The cost callback used to look for an existing `<batch id>_batch_cost` row before charging a
completed batch, which left a window where concurrent retrieves on any instance all charged the
key, and it would honor a row any request had written under that id. The spend update writer now
inserts the batch cost row itself with `create_many(skip_duplicates=True)` and only the retrieve
whose insert lands charges the key, team, and user. An existing row only takes the charge when it
is a successful `aretrieve_batch` row, so a client-chosen `x-litellm-call-id` on another endpoint
cannot suppress billing. Batch cost rows no longer get their own immediate flush path
`batch_cost_is_final` now treats the proxy's normalized `complete` status like `completed`, which
the enterprise batch cost poller relies on when it decides whether a completed batch is safe to
retire. Tests build that status with `model_copy` since the OpenAI `Batch` model rejects it
The `test-quality-ok` markers sit on the `patch(` lines the gate keys on, and the logging tests no
longer wrap the priced retrieve in `contextlib.suppress`
Every retrieve of a batch through the proxy shares one spend row, the batch id
plus the batch cost suffix, and spend log inserts skip duplicates. A poll that
landed while the batch was still validating or in progress wrote that row at
$0 and no later retrieve could overwrite it, and every completed retrieve after
the first added the cost to the key, team, and user counters again with no new
row to show for it.
The cost callback now writes nothing for a batch retrieve until the batch is
final, releasing the poll's budget reservation instead, and once it is final it
charges only when no spend row for that batch is queued for flush or already
stored. Batch cost rows are flushed to the database right away so a second
instance sees them, and the logger prices a batch only once it is final, which
also covers a failed batch that never produced an output file.
The test-quality gate counts every patch of a litellm internal against a ceiling, and the three patches this test needs pushed it over. The callback imports increment_spend_counters, update_cache and proxy_logging_obj from proxy_server inside its own body, so there is no seam to inject fakes through; every other test in this file uses the same three patches for the same reason
Claude-Session: https://claude.ai/code/session_01EX13mWex6RaBo9PYnkAtFW
The proxy cost callback zeroed response_cost whenever cache_hit was true. That rule dates from Jan 2024 when it was the only place cache hits were priced. The logging layer has priced the LLM share at 0 on a cache hit since Aug 2024, and since guardrail cost joined the standard logging payload the proxy-side zeroing has thrown away a real provider charge: a pre_call guardrail runs before the cache is consulted, so a cached response still cost whatever the guardrail billed. Drop the redundant zeroing so the payload's response_cost, which is already LLM 0 + guardrail cost, reaches spend logs, daily tables and budgets untouched
Claude-Session: https://claude.ai/code/session_01EX13mWex6RaBo9PYnkAtFW
The one-time seed for a budget window row subtracted the batch's own
LiteLLM_SpendLogs rows by request_id, and request_id is the client's
x-litellm-call-id whenever the response carries no id of its own. Carrying
that set through the queue meant an unbounded, client-controlled aggregate
that the commit-failure requeue kept alive across retries.
Every log row at or after a batch's earliest start is owed by an increment
that still reaches the row, so summing only rows before it needs nothing
from the request. That drops request_ids end to end and closes the
cross-pod double count the id list could not see.
The four model access group callback tests now share one helper, so nine
patches of proxy_server internals become three, and both mock-echo assertions
go with them. The delete_access_group tests share a context manager for the
same reason.
test_group_exactly_at_its_max_budget_passes gained the assertion it was
missing: it now proves the group reached the spend comparison, which a group
skipped for a missing budget row would not. The route-allowed patch beside it
was dead, so it is gone.
What is left is suppressed with the collaborator each one cannot inject.
The database writer already intersects the auth-matched groups with the ones
the served deployment declares, but the live spend counters got the unnarrowed
set. A caller granted two pools that both cover a model group debited both
counters while only one row moved, so the in-memory ceiling could block a pool
its persisted spend never touched.
Narrow once at the callback so both consumers read the same set.
Two problems, both caught in review.
The new table only landed in the root schema.prisma. Client generation reads
litellm/proxy/schema.prisma and packaging reads the copy under
litellm-proxy-extras, so the generated client had no
litellm_modelaccessgroupbudgettable and every budget read and write against
it would have failed at runtime. The root is the source of truth; both
copies are now byte-identical to it.
Nothing incremented spend:model_access_group:{group} after a call. Only the
reservation path ever wrote it, so with disable_budget_reservation the
read-time check was reading a counter nobody maintained and falling back to
the row's spend, which is cached for up to DEFAULT_MODEL_ACCESS_GROUP_CACHE_TTL.
A caller could run well past the pool inside that window, which is precisely
the case the read-time check exists to cover.
increment_spend_counters now takes the matched groups and charges them
through _init_and_increment_unreserved_spend_counter, so a group already
covered by a reservation is skipped rather than counted twice. The cost
callback sources the names with get_request_model_access_groups, the same
reader the spend writer uses.
request_id can be chosen by the client through x-litellm-call-id, so an
unbounded NOT (request_id = ANY(batch)) let a replayed old id drop that id's
historical LiteLLM_SpendLogs row from the one-time seed while its increment
still landed. The increment now carries the request start, the batch keeps
the earliest one, and the seed only excludes ids whose startTime is at or
after it.
`queued` is the Interactions API's not-started-yet status. It was in neither
the pollable set nor the terminal one, so a create returning it got no poll
task, counted as a response with nothing to charge for, and released its
budget reservation: billed nowhere, alerting nobody.
Poll it alongside `in_progress`, and pin the union of the pollable and
terminal sets against the generated spec enum so a status Google adds later
fails CI rather than shipping another unbilled path. A give-up on a status in
neither set now names the status and logs at error, instead of the warning
that reads as an interaction merely still running.
Also drop `model_id` and `litellm_model_name` from the settled body next to
the foreign `response_cost` already dropped there. All three come from the
poll's own throwaway client call, and left in place the two identity fields
overwrite the create's real deployment in the payload every logging
integration reads.
Rewrites the callback's per-status test to assert the observable outcome
(reservation held vs released) across all eight statuses rather than
comparing the gate to the function it delegates to, and pins the shipped
5-10-20-40-60 poll backoff and its timeout cutoff.
Silencing the cost-tracking alert for every usage-less interaction response
went one status too far. An interaction that stopped at failed, cancelled,
incomplete or budget_exceeded genuinely has nothing to charge for, so alerting
on it is noise. completed and requires_action are different: both mean the
model produced output, so a usage block is always expected, and one arriving
without it means the charge for real work was lost. That is precisely the case
failed_tracking_alert exists to surface, and swallowing it would let an
operator's interactions bill nothing with no signal that anything was wrong.
The status knowledge lives next to the other status predicates rather than in
the proxy callback. The reservation is still released on both paths, since
suppressing the alert was never what freed it.
The tightened gate correctly stopped deferring the reservation release for
InteractionsAPIResponses the scheduler will not poll (terminal status, or
in_progress without an id), but the response then fell through into the
generic 'Cost tracking failed' raise and the failed_tracking_alert path.
A create returning failed, cancelled, requires_action, incomplete,
budget_exceeded, or an id-less in_progress without usage therefore released
its reservation as intended and, in the same breath, alerted operators for a
legitimate no-usage response, both creating noise and masking real
cost-tracking failures.
The two gates are now nested under a single 'unbilled interaction response'
outer check, so any InteractionsAPIResponse with no usage takes either the
defer path (pollable, polling on) or the release-and-return path, and none
of them fall through to the generic failure raise. The two regression tests
also now assert failed_tracking_alert is not called, closing the observation
gap the report flagged.
Two review findings, both in the handoff between the create's success
callback and the background poll task.
The callback deferred its budget reservation release for any interactions
response with no usage, but the scheduler only starts a poll task when the
status is in_progress and an id is present. A create that came back
terminal without usage therefore matched the callback's test, got no poll
task, and left its reservation open forever: the pre-call estimate stayed
added to the key, user, team and org spend counters, and the key began
refusing traffic against budget it had never spent. The two conditions now
come from one shared gate so they cannot drift apart again.
Settling a background interaction re-runs the success handlers for a second
result on the same request, and OTEL dedupes span emission on a marker held
in that request's metadata. The in-progress create claimed the marker, so
the completion, the only event carrying usage and cost, was dropped as a
duplicate by OTEL and by every integration deriving from it. Clearing the
success-scoped markers alongside the existing dedup flag lets the cost span
through, leaving failure and guardrail markers untouched.
