diff --git a/litellm/integrations/anthropic_cache_control_hook.py b/litellm/integrations/anthropic_cache_control_hook.py index 4f9b18713d0..494d9e0935a 100644 --- a/litellm/integrations/anthropic_cache_control_hook.py +++ b/litellm/integrations/anthropic_cache_control_hook.py @@ -25,7 +25,10 @@ from litellm.integrations.prompt_management_base import PromptManagementClient from litellm.litellm_core_utils.prompt_templates.common_utils import ( with_prompt_cache_breakpoint, ) -from litellm.llms.anthropic.common_utils import is_claude_code_one_shot_subagent_request +from litellm.llms.anthropic.common_utils import ( + is_claude_code_one_shot_subagent_request, + supports_anthropic_cache_control, +) from litellm.types.integrations.anthropic_cache_control_hook import ( GATEWAY_INJECTED_CACHE_METADATA_KEY, GATEWAY_INJECTED_FOR_EVERY_DEPLOYMENT, @@ -574,8 +577,9 @@ class AnthropicCacheControlHook(CustomPromptManagement): custom_llm_provider: str | None, api_base: object, prompt_cache_options: object, + request_kwargs: object, ) -> Sequence[Mapping[str, object]] | None: - if AnthropicCacheControlHook._should_stand_down(points, messages, None, tools, cache_control): + if AnthropicCacheControlHook._should_stand_down(points, messages, None, tools, cache_control, request_kwargs): return None return AnthropicCacheControlHook._stamped_with_dialect( points, model, custom_llm_provider, api_base, prompt_cache_options @@ -612,6 +616,7 @@ class AnthropicCacheControlHook(CustomPromptManagement): system: str | list | None, tools: list | None, cache_control: object = None, + request_kwargs: object = None, ) -> bool: """Whether configured injection points must yield to client-set cache_control. @@ -624,7 +629,9 @@ class AnthropicCacheControlHook(CustomPromptManagement): """ if all(point.get("_litellm_judged") for point in points): return False - return AnthropicCacheControlHook._request_has_cache_control(messages, system, tools, cache_control) + return AnthropicCacheControlHook._request_has_cache_control( + messages, system, tools, cache_control, request_kwargs + ) @staticmethod def _request_has_cache_control( @@ -632,31 +639,29 @@ class AnthropicCacheControlHook(CustomPromptManagement): system: str | list | None, tools: list | None = None, cache_control: object = None, + request_kwargs: object = None, ) -> bool: - """Return True if the request already carries any client-supplied cache_control. - - When the client (e.g. Claude Code) already marks its own breakpoints we - stand down entirely rather than add more, per the auto-caching contract. - Tools count: they are a breakpoint the client can mark, they count toward - the provider's four-block limit, and caching only the tool definitions is - a common pattern, so injecting alongside them can exceed the cap. Tools - carry the mark either at the top level (Anthropic shape) or nested under - ``function`` (OpenAI shape); the Anthropic chat transform accepts both. - """ - if cache_control is not None: - return True - if AnthropicCacheControlHook.count_request_cache_breakpoints(messages, system) > 0: - return True - if tools is not None: - return any( - isinstance(tool, dict) - and ( - tool.get("cache_control") is not None - or (isinstance(tool.get("function"), dict) and tool["function"].get("cache_control") is not None) - ) - for tool in tools + """Client breakpoints own caching in both the request and its extra_body envelope.""" + bodies: Final = ( + {"messages": messages, "system": system, "tools": tools, "cache_control": cache_control}, + _validated_object_mapping(AnthropicCacheControlHook._request_value(request_kwargs, "extra_body")) or {}, + ) + return any( + body.get("cache_control") is not None + or AnthropicCacheControlHook.count_request_cache_breakpoints( + _validated_object_list(body.get("messages")) or (), body.get("system") ) - return False + > 0 + or any( + AnthropicCacheControlHook._request_value(tool, "cache_control") is not None + or AnthropicCacheControlHook._request_value( + AnthropicCacheControlHook._request_value(tool, "function"), "cache_control" + ) + is not None + for tool in (_validated_object_list(body.get("tools")) or ()) + ) + for body in bodies + ) @staticmethod def get_default_injection_points( @@ -676,36 +681,19 @@ class AnthropicCacheControlHook(CustomPromptManagement): even when the global flag is