Merge remote-tracking branch 'origin/main' into litellm_semantic_cache_gate_hook

.circleci/config.yml

@@ -1291,7 +1291,7 @@ jobs:
                 name: Install MCP SDK1 peer
                 command: |
                   uv venv --python 3.12 .venv-mcp-peer
-                  uv pip install --python .venv-mcp-peer 'mcp==1.28.1' 'langchain-mcp-adapters==0.2.1'
+                  uv pip install --no-config --exclude-newer "3 days" --python .venv-mcp-peer 'mcp==1.28.1' 'langchain-mcp-adapters==0.2.1'
                   echo "export MCP_TEST_PEER_PYTHON=$PWD/.venv-mcp-peer/bin/python" >> "$BASH_ENV"
       - when:
           condition:

.github/workflows/test-unit.yml

@@ -216,10 +216,12 @@ jobs:
 
           - shard: llms-providers
             test-path: |-
+              tests/unit/llms/test_api_base_trailing_slash_routes.py
               tests/unit/llms/test_cache_control_and_reasoning.py
               tests/unit/llms/test_custom_llm.py
               tests/unit/llms/test_file_content_block.py
               tests/unit/llms/test_file_search_responses.py
+              tests/unit/llms/test_get_complete_url_trailing_slash.py
               tests/unit/llms/test_lifecycle_fix.py
               tests/unit/llms/test_oss_decision.py
               tests/unit/llms/test_polling_url_origin_match.py

litellm/constants.py

@@ -611,6 +611,7 @@ NON_LLM_CONNECTION_TIMEOUT: Final = get_env_int(
 )  # timeout for adjacent services (e.g. jwt auth)
 MAX_EXCEPTION_MESSAGE_LENGTH: Final = get_env_int("MAX_EXCEPTION_MESSAGE_LENGTH", 2000)
 MAX_STRING_LENGTH_PROMPT_IN_DB: Final = get_env_int("MAX_STRING_LENGTH_PROMPT_IN_DB", 2048)
+MAX_STRING_LENGTH_COMPACTION_SUMMARY_IN_DB: Final = 64 * 1024
 BEDROCK_MAX_POLICY_SIZE: Final = get_env_int("BEDROCK_MAX_POLICY_SIZE", 75)
 # One entry per distinct AWS credential-argument set. Per-user cost attribution passes the attributed
 # identity as aws_session_name, so this bounds how many attributed identities keep a cached STS session.

litellm/cost_calculator.py

@@ -1673,8 +1673,10 @@ def completion_cost(
                     verbose_logger.debug("selected model name for cost calculation: %s", model)
 
                 if completion_response is not None and (
-                    isinstance(completion_response, BaseModel) or isinstance(completion_response, dict)
-                ):  # tts returns a custom class
+                    isinstance(completion_response, (BaseModel, dict))
+                    or isinstance(completion_response, HttpxBinaryResponseContent)
+                    and completion_response.usage is not None
+                ):
                     if isinstance(completion_response, dict):
                         usage_obj: dict | Usage | None = completion_response.get("usage", {})
                     else:

litellm/endpoints/speech/speech_to_completion_bridge/handler.py

@@ -122,6 +122,7 @@ class SpeechToCompletionBridgeHandler:
         if isinstance(result, ModelResponse):
             return self.transformation_handler.transform_response(
                 model_response=result,
+                model=model,
                 response_format=requested_response_format if isinstance(requested_response_format, str) else None,
             )
         else:

litellm/endpoints/speech/speech_to_completion_bridge/transformation.py

@@ -5,6 +5,7 @@ from typing import TYPE_CHECKING, Final, cast
 from typing_extensions import NotRequired, ReadOnly, TypedDict
 
 from litellm.constants import OPENAI_CHAT_COMPLETION_PARAMS
+from litellm.litellm_core_utils.core_helpers import drops_unsupported_params
 from litellm.litellm_core_utils.hidden_params import get_hidden_params, served_from_cache, set_hidden_param
 
 if TYPE_CHECKING:
@@ -38,6 +39,11 @@ class ChatAudioParam(TypedDict):
     format: ReadOnly[NotRequired[str]]
 
 
+class ChatSpeechConfigAudioParam(TypedDict):
+    speech_config: ReadOnly[Mapping[str, object]]
+    format: ReadOnly[str]
+
+
 class SpeechToCompletionBridgeTransformationHandler:
     def _validate_response_format(
         self, model: str, custom_llm_provider: str, optional_params: Mapping[str, object]
@@ -49,6 +55,20 @@ class SpeechToCompletionBridgeTransformationHandler:
             return
         from litellm.exceptions import BadRequestError
 
+        if custom_llm_provider.startswith("vertex_ai"):
+            from litellm.llms.vertex_ai.text_to_speech.transformation import VertexAITextToSpeechConfig
+
+            vertex_supported: Final = ", ".join(
+                sorted({*GEMINI_TTS_SUPPORTED_RESPONSE_FORMATS, *VertexAITextToSpeechConfig.GEMINI_FORMAT_MAPPINGS})
+            )
+            raise BadRequestError(
+                message=(
+                    f"Vertex AI Gemini TTS does not support response_format='{response_format}'."
+                    f" Supported response formats: {vertex_supported}."
+                ),
+                model=model,
+                llm_provider=custom_llm_provider,
+            )
         supported: Final = ", ".join(sorted(GEMINI_TTS_SUPPORTED_RESPONSE_FORMATS))
         raise BadRequestError(
             message=(
@@ -75,8 +95,27 @@ class SpeechToCompletionBridgeTransformationHandler:
         return response_format if isinstance(response_format, str) else None
 
     def _chat_audio_param(
-        self, model: str, voice: str | Mapping[str, object] | None, optional_params: Mapping[str, object]
-    ) -> ChatAudioParam | None:
+        self,
+        model: str,
+        voice: str | Mapping[str, object] | None,
+        optional_params: Mapping[str, object],
+        drop_params: bool,
+        custom_llm_provider: str,
+    ) -> ChatAudioParam | ChatSpeechConfigAudioParam | None:
+        import litellm
+
+        if isinstance(voice, Mapping) and self._is_gemini_tts_model(litellm.model_alias_map.get(model, model)):
+            from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
+                normalize_gemini_speech_config,
+                reject_unknown_gemini_tts_voice_keys,
+            )
+
+            reject_unknown_gemini_tts_voice_keys(model, voice, drop_params, custom_llm_provider)
+            speech_audio: Final[ChatSpeechConfigAudioParam] = {
+                "speech_config": normalize_gemini_speech_config(voice, drops_unsupported_params(drop_params)),
+                "format": GEMINI_TTS_CHAT_AUDIO_FORMAT,
+            }
+            return speech_audio
         if not isinstance(voice, str):
             return None
         audio_format: Final = self._chat_audio_format(model, optional_params)
@@ -90,9 +129,9 @@ class SpeechToCompletionBridgeTransformationHandler:
         self,
         model: str,
         input: str,
-        voice: str | dict | None,
-        optional_params: dict,
-        litellm_params: dict,
+        voice: str | Mapping[str, object] | None,
+        optional_params: Mapping[str, object],
+        litellm_params: Mapping[str, object],
         headers: dict,
         litellm_logging_obj: "LiteLLMLoggingObj",
         custom_llm_provider: str,
@@ -104,7 +143,13 @@ class SpeechToCompletionBridgeTransformationHandler:
             "messages": [user_message],
             "modalities": ["audio"],
             **self._chat_completion_params(optional_params),
-            "audio": self._chat_audio_param(model, voice, optional_params),
+            "audio": self._chat_audio_param(
+                model,
+                voice,
+                optional_params,
+                drops_unsupported_params(litellm_params.get("drop_params")),
+                custom_llm_provider,
+            ),
             **litellm_params,
             "headers": headers,
             "litellm_logging_obj": litellm_logging_obj,
@@ -153,8 +198,9 @@ class SpeechToCompletionBridgeTransformationHandler:
         return wav_header + pcm_data
 
     def _is_gemini_tts_model(self, model: str) -> bool:
-        """Check if the model is a Gemini TTS model that returns PCM16 data."""
-        return "gemini" in model.lower() and ("tts" in model.lower() or "preview-tts" in model.lower())
+        from litellm.utils import is_gemini_tts_model
+
+        return is_gemini_tts_model(model)
 
     def _gemini_tts_response_body(self, decoded_audio: bytes, response_format: str | None) -> tuple[bytes, str]:
         if response_format == GEMINI_TTS_RAW_RESPONSE_FORMAT:
@@ -162,7 +208,7 @@ class SpeechToCompletionBridgeTransformationHandler:
         return self._convert_pcm16_to_wav(decoded_audio), "audio/wav"
 
     def transform_response(
-        self, model_response: "ModelResponse", response_format: str | None
+        self, model_response: "ModelResponse", model: str, response_format: str | None
     ) -> "HttpxBinaryResponseContent":
         import base64
 
@@ -176,10 +222,9 @@ class SpeechToCompletionBridgeTransformationHandler:
             raise ValueError("No audio part found in the response")
         decoded_audio: Final = base64.b64decode(audio_part.data)
 
-        model: Final = getattr(model_response, "model", "")
         content, content_type = (
             self._gemini_tts_response_body(decoded_audio, response_format)
-            if self._is_gemini_tts_model(model)
+            if self._is_gemini_tts_model(model) or self._is_gemini_tts_model(model_response.model or "")
             else (decoded_audio, "audio/mpeg")
         )
         response: Final = httpx.Response(

litellm/harness/endpoint.py

@@ -37,6 +37,7 @@ from litellm.harness.context import GatewayTarget
 from litellm.harness.errors import HarnessError, HarnessInstallFailed
 from litellm.harness.types import Harness, Usage
 from litellm.llms.custom_httpx.http_handler import get_async_httpx_client
+from litellm.types.litellm_params import STREAM_USAGE_INJECTED_FIELD
 from litellm.types.llms.custom_http import httpxSpecialProvider
 
 if TYPE_CHECKING:
@@ -690,18 +691,21 @@ class ModelEndpoint:
     async def _invoke_sdk(
         self,
         route: str,
-        kwargs: dict[str, Any],  # mutable-ok: injects stream_options into the SDK kwargs
+        kwargs: dict[str, Any],  # mutable-ok: SDK call kwargs built by _sdk_kwargs
     ) -> object:
         if route == ROUTE_MESSAGES:
             return await litellm.anthropic.messages.acreate(**kwargs)
         if route == ROUTE_CHAT:
-            if kwargs.get("stream"):
-                stream_options = kwargs.get("stream_options") or {}  # mutable-ok: empty default for a JSON field
-                kwargs["stream_options"] = {  # mutable-ok: JSON field sent to litellm.acompletion
-                    "include_usage": True,
-                    **stream_options,
+            if not kwargs.get("stream"):
+                return await litellm.acompletion(**kwargs)
+            stream_options = kwargs.get("stream_options") or {}  # mutable-ok: empty default for a JSON field
+            return await litellm.acompletion(
+                **{
+                    **kwargs,
+                    "stream_options": {"include_usage": True, **stream_options},
+                    STREAM_USAGE_INJECTED_FIELD: "include_usage" not in stream_options,
                 }
-            return await litellm.acompletion(**kwargs)
+            )
         return await litellm.aresponses(**kwargs)
 
     async def _call_sdk(self, route: str, body: Mapping[str, object]) -> Response:

litellm/litellm_core_utils/core_helpers.py

@@ -329,11 +329,14 @@ def add_missing_spend_metadata_to_litellm_metadata(litellm_metadata: dict, metad
     Helper to get litellm metadata for spend tracking
 
     PATCH for issue where both `litellm_metadata` and `metadata` are present in the kwargs
-    and user_api_key values are in 'metadata'.
+    and user_api_key values are in 'metadata'. The proxy receive stamp rides along for the same
+    reason: a guardrail hook can plant `litellm_metadata` on a chat request after the proxy stamped
+    `metadata`, and the response timing headers anchor on that stamp.
     """
     potential_spend_tracking_metadata_substring: Final = "user_api_key"
     for key, value in dict(metadata).items():
-        if potential_spend_tracking_metadata_substring in key:
+        carry_receive_stamp = key == "litellm_received_at" and key not in litellm_metadata
+        if potential_spend_tracking_metadata_substring in key or carry_receive_stamp:
             litellm_metadata[key] = value
     return litellm_metadata
 

litellm/litellm_core_utils/litellm_logging.py

@@ -12,7 +12,7 @@ import subprocess
 import sys
 import time
 import traceback
-from collections.abc import Awaitable, Callable, Iterable, Iterator, Mapping, Sequence
+from collections.abc import Awaitable, Callable, Coroutine, Iterable, Iterator, Mapping, Sequence
 from dataclasses import dataclass
 from datetime import datetime as dt_object
 from functools import lru_cache
@@ -138,6 +138,7 @@ from litellm.types.interactions import (
     InteractionsAPIResponse,
     InteractionsAPIStreamingResponse,
 )
+from litellm.types.litellm_params import STREAM_USAGE_INJECTED_FIELD
 from litellm.types.llms.openai import (
     AllMessageValues,
     Batch,
@@ -574,6 +575,27 @@ class UncachedPrice:
     cost_breakdown: CostBreakdown | None
 
 
+@dataclass(frozen=True, slots=True)
+class BilledRoundsSpend:
+    usage: Usage
+    cost: float
+    cost_breakdown: CostBreakdown | None
+
+
+@dataclass(frozen=True, slots=True)
+class BilledRoundsDispatch:
+    """A parked success log for a stream bridged through chat completions that billed several rounds, and the spend
+    its failure log bills when a post-call hook rejects the response, as it does for a rejected one-round bridged
+    stream."""
+
+    dispatch: Coroutine[object, object, None]
+    spend: BilledRoundsSpend
+
+
+def _fixed_margin(cost_breakdown: CostBreakdown | None) -> float:
+    return 0.0 if cost_breakdown is None else cost_breakdown.get("margin_fixed_amount", 0.0)
+
+
 def _timestamp_seconds(moment: object) -> float | None:
     if isinstance(moment, datetime.datetime):
         return moment.timestamp()
@@ -582,6 +604,10 @@ def _timestamp_seconds(moment: object) -> float | None:
     return None
 
 
+def same_attempt_marker(first: object, second: object) -> bool:
+    return first is second or first == second
+
+
 class Logging(LiteLLMLoggingBaseClass):
     global \
         supabaseClient, \
@@ -595,6 +621,7 @@ class Logging(LiteLLMLoggingBaseClass):
         prometheusLogger, \
         slack_app
     custom_pricing: bool = False
+    replaced_global_callback_keys: frozenset[str] = frozenset()
     stream_options = None
     litellm_request_debug: bool = False
     streamed_anthropic_message_id: str | None = None
@@ -602,6 +629,7 @@ class Logging(LiteLLMLoggingBaseClass):
     classifier_input: Mapping[str, JsonValue] | None = None
     baseline_cache_context: "BaselineCacheContext | None" = None
     baseline_observation: "CapturedBaselineObservation | None" = None
+    _on_failure_payload_built: "Callable[[], None] | None" = None
 
     @property
     def _defer_async_logging(self) -> bool:
@@ -649,11 +677,11 @@ class Logging(LiteLLMLoggingBaseClass):
         litellm_call_id: str,
         function_id: str,
         litellm_trace_id: str | None = None,
-        dynamic_input_callbacks: list[str | Callable | CustomLogger] | None = None,
+        dynamic_input_callbacks: list[str | Callable[..., object] | CustomLogger] | None = None,
         dynamic_success_callbacks: list[str | Callable[..., object] | CustomLogger] | None = None,
         dynamic_async_success_callbacks: list[str | Callable[..., object] | CustomLogger] | None = None,
-        dynamic_failure_callbacks: list[str | Callable | CustomLogger] | None = None,
-        dynamic_async_failure_callbacks: list[str | Callable | CustomLogger] | None = None,
+        dynamic_failure_callbacks: list[str | Callable[..., object] | CustomLogger] | None = None,
+        dynamic_async_failure_callbacks: list[str | Callable[..., object] | CustomLogger] | None = None,
         applied_guardrails: list[str] | None = None,
         kwargs: dict | None = None,
         log_raw_request_response: bool = False,
@@ -698,6 +726,7 @@ class Logging(LiteLLMLoggingBaseClass):
         self._pre_call_session_id: str = session_id_var.get()
         _sid: Final = kwargs.get("litellm_session_id") if kwargs else None
         self.litellm_session_id: str = str(_sid) if _sid else ""
+        self.stream_usage_injected: Final = kwargs is not None and kwargs.get(STREAM_USAGE_INJECTED_FIELD) is True
         # supports_correlation_logging is False for calls originating from the
         # sync client entry point (wrapper() in utils.py): a plain OS thread
         # has no per-call context isolation the way an asyncio Task does, and
@@ -726,15 +755,17 @@ class Logging(LiteLLMLoggingBaseClass):
         self.raw_request_only = raw_request_only
 
         # Initialize dynamic callbacks
-        self.dynamic_input_callbacks: list[str | Callable | CustomLogger] | None = dynamic_input_callbacks
+        self.dynamic_input_callbacks: list[str | Callable[..., object] | CustomLogger] | None = dynamic_input_callbacks
         self.dynamic_success_callbacks: list[str | Callable[..., object] | CustomLogger] | None = (
             dynamic_success_callbacks
         )
         self.dynamic_async_success_callbacks: list[str | Callable[..., object] | CustomLogger] | None = (
             dynamic_async_success_callbacks
         )
-        self.dynamic_failure_callbacks: list[str | Callable | CustomLogger] | None = dynamic_failure_callbacks
-        self.dynamic_async_failure_callbacks: list[str | Callable | CustomLogger] | None = (
+        self.dynamic_failure_callbacks: list[str | Callable[..., object] | CustomLogger] | None = (
+            dynamic_failure_callbacks
+        )
+        self.dynamic_async_failure_callbacks: list[str | Callable[..., object] | CustomLogger] | None = (
             dynamic_async_failure_callbacks
         )
 
@@ -826,6 +857,7 @@ class Logging(LiteLLMLoggingBaseClass):
             "litellm_call_id": litellm_call_id,
             "start_time": start_time,
             "first_api_call_start_time": None,
+            "current_attempt_api_call_start_time": None,
             "litellm_params": litellm_params,
         }
         new_call.streaming_chunks = []
@@ -851,10 +883,32 @@ class Logging(LiteLLMLoggingBaseClass):
             self.dynamic_async_failure_callbacks, callback
         )
 