* test: drop the cwd-relative sys.path.insert calls from the test suite
TQ003 stands at 1,077 across 1,058 files, and 1,015 of them are the same shape:
sys.path.insert(0, os.path.abspath("../..")) and its deeper siblings. The
argument resolves against the working directory rather than the file, so from
the repo root, where every job runs pytest, it inserts the directory two levels
above the checkout. It has never pointed at litellm. The package is installed
into the environment anyway, which is what actually makes the import work, and
what the rule's message has said all along.
Removing them leaves 1,634 imports of sys and os with no remaining reference,
and those go too, except where another test module imports the name back out of
the file. The rest of TQ003 is 62 call sites that resolve against __file__ or a
variable, which are a different question and are left alone.
Collection is identical either way: 45,871 tests and the same 51 pre-existing
collection errors before and after, and ruff reports no new undefined name.
* test: drop the duplicate imports the sys.path sweep exposed to F811
* test(pre-call-utils): restore the os import the new bedrock tests need
* test: use monkeypatch.setenv for env writes in tests/test_litellm
`os.environ["X"] = v` inside a test leaks the value into every test that runs
after it in the same worker, so ordering decides the result. 262 of those
writes across 40 files now go through pytest's `monkeypatch` fixture, which
restores the previous value at teardown.
The rewrite skips any test that a mock.patch-family decorator wraps, any test
with defaulted positional parameters, any test whose own name is called
directly elsewhere, and rebinds nothing inside nested defs, because in each of
those cases appending a fixture parameter changes what pytest or mock binds.
Ratchets the TQ004 ceiling from 768 to 506.
* fix(test): delete the key through monkeypatch instead of popping it first
Five tests popped a key straight out of `os.environ`, ran, then restored it with
`monkeypatch.setenv`. By the time monkeypatch saw the name it was already gone,
so it recorded "absent" as the value to go back to and deleted the key at
teardown. On a worker that inherited a real `RESEND_API_KEY`, `SENDGRID_API_KEY`,
`UI_PASSWORD`, `LITELLM_SALT_KEY` or `OPENAI_API_KEY`, every test after the first
one ran without it.
`monkeypatch.delenv(..., raising=False)` removes the key and restores whatever
was there, so the try/finally the manual restore needed goes with it.
* chore(test): leave the two cost-calc files to the PR that rewrites them fully
Both files are also in #37815, which converts the module-global writes as well
as the env writes and folds them into one fixture. Two PRs rewriting the same
lines differently is a conflict nobody benefits from resolving, so this one
drops back to staging on those two and keeps the other 39.
TQ004 clears 200 here instead of 275; the rest moves with #37815.
A test that asserts on the error inside its own except block passes when the
call stops raising, because nothing runs the handler. That is the exact case
the test exists to catch, so the regression lands green.
Rewrites all 111 such blocks into pytest.raises, which fails when the call
succeeds, and selects PT017 in ruff-tests.toml so no new one lands.
A name bound twice keeps only the second binding. In `tests/` that is nearly
always a repeated import, harmless but misleading, and the same rule is what
catches the cases that are not harmless: a local that shadows an import the
module still calls, and a second `def test_x` that quietly replaces the first.
311 of the 344 sites were repeated imports and came out with ruff's own fix.
The remaining 33 needed a decision. Four modules imported a name they never
used because a local definition below already shadowed it. Two comprehensions
bound `call` over `unittest.mock.call`, which those modules import and use.
One test rebound the two module handles its nested reload closure had captured.
One class attribute shadowed an unused `status` import.
The load-test fixtures move to a conftest, which is how pytest is meant to share
them, so the test module no longer imports three fixture names it never calls.
The nine `prisma_client` parameters keep a narrow `noqa`: pytest resolves that
fixture by name before the body runs, so the parameter never shadows anything.
`pytest.raises(Exception)` with no `match=` passes on any error that broad. A
TypeError from a refactor, a botched fixture, an import that moved: all of them
read as the rejection the test claims to police, so the test goes green for the
wrong reason and stays green after the behaviour it guards is gone.
PT011 closes that gap for the 317 sites B017 could not reach, because B017 only
fires on a single-statement body with no `as e` binding. Each pattern here is the
message the code actually raised, recorded by running the sites under a plugin
that logged the concrete type and text per call site, so the assertions describe
observed behaviour rather than a guess. Where a site raises more than one message
across its parametrize cases, the pattern is an alternation of what was seen;
where the exception carries an empty `str()` and puts the text on `.message`, the
site keeps a narrow `noqa` with the reason.
