diff --git a/litellm/constants.py b/litellm/constants.py index da731cb5eb2..25fdaec20de 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -215,6 +215,10 @@ MAX_CALLBACKS: Final = get_env_int("LITELLM_MAX_CALLBACKS", 100) # so the deployment-level hook does not re-run them for the same request PRE_CALL_EXECUTED_GUARDRAILS_KEY: Final = "_pre_call_executed_guardrails" +# Metadata key listing compression guardrails an auto router's own compression +# policy suppresses for this request. See litellm.proxy.guardrails.auto_router_compression. +AUTO_ROUTER_SUPPRESSED_COMPRESSION_GUARDRAILS_KEY: Final = "_auto_router_suppressed_compression_guardrails" + # Generic fallback for unknown models DEFAULT_REASONING_EFFORT_MINIMAL_THINKING_BUDGET: Final = int( os.getenv("DEFAULT_REASONING_EFFORT_MINIMAL_THINKING_BUDGET", 128) diff --git a/litellm/integrations/custom_guardrail.py b/litellm/integrations/custom_guardrail.py index 9bb613654e4..7f8effa2317 100644 --- a/litellm/integrations/custom_guardrail.py +++ b/litellm/integrations/custom_guardrail.py @@ -45,6 +45,7 @@ dc: Final = DualCache() from litellm.constants import ( + AUTO_ROUTER_SUPPRESSED_COMPRESSION_GUARDRAILS_KEY, GUARDRAIL_SCANNED_MESSAGES_CACHE_TTL_SECONDS, PRE_CALL_EXECUTED_GUARDRAILS_KEY, ) @@ -940,6 +941,20 @@ class CustomGuardrail(CustomLogger): """ return False + def _suppressed_by_auto_router_compression(self, data: dict) -> bool: + """True when an auto router's own compression policy suppresses this guardrail. + + Set only by litellm.proxy.guardrails.auto_router_compression.arm_pre_call, never + by the caller, so a request cannot suppress its own guardrails this way. + """ + for meta_key in ("metadata", "litellm_metadata"): + meta = data.get(meta_key) + if isinstance(meta, dict): + suppressed = meta.get(AUTO_ROUTER_SUPPRESSED_COMPRESSION_GUARDRAILS_KEY) + if isinstance(suppressed, list) and self.guardrail_name in suppressed: + return True + return False + def should_run_guardrail( self, data, @@ -948,6 +963,9 @@ class CustomGuardrail(CustomLogger): """ Returns True if the guardrail should be run on the event_type """ + if self._suppressed_by_auto_router_compression(data): + return False + requested_guardrails: Final = self.get_guardrail_from_metadata(data) disable_global_guardrail: Final = self.get_disable_global_guardrail(data) opted_out_global_guardrails: Final = self.get_opted_out_global_guardrails_from_metadata(data) diff --git a/litellm/proxy/common_request_processing.py b/litellm/proxy/common_request_processing.py index f25fa46197e..534b2db3e61 100644 --- a/litellm/proxy/common_request_processing.py +++ b/litellm/proxy/common_request_processing.py @@ -61,6 +61,7 @@ from litellm.proxy.common_utils.sse_keepalive import ( wrap_sse_stream_with_keepalive_pings, ) from litellm.proxy.dd_span_tagger import DDSpanTagger +from litellm.proxy.guardrails.auto_router_compression import arm_pre_call as _arm_auto_router_compression from litellm.proxy.route_llm_request import route_request from litellm.proxy.utils import ProxyLogging, _check_and_merge_model_level_guardrails from litellm.router import Router @@ -2004,6 +2005,12 @@ class ProxyBaseLLMRequestProcessing: trust_client_model_info=False, ) + # An auto router with its own compression policy is authoritative for this + # request: suppress every other compression guardrail and arm whichever one + # the policy names for the model call, before those guardrails get a chance + # to run below. + self.data = await _arm_auto_router_compression(data=self.data, llm_router=llm_router) + self.data = await proxy_logging_obj.pre_call_hook( user_api_key_dict=user_api_key_dict, data=self.data, diff --git a/litellm/proxy/guardrails/auto_router_compression.py b/litellm/proxy/guardrails/auto_router_compression.py new file mode 100644 index 00000000000..c3fba937d22 --- /dev/null +++ b/litellm/proxy/guardrails/auto_router_compression.py @@ -0,0 +1,211 @@ +""" +Decouples prompt compression between an auto router's routing decision and the +model it routes to. An auto router marker deployment may set +``auto_router_routing_compression`` and/or ``auto_router_model_compression`` in its +``litellm_params`` to name the compression guardrail that hop should use, or +``"none"`` to run no compression on that hop. Neither key set means the request's +own compression guardrails (key/team/model-level, or an "Always on" guardrail) +apply to both hops unchanged, exactly as before this feature existed. + +Once either key is set, this auto router is authoritative: every other compression +guardrail is suppressed for that request, and only these two settings decide what +each hop sees. +""" + +import copy +from collections.abc import Mapping +from dataclasses import dataclass +from typing import TYPE_CHECKING, Any, Final + +from litellm._logging import verbose_proxy_logger +from litellm.constants import AUTO_ROUTER_SUPPRESSED_COMPRESSION_GUARDRAILS_KEY +from litellm.litellm_core_utils.core_helpers import ( + get_metadata_variable_name_from_kwargs, + get_or_create_metadata_bucket, +) +from litellm.router_utils.auto_router_model_naming import AUTO_ROUTER_MODEL_PREFIX +from litellm.types.utils import GenericGuardrailAPIInputs + +if TYPE_CHECKING: + from litellm.integrations.custom_guardrail import CustomGuardrail + from litellm.router import Router +else: + CustomGuardrail = Any + Router = Any + +COMPRESSION_GUARDRAIL_PROVIDERS: Final = frozenset({"headroom", "compresr"}) +_NO_COMPRESSION: Final = "none" + +# Metadata key stashing the pre-compression messages so a routing decision that +# names a different compression than the model call still compresses the +# original text, not whatever the model-side guardrail already rewrote it to. +AUTO_ROUTER_ROUTING_MESSAGES_SNAPSHOT_KEY: Final = "_auto_router_routing_messages_snapshot" + + +@dataclass(frozen=True, slots=True) +class AutoRouterCompressionPolicy: + """An auto router's compression choice for each hop. ``None`` means no compression.""" + + routing: str | None + model: str | None + + @property + def is_same(self) -> bool: + return self.routing == self.model + + +def _normalized_compression_choice(raw: object) -> str | None: + if not isinstance(raw, str) or not raw: + return None + return None if raw.strip().lower() == _NO_COMPRESSION else raw + + +def policy_from_litellm_params(litellm_params: Mapping[str, object]) -> AutoRouterCompressionPolicy | None: + raw_routing: Final = litellm_params.get("auto_router_routing_compression") + raw_model: Final = litellm_params.get("auto_router_model_compression") + if raw_routing is None and raw_model is None: + return None + return AutoRouterCompressionPolicy( + routing=_normalized_compression_choice(raw_routing), + model=_normalized_compression_choice(raw_model), + ) + + +def policy_for_model( + llm_router: "Router | None", model_alias: str, team_id: str | None +) -> AutoRouterCompressionPolicy | None: + """The compression policy declared by the auto router marker deployment `model_alias` resolves to. + + Mirrors the alias lookup in ``_check_and_merge_model_level_guardrails``: this runs + before routing has picked a strategy, so it takes the first marker deployment for + the alias rather than disambiguating by request tags. + """ + if llm_router is None: + return None + deployments: Final = llm_router.get_model_list(model_name=model_alias, team_id=team_id) or [] + for deployment in deployments: + litellm_params: Final = deployment.get("litellm_params") or {} + model_field = litellm_params.get("model") + if not isinstance(model_field, str) or not model_field.startswith(AUTO_ROUTER_MODEL_PREFIX): + continue + policy = policy_from_litellm_params(litellm_params) + if policy is not None: + return policy + return None + + +def _active_compression_guardrail_names() -> frozenset[str]: + """Names of every currently-active guardrail whose type is a compression guardrail.""" + import litellm + from litellm.integrations.custom_guardrail import CustomGuardrail + from litellm.proxy.guardrails.guardrail_registry import guardrail_class_registry + + compression_classes: Final = tuple( + cls for name, cls in guardrail_class_registry.items() if name in COMPRESSION_GUARDRAIL_PROVIDERS + ) + if not compression_classes: + return