+    def replace_global_callback(self, callback: CustomLogger) -> None:
+        self.add_dynamic_callback(callback)
+        self.replaced_global_callback_keys = self.replaced_global_callback_keys | {
+            litellm.logging_callback_manager.get_custom_logger_key(callback)
+        }
+
+    def add_dynamic_function_callbacks(
+        self,
+        *,
+        async_success: Callable[..., object],
+        failure: Callable[..., object],
+        async_failure: Callable[..., object],
+    ) -> None:
+        self.dynamic_async_success_callbacks = self._with_dynamic_callback(
+            self.dynamic_async_success_callbacks, async_success
+        )
+        self.dynamic_failure_callbacks = self._with_dynamic_callback(self.dynamic_failure_callbacks, failure)
+        self.dynamic_async_failure_callbacks = self._with_dynamic_callback(
+            self.dynamic_async_failure_callbacks, async_failure
+        )
+
     @staticmethod
     def _with_dynamic_callback(
-        callbacks: Sequence[str | Callable | CustomLogger] | None, callback: CustomLogger
-    ) -> list[str | Callable | CustomLogger]:
+        callbacks: Sequence[str | Callable[..., object] | CustomLogger] | None,
+        callback: Callable[..., object] | CustomLogger,
+    ) -> list[str | Callable[..., object] | CustomLogger]:
         existing: Final = tuple(callbacks or ())
         return [*existing, *(() if callback in existing else (callback,))]
 
@@ -891,9 +945,9 @@ class Logging(LiteLLMLoggingBaseClass):
 
     def _process_dynamic_callback_list(
         self,
-        callback_list: list[str | Callable | CustomLogger] | None,
+        callback_list: list[str | Callable[..., object] | CustomLogger] | None,
         dynamic_callbacks_type: Literal["input", "success", "failure", "async_success", "async_failure"],
-    ) -> list[str | Callable | CustomLogger] | None:
+    ) -> list[str | Callable[..., object] | CustomLogger] | None:
         """
         Helper function to initialize CustomLogger compatible callbacks in self.dynamic_* callbacks
 
@@ -905,7 +959,7 @@ class Logging(LiteLLMLoggingBaseClass):
         if callback_list is None:
             return None
 
-        processed_list: Final[list[str | Callable | CustomLogger]] = []
+        processed_list: Final[list[str | Callable[..., object] | CustomLogger]] = []
         for callback in callback_list:
             if isinstance(callback, str) and callback in litellm._known_custom_logger_compatible_callbacks:
                 for callback_instance in self._resolve_dynamic_callback_string(callback):
@@ -1554,7 +1608,9 @@ class Logging(LiteLLMLoggingBaseClass):
             # A proxy-gate failure (auth, rate limit, client disconnect) fires pre_call for its callbacks
             # without any provider handoff, so it must not stamp the handoff instants.
             if LITELLM_LOGGING_NO_UPSTREAM_LLM_CALL not in self.model_call_details:
-                self.model_call_details["api_call_start_time"] = datetime.datetime.now()
+                api_call_start_time: Final = datetime.datetime.now()
+                self.model_call_details["api_call_start_time"] = api_call_start_time
+                self.model_call_details["current_attempt_api_call_start_time"] = api_call_start_time
                 # Set-once first provider-handoff instant. api_call_start_time
                 # is overwritten on every retry, so it can't measure one-time
                 # preprocessing; pinning the first attempt excludes retry loops
@@ -1562,11 +1618,12 @@ class Logging(LiteLLMLoggingBaseClass):
                 # litellm_params["metadata"] (caller request metadata, typed
                 # Dict[str, str], echoed downstream; a datetime breaks it).
                 if self.model_call_details.get("first_api_call_start_time") is None:
-                    self.model_call_details["first_api_call_start_time"] = self.model_call_details[
-                        "api_call_start_time"
-                    ]
+                    self.model_call_details["first_api_call_start_time"] = api_call_start_time
             # Input Integration Logging -> If you want to log the fact that an attempt to call the model was made
-            callbacks: Final = litellm.input_callback + (self.dynamic_input_callbacks or [])
+            callbacks: Final = (
+                *self._without_replaced_global_callbacks(litellm.input_callback),
+                *(self.dynamic_input_callbacks or ()),
+            )
             for callback in callbacks:
                 try:
                     if callback == "supabase" and supabaseClient is not None:
@@ -1739,7 +1796,10 @@ class Logging(LiteLLMLoggingBaseClass):
             )
             # Input Integration Logging -> If you want to log the fact that an attempt to call the model was made
 
-            callbacks: Final = litellm.input_callback + (self.dynamic_input_callbacks or [])
+            callbacks: Final = (
+                *self._without_replaced_global_callbacks(litellm.input_callback),
+                *(self.dynamic_input_callbacks or ()),
+            )
             for callback in callbacks:
                 try:
                     if callback == "sentry" and add_breadcrumb:
@@ -2388,6 +2448,31 @@ class Logging(LiteLLMLoggingBaseClass):
         self.model_call_details["combined_usage_object"] = usage
         self.model_call_details["response_cost"] = response_cost
 
+    def record_failure_after_earlier_rounds(
+        self, earlier: BilledRoundsSpend, failed_usage: Usage, failed_cost: float
+    ) -> None:
+        """Bill a failed round on the failure log after the rounds of its request before it were billed on the success
+        log.
+
+        The fixed margin is a per-request fee, so it is charged once: the failed round's margin is dropped when the
+        earlier rounds already paid one."""
+        from litellm.responses.mcp.round_billing import breakdown_less_repeated_fixed_margin
+
+        failed_breakdown: Final = None if self.cost_breakdown is earlier.cost_breakdown else self.cost_breakdown
+        repeated_fixed_margin: Final = (
+            0.0
+            if failed_breakdown is None
+            else min(_fixed_margin(earlier.cost_breakdown), _fixed_margin(failed_breakdown), earlier.cost, failed_cost)
+        )
+        self.model_call_details["combined_usage_object"] = failed_usage
+        self.model_call_details["response_cost"] = failed_cost - repeated_fixed_margin
+        self.cost_breakdown = breakdown_less_repeated_fixed_margin(failed_breakdown, repeated_fixed_margin)
+
+    def record_billed_rounds_for_failure(self, spend: BilledRoundsSpend) -> None:
+        """Bill every round of a fully streamed request on the failure log when a post-call hook rejects it."""
+        self.record_partial_usage_for_failure(spend.usage, spend.cost)
+        self.cost_breakdown = spend.cost_breakdown
+
     def attempt_scoped_copy(self) -> "Logging":
         """A copy with its own model_call_details, so logging one attempt of a request through it leaves
         the state the request's other attempts share untouched."""
@@ -2395,12 +2480,37 @@ class Logging(LiteLLMLoggingBaseClass):
         attempt.model_call_details = dict(self.model_call_details)
         return attempt
 
+    def dispatch_attempt_failure_handlers(
+        self, exception: Exception, traceback_exception: str
+    ) -> Coroutine[None, None, None]:
+        shared_payload: Final = self.model_call_details.get("standard_logging_object")
+        attempt_marker: Final = self.model_call_details.get("current_attempt_api_call_start_time")
+        attempt: Final = self.attempt_scoped_copy()
+
+        def publish() -> None:
+            attempt_payload: Final = attempt.model_call_details.get("standard_logging_object")
+            current_marker: Final = self.model_call_details.get("current_attempt_api_call_start_time")
+            if (
+                attempt_payload is not None
+                and self.model_call_details.get("standard_logging_object") is shared_payload
+                and same_attempt_marker(current_marker, attempt_marker)
+            ):
+                self.model_call_details["standard_logging_object"] = attempt_payload
+
+        attempt._on_failure_payload_built = publish
+
+        async def dispatch() -> None:
+            await attempt.dispatch_failure_handlers(exception, traceback_exception, prefer_async_handlers=True)
+
+        return dispatch()
+
     def carry_failed_attempt_usage(self) -> None:
         """Move what a failed earlier attempt stashed off the retry sharing this logging object, so the retry logs its
         own usage while the request's failure log still bills every attempt."""
         usage: Final = self.model_call_details.pop("combined_usage_object", None)
         response_cost: Final = self.model_call_details.pop("response_cost", None)
         if isinstance(usage, Usage):
+            self.model_call_details["current_attempt_api_call_start_time"] = None
             self.model_call_details["failed_attempts_usage"] = (
                 *self.model_call_details.get("failed_attempts_usage", ()),
                 (usage, float(response_cost) if isinstance(response_cost, (int, float)) else 0.0),
@@ -2514,6 +2624,10 @@ class Logging(LiteLLMLoggingBaseClass):
             """
             return
         self.model_call_details[f"has_logged_{event_type}"] = True
+        if event_type == "async_failure":
+            self.model_call_details["async_failure_logged_attempt"] = self.model_call_details.get(
+                "current_attempt_api_call_start_time"
+            )
         return
 
     def has_run_logging(
@@ -4331,7 +4445,8 @@ class Logging(LiteLLMLoggingBaseClass):
             start_time=start_time,
             end_time=end_time,
         )
-
+        if self._on_failure_payload_built is not None:
+            self._on_failure_payload_built()
         callbacks: Final = self.get_combined_callback_list(
             dynamic_success_callbacks=self.dynamic_async_failure_callbacks,
             global_callbacks=litellm._async_failure_callback,
@@ -4516,14 +4631,26 @@ class Logging(LiteLLMLoggingBaseClass):
         _filtered_failure_callbacks = self._remove_internal_litellm_callbacks(_filtered_failure_callbacks)
         return len(_filtered_failure_callbacks) > 0
 
+    def _without_replaced_global_callbacks(
+        self, global_callbacks: Iterable[str | Callable[..., object] | CustomLogger]
+    ) -> tuple[str | Callable[..., object] | CustomLogger, ...]:
+        if not self.replaced_global_callback_keys:
+            return tuple(global_callbacks)
+        return tuple(
+            litellm.logging_callback_manager.without_custom_logger_keys(
+                global_callbacks, self.replaced_global_callback_keys
+            )
+        )
+
     def get_combined_callback_list(
         self,
         dynamic_success_callbacks: list[str | Callable[..., object] | CustomLogger] | None,
         global_callbacks: list[str | Callable[..., object] | CustomLogger],
     ) -> list[str | Callable[..., object] | CustomLogger]:
+        kept_global_callbacks: Final = self._without_replaced_global_callbacks(global_callbacks)
         if dynamic_success_callbacks is None:
-            return list(global_callbacks)
-        return list(dict.fromkeys(dynamic_success_callbacks + global_callbacks))
+            return list(kept_global_callbacks)
+        return list(dict.fromkeys((*dynamic_success_callbacks, *kept_global_callbacks)))
 
     def _remove_internal_litellm_callbacks(self, callbacks: list) -> list:
         """

litellm/litellm_core_utils/logging_callback_manager.py

@@ -1,4 +1,6 @@
-from collections.abc import Callable
+import threading
+from collections.abc import Callable, Iterable, Iterator
+from types import MethodType
 from typing import TYPE_CHECKING, Final
 
 import litellm
@@ -15,6 +17,7 @@ else:
     _custom_logger_compatible_callbacks_literal = str
 
 _generic_api_logger_cache: Final[dict[str, GenericAPILogger]] = {}
+_callback_lists_lock: Final = threading.RLock()
 
 
 class LoggingCallbackManager:
@@ -43,16 +46,18 @@ class LoggingCallbackManager:
         Add a input callback to litellm.input_callback.
         Auto-routes async callbacks to litellm._async_input_callback.
         """
-        if not isinstance(callback, str) and self._is_async_callable(callback):
-            self._safe_add_callback_to_list(callback=callback, parent_list=litellm._async_input_callback)
-        else:
-            self._safe_add_callback_to_list(callback=callback, parent_list=litellm.input_callback)
+        with _callback_lists_lock:
+            if not isinstance(callback, str) and self._is_async_callable(callback):
+                self._safe_add_callback_to_list(callback=callback, parent_list=litellm._async_input_callback)
+            else:
+                self._safe_add_callback_to_list(callback=callback, parent_list=litellm.input_callback)
 
     def add_litellm_service_callback(self, callback: CustomLogger | str | Callable[..., object]):
         """
         Add a service callback to litellm.service_callback
         """
-        self._safe_add_callback_to_list(callback=callback, parent_list=litellm.service_callback)
+        with _callback_lists_lock:
+            self._safe_add_callback_to_list(callback=callback, parent_list=litellm.service_callback)
 
     def add_litellm_callback(self, callback: CustomLogger | str | Callable[..., object]):
         """
@@ -60,10 +65,11 @@ class LoggingCallbackManager:
 
         Ensures no duplicates are added.
         """
-        self._safe_add_callback_to_list(
-            callback=callback,
-            parent_list=litellm.callbacks,
-        )
+        with _callback_lists_lock:
+            self._safe_add_callback_to_list(
+                callback=callback,
+                parent_list=litellm.callbacks,
+            )
 
     def add_litellm_success_callback(self, callback: CustomLogger | str | Callable[..., object]):
         """
@@ -71,37 +77,41 @@ class LoggingCallbackManager:
         Auto-routes async callbacks to litellm._async_success_callback.
         Special-cases 'dynamodb' and 'openmeter' as async callbacks.
         """
-        if (
-            isinstance(callback, str)
-            and callback in ("dynamodb", "openmeter")
-            or not isinstance(callback, str)
-            and self._is_async_callable(callback)
-        ):
-            self._safe_add_callback_to_list(callback=callback, parent_list=litellm._async_success_callback)
-        else:
-            self._safe_add_callback_to_list(callback=callback, parent_list=litellm.success_callback)
+        with _callback_lists_lock:
+            if (
+                isinstance(callback, str)
+                and callback in ("dynamodb", "openmeter")
+                or not isinstance(callback, str)
+                and self._is_async_callable(callback)
+            ):
+                self._safe_add_callback_to_list(callback=callback, parent_list=litellm._async_success_callback)
+            else:
+                self._safe_add_callback_to_list(callback=callback, parent_list=litellm.success_callback)
 
     def add_litellm_failure_callback(self, callback: CustomLogger | str | Callable[..., object]):
         """
         Add a failure callback to `litellm.failure_callback`.
         Auto-routes async callbacks to litellm._async_failure_callback.
         """
-        if not isinstance(callback, str) and self._is_async_callable(callback):
-            self._safe_add_callback_to_list(callback=callback, parent_list=litellm._async_failure_callback)
-        else:
-            self._safe_add_callback_to_list(callback=callback, parent_list=litellm.failure_callback)
+        with _callback_lists_lock:
+            if not isinstance(callback, str) and self._is_async_callable(callback):
+                self._safe_add_callback_to_list(callback=callback, parent_list=litellm._async_failure_callback)
+            else:
+                self._safe_add_callback_to_list(callback=callback, parent_list=litellm.failure_callback)
 
     def add_litellm_async_success_callback(self, callback: CustomLogger | Callable[..., object] | str):
         """
         Add a success callback to litellm._async_success_callback
         """
-        self._safe_add_callback_to_list(callback=callback, parent_list=litellm._async_success_callback)
+        with _callback_lists_lock:
+            self._safe_add_callback_to_list(callback=callback, parent_list=litellm._async_success_callback)
 
     def add_litellm_async_failure_callback(self, callback: CustomLogger | Callable[..., object] | str):
         """
         Add a failure callback to litellm._async_failure_callback
         """
-        self._safe_add_callback_to_list(callback=callback, parent_list=litellm._async_failure_callback)
+        with _callback_lists_lock:
+            self._safe_add_callback_to_list(callback=callback, parent_list=litellm._async_failure_callback)
 
     def remove_callback_from_list_by_object(self, callback_list, obj, require_self=True):
         """
@@ -110,13 +120,14 @@ class LoggingCallbackManager:
         if not isinstance(callback_list, list):  # Not list -> do nothing
             return
 
-        if require_self:
-            remove_list = [c for c in callback_list if hasattr(c, "__self__") and c.__self__ == obj]
-        else:
-            remove_list = [c for c in callback_list if c == obj]
+        with _callback_lists_lock:
+            if require_self:
+                remove_list = [c for c in callback_list if hasattr(c, "__self__") and c.__self__ == obj]
+            else:
+                remove_list = [c for c in callback_list if c == obj]
 
-        for c in remove_list:
-            callback_list.remove(c)
+            for c in remove_list:
+                callback_list.remove(c)
 
     def remove_callbacks_by_type(self, callback_list, callback_type):
         """
@@ -134,10 +145,11 @@ class LoggingCallbackManager:
         if not isinstance(callback_list, list):
             return
 
-        remove_list: Final = [c for c in callback_list if isinstance(c, callback_type)]
+        with _callback_lists_lock:
+            remove_list: Final = [c for c in callback_list if isinstance(c, callback_type)]
 
-        for c in remove_list:
-            callback_list.remove(c)
+            for c in remove_list:
+                callback_list.remove(c)
 
     def _add_string_callback_to_list(
         self, callback: str, parent_list: list[CustomLogger | Callable[..., object] | str]
@@ -336,6 +348,55 @@ class LoggingCallbackManager:
 
         return "-".join(key_parts)
 
+    get_custom_logger_key = _get_custom_logger_key
+    callback_lists_lock: Final = _callback_lists_lock
+
+    def drop_callbacks(self, callbacks: Iterable[object], method_owners: Iterable[object] = ()) -> frozenset[str]:
+        dropped_ids: Final = frozenset(map(id, callbacks))
+        owner_ids: Final = frozenset(map(id, method_owners))
+
+        def keep(callback: object) -> bool:
+            if id(callback) in dropped_ids:
+                return False
+            return not (isinstance(callback, MethodType) and id(callback.__self__) in owner_ids)
+
+        with _callback_lists_lock:
+            vacated_keys: Final = frozenset(
+                self._get_custom_logger_key(callback)
+                for callback in litellm.callbacks
+                if isinstance(callback, CustomLogger) and id(callback) in dropped_ids
+            )
+            litellm.callbacks = [callback for callback in litellm.callbacks if keep(callback)]
+            litellm.input_callback = [callback for callback in litellm.input_callback if keep(callback)]
+            litellm.success_callback = [callback for callback in litellm.success_callback if keep(callback)]
+            litellm.failure_callback = [callback for callback in litellm.failure_callback if keep(callback)]
+            litellm.service_callback = [callback for callback in litellm.service_callback if keep(callback)]
+            litellm._async_success_callback = [
+                callback for callback in litellm._async_success_callback if keep(callback)
+            ]
+            litellm._async_failure_callback = [
+                callback for callback in litellm._async_failure_callback if keep(callback)
+            ]
+        return vacated_keys
+
+    def with_custom_logger_keys(
+        self, candidates: Iterable[CustomLogger], keys: frozenset[str]
+    ) -> tuple[CustomLogger, ...]:
+        if not keys:
+            return ()
+        class_names: Final = frozenset(key.split("-", 1)[0] for key in keys)
+        same_class: Final = (candidate for candidate in candidates if type(candidate).__name__ in class_names)
+        return tuple(candidate for candidate in same_class if self._get_custom_logger_key(candidate) in keys)
+
+    def without_custom_logger_keys(
+        self, callbacks: Iterable[str | Callable[..., object] | CustomLogger], keys: frozenset[str]
+    ) -> Iterator[str | Callable[..., object] | CustomLogger]:
+        return (
+            callback
+            for callback in callbacks
+            if not (isinstance(callback, CustomLogger) and self._get_custom_logger_key(callback) in keys)
+        )
+
     def _reset_all_callbacks(self):
         """
         Reset all callbacks to an empty list
@@ -363,19 +424,21 @@ class LoggingCallbackManager:
 
     _get_all_callbacks = get_all_callbacks
 
-    def remove_callback_from_all_lists(self, obj, require_self=False) -> None:
+    def remove_callback_from_all_lists(self, obj: object, require_self: bool = False) -> None:
         """
         Remove a callback object from every callback list it may have been
         promoted into, so a re-initialized callback leaves no stale instance behind.
         """
-        for callback_list in (
-            litellm.callbacks,
-            litellm.success_callback,
-            litellm.failure_callback,
-            litellm._async_success_callback,
-            litellm._async_failure_callback,
-        ):
-            self.remove_callback_from_list_by_object(callback_list, obj, require_self=require_self)
+        with _callback_lists_lock:
+            for callback_list in (
+                litellm.callbacks,
+                litellm.input_callback,
+                litellm.success_callback,
+                litellm.failure_callback,
+                litellm._async_success_callback,
+                litellm._async_failure_callback,
+            ):
+                self.remove_callback_from_list_by_object(callback_list, obj, require_self=require_self)
 
     def get_active_additional_logging_utils_from_custom_logger(
         self,

litellm/litellm_core_utils/redact_messages.py

@@ -24,6 +24,7 @@ from litellm.litellm_core_utils.core_helpers import (
     get_metadata_variable_name_from_kwargs,
 )
 from litellm.litellm_core_utils.served_output_texts import SERVED_OUTPUT_TEXTS_KEY
+from litellm.llms.base_llm.guardrail_translation.utils import stream_item_field, stream_item_items
 from litellm.llms.vertex_ai.common_utils import (
     redact_vertex_ai_metadata_from_litellm_params,
     redact_vertex_ai_metadata_from_logged_object,
@@ -41,6 +42,8 @@ if TYPE_CHECKING:
 else:
     LiteLLMLoggingObject = Any
 
+MESSAGE_REDACTION_ENABLE_HEADERS: Final = ("litellm-enable-message-redaction", "x-litellm-enable-message-redaction")
+
 
 def _opted_out_of_message_logging(custom_logger: CustomLogger) -> bool:
     return hasattr(custom_logger, "message_logging") and custom_logger.message_logging is not True
@@ -195,33 +198,45 @@ def _redacted_provider_specific_fields(fields: Mapping[str, object]) -> Mapping[
     return {
         **fields,
         **({"reasoning_content": REDACTED_BY_LITELLM} if fields.get("reasoning_content") is not None else {}),
-        **{key: None for key in ("thinking_blocks", "reasoningContentBlocks", "citations") if key in fields},
+        **{
+            key: None
+            for key in (
+                "thinking_blocks",
+                "reasoningContentBlocks",
+                "citations",
+                "compaction_blocks",
+                "compaction_delta",
+            )
+            if key in fields
+        },
     }
 
 
+def _redact_field_when_present(item: object, field: str) -> None:
+    if getattr(item, field, None) is not None:
+        setattr(item, field, REDACTED_BY_LITELLM)
+
+
 def _redact_responses_api_output(output_items):
     """Helper to redact ResponsesAPIResponse output items."""
     for output_item in output_items:
+        output_type = stream_item_field(output_item, "type")
         if getattr(output_item, "text", None) is not None:
             output_item.text = REDACTED_BY_LITELLM
 
-        if hasattr(output_item, "content") and isinstance(output_item.content, list):
-            for content_part in output_item.content:
-                if getattr(content_part, "text", None) is not None:
-                    content_part.text = REDACTED_BY_LITELLM
-                if getattr(content_part, "refusal", None) is not None:
-                    content_part.refusal = REDACTED_BY_LITELLM
+        for content_part in stream_item_items(output_item, "content"):
+            _redact_field_when_present(content_part, "text")
+            _redact_field_when_present(content_part, "refusal")
 
-        # Redact reasoning items in output array
-        if hasattr(output_item, "type") and output_item.type == "reasoning":
-            if hasattr(output_item, "summary") and isinstance(output_item.summary, list):
-                for summary_item in output_item.summary:
-                    if getattr(summary_item, "text", None) is not None:
-                        summary_item.text = REDACTED_BY_LITELLM
+        if output_type == "reasoning":
+            for summary_item in stream_item_items(output_item, "summary"):
+                _redact_field_when_present(summary_item, "text")
 
-        if hasattr(output_item, "type") and output_item.type == "function_call" and hasattr(output_item, "arguments"):
+        if output_type == "compaction" and hasattr(output_item, "encrypted_content"):
+            output_item.encrypted_content = REDACTED_BY_LITELLM
+        if output_type == "function_call" and hasattr(output_item, "arguments"):
             output_item.arguments = REDACTED_BY_LITELLM
-        if hasattr(output_item, "type") and output_item.type == "custom_tool_call" and hasattr(output_item, "input"):
+        if output_type == "custom_tool_call" and hasattr(output_item, "input"):
             output_item.input = REDACTED_BY_LITELLM
 
 
@@ -262,6 +277,11 @@ def _redacted_responses_api_output_item(item: object, redacted_str: str) -> obje
             if item_mapping.get("type") == "reasoning" and isinstance(summary, list)
             else {}
         ),
+        **(
+            {"encrypted_content": redacted_str}
+            if item_mapping.get("type") == "compaction" and "encrypted_content" in item_mapping
+            else {}
+        ),
         **(
             {"arguments": redacted_str}
             if item_mapping.get("type") == "function_call" and "arguments" in item_mapping
@@ -556,13 +576,8 @@ def should_redact_message_logging(model_call_details: dict) -> bool:
         # User explicitly disabled redaction via header
         return False
 
-    possible_enable_headers: Final = [
-        "litellm-enable-message-redaction",  # old header. maintain backwards compatibility
-        "x-litellm-enable-message-redaction",  # new header
-    ]
-
     is_redaction_enabled_via_header = False
-    for header in possible_enable_headers:
+    for header in MESSAGE_REDACTION_ENABLE_HEADERS:
         if bool(request_headers.get(header, False)):
             is_redaction_enabled_via_header = True
             break

litellm/litellm_core_utils/streaming_handler.py

@@ -10,7 +10,7 @@ import traceback
 from collections.abc import AsyncIterator, Callable, Iterable, Iterator, Mapping, Sequence
 from dataclasses import dataclass
 from types import MappingProxyType
-from typing import Any, Final, NoReturn, Protocol, TypeVar, cast
+from typing import TYPE_CHECKING, Any, Final, NoReturn, Protocol, TypeVar, cast
 
 import anyio
 import httpx
@@ -60,6 +60,9 @@ from .exception_mapping_utils import exception_type, with_deployment_api_base
 from .llm_response_utils.get_api_base import get_api_base
 from .rules import Rules
 
+if TYPE_CHECKING:
+    from .litellm_logging import BilledRoundsSpend
+
 # Constants for special delta attribute names
 AUDIO_ATTRIBUTE: Final = "audio"
 IMAGE_ATTRIBUTE: Final = "images"
@@ -259,6 +262,8 @@ def _provider_hidden_params(
 
 
 class CustomStreamWrapper:
+    _earlier_rounds_spend: "BilledRoundsSpend | None" = None
+
     @property
     def hidden_params(self) -> dict[str, object]:  # mutable-ok: API requires mutation
         return self._hidden_params
@@ -342,6 +347,10 @@ class CustomStreamWrapper:
         self.messages = getattr(logging_obj, "messages", None)
         self.sent_stream_usage = False
         self.send_stream_usage = True if self.check_send_stream_usage(self.stream_options) else False
+        self.caller_requested_stream_usage: Final = (
+            self.check_send_stream_usage(getattr(logging_obj, "stream_options", None))
+            and getattr(logging_obj, "stream_usage_injected", False) is not True
+        )
         self.tool_call = False
         # keep track of the returned chunks - used for calculating the input/output tokens for stream options
         self.chunks: list[ModelResponseStream] = []
@@ -2070,7 +2079,7 @@ class CustomStreamWrapper:
         except StopIteration:
             if self.sent_last_chunk is True:
                 complete_streaming_response: Final = self._build_complete_streaming_response()
-                response = self.model_response_creator()
+                response = self._new_usage_chunk()
                 if complete_streaming_response is not None:
                     self._propagate_usage_cost_to_hidden_params(complete_streaming_response, self.custom_llm_provider)
 
@@ -2222,27 +2231,10 @@ class CustomStreamWrapper:
                     )
 
                     if _has_usage:
-                        # Store a copy ONLY when usage stripping below will mutate
-                        # the chunk. For non-usage chunks (vast majority), store
-                        # directly to avoid expensive model_copy() per chunk.
-                        self.chunks.append(processed_chunk.model_copy())
-
-                        # Strip usage from the outgoing chunk so it's not sent twice
-                        # (once in the chunk, once in _hidden_params).
-                        obj_dict = processed_chunk.model_dump()
-                        if "usage" in obj_dict:
-                            del obj_dict["usage"]
-                        processed_chunk = self.model_response_creator(
-                            chunk=obj_dict,
-                            hidden_params=cast(  # cast-ok: preserve mapping operations on dynamic storage
-                                Mapping[str, object], getattr(processed_chunk, HIDDEN_PARAMS_ATTR)
-                            ),
-                        )
-                        is_empty = is_model_response_stream_empty(
-                            model_response=cast(ModelResponseStream, processed_chunk)
-                        )
-                        if is_empty:
+                        usage_free_chunk = self._record_and_strip_usage_chunk(processed_chunk)
+                        if usage_free_chunk is None:
                             continue
+                        processed_chunk = usage_free_chunk
                     else:
                         # No usage data — safe to store directly without copying
                         self.chunks.append(processed_chunk)
@@ -2273,25 +2265,25 @@ class CustomStreamWrapper:
                         if processed_chunk is None:
                             continue
 
-                        choice = processed_chunk.choices[0]
-                        if isinstance(choice, StreamingChoices):
-                            self._accumulate_streamed_delta(choice.delta)
+                        if processed_chunk.choices:
+                            choice = processed_chunk.choices[0]
+                            if isinstance(choice, StreamingChoices):
+                                self._accumulate_streamed_delta(choice.delta)
                         self.rules.post_call_rules(input=self.response_uptil_now, model=self.model)
-                        # RETURN RESULT
-                        self.chunks.append(processed_chunk)
-                        return processed_chunk
+                        if not self.send_stream_usage or getattr(processed_chunk, "usage", None) is None:
+                            self.chunks.append(processed_chunk)
+                            return processed_chunk
+                        usage_free_chunk = self._record_and_strip_usage_chunk(processed_chunk)
+                        if usage_free_chunk is None:
+                            continue
+                        return usage_free_chunk
         except (StopAsyncIteration, StopIteration):
             return await self._finalize_completed_stream(cache_hit=cache_hit)
         except httpx.TimeoutException as e:  # if httpx read timeout error occues
             traceback_exception = traceback.format_exc()
             ## ADD DEBUG INFORMATION - E.G. LITELLM REQUEST TIMEOUT
             traceback_exception += f"\nLiteLLM Default Request Timeout - {litellm.request_timeout}"
-            if self.logging_obj is not None:
-                self._record_partial_usage_for_failure()
-                ## LOGGING
-                self._failure_logging = asyncio.create_task(
-                    self.logging_obj.dispatch_failure_handlers(e, traceback_exception, prefer_async_handlers=True)
-                )
+            self._dispatch_stream_failure(e, traceback_exception)
             await self.aclose()
             self._handle_stream_fallback_error(e)
         except (httpx.ReadError, httpx.RemoteProtocolError) as e:
@@ -2306,7 +2298,7 @@ class CustomStreamWrapper:
     async def _finalize_completed_stream(self, cache_hit: bool) -> "ModelResponseStream":
         if self.sent_last_chunk is True:
             complete_streaming_response: Final = await self._async_build_complete_streaming_response()
-            response: Final = self.model_response_creator()
+            response: Final = self._new_usage_chunk()
             if complete_streaming_response is not None:
                 self._propagate_usage_cost_to_hidden_params(complete_streaming_response, self.custom_llm_provider)
 
@@ -2392,29 +2384,36 @@ class CustomStreamWrapper:
             return processed_chunk
 
     def _log_stream_failure_and_raise(self, e: Exception) -> NoReturn:
-        traceback_exception: Final = traceback.format_exc()
-        if self.logging_obj is not None:
-            self._record_partial_usage_for_failure()
-            ## LOGGING
-            self._failure_logging = asyncio.create_task(
-                self.logging_obj.dispatch_failure_handlers(e, traceback_exception, prefer_async_handlers=True)
-            )
+        self._dispatch_stream_failure(e, traceback.format_exc())
         self._handle_stream_fallback_error(e)
 
-    async def wait_for_failure_logging(self, timeout: float) -> None:
+    def _dispatch_stream_failure(self, e: Exception, traceback_exception: str) -> None:
+        self._record_partial_usage_for_failure()
+        self._failure_logging = asyncio.create_task(
+            self.logging_obj.dispatch_attempt_failure_handlers(e, traceback_exception)
+        )
+
+    @property
+    def has_dispatched_failure_logging(self) -> bool:
+        return self._failure_logging is not None
+
+    async def wait_for_failure_logging(self, timeout: float | None) -> None:
         """Let this stream's failure log read its own deployment before a fallback reuses the logging object."""
         if self._failure_logging is None:
             return
         await asyncio.wait((self._failure_logging,), timeout=timeout)
 
+    def bill_failure_after_earlier_rounds(self, earlier: "BilledRoundsSpend") -> None:
+        self._earlier_rounds_spend = earlier
+
     def _record_partial_usage_for_failure(self) -> None:
         """
         A stream that breaks mid-flight still billed the provider for the chunks
         already delivered. Recover that partial usage from the chunks seen so
         far and stash it, with its cost, on the logging object so the failure
-        handler records the real partial spend instead of zero. A request that
-        later recovers via a router fallback overwrites this with the combined
-        success log on the same request id, so this never double counts.
+        handler records the real partial spend instead of zero. The router
+        moves it into terminal failure usage before opening the next attempt,
+        so a recovered attempt logs its own usage.
         """
         if self.logging_obj is None or not self.chunks:
             return
@@ -2433,10 +2432,12 @@ class CustomStreamWrapper:
             if self.model:
                 partial_response.model = self.model
             backfill_missing_cache_usage_fields(usage)
-            self.logging_obj.model_call_details["combined_usage_object"] = usage
-            self.logging_obj.model_call_details["response_cost"] = (
-                self.logging_obj.response_cost_calculator(result=partial_response) or 0.0
-            )
+            partial_cost: Final = self.logging_obj.response_cost_calculator(result=partial_response) or 0.0
+            if self._earlier_rounds_spend is None:
+                self.logging_obj.model_call_details["combined_usage_object"] = usage
+                self.logging_obj.model_call_details["response_cost"] = partial_cost
+                return
+            self.logging_obj.record_failure_after_earlier_rounds(self._earlier_rounds_spend, usage, partial_cost)
         except Exception as recover_error:
             verbose_logger.debug(
                 "could not recover partial usage for interrupted stream: %s",
@@ -2463,6 +2464,22 @@ class CustomStreamWrapper:
             return False
         return any(get(field) for field in cls._CONTINUATION_DISQUALIFYING_DELTA_FIELDS)
 
+    def _new_usage_chunk(self) -> ModelResponseStream:
+        usage_chunk: Final = self.model_response_creator()
+        if self.caller_requested_stream_usage:
+            usage_chunk.choices = []
+        return usage_chunk
+
+    def _record_and_strip_usage_chunk(self, chunk: ModelResponseStream) -> ModelResponseStream | None:
+        self.chunks.append(chunk.model_copy())
+        usage_free_chunk: Final = self.model_response_creator(
+            chunk={key: value for key, value in chunk.model_dump().items() if key != "usage"},
+            hidden_params=cast(  # cast-ok: preserve mapping operations on dynamic storage
+                Mapping[str, object], getattr(chunk, HIDDEN_PARAMS_ATTR)
+            ),
+        )
+        return None if is_model_response_stream_empty(model_response=usage_free_chunk) else usage_free_chunk
+
     def _accumulate_streamed_delta(self, delta: object) -> None:
         # Shared by the sync, async, and non-aiohttp iteration sites so answer
         # text and the disqualifying-content latch stay in step across all three.

litellm/litellm_core_utils/url_utils.py

@@ -23,7 +23,7 @@ import asyncio
 import socket
 from ipaddress import ip_address, ip_network
 from typing import Any, Final, Protocol
-from urllib.parse import quote, urlparse, urlunparse
+from urllib.parse import quote, urlparse, urlsplit, urlunparse
 
 import httpx
 from typing_extensions import ReadOnly, TypedDict
@@ -129,6 +129,28 @@ def encode_url_path_segments(value: object, *, field_name: str = "path") -> str:
     return "/".join(encode_url_path_segment(segment, field_name=field_name) for segment in value_str.split("/"))
 
 
+def _strip_trailing_slashes_keeping_the_scheme(api_base: str) -> str:
+    stripped: Final = api_base.rstrip("/")
+    parts: Final = urlsplit(stripped)
+    if parts.scheme and not (parts.netloc or parts.path or parts.query or parts.fragment):
+        return api_base
+    if not parts.query or not parts.path.endswith("/"):
+        return stripped
+    url_before_query, _, query = stripped.partition("?")
+    return f"{url_before_query.rstrip('/')}?{query}"
+
+
+def join_api_base(api_base: str, route: str) -> str:
+    return f"{_strip_trailing_slashes_keeping_the_scheme(api_base)}/{route.lstrip('/')}"
+
+
+def ensure_api_route(api_base: str, route: str) -> str:
+    base: Final = _strip_trailing_slashes_keeping_the_scheme(api_base)
+    if urlsplit(base).path.endswith(f"/{route.lstrip('/')}"):
+        return base
+    return join_api_base(base, route)
+
+
 def _is_blocked_ip(addr: str) -> bool:
     """Return True for any IP not safe to reach from a user-supplied URL.
 

litellm/llms/aiml/image_generation/transformation.py

@@ -4,6 +4,7 @@ from typing import TYPE_CHECKING, Any, Final
 import httpx
 from pydantic import ConfigDict, TypeAdapter
 
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.image_generation.transformation import (
     BaseImageGenerationConfig,
 )
@@ -121,7 +122,7 @@ class AimlImageGenerationConfig(BaseImageGenerationConfig):
         complete_url = complete_url.rstrip("/")
         # Strip /v1 suffix if present since IMAGE_GENERATION_ENDPOINT already includes v1
         complete_url = complete_url.removesuffix("/v1")
-        complete_url = f"{complete_url}/{self.IMAGE_GENERATION_ENDPOINT}"
+        complete_url = join_api_base(complete_url, self.IMAGE_GENERATION_ENDPOINT)
         return complete_url
 
     def validate_environment(

litellm/llms/aiohttp_openai/chat/transformation.py

@@ -13,6 +13,7 @@ from typing import TYPE_CHECKING, Any, Final
 from aiohttp import ClientResponse
 from pydantic import ConfigDict, TypeAdapter
 
+from litellm.litellm_core_utils.url_utils import ensure_api_route
 from litellm.llms.openai_like.chat.transformation import OpenAILikeChatConfig
 from litellm.types.llms.openai import AllMessageValues
 from litellm.types.utils import Choices, ModelResponse
@@ -44,12 +45,7 @@ class AiohttpOpenAIChatConfig(OpenAILikeChatConfig):
         Ensure - /v1/chat/completions is at the end of the url
 
         """
-        if api_base is None:
-            api_base = "https://api.openai.com"
-
-        if not api_base.endswith("/chat/completions"):
-            api_base += "/chat/completions"
-        return api_base
+        return ensure_api_route("https://api.openai.com" if api_base is None else api_base, "chat/completions")
 
     def validate_environment(
         self,

litellm/llms/anthropic/batches/transformation.py

@@ -7,7 +7,7 @@ import httpx
 from httpx import Headers, Response
 from typing_extensions import ReadOnly, TypedDict
 
-from litellm.litellm_core_utils.url_utils import encode_url_path_segment
+from litellm.litellm_core_utils.url_utils import encode_url_path_segment, ensure_api_route, join_api_base
 from litellm.llms.base_llm.batches.transformation import BaseBatchesConfig
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.types.llms.openai import AllMessageValues, CreateBatchRequest
@@ -106,10 +106,8 @@ class AnthropicBatchesConfig(BaseBatchesConfig):
         data: CreateBatchRequest,
     ) -> str:
         """Get the complete URL for batch creation request."""
-        api_base = api_base or self.anthropic_model_info.get_api_base(api_base)
-        if not api_base.endswith("/v1/messages/batches"):
-            api_base = f"{api_base.rstrip('/')}/v1/messages/batches"
-        return api_base
+        api_base = self.anthropic_model_info.resolve_anthropic_api_base(api_base)
+        return ensure_api_route(api_base, "v1/messages/batches")
 
     def transform_create_batch_request(
         self,
@@ -158,9 +156,9 @@ class AnthropicBatchesConfig(BaseBatchesConfig):
         Returns:
             Complete URL for Anthropic batch retrieval: {api_base}/v1/messages/batches/{batch_id}
         """
-        api_base = api_base or self.anthropic_model_info.get_api_base(api_base)
+        api_base = self.anthropic_model_info.resolve_anthropic_api_base(api_base)
         encoded_batch_id: Final = encode_url_path_segment(batch_id, field_name="batch_id")
-        return f"{api_base.rstrip('/')}/v1/messages/batches/{encoded_batch_id}"
+        return join_api_base(api_base, f"v1/messages/batches/{encoded_batch_id}")
 
     def transform_retrieve_batch_request(
         self,

litellm/llms/anthropic/chat/handler.py

@@ -856,11 +856,21 @@ class ModelResponseIterator:
                 if reasoning_content is None:
                     reasoning_content = ""
         elif "content" in content_block["delta"] and content_block["delta"].get("type") == "compaction_delta":
-            # Handle compaction delta
+            compaction_delta: Final = content_block["delta"]["content"]
             provider_specific_fields["compaction_delta"] = {
                 "type": "compaction_delta",
-                "content": content_block["delta"]["content"],
+                "content": compaction_delta,
             }
+            if self.compaction_blocks and isinstance(compaction_delta, str):
+                streamed_summary: Final = self.compaction_blocks[-1].get("content")
+                self.compaction_blocks = [
+                    *self.compaction_blocks[:-1],
+                    {
+                        **self.compaction_blocks[-1],
+                        "content": (streamed_summary if isinstance(streamed_summary, str) else "") + compaction_delta,
+                    },
+                ]
+                provider_specific_fields["compaction_blocks"] = self.compaction_blocks
 
         return (
             text,

litellm/llms/anthropic/common_utils.py

@@ -35,6 +35,7 @@ from litellm.litellm_core_utils.prompt_templates.mid_conversation_system import
     parts_of,
 )
 from litellm.litellm_core_utils.reasoning_effort_utils import non_numeric_thinking_budget_error
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.anthropic.wif import (
     aget_anthropic_wif_token,
     anthropic_base_without_chat_suffix,
@@ -468,7 +469,7 @@ def _fetch_anthropic_models_page(
     # which a read-only one cannot support
     query: Final = f"?after_id={quote(after_id)}" if after_id else ""
     response: Final = litellm.module_level_client.get(
-        url=f"{api_base}/v1/models{query}",
+        url=f"{join_api_base(api_base, 'v1/models')}{query}",
         headers=headers,
         follow_redirects=False,
     )
@@ -1406,7 +1407,7 @@ class AnthropicModelInfo(BaseLLMModelInfo):
         return {**caller_headers, **anthropic_headers}
 
     @staticmethod
-    def get_api_base(api_base: str | None = None) -> str | None:
+    def resolve_anthropic_api_base(api_base: str | None = None) -> str:
         from litellm.secret_managers.main import get_secret_str
 
         return (
@@ -1416,6 +1417,10 @@ class AnthropicModelInfo(BaseLLMModelInfo):
             or "https://api.anthropic.com"
         )
 
+    @staticmethod
+    def get_api_base(api_base: str | None = None) -> str | None:
+        return AnthropicModelInfo.resolve_anthropic_api_base(api_base)
+
     @staticmethod
     def get_api_key(api_key: str | None = None) -> str | None:
         """An empty or whitespace-only key counts as unset: it can never authenticate anything, and
@@ -1533,14 +1538,14 @@ class AnthropicModelInfo(BaseLLMModelInfo):
         api_base: str | None,
         litellm_params: Mapping[str, object] | None,
     ) -> list[str]:  # mutable-ok: matches get_models' list[str] contract shared by every provider override
-        resolved_api_base: Final = AnthropicModelInfo.get_api_base(api_base)
+        resolved_api_base: Final = AnthropicModelInfo.resolve_anthropic_api_base(api_base)
         auth_header: Final = AnthropicModelInfo.get_auth_header(
             api_key,
             resolved_api_base,
             litellm_params=litellm_params,
             allow_workload_identity=config_allows_workload_identity(self),
         )
-        if resolved_api_base is None or auth_header is None:
+        if auth_header is None:
             raise ValueError(
                 "ANTHROPIC_API_BASE/ANTHROPIC_BASE_URL or ANTHROPIC_API_KEY/ANTHROPIC_AUTH_TOKEN (or workload "
                 "identity federation via ANTHROPIC_FEDERATION_RULE_ID/ANTHROPIC_ORGANIZATION_ID/"

litellm/llms/anthropic/count_tokens/transformation.py

@@ -25,6 +25,7 @@ from litellm.litellm_core_utils.prompt_templates.mid_conversation_system import
     anthropic_system_blocks,
     split_leading_system_run,
 )
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.anthropic.common_utils import merge_anthropic_beta_headers
 from litellm.llms.anthropic.wif import resolve_anthropic_base
 from litellm.types.llms.openai import AllMessageValues, ChatCompletionImageObject, ValidUserMessageContentTypes
@@ -317,7 +318,7 @@ class AnthropicCountTokensConfig:
         Returns:
             The endpoint URL for the CountTokens API
         """
-        return resolve_anthropic_base(api_base) + "/v1/messages/count_tokens"
+        return join_api_base(resolve_anthropic_base(api_base), "v1/messages/count_tokens")
 
     def transform_request_to_count_tokens(
         self,

litellm/llms/anthropic/files/handler.py

@@ -10,7 +10,7 @@ import litellm
 from litellm._logging import verbose_logger
 from litellm._uuid import uuid
 from litellm.litellm_core_utils.litellm_logging import Logging
-from litellm.litellm_core_utils.url_utils import encode_url_path_segment
+from litellm.litellm_core_utils.url_utils import encode_url_path_segment, join_api_base
 from litellm.llms.custom_httpx.http_handler import get_async_httpx_client
 from litellm.types.llms.openai import (
     FileContentRequest,
@@ -74,7 +74,7 @@ class AnthropicFilesHandler:
             batch_id = file_id
 
         # Get Anthropic API credentials
-        api_base = self.anthropic_model_info.get_api_base(api_base)
+        api_base = self.anthropic_model_info.resolve_anthropic_api_base(api_base)
         auth_header: Final = await self.anthropic_model_info.aget_auth_header(
             api_key, api_base, litellm_params=litellm_params, allow_workload_identity=True
         )
@@ -84,7 +84,7 @@ class AnthropicFilesHandler:
 
         # Construct the Anthropic batch results URL
         encoded_batch_id: Final = encode_url_path_segment(batch_id, field_name="batch_id")
-        results_url: Final = f"{api_base.rstrip('/')}/v1/messages/batches/{encoded_batch_id}/results"
+        results_url: Final = join_api_base(api_base, f"v1/messages/batches/{encoded_batch_id}/results")
 
         # Prepare headers
         headers: Final = {

litellm/llms/anthropic/files/transformation.py

@@ -22,7 +22,7 @@ from openai.types.file_deleted import FileDeleted
 from pydantic import ConfigDict, TypeAdapter
 
 from litellm.litellm_core_utils.prompt_templates.common_utils import extract_file_data
-from litellm.litellm_core_utils.url_utils import encode_url_path_segment
+from litellm.litellm_core_utils.url_utils import encode_url_path_segment, join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.llms.base_llm.files.transformation import (
     BaseFilesConfig,
@@ -80,7 +80,7 @@ class AnthropicFilesConfig(BaseFilesConfig):
         stream: bool | None = None,
     ) -> str:
         api_base = AnthropicModelInfo.get_api_base(api_base) or ANTHROPIC_FILES_API_BASE
-        return f"{api_base.rstrip('/')}/v1/files"
+        return join_api_base(api_base, "v1/files")
 
     def get_error_class(
         self,
@@ -226,7 +226,7 @@ class AnthropicFilesConfig(BaseFilesConfig):
     ) -> tuple[str, dict]:
         api_base: Final = AnthropicModelInfo.get_api_base(litellm_params.get("api_base")) or ANTHROPIC_FILES_API_BASE
         encoded_file_id: Final = encode_url_path_segment(file_id, field_name="file_id")
-        return f"{api_base.rstrip('/')}/v1/files/{encoded_file_id}", {}
+        return join_api_base(api_base, f"v1/files/{encoded_file_id}"), {}
 
     def transform_retrieve_file_response(
         self,
@@ -245,7 +245,7 @@ class AnthropicFilesConfig(BaseFilesConfig):
     ) -> tuple[str, dict]:
         api_base: Final = AnthropicModelInfo.get_api_base(litellm_params.get("api_base")) or ANTHROPIC_FILES_API_BASE
         encoded_file_id: Final = encode_url_path_segment(file_id, field_name="file_id")
-        return f"{api_base.rstrip('/')}/v1/files/{encoded_file_id}", {}
+        return join_api_base(api_base, f"v1/files/{encoded_file_id}"), {}
 
     def transform_delete_file_response(
         self,
@@ -264,7 +264,7 @@ class AnthropicFilesConfig(BaseFilesConfig):
         litellm_params: dict,
     ) -> tuple[str, dict]:
         api_base: Final = AnthropicModelInfo.get_api_base(litellm_params.get("api_base")) or ANTHROPIC_FILES_API_BASE
-        url: Final = f"{api_base.rstrip('/')}/v1/files"
+        url: Final = join_api_base(api_base, "v1/files")
         params: Final[dict[str, str]] = {}
         if purpose:
             params["purpose"] = purpose
@@ -299,8 +299,8 @@ class AnthropicFilesConfig(BaseFilesConfig):
         api_base: Final = AnthropicModelInfo.get_api_base(litellm_params.get("api_base")) or ANTHROPIC_FILES_API_BASE
         encoded_file_id: Final = encode_url_path_segment(file_id, field_name="file_id")
         if file_id.startswith(ANTHROPIC_MESSAGE_BATCH_ID_PREFIX):
-            return f"{api_base.rstrip('/')}/v1/messages/batches/{encoded_file_id}/results", {}
-        return f"{api_base.rstrip('/')}/v1/files/{encoded_file_id}/content", {}
+            return join_api_base(api_base, f"v1/messages/batches/{encoded_file_id}/results"), {}
+        return join_api_base(api_base, f"v1/files/{encoded_file_id}/content"), {}
 
     def transform_file_content_response(
         self,

litellm/llms/anthropic/pass_through/adapters/streaming_iterator.py

@@ -16,6 +16,7 @@ from typing import (
     Protocol,
     cast,
     get_args,
+    runtime_checkable,
 )
 
 from pydantic import TypeAdapter
@@ -65,6 +66,14 @@ class _ChunkStream(Protocol):
     def __aiter__(self) -> "AsyncIterator[ModelResponseStream]": ...
 
 
+@runtime_checkable
+class _FailureLoggingWaiter(Protocol):
+    @property
+    def has_dispatched_failure_logging(self) -> bool: ...
+
+    async def wait_for_failure_logging(self, timeout: float | None) -> None: ...
+
+
 def _optional_attr(obj: object, name: str) -> object:
     return getattr(obj, name, None)
 
@@ -189,6 +198,14 @@ class _CombinedChunkSplitter:
         self._held_calls: Mapping[int, tuple[ModelResponseStream, _ToolCallPart]] = MappingProxyType({})
         self._live_call_index: int | None = None
 
+    @property
+    def has_dispatched_failure_logging(self) -> bool:
+        return isinstance(self._stream, _FailureLoggingWaiter) and self._stream.has_dispatched_failure_logging is True
+
+    async def wait_for_failure_logging(self, timeout: float | None) -> None:
+        if isinstance(self._stream, _FailureLoggingWaiter):
+            await self._stream.wait_for_failure_logging(timeout=timeout)
+
     @property
     def chunks(self) -> "list[ModelResponseStream] | None":
         return cast(  # cast-ok: chunks is a list of ModelResponseStream on the inner stream
@@ -522,7 +539,9 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
         # Wrap the upstream stream so chunks that carry both content and a
         # finish_reason (fake-streamed providers) are split into two — see
         # _CombinedChunkSplitter.
-        super().__init__(_CombinedChunkSplitter(completion_stream))
+        combined_chunk_splitter: Final = _CombinedChunkSplitter(completion_stream)
+        super().__init__(combined_chunk_splitter)
+        self._combined_chunk_splitter: Final = combined_chunk_splitter
         self.model = model
         self._message_id: str = f"msg_{uuid.uuid4()}"
         if litellm_logging_obj is not None:
@@ -1244,7 +1263,10 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
                 if provider_error is e:
                     raise
                 raise provider_error from e
-            if logging_obj is not None:
+            stream_has_dispatched_failure_logging: Final = self._combined_chunk_splitter.has_dispatched_failure_logging
+            if stream_has_dispatched_failure_logging:
+                await self._combined_chunk_splitter.wait_for_failure_logging(timeout=None)
+            elif logging_obj is not None:
                 try:
                     await logging_obj.dispatch_failure_handlers(
                         exception=provider_error,

litellm/llms/anthropic/pass_through/messages/transformation.py

@@ -9,6 +9,7 @@ from litellm.anthropic_beta_headers_manager import get_unsupported_headers, is_b
 from litellm.exceptions import AuthenticationError
 from litellm.litellm_core_utils.core_helpers import drops_unsupported_params
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+from litellm.litellm_core_utils.url_utils import ensure_api_route
 from litellm.llms.anthropic.pass_through.convertible_blocks import drop_non_string_tool_result_texts
 from litellm.llms.anthropic.pass_through.safeguards import proxy_serves_safeguards
 from litellm.llms.base_llm.anthropic_messages.transformation import (
@@ -287,10 +288,7 @@ class AnthropicMessagesConfig(BaseAnthropicMessagesConfig):
         litellm_params: dict,
         stream: bool | None = None,
     ) -> str:
-        api_base = AnthropicModelInfo.get_api_base(api_base) or "https://api.anthropic.com"
-        if not api_base.endswith("/v1/messages"):
-            api_base = f"{api_base}/v1/messages"
-        return api_base
+        return ensure_api_route(AnthropicModelInfo.get_api_base(api_base) or "https://api.anthropic.com", "v1/messages")
 
     def validate_anthropic_messages_environment(
         self,

litellm/llms/anthropic/skills/transformation.py

@@ -9,7 +9,7 @@ import httpx
 from pydantic import TypeAdapter
 
 from litellm._logging import verbose_logger
-from litellm.litellm_core_utils.url_utils import encode_url_path_segment
+from litellm.litellm_core_utils.url_utils import encode_url_path_segment, join_api_base
 from litellm.llms.base_llm.skills.transformation import (
     BaseSkillsAPIConfig,
     LiteLLMLoggingObj,
@@ -75,13 +75,11 @@ class AnthropicSkillsConfig(BaseSkillsAPIConfig):
         """Get complete URL for Anthropic Skills API"""
         from litellm.llms.anthropic.common_utils import AnthropicModelInfo
 
