fix(gemini): drop ttl values outside the protobuf Duration range and the explanatory docstrings

litellm/llms/vertex_ai/context_caching/transformation.py

@@ -89,24 +89,18 @@ def extract_ttl_from_cached_messages(messages: list[AllMessageValues]) -> str |
 
 _TTL_PATTERN: Final = re.compile(r"^([0-9]*\.?[0-9]+)([smh])$")
 _TTL_UNIT_SECONDS: Final = MappingProxyType({"s": 1, "m": 60, "h": 3600})
+_PROTOBUF_DURATION_MAX_SECONDS: Final = 315_576_000_000
 
 
 def _normalize_ttl_to_seconds(ttl: object) -> str | None:
-    """
-    Gemini's cachedContents API only takes a TTL as "<seconds>s", while Anthropic clients
-    (Claude Code among them) send the minute and hour units the Anthropic API defines, "5m"
-    and "1h". Returns the Gemini form for any of the three units, or None for a missing,
-    non-positive, or unparseable value so the cache falls back to Gemini's default TTL.
-    """
     if not isinstance(ttl, str):
         return None
     match: Final = _TTL_PATTERN.match(ttl)
     if match is None:
         return None
-    value: Final = float(match.group(1))
-    if value <= 0:
+    seconds: Final = round(float(match.group(1)) * _TTL_UNIT_SECONDS[match.group(2)], 9)
+    if not 0 < seconds <= _PROTOBUF_DURATION_MAX_SECONDS:
         return None
-    seconds: Final = round(value * _TTL_UNIT_SECONDS[match.group(2)], 9)
     return f"{seconds:.9f}".rstrip("0").rstrip(".") + "s"
 
 

tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py

@@ -2102,11 +2102,7 @@ def test_should_add_cache_control_for_anthropic_model():
 
 
 def test_should_not_add_cache_control_for_non_anthropic_model():
-    """Should not add cache_control for providers that reject an explicit cache_control field.
-
-    OpenAI/Azure do prompt caching implicitly and 400 on an unexpected
-    cache_control field, so it must not be forwarded to them.
-    """
+    """Should not add cache_control for non-Anthropic models."""
     adapter = LiteLLMAnthropicMessagesAdapter()
     cache_control = {"type": "ephemeral"}
 
@@ -2122,13 +2118,6 @@ def test_should_not_add_cache_control_for_non_anthropic_model():
 
 
 def test_should_add_cache_control_for_gemini_model():
-    """Should add cache_control for Gemini / Vertex Gemini targets.
-
-    These consume anthropic-style cache_control blocks via the Gemini context
-    caching path, so /v1/messages requests (e.g. Claude Code) routed to a
-    Gemini model must keep it. Regression for the adapter dropping the field
-    before it reaches the Gemini transformation.
-    """
     adapter = LiteLLMAnthropicMessagesAdapter()
     cache_control = {"type": "ephemeral", "ttl": "1h"}
 
@@ -2146,7 +2135,6 @@ def test_should_add_cache_control_for_gemini_model():
 
 
 def test_cache_control_preserved_in_text_content_for_gemini():
-    """cache_control must survive message translation for a Gemini target."""
     anthropic_messages = [
         AnthropicMessagesUserMessageParam(
             role="user",

tests/test_litellm/llms/vertex_ai/context_caching/test_context_caching_ttl.py

@@ -7,8 +7,6 @@ from litellm.llms.vertex_ai.context_caching.transformation import (
 
 
 class TestTTLNormalization:
-    """Gemini only takes "<seconds>s"; Anthropic clients send "5m" and "1h" too"""
-
     @pytest.mark.parametrize(
         "ttl, expected",
         [
@@ -23,6 +21,8 @@ class TestTTLNormalization:
             ("1h", "3600s"),
             ("0.5h", "1800s"),
             ("48h", "172800s"),
+            ("315576000000s", "315576000000s"),
+            ("87660000h", "315576000000s"),
         ],
     )
     def test_normalizes_supported_units_to_seconds(self, ttl, expected):
@@ -44,6 +44,10 @@ class TestTTLNormalization:
             "3600 s",
             "3600ss",
             "1 h",
+            "0.0000000001s",
+            "315576000001s",
+            "87660001h",
+            "9" * 400 + "h",
             None,
             123,
         ],

tests/test_litellm/llms/vertex_ai/context_caching/test_vertex_ai_context_caching.py

@@ -1399,11 +1399,6 @@ class TestContextCachingEndpoints:
     def test_check_and_create_cache_skips_between_default_and_gemini_2_5_pro_minimum(
         self, local_model_cost_map
     ):
-        """Gemini 2.5 Pro needs 2048 cached tokens, twice the provider-agnostic default.
-
-        Content between the two used to reach Google's cachedContents endpoint and 400
-        with "Cached content is too small".
-        """
         model = "gemini-2.5-pro"
         self._token_check_patcher.stop()