frozenset() + active: Final = litellm.logging_callback_manager.get_custom_loggers_for_type(callback_type=CustomGuardrail) + return frozenset( + cb.guardrail_name for cb in active if isinstance(cb, compression_classes) and cb.guardrail_name + ) + + +async def arm_pre_call(data: dict, llm_router: "Router | None") -> dict: + """Apply an auto router's compression policy, if any, before guardrails run. + + Suppresses every other compression guardrail, re-enables the model-side + guardrail the policy names (if any) even when it isn't ``default_on``, and + snapshots the pre-compression messages so the routing decision can compress + them independently of whatever the model-side guardrail does to `data`. + """ + if llm_router is None: + return data + + model_alias: Final = data.get("model") + if not isinstance(model_alias, str) or not model_alias: + return data + + # Read-only until a policy is confirmed: creating the metadata bucket for every + # request, including the vast majority with no auto-router compression policy, + # would be an unwanted side effect of merely checking for one. + metadata_key: Final = get_metadata_variable_name_from_kwargs(data) + existing_bucket: Final = data.get(metadata_key) + other_bucket: Final = data.get("metadata" if metadata_key == "litellm_metadata" else "litellm_metadata") + team_id: Final = (existing_bucket.get("user_api_key_team_id") if isinstance(existing_bucket, dict) else None) or ( + other_bucket.get("user_api_key_team_id") if isinstance(other_bucket, dict) else None + ) + + policy: Final = policy_for_model(llm_router=llm_router, model_alias=model_alias, team_id=team_id) + if policy is None: + return data + + _, metadata = get_or_create_metadata_bucket(data) + suppressed: Final = _active_compression_guardrail_names() - ({policy.model} if policy.model else set()) + if suppressed: + metadata[AUTO_ROUTER_SUPPRESSED_COMPRESSION_GUARDRAILS_KEY] = sorted(suppressed) + + if policy.model is not None: + requested = metadata.get("guardrails") + if isinstance(requested, list): + if policy.model not in requested: + requested.append(policy.model) + else: + metadata["guardrails"] = [policy.model] + + from litellm.litellm_core_utils.prompt_templates.factory import resolve_structured_messages + + snapshot: Final = resolve_structured_messages(messages=data.get("messages"), request_kwargs=data) + if snapshot is not None: + metadata[AUTO_ROUTER_ROUTING_MESSAGES_SNAPSHOT_KEY] = copy.deepcopy(snapshot) + + return data + + +async def messages_for_routing( + policy: AutoRouterCompressionPolicy | None, + messages: list[dict[str, Any]] | None, + request_kwargs: Mapping[str, object], +) -> list[dict[str, Any]] | None: + """Messages to use for a routing decision, compressed per `policy.routing`. + + Returns None when there is no policy or the policy's routing side names no + compression, meaning the caller should route on whatever messages it already + has. The model call is untouched by this function either way: model-side + compression, if any, already ran as an ordinary pre-call guardrail before the + router was ever reached. + """ + if policy is None or policy.routing is None: + return None + + from litellm.proxy.common_utils.registry_read_through import ( + get_initialized_guardrail_with_read_through, + ) + + metadata_key: Final = get_metadata_variable_name_from_kwargs(request_kwargs) + metadata: Final = request_kwargs.get(metadata_key) + snapshot: Final = metadata.get(AUTO_ROUTER_ROUTING_MESSAGES_SNAPSHOT_KEY) if isinstance(metadata, dict) else None + original: Final = snapshot if isinstance(snapshot, list) else messages + if not original: + return None + + guardrail: Final = await get_initialized_guardrail_with_read_through(policy.routing) + if guardrail is None: + verbose_proxy_logger.warning( + "AutoRouter compression: guardrail '%s' not found; routing on uncompressed messages", policy.routing + ) + return None + + inputs: GenericGuardrailAPIInputs = {"structured_messages": [dict(m) for m in original]} + # A throwaway request_data: apply_guardrail writes its stats onto this dict, not + # the real request's metadata, so routing-side compression never double-counts + # against extract_compression_saved_tokens's model-savings accounting. + throwaway_request_data: Final[dict[str, object]] = { + "messages": original, + "model": request_kwargs.get("model"), + } + result: Final = await guardrail.apply_guardrail( + inputs=inputs, request_data=throwaway_request_data, input_type="request" + ) + compressed = result.get("structured_messages") + return compressed if isinstance(compressed, list) else original diff --git a/litellm/router.py b/litellm/router.py index 6c7611c6236..8149ec60ddc 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -13037,13 +13037,46 @@ class Router: ) return None + from litellm.proxy.guardrails.auto_router_compression import ( + messages_for_routing, + policy_from_litellm_params, + ) + + marker_params: Final = self._alias_marker_litellm_params(registered_model_name, selected_strategy.tags) + compression_policy: Final = policy_from_litellm_params(marker_params) if marker_params else None + # When both hops share the same compression, the model-side guardrail already + # ran in the proxy's ordinary pre-call hook and compressed `messages` in place + # (arm_pre_call armed it whether or not it is `default_on`); reuse that result + # for routing too instead of paying for a second compression call against the + # same content. + needs_independent_routing_compression: Final = compression_policy is not None and not ( + compression_policy.is_same and compression_policy.model is not None + ) + routing_messages: Final = ( + await messages_for_routing(policy=compression_policy, messages=messages, request_kwargs=request_kwargs) + if needs_independent_routing_compression + else None + ) + pre_routing_hook_response: Final = await selected_strategy.strategy.async_pre_routing_hook( model=registered_model_name, request_kwargs=request_kwargs, - messages=messages, + messages=routing_messages if routing_messages is not None else messages, input=input, specific_deployment=specific_deployment, ) + # The strategy only echoes back whatever `messages` it was handed, so a + # routing-only compression must not leak into the response: the model call + # and downstream deployment-context filtering both key off this field. + # Compared by value, not identity: PreRoutingHookResponse is a pydantic model, + # and pydantic reconstructs a validated list field rather than keeping the + # exact object passed in, even when nothing about it changed. + if ( + pre_routing_hook_response is not None + and routing_messages is not None + and pre_routing_hook_response.messages == routing_messages + ): + pre_routing_hook_response = pre_routing_hook_response.model_copy(update={"messages": messages}) self._record_routing_decision( request_kwargs=request_kwargs, routing_decision=(pre_routing_hook_response.routing_decision if pre_routing_hook_response else None), @@ -13100,9 +13133,16 @@ class Router: return pre_routing_hook_response - def _forwardable_alias_marker_params( + def _alias_marker_litellm_params( self, model: str, strategy_tags: tuple[str, ...] - ) -> tuple[tuple[str, object], ...]: + ) -> Mapping[str, object] | None: + """The auto-router marker deployment's own `litellm_params` for `model`, tag-scoped. + + Shared by `_forwardable_alias_marker_params` (forwarding api_base/api_key/... + gaps onto the routed deployment) and the auto-router compression policy lookup + (reading `auto_router_routing_compression`/`auto_router_model_compression`), so + both read the same marker row when an alias has more than one, tag-scoped marker. + """ marker_params: Final = tuple( litellm_params for idx in self.model_name_to_deployment_indices.get(model, ()) @@ -13112,7 +13152,12 @@ class Router: tag_matched: Final = tuple( params for params in marker_params if tuple(params.get("tags") or ()) == strategy_tags ) - selected: Final = tag_matched[0] if tag_matched else (marker_params[0] if marker_params else None) + return tag_matched[0] if tag_matched else (marker_params[0] if marker_params else None) + + def _forwardable_alias_marker_params( + self, model: str, strategy_tags: tuple[str, ...] + ) -> tuple[tuple[str, object], ...]: + selected: Final = self._alias_marker_litellm_params(model, strategy_tags) if selected is None: return () return tuple( diff --git a/litellm/types/router.py b/litellm/types/router.py index 7ebd50f1328..f5295d6569c 100644 --- a/litellm/types/router.py +++ b/litellm/types/router.py @@ -359,6 +359,10 @@ class GenericLiteLLMParams(CredentialLiteLLMParams, CustomPricingLiteLLMParams): auto_router_default_model: str | None = None auto_router_embedding_model: str | None = None auto_router_max_input_chars: int | None = None + # Compression policy for the two hops of a routed request. Both unset means the + # request's own compression guardrails apply to both, as they always have. + auto_router_routing_compression: str | None = None + auto_router_model_compression: str | None = None # complexity-router params complexity_router_config: dict | None = None diff --git a/litellm/types/utils.py b/litellm/types/utils.py index c1d90694f14..a1fda3d0524 100644 --- a/litellm/types/utils.py +++ b/litellm/types/utils.py @@ -3713,6 +3713,8 @@ all_litellm_params = ( "auto_router_default_model", "auto_router_embedding_model", "auto_router_max_input_chars", + "auto_router_routing_compression", + "auto_router_model_compression", "complexity_router_config", "complexity_router_default_model", "adaptive_router_config", diff --git a/tests/test_litellm/integrations/test_custom_guardrail.py b/tests/test_litellm/integrations/test_custom_guardrail.py index 6edc2c9bf77..1fb4299cb56 100644 --- a/tests/test_litellm/integrations/test_custom_guardrail.py +++ b/tests/test_litellm/integrations/test_custom_guardrail.py @@ -518,6 +518,50 @@ class TestCustomGuardrailShouldRunGuardrail: is True ) + def test_should_run_guardrail_suppressed_by_auto_router_compression(self): + """An auto router's own compression policy can suppress an otherwise-eligible + guardrail, even one that is default_on and explicitly requested.""" + from litellm.constants import AUTO_ROUTER_SUPPRESSED_COMPRESSION_GUARDRAILS_KEY + from litellm.types.guardrails import GuardrailEventHooks + + always_on = CustomGuardrail( + guardrail_name="headroom-default", + default_on=True, + event_hook=GuardrailEventHooks.pre_call, + ) + data = { + "model": "smart-router", + "metadata": { + AUTO_ROUTER_SUPPRESSED_COMPRESSION_GUARDRAILS_KEY: ["headroom-default"], + }, + } + + assert ( + always_on.should_run_guardrail(data=data, event_type=GuardrailEventHooks.pre_call) + is False + ) + + def test_should_run_guardrail_suppression_list_does_not_affect_other_names(self): + from litellm.constants import AUTO_ROUTER_SUPPRESSED_COMPRESSION_GUARDRAILS_KEY + from litellm.types.guardrails import GuardrailEventHooks + + always_on = CustomGuardrail( + guardrail_name="headroom-default", + default_on=True, + event_hook=GuardrailEventHooks.pre_call, + ) + data = { + "model": "smart-router", + "metadata": { + AUTO_ROUTER_SUPPRESSED_COMPRESSION_GUARDRAILS_KEY: ["some-other-guardrail"], + }, + } + + assert ( + always_on.should_run_guardrail(data=data, event_type=GuardrailEventHooks.pre_call) + is True + ) + class TestApplyGuardrailCheck: def test_apply_guardrail_check_only_on_direct_implementation(self): diff --git a/tests/test_litellm/proxy/guardrails/test_auto_router_compression.py b/tests/test_litellm/proxy/guardrails/test_auto_router_compression.py new file mode 100644 index 00000000000..b2e83e75768 --- /dev/null +++ b/tests/test_litellm/proxy/guardrails/test_auto_router_compression.py @@ -0,0 +1,285 @@ +""" +Unit tests for litellm.proxy.guardrails.auto_router_compression. + +Covers: +- policy_from_litellm_params: absent keys mean no policy; the "none" sentinel + normalizes to explicit no-compression within an active policy; is_same +- policy_for_model: finds the auto-router marker deployment for an alias +- arm_pre_call: no-op without a policy; suppresses active compression guardrails; + arms the model-side guardrail even when it isn't default_on; snapshots messages +- messages_for_routing: no-op without a policy or an unset routing side; compresses + via the named guardrail's apply_guardrail; never writes stats onto the caller's + own request_kwargs (regression for double-counted compression savings) +""" + +from typing import Any + +import pytest + +from litellm.integrations.custom_guardrail import CustomGuardrail +from litellm.proxy.guardrails.auto_router_compression import ( + AUTO_ROUTER_ROUTING_MESSAGES_SNAPSHOT_KEY, + AutoRouterCompressionPolicy, + arm_pre_call, + messages_for_routing, + policy_for_model, + policy_from_litellm_params, +) +from litellm.constants import AUTO_ROUTER_SUPPRESSED_COMPRESSION_GUARDRAILS_KEY +from litellm.types.utils import GenericGuardrailAPIInputs + + +class TestPolicyFromLitellmParams: + def test_neither_key_set_is_no_policy(self): + assert policy_from_litellm_params({}) is None + + def test_routing_only(self): + policy = policy_from_litellm_params({"auto_router_routing_compression": "headroom-a"}) + assert policy == AutoRouterCompressionPolicy(routing="headroom-a", model=None) + + def test_none_sentinel_normalizes_to_no_compression(self): + policy = policy_from_litellm_params( + {"auto_router_routing_compression": "headroom-a", "auto_router_model_compression": "none"} + ) + assert policy == AutoRouterCompressionPolicy(routing="headroom-a", model=None) + + def test_none_sentinel_is_case_insensitive(self): + policy = policy_from_litellm_params({"auto_router_routing_compression": "NONE"}) + assert policy == AutoRouterCompressionPolicy(routing=None, model=None) + + def test_is_same_true_for_matching_names(self): + policy = policy_from_litellm_params( + {"auto_router_routing_compression": "x", "auto_router_model_compression": "x"} + ) + assert policy.is_same is True + + def test_is_same_false_for_different_names(self): + policy = policy_from_litellm_params( + {"auto_router_routing_compression": "x", "auto_router_model_compression": "y"} + ) + assert policy.is_same is False + + def test_is_same_true_when_both_no_compression(self): + policy = policy_from_litellm_params( + {"auto_router_routing_compression": "none", "auto_router_model_compression": "none"} + ) + assert policy.is_same is True + + +class _FakeRouter: + """Minimal stand-in for litellm.Router.get_model_list, for policy_for_model.""" + + def __init__(self, deployments: list[dict[str, Any]]): + self._deployments = deployments + + def get_model_list(self, model_name, team_id=None): + return [d for d in self._deployments if d.get("model_name") == model_name] + + +class TestPolicyForModel: + def test_no_router_returns_none(self): + assert policy_for_model(llm_router=None, model_alias="smart-router", team_id=None) is None + + def test_no_marker_deployment_returns_none(self): + router = _FakeRouter( + [{"model_name": "smart-router", "litellm_params": {"model": "openai/gpt-4o-mini"}}] + ) + assert policy_for_model(llm_router=router, model_alias="smart-router", team_id=None) is None + + def test_marker_deployment_without_policy_returns_none(self): + router = _FakeRouter( + [{"model_name": "smart-router", "litellm_params": {"model": "auto_router/complexity_router"}}] + ) + assert policy_for_model(llm_router=router, model_alias="smart-router", team_id=None) is None + + def test_marker_deployment_with_policy_is_found(self): + router = _FakeRouter( + [ + { + "model_name": "smart-router", + "litellm_params": { + "model": "auto_router/complexity_router", + "auto_router_routing_compression": "headroom-a", + "auto_router_model_compression": "none", + }, + } + ] + ) + policy = policy_for_model(llm_router=router, model_alias="smart-router", team_id=None) + assert policy == AutoRouterCompressionPolicy(routing="headroom-a", model=None) + + +class _RecordingCompressionGuardrail(CustomGuardrail): + """A guardrail whose apply_guardrail marks every text message as compressed.""" + + def __init__(self, guardrail_name: str): + super().