-        if api_base is None:
-            api_base = AnthropicModelInfo.get_api_base()
-
+        resolved_api_base: Final = AnthropicModelInfo.get_api_base(api_base) or "https://api.anthropic.com"
         if skill_id:
             encoded_skill_id: Final = encode_url_path_segment(skill_id, field_name="skill_id")
-            return f"{api_base}/v1/skills/{encoded_skill_id}"
-        return f"{api_base}/v1/{endpoint}"
+            return join_api_base(resolved_api_base, f"v1/skills/{encoded_skill_id}")
+        return join_api_base(resolved_api_base, f"v1/{endpoint}")
 
     def transform_create_skill_request(
         self,

litellm/llms/anthropic/wif.py

@@ -15,6 +15,7 @@ from typing_extensions import assert_never
 
 import litellm
 from litellm._logging import verbose_logger
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.auth.client_credentials import keycloak_assertion_source
 from litellm.llms.base_llm.auth.identity_source import (
     AnthropicIdentitySourceKind,
@@ -324,7 +325,7 @@ def anthropic_internal_issuer_jwks(credential_values: Mapping[str, object]) -> A
 
 def build_anthropic_wif_spec(params: AnthropicWifParams, api_base: str) -> TokenExchangeSpec:
     return TokenExchangeSpec(
-        token_url=api_base.rstrip("/") + ANTHROPIC_TOKEN_EXCHANGE_PATH,
+        token_url=join_api_base(api_base, ANTHROPIC_TOKEN_EXCHANGE_PATH),
         assertion_ref=params.assertion_ref,
         assertion_field="assertion",
         static_body=MappingProxyType(

litellm/llms/apiserpent/search/transformation.py

@@ -14,6 +14,7 @@ from urllib.parse import urlencode
 import httpx
 
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+from litellm.litellm_core_utils.url_utils import ensure_api_route
 from litellm.llms.apiserpent.search.defaults import (
     DEEP_SEARCH_PATH,
     NUM_MAX,
@@ -77,13 +78,11 @@ class APISerpentSearchConfig(BaseSearchConfig):
         Build the search URL. APISerpent uses GET, so the transformed request is
         serialized into the query string. The endpoint path (quick vs deep) is
         always applied; an ``api_base`` / ``APISERPENT_API_BASE`` override only
-        changes the host. The ``endswith`` guard keeps this idempotent, since the
+        changes the host. ``ensure_api_route`` keeps this idempotent, since the
         handler re-invokes this method with the already-resolved URL as api_base.
         """
-        base = (api_base or get_secret_str("APISERPENT_API_BASE") or APISERPENT_BASE).rstrip("/")
         path: Final = DEEP_SEARCH_PATH if self._is_deep_search(optional_params) else QUICK_SEARCH_PATH
-        if not base.endswith(path):
-            base = f"{base}{path}"
+        base: Final = ensure_api_route(api_base or get_secret_str("APISERPENT_API_BASE") or APISERPENT_BASE, path)
 
         if data and isinstance(data, dict) and APISERPENT_PARAMS_KEY in data:
             query_string: Final = urlencode(data[APISERPENT_PARAMS_KEY], doseq=True)

litellm/llms/azure/audio_transcription/transformation.py

@@ -13,6 +13,7 @@ import httpx
 from pydantic import ConfigDict, TypeAdapter
 
 from litellm.litellm_core_utils.audio_utils.utils import process_audio_file
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.audio_transcription.transformation import (
     AudioTranscriptionRequestData,
     BaseAudioTranscriptionConfig,
@@ -110,7 +111,7 @@ class AzureSpeechAudioTranscriptionConfig(BaseAudioTranscriptionConfig):
             "language": optional_params.get("language", self.DEFAULT_LANGUAGE),
             "format": self._get_azure_response_format(optional_params.get("response_format")),
         }
-        return f"{base_url}{self.STT_ENDPOINT_PATH}?{urlencode(query_params)}"
+        return f"{join_api_base(base_url, self.STT_ENDPOINT_PATH)}?{urlencode(query_params)}"
 
     def transform_audio_transcription_request(
         self,
@@ -153,7 +154,6 @@ class AzureSpeechAudioTranscriptionConfig(BaseAudioTranscriptionConfig):
         )
 
     def _resolve_stt_base_url(self, api_base: str) -> str:
-        api_base = api_base.rstrip("/")
         parsed_url: Final = urlparse(api_base)
         hostname: Final = parsed_url.hostname or ""
 

litellm/llms/azure/common_utils.py

@@ -18,6 +18,7 @@ import litellm
 from litellm._logging import verbose_logger
 from litellm.caching.caching import DualCache
 from litellm.constants import DEFAULT_MAX_RETRIES
+from litellm.litellm_core_utils.url_utils import ensure_api_route
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.llms.custom_httpx.http_handler import HTTPHandler
 from litellm.llms.openai.common_utils import BaseOpenAILLM, OpenAIAsyncHTTPClient, OpenAIHTTPClient
@@ -616,7 +617,7 @@ class BaseAzureLLM(BaseOpenAILLM):
 
             v1_params: Final[_AzureV1ClientParams] = {
                 "api_key": v1_api_key,
-                "base_url": f"{api_base}/openai/v1/",
+                "base_url": f"{ensure_api_route(api_base or '', 'openai/v1')}/",
             }
             if "timeout" in azure_client_params:
                 v1_params["timeout"] = azure_client_params["timeout"]

litellm/llms/azure/realtime/handler.py

@@ -10,6 +10,7 @@ from typing import Final, Protocol, cast
 
 from litellm._logging import verbose_proxy_logger
 from litellm.constants import REALTIME_WEBSOCKET_MAX_MESSAGE_SIZE_BYTES
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.types.realtime import RealtimeQueryParams
 
 from ....litellm_core_utils.litellm_logging import Logging as LiteLLMLogging
@@ -136,7 +137,8 @@ class AzureOpenAIRealtime(AzureChatCompletion):
             query_parts.append(urlencode({"intent": intent}))
 
         qs: Final = "&".join(query_parts)
-        return f"{api_base}{path}?{qs}" if qs else f"{api_base}{path}"
+        endpoint_url: Final = join_api_base(api_base, path)
+        return f"{endpoint_url}?{qs}" if qs else endpoint_url
 
     def construct_url(
         self,

litellm/llms/azure/realtime/http_transformation.py

@@ -3,6 +3,7 @@
 from typing import Final
 
 import litellm
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.realtime.http_transformation import BaseRealtimeHTTPConfig
 from litellm.secret_managers.main import get_secret_str
 
@@ -39,9 +40,9 @@ class AzureRealtimeHTTPConfig(BaseRealtimeHTTPConfig):
         return _azure_realtime_v1_url(self.get_api_base(api_base), "calls")
 
     def get_transcription_session_url(self, api_base: str | None, model: str, api_version: str | None = None) -> str:
-        base: Final = self.get_api_base(api_base).rstrip("/")
+        base: Final = self.get_api_base(api_base)
         version: Final = api_version or get_secret_str("AZURE_API_VERSION") or "2024-12-17"
-        return f"{base}/openai/realtime/transcription_sessions?api-version={version}"
+        return f"{join_api_base(base, 'openai/realtime/transcription_sessions')}?api-version={version}"
 
     def get_realtime_calls_headers(self, ephemeral_key: str) -> dict[str, str]:
         return {

litellm/llms/azure/search/transformation.py

@@ -35,6 +35,7 @@ from typing import TYPE_CHECKING, Final, Literal
 import httpx
 from pydantic import ConfigDict, ValidationError
 
+from litellm.litellm_core_utils.url_utils import ensure_api_route
 from litellm.llms.azure.common_utils import cached_azure_ad_token_refresh_provider
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.llms.base_llm.search.transformation import (
@@ -340,10 +341,7 @@ class BingGroundingSearchConfig(BaseSearchConfig):
                 f"{PROJECT_ENDPOINT_ENV} is not set. Set it to your Microsoft Foundry project "
                 f"endpoint, e.g. https://<account>.services.ai.azure.com/api/projects/<project>."
             )
-        trimmed: Final = resolved_base.rstrip("/")
-        if trimmed.endswith(_RESPONSES_PATH):
-            return trimmed
-        return f"{trimmed}{_RESPONSES_PATH}"
+        return ensure_api_route(resolved_base, _RESPONSES_PATH)
 
     def transform_search_request(
         self,

litellm/llms/azure/text_to_speech/transformation.py

@@ -11,6 +11,7 @@ from urllib.parse import urlparse
 import httpx
 
 import litellm
+from litellm.litellm_core_utils.url_utils import ensure_api_route, join_api_base
 from litellm.llms.base_llm.text_to_speech.transformation import (
     BaseTextToSpeechConfig,
     TextToSpeechRequestData,
@@ -329,8 +330,6 @@ class AzureAVATextToSpeechConfig(BaseTextToSpeechConfig):
                 f"https://{{region}}.{self.TTS_SPEECH_DOMAIN}"
             )
 
-        # Remove trailing slash and parse URL
-        api_base = api_base.rstrip("/")
         parsed_url: Final = urlparse(api_base)
         hostname: Final = parsed_url.hostname or ""
 
@@ -339,14 +338,10 @@ class AzureAVATextToSpeechConfig(BaseTextToSpeechConfig):
             region: Final = self._extract_region_from_hostname(hostname=hostname, domain=self.COGNITIVE_SERVICES_DOMAIN)
             return self._build_tts_url(region=region)
 
-        # Check if it's already a TTS endpoint
         if self._is_tts_endpoint(hostname=hostname):
-            if not api_base.endswith(self.TTS_ENDPOINT_PATH):
-                return f"{api_base}{self.TTS_ENDPOINT_PATH}"
-            return api_base
+            return ensure_api_route(api_base, self.TTS_ENDPOINT_PATH)
 
-        # Assume it's a custom endpoint, append the path
-        return f"{api_base}{self.TTS_ENDPOINT_PATH}"
+        return join_api_base(api_base, self.TTS_ENDPOINT_PATH)
 
     def _is_cognitive_services_endpoint(self, hostname: str) -> bool:
         """Check if hostname is a Cognitive Services endpoint"""

litellm/llms/azure_ai/agents/handler.py

@@ -30,7 +30,7 @@ from typing_extensions import ReadOnly
 
 from litellm._logging import verbose_logger
 from litellm.litellm_core_utils.hidden_params import set_hidden_param
-from litellm.litellm_core_utils.url_utils import encode_url_path_segment
+from litellm.litellm_core_utils.url_utils import encode_url_path_segment, join_api_base
 from litellm.llms.azure_ai.agents.transformation import (
     AzureAIAgentsConfig,
     AzureAIAgentsError,
@@ -130,28 +130,29 @@ class AzureAIAgentsHandler:
     # See: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/quickstart
     # -------------------------------------------------------------------------
     def _build_thread_url(self, api_base: str, api_version: str) -> str:
-        return f"{api_base}/threads?api-version={api_version}"
+        return f"{join_api_base(api_base, 'threads')}?api-version={api_version}"
 
     def _build_messages_url(self, api_base: str, thread_id: str, api_version: str) -> str:
         encoded_thread_id: Final = encode_url_path_segment(thread_id, field_name="thread_id")
-        return f"{api_base}/threads/{encoded_thread_id}/messages?api-version={api_version}"
+        return f"{join_api_base(api_base, f'threads/{encoded_thread_id}/messages')}?api-version={api_version}"
 
     def _build_runs_url(self, api_base: str, thread_id: str, api_version: str) -> str:
         encoded_thread_id: Final = encode_url_path_segment(thread_id, field_name="thread_id")
-        return f"{api_base}/threads/{encoded_thread_id}/runs?api-version={api_version}"
+        return f"{join_api_base(api_base, f'threads/{encoded_thread_id}/runs')}?api-version={api_version}"
 
     def _build_run_status_url(self, api_base: str, thread_id: str, run_id: str, api_version: str) -> str:
         encoded_thread_id: Final = encode_url_path_segment(thread_id, field_name="thread_id")
         encoded_run_id: Final = encode_url_path_segment(run_id, field_name="run_id")
-        return f"{api_base}/threads/{encoded_thread_id}/runs/{encoded_run_id}?api-version={api_version}"
+        run_status_route: Final = f"threads/{encoded_thread_id}/runs/{encoded_run_id}"
+        return f"{join_api_base(api_base, run_status_route)}?api-version={api_version}"
 
     def _build_list_messages_url(self, api_base: str, thread_id: str, api_version: str) -> str:
         encoded_thread_id: Final = encode_url_path_segment(thread_id, field_name="thread_id")
-        return f"{api_base}/threads/{encoded_thread_id}/messages?api-version={api_version}"
+        return f"{join_api_base(api_base, f'threads/{encoded_thread_id}/messages')}?api-version={api_version}"
 
     def _build_create_thread_and_run_url(self, api_base: str, api_version: str) -> str:
         """URL for the create-thread-and-run endpoint (supports streaming)."""
-        return f"{api_base}/threads/runs?api-version={api_version}"
+        return f"{join_api_base(api_base, 'threads/runs')}?api-version={api_version}"
 
     # -------------------------------------------------------------------------
     # Response Helpers
@@ -281,8 +282,6 @@ class AzureAIAgentsHandler:
         api_version: Final = optional_params.get("api_version", self.config.DEFAULT_API_VERSION)
         agent_id: Final = self.config.get_agent_id(model, optional_params)
         thread_id: Final = optional_params.get("thread_id")
-        api_base = api_base.rstrip("/")
-
         verbose_logger.debug("Azure AI Agents completion - api_base: %s, agent_id: %s", api_base, agent_id)
 
         return headers, api_version, agent_id, thread_id, api_base

litellm/llms/azure_ai/anthropic/count_tokens/transformation.py

@@ -5,8 +5,10 @@ Extends the base Anthropic CountTokens transformation with Azure authentication.
 """
 
 from typing import Final
+from urllib.parse import urlsplit
 
 from litellm.constants import ANTHROPIC_TOKEN_COUNTING_BETA_VERSION
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.anthropic.count_tokens.transformation import (
     AnthropicCountTokensConfig,
 )
@@ -76,14 +78,8 @@ class AzureAIAnthropicCountTokensConfig(AnthropicCountTokensConfig):
         Returns:
             The endpoint URL for the CountTokens API
         """
-        # Azure AI Anthropic endpoint format:
-        # https://<resource>.services.ai.azure.com/anthropic/v1/messages/count_tokens
-        api_base = api_base.rstrip("/")
-
-        # Ensure the URL has /anthropic path
-        if not api_base.endswith("/anthropic"):
-            if "/anthropic" not in api_base:
-                api_base = f"{api_base}/anthropic"
-
-        # Add the count_tokens path
-        return f"{api_base}/v1/messages/count_tokens"
+        base: Final = api_base.rstrip("/")
+        if base.endswith("/v1/messages"):
+            return join_api_base(base, "count_tokens")
+        anthropic_base: Final = base if "/anthropic" in urlsplit(base).path else join_api_base(base, "anthropic")
+        return join_api_base(anthropic_base, "v1/messages/count_tokens")

litellm/llms/azure_ai/anthropic/messages_transformation.py

@@ -9,6 +9,7 @@ from litellm.llms.anthropic.pass_through.messages.transformation import (
     AnthropicMessagesConfig,
 )
 from litellm.llms.azure.common_utils import BaseAzureLLM
+from litellm.llms.azure_ai.anthropic.transformation import azure_anthropic_messages_url
 from litellm.types.router import GenericLiteLLMParams
 
 
@@ -96,25 +97,7 @@ class AzureAnthropicMessagesConfig(AnthropicMessagesConfig):
                 "Expected format: https://<resource-name>.services.ai.azure.com/anthropic"
             )
 
-        # Ensure the URL ends with /v1/messages
-        api_base = api_base.rstrip("/")
-        if api_base.endswith("/v1/messages") or api_base.endswith("/anthropic/v1/messages"):
-            # Already correct
-            pass
-        else:
-            # Check if /anthropic is already in the path
-            if "/anthropic" in api_base:
-                # /anthropic exists, ensure we end with /anthropic/v1/messages
-                # Extract the base URL up to and including /anthropic
-                parts: Final = api_base.split("/anthropic", 1)
-                api_base = parts[0] + "/anthropic"
-            else:
-                # /anthropic not in path, add it
-                api_base = api_base + "/anthropic"
-            # Add /v1/messages
-            api_base = api_base + "/v1/messages"
-
-        return api_base
+        return azure_anthropic_messages_url(api_base)
 
     def _remove_scope_from_cache_control(self, anthropic_messages_request: dict) -> None:
         """

litellm/llms/azure_ai/anthropic/transformation.py

@@ -4,12 +4,23 @@ Azure Anthropic transformation config - extends AnthropicConfig with Azure authe
 
 from typing import Final
 
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.anthropic.chat.transformation import AnthropicConfig
 from litellm.llms.azure.common_utils import BaseAzureLLM
 from litellm.types.llms.openai import AllMessageValues
 from litellm.types.router import GenericLiteLLMParams
 
 
+def azure_anthropic_messages_url(api_base: str) -> str:
+    base: Final = api_base.rstrip("/")
+    if base.endswith("/v1/messages"):
+        return base
+    path_start: Final = base.find("/", base.find("://") + len("://")) if "://" in base else 0
+    anthropic_start: Final = base.find("/anthropic", path_start) if path_start >= 0 else -1
+    anthropic_root: Final = base[:anthropic_start] if anthropic_start >= 0 else base
+    return join_api_base(anthropic_root, "anthropic/v1/messages")
+
+
 def _promote_extra_body_to_optional_params(optional_params: dict) -> None:
     """Promote anthropic-native passthrough keys out of ``extra_body``.
 

litellm/llms/azure_ai/embed/handler.py

@@ -5,6 +5,7 @@ from urllib.parse import urlsplit, urlunsplit
 from openai import OpenAI
 
 import litellm
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.azure_ai.common_utils import is_foundry_model_inference_base
 from litellm.llms.custom_httpx.http_handler import (
     AsyncHTTPHandler,
@@ -74,13 +75,18 @@ class AzureAIEmbedding(OpenAIChatCompletion):
         api_base: str | None,
         client: HTTPHandler | AsyncHTTPHandler | None = None,
     ) -> EmbeddingResponse:
+        if api_base is None:
+            raise ValueError(
+                "api_base is None. Please set AZURE_AI_API_BASE or dynamically via `api_base` param, "
+                "to make the request."
+            )
         if client is None or not isinstance(client, AsyncHTTPHandler):
             client = get_async_httpx_client(
                 llm_provider=litellm.LlmProviders.AZURE_AI,
                 params={"timeout": timeout},
             )
 
-        url: Final = f"{api_base}/images/embeddings"
+        url: Final = join_api_base(api_base, "images/embeddings")
 
         response: Final = await client.post(
             url=url,
@@ -123,7 +129,7 @@ class AzureAIEmbedding(OpenAIChatCompletion):
         if client is None or not isinstance(client, HTTPHandler):
             client = HTTPHandler(timeout=timeout, concurrent_limit=1)
 
-        url: Final = f"{api_base}/images/embeddings"
+        url: Final = join_api_base(api_base, "images/embeddings")
 
         response: Final = client.post(
             url=url,

litellm/llms/azure_ai/image_generation/flux_transformation.py

@@ -3,6 +3,7 @@ from types import MappingProxyType
 from typing import Final
 
 from litellm.exceptions import BadRequestError, UnsupportedParamsError
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.openai.image_generation import GPTImageGenerationConfig
 from litellm.types.llms.openai import OpenAIImageGenerationOptionalParams
 
@@ -52,7 +53,8 @@ class AzureFoundryFluxImageGenerationConfig(GPTImageGenerationConfig):
             return f"{api_base}?api-version={api_version}"
 
         provider_model_path: Final = AzureFoundryFluxImageGenerationConfig.get_flux2_provider_model_path(model)
-        return f"{api_base}/providers/blackforestlabs/v1/{provider_model_path}?api-version={api_version}"
+        flux_route: Final = f"providers/blackforestlabs/v1/{provider_model_path}"
+        return f"{join_api_base(api_base, flux_route)}?api-version={api_version}"
 