off. Caches the system prompt and the trailing turn, so the stable prefix (system + tools + history) is reused while the breakpoint advances with the conversation. Returns [] - (stand down) when neither flag is on, the provider does not consume - cache_control breakpoints (only anthropic / bedrock do), the model - lacks prompt-caching support, or the request already carries - client-supplied cache_control. + (stand down) when neither flag is on, the model is not Claude on a + supported explicit-cache transport, the model lacks prompt-caching + support, or the request already carries client-supplied cache_control. """ import litellm if litellm.enable_anthropic_prompt_caching is not True and enable_prompt_caching is not True: return [] - provider = custom_llm_provider - if provider is None: - from litellm.litellm_core_utils.get_llm_provider_logic import ( - get_llm_provider, - ) - - try: - _, provider, _, _ = get_llm_provider(model=model) - except Exception: # noqa: BLE001 # unroutable model must never block the call, just skip auto-caching - return [] - - if provider not in ("anthropic", "bedrock"): + if not supports_anthropic_cache_control(model, custom_llm_provider): return [] - from litellm.utils import supports_prompt_caching - - if not supports_prompt_caching(model=model, custom_llm_provider=provider): - return [] - - if AnthropicCacheControlHook._request_has_cache_control(messages, system, tools, cache_control): + if AnthropicCacheControlHook._request_has_cache_control(messages, system, tools, cache_control, request_kwargs): return [] if is_claude_code_one_shot_subagent_request( @@ -737,13 +725,15 @@ class AnthropicCacheControlHook(CustomPromptManagement): prompt and trailing turn) do not depend on which deployment serves the call. Returns the input list itself when auto-injection would not apply """ + import litellm + points: Final = next( ( candidate for candidate in ( AnthropicCacheControlHook.get_default_injection_points( messages=messages, - model=model, + model=litellm.model_alias_map.get(model, model), custom_llm_provider=None, tools=tools, enable_prompt_caching=enable_prompt_caching, @@ -789,6 +779,8 @@ class AnthropicCacheControlHook(CustomPromptManagement): prompt-management gate and the AnthropicCacheControlHook run unchanged. """ + import litellm + if non_default_params.get("cache_control_injection_points"): judged: Final = AnthropicCacheControlHook._judged_configured_points( non_default_params["cache_control_injection_points"], @@ -799,6 +791,7 @@ class AnthropicCacheControlHook(CustomPromptManagement): custom_llm_provider, api_base, non_default_params.get("prompt_cache_options"), + non_default_params, ) if judged is None: non_default_params.pop("cache_control_injection_points") @@ -808,7 +801,7 @@ class AnthropicCacheControlHook(CustomPromptManagement): points: Final = AnthropicCacheControlHook.get_default_injection_points( messages=messages, system=None, - model=model, + model=litellm.model_alias_map.get(model, model), custom_llm_provider=custom_llm_provider, tools=tools, enable_prompt_caching=enable_prompt_caching, @@ -925,7 +918,7 @@ class AnthropicCacheControlHook(CustomPromptManagement): list[CacheControlInjectionPoint] | None, kwargs.pop("cache_control_injection_points", None) ) if configured and AnthropicCacheControlHook._should_stand_down( - configured, typed_messages, system, tools, cache_control + configured, typed_messages, system, tools, cache_control, kwargs ): return messages, system injection_points: list[CacheControlInjectionPoint] = configured or [] diff --git a/litellm/llms/anthropic/common_utils.py b/litellm/llms/anthropic/common_utils.py index 3328e20e2ea..98e2f6d5bde 100644 --- a/litellm/llms/anthropic/common_utils.py +++ b/litellm/llms/anthropic/common_utils.py @@ -77,6 +77,21 @@ _CLAUDE_CODE_OBJECT_LIST_ADAPTER: Final = TypeAdapter(list[object]) _CLAUDE_CODE_USER_AGENT_PREFIXES: Final = ("claude-cli/", "claude-code/") +def supports_anthropic_cache_control(model: str, custom_llm_provider: str | None) -> bool: + from litellm.litellm_core_utils.get_llm_provider_logic import get_llm_provider + from litellm.utils import supports_prompt_caching + + try: + provider: Final = custom_llm_provider if custom_llm_provider is not None else get_llm_provider(model=model)[1] + except Exception: # noqa: BLE001 # Optional caching must