__init__(guardrail_name=guardrail_name) + self.request_data_seen: list[dict] = [] + + async def apply_guardrail( + self, inputs: GenericGuardrailAPIInputs, request_data: dict, input_type: str, logging_obj=None + ) -> GenericGuardrailAPIInputs: + self.request_data_seen.append(request_data) + structured_messages = inputs.get("structured_messages") or [] + compressed = [ + {**m, "content": f"[COMPRESSED] {m.get('content')}"} for m in structured_messages + ] + return {**inputs, "structured_messages": compressed} + + +@pytest.fixture +def registered_guardrail(): + import litellm + + guardrail = _RecordingCompressionGuardrail(guardrail_name="fake-compress") + litellm.logging_callback_manager.add_litellm_callback(guardrail) + yield guardrail + litellm.logging_callback_manager.remove_callback_from_all_lists(guardrail) + + +class TestArmPreCall: + @pytest.mark.asyncio + async def test_no_router_is_noop(self): + data = {"model": "smart-router", "messages": [{"role": "user", "content": "hi"}]} + result = await arm_pre_call(data=data, llm_router=None) + assert result == data + assert "metadata" not in result + + @pytest.mark.asyncio + async def test_no_policy_does_not_create_metadata_bucket(self): + router = _FakeRouter( + [{"model_name": "smart-router", "litellm_params": {"model": "openai/gpt-4o-mini"}}] + ) + data = {"model": "smart-router", "messages": [{"role": "user", "content": "hi"}]} + result = await arm_pre_call(data=data, llm_router=router) + assert "metadata" not in result + assert "litellm_metadata" not in result + + @pytest.mark.asyncio + async def test_policy_suppresses_active_compression_guardrails(self, monkeypatch): + from litellm.proxy.guardrails import guardrail_registry + + monkeypatch.setitem( + guardrail_registry.guardrail_class_registry, "fake-provider", _RecordingCompressionGuardrail + ) + monkeypatch.setattr( + "litellm.proxy.guardrails.auto_router_compression.COMPRESSION_GUARDRAIL_PROVIDERS", + frozenset({"fake-provider"}), + ) + import litellm + + always_on = _RecordingCompressionGuardrail(guardrail_name="always-on-compression") + litellm.logging_callback_manager.add_litellm_callback(always_on) + try: + router = _FakeRouter( + [ + { + "model_name": "smart-router", + "litellm_params": { + "model": "auto_router/complexity_router", + "auto_router_routing_compression": "headroom-a", + "auto_router_model_compression": "none", + }, + } + ] + ) + data = {"model": "smart-router", "messages": [{"role": "user", "content": "hi"}]} + result = await arm_pre_call(data=data, llm_router=router) + suppressed = result["metadata"][AUTO_ROUTER_SUPPRESSED_COMPRESSION_GUARDRAILS_KEY] + assert "always-on-compression" in suppressed + finally: + litellm.logging_callback_manager.remove_callback_from_all_lists(always_on) + + @pytest.mark.asyncio + async def test_model_side_guardrail_is_requested_even_when_not_default_on(self): + router = _FakeRouter( + [ + { + "model_name": "smart-router", + "litellm_params": { + "model": "auto_router/complexity_router", + "auto_router_routing_compression": "none", + "auto_router_model_compression": "headroom-b", + }, + } + ] + ) + data = {"model": "smart-router", "messages": [{"role": "user", "content": "hi"}]} + result = await arm_pre_call(data=data, llm_router=router) + assert result["metadata"]["guardrails"] == ["headroom-b"] + + @pytest.mark.asyncio + async def test_snapshots_original_messages(self): + router = _FakeRouter( + [ + { + "model_name": "smart-router", + "litellm_params": { + "model": "auto_router/complexity_router", + "auto_router_routing_compression": "headroom-a", + "auto_router_model_compression": "none", + }, + } + ] + ) + original_messages = [{"role": "user", "content": "hi"}] + data = {"model": "smart-router", "messages": original_messages} + result = await arm_pre_call(data=data, llm_router=router) + snapshot = result["metadata"][AUTO_ROUTER_ROUTING_MESSAGES_SNAPSHOT_KEY] + assert snapshot == original_messages + assert snapshot is not original_messages # a copy, not the live reference + + +class TestMessagesForRouting: + @pytest.mark.asyncio + async def test_no_policy_returns_none(self): + assert await messages_for_routing(policy=None, messages=[], request_kwargs={}) is None + + @pytest.mark.asyncio + async def test_routing_side_unset_returns_none(self): + policy = AutoRouterCompressionPolicy(routing=None, model="headroom-a") + assert await messages_for_routing(policy=policy, messages=[], request_kwargs={}) is None + + @pytest.mark.asyncio + async def test_unknown_guardrail_name_returns_none(self): + policy = AutoRouterCompressionPolicy(routing="does-not-exist", model=None) + messages = [{"role": "user", "content": "hi"}] + result = await messages_for_routing(policy=policy, messages=messages, request_kwargs={}) + assert result is None + + @pytest.mark.asyncio + async def test_compresses_via_the_named_guardrail(self, registered_guardrail): + policy = AutoRouterCompressionPolicy(routing="fake-compress", model=None) + messages = [{"role": "user", "content": "hello world"}] + result = await messages_for_routing(policy=policy, messages=messages, request_kwargs={}) + assert result == [{"role": "user", "content": "[COMPRESSED] hello world"}] + + @pytest.mark.asyncio + async def test_uses_the_snapshot_when_present(self, registered_guardrail): + policy = AutoRouterCompressionPolicy(routing="fake-compress", model=None) + snapshot = [{"role": "user", "content": "original"}] + request_kwargs = {"metadata": {AUTO_ROUTER_ROUTING_MESSAGES_SNAPSHOT_KEY: snapshot}} + # `messages` here stands in for whatever a model-side guardrail already + # rewrote `data["messages"]` to -- routing must ignore it and compress the + # pristine snapshot instead. + already_rewritten = [{"role": "user", "content": "rewritten by another guardrail"}] + result = await messages_for_routing( + policy=policy, messages=already_rewritten, request_kwargs=request_kwargs + ) + assert result == [{"role": "user", "content": "[COMPRESSED] original"}] + + @pytest.mark.asyncio + async def test_guardrail_receives_a_throwaway_request_data_not_the_real_request_kwargs( + self, registered_guardrail + ): + """Regression: a real compression guardrail writes its stats onto whatever + `request_data` dict it's given (`add_standard_logging_guardrail_information_to_ + request_data`). If that were the caller's own `request_kwargs`, routing-side + compression would double-count into extract_compression_saved_tokens, which + sums every guardrail_information entry on the real request's metadata.""" + policy = AutoRouterCompressionPolicy(routing="fake-compress", model=None) + messages = [{"role": "user", "content": "hi"}] + request_kwargs = {"metadata": {}} + await messages_for_routing(policy=policy, messages=messages, request_kwargs=request_kwargs) + assert registered_guardrail.request_data_seen[0] is not request_kwargs + assert request_kwargs == {"metadata": {}} diff --git a/tests/test_litellm/proxy/test_common_request_processing.py b/tests/test_litellm/proxy/test_common_request_processing.py index f7fe6ad9d39..c0809e53d2e 100644 --- a/tests/test_litellm/proxy/test_common_request_processing.py +++ b/tests/test_litellm/proxy/test_common_request_processing.py @@ -376,6 +376,60 @@ class TestProxyBaseLLMRequestProcessing: assert "litellm_logging_obj" not in persisted_body json.dumps(persisted_body) + @pytest.mark.asyncio + async def test_common_processing_pre_call_logic_arms_auto_router_compression_before_guardrails( + self, monkeypatch + ): + """arm_pre_call must run before pre_call_hook: an auto router's own compression + policy has to be in `data["metadata"]` (naming the model-side guardrail so it + runs even if it isn't default_on) by the time guardrails see the request.""" + processing_obj = ProxyBaseLLMRequestProcessing(data={}) + mock_request = MagicMock(spec=Request) + mock_request.headers = {} + + async def mock_add_litellm_data_to_request(*args, **kwargs): + return {"model": "smart-router", "messages": [{"role": "user", "content": "hi"}]} + + seen_metadata: dict = {} + + async def mock_pre_call_hook(user_api_key_dict, data, call_type): + seen_metadata.update(data.get("metadata") or {}) + return data + + mock_proxy_logging_obj = MagicMock(spec=ProxyLogging) + mock_proxy_logging_obj.pre_call_hook = AsyncMock(side_effect=mock_pre_call_hook) + monkeypatch.setattr( + litellm.proxy.common_request_processing, + "add_litellm_data_to_request", + mock_add_litellm_data_to_request, + ) + + fake_llm_router = MagicMock() + fake_llm_router.get_model_list.return_value = [ + { + "model_name": "smart-router", + "litellm_params": { + "model": "auto_router/complexity_router", + "auto_router_routing_compression": "none", + "auto_router_model_compression": "headroom-model", + }, + } + ] + mock_proxy_config = MagicMock(spec=ProxyConfig) + mock_proxy_config._get_hierarchical_router_settings = AsyncMock(return_value=None) + + await processing_obj.common_processing_pre_call_logic( + request=mock_request, + general_settings={}, + user_api_key_dict=ProxyUserAPIKeyAuth(api_key="sk-test"), + proxy_logging_obj=mock_proxy_logging_obj, + proxy_config=mock_proxy_config, + route_type="acompletion", + llm_router=fake_llm_router, + ) + + assert seen_metadata.get("guardrails") == ["headroom-model"] + def test_add_dd_apm_tags_for_litellm_call_id_uses_dd_tracing_helper(self, monkeypatch): mock_set_active_span_tag = MagicMock(return_value=True) import litellm.proxy.dd_span_tagger diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py index 5fc96bcfbb1..cdae3b131ae 100644 --- a/tests/test_litellm/test_router.py +++ b/tests/test_litellm/test_router.py @@ -18,6 +18,7 @@ import pytest import litellm from litellm import Router from litellm.exceptions import MidStreamFallbackError +from litellm.integrations.custom_guardrail import CustomGuardrail from litellm.integrations.custom_logger import CustomLogger from litellm.llms.bedrock.common_utils import BedrockError from litellm.llms.anthropic.experimental_pass_through.messages.agentic_streaming_iterator import ( @@ -9995,6 +9996,156 @@ class TestModelGroupAliasReachesPreRoutingStrategies: ) +class TestAutoRouterCompressionDecoupling: + """An auto router's `auto_router_routing_compression` / `auto_router_model_compression` + decouple what the routing decision sees from what the model call sees. The one + assertion that must hold under any mutation: the strategy can be routed on + compressed text while the caller's own `messages` list - the one that would reach + the model - is never touched.""" + + class _RecordingStrategy: + """Echoes back whatever `messages` it was handed, like every real strategy does.""" + + def __init__(self): + self.received_messages: list[dict] | None = None + + async def async_pre_routing_hook( + self, model, request_kwargs, messages=None, input=None, specific_deployment=False + ): + from litellm.types.router import PreRoutingHookResponse + + self.received_messages = messages + return PreRoutingHookResponse(model="gemini-flash", messages=messages) + + class _CompressingGuardrail(CustomGuardrail): + def __init__(self, guardrail_name: str): + super().__init__(guardrail_name=guardrail_name) + self.call_count = 0 + + async def apply_guardrail(self, inputs, request_data, input_type, logging_obj=None): + self.call_count += 1 + structured_messages = inputs.get("structured_messages") or [] + compressed = [ + {**m, "content": f"[COMPRESSED] {m.get('content')}"} for m in structured_messages + ] + return {**inputs, "structured_messages": compressed} + + @staticmethod + def _messages() -> list[dict[str, str]]: + return [{"role": "user", "content": "What is the capital of France?"}] + + def _router(self, marker_litellm_params: dict) -> tuple[litellm.Router, "_RecordingStrategy"]: + from litellm.types.router import TaggedPreRoutingStrategy + + tiers = dict.fromkeys(("SIMPLE", "MEDIUM", "COMPLEX", "REASONING"), "gemini-flash") + router = litellm.Router( + model_list=[ + { + "model_name": "smart-router", + "litellm_params": { + "model": "auto_router/complexity_router", + "complexity_router_config": {"tiers": tiers}, + "complexity_router_default_model": "gemini-flash", + **marker_litellm_params, + }, + }, + { + "model_name": "gemini-flash", + "litellm_params": {"model": "gemini/gemini-3.6-flash", "mock_response": "routed by the tier"}, + }, + ], + ) + for name in ("auto_routers", "complexity_routers", "adaptive_routers", "quality_routers"): + setattr(router, name, {}) + strategy = self._RecordingStrategy() + router.complexity_routers = {"smart-router": [TaggedPreRoutingStrategy(tags=(), strategy=strategy)]} + return router, strategy + + @pytest.fixture + def registered_guardrail(self): + guardrail = self._CompressingGuardrail(guardrail_name="fake-compress") + litellm.logging_callback_manager.add_litellm_callback(guardrail) + yield guardrail + litellm.logging_callback_manager.remove_callback_from_all_lists(guardrail) + + @pytest.mark.asyncio + async def test_routing_side_compression_never_reaches_the_caller_messages(self, registered_guardrail): + router, strategy = self._router( + { + "auto_router_routing_compression": "fake-compress", + "auto_router_model_compression": "none", + } + ) + original_messages = self._messages() + + response = await router.async_pre_routing_hook( + model="smart-router", request_kwargs={"metadata": {}}, messages=original_messages + ) + + assert strategy.received_messages == [ + {"role": "user", "content": "[COMPRESSED] What is the capital of France?"} + ] + assert response.messages == original_messages + + @pytest.mark.asyncio + async def test_model_side_compression_alone_leaves_routing_uncompressed(self, registered_guardrail): + router, strategy = self._router( + { + "auto_router_routing_compression": "none", + "auto_router_model_compression": "fake-compress", + } + ) + original_messages = self._messages() + + response = await router.async_pre_routing_hook( + model="smart-router", request_kwargs={"metadata": {}}, messages=original_messages + ) + + assert strategy.received_messages == original_messages + assert response.messages == original_messages + assert registered_guardrail.call_count == 0 + + @pytest.mark.asyncio + async def test_same_compression_on_both_hops_compresses_once(self, registered_guardrail): + """The same/different distinction exists so a shared choice does not pay for + compression twice: by the time the router runs, `messages` already reflects + whatever the ordinary pre-call guardrail pipeline did for the model call, so + the routing decision must reuse it rather than calling the guardrail again.""" + router, strategy = self._router( + { + "auto_router_routing_compression": "fake-compress", + "auto_router_model_compression": "fake-compress", + } + ) + # Stands in for what the proxy's ordinary pre-call guardrail pipeline would + # have already produced for the model call, since `auto_router_model_compression` + # names a guardrail: the router never triggers that pipeline itself. + already_compressed_messages = [ + {"role": "user", "content": "[COMPRESSED] What is the capital of France?"} + ] + + response = await router.async_pre_routing_hook( + model="smart-router", request_kwargs={"metadata": {}}, messages=already_compressed_messages + ) + + assert strategy.received_messages == already_compressed_messages + assert response.messages == already_compressed_messages + assert registered_guardrail.call_count == 0 + + @pytest.mark.asyncio + async def test_no_policy_is_fully_unaffected(self, registered_guardrail): + router, strategy = self._router({}) + original_messages = self._messages() + + response = await router.async_pre_routing_hook( + model="smart-router", request_kwargs={"metadata": {}}, messages=original_messages + ) + + assert strategy.received_messages is original_messages + assert response.messages == original_messages + assert registered_guardrail.call_count == 0 + + @pytest.mark.usefixtures("local_model_cost_map") class TestAzureBaseModelFallbackLogging: """When an azure deployment has no base_model but its model name is a known diff --git a/ui/litellm-dashboard/src/components/add_model/ComplexityRouterConfig.tsx b/ui/litellm-dashboard/src/components/add_model/ComplexityRouterConfig.tsx index 2a024ab7fdf..42115265034 100644 --- a/ui/litellm-dashboard/src/components/add_model/ComplexityRouterConfig.tsx +++ b/ui/litellm-dashboard/src/components/add_model/ComplexityRouterConfig.tsx @@ -48,6 +48,8 @@ import EscalationKeywords from "./EscalationKeywords"; import KeywordTierRules, { KeywordTierRule } from "./KeywordTierRules"; import SemanticKeywordMatching from "./SemanticKeywordMatching"; import { type DimensionWeights, type TierBoundaries, type TokenThresholds } from "./heuristic_scoring_knobs"; +import CompressionControls from "./CompressionControls"; +import { type AutoRouterCompressionState, DEFAULT_AUTO_ROUTER_COMPRESSION } from "./buildAutoRouterCompression"; export type { DimensionWeights, TierBoundaries, TokenThresholds }; export type { CustomTierSet, TierRow } from "./tier_rows"; @@ -490,6 +492,10 @@ interface ComplexityRouterConfigProps { onMatchThresholdChange?: (threshold: number) => void; escalationKeywords?: string[]; onEscalationKeywordsChange?: (keywords: string[]) => void; + // Optional: not part of complexity_router_config, since it applies to every + // pre-routing strategy, not just the complexity router. + autoRouterCompression?: AutoRouterCompressionState; + onAutoRouterCompressionChange?: (state: AutoRouterCompressionState) => void; showValidationErrors?: boolean; } @@ -611,6 +617,8 @@ const ComplexityRouterConfig: React.FC = ({ onMatchThresholdChange = () => {}, escalationKeywords = [], onEscalationKeywordsChange, + autoRouterCompression = DEFAULT_AUTO_ROUTER_COMPRESSION, + onAutoRouterCompressionChange, showValidationErrors = false, }) => { const customTierSet = value.custom_tier_set; @@ -875,6 +883,32 @@ const ComplexityRouterConfig: React.FC = ({ }, ] : []), + ...