* fix(hosted_vllm): route through base_llm_http_handler to support ssl_verify
The hosted_vllm provider was falling through to the OpenAI catch-all path
which doesn't pass ssl_verify to the HTTP client. This adds an explicit
elif branch that routes hosted_vllm through base_llm_http_handler.completion()
which properly passes ssl_verify to the httpx client.
- Add explicit hosted_vllm branch in main.py completion()
- Add ssl_verify tests for sync and async completion
- Update existing audio_url test to mock httpx instead of OpenAI client
* feat(hosted_vllm): add embedding support with ssl_verify
- Add HostedVLLMEmbeddingConfig for embedding transformations
- Register hosted_vllm embedding config in utils.py
- Add lazy import for embedding transformation module
- Add unit test for ssl_verify parameter handling
* fix(proxy): prevent provider-prefixed model leaks
Proxy clients should not see LiteLLM internal provider prefixes (e.g. hosted_vllm/...) in the OpenAI-compatible response model field.
This patch sanitizes the client-facing model name for both:
- Non-streaming responses returned from base_process_llm_request
- Streaming SSE chunks emitted by async_data_generator
Adds regression tests covering vLLM-style hosted_vllm routing for both streaming and non-streaming paths.
* chore(lint): suppress PLR0915 in proxy handler
Ruff started flagging ProxyBaseLLMRequestProcessing.base_process_llm_request() for too many statements after the hotpatch changes.
Add an explicit '# noqa: PLR0915' on the function definition to avoid a large refactor in a hotpatch.
* refactor(proxy): make model restamp explicit
Replace silent try/except/pass and type ignores with explicit model restamping.
- Logs an error when the downstream response model differs from the client-requested model
- Overwrites the OpenAI `model` field to the client-requested value to avoid leaking internal provider-prefixed identifiers
- Applies the same behavior to streaming chunks, logging the mismatch only once per stream
* chore(lint): drop PLR0915 suppression
The model restamping bugfix made `base_process_llm_request()` slightly exceed Ruff's
PLR0915 (too-many-statements) threshold, requiring a `# noqa` suppression.
Collapse consecutive `hidden_params` extractions into tuple unpacking so the
function falls back under the lint limit and remove the suppression.
No functional change intended; this keeps the proxy model-field bugfix intact
while aligning with project linting rules.
* chore(proxy): log model mismatches as warnings
These model-restamping logs are intentionally verbose: a mismatch is a useful signal
that an internal provider/deployment identifier may be leaking into the public
OpenAI response `model` field.
- Downgrade model mismatch logs from error -> warning
- Keep error logs only for cases where the proxy cannot read/override the model
* fix(proxy): preserve client model for streaming aliasing
Pre-call processing can rewrite request_data['model'] via model alias maps.\n\nOur streaming SSE generator was using the rewritten value when restamping chunk.model, which caused the public 'model' field to differ between streaming and non-streaming responses for alias-based requests.\n\nStash the original client model in request_data as _litellm_client_requested_model after the model has been routed, and prefer it when overriding the outgoing chunk model. Add a regression test for the alias-mapping case.
* chore(lint): satisfy PLR0915 in streaming generator
Ruff started flagging async_data_generator() for too many statements after adding model restamping logic.\n\nExtract the client-model selection + chunk restamping into small helpers to keep behavior unchanged while meeting the project's PLR0915 threshold.
The /health/services endpoint rejected datadog_llm_observability as an
unknown service, even though it was registered in the core callback
registry and __init__.py. Added it to both the Literal type hint and
the hardcoded validation list in the health endpoint.
* fix(vertex_ai): convert image URLs to base64 in tool messages for Anthropic
Fixes#19891
Vertex AI Anthropic models don't support URL sources for images. LiteLLM
already converted image URLs to base64 for user messages, but not for tool
messages (role='tool'). This caused errors when using ToolOutputImage with
image_url in tool outputs.
Changes:
- Add force_base64 parameter to convert_to_anthropic_tool_result()
- Pass force_base64 to create_anthropic_image_param() for tool message images
- Calculate force_base64 in anthropic_messages_pt() based on llm_provider
- Add unit tests for tool message image handling
* chore: remove extra comment from test file header
* fix(vertex_ai): replace custom model names with actual Vertex AI model names in passthrough URLs (#19948)
When the passthrough URL already contains project and location, the code
was skipping the deployment lookup and forwarding the URL as-is to Vertex AI.
For custom model names like gcp/google/gemini-2.5-flash, Vertex AI returned
404 because it only knows the actual model name (gemini-2.5-flash).
The fix makes the deployment lookup always run, so the custom model name
gets replaced with the actual Vertex AI model name before forwarding.
* add _resolve_vertex_model_from_router
* fix: get_llm_provider
* Potential fix for code scanning alert no. 4020: Clear-text logging of sensitive information
Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com>
---------
Co-authored-by: michelligabriele <gabriele.michelli@icloud.com>
Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com>
The regex in get_vertex_model_id_from_url() was using [^/:]+
which stopped at the first slash, truncating model names like
'gcp/google/gemini-2.5-flash' to just 'gcp'. This caused
access_groups checks to fail for custom model names.
Changed the pattern to [^:]+ to allow slashes in model names,
only stopping at the colon before the action (e.g., :generateContent).
* Cleanup code for user cli auth, and make sure not to prompt user for team multiple times while polling
* Adding tests
* Cleanup normalize teams some more
When Gemini uses implicit caching, it returns cachedContentTokenCount but
NOT cacheTokensDetails. Previously, text_tokens was not adjusted in this case,
causing costs to be calculated as if all tokens were non-cached.
This fix subtracts cachedContentTokenCount from text_tokens when no
cacheTokensDetails is present (implicit caching), ensuring correct cost
calculation with the reduced cache_read pricing.
The lazy loading implementation for encoding in __getattr__ was calling
tiktoken.get_encoding() directly without first setting TIKTOKEN_CACHE_DIR.
This caused tiktoken to attempt downloading the encoding file from the
internet instead of using the local copy bundled with litellm.
This fix uses _get_default_encoding() from _lazy_imports which properly
sets TIKTOKEN_CACHE_DIR before loading tiktoken, ensuring the local cache
is used.
- Add whitelist-based filtering for anthropic_beta headers
- Only allow Bedrock-supported beta flags (computer-use, tool-search, etc.)
- Filter out unsupported flags like mcp-servers, structured-outputs
- Remove output_format parameter from Bedrock Invoke requests
- Force tool-based structured outputs when response_format is used
Fixes#16726
Extend advanced-tool-use header translation to include Claude Sonnet 4.5
in addition to Opus 4.5 on Bedrock Invoke API.
When Claude Code sends the advanced-tool-use-2025-11-20 header, it now
gets correctly translated to Bedrock-specific headers for both:
- Claude Opus 4.5
- Claude Sonnet 4.5
Headers translated:
- tool-search-tool-2025-10-19
- tool-examples-2025-10-29
Fixes defer_loading validation error on Bedrock with Sonnet 4.5.
Ref: https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool
