scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-11-02 07:13:02 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 49
litellm/tests/unit/caching/test_caching.py
yuneng-jiang 24afa12253
chore(caching)!: remove the disk cache backend (#48114) 
* chore(caching)!: remove the disk cache backend

Drop the disk cache type and its diskcache dependency, along with the
caching install extra that existed only to pull it in. diskcache has not
had a release since August 2023

Cache now raises a ValueError listing the supported types when it gets a
type it does not know. Before, an unknown type built a Cache with no
backend, which only failed on first use

BREAKING CHANGE: cache_params type "disk" and the litellm[caching] extra
are gone. A proxy configured with type disk now refuses to start. Use
local or redis instead

* test(proxy): expect the unsupported cache type error from _init_cache

The test pinned the old fallthrough, where an unknown type left the cache
without a backend and the proxy hit an AttributeError reading it

* chore(deps): drop the diskcache scan ignore main extended

diskcache is no longer in the lock, so the ignore matches nothing

* fix(caching): ignore a leftover disk_cache_dir instead of crashing redis and s3

Dropping the disk_cache_dir parameter sent the key into **kwargs, which
the redis and s3 branches hand to their clients, so a config that moved
from the disk cache to redis or s3 and kept the key failed with a
TypeError. Accept the key again, ignore it, and log a warning

The unsupported type test now checks the whole supported list, and two
proxy docstrings no longer name disk as a cache backend

* fix(caching): warn about a leftover disk_cache_dir only once the cache is built

The warning ran before the type check, so a config with type disk logged
that the key was ignored and then failed on the type anyway. It now logs
only after a supported backend is built. The test also checks that a
cache without the key logs nothing
2026-10-29 18:02:31 -07:00

1593 lines
64 KiB
Python
Raw Blame History

import asyncio
import hashlib
import json
import logging
import re
import traceback
import uuid
from datetime import timedelta
from typing import Final
from unittest.mock import AsyncMock, MagicMock, call, patch
import httpx
import pytest
import respx
import litellm
from litellm import acompletion, aembedding, completion, embedding
import litellm.caching.redis_cache as redis_cache_module
from litellm._internal_context import current_service_target
from litellm.caching.caching import Cache, CacheMode, response_cache_phase
from litellm.caching.caching_handler import _PENDING_CACHE_WRITES
from litellm.caching.dual_cache import DualCache
from litellm.caching.in_memory_cache import InMemoryCache
from litellm.caching.redis_cache import RedisCache, _RedisTimeoutLogThrottle
from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER
from litellm.types.caching import EMBEDDING_CACHE_FORMAT_VERSION, LiteLLMCacheType, SemanticCacheScope
from litellm.types.utils import Embedding, EmbeddingResponse, Usage
_CACHING_TEST_MESSAGES: Final = [{"role": "user", "content": "who is ishaan 5222"}]
messages = [{"role": "user", "content": "who is ishaan 5222"}]
@pytest.fixture
def preserve_litellm_set_verbose(monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setattr(litellm, "set_verbose", litellm.set_verbose)
@pytest.fixture
def redis_client_without_network(monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setattr("redis.Redis.info", lambda *args, **kwargs: {"redis_version": "7.2.0"})
monkeypatch.setattr("redis.Redis.ping", lambda *args, **kwargs: True)
monkeypatch.setattr("redis.asyncio.Redis.ping", AsyncMock(return_value=True))
def test_cache_key_debug_log_does_not_include_prompt_material(caplog):
cache = Cache(type=LiteLLMCacheType.LOCAL)
prompt_marker = "secret prompt material "
with caplog.at_level(logging.DEBUG, logger="LiteLLM"):
cache_key = cache.get_cache_key(
model="gpt-4.1-mini",
messages=[
{"role": "system", "content": prompt_marker * 100},
{"role": "user", "content": "hello"},
],
tools=[
{
"type": "function",
"function": {
"name": "lookup",
"parameters": {
"type": "object",
"properties": {"query": {"type": "string"}},
},
},
}
],
response_format={
"type": "json_schema",
"json_schema": {
"name": "lookup_response",
"schema": {"type": "object"},
},
},
stream=True,
)
assert re.fullmatch(r"[0-9a-f]{64}", cache_key)
created_cache_key_logs = [
record.getMessage() for record in caplog.records if "Created cache key:" in record.getMessage()
]
assert created_cache_key_logs
assert all(prompt_marker not in message for message in created_cache_key_logs)
assert any(cache_key in message for message in created_cache_key_logs)
@pytest.mark.parametrize(
("backend", "expected_level"),
[
pytest.param(MagicMock(spec=RedisCache), logging.DEBUG, id="redis_backend_is_throttled"),
pytest.param(MagicMock(), logging.ERROR, id="other_backend_logs_every_timeout"),
],
)
def test_add_cache_timeout_only_joins_redis_throttle_for_redis_backends(backend, expected_level, caplog, monkeypatch):
throttle = _RedisTimeoutLogThrottle(interval=5.0, clock=MagicMock(return_value=1_000.0))
assert throttle.admit() == 0
monkeypatch.setattr(redis_cache_module, "_redis_timeout_log_throttle", throttle)
cache = Cache(type=LiteLLMCacheType.LOCAL)
backend.set_cache.side_effect = TimeoutError("lit7520 backend timed out")
cache.cache = backend
with caplog.at_level(logging.DEBUG, logger="LiteLLM"):
cache.add_cache("result", model="gpt-4.1-mini", messages=[{"role": "user", "content": "hi"}])
records = [r for r in caplog.records if "lit7520 backend timed out" in r.getMessage()]
assert [r.levelno for r in records] == [expected_level]
def _embedding_response(prompt_tokens, num_items):
return EmbeddingResponse(
model="amazon.titan-embed-image-v1",
data=[Embedding(embedding=[0.0], index=i, object="embedding") for i in range(num_items)],
usage=Usage(prompt_tokens=prompt_tokens, completion_tokens=0, total_tokens=prompt_tokens),
)
def test_get_per_item_prompt_tokens_single_item_returns_full_value():
cache = Cache(type=LiteLLMCacheType.LOCAL)
result = _embedding_response(prompt_tokens=0, num_items=1)
assert cache._get_per_item_prompt_tokens(result, 0) == 0
def test_get_per_item_prompt_tokens_distributes_with_remainder():
cache = Cache(type=LiteLLMCacheType.LOCAL)
result = _embedding_response(prompt_tokens=10, num_items=3)
per_item = [cache._get_per_item_prompt_tokens(result, i) for i in range(3)]
assert sum(per_item) == 10 # 4 + 3 + 3
assert per_item == [4, 3, 3]
def _semantic_cache(**cache_kwargs):
return Cache(
type=LiteLLMCacheType.VALKEY_SEMANTIC,
host="localhost",
port="6379",
similarity_threshold=0.8,
**cache_kwargs,
)
@pytest.mark.parametrize(
"cache_type",
[LiteLLMCacheType.REDIS_SEMANTIC, LiteLLMCacheType.VALKEY_SEMANTIC],
)
def test_semantic_cache_embedding_max_input_tokens_reaches_backend(cache_type):
cache = Cache(
type=cache_type,
redis_url="redis://localhost:6379",
similarity_threshold=0.8,
semantic_cache_embedding_max_input_tokens=2048,
)
assert cache.cache.embedding_max_input_tokens == 2048
def test_semantic_cache_key_excludes_prompt_so_paraphrases_share_a_bucket():
cache = _semantic_cache()
tenant = {"user_api_key": "hash-abc"}
key_a = cache.get_cache_key(
model="gpt-4o-mini",
messages=[{"role": "user", "content": "What color is the sky?"}],
metadata=dict(tenant),
)
key_b = cache.get_cache_key(
model="gpt-4o-mini",
messages=[{"role": "user", "content": "Tell me the colour of the daytime sky."}],
metadata=dict(tenant),
)
assert key_a == key_b
def test_semantic_cache_key_isolates_tenants():
messages = [{"role": "user", "content": "What color is the sky?"}]
cache = _semantic_cache()
key_a = cache.get_cache_key(model="gpt-4o-mini", messages=messages, metadata={"user_api_key": "hash-A"})
key_b = cache.get_cache_key(model="gpt-4o-mini", messages=messages, metadata={"user_api_key": "hash-B"})
key_team = cache.get_cache_key(
model="gpt-4o-mini",
messages=messages,
metadata={"user_api_key": "hash-A", "user_api_key_team_id": "team-1"},
)
assert key_a != key_b
assert key_a != key_team
_SEMANTICALLY_IDENTICAL_PROMPTS = (
[{"role": "user", "content": "What color is the sky?"}],
[{"role": "user", "content": "Tell me the colour of the daytime sky."}],
)
def _end_user_keys(cache, metadata_field, *end_user_ids):
return [
cache.get_cache_key(
model="gpt-4o-mini",
messages=messages,
**{metadata_field: {"user_api_key": "hash-A", "user_api_key_end_user_id": end_user_id}},
)
for messages, end_user_id in zip(_SEMANTICALLY_IDENTICAL_PROMPTS, end_user_ids)
]
@pytest.mark.parametrize("metadata_field", ["metadata", "litellm_metadata"])
def test_semantic_cache_key_shares_bucket_across_end_users_by_default(metadata_field):
key_alice, key_bob = _end_user_keys(_semantic_cache(), metadata_field, "alice", "bob")
assert key_alice == key_bob
@pytest.mark.parametrize("metadata_field", ["metadata", "litellm_metadata"])
def test_semantic_cache_key_isolates_end_users_under_end_user_scope(metadata_field):
cache = _semantic_cache(semantic_cache_scope="end_user")
key_alice, key_bob = _end_user_keys(cache, metadata_field, "alice", "bob")
key_alice_again, _ = _end_user_keys(cache, metadata_field, "alice", "alice")
assert key_alice != key_bob
assert key_alice == key_alice_again
def test_semantic_cache_key_end_user_scope_without_end_user_falls_back_to_key_scope():
cache = _semantic_cache(semantic_cache_scope=SemanticCacheScope.END_USER)
messages = [{"role": "user", "content": "What color is the sky?"}]
key_scope_only = cache.get_cache_key(model="gpt-4o-mini", messages=messages, metadata={"user_api_key": "hash-A"})
end_user_absent = cache.get_cache_key(
model="gpt-4o-mini",
messages=messages,
metadata={"user_api_key": "hash-A", "user_api_key_end_user_id": None},
)
other_key = cache.get_cache_key(model="gpt-4o-mini", messages=messages, metadata={"user_api_key": "hash-B"})
key_alice, _ = _end_user_keys(cache, "metadata", "alice", "alice")
default_scope_key = _semantic_cache().get_cache_key(
model="gpt-4o-mini", messages=messages, metadata={"user_api_key": "hash-A"}
)
assert key_scope_only == end_user_absent == default_scope_key
assert key_scope_only != other_key
assert key_scope_only != key_alice
def test_semantic_cache_key_reads_tenant_identity_from_litellm_metadata():
cache = _semantic_cache()
messages = [{"role": "user", "content": "What color is the sky?"}]
key_a = cache.get_cache_key(model="gpt-4o-mini", messages=messages, litellm_metadata={"user_api_key": "hash-A"})
key_b = cache.get_cache_key(model="gpt-4o-mini", messages=messages, litellm_metadata={"user_api_key": "hash-B"})
key_a_in_litellm_params = cache.get_cache_key(
model="gpt-4o-mini",
messages=messages,
litellm_params={"litellm_metadata": {"user_api_key": "hash-A"}},
)
assert key_a != key_b
assert key_a == key_a_in_litellm_params
def test_cache_rejects_unsupported_type() -> None:
cache_params: Final = {"type": "not-a-cache"}
with pytest.raises(
ValueError, match=re.escape("Unsupported cache type 'not-a-cache'. Supported types: ")
) as raised:
Cache(**cache_params)
assert str(raised.value).split("Supported types: ")[1].split(", ") == [member.value for member in LiteLLMCacheType]
def test_semantic_cache_scope_rejects_unknown_value():
with pytest.raises(ValueError, match="'team' is not a valid SemanticCacheScope"):
_semantic_cache(semantic_cache_scope="team")
def test_semantic_cache_key_still_separates_models_and_params():
cache = _semantic_cache()
messages = [{"role": "user", "content": "hi"}]
tenant = {"user_api_key": "hash-A"}
assert cache.get_cache_key(model="gpt-4o-mini", messages=messages, metadata=dict(tenant)) != cache.get_cache_key(
model="gpt-4o", messages=messages, metadata=dict(tenant)
)
assert cache.get_cache_key(
model="gpt-4o-mini", messages=messages, temperature=0, metadata=dict(tenant)
) != cache.get_cache_key(model="gpt-4o-mini", messages=messages, temperature=1, metadata=dict(tenant))
def test_exact_cache_key_still_includes_prompt():
cache = Cache(type=LiteLLMCacheType.LOCAL)
key_a = cache.get_cache_key(model="gpt-4o-mini", messages=[{"role": "user", "content": "a"}])
key_b = cache.get_cache_key(model="gpt-4o-mini", messages=[{"role": "user", "content": "b"}])
assert key_a != key_b
@pytest.mark.parametrize(
"anthropic_param",
[
{"system": "answer ALPHA"},
{"top_k": 5},
{"stop_sequences": ["STOP"]},
],
)
def test_exact_cache_key_includes_anthropic_messages_params(anthropic_param):
"""Anthropic /v1/messages params with no OpenAI equivalent must still key the
cache; without them two requests that differ only by system prompt collide."""
cache = Cache(type=LiteLLMCacheType.LOCAL)
messages = [{"role": "user", "content": "which greek letter?"}]
baseline = cache.get_cache_key(model="claude-sonnet-4-5", messages=messages)
assert baseline != cache.get_cache_key(model="claude-sonnet-4-5", messages=messages, **anthropic_param)
@pytest.mark.parametrize(
"summary_ask",
[
{"reasoning_summary": "auto"},
{"reasoningSummary": "auto"},
{"extra_body": {"reasoning_summary": "auto"}},
{"extra_body": {"reasoningSummary": "auto"}},
],
ids=["top_level", "top_level_camel", "extra_body", "extra_body_camel"],
)
def test_exact_cache_key_includes_the_reasoning_summary_alias(summary_ask: dict[str, object]) -> None:
"""The AI SDK summary alias decides whether thought summaries come back, wherever the request
carries it, so a reply cached without it must not answer a request that asks for one."""
cache: Final = Cache(type=LiteLLMCacheType.LOCAL)
messages: Final = [{"role": "user", "content": "how much is the ball?"}]
no_ask: Final = cache.get_cache_key(model="gemini/gemini-3.8-flash", messages=messages, reasoning_effort="medium")
ask: Final = cache.get_cache_key(
model="gemini/gemini-3.8-flash", messages=messages, reasoning_effort="medium", **summary_ask
)
assert no_ask != ask
assert ask == cache.get_cache_key(
model="gemini/gemini-3.8-flash", messages=messages, reasoning_effort="medium", **summary_ask
)
assert ask != cache.get_cache_key(
model="gemini/gemini-3.8-flash",
messages=messages,
reasoning_effort="medium",
reasoning_summary="detailed",
)
@pytest.mark.asyncio
async def test_embedding_cache_skips_write_when_one_input_yields_many_embeddings(monkeypatch):
"""A cross-encoder behind /embeddings returns one score per document for a single
input string; caching data[0] per input would make the second call return 1 score."""
import litellm
from litellm import CustomLLM
class ScoreEveryDocument(CustomLLM):
provider_calls: int = 0
async def aembedding(self, model, input, model_response, **kwargs) -> EmbeddingResponse:
self.provider_calls += 1
return EmbeddingResponse(
model=model,
data=[Embedding(embedding=[float(i)], index=i, object="embedding") for i in range(5)],
)
scorer = ScoreEveryDocument()
monkeypatch.setattr(litellm, "custom_provider_map", [{"provider": "score-every-doc", "custom_handler": scorer}])
monkeypatch.setattr(litellm, "provider_list", [*litellm.provider_list, "score-every-doc"])
monkeypatch.setattr(litellm, "_custom_providers", [*litellm._custom_providers, "score-every-doc"])
monkeypatch.setattr(litellm, "cache", Cache(type=LiteLLMCacheType.LOCAL))
batch = '{"query": "q", "documents": ["a", "b", "c", "d", "e"]}'
first = await litellm.aembedding(model="score-every-doc/m", input=[batch])
await asyncio.gather(*_PENDING_CACHE_WRITES)
second = await litellm.aembedding(model="score-every-doc/m", input=[batch])
assert scorer.provider_calls == 2
assert [len(first.data), len(second.data)] == [5, 5]
@pytest.mark.asyncio
async def test_embedding_cache_refetches_entries_written_without_format_version(monkeypatch):
import litellm
from litellm import CustomLLM
class EmbedLength(CustomLLM):
provider_calls: int = 0
async def aembedding(self, model, input, model_response, **kwargs) -> EmbeddingResponse:
self.provider_calls += 1
return EmbeddingResponse(
model=model,
data=[
Embedding(embedding=[float(len(text))], index=idx, object="embedding")
for idx, text in enumerate(input)
],
)
embedder = EmbedLength()
monkeypatch.setattr(litellm, "custom_provider_map", [{"provider": "embed-length", "custom_handler": embedder}])
monkeypatch.setattr(litellm, "provider_list", [*litellm.provider_list, "embed-length"])
monkeypatch.setattr(litellm, "_custom_providers", [*litellm._custom_providers, "embed-length"])
monkeypatch.setattr(litellm, "cache", Cache(type=LiteLLMCacheType.LOCAL))
await litellm.aembedding(model="embed-length/m", input=["abcd"])
await asyncio.gather(*_PENDING_CACHE_WRITES)
store = litellm.cache.cache.cache_dict
stored = [entry["response"] for entry in store.values()]
assert [entry["format_version"] for entry in stored] == [EMBEDDING_CACHE_FORMAT_VERSION], stored
legacy_store = {
key: {
**entry,
"response": {
field: value
for field, value in {**entry["response"], "embedding": [-1.0]}.items()
if field != "format_version"
},
}
for key, entry in store.items()
}
monkeypatch.setattr(litellm.cache.cache, "cache_dict", legacy_store)
refetched = await litellm.aembedding(model="embed-length/m", input=["abcd"])
assert embedder.provider_calls == 2, "an entry written without format_version must be a cache miss"
assert [item["embedding"] for item in refetched.data] == [[4.0]]
@pytest.mark.asyncio
async def test_embedding_cache_serves_base64_string_embeddings_on_repeat(monkeypatch):
import litellm
from litellm import CustomLLM
class Base64Embedder(CustomLLM):
provider_calls: int = 0
async def aembedding(self, model, input, model_response, **kwargs) -> EmbeddingResponse:
self.provider_calls += 1
return EmbeddingResponse(
model=model,
data=[
Embedding(embedding="AACAPwAAAEA=", index=idx, object="embedding") for idx, _ in enumerate(input)
],
)
embedder = Base64Embedder()
monkeypatch.setattr(litellm, "custom_provider_map", [{"provider": "embed-b64", "custom_handler": embedder}])
monkeypatch.setattr(litellm, "provider_list", [*litellm.provider_list, "embed-b64"])
monkeypatch.setattr(litellm, "_custom_providers", [*litellm._custom_providers, "embed-b64"])
monkeypatch.setattr(litellm, "cache", Cache(type=LiteLLMCacheType.LOCAL))
first = await litellm.aembedding(model="embed-b64/m", input=["abcd"])
await asyncio.gather(*_PENDING_CACHE_WRITES)
second = await litellm.aembedding(model="embed-b64/m", input=["abcd"])
assert embedder.provider_calls == 1, "a string embedding written to the cache must be served on repeat"
assert [item["embedding"] for item in second.data] == [item["embedding"] for item in first.data] == ["AACAPwAAAEA="]
def test_provider_specific_cache_key_ignores_litellm_owned_kwargs(monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setattr(litellm, "enable_caching_on_provider_specific_optional_params", True)
cache: Final = Cache(type=LiteLLMCacheType.LOCAL)
request: Final = {"model": "gpt-4.1-mini", "messages": [{"role": "user", "content": "hi"}], "top_k": 5}
base_key: Final = cache.get_cache_key(**request)
assert cache.get_cache_key(**request, _litellm_control={"stream_chunk_size": 64}) == base_key
assert cache.get_cache_key(**request, litellm_trace_id="trace-1") == base_key
assert cache.get_cache_key(**{**request, "top_k": 6}) != base_key
def test_cache_key_tells_apart_requests_that_differ_only_in_continuation_token() -> None:
cache: Final = Cache(type=LiteLLMCacheType.LOCAL)
request: Final = {
"model": "gemini/gemini-4-argon",
"messages": [{"role": "user", "content": "hi"}],
"max_tokens": 16000,
}
first_segment_key: Final = cache.get_cache_key(**request)
resumed_key: Final = cache.get_cache_key(**request, continuation_token="dG9rLTE=")
assert resumed_key != first_segment_key
assert cache.get_cache_key(**request, continuation_token="dG9rLTI=") not in (first_segment_key, resumed_key)
assert cache.get_cache_key(**request, continuation_token="dG9rLTE=") == resumed_key
assert cache.get_cache_key(**request, continuation_token=None) == first_segment_key
assert cache.get_cache_key(**request, continuation_token="") == first_segment_key
class PhaseRecordingCache(InMemoryCache):
"""Records the target and the active span each read / write ran under, as a Redis span would."""
def __init__(self) -> None:
super().__init__()
self.seen: list[tuple[str | None, str]] = []
def _record(self) -> None:
from opentelemetry import trace
span = trace.get_current_span()
self.seen.append((current_service_target(), getattr(span, "name", "")))
def get_cache(self, key, **kwargs):
self._record()
return super().get_cache(key, **kwargs)
def set_cache(self, key, value, **kwargs):
self._record()
super().set_cache(key, value, **kwargs)
@pytest.fixture
def v2_span_exporter(monkeypatch):
from opentelemetry.sdk.trace.export.in_memory_span_exporter import InMemorySpanExporter
from litellm.integrations.otel import OpenTelemetryV2Config
from litellm.integrations.otel.logger import OpenTelemetryV2
from litellm.integrations.otel.plumbing import providers
from litellm.proxy import proxy_server
config = OpenTelemetryV2Config(exporter="in_memory")
exporter = InMemorySpanExporter()
logger = OpenTelemetryV2(config=config, tracer_provider=providers.build_tracer_provider(config, exporter=exporter))
monkeypatch.setattr(proxy_server, "open_telemetry_logger", logger)
return exporter
_REQUEST: Final = {"model": "gpt-5.4-mini", "messages": [{"role": "user", "content": "phase me"}]}
@pytest.mark.asyncio
async def test_facade_lookup_and_store_run_inside_the_response_cache_phases(v2_span_exporter):
"""The native bridge calls ``Cache.async_get_cache`` / ``async_add_cache`` straight, never through
``caching_handler``, so the ``cache.get llm_response`` / ``cache.set llm_response`` phase and the
``llm_response`` target come from the facade: the store runs under them too, and a hit reads back."""
cache = Cache(type=LiteLLMCacheType.LOCAL)
backend = PhaseRecordingCache()
assert await cache.async_get_cache(dynamic_cache_object=backend, **_REQUEST) is None
await cache.async_add_cache({"id": "resp-1"}, dynamic_cache_object=backend, **_REQUEST)
assert await cache.async_get_cache(dynamic_cache_object=backend, **_REQUEST) == {"id": "resp-1"}
assert backend.seen == [
("llm_response", "cache.get llm_response"),
("llm_response", "cache.set llm_response"),
("llm_response", "cache.get llm_response"),
]
assert [s.name for s in v2_span_exporter.get_finished_spans()] == [
"cache.get llm_response",
"cache.set llm_response",
"cache.get llm_response",
]
assert current_service_target() is None
def test_sync_facade_lookup_and_store_run_inside_the_response_cache_phases(v2_span_exporter):
cache = Cache(type=LiteLLMCacheType.LOCAL)
backend = PhaseRecordingCache()
assert cache.get_cache(dynamic_cache_object=backend, **_REQUEST) is None
cache.add_cache({"id": "resp-1"}, **_REQUEST)
assert backend.seen == [("llm_response", "cache.get llm_response")]
assert [s.name for s in v2_span_exporter.get_finished_spans()] == [
"cache.get llm_response",
"cache.set llm_response",
]
@pytest.mark.asyncio
async def test_a_lookup_already_inside_the_phase_does_not_open_a_second_one(v2_span_exporter):
"""``caching_handler`` opens the phase around the facade call; the facade joins it."""