(onAutoRouterCompressionChange + ? [ + { + key: "compression", + label: Advanced: Compression, + children: ( + + onAutoRouterCompressionChange({ + ...autoRouterCompression, + routing, + sameAsRouting: routing === undefined ? true : autoRouterCompression.sameAsRouting, + }) + } + sameAsRouting={autoRouterCompression.sameAsRouting} + onSameAsRoutingChange={(sameAsRouting) => + onAutoRouterCompressionChange({ ...autoRouterCompression, sameAsRouting }) + } + model={autoRouterCompression.model} + onModelChange={(model) => onAutoRouterCompressionChange({ ...autoRouterCompression, model })} + /> + ), + }, + ] + : []), ...(onKeywordTierRulesChange || onSemanticMatchingEnabledChange ? [ { diff --git a/ui/litellm-dashboard/src/components/add_model/CompressionControls.tsx b/ui/litellm-dashboard/src/components/add_model/CompressionControls.tsx new file mode 100644 index 00000000000..a0a240f76b0 --- /dev/null +++ b/ui/litellm-dashboard/src/components/add_model/CompressionControls.tsx @@ -0,0 +1,93 @@ +import { SimpleTooltip } from "@/components/ui/tooltip"; +import { SearchSelect, SearchSelectOption } from "@/components/shared/SearchSelect"; +import { Label } from "@/components/ui/label"; +import { RadioGroup, RadioGroupItem } from "@/components/ui/radio-group"; +import { Info } from "lucide-react"; +import React from "react"; +import { useGuardrails } from "@/app/(dashboard)/hooks/guardrails/useGuardrails"; +import { COMPRESSION_GUARDRAIL_PROVIDER } from "@/app/(dashboard)/cost-optimization/_components/helpers"; +import { NO_COMPRESSION } from "./buildAutoRouterCompression"; + +interface CompressionControlsProps { + routing: string | undefined; + onRoutingChange: (value: string | undefined) => void; + sameAsRouting: boolean; + onSameAsRoutingChange: (same: boolean) => void; + model: string | undefined; + onModelChange: (value: string | undefined) => void; +} + +const NONE_OPTION: SearchSelectOption = { label: "None (no compression)", value: NO_COMPRESSION }; + +const CompressionControls: React.FC = ({ + routing, + onRoutingChange, + sameAsRouting, + onSameAsRoutingChange, + model, + onModelChange, +}) => { + const { data } = useGuardrails(); + const compressionOptions: SearchSelectOption[] = (data?.guardrails ?? []) + .filter((g) => (g.litellm_params?.guardrail ?? "").toString().toLowerCase() === COMPRESSION_GUARDRAIL_PROVIDER) + .map((g) => ({ label: g.guardrail_name, value: g.guardrail_name })); + const options: SearchSelectOption[] = [NONE_OPTION, ...compressionOptions]; + + return ( +
+
+
+ Routing decision + + + +
+ onRoutingChange(value === "" ? undefined : value)} + placeholder="Inherit from the request's own compression guardrails" + emptyText="No compression guardrails found" + aria-label="Routing decision compression" + /> +
+ + {routing !== undefined && ( +
+ Model call + onSameAsRoutingChange(value === "same")} + className="w-full" + > +
+ + +
+ + + {!sameAsRouting && ( +
+ onModelChange(value === "" ? undefined : value)} + placeholder="None (no compression)" + emptyText="No compression guardrails found" + aria-label="Model call compression" + /> +
+ )} +
+ )} +
+ ); +}; + +export default CompressionControls; diff --git a/ui/litellm-dashboard/src/components/add_model/add_auto_router_tab.test.tsx b/ui/litellm-dashboard/src/components/add_model/add_auto_router_tab.test.tsx index d2f6b10c3a6..5605a993ded 100644 --- a/ui/litellm-dashboard/src/components/add_model/add_auto_router_tab.test.tsx +++ b/ui/litellm-dashboard/src/components/add_model/add_auto_router_tab.test.tsx @@ -1,4 +1,12 @@ -import { renderWithProviders, screen, waitFor, within, fireEvent, testQueryClient } from "../../../tests/test-utils"; +import { + renderWithProviders, + screen, + waitFor, + within, + fireEvent, + testQueryClient, + chooseSelectOption, +} from "../../../tests/test-utils"; import userEvent from "@testing-library/user-event"; import { vi } from "vitest"; import AddAutoRouterTab from "./add_auto_router_tab"; @@ -522,6 +530,71 @@ describe("AddAutoRouterTab", () => { ); }); + describe("prompt compression", () => { + it("leaves both compression keys out of the create payload when the section is untouched", async () => { + const user = userEvent.setup(); + vi.mocked(getMissingTiersError).mockReturnValue(null); + + renderWithProviders(); + + await user.type(screen.getByPlaceholderText(/smart_router/i), "no-compression-router"); + await user.click(screen.getByRole("button", { name: /add auto router/i })); + + await waitFor(() => expect(handleAddAutoRouterSubmit).toHaveBeenCalled()); + const submitted = vi.mocked(handleAddAutoRouterSubmit).mock.calls.at(-1)?.[0]; + expect(submitted).not.toHaveProperty("auto_router_routing_compression"); + expect(submitted).not.toHaveProperty("auto_router_model_compression"); + }); + + it("mirrors an explicit no-compression routing choice onto the model call by default", async () => { + const user = userEvent.setup(); + vi.mocked(getMissingTiersError).mockReturnValue(null); + + renderWithProviders(); + + await user.type(screen.getByPlaceholderText(/smart_router/i), "no-compression-explicit-router"); + expandDetailedConfiguration(); + await user.click(screen.getByText("Advanced: Compression")); + await chooseSelectOption( + user, + screen.getByRole("combobox", { name: "Routing decision compression" }), + "None (no compression)", + ); + + await user.click(screen.getByRole("button", { name: /add auto router/i })); + + await waitFor(() => expect(handleAddAutoRouterSubmit).toHaveBeenCalled()); + const submitted = vi.mocked(handleAddAutoRouterSubmit).mock.calls.at(-1)?.[0]; + expect(submitted?.auto_router_routing_compression).toBe("none"); + expect(submitted?.auto_router_model_compression).toBe("none"); + }); + + it("defaults the model call to none when different is chosen but nothing is picked there", async () => { + const user = userEvent.setup(); + vi.mocked(getMissingTiersError).mockReturnValue(null); + + renderWithProviders(); + + await user.type(screen.getByPlaceholderText(/smart_router/i), "different-compression-router"); + expandDetailedConfiguration(); + await user.click(screen.getByText("Advanced: Compression")); + await chooseSelectOption( + user, + screen.getByRole("combobox", { name: "Routing decision compression" }), + "None (no compression)", + ); + await user.click(screen.getByText("Use a different compression")); + expect(screen.getByRole("combobox", { name: "Model call compression" })).toBeInTheDocument(); + + await user.click(screen.getByRole("button", { name: /add auto router/i })); + + await waitFor(() => expect(handleAddAutoRouterSubmit).toHaveBeenCalled()); + const submitted = vi.mocked(handleAddAutoRouterSubmit).mock.calls.at(-1)?.