     @staticmethod
     def is_flux2_model(model: str) -> bool:

litellm/llms/azure_ai/image_generation/mai_transformation.py

@@ -3,6 +3,7 @@ from typing import TYPE_CHECKING, Any, Final
 import httpx
 
 from litellm.exceptions import UnsupportedParamsError
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.image_generation.transformation import (
     BaseImageGenerationConfig,
 )
@@ -41,9 +42,9 @@ class AzureFoundryMAIImageGenerationConfig(BaseImageGenerationConfig):
 
         if "/mai/" in path:
             prefix, _, _ = path.partition("/images/")
-            path = f"{prefix}/images/generations"
+            path = join_api_base(prefix, "images/generations")
         else:
-            path = f"{path}/mai/v1/images/generations"
+            path = join_api_base(path, "mai/v1/images/generations")
 
         if separator:
             return f"{path}?{query}"
@@ -63,9 +64,9 @@ class AzureFoundryMAIImageGenerationConfig(BaseImageGenerationConfig):
 
         if "/mai/" in path:
             prefix, _, _ = path.partition("/images/")
-            path = f"{prefix}/images/edits"
+            path = join_api_base(prefix, "images/edits")
         else:
-            path = f"{path}/mai/v1/images/edits"
+            path = join_api_base(path, "mai/v1/images/edits")
 
         if separator:
             return f"{path}?{query}"

litellm/llms/azure_ai/vector_stores/transformation.py

@@ -5,6 +5,7 @@ from typing import TYPE_CHECKING, Any, Final
 
 import httpx
 
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.azure.common_utils import BaseAzureLLM
 from litellm.llms.base_llm.vector_store.transformation import (
     BaseQueryEmbeddingVectorStoreConfig,
@@ -172,7 +173,8 @@ class AzureAIVectorStoreConfig(BaseQueryEmbeddingVectorStoreConfig, BaseAzureLLM
         litellm_logging_obj.model_call_details["input"] = query_text
         litellm_logging_obj.model_call_details["embedding_model"] = litellm_params.get("litellm_embedding_model")
         litellm_logging_obj.model_call_details["top_k"] = top_k
-        return f"{api_base}/indexes/{vector_store_id}/docs/search?api-version=2024-07-01", {
+        search_route: Final = f"indexes/{vector_store_id}/docs/search"
+        return f"{join_api_base(api_base, search_route)}?api-version=2024-07-01", {
             "search": "*",
             "vectorQueries": [{"vector": query_vector, "fields": vector_field, "kind": "vector", "k": top_k}],
             "select": "id,content",

litellm/llms/base_llm/decisions/transformation.py

@@ -12,6 +12,7 @@ from pydantic_core import ErrorDetails
 from typing_extensions import assert_never
 
 import litellm
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.secret_managers.main import get_secret_str
 from litellm.types.decisions import (
@@ -582,7 +583,7 @@ class BaseDecisionsConfig(ABC):
         )
 
     def get_complete_url(self, api_base: str, model: str) -> str:
-        return f"{api_base.rstrip('/').removesuffix('/v1')}{self.path}"
+        return join_api_base(api_base.rstrip("/").removesuffix("/v1"), self.path)
 
     def signs_with_aws(self, api_base: str) -> bool:
         return False

litellm/llms/base_llm/evals/transformation.py

@@ -7,6 +7,7 @@ from typing import TYPE_CHECKING, Any
 
 import httpx
 
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.types.llms.openai_evals import (
     CancelEvalResponse,
@@ -79,7 +80,7 @@ class BaseEvalsAPIConfig(ABC):
         """
         if api_base is None:
             raise ValueError("api_base is required")
-        return f"{api_base}/v1/{endpoint}"
+        return join_api_base(api_base, f"v1/{endpoint}")
 
     @abstractmethod
     def transform_create_eval_request(

litellm/llms/base_llm/realtime/http_transformation.py

@@ -7,10 +7,11 @@ These are HTTP (not WebSocket) endpoints used by the WebRTC flow:
 """
 
 from abc import ABC, abstractmethod
-from typing import Final
 
 import httpx
 
+from litellm.litellm_core_utils.url_utils import join_api_base
+
 
 class BaseRealtimeHTTPConfig(ABC):
     """
@@ -61,7 +62,7 @@ class BaseRealtimeHTTPConfig(ABC):
         """Return the full URL for POST /realtime/transcription_sessions."""
         base = (api_base or "").rstrip("/")
         base = base.removesuffix("/v1")
-        return f"{base}/v1/realtime/transcription_sessions"
+        return join_api_base(base, "v1/realtime/transcription_sessions")
 
     @abstractmethod
     def validate_environment(
@@ -83,8 +84,7 @@ class BaseRealtimeHTTPConfig(ABC):
 
     def get_realtime_calls_url(self, api_base: str | None, model: str, api_version: str | None = None) -> str:
         """Return the full URL for POST /realtime/calls (SDP exchange)."""
-        base: Final = (api_base or "").rstrip("/")
-        return f"{base}/v1/realtime/calls"
+        return join_api_base(api_base or "", "v1/realtime/calls")
 
     def get_realtime_calls_headers(self, ephemeral_key: str) -> dict:
         """

litellm/llms/base_llm/skills/transformation.py

@@ -7,6 +7,7 @@ from typing import TYPE_CHECKING, Any
 
 import httpx
 
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.types.llms.anthropic_skills import (
     CreateSkillRequest,
@@ -71,7 +72,7 @@ class BaseSkillsAPIConfig(ABC):
         """
         if api_base is None:
             raise ValueError("api_base is required")
-        return f"{api_base}/v1/{endpoint}"
+        return join_api_base(api_base, f"v1/{endpoint}")
 
     @abstractmethod
     def transform_create_skill_request(

litellm/llms/bedrock/chat/agentcore/transformation.py

@@ -18,6 +18,7 @@ from litellm.litellm_core_utils.prompt_templates.common_utils import (
     convert_content_list_to_str,
 )
 from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.a2a.common_utils import extract_text_from_a2a_response
 from litellm.llms.base_llm.chat.transformation import BaseConfig, BaseLLMException
 from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
@@ -98,14 +99,14 @@ class AmazonAgentCoreConfig(BaseConfig, BaseAWSLLM):
         # Build the base endpoint URL for AgentCore
         # Note: We don't use get_runtime_endpoint as AgentCore has its own endpoint structure
         if aws_bedrock_runtime_endpoint:
-            base_url = aws_bedrock_runtime_endpoint
+            base_url = f"{aws_bedrock_runtime_endpoint}"
         else:
             base_url = f"https://bedrock-agentcore.{region}.{get_aws_dns_suffix(region)}"
 
         # Based on boto3 client.invoke_agent_runtime, the path is:
         # /runtimes/{URL-ENCODED-ARN}/invocations?qualifier=<value>
         encoded_arn: Final = quote(agent_runtime_arn, safe="")
-        endpoint_url = f"{base_url}/runtimes/{encoded_arn}/invocations"
+        endpoint_url = join_api_base(base_url, f"runtimes/{encoded_arn}/invocations")
 
         # Add qualifier as query parameter if provided
         if "qualifier" in optional_params:

litellm/llms/bedrock/chat/chat_completions/transformation.py

@@ -33,6 +33,7 @@ from litellm.litellm_core_utils.prompt_templates.image_handling import (
     inline_remote_image_urls,
     inline_remote_media,
 )
+from litellm.litellm_core_utils.url_utils import ensure_api_route, join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM, bedrock_bearer_token
 from litellm.llms.bedrock.common_utils import (
@@ -371,12 +372,9 @@ class AmazonBedrockRuntimeChatCompletionsConfig(OpenAILikeChatConfig):
             ),
             aws_region_name=aws_region_name,
         )
-        base: Final = proxy_endpoint_url.rstrip("/")
-        if base.endswith("/openai/v1/chat/completions"):
-            return base
-        if base.endswith("/openai/v1"):
-            return f"{base}/chat/completions"
-        return f"{base}/openai/v1/chat/completions"
+        if proxy_endpoint_url.rstrip("/").endswith("/openai/v1"):
+            return join_api_base(proxy_endpoint_url, "chat/completions")
+        return ensure_api_route(proxy_endpoint_url, "openai/v1/chat/completions")
 
     def _params_with_region_from_path(
         self, optional_params: dict, model: str | None

litellm/llms/bedrock/chat/converse_handler.py

@@ -10,6 +10,7 @@ from litellm.anthropic_beta_headers_manager import (
 )
 from litellm.litellm_core_utils.get_litellm_params import stored_control_options
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObject
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm import base_model_iterator
 from litellm.llms.custom_httpx.http_handler import (
     AsyncHTTPHandler,
@@ -363,11 +364,11 @@ class BedrockConverseLLM(BaseAWSLLM):
             aws_region_name=aws_region_name,
         )
         if (stream is not None and stream is True) and not fake_stream:
-            endpoint_url = f"{endpoint_url}/model/{modelId}/converse-stream"
-            proxy_endpoint_url = f"{proxy_endpoint_url}/model/{modelId}/converse-stream"
+            endpoint_url = join_api_base(endpoint_url, f"model/{modelId}/converse-stream")
+            proxy_endpoint_url = join_api_base(proxy_endpoint_url, f"model/{modelId}/converse-stream")
         else:
-            endpoint_url = f"{endpoint_url}/model/{modelId}/converse"
-            proxy_endpoint_url = f"{proxy_endpoint_url}/model/{modelId}/converse"
+            endpoint_url = join_api_base(endpoint_url, f"model/{modelId}/converse")
+            proxy_endpoint_url = join_api_base(proxy_endpoint_url, f"model/{modelId}/converse")
 
         ## COMPLETION CALL
         headers = {"Content-Type": "application/json"}

litellm/llms/bedrock/chat/invoke_agent/transformation.py

@@ -15,7 +15,7 @@ from litellm._uuid import uuid
 from litellm.litellm_core_utils.prompt_templates.common_utils import (
     convert_content_list_to_str,
 )
-from litellm.litellm_core_utils.url_utils import encode_url_path_segment
+from litellm.litellm_core_utils.url_utils import encode_url_path_segment, join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseConfig, BaseLLMException
 from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
 from litellm.llms.bedrock.common_utils import BedrockError
@@ -101,9 +101,10 @@ class AmazonInvokeAgentConfig(BaseConfig, BaseAWSLLM):
         encoded_agent_alias_id: Final = encode_url_path_segment(agent_alias_id, field_name="agent_alias_id")
         encoded_session_id: Final = encode_url_path_segment(session_id, field_name="session_id")
 
-        endpoint_url = f"{endpoint_url}/agents/{encoded_agent_id}/agentAliases/{encoded_agent_alias_id}/sessions/{encoded_session_id}/text"
-
-        return endpoint_url
+        return join_api_base(
+            endpoint_url,
+            f"agents/{encoded_agent_id}/agentAliases/{encoded_agent_alias_id}/sessions/{encoded_session_id}/text",
+        )
 
     def sign_request(
         self,

litellm/llms/bedrock/chat/invoke_transformations/amazon_openai_transformation.py

@@ -11,6 +11,7 @@ from typing import TYPE_CHECKING, Any, Final
 
 import httpx
 
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
 from litellm.llms.bedrock.common_utils import BedrockError
 from litellm.llms.bedrock.request_metadata import (
@@ -97,13 +98,8 @@ class AmazonBedrockOpenAIConfig(OpenAIGPTConfig, BaseAWSLLM):
         # Encode model ID for ARNs (e.g., :imported-model/ -> :imported-model%2F)
         model_id = CommonUtils.encode_bedrock_runtime_modelid_arn(model_id)
 
-        # Build the invoke URL
-        if stream:
-            endpoint_url = f"{endpoint_url}/model/{model_id}/invoke-with-response-stream"
-        else:
-            endpoint_url = f"{endpoint_url}/model/{model_id}/invoke"
-
-        return endpoint_url
+        invoke_route: Final = "invoke-with-response-stream" if stream else "invoke"
+        return join_api_base(endpoint_url, f"model/{model_id}/{invoke_route}")
 
     def sign_request(
         self,

litellm/llms/bedrock/chat/invoke_transformations/base_invoke_transformation.py

@@ -18,6 +18,7 @@ from litellm.litellm_core_utils.prompt_templates.factory import (
     prompt_factory,
     render_custom_prompt_template,
 )
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseConfig, BaseLLMException
 from litellm.llms.bedrock.chat.invoke_handler import make_call, make_sync_call
 from litellm.llms.bedrock.common_utils import BedrockError
@@ -136,20 +137,13 @@ class AmazonInvokeConfig(BaseConfig, BaseAWSLLM):
         aws_bedrock_runtime_endpoint: Final = optional_params.get(
             "aws_bedrock_runtime_endpoint", None
         )  # https://bedrock-runtime.{region_name}.amazonaws.com
-        endpoint_url, proxy_endpoint_url = self.get_runtime_endpoint(
+        endpoint_url, _ = self.get_runtime_endpoint(
             api_base=api_base,
             aws_bedrock_runtime_endpoint=aws_bedrock_runtime_endpoint,
             aws_region_name=self._get_aws_region_name(optional_params=optional_params, model=model),
         )
-
-        if (stream is not None and stream is True) and provider != "ai21":
-            endpoint_url = f"{endpoint_url}/model/{modelId}/invoke-with-response-stream"
-            proxy_endpoint_url = f"{proxy_endpoint_url}/model/{modelId}/invoke-with-response-stream"
-        else:
-            endpoint_url = f"{endpoint_url}/model/{modelId}/invoke"
-            proxy_endpoint_url = f"{proxy_endpoint_url}/model/{modelId}/invoke"
-
-        return endpoint_url
+        invoke_route: Final = "invoke-with-response-stream" if stream is True and provider != "ai21" else "invoke"
+        return join_api_base(endpoint_url, f"model/{modelId}/{invoke_route}")
 
     def sign_request(
         self,

litellm/llms/bedrock/claude_platform/common_utils.py

@@ -5,6 +5,7 @@ import httpx
 
 import litellm
 from litellm._logging import verbose_logger
+from litellm.litellm_core_utils.url_utils import ensure_api_route
 from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
 from litellm.llms.bedrock.common_utils import BedrockError
 from litellm.secret_managers.main import get_secret_str
@@ -140,9 +141,7 @@ class BedrockClaudePlatformMixin(BaseAWSLLM):
         if api_base is None:
             aws_region_name: Final = self._get_required_aws_region_name(optional_params)
             api_base = f"https://{CLAUDE_PLATFORM_SERVICE_NAME}.{aws_region_name}.api.aws"
-        if not api_base.endswith("/v1/messages"):
-            api_base = f"{api_base.rstrip('/')}/v1/messages"
-        return api_base
+        return ensure_api_route(api_base, "v1/messages")
 
     def sign_request(
         self,

litellm/llms/bedrock/common_utils.py

@@ -31,6 +31,7 @@ from litellm import verbose_logger
 from litellm.litellm_core_utils.aws_partition import get_aws_dns_suffix
 from litellm.litellm_core_utils.optional_imports import ensure_optional_import
 from litellm.litellm_core_utils.prompt_templates.common_utils import schema_property_count
+from litellm.litellm_core_utils.url_utils import ensure_api_route
 from litellm.llms.base_llm.anthropic_messages.transformation import (
     BaseAnthropicMessagesConfig,
 )
@@ -1305,10 +1306,7 @@ def build_mantle_messages_url(
     """
     override: Final = api_base or aws_bedrock_runtime_endpoint or _mantle_api_base_from_env()
     if override:
-        base: Final = override.rstrip("/")
-        if base.endswith(MANTLE_MESSAGES_PATH):
-            return base
-        return f"{base}{MANTLE_MESSAGES_PATH}"
+        return ensure_api_route(override, MANTLE_MESSAGES_PATH)
     return f"https://bedrock-mantle.{region}.api.aws{MANTLE_MESSAGES_PATH}"
 
 

litellm/llms/bedrock/count_tokens/mantle_handler.py

@@ -7,6 +7,7 @@ from typing_extensions import NotRequired, ReadOnly, TypedDict
 
 import litellm
 from litellm._logging import verbose_logger
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.anthropic.count_tokens.transformation import COUNT_TOKEN_OPTION_NAMES, AnthropicCountTokensConfig
 from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM, run_aws_signing
 from litellm.llms.bedrock.common_utils import BedrockError, build_mantle_messages_url
@@ -43,7 +44,7 @@ class BedrockMantleCountTokensHandler(BaseAWSLLM):
         messages_url: Final = build_mantle_messages_url(
             api_base=None, aws_bedrock_runtime_endpoint=None, region=aws_region_name
         )
-        return f"{messages_url}{MANTLE_COUNT_TOKENS_SUFFIX}"
+        return join_api_base(messages_url, MANTLE_COUNT_TOKENS_SUFFIX)
 
     async def handle_count_tokens_request(
         self,

litellm/llms/bedrock/count_tokens/transformation.py

@@ -11,6 +11,7 @@ from typing import Final, Literal
 
 from pydantic import JsonValue
 
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
 from litellm.llms.bedrock.common_utils import get_bedrock_base_model
 
@@ -236,9 +237,7 @@ class BedrockCountTokensConfig(BaseAWSLLM):
             aws_bedrock_runtime_endpoint=aws_bedrock_runtime_endpoint,
             aws_region_name=aws_region_name,
         )
-        endpoint: Final = f"{base_url}/model/{encoded_model_id}/count-tokens"
-
-        return endpoint
+        return join_api_base(base_url, f"model/{encoded_model_id}/count-tokens")
 
     def transform_bedrock_response_to_anthropic(
         self, bedrock_response: Mapping[str, JsonValue]

litellm/llms/bedrock/embed/embedding.py

@@ -12,6 +12,7 @@ import httpx
 
 import litellm
 from litellm.constants import BEDROCK_EMBEDDING_PROVIDERS_LITERAL
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.cohere.embed.handler import embedding as cohere_embedding
 from litellm.llms.custom_httpx.http_handler import (
     AsyncHTTPHandler,
@@ -512,9 +513,9 @@ class BedrockEmbedding(BaseAWSLLM):
             aws_region_name=aws_region_name,
         )
         if has_async_invoke:
-            endpoint_url = f"{endpoint_url}/async-invoke"
+            endpoint_url = join_api_base(endpoint_url, "async-invoke")
         else:
-            endpoint_url = f"{endpoint_url}/model/{modelId}/invoke"
+            endpoint_url = join_api_base(endpoint_url, f"model/{modelId}/invoke")
 
         if batch_data is not None:
             if aembedding:
@@ -611,7 +612,7 @@ class BedrockEmbedding(BaseAWSLLM):
 
         # Encode the ARN for use in URL path
         encoded_arn: Final = quote(invocation_arn, safe="")
-        status_url: Final = f"{endpoint_url.rstrip('/')}/async-invoke/{encoded_arn}"
+        status_url: Final = join_api_base(endpoint_url, f"async-invoke/{encoded_arn}")
 