cache = Cache(type=LiteLLMCacheType.LOCAL)
backend = PhaseRecordingCache()
with response_cache_phase("get"):
await cache.async_get_cache(dynamic_cache_object=backend, **_REQUEST)
assert backend.seen == [("llm_response", "cache.get llm_response")]
assert [s.name for s in v2_span_exporter.get_finished_spans()] == ["cache.get llm_response"]
def test_cache_override():
# test if we can override the cache, when `caching=False` but litellm.cache = Cache() is set
# in this case it should not return cached responses
litellm.cache = Cache()
print("Testing cache override")
litellm.set_verbose = True
# test embedding
response1 = embedding(
model="text-embedding-ada-002",
input=["hello who are you"],
caching=False,
mock_response="0.1,0.2,0.3,0.4,0.5",
)
response2 = embedding(
model="text-embedding-ada-002",
input=["hello who are you"],
caching=False,
mock_response="0.6,0.7,0.8,0.9,1.0",
)
# When caching=False, responses should have different IDs
assert response1.data[0].embedding != response2.data[0].embedding
def test_caching_v2(): # test in memory cache
try:
litellm.set_verbose = True
litellm.cache = Cache()
response1 = completion(
model="gpt-3.5-turbo",
messages=messages,
caching=True,
mock_response="Hello world from cache test",
)
response2 = completion(model="gpt-3.5-turbo", messages=messages, caching=True)
print(f"response1: {response1}")
print(f"response2: {response2}")
litellm.cache = None # disable cache
litellm.success_callback = []
litellm._async_success_callback = []
if response2["choices"][0]["message"]["content"] != response1["choices"][0]["message"]["content"]:
print(f"response1: {response1}")
print(f"response2: {response2}")
pytest.fail(f"Error occurred:")
except Exception as e:
print(f"error occurred: {traceback.format_exc()}")
pytest.fail(f"Error occurred: {e}")
def test_caching_with_ttl():
try:
litellm.set_verbose = True
litellm.cache = Cache()
response1 = completion(
model="gpt-3.5-turbo",
messages=messages,
caching=True,
ttl=0,
mock_response="Hello world from cache test 1",
)
response2 = completion(
model="gpt-3.5-turbo",
messages=messages,
caching=True,
mock_response="Hello world from cache test 2",
)
print(f"response1: {response1}")
print(f"response2: {response2}")
litellm.cache = None # disable cache
litellm.success_callback = []
litellm._async_success_callback = []
assert response2["choices"][0]["message"]["content"] != response1["choices"][0]["message"]["content"]
except Exception as e:
print(f"error occurred: {traceback.format_exc()}")
pytest.fail(f"Error occurred: {e}")
def test_caching_with_default_ttl():
try:
litellm.set_verbose = True
litellm.cache = Cache(ttl=0)
response1 = completion(
model="gpt-3.5-turbo",
messages=messages,
caching=True,
mock_response="Hello world from cache test",
)
response2 = completion(
model="gpt-3.5-turbo",
messages=messages,
caching=True,
mock_response="Hello world from cache test",
)
print(f"response1: {response1}")
print(f"response2: {response2}")
litellm.cache = None # disable cache
litellm.success_callback = []
litellm._async_success_callback = []
assert response2["id"] != response1["id"]
except Exception as e:
print(f"error occurred: {traceback.format_exc()}")
pytest.fail(f"Error occurred: {e}")
def test_caching_with_models_v2():
messages = [{"role": "user", "content": "who is ishaan CTO of litellm from litellm 2023"}]
litellm.cache = Cache()
print("test2 for caching")
litellm.set_verbose = True
response1 = completion(
model="gpt-3.5-turbo",
messages=messages,
caching=True,
mock_response="Hello world from cache test",
)
response2 = completion(model="gpt-3.5-turbo", messages=messages, caching=True)
response3 = completion(
model="gpt-4.1-nano",
messages=messages,
caching=True,
mock_response="Different model response",
)
print(f"response1: {response1}")
print(f"response2: {response2}")
print(f"response3: {response3}")
litellm.cache = None
litellm.success_callback = []
litellm._async_success_callback = []
if response3["choices"][0]["message"]["content"] == response2["choices"][0]["message"]["content"]:
# if models are different, it should not return cached response
print(f"response2: {response2}")
print(f"response3: {response3}")
pytest.fail(f"Error occurred:")
if response1["choices"][0]["message"]["content"] != response2["choices"][0]["message"]["content"]:
print(f"response1: {response1}")
print(f"response2: {response2}")
pytest.fail(f"Error occurred:")
@pytest.mark.asyncio
async def test_dual_cache_caching_batch_get_cache():
"""
- check redis cache called for initial batch get cache
- check redis cache not called for consecutive batch get cache with same keys
"""
from litellm.caching.dual_cache import DualCache
from litellm.caching.redis_cache import RedisCache
dc = DualCache(redis_cache=MagicMock(spec=RedisCache))
with patch.object(
dc.redis_cache,
"async_batch_get_cache",
new=AsyncMock(return_value={"test_key1": "test_value1", "test_key2": "test_value2"}),
) as mock_async_get_cache:
await dc.async_batch_get_cache(keys=["test_key1", "test_key2"])
assert mock_async_get_cache.call_count == 1
await dc.async_batch_get_cache(keys=["test_key1", "test_key2"])
assert mock_async_get_cache.call_count == 1
def test_get_cache_key():
from litellm.caching.caching import Cache
try:
print("Testing get_cache_key")
cache_instance = Cache()
cache_key = cache_instance.get_cache_key(
**{
"model": "gpt-3.5-turbo",
"messages": [{"role": "user", "content": "write a one sentence poem about: 7510"}],
"max_tokens": 40,
"temperature": 0.2,
"stream": True,
"litellm_call_id": "ffe75e7e-8a07-431f-9a74-71a5b9f35f0b",
"litellm_logging_obj": {},
}
)
cache_key_2 = cache_instance.get_cache_key(
**{
"model": "gpt-3.5-turbo",
"messages": [{"role": "user", "content": "write a one sentence poem about: 7510"}],
"max_tokens": 40,
"temperature": 0.2,
"stream": True,
"litellm_call_id": "ffe75e7e-8a07-431f-9a74-71a5b9f35f0b",
"litellm_logging_obj": {},
}
)
cache_key_str = "model: gpt-3.5-turbomessages: [{'role': 'user', 'content': 'write a one sentence poem about: 7510'}]max_tokens: 40temperature: 0.2stream: True"
hash_object = hashlib.sha256(cache_key_str.encode())
# Hexadecimal representation of the hash
hash_hex = hash_object.hexdigest()
assert cache_key == hash_hex
assert cache_key_2 == hash_hex, (
f"{cache_key} != {cache_key_2}. The same kwargs should have the same cache key across runs"
)
embedding_cache_key = cache_instance.get_cache_key(
**{
"model": "azure/text-embedding-ada-002",
"api_base": "https://openai-gpt-4-test-v-1.openai.azure.com/",
"api_key": "",
"api_version": "2023-07-01-preview",
"timeout": None,
"max_retries": 0,
"input": ["hi who is ishaan"],
"caching": True,
"client": "<openai.lib.azure.AsyncAzureOpenAI object at 0x12b6a1060>",
}
)
print(embedding_cache_key)
embedding_cache_key_str = "model: azure/text-embedding-ada-002input: ['hi who is ishaan']"
hash_object = hashlib.sha256(embedding_cache_key_str.encode())
# Hexadecimal representation of the hash
hash_hex = hash_object.hexdigest()
assert embedding_cache_key == hash_hex, (
f"{embedding_cache_key} != 'model: azure/text-embedding-ada-002input: ['hi who is ishaan']'. The same kwargs should have the same cache key across runs"
)
# Proxy - embedding cache, test if embedding key, gets model_group and not model
embedding_cache_key_2 = cache_instance.get_cache_key(
**{
"model": "azure/text-embedding-ada-002",
"api_base": "https://openai-gpt-4-test-v-1.openai.azure.com/",
"api_key": "",
"api_version": "2023-07-01-preview",
"timeout": None,
"max_retries": 0,
"input": ["hi who is ishaan"],
"caching": True,
"client": "<openai.lib.azure.AsyncAzureOpenAI object at 0x12b6a1060>",
"proxy_server_request": {
"url": "http://0.0.0.0:8000/embeddings",
"method": "POST",
"headers": {
"host": "0.0.0.0:8000",
"user-agent": "curl/7.88.1",
"accept": "*/*",
"content-type": "application/json",
"content-length": "80",
},
"body": {
"model": "azure-embedding-model",
"input": ["hi who is ishaan"],
},
},
"user": None,
"metadata": {
"user_api_key": None,
"headers": {
"host": "0.0.0.0:8000",
"user-agent": "curl/7.88.1",
"accept": "*/*",
"content-type": "application/json",
"content-length": "80",
},
"model_group": "EMBEDDING_MODEL_GROUP",
"deployment": "azure/text-embedding-ada-002-ModelID-azure/text-embedding-ada-002https://openai-gpt-4-test-v-1.openai.azure.com/2023-07-01-preview",
},
"model_info": {
"mode": "embedding",
"base_model": "text-embedding-ada-002",
"id": "20b2b515-f151-4dd5-a74f-2231e2f54e29",
},
"litellm_call_id": "2642e009-b3cd-443d-b5dd-bb7d56123b0e",
"litellm_logging_obj": "<litellm.utils.Logging object at 0x12f1bddb0>",
}
)
print(embedding_cache_key_2)
embedding_cache_key_str_2 = "model: EMBEDDING_MODEL_GROUPinput: ['hi who is ishaan']"
hash_object = hashlib.sha256(embedding_cache_key_str_2.encode())
# Hexadecimal representation of the hash
hash_hex = hash_object.hexdigest()
assert embedding_cache_key_2 == hash_hex
print("passed!")
