests/logging_callback_tests/test_posthog.py:
- Added test_async_callback_atexit_handler_exists()
- Added test_posthog_atexit_flushes_internal_queue()
- Added test_sync_callback_not_affected_by_atexit()
## Testing
- All existing tests pass
- Manual end-to-end testing confirms fix:
- Async events now arrive in PostHog
- Sync events continue working (no regression)
- Unit tests verify atexit handlers registered and functional
## Impact
- Fixes async callback delivery for ALL integrations using GLOBAL_LOGGING_WORKER
- No breaking changes - only adds missing functionality
- Sync path unchanged - no performance impact
* Use end user budget instead of key budget when creating new team
* Fixed implementation to use user's max budget from the UserTable instead of EndUserTable
* Address a bug where cloudzero spend is not sent to cloudzero if a spend
update happens
* revert change unrelated to PR
* use polars for mocking instead of sqlite
YutaSaito