[0]; + expect(submitted?.auto_router_routing_compression).toBe("none"); + expect(submitted?.auto_router_model_compression).toBe("none"); + }); + }); + // The scalar floor is the one scorer knob with no group dict behind it, so its wiring into the create // payload is only proven end to end. 0 is the case a truthy check would silently drop. it("carries a reasoning override floor of 0 through to the create payload", async () => { diff --git a/ui/litellm-dashboard/src/components/add_model/add_auto_router_tab.tsx b/ui/litellm-dashboard/src/components/add_model/add_auto_router_tab.tsx index a548c2c6533..decacac6501 100644 --- a/ui/litellm-dashboard/src/components/add_model/add_auto_router_tab.tsx +++ b/ui/litellm-dashboard/src/components/add_model/add_auto_router_tab.tsx @@ -32,6 +32,11 @@ import ComplexityRouterConfig, { } from "./ComplexityRouterConfig"; import { KeywordTierRule } from "./KeywordTierRules"; import { DEFAULT_ESCALATION_KEYWORDS } from "./EscalationKeywords"; +import { + type AutoRouterCompressionState, + buildAutoRouterCompressionParams, + DEFAULT_AUTO_ROUTER_COMPRESSION, +} from "./buildAutoRouterCompression"; import { DEFAULT_MATCH_THRESHOLD } from "./SemanticKeywordMatching"; import { BuildComplexityRouterConfigParams, @@ -194,6 +199,9 @@ const AddAutoRouterTab: React.FC = ({ const [embeddingModel, setEmbeddingModel] = useState(undefined); const [matchThreshold, setMatchThreshold] = useState(DEFAULT_MATCH_THRESHOLD); const [escalationKeywords, setEscalationKeywords] = useState(DEFAULT_ESCALATION_KEYWORDS); + const [autoRouterCompression, setAutoRouterCompression] = useState( + DEFAULT_AUTO_ROUTER_COMPRESSION, + ); const [showValidationErrors, setShowValidationErrors] = useState(false); const [editingTiers, setEditingTiers] = useState(false); const [isSubmitting, setIsSubmitting] = useState(false); @@ -461,6 +469,7 @@ const AddAutoRouterTab: React.FC = ({ model_type: "complexity_router", complexity_router_config: complexityRouterConfigPayload, model_access_group: form.getValues("model_access_group"), + ...buildAutoRouterCompressionParams(autoRouterCompression), }; await handleAddAutoRouterSubmit(submitValues, accessToken, () => form.reset(EMPTY_FORM_VALUES), handleOk); @@ -666,6 +675,8 @@ const AddAutoRouterTab: React.FC = ({ onMatchThresholdChange={setMatchThreshold} escalationKeywords={escalationKeywords} onEscalationKeywordsChange={setEscalationKeywords} + autoRouterCompression={autoRouterCompression} + onAutoRouterCompressionChange={setAutoRouterCompression} showValidationErrors={showValidationErrors} /> diff --git a/ui/litellm-dashboard/src/components/add_model/buildAutoRouterCompression.test.ts b/ui/litellm-dashboard/src/components/add_model/buildAutoRouterCompression.test.ts new file mode 100644 index 00000000000..b917fcedaa2 --- /dev/null +++ b/ui/litellm-dashboard/src/components/add_model/buildAutoRouterCompression.test.ts @@ -0,0 +1,93 @@ +import { + buildAutoRouterCompressionParams, + DEFAULT_AUTO_ROUTER_COMPRESSION, + hydrateAutoRouterCompression, + NO_COMPRESSION, +} from "./buildAutoRouterCompression"; + +describe("buildAutoRouterCompressionParams", () => { + it("omits both keys when routing was never configured", () => { + expect(buildAutoRouterCompressionParams(DEFAULT_AUTO_ROUTER_COMPRESSION)).toEqual({}); + }); + + it("mirrors routing onto model when same-as-routing is chosen", () => { + const params = buildAutoRouterCompressionParams({ + routing: "headroom-a", + sameAsRouting: true, + model: undefined, + }); + expect(params).toEqual({ + auto_router_routing_compression: "headroom-a", + auto_router_model_compression: "headroom-a", + }); + }); + + it("uses the explicit model choice when different is chosen", () => { + const params = buildAutoRouterCompressionParams({ + routing: "headroom-a", + sameAsRouting: false, + model: "headroom-b", + }); + expect(params).toEqual({ + auto_router_routing_compression: "headroom-a", + auto_router_model_compression: "headroom-b", + }); + }); + + it("defaults the model side to none when different is chosen but nothing is picked", () => { + const params = buildAutoRouterCompressionParams({ + routing: "headroom-a", + sameAsRouting: false, + model: undefined, + }); + expect(params).toEqual({ + auto_router_routing_compression: "headroom-a", + auto_router_model_compression: NO_COMPRESSION, + }); + }); + + it("sends the none sentinel when routing itself is explicitly turned off", () => { + const params = buildAutoRouterCompressionParams({ + routing: NO_COMPRESSION, + sameAsRouting: true, + model: undefined, + }); + expect(params).toEqual({ + auto_router_routing_compression: NO_COMPRESSION, + auto_router_model_compression: NO_COMPRESSION, + }); + }); +}); + +describe("hydrateAutoRouterCompression", () => { + it("returns the default state when neither key is set", () => { + expect(hydrateAutoRouterCompression({})).toEqual(DEFAULT_AUTO_ROUTER_COMPRESSION); + }); + + it("is same-as-routing when the model value matches routing", () => { + const state = hydrateAutoRouterCompression({ + auto_router_routing_compression: "headroom-a", + auto_router_model_compression: "headroom-a", + }); + expect(state).toEqual({ routing: "headroom-a", sameAsRouting: true, model: undefined }); + }); + + it("is different when the model value diverges from routing", () => { + const state = hydrateAutoRouterCompression({ + auto_router_routing_compression: "headroom-a", + auto_router_model_compression: "headroom-b", + }); + expect(state).toEqual({ routing: "headroom-a", sameAsRouting: false, model: "headroom-b" }); + }); + + it("treats a missing model key as same-as-routing", () => { + const state = hydrateAutoRouterCompression({ auto_router_routing_compression: "headroom-a" }); + expect(state).toEqual({ routing: "headroom-a", sameAsRouting: true, model: undefined }); + }); + + it("round-trips through buildAutoRouterCompressionParams", () => { + const original = { auto_router_routing_compression: "headroom-a", auto_router_model_compression: "none" }; + const rebuilt = buildAutoRouterCompressionParams(hydrateAutoRouterCompression(original)); + expect(rebuilt).toEqual(original); + }); +}); diff --git a/ui/litellm-dashboard/src/components/add_model/buildAutoRouterCompression.ts b/ui/litellm-dashboard/src/components/add_model/buildAutoRouterCompression.ts new file mode 100644 index 00000000000..49180d5b4ce --- /dev/null +++ b/ui/litellm-dashboard/src/components/add_model/buildAutoRouterCompression.ts @@ -0,0 +1,52 @@ +/** + * Maps the auto router's compression form state to the two flat litellm_params keys + * the backend reads (litellm.proxy.guardrails.auto_router_compression), and back. + * + * `routing` being undefined means the section was never touched: both keys are + * omitted from the payload, and the request's own compression guardrails apply to + * both hops unchanged. Once `routing` has a value (a guardrail name, or the "none" + * sentinel for explicit no-compression), the auto router is authoritative and the + * model side always gets a concrete value too, mirroring `routing` when same-as + * is chosen and defaulting to "none" otherwise. + */ + +export const NO_COMPRESSION = "none"; + +export interface AutoRouterCompressionState { + routing: string | undefined; + sameAsRouting: boolean; + model: string | undefined; +} + +export interface AutoRouterCompressionLitellmParams { + auto_router_routing_compression?: string; + auto_router_model_compression?: string; +} + +export const DEFAULT_AUTO_ROUTER_COMPRESSION: AutoRouterCompressionState = { + routing: undefined, + sameAsRouting: true, + model: undefined, +}; + +export const buildAutoRouterCompressionParams = ( + state: AutoRouterCompressionState, +): AutoRouterCompressionLitellmParams => { + if (state.routing === undefined) return {}; + return { + auto_router_routing_compression: state.routing, + auto_router_model_compression: state.sameAsRouting ? state.routing : (state.model ?? NO_COMPRESSION), + }; +}; + +export const hydrateAutoRouterCompression = (litellmParams: { + auto_router_routing_compression?: string | null; + auto_router_model_compression?: string | null; +}): AutoRouterCompressionState => { + const routing = litellmParams.auto_router_routing_compression ?? undefined; + if (routing === undefined) return DEFAULT_AUTO_ROUTER_COMPRESSION; + + const model = litellmParams.auto_router_model_compression ?? undefined; + const sameAsRouting = model === undefined || model === routing; + return { routing, sameAsRouting, model: sameAsRouting ? undefined : model }; +}; diff --git a/ui/litellm-dashboard/src/components/add_model/handle_add_auto_router_submit.tsx b/ui/litellm-dashboard/src/components/add_model/handle_add_auto_router_submit.tsx index 9385836ce1a..59d9ecf205e 100644 --- a/ui/litellm-dashboard/src/components/add_model/handle_add_auto_router_submit.tsx +++ b/ui/litellm-dashboard/src/components/add_model/handle_add_auto_router_submit.tsx @@ -1,8 +1,9 @@ import { modelCreateCall } from "../networking"; import { toast } from "@/lib/toast"; import type { ComplexityRouterConfigPayload } from "./build_complexity_router_config"; +import type { AutoRouterCompressionLitellmParams } from "./buildAutoRouterCompression"; -export interface AddAutoRouterValues { +export interface AddAutoRouterValues extends AutoRouterCompressionLitellmParams { auto_router_name: string; auto_router_default_model: string | undefined; model_type: "complexity_router"; @@ -24,6 +25,8 @@ export const handleAddAutoRouterSubmit = async ( model: "auto_router/complexity_router", complexity_router_config: values.complexity_router_config, complexity_router_default_model: values.auto_router_default_model, + auto_router_routing_compression: values.auto_router_routing_compression, + auto_router_model_compression: values.auto_router_model_compression, }, model_info: { ...