except Exception as e:
traceback.print_exc()
pytest.fail(f"Error occurred:", e)
def test_redis_caching_multiple_namespaces():
"""
Test that redis caching works with multiple namespaces
If client side request specifies a namespace, it should be used for caching
The same request with different namespaces should not be cached under the same key
"""
from unittest.mock import MagicMock, patch
import litellm
from litellm import completion
from litellm._uuid import uuid
from litellm.caching import Cache
# Use a fixed uuid to ensure consistent cache keys
test_uuid = "12345678-1234-1234-1234-123456789abc"
messages = [{"role": "user", "content": f"what is litellm? {test_uuid}"}]
# Mock the Redis client creation from the _redis module
with (
patch("litellm._redis.get_redis_client") as mock_get_redis_client,
patch("litellm._redis.get_redis_connection_pool") as mock_get_redis_connection_pool,
):
# Create a mock Redis client that simulates real Redis behavior
mock_redis_client = MagicMock()
mock_get_redis_client.return_value = mock_redis_client
# Mock the connection pool
mock_connection_pool = MagicMock()
mock_get_redis_connection_pool.return_value = mock_connection_pool
# Dictionary to simulate Redis storage with namespace support
redis_storage = {}
def mock_redis_get(key):
print(f"Redis GET: {key}")
value = redis_storage.get(key, None)
# Convert to bytes to match real Redis behavior
if value is not None:
import json
return json.dumps(value).encode("utf-8")
return None
def mock_redis_set(name, value, ex=None, **kwargs):
print(f"Redis SET: {name} = {value}")
redis_storage[name] = value
return True
def mock_redis_ping():
return True
def mock_redis_info():
return {"redis_version": "7.0.0"}
mock_redis_client.get = mock_redis_get
mock_redis_client.set = mock_redis_set
mock_redis_client.ping = mock_redis_ping
mock_redis_client.info = mock_redis_info
# Initialize the cache
litellm.cache = Cache(type="redis")
namespace_1 = "org-id1"
namespace_2 = "org-id2"
# Use mock_response to ensure deterministic responses without external API calls
response_1 = completion(
model="gpt-3.5-turbo",
messages=messages,
cache={"namespace": namespace_1},
mock_response="Response for namespace 1",
)
response_2 = completion(
model="gpt-3.5-turbo",
messages=messages,
cache={"namespace": namespace_2},
mock_response="Response for namespace 2",
)
response_3 = completion(
model="gpt-3.5-turbo",
messages=messages,
cache={"namespace": namespace_1},
mock_response="This should be cached",
)
response_4 = completion(
model="gpt-3.5-turbo",
messages=messages,
mock_response="Response without namespace",
)
print(f"Response 1 type: {type(response_1)} - ID: {getattr(response_1, 'id', 'N/A')}")
print(f"Response 2 type: {type(response_2)} - ID: {getattr(response_2, 'id', 'N/A')}")
print(f"Response 3 type: {type(response_3)} - Cache hit: {isinstance(response_3, str)}")
print(f"Response 4 type: {type(response_4)} - ID: {getattr(response_4, 'id', 'N/A')}")
print(f"Redis storage keys: {list(redis_storage.keys())}")
# Verify that different namespaces created different cache keys
cache_keys = list(redis_storage.keys())
namespace_1_keys = [k for k in cache_keys if k.startswith(f"{namespace_1}:")]
namespace_2_keys = [k for k in cache_keys if k.startswith(f"{namespace_2}:")]
no_namespace_keys = [
k for k in cache_keys if not k.startswith(f"{namespace_1}:") and not k.startswith(f"{namespace_2}:")
]
print(f"Namespace 1 keys: {namespace_1_keys}")
print(f"Namespace 2 keys: {namespace_2_keys}")
print(f"No namespace keys: {no_namespace_keys}")
# Should have at least one key for each namespace
assert len(namespace_1_keys) > 0, "Should have cache keys for namespace 1"
assert len(namespace_2_keys) > 0, "Should have cache keys for namespace 2"
assert len(no_namespace_keys) > 0, "Should have cache keys for no namespace"
# The main test: response 3 should be a cache hit (string) because it uses same namespace as response 1
assert isinstance(response_3, str), "Response 3 should be a cache hit (string) for same namespace"
# response 1 & 2 should be ModelResponse objects (cache misses)
assert hasattr(response_1, "id"), "Response 1 should be a ModelResponse object"
assert hasattr(response_2, "id"), "Response 2 should be a ModelResponse object"
assert hasattr(response_4, "id"), "Response 4 should be a ModelResponse object"
# response 1 & 2 should have different IDs (different namespaces)
assert response_1.id != response_2.id, (
f"Expected different response ID for different namespace. Got {response_1.id} and {response_2.id}"
)
# response 1 & 4 should have different IDs (different namespaces)
assert response_1.id != response_4.id, (
f"Expected different response ID for no namespace vs namespaced. Got {response_1.id} and {response_4.id}"
)
_TOOL_TURN_ITEM: Final = {"role": "user", "content": "hi"}
_FILE_BLOCK_ITEM: Final = {"type": "file", "file": {"file_data": "data:video/mp4;base64,AAAA", "format": "video/mp4"}}
@pytest.mark.parametrize(
("kwargs", "expected"),
[
pytest.param({"messages": [_TOOL_TURN_ITEM] * 4}, True, id="four-messages-are-cached"),
pytest.param({"messages": [_TOOL_TURN_ITEM] * 5}, False, id="five-messages-skip-the-cache"),
pytest.param({"input": [_TOOL_TURN_ITEM] * 4}, True, id="four-responses-items-are-cached"),
pytest.param({"input": [_TOOL_TURN_ITEM] * 5}, False, id="five-responses-items-skip-the-cache"),
pytest.param({"input": "one prompt"}, True, id="string-input-is-one-message"),
pytest.param({"input": ["a", "b", "c", "d", "e"]}, True, id="embedding-strings-are-not-messages"),
pytest.param({"input": [_FILE_BLOCK_ITEM] * 5}, True, id="embedding-file-blocks-are-not-messages"),
],
)
def test_should_use_cache_stops_past_the_default_max_messages(kwargs: dict[str, object], expected: bool) -> None:
assert Cache(type=LiteLLMCacheType.LOCAL).should_use_cache(**kwargs) is expected
def test_responses_sdk_items_count_toward_max_messages() -> None:
from openai.types.responses import ResponseFunctionToolCall
call: Final = ResponseFunctionToolCall(type="function_call", call_id="c1", name="ls", arguments="{}")
assert Cache(type=LiteLLMCacheType.LOCAL).should_use_cache(input=[_TOOL_TURN_ITEM, call, call, call, call]) is False
def test_max_messages_is_configurable_and_none_disables_it() -> None:
three: Final = [_TOOL_TURN_ITEM] * 3
assert Cache(type=LiteLLMCacheType.LOCAL, max_messages=2).should_use_cache(messages=three) is False
assert Cache(type=LiteLLMCacheType.LOCAL, max_messages=3).should_use_cache(messages=three) is True
assert Cache(type=LiteLLMCacheType.LOCAL, max_messages=None).should_use_cache(messages=three * 50) is True
def test_max_messages_beats_an_explicit_use_cache_opt_in() -> None:
cache: Final = Cache(type=LiteLLMCacheType.LOCAL, mode=CacheMode.default_off)
assert cache.should_use_cache(messages=[_TOOL_TURN_ITEM] * 4, cache={"use-cache": True}) is True