         # Prepare headers for GET request
         headers: Final = {"Content-Type": "application/json"}

litellm/llms/bedrock/image_edit/handler.py

@@ -14,6 +14,7 @@ import httpx
 import litellm
 from litellm._logging import verbose_logger
 from litellm.litellm_core_utils.litellm_logging import Logging as LitellmLogging
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.bedrock.image_edit.amazon_nova_canvas_image_edit_transformation import (
     BedrockAmazonNovaCanvasImageEditConfig,
 )
@@ -226,7 +227,7 @@ class BedrockImageEdit(BaseAWSLLM):
             aws_bedrock_runtime_endpoint=boto3_credentials_info.aws_bedrock_runtime_endpoint,
             aws_region_name=boto3_credentials_info.aws_region_name,
         )
-        proxy_endpoint_url = f"{proxy_endpoint_url}/model/{modelId}/invoke"
+        proxy_endpoint_url = join_api_base(proxy_endpoint_url, f"model/{modelId}/invoke")
         data: Final = self._get_request_body(
             model=model,
             image=image,

litellm/llms/bedrock/image_generation/image_handler.py

@@ -8,6 +8,7 @@ import httpx
 import litellm
 from litellm._logging import verbose_logger
 from litellm.litellm_core_utils.litellm_logging import Logging as LitellmLogging
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.bedrock.image_generation.amazon_nova_canvas_transformation import (
     AmazonNovaCanvasConfig,
 )
@@ -248,7 +249,7 @@ class BedrockImageGeneration(BaseAWSLLM):
             aws_bedrock_runtime_endpoint=boto3_credentials_info.aws_bedrock_runtime_endpoint,
             aws_region_name=boto3_credentials_info.aws_region_name,
         )
-        proxy_endpoint_url = f"{proxy_endpoint_url}/model/{modelId}/invoke"
+        proxy_endpoint_url = join_api_base(proxy_endpoint_url, f"model/{modelId}/invoke")
         data: Final = self._get_request_body(
             model=model,
             prompt=prompt,

litellm/llms/bedrock/rerank/handler.py

@@ -6,6 +6,7 @@ from pydantic import ConfigDict, TypeAdapter
 
 import litellm
 from litellm.litellm_core_utils.litellm_logging import Logging as LitellmLogging
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.custom_httpx.http_handler import (
     AsyncHTTPHandler,
     HTTPHandler,
@@ -160,7 +161,7 @@ class BedrockRerankHandler(BaseAWSLLM):
             aws_region_name=boto3_credentials_info.aws_region_name,
         )
         proxy_endpoint_url = proxy_endpoint_url.replace("bedrock-runtime", "bedrock-agent-runtime")
-        proxy_endpoint_url = f"{proxy_endpoint_url}/rerank"
+        proxy_endpoint_url = join_api_base(proxy_endpoint_url, "rerank")
 
         body: Final = json.dumps(data).encode("utf-8")
         headers = {"Content-Type": "application/json"}

litellm/llms/bedrock/responses/transformation.py

@@ -46,6 +46,7 @@ from litellm.litellm_core_utils.prompt_templates.image_handling import (
     async_convert_url_to_base64,
     convert_url_to_base64,
 )
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.llms.base_llm.responses.codex_compat import drop_unsupported_tools, normalize_codex_input_items
 from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
@@ -223,9 +224,10 @@ class BedrockOpenAIResponsesConfig(BaseAWSLLM, OpenAIResponsesAPIConfig):
         litellm_params: dict,  # mutable-ok: signature fixed by the BaseResponsesAPIConfig override contract
     ) -> str:
         region: Final = self._get_aws_region_name(optional_params=litellm_params, model=None)
+        configured_endpoint: Final = litellm_params.get("aws_bedrock_runtime_endpoint")
         override: Final = (
             api_base
-            or litellm_params.get("aws_bedrock_runtime_endpoint")
+            or (f"{configured_endpoint}" if configured_endpoint else None)
             or get_secret_str("AWS_BEDROCK_RUNTIME_ENDPOINT")
         )
         # Partition-aware: bedrock-runtime is amazonaws.com.cn in China, and other
@@ -237,7 +239,7 @@ class BedrockOpenAIResponsesConfig(BaseAWSLLM, OpenAIResponsesAPIConfig):
             (host[: -len(suffix)] for suffix in BEDROCK_RUNTIME_OPENAI_BASE_SUFFIXES if host.endswith(suffix)),
             host,
         )
-        return f"{base}{BEDROCK_RUNTIME_OPENAI_RESPONSES_PATH}"
+        return join_api_base(base, BEDROCK_RUNTIME_OPENAI_RESPONSES_PATH)
 
     def supports_native_file_search(self) -> bool:
         return False

litellm/llms/bedrock/vector_stores/transformation.py

@@ -6,7 +6,7 @@ from urllib.parse import urlparse
 import httpx
 
 from litellm._logging import verbose_logger
-from litellm.litellm_core_utils.url_utils import encode_url_path_segment
+from litellm.litellm_core_utils.url_utils import encode_url_path_segment, join_api_base
 from litellm.llms.base_llm.vector_store.transformation import BaseVectorStoreConfig
 from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
 from litellm.llms.bedrock.common_utils import BedrockError
@@ -196,7 +196,7 @@ class BedrockVectorStoreConfig(BaseVectorStoreConfig, BaseAWSLLM):
             aws_region_name=self.get_aws_region_name_for_non_llm_api_calls(aws_region_name=aws_region_name),
             endpoint_type="agent",
         )
-        return f"{endpoint_url}/knowledgebases"
+        return join_api_base(endpoint_url, "knowledgebases")
 
     def transform_search_vector_store_request(
         self,
@@ -212,7 +212,7 @@ class BedrockVectorStoreConfig(BaseVectorStoreConfig, BaseAWSLLM):
             query = " ".join(query)
 
         encoded_vector_store_id: Final = encode_url_path_segment(vector_store_id, field_name="vector_store_id")
-        url: Final = f"{api_base}/{encoded_vector_store_id}/retrieve"
+        url: Final = join_api_base(api_base, f"{encoded_vector_store_id}/retrieve")
 
         request_body: Final[dict[str, object]] = {
             "retrievalQuery": BedrockKBRetrievalQuery(text=query),

litellm/llms/bedrock_mantle/messages/transformation.py

@@ -2,6 +2,7 @@ from collections.abc import Mapping
 from types import MappingProxyType
 from typing import Final
 
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
 from litellm.llms.bedrock.common_utils import MANTLE_MESSAGES_PATH
 from litellm.llms.bedrock.messages.mantle_transformation import AmazonMantleMessagesConfig
@@ -33,7 +34,7 @@ def build_mantle_native_messages_url(api_base: str | None, litellm_params: Mappi
         configured,
     )
     host: Final = f"https://bedrock-mantle.{region}.api.aws" if MANTLE_HOST_RE.match(stripped) else stripped
-    return f"{host}{MANTLE_MESSAGES_PATH}"
+    return join_api_base(host, MANTLE_MESSAGES_PATH)
 
 
 class BedrockMantleAnthropicMessagesConfig(BedrockMantleAuthMixin, AmazonMantleMessagesConfig):

litellm/llms/bedrock_mantle/responses/transformation.py

@@ -22,6 +22,7 @@ import httpx
 
 import litellm
 from litellm._logging import verbose_logger
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.llms.base_llm.responses.codex_compat import drop_unsupported_tools, normalize_codex_input_items
 from litellm.llms.base_llm.responses.transformation import ReasoningSummaryLevel
@@ -100,7 +101,7 @@ class BedrockMantleResponsesAPIConfig(BedrockMantleAuthMixin, OpenAIResponsesAPI
         if MANTLE_HOST_RE.match(base):
             base = f"https://bedrock-mantle.{region}.api.aws"
         path: Final = "/openai/v1/responses" if self.use_openai_path else "/v1/responses"
-        return f"{base}{path}"
+        return join_api_base(base, path)
 
     def validate_environment(self, headers: dict, model: str, litellm_params: GenericLiteLLMParams | None) -> dict:
         litellm_params = litellm_params or GenericLiteLLMParams()

litellm/llms/black_forest_labs/image_edit/transformation.py

@@ -19,7 +19,7 @@ from typing_extensions import NotRequired, ReadOnly, TypedDict
 
 import litellm
 from litellm.constants import DEFAULT_MAX_RECURSE_DEPTH
-from litellm.litellm_core_utils.url_utils import async_safe_get, safe_get
+from litellm.litellm_core_utils.url_utils import async_safe_get, join_api_base, safe_get
 from litellm.llms.base_llm.image_edit.transformation import BaseImageEditConfig
 from litellm.secret_managers.main import get_secret_str
 from litellm.types.images.main import ImageEditOptionalRequestParams
@@ -189,11 +189,10 @@ class BlackForestLabsImageEditConfig(BaseImageEditConfig):
         """
         Get the complete URL for the Black Forest Labs API request.
         """
-        base_url: str = api_base or get_secret_str("BFL_API_BASE") or DEFAULT_API_BASE
-        base_url = base_url.rstrip("/")
+        base_url: Final = api_base or get_secret_str("BFL_API_BASE") or DEFAULT_API_BASE
 
         endpoint: Final = self._get_model_endpoint(model)
-        return f"{base_url}{endpoint}"
+        return join_api_base(base_url, endpoint)
 
     def _read_image_bytes(
         self,

litellm/llms/black_forest_labs/image_generation/transformation.py

@@ -14,6 +14,7 @@ from typing import TYPE_CHECKING, Any, Final
 import httpx
 from pydantic import ConfigDict, TypeAdapter
 
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.image_generation.transformation import (
     BaseImageGenerationConfig,
 )
@@ -202,11 +203,10 @@ class BlackForestLabsImageGenerationConfig(BaseImageGenerationConfig):
         """
         Get the complete URL for the Black Forest Labs API request.
         """
-        base_url: str = api_base or get_secret_str("BFL_API_BASE") or DEFAULT_API_BASE
-        base_url = base_url.rstrip("/")
+        base_url: Final = api_base or get_secret_str("BFL_API_BASE") or DEFAULT_API_BASE
 
         endpoint: Final = self._get_model_endpoint(model)
-        return f"{base_url}{endpoint}"
+        return join_api_base(base_url, endpoint)
 
     def transform_image_generation_request(
         self,

litellm/llms/chatgpt/responses/transformation.py

@@ -9,6 +9,7 @@ from litellm.litellm_core_utils.hidden_params import HIDDEN_PARAMS_ATTR, set_hid
 from litellm.litellm_core_utils.llm_response_utils.convert_dict_to_response import (
     safe_convert_created_field,
 )
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.openai.common_utils import OpenAIError
 from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig
 from litellm.responses.sse_output_recovery import (
@@ -253,8 +254,7 @@ class ChatGPTResponsesAPIConfig(OpenAIResponsesAPIConfig):
         litellm_params: dict,
     ) -> str:
         api_base = api_base or self.authenticator.get_api_base() or CHATGPT_API_BASE
-        api_base = api_base.rstrip("/")
-        return f"{api_base}/responses"
+        return join_api_base(api_base, "responses")
 
     def supports_native_websocket(self) -> bool:
         """ChatGPT does not support native WebSocket for Responses API"""

litellm/llms/cloudflare/chat/transformation.py

@@ -3,6 +3,7 @@ from typing import Final
 import httpx
 
 from litellm._logging import verbose_logger
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.llms.openai.chat.gpt_transformation import OpenAIGPTConfig
 from litellm.secret_managers.main import (
@@ -59,7 +60,7 @@ class CloudflareChatConfig(OpenAIGPTConfig):
             verbose_logger.warning(
                 "Cloudflare api_base ending in '/ai/run' is the legacy Workers AI path and no longer serves OpenAI-compatible requests; rewriting to the '/ai/v1' endpoint"
             )
-            return f"{trimmed[: -len('/ai/run')]}/ai/v1"
+            return join_api_base(trimmed[: -len("/ai/run")], "ai/v1")
         return api_base
 
     def validate_environment(

litellm/llms/cloudflare/common_utils.py

@@ -3,6 +3,7 @@ from typing import Final
 
 from pydantic import TypeAdapter, ValidationError
 
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.secret_managers.main import (
     get_secret_str,
     normalize_nonempty_secret_str,
@@ -34,10 +35,10 @@ def workers_ai_request_model(model: str) -> str:
 def workers_ai_run_url(api_base: str, model: str) -> str:
     normalized_api_base: Final = api_base.rstrip("/")
     if normalized_api_base.endswith("/ai/v1"):
-        return f"{normalized_api_base.removesuffix('/ai/v1')}/ai/run/{model}"
+        return join_api_base(normalized_api_base.removesuffix("/ai/v1"), f"ai/run/{model}")
     if normalized_api_base.endswith("/ai/run"):
         return f"{normalized_api_base}/{model}"
-    return f"{normalized_api_base}/ai/run/{model}"
+    return join_api_base(normalized_api_base, f"ai/run/{model}")
 
 
 def unwrap_workers_ai_result(payload: object) -> object:

litellm/llms/cohere/rerank/transformation.py

@@ -6,6 +6,7 @@ from pydantic import ConfigDict, TypeAdapter
 
 import litellm
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+from litellm.litellm_core_utils.url_utils import ensure_api_route
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.llms.base_llm.rerank.transformation import BaseRerankConfig
 from litellm.secret_managers.main import get_secret_str
@@ -31,11 +32,7 @@ class CohereRerankConfig(BaseRerankConfig):
         optional_params: dict | None = None,
     ) -> str:
         if api_base:
-            # Remove trailing slashes and ensure clean base URL
-            api_base = api_base.rstrip("/")
-            if not api_base.endswith("/v1/rerank"):
-                api_base = f"{api_base}/v1/rerank"
-            return api_base
+            return ensure_api_route(api_base, "v1/rerank")
         return "https://api.cohere.ai/v1/rerank"
 
     def get_supported_cohere_rerank_params(self, model: str) -> list:

litellm/llms/cohere/rerank_v2/transformation.py

@@ -1,5 +1,6 @@
 from typing import Any, Final
 
+from litellm.litellm_core_utils.url_utils import ensure_api_route
 from litellm.llms.cohere.rerank.transformation import CohereRerankConfig
 from litellm.types.rerank import OptionalRerankParams, RerankRequest
 
@@ -19,11 +20,7 @@ class CohereRerankV2Config(CohereRerankConfig):
         optional_params: dict | None = None,
     ) -> str:
         if api_base:
-            # Remove trailing slashes and ensure clean base URL
-            api_base = api_base.rstrip("/")
-            if not api_base.endswith("/v2/rerank"):
-                api_base = f"{api_base}/v2/rerank"
-            return api_base
+            return ensure_api_route(api_base, "v2/rerank")
         return "https://api.cohere.ai/v2/rerank"
 
     def get_supported_cohere_rerank_params(self, model: str) -> list:

litellm/llms/cometapi/chat/transformation.py

@@ -10,6 +10,7 @@ from typing import Any, Final
 
 import httpx
 
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.base_model_iterator import BaseModelResponseIterator
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.types.llms.openai import AllMessageValues, ChatCompletionToolParam
@@ -112,16 +113,14 @@ class CometAPIConfig(OpenAIGPTConfig):
 
         # Ensure we include /v1 prefix when missing
         if api_base.endswith("/v1"):
-            return f"{api_base}/{endpoint}"
-        if api_base.endswith("/v1/"):
-            return f"{api_base}{endpoint}"
+            return join_api_base(api_base, endpoint)
         # If user provided https://api.cometapi.com, add /v1
         if api_base == "https://api.cometapi.com":
-            return f"{api_base}/v1/{endpoint}"
+            return join_api_base(api_base, f"v1/{endpoint}")
         # Generic fallback: if '/v1' not in path, add it
         if "/v1" not in api_base.split("//", 1)[-1]:
-            return f"{api_base}/v1/{endpoint}"
-        return f"{api_base}/{endpoint}"
+            return join_api_base(api_base, f"v1/{endpoint}")
+        return join_api_base(api_base, endpoint)
 
     def get_error_class(self, error_message: str, status_code: int, headers: dict | httpx.Headers) -> BaseLLMException:
         """

litellm/llms/cometapi/embed/transformation.py

@@ -9,6 +9,7 @@ import httpx
 from pydantic import ConfigDict, TypeAdapter
 
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.llms.base_llm.embedding.transformation import BaseEmbeddingConfig
 from litellm.secret_managers.main import get_secret_str
@@ -44,9 +45,7 @@ class CometAPIEmbeddingConfig(BaseEmbeddingConfig):
         """
         Get the complete URL for the CometAPI embedding endpoint.
         """
-        api_base = "https://api.cometapi.com/v1" if api_base is None else api_base.rstrip("/")
-        complete_url: Final = f"{api_base}/embeddings"
-        return complete_url
+        return join_api_base("https://api.cometapi.com/v1" if api_base is None else api_base, "embeddings")
 
     def validate_environment(
         self,

litellm/llms/cometapi/image_generation/transformation.py

@@ -4,6 +4,7 @@ from typing import TYPE_CHECKING, Any, Final
 import httpx
 from pydantic import ConfigDict, TypeAdapter
 
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.image_generation.transformation import (
     BaseImageGenerationConfig,
 )
@@ -77,16 +78,13 @@ class CometAPIImageGenerationConfig(BaseImageGenerationConfig):
         """
         Get the complete url for the request
         """
-        complete_url: str = (
+        base_url: Final = (
             api_base
             or get_secret_str("COMETAPI_BASE_URL")
             or get_secret_str("COMETAPI_API_BASE")
             or self.DEFAULT_BASE_URL
         )
-
-        complete_url = complete_url.rstrip("/")
-        complete_url = f"{complete_url}/{self.IMAGE_GENERATION_ENDPOINT}"
-        return complete_url
+        return join_api_base(base_url, self.IMAGE_GENERATION_ENDPOINT)
 
     def validate_environment(
         self,

litellm/llms/custom_httpx/llm_http_handler.py

@@ -59,7 +59,7 @@ from litellm.litellm_core_utils.realtime_errors import (
     websocket_close_reason,
 )
 from litellm.litellm_core_utils.realtime_streaming import RealTimeStreaming
-from litellm.litellm_core_utils.url_utils import encode_url_path_segment
+from litellm.litellm_core_utils.url_utils import encode_url_path_segment, join_api_base
 from litellm.llms.base_llm.anthropic_messages.transformation import (
     BaseAnthropicMessagesConfig,
 )
@@ -6937,7 +6937,7 @@ class BaseLLMHTTPHandler:
                 headers={}, model=model or "", api_key=api_key
             )
         else:
-            url = f"{api_base.rstrip('/')}/v1/realtime/{endpoint}"
+            url = join_api_base(api_base, f"v1/realtime/{endpoint}")
             headers = {
                 "Authorization": f"Bearer {api_key}",
                 "Content-Type": "application/json",
@@ -7006,7 +7006,7 @@ class BaseLLMHTTPHandler:
             url = provider_config.get_realtime_calls_url(api_base=api_base, model=model or "", api_version=api_version)
             headers: dict[str, object] = provider_config.get_realtime_calls_headers(ephemeral_key=openai_ephemeral_key)
         else:
-            url = f"{api_base.rstrip('/')}/v1/realtime/calls"
+            url = join_api_base(api_base, "v1/realtime/calls")
             headers = {
                 "Authorization": f"Bearer {openai_ephemeral_key}",
             }

litellm/llms/dashscope/chat/transformation.py

@@ -5,6 +5,7 @@ Translates from OpenAI's `/v1/chat/completions` to DashScope's `/v1/chat/complet
 from collections.abc import Coroutine
 from typing import Final, Literal, overload
 
+from litellm.litellm_core_utils.url_utils import ensure_api_route
 from litellm.secret_managers.main import get_secret_str
 from litellm.types.llms.openai import AllMessageValues, ChatCompletionToolParam
 
