Keeps the original CustomStreamWrapper so response headers and the
correlation-context cleanup in __del__ are untouched when a deployment
hook rewrites the converted response. Covers the early-return branches
for real provider streams and unmapped call types
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
A pre-call deployment hook can turn a requested stream into a non-streaming provider call and the result is wrapped back into a fake stream. The async client wrapper treated that wrapper like a caller-requested stream and returned before async_post_call_success_deployment_hook, so SDK callers lost post-call deployment processing (including CustomGuardrail post_call enforcement) on converted streams. Run the hook on the complete ModelResponse behind the wrapper and rewrap a modified response so it reaches the emitted chunks.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Resolve the azure_ai common_utils conflict onto main's api_key_header_for_base
helper, gate native Responses routing on the resolved api_base host, derive the
/openai/v1/responses URL from the host and project prefix, keep websocket mode
on the managed emulation, and cover the routing end to end with respx
A deployment hook (Headroom, code interpreter, web search) can downgrade
kwargs["stream"] to False while the caller still expects to iterate the
result. The cache handler keyed stream replay and callback deferral off
the raw flag, so a cache hit returned a plain object to a caller that
iterates, and the Responses iterator never persisted the converted
stream in the first place. Key both off the conversion marker as well
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Deployment hooks such as Headroom downgrade stream=True to a non-streaming provider call and the agentic loop then hands back a CustomStreamWrapper (or MockResponsesAPIStreamingIterator for Responses). wrapper_async still saw kwargs["stream"] is False, so it took the non-streaming success path with a lazy stream object: no standard_logging_object was built, the proxy cost callback raised failed_tracking_spend, and the wrapper's own end-of-stream dispatch was deduped away. Treat a lazy stream result as streaming for logging regardless of the downgraded kwarg. Regression in v1.99.0 via #35017
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
With a post-call guardrail the proxy defers async success logging, and every nested wrapper on a
/v1/messages call bridged to the Responses API overwrote the stored closure, so the spend log was
built from the outermost Anthropic-shaped reply under Responses semantics and recorded the prompt
tokens without the cache hit. The first wrapper to exit now keeps the slot, which is the innermost
provider response, the same one the non-deferred path logs.
The flat cost path also billed cache-creation tokens at 0 when the model had no
cache_creation_input_token_cost. It now falls back to the input rate, and the 1h rate to the
creation rate, matching the tiered path and the custom pricing helper.
_client_async_logging_helper re-submitted logging_obj.success_handler to the
executor after _dispatch_success_logging had already done so, running the same
success pipeline twice per async request and racing on shared logging state.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Zeroes the remaining violations for 12 rules so they can hard-fail
in the main ruff config instead of being budget-ratcheted, and drops
their strict-gate budgets to 0:
- B021: drop useless f-prefix on the Javelin docstring
- C404 / C419: dict()/any() around unnecessary list comprehension
- PLR0124: replace the 'value == value' NaN idiom (and the separate
+/-inf exclusion) with math.isfinite in _validate_response_time
- SIM201: 'not X == "function"' -> 'X != "function"'
- SIM211: 'False if x is False else True' -> 'x is not False'
- SIM222: drop literal 'None or' before "success"
- UP036: remove the dead sys.version_info < (3, 8) branch (and the
now-unused sys import) in the weights_biases TYPE_CHECKING block
- B018 x2: keep the deliberate property side-effect access but assign
it ('_ = self.prompt_manager') as the rule requires
- PLR0206: the unusable '@property def api_version(self, api_version)'
(a property getter cannot take extra args) becomes a @staticmethod
matching its siblings get_api_base/get_api_key; it had no callers
- PLR1704: rename the loop variable (and the nested helper parameter)
that shadowed abatch_completion_fastest_response's 'model' argument
- B004 x2: scoped noqa with rationale — both sites retrieve __call__
to unwrap functors for iscoroutinefunction, which is a value use,
not the callability test B004 assumes; the callable() autofix would
break them
N999 intentionally stays on the ratchet (limit 1): it flags the
'litellm/proxy/lambda.py' filename, which needs a module rename.
Verified: full-tree 'ruff check litellm' green with the graduated
rules enforced; ruff-strict counts for all 12 rules are 0; budget
JSON regenerated in the gate script's json.dumps style.
Router.log_retry used to copy the failed attempt's kwargs and metadata into
metadata.previous_models. Nothing downstream read those copies, but they carried
client credentials into spend logs and grew the payload on every retry. Each
attempt now leaves a flat record (model group, deployment id, exception type and
string, attempt number), which drops RETRY_BREADCRUMB_EXCLUDED_KWARGS and the
per-retry credential masking.
num_retries_per_request was enforced from len(previous_models), which only
looked at the metadata bucket and never exceeded four records. The sync and
async client wrappers and the Rust lifecycle guard now read attempted_retries
from whichever metadata bucket the call carries.