not block an unroutable request + return False + return ( + provider in ("anthropic", "bedrock", "vertex_ai", "azure_ai") + and "claude" in model.lower() + and supports_prompt_caching(model=model, custom_llm_provider=provider) + ) + + def is_claude_code_user_agent(user_agent: str) -> bool: """Claude Code sends its API calls through the Anthropic SDK as `claude-cli/` and its own fetches, such as gateway model discovery, as `claude-code/`""" diff --git a/litellm/proxy/management_endpoints/key_management_endpoints.py b/litellm/proxy/management_endpoints/key_management_endpoints.py index 4554a85b225..40bd496fdce 100644 --- a/litellm/proxy/management_endpoints/key_management_endpoints.py +++ b/litellm/proxy/management_endpoints/key_management_endpoints.py @@ -1968,7 +1968,7 @@ async def generate_key_fn( - policies: Optional[List[str]] - List of policy names to apply to the key. Policies define guardrails, conditions, and inheritance rules. - disable_global_guardrails: Optional[bool] - Whether to disable global guardrails for the key. - throttle_on_budget_exceeded: Optional[bool] - When the key exceeds its max_budget, throttle its tpm/rpm to the global budget_exceeded_throttle_percentage instead of blocking the key entirely. - - enable_prompt_caching: Optional[bool] - Auto-inject prompt caching breakpoints (Anthropic cache_control markers) on requests made with this key. Anthropic and Bedrock Claude models only. + - enable_prompt_caching: Optional[bool] - Auto-inject prompt caching breakpoints (Anthropic cache_control markers) on requests made with this key. Supported Claude models on Anthropic, Bedrock, Vertex AI, and Azure AI only. - permissions: Optional[dict] - key-specific permissions. Currently just used for turning off pii masking (if connected). Example - {"pii": false} - model_max_budget: Optional[Dict[str, BudgetConfig]] - Model-specific budgets {"gpt-4": {"budget_limit": 0.0005, "time_period": "30d"}}}. IF null or {} then no model specific budget. - budget_fallbacks: Optional[Dict[str, List[str]]] - Per-model fallback chain tried in order when that model's own `model_max_budget` is exceeded, e.g. {"gpt-4o": ["gpt-4o-mini"]}. @@ -3323,7 +3323,7 @@ async def update_key_fn( - policies: Optional[List[str]] - List of policy names to apply to the key. Policies define guardrails, conditions, and inheritance rules. - disable_global_guardrails: Optional[bool] - Whether to disable global guardrails for the key. - throttle_on_budget_exceeded: Optional[bool] - When the key exceeds its max_budget, throttle its tpm/rpm to the global budget_exceeded_throttle_percentage instead of blocking the key entirely. - - enable_prompt_caching: Optional[bool] - Auto-inject prompt caching breakpoints (Anthropic cache_control markers) on requests made with this key. Anthropic and Bedrock Claude models only. + - enable_prompt_caching: Optional[bool] - Auto-inject prompt caching breakpoints (Anthropic cache_control markers) on requests made with this key. Supported Claude models on Anthropic, Bedrock, Vertex AI, and Azure AI only. - prompts: Optional[List[str]] - List of prompts that the key is allowed to use. - blocked: Optional[bool] - Whether the key is blocked - aliases: Optional[dict] - Model aliases for the key - [Docs](https://litellm.vercel.app/docs/proxy/virtual_keys#model-aliases) diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index c6ce2e61b3c..3c7d06268ad 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -17708,8 +17708,8 @@ _GENERAL_SETTINGS_UI_LITELLM_FIELDS: Final[dict[str, GeneralSettingsUILiteLLMFie "type": "Boolean", "tab": "prompt_caching", "description": ( - "Auto-adds cache_control to the system prompt and trailing turn for supported Anthropic " - "and Bedrock Claude models. The cache is shared across callers on the same upstream credentials." + "Auto-adds cache_control to the system prompt and trailing turn for supported Claude models on " + "Anthropic, Bedrock, Vertex AI, and Azure AI. The cache is shared across callers on the same upstream credentials." ), }, "anthropic_prompt_caching_ttl": { diff --git a/tests/test_litellm/integrations/test_anthropic_cache_control_hook.py b/tests/test_litellm/integrations/test_anthropic_cache_control_hook.py index 92b1185e542..83649c3386a 100644 --- a/tests/test_litellm/integrations/test_anthropic_cache_control_hook.py +++ b/tests/test_litellm/integrations/test_anthropic_cache_control_hook.py @@ -1595,6 +1595,178 @@ class TestEnableAnthropicPromptCaching: assert supports_prompt_caching(model=model, custom_llm_provider=provider) is True assert self._points(model=model, provider=provider) == [] + @pytest.mark.parametrize("family", ["haiku-4-5", "sonnet-5", "opus-5", "fable-5", "fable-5-1"]) + @pytest.mark.parametrize( + "provider, template", + [("anthropic", "{}"), ("vertex_ai", "{}"), ("azure_ai", "{}"), ("bedrock", "us.anthropic.{}-v1:0")], + ) + @pytest.mark.parametrize("infer_provider", [False, True]) + @pytest.mark.parametrize("supported", [False, True]) + def test_claude_transport_defaults(self, monkeypatch, local_model_cost_map, family, provider, template, infer_provider, supported): + from litellm.utils import supports_prompt_caching + + model = template.format(f"claude-{family}") + qualified = f"{provider}/{model}" + entry = {"litellm_provider": provider, "mode": "chat", "supports_prompt_caching": supported} + monkeypatch.setitem(litellm.model_cost, model, entry) + monkeypatch.setitem(litellm.model_cost, qualified, entry) + monkeypatch.setattr(litellm, "enable_anthropic_prompt_caching", False) + target = qualified if infer_provider else model + resolved_provider = None if infer_provider else provider + assert supports_prompt_caching(model=target, custom_llm_provider=resolved_provider) is supported + points = AnthropicCacheControlHook.get_default_injection_points( + messages=copy.deepcopy(self.MESSAGES), system=None, model=target, + custom_llm_provider=resolved_provider, enable_prompt_caching=True, + ) + assert [point["index"] for point in points] == ([None, -1] if supported else []) + affinity_messages = AnthropicCacheControlHook.messages_with_default_injections( + copy.deepcopy(self.MESSAGES), models=[qualified], enable_prompt_caching=True, + ) + assert sum(AnthropicCacheControlHook._count_cache_control_blocks(m) for m in affinity_messages) == (2 if supported else 0) + + @pytest.mark.parametrize( + "provider, model", + [ + ("bedrock", "us.openai.gpt-6-astra"), + ("bedrock", "amazon.nova-pro-v1:0"), + ("bedrock", "us.xai.grok-4.6"), + ("bedrock", "arn:aws:bedrock:us-east-1:123456789012:application-inference-profile/opaque"), + ("vertex_ai", "gemini-3.8-flash"), + ("azure_ai", "gpt-6-astra"), + ("anthropic", "unknown-model"), + ], + ) + def test_non_claude_caching_capability_does_not_enable_defaults(self, monkeypatch, local_model_cost_map, provider, model): + from litellm.utils import supports_prompt_caching + + qualified = f"{provider}/{model}" + entry = {"litellm_provider": provider, "mode": "chat", "supports_prompt_caching": True} + monkeypatch.setitem(litellm.model_cost, model, entry) + monkeypatch.setitem(litellm.model_cost, qualified, entry) + monkeypatch.setattr(litellm, "enable_anthropic_prompt_caching", True) + assert supports_prompt_caching(model=model, custom_llm_provider=provider) + assert self._points(model=model, provider=provider) == [] + assert self._points(model=qualified, provider=None) == [] + assert AnthropicCacheControlHook.messages_with_default_injections(self.MESSAGES, [qualified]) == self.MESSAGES + + @pytest.mark.parametrize("provider", ["vertex_ai", "azure_ai"]) + @pytest.mark.parametrize("client_control", ["none", "message", "system", "tool", "function", "top_level"]) + @pytest.mark.parametrize("envelope", ["request", "extra_body"]) + @pytest.mark.parametrize("configured", [False, True]) + def test_new_transports_preserve_client_controls(self, monkeypatch, local_model_cost_map, provider, client_control, envelope, configured): + from litellm.llms.vertex_ai.vertex_ai_partner_models.anthropic.transformation import VertexAIAnthropicConfig + + model = "claude-sonnet-5" + monkeypatch.setattr(litellm, "enable_anthropic_prompt_caching", True) + monkeypatch.setitem(litellm.model_cost, f"{provider}/{model}", { + **litellm.model_cost[f"{provider}/{model}"], "supports_prompt_caching": True, + }) + control = {"type": "ephemeral"} + messages = [{"role": "user", "content": [{"type": "text", "text": "question", **({"cache_control": control} if client_control == "message" else {})}]}] + system = [{"type": "text", "text": "stable context", **({"cache_control": control} if client_control == "system" else {})}] + tools = [{"name": "lookup", "description": "Lookup", "input_schema": {"type": "object", "properties": {}}, **({"cache_control": control} if client_control == "tool" else {})}] + if client_control == "function": + tools = [{"type": "function", "function": {"name": "lookup", "parameters": {}, "cache_control": control}}] + kwargs = {"metadata": {}, "model_info": {"id": "selected-deployment"}, **({"cache_control": control} if client_control == "top_level" else {})} + if envelope == "extra_body": + kwargs["extra_body"] = {"messages": messages, "system": system, "tools": tools} + if "cache_control" in kwargs: + kwargs["extra_body"]["cache_control"] = kwargs.pop("cache_control") + messages, system, tools = [{"role": "user", "content": "question"}], "stable context", [] + if configured: + kwargs["cache_control_injection_points"] = [ + {"location": "message", "role": "system", "index": None, "control": control}, + {"location": "message", "role": None, "index": -1, "control": control}, + ] + seeded = copy.deepcopy(kwargs) + original = copy.deepcopy((messages, system, tools)) + result_messages, result_system = AnthropicCacheControlHook.maybe_inject_cache_control( + messages, system, kwargs, model, provider, tools=tools, + ) + if client_control != "none": + assert (result_messages, result_system, tools) == original + assert kwargs["metadata"] == {} + else: + assert kwargs["metadata"]["litellm_gateway_injected_cache"] == "selected-deployment" + assert sum(AnthropicCacheControlHook._count_cache_control_blocks(m) for m in result_messages) == 1 + assert result_system[0]["cache_control"] == control + if provider == "vertex_ai": + wire = VertexAIAnthropicConfig().transform_request( + model=model, messages=[{"role": "system", "content": result_system}, *result_messages], + optional_params={"max_tokens": 8}, litellm_params={}, headers={}, + ) + assert wire["system"][0]["cache_control"] == control + assert wire["messages"][-1]["content"][-1]["cache_control"] == control + affinity = AnthropicCacheControlHook.messages_with_default_injections( + [{"role": "system", "content": original[1]}, *original[0]], [f"{provider}/{model}"], + tools=tools, request_kwargs=seeded, + ) + if client_control != "none": + assert affinity == [{"role": "system", "content": original[1]}, *original[0]] + AnthropicCacheControlHook.maybe_seed_default_injection_points( + seeded, [{"role": "system", "content": original[1]}, *original[0]], model, provider, tools=tools, + ) + assert bool(seeded.get("cache_control_injection_points")) == (client_control == "none") + + @pytest.mark.asyncio + @pytest.mark.parametrize("asynchronous", [False, True]) + @pytest.mark.parametrize("model, target, client_control, expected", [ + ("vertex_ai/claude-sonnet-5", "bedrock/amazon.nova-pro-v1:0", False, 0), + ("azure_ai/gpt-6-astra", "azure_ai/claude-sonnet-5", False, 2), + ("azure_ai/claude-sonnet-5", None, False, 2), + ("azure_ai/claude-sonnet-5", None, True, 1), + ("azure_ai/model_router/claude-replacement", None, False, 2), + ]) + async def test_public_completion_cache_ownership(self, monkeypatch, local_model_cost_map, asynchronous, model, target, client_control, expected): + import httpx + from litellm.llms.custom_httpx.http_handler import HTTPHandler + + monkeypatch.setattr(litellm, "enable_anthropic_prompt_caching", True) + monkeypatch.setattr(litellm, "model_alias_map", {model: target} if target else {}) + for qualified in (model, target): + if qualified: + provider = qualified.split("/")[0] + entry = {"litellm_provider": provider, "mode": "chat", "supports_prompt_caching": True} + monkeypatch.setitem(litellm.model_cost, qualified, entry) + monkeypatch.setitem(litellm.model_cost, qualified.split("/", 1)[-1], entry) + sent = [] + def respond(request): + sent.append(json.loads(request.content)) + return httpx.Response(200, request=request, json={ + "id": "msg-test", "type": "message", "role": "assistant", "model": "claude-sonnet-5", + "content": [{"type": "text", "text": "ok"}], "stop_reason": "end_turn", "stop_sequence": None, + "output": {"message": {"role": "assistant", "content": [{"text": "ok"}]}}, "stopReason": "end_turn", + "usage": {"input_tokens": 10, "output_tokens": 1, "inputTokens": 10, "outputTokens": 1, "totalTokens": 11}, + }) + control = {"type": "ephemeral", "ttl": "1h"} + messages = [{"role": "system", "content": "stable