The reservation path (PR #26845) atomically pre-fills `spend:user:{user_id}`
and admits at the strict-`<` boundary. The legacy `_PROXY_MaxBudgetLimiter`
pre-call hook re-reads the same counter with `>=`, so a reservation that
fills the counter to exactly `max_budget` (e.g. a request without a
`max_tokens` cap that falls back to reserving the smallest remaining
headroom) is rejected by the hook even though the reservation already
admitted it.
Skip the hook when the request's active `budget_reservation` covers
`spend:user:{user_id}`. The reservation is the source of truth for that
counter cross-pod; the legacy `>=` path remains in place for requests
without a reservation (e.g. paths that bypass the reservation entirely).
Reproduces as `tests/otel_tests/test_prometheus.py::test_user_budget_metrics`
on a fresh user with `max_budget=10` calling `fake-openai-endpoint` without
`max_tokens`. Adds focused unit coverage in
`tests/test_litellm/proxy/hooks/test_max_budget_limiter.py`.
The endpoint builder in BedrockCountTokensConfig.get_bedrock_count_tokens_endpoint
percent-encodes the model id as a single path segment (d4dd865b1a, path-traversal
hardening). Update the four endpoint-URL assertions in TestBedrockCountTokensEndpoint
to expect `amazon.nova-lite-v1%3A0` instead of the literal `:0`, matching production
behavior already covered by test_count_tokens_endpoint_encodes_model_id.
_normalize_operation_ids referenced HTTP_METHODS but only HTTP_METHOD_SUFFIXES is defined, raising NameError on snapshot generation and failing test_lazy_openapi_snapshot. The constant was renamed in an earlier merge without updating these two references; values are identical sets of HTTP method names.
Merge of #26845 kept the PR's _should_skip_budget_checks helper but lost staging's upgrade to _get_model_from_request_context, so zero-cost models resolved from request headers/query params no longer skipped budget checks. Route the helper through _get_model_from_request_context so this path matches the other 8 model-resolution sites in the file.
Remove explanatory comments that restated what the code already says.
Kept only those that document non-obvious external contracts (the aiohttp
record-path patch's reason for re-feeding the body, and the warning
messages inside save_cassette that reach the user).
Previous attempt wrote to sys.__stderr__ from the test fixture. Under
xdist, fixtures run inside worker subprocesses whose stderr is captured
by the controller and only released to the live log on test failure —
so passing tests' verdicts were silently swallowed.
Round-trip via report.user_properties: the worker-side fixture stashes
the verdict on user_properties, xdist serializes it onto the report,
and a controller-side pytest_runtest_logreport hook writes it via the
TerminalReporter (the same plugin that emits PASSED/FAILED markers).
TerminalReporter is resolved lazily on first hook call because it's
not yet registered when conftest's pytest_configure runs.
Verified locally in both serial and xdist modes.
The GitHub merge conflict resolver concatenated both test sets but left
`from litellm.proxy._lazy_openapi_snapshot import _normalize_operation_ids`
stranded between functions instead of at the top of the file.
Raw github serves application/octet-stream which OpenAI/Gemini reject
when LiteLLM fetches the URL client-side. jsDelivr serves the same
file with content-type: application/pdf. Pin to a commit SHA so the
asset is immutable and jsDelivr can cache it for a year.
Previously, the per-test [VCR HIT/MISS/...] line was written via
TerminalReporter.write_line from inside fixture teardown. Pytest
captures that stream by default and only surfaces it on FAILED tests
(under 'Captured stdout teardown'), so passing tests' verdicts were
invisible in CI logs and the user couldn't tell whether the cache
was working.
Write directly to sys.__stderr__ so the line bypasses pytest's
capture entirely. Under xdist each worker has its own __stderr__
which CircleCI aggregates into the live job log alongside the
PASSED/FAILED markers.
A non-admin scoped to ["model-a"] could call /health?model_id=id-b
(where id-b belongs to a deployment outside their scope) and the
background-cache code path would return id-b's cached health entry. The
helper returned {model_id} unconditionally, so the cache filter was
driven by an unvalidated id and the global cache leaked the entry — the
ternary `targeted_ids if not None else allowed_model_ids` skipped any
intersection with the caller's allowed deployments.
Make _resolve_targeted_model_ids walk the supplied model_list for both
the model and model_id branches. Callers pass an already-scoped list
(filtered to allowed model_names for non-admins, full list for admins),
so an out-of-scope model_id resolves to an empty set and the cache
filter drops every entry — matching the live path's existing behavior.