(values.team_id ? { team_id: values.team_id } : {}), diff --git a/ui/litellm-dashboard/src/components/edit_auto_router/edit_auto_router_modal.test.tsx b/ui/litellm-dashboard/src/components/edit_auto_router/edit_auto_router_modal.test.tsx index 970bcaa545f..b93db8d963e 100644 --- a/ui/litellm-dashboard/src/components/edit_auto_router/edit_auto_router_modal.test.tsx +++ b/ui/litellm-dashboard/src/components/edit_auto_router/edit_auto_router_modal.test.tsx @@ -1029,3 +1029,84 @@ describe("EditAutoRouterModal with a stored custom tier set", () => { expect(savedConfig().tier_model_configs).toEqual(CUSTOM_STORED.tier_model_configs); }); }); + +describe("EditAutoRouterModal prompt compression", () => { + beforeEach(() => { + modelPatchUpdateCall.mockClear(); + }); + + const savedLitellmParams = () => { + const [, payload] = modelPatchUpdateCall.mock.calls.at(-1) ?? []; + return payload?.litellm_params; + }; + + const renderWithStoredCompression = ( + compression?: { auto_router_routing_compression?: string; auto_router_model_compression?: string }, + ) => + renderWithProviders( + , + ); + + it("leaves both compression keys out of an untouched save when none were stored", async () => { + const user = userEvent.setup(); + renderWithStoredCompression(); + + await user.click(await screen.findByRole("button", { name: /save changes/i })); + + await waitFor(() => expect(modelPatchUpdateCall).toHaveBeenCalled()); + expect(savedLitellmParams()).not.toHaveProperty("auto_router_routing_compression"); + expect(savedLitellmParams()).not.toHaveProperty("auto_router_model_compression"); + }); + + it("preserves a stored same-as-routing compression through an untouched open-and-save", async () => { + const user = userEvent.setup(); + renderWithStoredCompression({ + auto_router_routing_compression: "headroom-a", + auto_router_model_compression: "headroom-a", + }); + + await user.click(await screen.findByRole("button", { name: /save changes/i })); + + await waitFor(() => expect(modelPatchUpdateCall).toHaveBeenCalled()); + expect(savedLitellmParams()?.auto_router_routing_compression).toBe("headroom-a"); + expect(savedLitellmParams()?.auto_router_model_compression).toBe("headroom-a"); + }); + + it("shows a stored different-compression choice as Use a different compression, not Same", async () => { + const user = userEvent.setup(); + renderWithStoredCompression({ + auto_router_routing_compression: "headroom-a", + auto_router_model_compression: "none", + }); + + await user.click(await screen.findByText("Advanced: Compression")); + + expect(await screen.findByRole("combobox", { name: "Routing decision compression" })).toHaveValue("headroom-a"); + expect(screen.getByRole("radio", { name: "Use a different compression" })).toBeChecked(); + expect(screen.getByRole("combobox", { name: "Model call compression" })).toHaveValue("None (no compression)"); + }); + + it("preserves a stored different-compression choice through an untouched open-and-save", async () => { + const user = userEvent.setup(); + renderWithStoredCompression({ + auto_router_routing_compression: "headroom-a", + auto_router_model_compression: "none", + }); + + await user.click(await screen.findByRole("button", { name: /save changes/i })); + + await waitFor(() => expect(modelPatchUpdateCall).toHaveBeenCalled()); + expect(savedLitellmParams()?.auto_router_routing_compression).toBe("headroom-a"); + expect(savedLitellmParams()?.auto_router_model_compression).toBe("none"); + }); +}); diff --git a/ui/litellm-dashboard/src/components/edit_auto_router/edit_auto_router_modal.tsx b/ui/litellm-dashboard/src/components/edit_auto_router/edit_auto_router_modal.tsx index ea1e5cba6a3..a4852f9e784 100644 --- a/ui/litellm-dashboard/src/components/edit_auto_router/edit_auto_router_modal.tsx +++ b/ui/litellm-dashboard/src/components/edit_auto_router/edit_auto_router_modal.tsx @@ -41,6 +41,12 @@ import { } from "../add_model/build_complexity_router_config"; import { KeywordTierRule } from "../add_model/KeywordTierRules"; import { DEFAULT_MATCH_THRESHOLD } from "../add_model/SemanticKeywordMatching"; +import { + type AutoRouterCompressionState, + buildAutoRouterCompressionParams, + DEFAULT_AUTO_ROUTER_COMPRESSION, + hydrateAutoRouterCompression, +} from "../add_model/buildAutoRouterCompression"; import { hydrateKeywordTierRules } from "../add_model/complexity_router_keywords"; import { hydrateDimensionWeights, @@ -424,6 +430,9 @@ const EditAutoRouterModal: React.FC = ({ const [semanticMatchingEnabled, setSemanticMatchingEnabled] = useState(false); const [embeddingModel, setEmbeddingModel] = useState(undefined); const [matchThreshold, setMatchThreshold] = useState(DEFAULT_MATCH_THRESHOLD); + const [autoRouterCompression, setAutoRouterCompression] = useState( + DEFAULT_AUTO_ROUTER_COMPRESSION, + ); const [complexityRouterConfig, setComplexityRouterConfig] = useState({ tiers: { SIMPLE: [], MEDIUM: [], COMPLEX: [], REASONING: [] }, classifier_type: "heuristic", @@ -516,6 +525,12 @@ const EditAutoRouterModal: React.FC = ({ setMatchThreshold( typeof parsedConfig.match_threshold === "number" ? parsedConfig.match_threshold : DEFAULT_MATCH_THRESHOLD, ); + setAutoRouterCompression( + hydrateAutoRouterCompression({ + auto_router_routing_compression: modelData.litellm_params?.auto_router_routing_compression, + auto_router_model_compression: modelData.litellm_params?.auto_router_model_compression, + }), + ); form.reset({ ...EMPTY_FORM_VALUES, @@ -628,6 +643,7 @@ const EditAutoRouterModal: React.FC = ({ ...modelData.litellm_params, complexity_router_config: updatedConfig, complexity_router_default_model: defaultModel, + ...buildAutoRouterCompressionParams(autoRouterCompression), }; const updatedModelInfo = { ...modelData.model_info, @@ -749,6 +765,8 @@ const EditAutoRouterModal: React.FC = ({ onMatchThresholdChange={setMatchThreshold} escalationKeywords={escalationKeywords} onEscalationKeywordsChange={setEscalationKeywords} + autoRouterCompression={autoRouterCompression} + onAutoRouterCompressionChange={setAutoRouterCompression} /> ) : ( diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index 5f3cca49644..8f6a0700517 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -29222,6 +29222,10 @@ export interface components { auto_router_embedding_model?: string | null; /** Auto Router Max Input Chars */ auto_router_max_input_chars?: number | null; + /** Auto Router Model Compression */ + auto_router_model_compression?: string | null; + /** Auto Router Routing Compression */ + auto_router_routing_compression?: string | null; /** Aws Access Key Id */ aws_access_key_id?: string | null; /** Aws Batch Role Arn */ @@ -39275,6 +39279,10 @@ export interface components { auto_router_embedding_model?: string | null; /** Auto Router Max Input Chars */ auto_router_max_input_chars?: number | null; + /** Auto Router Model Compression */ + auto_router_model_compression?: string | null; + /** Auto Router Routing Compression */ + auto_router_routing_compression?: string | null; /** Aws Access Key Id */ aws_access_key_id?: string | null; /** Aws Batch Role Arn */