* fix(azure.py): ensure max_retries=0 is respected
Fixes https://github.com/BerriAI/litellm/issues/6129
* fix(test_openai.py): add unit test to ensure openai sdk calls always respect max_retries = 0
* test(test_azure_openai.py): add unit testing for azure_text/ route
* fix(azure.py): fix passing max retries on streaming
* fix(azure.py): fix azure max retries on async completion + streaming
* fix(completion/handler.py): fix azure text async completion + streaming
* test(test_azure_openai.py): ensure azure openai max retries always respected
* test(test_azure_o_series.py): add testing to ensure max retries always respected
* Added gemini providers for 2.0-flash and 2.0-flash lite (#8321)
* Update model_prices_and_context_window.json
added gemini providers for 2.0-flash and 2.0-flash light
* Update model_prices_and_context_window.json
fixed URL
---------
Co-authored-by: Krish Dholakia <krrishdholakia@gmail.com>
* Convert tool use arguments to string before counting tokens (#6989)
In at least some cases the `messages["tool_calls"]["function"]["arguments"]` is a dict, not a string. In order to tokenize it properly it needs to be a string. In the case that it is already a string this is a noop, which is also fine.
* build(model_prices_and_context_window.json): add gemini 2.0 flash lite pricing
* build(model_prices_and_context_window.json): add gemini commercial rate limits
* fix(utils.py): fix linting error
* refactor(utils.py): refactor to maintain function size
---------
Co-authored-by: Bardia Khosravi <bardiakhosravi95@gmail.com>
Co-authored-by: Josh Morrow <josh@jcmorrow.com>
* fix(convert_dict_to_response.py): only convert if response is the response_format tool call passed in
Fixes https://github.com/BerriAI/litellm/issues/8241
* fix(gpt_transformation.py): makes sure response format / tools conversion doesn't remove previous tool calls
* refactor(gpt_transformation.py): refactor out json schema converstion to base config
keeps logic consistent across providers
* fix(o_series_transformation.py): support o3 mini native streaming
Fixes https://github.com/BerriAI/litellm/issues/8274
* fix(gpt_transformation.py): remove unused variables
* test: update test
* initial transform for invoke
* invoke transform_response
* working - able to make request
* working get_complete_url
* working - invoke now runs on llm_http_handler
* fix unused imports
* track litellm overhead ms
* working stream request
* sign_request transform
* sign_request update
* use has_async_custom_stream_wrapper property
* use get_async_custom_stream_wrapper in base llm http handler
* fix make_call in invoke handler
* fix invoke with streaming get_async_custom_stream_wrapper
* working bedrock async streaming with invoke
* fix make call handler for bedrock
* test_all_model_configs
* fix test_bedrock_custom_prompt_template
* sync streaming for bedrock invoke
* fix _add_stream_param_to_request_body
* test_async_text_completion_bedrock
* fix transform_request
* fix get_supported_openai_params
* fix test supports tool choice
* fix test_supports_tool_choice
* add unit test coverage for bedrock invoke transform
* fix location of transformation files
* update import loc
* fix bedrock invoke unit tests
* fix import for max completion tokens
* refactor(deepseek/): move deepseek to base llm http handler
Fixes https://github.com/BerriAI/litellm/issues/8128#issuecomment-2635430457
* fix(gpt_transformation.py): support stream parsing for gpt-like calls
* test(test_deepseek_completion.py): add async streaming test
* fix(gpt_transformation.py): fix import
* fix(gpt_transformation.py): return full api base and content type
* test(base_llm_unit_tests.py): add test to ensure drop params is respected
* fix(types/prometheus.py): use typing_extensions for python3.8 compatibility
* build: add cherry picked commits
* fix(vertex_ai/gemini/transformation.py): handle 'http://' image urls
* test: add base test for `http:` url's
* fix(factory.py/get_image_details): follow redirects
allows http calls to work
* fix(codestral/): fix stream chunk parsing on last chunk of stream
* Azure ad token provider (#6917)
* Update azure.py
Added optional parameter azure ad token provider
* Added parameter to main.py
* Found token provider arg location
* Fixed embeddings
* Fixed ad token provider
---------
Co-authored-by: Krish Dholakia <krrishdholakia@gmail.com>
* fix: fix linting errors
* fix(main.py): leave out o1 route for azure ad token provider, for now
get v0 out for sync azure gpt route to begin with
* test: skip http:// test for fireworks ai
model does not support it
* refactor: cleanup dead code
* fix: revert http:// url passthrough for gemini
google ai studio raises errors
* test: fix test
---------
Co-authored-by: bahtman <anton@baht.dk>