PT014 removes four parametrize cases that were listed twice. The duplicate re-runs
an assertion that already passed, and it usually marks a case someone meant to
vary and forgot to edit.
A Redis over-limit verdict now survives a failing rollback DECRBY instead of
escaping into the in-memory fallback and granting tokens the counter already
rejected; the unrolled increments expire with the TTL. pop_reservation now
falls through to the local record when the Redis pop succeeds but finds
nothing, so a reservation saved in memory after a transient Redis save
failure still refunds on cancel or completion.
In-memory grants now record an owner token, and a refund only debits local
counters when the popping worker is the one that granted them, so a terminal
response handled elsewhere can no longer shrink another worker's unrelated
fallback reservations. A Redis-granted refund that fails no longer falls back
to decrementing local counters either: the leaked Redis increments expire
with the TTL and only tighten the allowance.
Reserve-script failures now roll back the scopes already incremented before
re-raising into the in-memory fallback, so a partial redis outage no longer
leaks counter increments that shrink the shared allowance. Reservations
record which backend granted them, so a refund never debits redis counters
an in-memory grant did not charge. Terminal-status matching is now
case-insensitive because the Bedrock async-invoke retrieve path returns raw
AWS-cased statuses like Completed.
- scale batch output-token reservations by the row's n / best_of candidate count
- parse client-supplied output caps defensively instead of 500ing on unparseable values
- exclude project IO descriptors from the first should_rate_limit pass when TPM
reservation is disabled so their buckets are not double-charged
Price ApplyGuardrail usage units recorded by PR #37225 with a new
bedrock/guardrails entry in the model cost map (regional override via
bedrock/{region}/guardrails), add the per-request guardrail_cost to the
standard logging payload's response_cost and CostBreakdown, surface it in
the x-litellm-response-cost header, and bill blocked requests through the
failure hook so key and team budgets see what AWS bills
Image, file, video, and previous_response_id requests reserved the whole
project ITPM limit up front, so any window with existing usage rejected
them and one in-flight multimodal request blocked the entire project.
Reserve the token_counter estimate instead, like every other request;
post-call reconciliation already charges actual usage.
Embeddings rows were identified by body shape (has `input`, no
`messages`/`prompt`), which also matches a `/v1/responses` batch row
and reserved zero output tokens for it -- letting a project caller run
large Responses generations against a quota-limited model without
consuming OTPM. Classify embeddings by the row's own `url` instead,
and read `max_output_tokens` as a Responses output cap alongside
`max_tokens`/`max_completion_tokens`.
Co-authored-by: Cursor <cursoragent@cursor.com>
Resolves conflicts from the upstream merge and addresses the Veria-AI
review comment on this PR: batch rows could bypass a project's
per-model ITPM/OTPM quota when the batch's file-bound/routing model
had no quota configured. Charges each row's own model against its own
project quota instead of only the routing model's, and fixes rate
limit error messages to attribute the correct model via a new
descriptor_value field on RateLimitStatus/AtomicCounterMeta. Also
re-syncs the ruff-strict, type-discipline, and basedpyright budgets
against the correct (non-stale) merge base.
Co-authored-by: Cursor <cursoragent@cursor.com>
The TPM pre-call reservation read `max_tokens or max_completion_tokens`, so a
request declaring both was charged for whichever field came first. A caller
sending `max_tokens=1` with `max_completion_tokens=10000` reserved 2 tokens and
was then free to consume ten thousand, since the provider honours the modern
field and litellm's own param mapping drops the legacy one for the gpt-5 and
o-series families.
Reserve against the larger of the declared budgets instead. Over-reserving is
the safe direction for a limiter: post-call reconciliation refunds the
difference between the reservation and actual usage, while under-reserving lets
the window be exceeded before anything notices.
The connection-level pre-call hook only ran once per WebSocket
connection, so a project caller could send unlimited high-token
response.create frames after a single minimal reservation. Adds
enforce_project_io_token_quota_for_frame to the v3 rate limiter and
wires it into both the native and managed WebSocket handlers via a
duck-typed litellm.callbacks lookup, so the SDK layer stays free of
proxy imports. A rejected frame gets an error event; the connection
stays open for the client to retry.
Also fixes the RET504 and BLE001 strict-lint-budget violations the
litellm_internal_staging merge introduced in
parallel_request_limiter_v3.py, which were failing the lint check.