Resolves LIT-7505
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
/model/info fills a deployment's missing pricing in from the model cost map so the
Admin UI has a rate to display. Clients echo that whole model_info blob back on save,
and update_db_model merged it into the row, so editing an unrelated setting turned
that day's catalog price into a real per-deployment override. After that the
deployment ignored the cost map and Reload Price Data could no longer move it,
because the reload replays each deployment's stored pricing over the fresh catalog.
Drop the derived pricing from incoming model_info on the two write paths. The
drop-set is read off the same objects the read path uses, CustomPricingLiteLLMParams
plus the tiered *_above_N_tokens pattern that get_model_info passes through and no
model declares, so it cannot drift as new rates are added. output_vector_size is
exempt: it lives on the pricing model but is an embedding dimension, not a rate.
A deployment's own pricing still rides litellm_params, which is untouched, as is the
explicit-null clear, which reads the incoming model rather than the filtered dict.
The filter sits in the endpoint bodies rather than _add_model_to_db, which master-key
rotation reuses to re-serialize every stored deployment.
Every proxy and router cost lookup goes through get_model_info, which copies
cost map keys explicitly, so the new audio cache-read branch always fell back
to the text cache-read rate there. Copy the key so models whose audio
cache-read rate differs from the text one bill cached audio correctly.
Gate the reasoning_effort translation on the cost-map flag or the model name containing
claude, so unmapped Claude serving endpoints keep translating. Flag the newer Claude
entries that were missing it. Expose supports_anthropic_thinking_payload as a public
helper next to the other supports_* wrappers instead of importing the private factory.
Drop the adaptive-only guard, since the adaptive flags only ever match Claude ids, and
add regression tests for an unmapped Claude endpoint and an adaptive Claude model
Replace the hand-rolled Meta realtime handler with a MetaRealtimeConfig
that plugs into the shared realtime handler and RealTimeStreaming relay.
Clients keep speaking the OpenAI realtime wire: session.update,
input_audio_buffer.append/commit and the OpenAI transcription events.
Unsupported transcription settings are logged and dropped, matching the
Gemini realtime precedent, and the Meta-specific session.mode, keywords,
language_bias, DIARIZATION and speaker extensions are removed.
Drop the MODEL_API_KEY env var in favor of the standard META_API_KEY,
remove the private-logging flag so spend logs record the transcript the
same way other realtime models do, and add per-second pricing for
muse-voice-transcribe-1.0.
The relay now sends raw bytes from transform_realtime_request straight to
the backend after pace_backend_send, and transcription sessions never
trigger response.create.
* feat(rust): count tiktoken cl100k_base admission tokens in Rust
The Rust admission token counter only had the Anthropic tokenizer, so every
other model (OpenAI gpt-4 family, Azure, Gemini, Bedrock non-Claude, Mistral)
tokenized with tiktoken on the Python inference worker.
Add an exact cl100k_base counter to litellm-token-counter: the vendored rank
file (base64 token / rank lines, the bytes Python's tiktoken uses) is parsed
into a byte-level BPE model and the cl100k split pattern is a handwritten
scanner over the shared Unicode classes, so no regex engine runs per request.
Both tokenizers share the message, tool and reply-priming accounting.
The PyO3 TokenCounter gains a from_cl100k_ranks constructor; Python reads the
rank file and passes it in, the way claude_json_str already works. The bridge
selects the counter through the same predicates litellm.token_counter uses
(huggingface_tokenizer_kind, openai_tokenizer_encoding), declines o200k_base,
downloaded HuggingFace and custom tokenizers to Python, and budget reservation
counts once per distinct tokenizer a request names.
The legacy gpt-3.5-turbo-0301 message accounting (4 per message, -1 per name)
stays in Python: the selector declines it through the predicate token_counter
itself uses.
* feat(rust): count tiktoken o200k_base admission tokens in Rust (#40794)
Add a handwritten o200k_base split scanner and TokenCounter::from_o200k_ranks
next to the cl100k_base counter, sharing MergeRanks and the request
accounting. The Python bridge selects it when openai_tokenizer_encoding
names o200k_base, so gpt-4o, gpt-4.1, gpt-5, o1/o3/o4 and chatgpt-4o
requests stop tokenizing on the Python worker under LITELLM_RUST=true
Co-authored-by: yassin <yassin@berri.ai>
---------
Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(logging): finish response metadata before the sync logging thread reads it
The async and sync client wrappers handed the response to the threaded success handler before computing its cost, call id, and api_base, so that thread inserted into the same metadata dict the request coroutine was still iterating and a finished chat completion turned into a 500 (dictionary changed size during iteration). Metadata is now finalized first, and the merge and header copies snapshot their dicts before iterating.