* fix(o_series_transformation.py): add 'reasoning_effort' as o series model param
Closes https://github.com/BerriAI/litellm/issues/8182
* fix(main.py): ensure `reasoning_effort` is a mapped openai param
* refactor(azure/): rename o1_[x] files to o_series_[x]
* refactor(base_llm_unit_tests.py): refactor testing for o series reasoning effort
* test(test_azure_o_series.py): have azure o series tests correctly inherit from base o series model tests
* feat(base_utils.py): support translating 'developer' role to 'system' role for non-openai providers
Makes it easy to switch from openai to anthropic
* fix: fix linting errors
* fix(base_llm_unit_tests.py): fix test
* fix(main.py): add missing param
* fix: support azure o3 model family for fake streaming workaround (#8162)
* fix: support azure o3 model family for fake streaming workaround
* refactor: rename helper to is_o_series_model for clarity
* update function calling parameters for o3 models (#8178)
* refactor(o1_transformation.py): refactor o1 config to be o series config, expand o series model check to o3
ensures max_tokens is correctly translated for o3
* feat(openai/): refactor o1 files to be 'o_series' files
expands naming to cover o3
* fix(azure/chat/o1_handler.py): azure openai is an instance of openai - was causing resets
* test(test_azure_o_series.py): assert stream faked for azure o3 mini
Resolves https://github.com/BerriAI/litellm/pull/8162
* fix(o1_transformation.py): fix o1 transformation logic to handle explicit o1_series routing
* docs(azure.md): update doc with `o_series/` model name
---------
Co-authored-by: byrongrogan <47910641+byrongrogan@users.noreply.github.com>
Co-authored-by: Low Jian Sheng <15527690+lowjiansheng@users.noreply.github.com>
* add support for using llama spec with bedrock
* fix get_bedrock_invoke_provider
* add support for using bedrock provider in mappings
* working request
* test_bedrock_custom_deepseek
* test_bedrock_custom_deepseek
* fix _get_model_id_for_llama_like_model
* test_bedrock_custom_deepseek
* doc DeepSeek-R1-Distill-Llama-70B
* test_bedrock_custom_deepseek
* feat(databricks/chat/transformation.py): add tools and 'tool_choice' param support
Closes https://github.com/BerriAI/litellm/issues/7788
* refactor: cleanup redundant file
* test: mark flaky test
* test: mark all parallel request tests as flaky
* docs: cleanup doc
* feat(bedrock/): initial commit adding bedrock/converse_like/<model> route support
allows routing to a converse like endpoint
Resolves https://github.com/BerriAI/litellm/issues/8085
* feat(bedrock/chat/converse_transformation.py): make converse config base config compatible
enables new 'converse_like' route
* feat(converse_transformation.py): enables using the proxy with converse like api endpoint
Resolves https://github.com/BerriAI/litellm/issues/8085
* refactor(factory.py): refactor async bedrock message transformation to use async get request for image url conversion
improve latency of bedrock call
* test(test_bedrock_completion.py): add unit testing to ensure async image url get called for async bedrock call
* refactor(factory.py): refactor bedrock translation to use BedrockImageProcessor
reduces duplicate code
* fix(factory.py): fix bug not allowing pdf's to be processed
* fix(factory.py): fix bedrock converse document understanding with image url
* docs(bedrock.md): clarify all bedrock document types are supported
* refactor: cleanup redundant test + unused imports
* perf: improve perf with reusable clients
* test: fix test
* fix(utils.py): initial commit fixing custom cost tracking
refactors out provider specific model info from `get_model_info` - this was causing custom costs to be registered incorrectly
* fix(utils.py): cleanup `_supports_factory` to check provider info, if model info is None
some providers support features like vision across all models
* fix(utils.py): refactor to use _supports_factory
* test: update testing
* fix: fix linting errors
* test: fix testing
* fix(base_utils.py): supported nested json schema passed in for anthropic calls
* refactor(base_utils.py): refactor ref parsing to prevent infinite loop
* test(test_openai_endpoints.py): refactor anthropic test to use bedrock
* fix(langfuse_prompt_management.py): add unit test for sync langfuse calls
Resolves https://github.com/BerriAI/litellm/issues/7938#issuecomment-2613293757
* fix(http_handler.py): support passing ssl verify dynamically and using the correct httpx client based on passed ssl verify param
Fixes https://github.com/BerriAI/litellm/issues/6499
* feat(llm_http_handler.py): support passing `ssl_verify=False` dynamically in call args
Closes https://github.com/BerriAI/litellm/issues/6499
* fix(proxy/utils.py): prevent bad logs from breaking all cost tracking + reset list regardless of success/failure
