A request whose model field matched no configured model was rejected with 400 but its failure row still persisted the raw client string as the model, so a client that concatenated its prompt into the model field wrote that prompt into LiteLLM_SpendLogs and the daily spend tables, where /user/daily/activity/aggregated returned it as a breakdown.models key. The spend log payload now records such rejections under the constant unknown-model, keeping the failed request counted without persisting client input as a model name.
An open breaker raised a generic Exception on every skipped call, and DualCache caught
it and logged a full ERROR traceback each time. Under load that became hundreds of
traceback formats per second on every replica and pinned the proxies at 100% CPU.
Raise a typed RedisCircuitBreakerOpenError instead and have DualCache return its
in-memory result without logging for it. Classify redis-py pool exhaustion
(ConnectionError chained from TimeoutError) as a timeout so a latency blip goes
through the duration gate. Track a breaker generation so a call admitted before the
breaker opened cannot close it, leaving that to the HALF_OPEN probe. Rate limit the
LoggingWorker callback-error traceback to one per interval so a stalled logging
backend cannot start a second traceback storm.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(rust_bridge): count budget-check input tokens in Rust on all LLM routes
Rust counts input tokens from the raw JSON body with the GIL released inside the existing budget reservation, covering every LLM route the auth dependency guards. It only fires for models on the Anthropic tokenizer when a budget is set, and Python counts whenever Rust is off, missing, or declines a body shape.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* perf(rust): count byte-level BPE tokens without the GPT-2 split regex (#40594)
The oniguruma run of the ByteLevel pre-tokenizer regex is about 90% of
encode_fast on a 100k token body (100 ms of the ~110 ms Rust admission
count in the gateway pod). A hand-written scanner that yields the same
pieces, then feeds the model directly, counts the same text in 10 ms.
It only engages for tokenizers with the Anthropic shape (optional NFKC,
ByteLevel without prefix space, no post-processor) and falls back to the
full encoder when the text contains an added token. Parity with
encode_fast is tested on random texts, the pieces are compared with the
real pre-tokenizer, and the \p{L}/\p{N}/\s tables are checked against
oniguruma for every code point.
NFKC runs through unicode-normalization-alignments, the crate and
Unicode tables NormalizedString::nfkc already uses, so the fast path
normalizes exactly what the full encoder would. Using the newer
unicode-normalization crate changed the count for 171 code points that
gained compatibility decompositions after Unicode 9 (U+32FF, U+A7F1..).
The fast normalizer is compared with the tokenizer's for every scalar
value and on random texts.
The scanner is built without mutable state: byte_char and mapped_len replace the const table builders and the reusable mapped buffer, and iter::successors replaces the stateful piece iterator. byte_chars_match_the_byte_level_alphabet checks the byte mapping against ByteLevel for every scalar value.
Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(rust_bridge): bound concurrent token-count encodes and share the Anthropic tokenizer predicate
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>
The writer health probe only ran SELECT 1, which a read-only Postgres
session answers fine, so a pooled connection left pointing at a demoted
primary kept failing every write with SQLSTATE 25006 until the pod was
restarted. Probe transaction_read_only instead, treat a 25006 on the
request path as a signal to recreate the client, and back off
exponentially while the database as a whole stays read-only so a replica
or an in-progress failover does not get its engine killed every cycle.
Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
A fix for a perf, memory, or crash regression names the released or rc
version it regressed in and carries the `backport-stable` label, so the
stable release gate cherry-picks it onto the rc line before tagging.
* feat(terraform): Prometheus metrics sidecar for the GCP Cloud Run gateway
gateway_metrics_port adds a metrics container running
litellm.proxy.prometheus_metrics_server next to the gateway, sharing the
PROMETHEUS_MULTIPROC_DIR over an in-memory volume, plus a Managed Service
for Prometheus collector sidecar that scrapes it over localhost and writes
to Cloud Monitoring. The gateway stays on port 4000 and the load balancer
routing is unchanged.
Resolves LIT-7502
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(terraform): reject fractional and collector health ports for gateway_metrics_port
Greptile review on #40614: 4000.5 fails integer port parsing and 13133 collides with the
gmp sidecar liveness listener. Also stop claiming the load balancer's own /metrics goes away
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>
A flat list[str] sent to voyage-context-4 is treated as independent inputs and forwarded flat with
enable_auto_chunking=True, chunk_size=32000, and input_type=document unless the caller already set
input_type=query. Caller-supplied params now override the defaults instead of being clobbered.