context"}, {"role": "user", "content": "question"}] + metadata = {} + kwargs = { + "model": model, "messages": copy.deepcopy(messages), "max_tokens": 32, "num_retries": 0, + "litellm_metadata": metadata, + "api_base": "https://rig.services.ai.azure.com/anthropic", "api_key": "synthetic-test-key", + "aws_access_key_id": "synthetic", "aws_secret_access_key": "synthetic", "aws_region_name": "us-east-1", + **({"extra_body": {"cache_control": control}} if client_control else {}), + } + if asynchronous: + handler = AsyncHTTPHandler() + await handler.client.aclose() + async with httpx.AsyncClient(transport=httpx.MockTransport(respond)) as client: + handler.client = client + response = await litellm.acompletion(**kwargs, client=handler) + else: + with httpx.Client(transport=httpx.MockTransport(respond)) as client: + response = litellm.completion(**kwargs, client=HTTPHandler(client=client)) + assert response.choices[0].message.content == "ok" + assert len(sent) == 1 + assert ("litellm_gateway_injected_cache" in metadata) == (expected == 2) + serialized = json.dumps(sent[0]) + assert serialized.count('"cache_control"') + serialized.count('"cachePoint"') == expected + if client_control: + assert sent[0]["cache_control"] == control + affinity = AnthropicCacheControlHook.messages_with_default_injections(messages, [model], request_kwargs=kwargs) + assert AnthropicCacheControlHook.count_request_cache_breakpoints(affinity) == (2 if expected == 2 else 0) + def test_databricks_claude_not_injected_despite_caching_support(self, monkeypatch, local_model_cost_map): from litellm.utils import supports_prompt_caching diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index d43adfe1ae4..4fe8bff3da8 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -7805,7 +7805,7 @@ export interface paths { * - policies: Optional[List[str]] - List of policy names to apply to the key. Policies define guardrails, conditions, and inheritance rules. * - disable_global_guardrails: Optional[bool] - Whether to disable global guardrails for the key. * - throttle_on_budget_exceeded: Optional[bool] - When the key exceeds its max_budget, throttle its tpm/rpm to the global budget_exceeded_throttle_percentage instead of blocking the key entirely. - * - enable_prompt_caching: Optional[bool] - Auto-inject prompt caching breakpoints (Anthropic cache_control markers) on requests made with this key. Anthropic and Bedrock Claude models only. + * - enable_prompt_caching: Optional[bool] - Auto-inject prompt caching breakpoints (Anthropic cache_control markers) on requests made with this key. Supported Claude models on Anthropic, Bedrock, Vertex AI, and Azure AI only. * - permissions: Optional[dict] - key-specific permissions. Currently just used for turning off pii masking (if connected). Example - {"pii": false} * - model_max_budget: Optional[Dict[str, BudgetConfig]] - Model-specific budgets {"gpt-4": {"budget_limit": 0.0005, "time_period": "30d"}}}. IF null or {} then no model specific budget. * - budget_fallbacks: Optional[Dict[str, List[str]]] - Per-model fallback chain tried in order when that model's own `model_max_budget` is exceeded, e.g. {"gpt-4o": ["gpt-4o-mini"]}. @@ -8286,7 +8286,7 @@ export interface paths { * - policies: Optional[List[str]] - List of policy names to apply to the key. Policies define guardrails, conditions, and inheritance rules. * - disable_global_guardrails: Optional[bool] - Whether to disable global guardrails for the key. * - throttle_on_budget_exceeded: Optional[bool] - When the key exceeds its max_budget, throttle its tpm/rpm to the global budget_exceeded_throttle_percentage instead of blocking the key entirely. - * - enable_prompt_caching: Optional[bool] - Auto-inject prompt caching breakpoints (Anthropic cache_control markers) on requests made with this key. Anthropic and Bedrock Claude models only. + * - enable_prompt_caching: Optional[bool] - Auto-inject prompt caching breakpoints (Anthropic cache_control markers) on requests made with this key. Supported Claude models on Anthropic, Bedrock, Vertex AI, and Azure AI only. * - prompts: Optional[List[str]] - List of prompts that the key is allowed to use. * - blocked: Optional[bool] - Whether the key is blocked * - aliases: Optional[dict] - Model aliases for the key - [Docs](https://litellm.vercel.app/docs/proxy/virtual_keys#model-aliases)