prevents malformed logs from causing all spend tracking to break since they're constantly retried
* test(test_proxy_utils.py): add test to ensure bad log is dropped
* test(test_proxy_utils.py): ensure in-memory spend logs reset after bad log error
* test(test_user_api_key_auth.py): add unit test to ensure end user id as str works
* fix(auth_utils.py): ensure extracted end user id is always a str
prevents db cost tracking errors
* test(test_auth_utils.py): ensure get end user id from request body always returns a string
* test: update tests
* test: skip bedrock test- behaviour now supported
* test: fix testing
* refactor(spend_tracking_utils.py): reduce size of get_logging_payload
* test: fix test
* bump: version 1.59.4 → 1.59.5
* Revert "bump: version 1.59.4 → 1.59.5"
This reverts commit 1182b46b2e.
* fix(utils.py): fix spend logs retry logic
* fix(spend_tracking_utils.py): fix get tags
* fix(spend_tracking_utils.py): fix end user id spend tracking on pass-through endpoints
We use the DEFAULT_REPLICATE_ constants for retry count and initial
delay. If the completion response returns status=processing, we
loop to retry.
Fixes https://github.com/BerriAI/litellm/issues/7900
Signed-off-by: BJ Hargrave <hargrave@us.ibm.com>
Co-authored-by: BJ Hargrave <bj@hargrave.dev>
* fix(bedrock/converse_handler.py): fix bedrock region name on async calls
* fix(utils.py): fix split model handling
Fixes bedrock cost calculation when region name is given
* feat(_health_endpoints.py): support health checking datadog integration
Closes https://github.com/BerriAI/litellm/issues/7921
* fix(types/utils.py): support returning 'reasoning_content' for deepseek models
Fixes https://github.com/BerriAI/litellm/issues/7877#issuecomment-2603813218
* fix(convert_dict_to_response.py): return deepseek response in provider_specific_field
allows for separating openai vs. non-openai params in model response
* fix(utils.py): support 'provider_specific_field' in delta chunk as well
allows deepseek reasoning content chunk to be returned to user from stream as well
Fixes https://github.com/BerriAI/litellm/issues/7877#issuecomment-2603813218
* fix(watsonx/chat/handler.py): fix passing space id to watsonx on chat route
* fix(watsonx/): fix watsonx_text/ route with space id
* fix(watsonx/): qa item - also adds better unit testing for watsonx embedding calls
* fix(utils.py): rename to '..fields'
* fix: fix linting errors
* fix(utils.py): fix typing - don't show provider-specific field if none or empty - prevents default respons
e from being non-oai compatible
* fix: cleanup unused imports
* docs(deepseek.md): add docs for deepseek reasoning model
* fix(proxy_server.py): fix get model info when litellm_model_id is set
Fixes https://github.com/BerriAI/litellm/issues/7873
* test(test_models.py): add test to ensure get model info on specific deployment has same value as all model info
Fixes https://github.com/BerriAI/litellm/issues/7873
* fix(usage.tsx): make model analytics free
Fixes @iqballx's feedback
* fix(fix(invoke_handler.py):-fix-bedrock-error-chunk-parsing): return correct bedrock status code and error message if chunk in stream
Improves bedrock stream error handling
* fix(proxy_server.py): fix linting errors
* test(test_auth_checks.py): remove redundant test
* fix(proxy_server.py): fix linting errors
* test: fix flaky test
* test: fix test
* fix(router.py): pass stream timeout correctly for non openai / azure models
Fixes https://github.com/BerriAI/litellm/issues/7870
* test(test_router_timeout.py): add test for streaming
* test(test_router_timeout.py): add unit testing for new router functions
* docs(ollama.md): link to section on calling ollama within docker container
* test: remove redundant test
* test: fix test to include timeout value
* docs(config_settings.md): document new router settings param
* fix(initial-test-to-return-api-timeout-value-in-openai-timeout-exception): Makes it easier for user to debug why request timed out
* feat(openai.py): return timeout value + time taken on openai timeout errors
helps debug timeout errors
* fix(utils.py): fix num retries extraction logic when num_retries = 0
* fix(config_settings.md): litellm_logging.py
support printing payload to console if 'LITELLM_PRINT_STANDARD_LOGGING_PAYLOAD' is true
Enables easier debug
* test(test_auth_checks.py'): remove common checks userapikeyauth enforcement check
* fix(litellm_logging.py): fix linting error
* fix(lm_studio/chat/transformation.py): Fix https://github.com/BerriAI/litellm/issues/7811
* fix(router.py): fix mock timeout check
* fix: drop model name from fallback args since it causes a conflict with the model=model that is provided later on. (#7806)
This error happens if you provide multiple fallback models to the completion function with model name defined in each one.