GitHub serves PDFs from raw.githubusercontent.com, github.com/.../raw/..., LFS, and Releases as application/octet-stream by deliberate anti-hotlinking policy. Anyone who passes a GitHub-hosted PDF URL as an OpenAI / Gemini / Bedrock file_id hits "unsupported MIME type 'application/octet-stream'" because _process_image_response inlines the URL with whatever Content-Type the server sent.
When the server-provided Content-Type is application/octet-stream or binary/octet-stream and the URL extension maps to a known MIME type (.pdf, .png, .jpg, etc.), trust the extension instead. Specific Content-Types (image/png, application/pdf) still win over the extension; the override only applies to generic binary types.
Also restores the Greptile SHA-pinned raw.githubusercontent.com URL on the file_id integration test so we test against the same hosting real users hit, no third-party CDN.
Anthropic's URL fetcher intermittently returns 400 'Unable to download
the file' for the Wikipedia URL the test was using. Point it at the
repo's existing tests/llm_translation/fixtures/dummy.pdf via raw
GitHub instead — small, deterministic, reliably fetchable.
With a stable URL the test no longer needs to be opted out of VCR;
remove it from the incompatible list so it can replay from cassette.
Resolve URL conflict: keep Greptile's commit-SHA pin for immutability, but route through jsDelivr (cdn.jsdelivr.net/gh/BerriAI/litellm@<sha>/...) so the response Content-Type is application/pdf instead of application/octet-stream. Without this, OpenAI / Gemini / Router PDF tests reject the inlined file_data with "unsupported MIME type 'application/octet-stream'".
Per review: `assert response.status_code != 503` is satisfied by 404,
500, or any other non-503 code, so a regression that returned the wrong
non-503 status would slip through. Switch to `== 200` so the assertions
verify the actual expected status, not just the absence of one specific
failure.
A test that produces non-deterministic request bodies (e.g. uuid in
the prompt) under record_mode=new_episodes never replays — every CI
run appends fresh unmatched episodes. The cassette grows unbounded
over time and silently inflates Redis (we observed one cassette at
22 episodes / ~860KB after ~5 CI runs).
Refuse the save when episode count exceeds MAX_EPISODES_PER_CASSETTE
so the pathology surfaces with a loud warning that points to the
opt-out fix instead of festering invisibly.
The previous URL switch to raw.githubusercontent.com fixed Anthropic's "Unable to download" failure but caused OpenAI / Gemini / Router PDF tests to fail with "unsupported MIME type 'application/octet-stream'": those providers download the URL and inline it as data:<Content-Type>;base64,..., and raw.githubusercontent.com serves PDFs as application/octet-stream.
jsDelivr proxies the same in-repo fixture (cdn.jsdelivr.net/gh/BerriAI/litellm@main/...) and returns the correct Content-Type: application/pdf, so all providers (Anthropic forwards the URL natively; OpenAI/Gemini/Bedrock fetch and inline) get the right MIME type without changing transformer code.
Some tests can't benefit from cassette replay because they assert on
state that only exists in the live provider between two calls (e.g.
prompt-cache propagation, intermittent provider quirks). Marking them
with @pytest.mark.vcr just wastes cycles trying to record cassettes
they will never replay against successfully.
Opt-out by nodeid suffix so subclassed/parametrized variants are
covered:
- ::test_prompt_caching — Anthropic/Bedrock prompt-cache propagation
isn't deterministic in the 0–1s window the test gives it.
- ::test_async_pdf_handling_with_file_id — flaky upstream Wikipedia
fetch through the Anthropic Files API.
- TestBedrockInvokeNovaJson::test_json_response_pydantic_obj —
Bedrock Nova returns tool_call vs JSON nondeterministically (other
providers' subclasses are healthy).
- ::test_bedrock_converse__streaming_passthrough — Bedrock streaming
response_cost calc returns None intermittently.
These tests keep their existing @pytest.mark.flaky retry behavior.
When use_background_health_checks is enabled, /health?model=foo returned
the full cached aggregate across every model — so an unhealthy foo
combined with any other healthy deployment kept healthy_count > 0 and
the targeted-503 path never fired.
Resolve the targeted model/model_id to a deployment-id set first
(mirroring perform_health_check's match-on-model_name-or-litellm_model
semantics) and narrow the cache to those IDs before _post_process
evaluates healthy_count, so the 503 contract holds for both the live
and cache code paths.