assert cache.should_use_cache(messages=[_TOOL_TURN_ITEM] * 5, cache={"use-cache": True}) is False
def test_completion_past_max_messages_is_neither_served_from_nor_written_to_the_cache(
monkeypatch: pytest.MonkeyPatch,
) -> None:
monkeypatch.setattr(litellm, "cache", Cache(type=LiteLLMCacheType.LOCAL))
tag: Final = uuid.uuid4().hex
four: Final = [{"role": "user", "content": f"{tag} turn {index}"} for index in range(4)]
five: Final = [*four, {"role": "user", "content": f"{tag} turn 4"}]
def answer(messages: list[dict[str, str]], mock_response: str) -> str:
response: Final = litellm.completion(model="gpt-4o-mini", messages=messages, mock_response=mock_response)
assert isinstance(response, litellm.ModelResponse), response
choice: Final = response.choices[0]
assert isinstance(choice, litellm.Choices), choice
return str(choice.message.content)
assert answer(four, "four first") == "four first"
assert answer(four, "four second") == "four first", "a 4-message repeat missed the cache"
assert answer(five, "five first") == "five first"
assert answer(five, "five second") == "five second", "a 5-message repeat was served from the cache"
@respx.mock
def test_enable_cache_serves_a_repeated_bedrock_titan_embedding_from_the_local_cache(
monkeypatch: pytest.MonkeyPatch, capsys: pytest.CaptureFixture[str], fake_provider_credentials: None
) -> None:
monkeypatch.setattr(litellm, "cache", None)
monkeypatch.setattr(litellm, "set_verbose", True)
monkeypatch.setattr(litellm, "input_callback", [])
monkeypatch.setattr(litellm, "success_callback", [])
monkeypatch.setattr(litellm, "_async_success_callback", [])
route: Final = respx.post(
"https://bedrock-runtime.us-west-2.amazonaws.com/model/amazon.titan-embed-text-v2:0/invoke"
).mock(return_value=httpx.Response(200, json={"embedding": [0.25, -0.5, 0.75], "inputTextTokenCount": 6}))
request: Final = {
"model": "bedrock/amazon.titan-embed-text-v2:0",
"input": "good morning from litellm, attempting to embed data",
"aws_region_name": "us-west-2",
}
litellm.enable_cache()
enabled_cache: Final = litellm.cache
monkeypatch.setattr(litellm, "input_callback", [])
monkeypatch.setattr(litellm, "success_callback", [])
monkeypatch.setattr(litellm, "_async_success_callback", [])
litellm.enable_cache()
first: Final = embedding(**request)
second: Final = embedding(**request)
assert isinstance(enabled_cache, Cache)
assert enabled_cache.type == LiteLLMCacheType.LOCAL
assert litellm.cache is enabled_cache
assert litellm.input_callback == ["cache"]
assert litellm.success_callback == ["cache"]
assert litellm._async_success_callback == ["cache"]
assert "LiteLLM: Enabling Cache" in capsys.readouterr().out
assert route.call_count == 1
assert first.data[0]["embedding"] == [0.25, -0.5, 0.75]
assert second.data[0]["embedding"] == [0.25, -0.5, 0.75]
assert second._hidden_params["cache_hit"] is True
def test_basic_caching_import() -> None:
from litellm.caching import Cache as PublicCache
assert PublicCache is Cache
cache: Final = PublicCache(type=LiteLLMCacheType.LOCAL)
assert cache.type == LiteLLMCacheType.LOCAL
def test_cache_context_managers(monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setattr(litellm, "cache", Cache(type=LiteLLMCacheType.LOCAL))
monkeypatch.setattr(litellm, "input_callback", list(litellm.input_callback))
monkeypatch.setattr(litellm, "success_callback", list(litellm.success_callback))
monkeypatch.setattr(litellm, "_async_success_callback", list(litellm._async_success_callback))
litellm.disable_cache()
assert litellm.cache is None
assert "cache" not in litellm.success_callback
assert "cache" not in litellm._async_success_callback
litellm.enable_cache(type=LiteLLMCacheType.LOCAL)
assert litellm.cache is not None
assert litellm.cache.type == LiteLLMCacheType.LOCAL
@pytest.mark.asyncio
async def test_cache_default_off_acompletion(monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setattr(litellm, "cache", Cache(type=LiteLLMCacheType.LOCAL, mode=CacheMode.default_off))
request: Final = {
"model": "gpt-4o-mini",
"messages": [{"role": "user", "content": f"cache mode {uuid.uuid4().hex}"}],
}
first: Final = await acompletion(**request, mock_response="first")
second: Final = await acompletion(**request, mock_response="second")
opted_in_first: Final = await acompletion(**request, cache={"use-cache": True}, mock_response="third")
await asyncio.gather(*_PENDING_CACHE_WRITES)
opted_in_second: Final = await acompletion(**request, cache={"use-cache": True}, mock_response="fourth")
assert first.id != second.id
assert opted_in_first.id == opted_in_second.id
def test_caching_kwargs_input(monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setattr(litellm, "cache", Cache(type=LiteLLMCacheType.LOCAL))
request: Final = {
"model": "gpt-4o-mini",
"messages": [{"role": "user", "content": f"kwargs input {uuid.uuid4().hex}"}],
"caching": True,
}
first: Final = completion(**request, mock_response="stored")
second: Final = completion(**request, mock_response="not stored")
assert first.id == second.id
def test_caching_reasoning_args_hit(monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setattr(litellm, "cache", Cache(type=LiteLLMCacheType.LOCAL))
request: Final = {
"model": "anthropic/claude-sonnet-4-6",
"messages": [{"role": "user", "content": f"reasoning hit {uuid.uuid4().hex}"}],
"reasoning_effort": "low",
"caching": True,
}
first: Final = completion(**request, mock_response="cached")
second: Final = completion(**request, mock_response="uncached")
assert first.id == second.id
def test_caching_reasoning_args_miss(monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setattr(litellm, "cache", Cache(type=LiteLLMCacheType.LOCAL))
messages: Final = [{"role": "user", "content": f"reasoning miss {uuid.uuid4().hex}"}]
first: Final = completion(
model="anthropic/claude-sonnet-4-6",
messages=messages,
reasoning_effort="low",
caching=True,
mock_response="low",
)
second: Final = completion(
model="anthropic/claude-sonnet-4-6",
messages=messages,
caching=True,
mock_response="default",
)
assert first.id != second.id
def test_caching_thinking_args_hit(monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setattr(litellm, "cache", Cache(type=LiteLLMCacheType.LOCAL))
request: Final = {
"model": "anthropic/claude-sonnet-4-5-20250929",
"messages": [{"role": "user", "content": f"thinking hit {uuid.uuid4().hex}"}],
"thinking": {"type": "enabled", "budget_tokens": 1024},
"caching": True,
}
first: Final = completion(**request, mock_response="cached")
second: Final = completion(**request, mock_response="uncached")
assert first.id == second.id
def test_caching_thinking_args_miss(monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setattr(litellm, "cache", Cache(type=LiteLLMCacheType.LOCAL))
messages: Final = [{"role": "user", "content": f"thinking miss {uuid.uuid4().hex}"}]
first: Final = completion(
model="anthropic/claude-sonnet-4-5-20250929",
messages=messages,
