- Introduce `_CallbackCapabilities` dataclass and `ProxyLogging._callback_capabilities()` static method that inspects `litellm.callbacks` once and caches capability flags keyed on (list length, member ids); invalidates automatically when the callback list mutates without per-request iteration overhead
- Replace O(n) `litellm.callbacks` walks in `async_pre_call_hook`, `during_call_hook`, `async_post_call_streaming_iterator_hook`, `async_post_call_streaming_hook`, and `post_call_response_headers_hook` with fast-path exits when no relevant callbacks are registered
- Add `needs_iterator_wrap()` and `needs_per_chunk_streaming_hook()` instance methods to decouple iterator-level wrapping from per-chunk hook execution; avoids `get_response_string` materialization per chunk when no guardrail or chunk-hook callback is active
- Introduce `_fast_serialize_simple_model_response_stream()` using `orjson` for common single-choice text streaming chunks, bypassing the full Pydantic serializer; falls back to `model_dump_json` for tool calls, logprobs, usage, and provider-specific fields
- Add early-return in `_restamp_streaming_chunk_model` when downstream model already matches the requested model, avoiding unnecessary string comparisons on every chunk
- Fix stale zero-cost cache bug in `_is_model_cost_zero`: move the per-router `_zero_cost_cache` dict onto the `Router` instance and clear it in `_invalidate_model_group_info_cache` so in-place pricing updates via `upsert_deployment` immediately resume budget enforcement
- Add `scripts/benchmark_chat_completions_perf.py`: standalone async benchmarking tool with a mock OpenAI provider, LiteLLM proxy process management, non-streaming RPS, streaming TTFT, and full-stream latency measurements with repeat/median run support
- Add comprehensive unit tests covering capability detection, cache invalidation, fast-path correctness, zero-cost cache regression, and the no-callback streaming fast path
Co-authored-by: Yassin Kortam <yassinkortam@g.ucla.edu>
When fastest_response=true with comma-separated models, the response
model field was stamped with the entire comma-separated string. Now
uses the x-litellm-model-group header from the winning response to
return the correct model name.
Made-with: Cursor
- Azure Model Router transform_response: let parent extract actual model from raw response
- common_request_processing: skip model override for Azure Model Router requests
- proxy_server: skip streaming chunk model restamp for Azure Model Router
- Add _is_azure_model_router_request helper
- Add tests for non-streaming and streaming
Made-with: Cursor
* 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.