* fix(router.py): remove mock_timeout before sending to request
prevents reuse in fallbacks
* test: update test
* test: revert test change - wrong pr
---------
Co-authored-by: Dudu Lasry <david1542@users.noreply.github.com>
* refactor: initial commit for using separate sync vs. async transformation routes for bedrock
ensures no blocking calls e.g. when converting image url to b64
* perf(converse_transformation.py): make bedrock converse transformation async
asyncify's the bedrock message transformation - useful for handling image urls for bedrock
* fix(converse_handler.py): fix logging for async streaming
* style: cleanup unused imports
* build: ensure all regional bedrock models have same supported values as base bedrock model
prevents drift
* test(base_llm_unit_tests.py): add testing for nested pydantic objects
* fix(test_utils.py): add test_get_potential_model_names
* fix(anthropic/chat/transformation.py): support nested pydantic objects
Fixes https://github.com/BerriAI/litellm/issues/7755
* feat(pass_through_endpoints.py): fix anthropic end user cost tracking
* fix(anthropic/chat/transformation.py): use returned provider model for anthropic
handles anthropic `-latest` tag in request body throwing cost calculation errors
ensures we can be accurate in our model cost tracking
* feat(model_prices_and_context_window.json): add gemini-2.0-flash-thinking-exp pricing
* test: update test to use assumption that user_api_key_dict can get anthropic user id
* test: fix test
* fix: fix test
* fix(anthropic_pass_through.py): uncomment previous anthropic end-user cost tracking code block
can't guarantee user api key dict always has end user id - too many code paths
* fix(user_api_key_auth.py): this allows end user id from request body to always be read and set in auth object
* fix(auth_check.py): fix linting error
* test: fix auth check
* fix(auth_utils.py): fix get end user id to handle metadata = None
* fix(gemini/): support gemini 'frequency_penalty' and 'presence_penalty'
Closes https://github.com/BerriAI/litellm/issues/7748
* feat(proxy_server.py): new env var to disable prisma health check on startup
* test: fix test
* fix base aws llm
* fix auth with aws role
* test aws base llm
* fix base aws llm init
* run ci/cd again
* fix get_credentials
* ci/cd run again
* _auth_with_aws_role
* fix(gpt_transformation.py): fix response_format translation check for 4o models
Fixes https://github.com/BerriAI/litellm/issues/7616
* feat(key_management_endpoints.py): support 'temp_budget_increase' and 'temp_budget_expiry' fields
Allow proxy admin to grant temporary budget increases to keys
* fix(proxy/_types.py): enforce temp_budget_increase and temp_budget_expiry are always passed together
* feat(user_api_key_auth.py): initial working temp budget increase logic
ensures key budget exceeded error checks for temp budget in key metadata
* feat(proxy_server.py): return the key max budget and key spend in the response headers
Allows clientside user to know their remaining limits
* test: add unit testing for new proxy utils
Ensures new key budget is correctly handled
* docs(temporary_budget_increase.md): add doc on temporary budget increase
* fix(utils.py): remove 3.5 from response_format check for now
not all azure 3.5 models support response_format
* fix(user_api_key_auth.py): return valid user api key auth object on all paths
* build(model_prices_and_context_window.json): add azure o1 pricing
Closes https://github.com/BerriAI/litellm/issues/7712
* refactor: replace regex with string method for whitespace check in stop-sequences handling (#7713)
* Allows overriding keep_alive time in ollama (#7079)
* Allows overriding keep_alive time in ollama
* Also adds to ollama_chat
* Adds some info on the docs about this parameter
* fix: together ai warning (#7688)
Co-authored-by: Carl Senze <carl.senze@aleph-alpha.com>
* fix(proxy_server.py): handle config containing thread locked objects when using get_config_state
* fix(proxy_server.py): add exception to debug
* build(model_prices_and_context_window.json): update 'supports_vision' for azure o1
---------
Co-authored-by: Wolfram Ravenwolf <52386626+WolframRavenwolf@users.noreply.github.com>
Co-authored-by: Regis David Souza Mesquita <github@rdsm.dev>
Co-authored-by: Carl <45709281+capsenz@users.noreply.github.com>
Co-authored-by: Carl Senze <carl.senze@aleph-alpha.com>
* feat(main.py): initial commit for `/image/variations` endpoint support
* refactor(base_llm/): introduce new base llm base config for image variation endpoints
* refactor(openai/image_variations/transformation.py): implement openai image variation transformation handler
* fix: test
* feat(openai/): working openai `/image/variation` endpoint calls via sdk
* feat(topaz/): topaz sync image variation call support
Addresses https://github.com/BerriAI/litellm/issues/7593
'
* fix(topaz/transformation.py): fix linting errors
* fix(openai/image_variations/handler.py): fix passing json data
* fix(main.py): image_variation/
support async image variation route - `aimage_variation`
* fix(test_get_model_info.py): fix test
* fix: cleanup unused imports
* feat(openai/): add async `/image/variations` endpoint support
* feat(topaz/): support async `/image/variations` calls
* fix: test
* fix(utils.py): fix get_model_info_helper for no model info w/ provider config
handles situation where model info is not known but provider config exists
* test(test_router_fallbacks.py): mark flaky test
* fix: fix unused imports
* test: bump otel load test perf threshold - accounts for current load tests hitting same server
* feat(langfuse.py): log the used prompt when prompt management used
* test: fix test
* docs(self_serve.md): add doc on restricting personal key creation on ui
* feat(s3.py): support s3 logging with team alias prefixes (if available)
New preview feature
* fix(main.py): remove old if block - simplify to just await if coroutine returned
fixes lm_studio async embedding error
* fix(langfuse.py): handle get prompt check
* fix(vertex_ai/gemini/transformation.py): handle 'http://' in gemini process url
* refactor(router.py): refactor '_prompt_management_factory' to use logging obj get_chat_completion logic
deduplicates code
* fix(litellm_logging.py): update 'get_chat_completion_prompt' to update logging object messages
* docs(prompt_management.md): update prompt management to be in beta
given feedback - this still needs to be revised (e.g. passing in user message, not ignoring)
* refactor(prompt_management_base.py): introduce base class for prompt management
allows consistent behaviour across prompt management integrations
* feat(prompt_management_base.py): support adding client message to template message + refactor langfuse prompt management to use prompt management base
* fix(litellm_logging.py): log prompt id + prompt variables to langfuse if set
allows tracking what prompt was used for what purpose
* feat(litellm_logging.py): log prompt management metadata in standard logging payload + use in langfuse
allows logging prompt id / prompt variables to langfuse
* test: fix test
* fix(router.py): cleanup unused imports
* fix: fix linting error
* fix: fix trace param typing
* fix: fix linting errors
* fix: fix code qa check
* fix(streaming_chunk_builder_utils.py): add test for groq tool calling + streaming + combine chunks
Addresses https://github.com/BerriAI/litellm/issues/7621
* fix(streaming_utils.py): fix modelresponseiterator for openai like chunk parser
ensures chunk parser uses the correct tool call id when translating the chunk
Fixes https://github.com/BerriAI/litellm/issues/7621
* build(model_hub.tsx): display cost pricing on model hub
* build(model_hub.tsx): show cost per token pricing + complete model information
* fix(types/utils.py): fix usage object handling
* build(ui/): update ui
* fix: drop unsupported non-whitespace characters for real when calling… (#7484)
* fix: drop unsupported non-whitespace characters for real when calling anthropic with stop sequences
* test: add parameterized test for _map_stop_sequences method in AnthropicConfig
---------
Co-authored-by: Wolfram Ravenwolf <52386626+WolframRavenwolf@users.noreply.github.com>