thinking={"type": "enabled", "budget_tokens": 1024},
caching=True,
mock_response="thinking",
)
second: Final = completion(
model="anthropic/claude-sonnet-4-5-20250929",
messages=messages,
caching=True,
mock_response="not thinking",
)
assert first.id != second.id
def test_caching_with_reasoning_content(monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setattr(litellm, "cache", Cache(type=LiteLLMCacheType.LOCAL))
request: Final = {
"model": "anthropic/claude-sonnet-4-5-20250929",
"messages": [{"role": "user", "content": f"reasoning content {uuid.uuid4().hex}"}],
"thinking": {"type": "enabled", "budget_tokens": 1024},
}
completion(**request, mock_response="reasoned")
cached: Final = completion(**request, mock_response="different")
assert cached._hidden_params["cache_hit"] is True
def test_custom_redis_cache_params(redis_client_without_network: None) -> None:
cache: Final = Cache(type=LiteLLMCacheType.REDIS, host="127.0.0.1", port="6379", db=4)
assert cache.cache.redis_client.connection_pool.connection_kwargs["db"] == 4
def test_redis_cache_ignores_a_leftover_disk_cache_dir(
redis_client_without_network: None, caplog: pytest.LogCaptureFixture
) -> None:
with caplog.at_level(logging.WARNING, logger="LiteLLM"):
Cache(type=LiteLLMCacheType.REDIS, host="127.0.0.1", port="6379")
assert "disk_cache_dir" not in caplog.text
with caplog.at_level(logging.WARNING, logger="LiteLLM"):
cache: Final = Cache(
type=LiteLLMCacheType.REDIS, host="127.0.0.1", port="6379", disk_cache_dir="/tmp/litellm-cache"
)
assert isinstance(cache.cache, RedisCache)
assert "disk_cache_dir" in caplog.text
def test_custom_redis_cache_with_key(monkeypatch: pytest.MonkeyPatch) -> None:
cache: Final = Cache(type=LiteLLMCacheType.LOCAL)
monkeypatch.setattr(litellm, "cache", cache)
request: Final = {
"model": "gpt-4o-mini",
"messages": [{"role": "user", "content": f"custom key {uuid.uuid4().hex}"}],
"caching": True,
"cache_key": f"custom-{uuid.uuid4().hex}",
}
first: Final = completion(**request, mock_response="stored")
second: Final = completion(
**{**request, "messages": [{"role": "user", "content": "different prompt, same key"}]},
mock_response="not stored",
)
uncached: Final = completion(
**{**request, "cache_key": f"other-{uuid.uuid4().hex}"},
mock_response="fresh",
)
assert second.id == first.id
assert second.choices[0].message.content == "stored"
assert uncached.id != first.id
assert uncached.choices[0].message.content == "fresh"
@pytest.mark.asyncio
async def test_dual_cache_async_batch_get_cache_returns_memory_values() -> None:
cache: Final = InMemoryCache()
cache.set_cache("test-value", "cached")
dual_cache: Final = DualCache(in_memory_cache=cache)
result: Final = await dual_cache.async_batch_get_cache(keys=["test-value"])
assert result == ["cached"]
def test_embedding_caching(monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setattr(litellm, "cache", Cache(type=LiteLLMCacheType.LOCAL))
inputs: Final = [f"embedding {uuid.uuid4().hex}"]
first: Final = embedding(model="text-embedding-3-small", input=inputs, caching=True, mock_response="0.1,0.2")
second: Final = embedding(model="text-embedding-3-small", input=inputs, caching=True, mock_response="0.3,0.4")
assert second.data[0]["embedding"] == first.data[0]["embedding"]
def test_embedding_caching_azure(monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setattr(litellm, "cache", Cache(type=LiteLLMCacheType.LOCAL))
request: Final = {
"model": "azure/text-embedding-3-small",
"input": [f"azure embedding {uuid.uuid4().hex}"],
"api_key": "test-key",
"api_base": "https://example.test",
"api_version": "2024-01-01",
"caching": True,
}
first: Final = embedding(**request, mock_response="0.1,0.2")
second: Final = embedding(**request, mock_response="0.3,0.4")
assert second.data[0]["embedding"] == first.data[0]["embedding"]
@pytest.mark.asyncio
async def test_embedding_caching_individual_items(monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setattr(litellm, "cache", Cache(type=LiteLLMCacheType.LOCAL))
item: Final = f"individual embedding {uuid.uuid4().hex}"
first: Final = await aembedding(model="text-embedding-3-small", input=item, caching=True, mock_response="0.1,0.2")
second: Final = await aembedding(model="text-embedding-3-small", input=item, caching=True, mock_response="0.3,0.4")
assert second.data[0].embedding == first.data[0].embedding
assert second._hidden_params["cache_hit"] is True
@pytest.mark.asyncio
async def test_embedding_caching_azure_individual_items(monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setattr(litellm, "cache", Cache(type=LiteLLMCacheType.LOCAL))
shared: Final = f"shared azure embedding {uuid.uuid4().hex}"
await aembedding(
model="azure/text-embedding-3-small",
input=[shared],
api_key="test-key",
api_base="https://example.test",
api_version="2024-01-01",
caching=True,
mock_response="0.1,0.2",
)
result: Final = await aembedding(
model="azure/text-embedding-3-small",
input=[shared, f"new {uuid.uuid4().hex}"],
api_key="test-key",
api_base="https://example.test",
api_version="2024-01-01",
caching=True,
mock_response="0.3,0.4",
)
assert result._hidden_params["cache_hit"] is True
@pytest.mark.asyncio
async def test_embedding_caching_azure_individual_items_reordered(monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setattr(litellm, "cache", Cache(type=LiteLLMCacheType.LOCAL))
first_item: Final = f"first reordered embedding {uuid.uuid4().hex}"
second_item: Final = f"second reordered embedding {uuid.uuid4().hex}"
first: Final = await aembedding(
model="azure/text-embedding-3-small",
input=first_item,
api_key="test-key",
api_base="https://example.test",
api_version="2024-01-01",
caching=True,
mock_response="0.1,0.2",
)
second: Final = await aembedding(
model="azure/text-embedding-3-small",
input=second_item,
api_key="test-key",
api_base="https://example.test",
api_version="2024-01-01",
caching=True,
mock_response="0.3,0.4",
)
await asyncio.gather(*_PENDING_CACHE_WRITES)
reordered: Final = await aembedding(
model="azure/text-embedding-3-small",
input=[second_item, first_item],
api_key="test-key",
api_base="https://example.test",
api_version="2024-01-01",
caching=True,
mock_response="0.5,0.6",
)
assert reordered._hidden_params["cache_hit"] is True
assert [item.index for item in reordered.data] == [0, 1]
assert reordered.data[0].embedding == second.data[0].embedding
assert reordered.data[1].embedding == first.data[0].embedding
@pytest.mark.asyncio
async def test_embedding_caching_individual_items_and_then_list(monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setattr(litellm, "cache", Cache(type=LiteLLMCacheType.LOCAL))
first_item: Final = f"first embedding {uuid.uuid4().hex}"
second_item: Final = f"second embedding {uuid.uuid4().hex}"
first: Final = await aembedding(
model="text-embedding-3-small", input=first_item, caching=True, mock_response="0.1,0.2"
)
second: Final = await aembedding(