The voyage-4 family and voyage-context-4 cost map entries already exist on litellm_internal_staging,
and voyage-4-nano is not served by the Voyage API, so those additions and their pricing test are dropped.
While the Redis circuit breaker is open every guarded call was refused with a bare
Exception that each swallowing catch site logged as an ERROR traceback, so a
sub-second latency blip turned into thousands of tracebacks per minute and pinned
every replica's CPU. The sync guard also recorded socket timeouts as hard
connectivity failures, so with least-busy routing the breaker opened on the first
slow replies and the timeout-only min-duration guard never applied.
Refusals now raise RedisCircuitBreakerOpenError, and the catch sites route it
through log_redis_failure, which logs a refusal at DEBUG and everything else at
the caller's level. The sync guard passes is_timeout like the async one.
* feat(infra): scale gateway on per-pod RPM and TPM in Helm and Terraform
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(helm): require the metrics server before rendering the gateway ServiceMonitor
The http port serves /metrics/ behind virtual-key auth, so a ServiceMonitor
pointed at it only collects 401s and the RPM/TPM HPA metrics never appear
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(infra): express gateway HPA, KEDA and ECS workload targets per second
Rename the per-pod request and token targets in both Helm charts and the
AWS module from per minute to per second, and shorten the recommended
Prometheus rate window to [1m] with no * 60 so the adapter and KEDA
signals are what the HPA compares against. ECS keeps CloudWatch's
60-second aggregation: the ALB target is 60x the per-second variable and
the token metric math divides the period Sum by 60 before dividing by
the running task count.
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>
Azure realtime defaulted to the beta upstream whenever realtime_protocol
was not configured, so a GA client's session.update (session.type,
output_modalities, nested audio) was forwarded unchanged to
/openai/realtime and Azure rejected it with "Unknown parameter:
'session.type'" on gpt-realtime and gpt-realtime-1.5. The unset default
now follows the client the way the OpenAI handler already does: a client
that sends OpenAI-Beta: realtime=v1 keeps the beta upstream, any other
client gets /openai/v1/realtime. An explicit realtime_protocol in
litellm_params or LITELLM_AZURE_REALTIME_PROTOCOL still wins.
The Admin UI and the docs already present Organizations as an enterprise
feature, but every /organization route served unlicensed proxies. A router
level dependency now enforces the license on all of them, and it resolves
the auth dependency first so a bad key still gets 401 rather than 403.
Claude-Session: https://claude.ai/code/session_01Se8ERtsqMQ3eVzWiLVMyNS
Profiling the sidecar-enabled gateway at 700 rps showed ~2.4% of all samples
in get_model_group_info called per request from budget reservation, plus
get_deployment_model_info for tiered pricing tables. Both are read-only lookups
over the model list, so serve them from the Router's lru caches and clear the
deployment cache alongside the group cache when the model list changes.
The deployment-info cache is a per-router lru_cache built in __init__ rather
than a class-level decorated method, so it does not pin Router instances in a
process-wide cache and is dropped with the router.
A price data reload replaces litellm.model_cost without touching model_list, so
the reload replay also clears both caches; otherwise reservation would keep
pricing against the old catalog until an unrelated model-list change.
Co-authored-by: yassin <yassin@berri.ai>
* feat(mock): report admission-time input token count in mock_response usage
Mock completions always reported prompt_tokens=10, so spend tracking, TPM metrics, budgets and the tokens-per-minute autoscaling signal saw 10 tokens for a 100k-token request. Budget reservation now carries the admission-time input token count in the reservation record, and mock_completion reads it back so mock traffic exercises the same spend and TPM paths as real traffic without any extra tokenizer work.
* fix(mock): keep a zero admission input token count instead of falling back to 10
---------
Co-authored-by: yassin <yassin@berri.ai>
The squash of #40512 onto a base that already carried #40514 left the Partial call on the metadata pre-read error path only, so a rejected /key/update again persisted the planned values into state and TestResourceKeyUpdateFailureKeepsPriorState fails on the default branch.
Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