* fix(logging): snapshot metadata with a dict copy and drop redundant comment
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(logging): copy metadata via dict.copy and dedupe Final import
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* perf(mock): emit admission-time usage chunk on streaming mock_response
Streaming mock_response chunks carried no usage, so the chunk builder re-tokenized the whole prompt in Python after the stream ended even when budget reservation had already counted it at admission. The mock streaming generators now yield a final usage-only chunk carrying the admission prompt count (same completion count as the non-streaming path). Without an admission count the old tokenizer fallback stays.
* fix(mock): type the mock stream generators and keep the usage chunk on the content stream id
Review follow-up: the usage-only chunk was built with a fresh id, so CustomStreamWrapper switched response_id for the finish-reason and usage chunks. It now copies the content stream id. The generators also get full parameter and return annotations.
---------
Co-authored-by: yassin <yassin@berri.ai>
* feat(proxy): offload spend tracking to a pod-local spend worker sidecar
py-spy on the gateway showed the post-response _PROXY_track_cost_callback,
spend-log and DBSpendUpdateWriter work running on the inference workers'
event loop, so a DB or Redis stall backed up the request path.
When LITELLM_SPEND_WORKER_ENABLED=true, _ProxyDBLogger serializes one compact
typed SpendEvent per success and hands it to a SpendEventProducer that ships
it over a unix socket (default) or loopback-only TCP to a sidecar started as
`python -m gateway.spend_worker`. The sidecar runs the unchanged
_ProxyDBLogger pipeline against the pod's PgBouncer (pooled_database_url).
When the sidecar is unreachable, the buffer is full, or the gateway shuts
down with events still queued or in flight, the producer applies
LITELLM_SPEND_WORKER_ON_UNAVAILABLE (fallback in-process, or drop). The
sidecar half-closes producers on SIGTERM and drains, the producer treats
EOF as unavailable, and the gateway flushes buffered spend counters on
shutdown. The sidecar honors LITELLM_LOG so its writes are visible in its
own process log.
Helm: both charts gain an opt-in spend-worker sidecar container sharing an
emptyDir socket dir, and the componentized chart's HPA uses a
ContainerResource CPU metric scoped to the gateway container so sidecar
CPU does not drive inference scaling.
* feat(terraform): opt-in spend-worker sidecar for the AWS and GCP gateway stacks
Adds spend_worker_* inputs to both modules. On ECS Fargate the sidecar is a second, non-essential container in the gateway task; on Cloud Run it is a second container in the gateway service. Both listen on loopback TCP, share the gateway's DB/Redis/secret env, and set LITELLM_JOB_ROLE=spend_worker. Disabled by default. Plan-only tests cover both, and the terraform CI workflow now runs the gcp module too
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(proxy): retrieve a completed batch in the in-process spend path test
The base now defers cost tracking for batches that are still in flight, so an in_progress batch never reaches update_database
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(proxy): rename the spend worker sidecar to collector
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(proxy): run the collector from the installed litellm package and finish in-flight fallbacks on shutdown
The sidecar command becomes python -m litellm.proxy.collector so the classic image, whose runtime
stage copies only the installed package, can run it. The module now assembles DATABASE_URL and the
pod-local pgbouncer URL itself, replacing gateway/collector.py
The componentized collector sidecar inherits gateway.volumeMounts so custom CA mounts reach it.
SpendEventProducer shields an in-progress fallback from the writer task cancellation so close()
no longer loses an event already handed to the in-process pipeline
Helpers used across modules (address_argument, should_store_prompts_and_responses_in_spend_logs,
flush_spend_counters_on_shutdown) become public so the change adds no reportPrivateUsage errors
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* ci(terraform): drop the gcp job duplicated by the aws/gcp matrix
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(collector): keep metrics env off the classic sidecar and reject shared loopback ports
The classic chart no longer hands PROMETHEUS_METRICS_PORT and the billing metrics env to the collector container, and gives it the same /.npm scratch mount as the proxy on a read-only root. AWS and GCP now refuse a plan where the spend collector and the metrics sidecar bind the same loopback port. A regression test drives a sidecar crash mid-stream on asyncio and uvloop and checks no event is billed by both the sidecar and the in-process fallback; the producer docstring spells out why a failed drain() cannot double count
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* style(proxy): format pooled_database_url after the pgbouncer rebase
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(proxy): keep the cache-hit preset key and survive dead producers on collector drain
Cache hits updated the logging object after the early return, so the offloaded spend event carried
preset_cache_key=None and the collector re-hashed reconstructed kwargs. Also guard write_eof() against
producer transports uvloop already closed so one dead connection cannot abort the drain
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(terraform): keep the gcp collector port off the metrics sidecar health port
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(proxy): collector connects to Postgres directly under IAM or Entra token auth
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(proxy): mark the collector's DATABASE_URL as pooled when it uses the pod's pgbouncer
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