model="text-embedding-3-small", input=second_item, caching=True, mock_response="0.3,0.4"
)
combined: Final = await aembedding(
model="text-embedding-3-small",
input=[first_item, second_item],
caching=True,
mock_response="0.5,0.6",
)
assert combined.data[0].embedding == first.data[0].embedding
assert combined.data[1].embedding == second.data[0].embedding
assert combined._hidden_params["cache_hit"] is True
@pytest.mark.asyncio
async def test_caching_with_cache_controls(monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setattr(litellm, "cache", Cache(type=LiteLLMCacheType.LOCAL))
sync_no_ttl_messages: Final = [{"role": "user", "content": f"sync controls {uuid.uuid4().hex}"}]
sync_no_ttl_first: Final = completion(
model="gpt-4o-mini", messages=sync_no_ttl_messages, cache={"ttl": 0}, mock_response="first"
)
sync_no_ttl_second: Final = completion(
model="gpt-4o-mini", messages=sync_no_ttl_messages, cache={"s-maxage": 10}, mock_response="second"
)
sync_ttl_messages: Final = [{"role": "user", "content": f"sync ttl {uuid.uuid4().hex}"}]
sync_ttl_first: Final = completion(
model="gpt-4o-mini", messages=sync_ttl_messages, cache={"ttl": 25}, mock_response="third"
)
sync_ttl_second: Final = completion(
model="gpt-4o-mini", messages=sync_ttl_messages, cache={"s-maxage": 25}, mock_response="fourth"
)
async_no_ttl_messages: Final = [{"role": "user", "content": f"async controls {uuid.uuid4().hex}"}]
async_no_ttl_first: Final = await acompletion(
model="gpt-4o-mini", messages=async_no_ttl_messages, cache={"ttl": 0}, mock_response="first"
)
await asyncio.gather(*_PENDING_CACHE_WRITES)
async_no_ttl_second: Final = await acompletion(
model="gpt-4o-mini", messages=async_no_ttl_messages, cache={"s-maxage": 10}, mock_response="second"
)
async_ttl_messages: Final = [{"role": "user", "content": f"async ttl {uuid.uuid4().hex}"}]
async_ttl_first: Final = await acompletion(
model="gpt-4o-mini", messages=async_ttl_messages, cache={"ttl": 25}, mock_response="third"
)
await asyncio.gather(*_PENDING_CACHE_WRITES)
async_ttl_second: Final = await acompletion(
model="gpt-4o-mini", messages=async_ttl_messages, cache={"s-maxage": 25}, mock_response="fourth"
)
assert sync_no_ttl_first.id != sync_no_ttl_second.id
assert sync_ttl_first.id == sync_ttl_second.id
assert async_no_ttl_first.id != async_no_ttl_second.id
assert async_ttl_first.id == async_ttl_second.id
@pytest.mark.asyncio
async def test_embedding_caching_redis_ttl(monkeypatch: pytest.MonkeyPatch, redis_client_without_network: None) -> None:
mock_pipeline: Final = AsyncMock()
mock_set: Final = AsyncMock()
mock_pipeline.__aenter__.return_value.set = mock_set
monkeypatch.setattr("redis.asyncio.Redis.pipeline", lambda *args, **kwargs: mock_pipeline)
monkeypatch.setattr(
litellm,
"cache",
Cache(type=LiteLLMCacheType.REDIS, host="unused", port="6379", default_in_redis_ttl=2),
)
await aembedding(
model="text-embedding-3-small",
input=[f"redis ttl embedding {uuid.uuid4().hex}"],
encoding_format="base64",
caching=True,
mock_response="0.1,0.2",
)
await asyncio.gather(*_PENDING_CACHE_WRITES)
assert [call.kwargs["ex"] for call in mock_set.call_args_list] == [timedelta(seconds=2)]
@pytest.mark.asyncio
@pytest.mark.parametrize("sync_mode", [True, False])
async def test_logging_turn_off_message_logging_streaming(
sync_mode: bool, monkeypatch: pytest.MonkeyPatch, drained_logging_worker: None
) -> None:
cache: Final = Cache(type=LiteLLMCacheType.LOCAL)
monkeypatch.setattr(litellm, "cache", cache)
monkeypatch.setattr(litellm, "turn_off_message_logging", True)
messages: Final = [{"role": "user", "content": f"stream {uuid.uuid4().hex}"}]
if sync_mode:
stream: Final = completion(
model="gpt-4o-mini", messages=messages, caching=True, stream=True, mock_response="cached"
)
tuple(stream)
else:
stream: Final = await acompletion(
model="gpt-4o-mini", messages=messages, caching=True, stream=True, mock_response="cached"
)
tuple([chunk async for chunk in stream])
for _ in range(50):
await asyncio.sleep(0)
await GLOBAL_LOGGING_WORKER.flush()
await asyncio.gather(*_PENDING_CACHE_WRITES)
if cache.cache.cache_dict:
break
assert len(cache.cache.cache_dict) == 1
cached_entry: Final = next(iter(cache.cache.cache_dict.values()))
cached_response: Final = json.loads(cached_entry["response"])
assert cached_response["choices"][0]["message"]["content"] == "cached"
def test_redis_caching_default_ttl(monkeypatch: pytest.MonkeyPatch, redis_client_without_network: None) -> None:
monkeypatch.setattr(litellm, "default_redis_ttl", 120)
cache: Final = RedisCache(host="127.0.0.1", port="6379")
assert cache.default_ttl == 120
def test_redis_caching_llm_caching_ttl(monkeypatch: pytest.MonkeyPatch, redis_client_without_network: None) -> None:
monkeypatch.setattr(litellm, "default_redis_ttl", 120)
cache: Final = RedisCache(host="127.0.0.1", port="6379")
set_cache: Final = MagicMock()
monkeypatch.setattr(cache.redis_client, "set", set_cache)
cache.set_cache(key="cache-key", value="cache-value")
assert set_cache.call_args.kwargs["ex"] == 120
@pytest.mark.asyncio
async def test_redis_caching_ttl_pipeline(monkeypatch: pytest.MonkeyPatch, redis_client_without_network: None) -> None:
monkeypatch.setattr(litellm, "default_redis_ttl", 120)
cache: Final = RedisCache(host="127.0.0.1", port="6379")
pipeline: Final = AsyncMock()
set_pipeline: Final = MagicMock()
monkeypatch.setattr(pipeline, "set", set_pipeline)
await cache._pipeline_helper(
pipe=pipeline,
cache_list=[("first", "one"), ("second", "two")],
ttl=None,
)
assert set_pipeline.call_args_list == [
call(name="first", value='"one"', ex=timedelta(seconds=120)),
call(name="second", value='"two"', ex=timedelta(seconds=120)),
]
@pytest.mark.asyncio
async def test_redis_caching_ttl_sadd(monkeypatch: pytest.MonkeyPatch, redis_client_without_network: None) -> None:
monkeypatch.setattr(litellm, "default_redis_ttl", 120)
cache: Final = RedisCache(host="127.0.0.1", port="6379")
redis_client: Final = AsyncMock()
await cache._set_cache_sadd_helper(
redis_client=redis_client,
key="cache-key",
value=["cache-value"],
ttl=None,
)
redis_client.expire.assert_awaited_once_with("cache-key", timedelta(seconds=120))
def test_cache_key_separates_model_ids_unless_a_model_group_owns_the_key() -> None:
cache: Final = Cache(type=LiteLLMCacheType.LOCAL)
variants: Final[tuple[dict[str, object], ...]] = (
{"model_id": "component-a"},
{"model_id": "component-b"},
{},
{"model_id": "component-a"},
{"model_id": "component-a", "metadata": {"model_group": "embed"}},
{"model_id": "component-b", "metadata": {"model_group": "embed"}},
)
key_a, key_b, bare, key_a_again, grouped_a, grouped_b = (
cache.get_cache_key(model="sagemaker/openai/embed", input=["hello"], **extra) for extra in variants
)
assert key_a != key_b
assert key_a != bare
assert key_a == key_a_again
assert grouped_a == grouped_b
Reference in a new issue View git blame Copy permalink