@@ -82,7 +83,4 @@ class DashScopeChatConfig(OpenAIGPTConfig):
         """
         If api_base is not provided, use the default DashScope /chat/completions endpoint.
         """
-        resolved_api_base: Final = self._resolve_chat_api_base(api_base).rstrip("/")
-        if resolved_api_base.endswith("/chat/completions"):
-            return resolved_api_base
-        return f"{resolved_api_base}/chat/completions"
+        return ensure_api_route(self._resolve_chat_api_base(api_base), "chat/completions")

litellm/llms/dashscope/embed/transformation.py

@@ -16,6 +16,7 @@ from typing import Final
 import httpx
 
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+from litellm.litellm_core_utils.url_utils import ensure_api_route
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.llms.base_llm.embedding.transformation import BaseEmbeddingConfig
 from litellm.secret_managers.main import get_secret_str
@@ -98,10 +99,7 @@ class DashScopeEmbeddingConfig(BaseEmbeddingConfig):
         litellm_params: dict,
         stream: bool | None = None,
     ) -> str:
-        base: Final = self._resolve_embedding_api_base(api_base).rstrip("/")
-        if base.endswith("/embeddings"):
-            return base
-        return f"{base}/embeddings"
+        return ensure_api_route(self._resolve_embedding_api_base(api_base), "embeddings")
 
     def transform_embedding_request(
         self,

litellm/llms/dashscope/rerank/transformation.py

@@ -34,6 +34,7 @@ from pydantic import ConfigDict, TypeAdapter
 
 from litellm._uuid import uuid
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.llms.base_llm.rerank.transformation import BaseRerankConfig
 from litellm.secret_managers.main import get_secret_str
@@ -94,10 +95,10 @@ class DashScopeRerankConfig(BaseRerankConfig):
             return cleaned
 
         if cleaned.endswith("/v1"):
-            return f"{cleaned}/reranks"
+            return join_api_base(cleaned, "reranks")
 
         # Unknown base: append /reranks rather than silently ignoring the caller's api_base.
-        return f"{cleaned}/reranks"
+        return join_api_base(cleaned, "reranks")
 
     def validate_environment(
         self,

litellm/llms/dashscope/videos/transformation.py

@@ -16,7 +16,7 @@ from pydantic import BaseModel, Field, StrictBool, StrictInt, StrictStr, TypeAda
 
 import litellm
 from litellm.exceptions import UnsupportedParamsError
-from litellm.litellm_core_utils.url_utils import encode_url_path_segment
+from litellm.litellm_core_utils.url_utils import encode_url_path_segment, join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.llms.base_llm.videos.async_task_utils import (
     input_reference_url,
@@ -708,7 +708,7 @@ class DashScopeVideoConfig(BaseVideoConfig):
         if parameters:
             request_data["parameters"] = parameters
 
-        return request_data, (), f"{api_base}{DASHSCOPE_VIDEO_SYNTHESIS_PATH}"
+        return request_data, (), join_api_base(api_base, DASHSCOPE_VIDEO_SYNTHESIS_PATH)
 
     @staticmethod
     def _reference_input(
@@ -773,7 +773,7 @@ class DashScopeVideoConfig(BaseVideoConfig):
     def _task_url(video_id: str, api_base: str) -> str:
         original_task_id: Final = extract_original_video_id(video_id)
         encoded_task_id: Final = encode_url_path_segment(original_task_id, field_name="video_id")
-        return f"{api_base}{DASHSCOPE_TASKS_PATH}/{encoded_task_id}"
+        return join_api_base(api_base, f"{DASHSCOPE_TASKS_PATH}/{encoded_task_id}")
 
     def transform_video_status_retrieve_response(
         self,

litellm/llms/databricks/chat/transformation.py

@@ -22,6 +22,7 @@ from litellm.litellm_core_utils.prompt_templates.common_utils import (
     strip_name_from_message,
     strip_prefill_prefix_field,
 )
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.base_model_iterator import BaseModelResponseIterator
 from litellm.llms.base_llm.base_utils import (
     type_to_response_format_param,  # pyright: ignore[reportUnknownVariableType]  # base_utils helper returns an untyped dict
@@ -339,10 +340,7 @@ class DatabricksConfig(DatabricksBase, OpenAILikeChatConfig, AnthropicConfig):
         stream: bool | None = None,
     ) -> str:
         use_ai_gateway: Final = self.uses_ai_gateway(model)
-        api_base = self._get_api_base(api_base, use_ai_gateway=use_ai_gateway)
-        url_base: Final = api_base.rstrip("/") if use_ai_gateway else api_base
-        complete_url: Final = f"{url_base}/chat/completions"
-        return complete_url
+        return join_api_base(self._get_api_base(api_base, use_ai_gateway=use_ai_gateway), "chat/completions")
 
     def get_supported_openai_params(self, model: str | None = None) -> list:
         return [

litellm/llms/databricks/common_utils.py

@@ -20,6 +20,7 @@ from urllib.parse import urlsplit, urlunsplit
 from pydantic import BaseModel, ConfigDict
 
 import litellm
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.secret_managers.main import get_secret_str
 
@@ -278,10 +279,10 @@ class DatabricksBase:
         if normalized_api_base.endswith("/ai-gateway/mlflow/v1"):
             return normalized_api_base
         if normalized_api_base.endswith("/serving-endpoints"):
-            return f"{normalized_api_base.removesuffix('/serving-endpoints')}/ai-gateway/mlflow/v1"
+            return join_api_base(normalized_api_base.removesuffix("/serving-endpoints"), "ai-gateway/mlflow/v1")
         api_base_parts: Final = urlsplit(normalized_api_base)
         if api_base_parts.path in ("", "/"):
-            return f"{normalized_api_base}/ai-gateway/mlflow/v1"
+            return join_api_base(normalized_api_base, "ai-gateway/mlflow/v1")
         return api_base
 
     def get_oauth_m2m_token(
@@ -319,7 +320,7 @@ class DatabricksBase:
 
         api_base_parts: Final = urlsplit(api_base)
         workspace_url: Final = urlunsplit((api_base_parts.scheme, api_base_parts.netloc, "", "", ""))
-        token_url: Final = f"{workspace_url}/oidc/v1/token"
+        token_url: Final = join_api_base(workspace_url, "oidc/v1/token")
 
         try:
             response: Final = requests.post(
@@ -477,5 +478,6 @@ class DatabricksBase:
         # Debug logging with redaction (never log actual tokens)
         verbose_logger.debug("Databricks request headers: %s", self.redact_headers_for_logging(headers))
 
-        endpoint_path: Final = "" if custom_endpoint is True else DATABRICKS_ENDPOINT_PATHS[endpoint_type]
-        return f"{api_base}{endpoint_path}", headers
+        if custom_endpoint is True:
+            return api_base, headers
+        return join_api_base(api_base, DATABRICKS_ENDPOINT_PATHS[endpoint_type]), headers

litellm/llms/databricks/decisions/transformation.py

@@ -6,6 +6,7 @@ from typing import Final
 
 from pydantic import TypeAdapter, ValidationError
 
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.decisions.transformation import BaseDecisionsConfig
 from litellm.types.decisions import DecisionsIRRequest, UnsupportedDecisionsRequest
 
@@ -84,8 +85,8 @@ class DatabricksDecisionsConfig(BaseDecisionsConfig):
 
     def get_complete_url(self, api_base: str, model: str) -> str:
         if is_ai_decide_model(model):
-            return f"{databricks_workspace_host(api_base)}{_AI_DECIDE_PATH}"
-        return f"{api_base.rstrip('/')}/{validate_databricks_decisions_model(model)}/invocations"
+            return join_api_base(databricks_workspace_host(api_base), _AI_DECIDE_PATH)
+        return join_api_base(api_base, f"{validate_databricks_decisions_model(model)}/invocations")
 
     def transform_decisions_request(
         self,

litellm/llms/databricks/responses/transformation.py

@@ -10,6 +10,7 @@ Reference: https://docs.databricks.com/aws/en/machine-learning/foundation-model-
 import os
 from typing import TYPE_CHECKING, Any, Final
 
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.databricks.common_utils import DatabricksBase
 from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig
 from litellm.types.llms.openai import ResponseInputParam
@@ -70,8 +71,7 @@ class DatabricksResponsesAPIConfig(DatabricksBase, OpenAIResponsesAPIConfig):
     ) -> str:
         api_base = api_base or os.getenv("DATABRICKS_API_BASE")
         api_base = self._get_api_base(api_base)
-        api_base = api_base.rstrip("/")
-        return f"{api_base}/responses"
+        return join_api_base(api_base, "responses")
 
     def transform_responses_api_request(
         self,

litellm/llms/deepgram/audio_transcription/transformation.py

@@ -10,6 +10,7 @@ from httpx import Headers, Response
 from pydantic import ConfigDict, TypeAdapter
 
 from litellm.litellm_core_utils.audio_utils.utils import process_audio_file
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.secret_managers.main import get_secret_str
 from litellm.types.llms.openai import (
@@ -179,7 +180,6 @@ class DeepgramAudioTranscriptionConfig(BaseAudioTranscriptionConfig):
     ) -> str:
         if api_base is None:
             api_base = get_secret_str("DEEPGRAM_API_BASE") or "https://api.deepgram.com/v1"
-        api_base = api_base.rstrip("/")  # Remove trailing slash if present
 
         # Build query parameters including the model
         all_query_params: Final = {"model": model}
@@ -189,7 +189,7 @@ class DeepgramAudioTranscriptionConfig(BaseAudioTranscriptionConfig):
         all_query_params.update(additional_params)
 
         # Construct URL with proper query string encoding
-        base_url: Final = f"{api_base}/listen"
+        base_url: Final = join_api_base(api_base, "listen")
         query_string: Final = urlencode(all_query_params)
         url: Final = f"{base_url}?{query_string}"
 

litellm/llms/deepgram/common_utils.py

@@ -8,6 +8,7 @@ import httpx
 
 import litellm
 from litellm.constants import DEEPGRAM_DEFAULT_API_BASE, DEEPGRAM_LISTEN_DEFAULT_MODEL
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.types.utils import LlmProviders
 
@@ -50,7 +51,7 @@ def _first_occurrences(query_string: str) -> httpx.QueryParams:
 
 
 def deepgram_listen_websocket_target(api_base: str | None, query_string: str) -> str:
-    listen_url: Final = httpx.URL(f"{(api_base or DEEPGRAM_DEFAULT_API_BASE).rstrip('/')}/listen")
+    listen_url: Final = httpx.URL(join_api_base(api_base or DEEPGRAM_DEFAULT_API_BASE, "listen"))
     websocket_url: Final = listen_url.copy_with(scheme=_WEBSOCKET_SCHEMES.get(listen_url.scheme, listen_url.scheme))
     params: Final = _first_occurrences(query_string)
     query: Final = params if params.get("model") else params.remove("model").add("model", DEEPGRAM_LISTEN_DEFAULT_MODEL)

litellm/llms/deepseek/chat/transformation.py

@@ -10,6 +10,7 @@ from litellm.litellm_core_utils.prompt_templates.common_utils import (
     convert_content_list_to_str,
     extract_search_results_text,
 )
+from litellm.litellm_core_utils.url_utils import ensure_api_route
 from litellm.secret_managers.main import get_secret_str
 from litellm.types.llms.openai import AllMessageValues
 from litellm.utils import supports_reasoning, supports_vision
@@ -363,9 +364,4 @@ class DeepSeekChatConfig(OpenAIGPTConfig):
         """
         If api_base is not provided, use the default DeepSeek /chat/completions endpoint.
         """
-        clean_api_base: Final = (api_base or "https://api.deepseek.com/beta").rstrip("/")
-
-        if not clean_api_base.endswith("/chat/completions"):
-            return f"{clean_api_base}/chat/completions"
-
-        return clean_api_base
+        return ensure_api_route(api_base or "https://api.deepseek.com/beta", "chat/completions")

litellm/llms/deepseek/messages/transformation.py

@@ -6,6 +6,7 @@ from collections.abc import Sequence
 from typing import Any, Final
 
 import litellm
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.anthropic.pass_through.messages.transformation import (
     AnthropicMessagesConfig,
 )
@@ -90,9 +91,9 @@ class DeepSeekAnthropicMessagesConfig(AnthropicMessagesConfig):
         base_url = base_url.removesuffix("/beta")
 
         if not base_url.endswith("/anthropic") and "/anthropic/" not in base_url:
-            base_url = f"{base_url}/anthropic"
+            base_url = join_api_base(base_url, "anthropic")
 
-        return f"{base_url}/v1/messages"
+        return join_api_base(base_url, "v1/messages")
 
     @staticmethod
     def _sanitize_tools_for_deepseek(tools: Any) -> object:

litellm/llms/docker_model_runner/chat/transformation.py

@@ -10,6 +10,7 @@ from typing import Final, Literal, overload
 from litellm.litellm_core_utils.prompt_templates.common_utils import (
     handle_messages_with_content_list_to_str_conversion,
 )
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.secret_managers.main import get_secret_str
 from litellm.types.llms.openai import AllMessageValues
 
@@ -104,12 +105,9 @@ class DockerModelRunnerChatConfig(OpenAIGPTConfig):
         if not api_base:
             api_base = "http://localhost:22088/engines/llama.cpp"
 
-        # Remove trailing slashes from api_base
-        api_base = api_base.rstrip("/")
-
         # Build the URL: {api_base}/v1/chat/completions
         # api_base is expected to already contain the engine path
-        complete_url: Final = f"{api_base}/v1/chat/completions"
+        complete_url: Final = join_api_base(api_base, "v1/chat/completions")
 
         return complete_url
 

litellm/llms/duckduckgo/search/transformation.py

@@ -12,6 +12,7 @@ import httpx
 from pydantic import ConfigDict, TypeAdapter
 
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.search.transformation import (
     BaseSearchConfig,
     SearchResponse,
@@ -91,7 +92,7 @@ class DuckDuckGoSearchConfig(BaseSearchConfig):
         if data and isinstance(data, dict) and "_duckduckgo_params" in data:
             params: Final = data["_duckduckgo_params"]
             query_string: Final = urlencode(params, doseq=True)
-            return f"{api_base}/?{query_string}"
+            return f"{join_api_base(api_base, '')}?{query_string}"
 
         return api_base
 

litellm/llms/e2b/sandbox/transformation.py

@@ -15,6 +15,7 @@ import httpx
 from pydantic import ConfigDict, TypeAdapter
 
 from litellm.litellm_core_utils.hidden_params import HIDDEN_PARAMS_ATTR
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.sandbox.transformation import (
     SANDBOX_MAX_OUTPUT_BYTES,
     BaseSandboxConfig,
@@ -78,7 +79,7 @@ class E2BSandboxConfig(BaseSandboxConfig):
             body["metadata"] = metadata
 
         response: Final = await self._http(client).post(
-            url=f"{base}/sandboxes",
+            url=join_api_base(base, "sandboxes"),
             headers={"X-API-Key": key, "Content-Type": "application/json"},
             json=body,
         )
@@ -151,10 +152,10 @@ class E2BSandboxConfig(BaseSandboxConfig):
             dict[str, object], getattr(handle, HIDDEN_PARAMS_ATTR)
         )
         key: Final = api_key or hidden_params.get("api_key") or self.validate_environment()
-        base: Final = api_base or hidden_params.get("api_base") or E2B_API_BASE
+        base: Final = str(api_base or hidden_params.get("api_base") or E2B_API_BASE)
         try:
             response: Final = await self._http(client).delete(
-                url=f"{base}/sandboxes/{handle.id}",
+                url=join_api_base(base, f"sandboxes/{handle.id}"),
                 headers={"X-API-Key": key},
             )
         except httpx.HTTPStatusError as e:

litellm/llms/edenai/chat/transformation.py

@@ -23,7 +23,7 @@ from litellm.types.llms.base import LiteLLMBaseModel
 from litellm.types.llms.openai import AllMessageValues
 from litellm.types.utils import ModelResponse, ModelResponseStream, Usage
 
-from ..common_utils import EdenAIException, reported_cost, resolve_api_base, resolve_api_key
+from ..common_utils import EdenAIException, endpoint_url, reported_cost, resolve_api_base, resolve_api_key
 
 if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
@@ -138,7 +138,7 @@ class EdenAIChatConfig(OpenAIGPTConfig):
     def get_models(
         self, api_key: str | None = None, api_base: str | None = None
     ) -> list[str]:  # mutable-ok: inherited contract
-        response: Final = litellm.module_level_client.get(url=f"{self.get_api_base(api_base)}/models")
+        response: Final = litellm.module_level_client.get(url=endpoint_url(api_base, "models"))
         if not response.is_success:
             raise EdenAIException(status_code=response.status_code, message=response.text, headers=response.headers)
         catalog: Final = _EdenAIModelCatalog.model_validate(response.json())

litellm/llms/edenai/common_utils.py

@@ -11,6 +11,7 @@ from pydantic import AliasChoices, Field, ValidationError
 
 import litellm
 from litellm.exceptions import AuthenticationError
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.secret_managers.main import get_secret_str
 from litellm.types.llms.base import LiteLLMBaseModel
@@ -74,7 +75,7 @@ def json_headers(
 
 
 def endpoint_url(api_base: str | None, path: str) -> str:
-    return f"{resolve_api_base(api_base).rstrip('/')}/{path}"
+    return join_api_base(resolve_api_base(api_base), path)
 
 
 def pick(params: Mapping[str, object], keys: Container[str]) -> Mapping[str, object]:

litellm/llms/elevenlabs/audio_transcription/transformation.py

@@ -10,6 +10,7 @@ from pydantic import ConfigDict, TypeAdapter
 
 import litellm
 from litellm.litellm_core_utils.audio_utils.utils import process_audio_file
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.secret_managers.main import get_secret_str
 from litellm.types.llms.openai import (
@@ -165,12 +166,7 @@ class ElevenLabsAudioTranscriptionConfig(BaseAudioTranscriptionConfig):
     ) -> str:
         if api_base is None:
             api_base = get_secret_str("ELEVENLABS_API_BASE") or "https://api.elevenlabs.io"
-        api_base = api_base.rstrip("/")  # Remove trailing slash if present
-
-        # ElevenLabs speech-to-text endpoint
-        url: Final = f"{api_base}/v1/speech-to-text"
-
-        return url
+        return join_api_base(api_base, "v1/speech-to-text")
 
     def validate_environment(
         self,

litellm/llms/elevenlabs/text_to_speech/transformation.py

@@ -11,7 +11,7 @@ import httpx
 from httpx import Headers
 
 import litellm
-from litellm.litellm_core_utils.url_utils import encode_url_path_segment
+from litellm.litellm_core_utils.url_utils import encode_url_path_segment, join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.llms.base_llm.text_to_speech.transformation import (
     BaseTextToSpeechConfig,
@@ -297,15 +297,14 @@ class ElevenLabsTextToSpeechConfig(BaseTextToSpeechConfig):
         """
         Construct the ElevenLabs endpoint URL, including path voice_id and query params.
         """
-        base_url = api_base or get_secret_str("ELEVENLABS_API_BASE") or self.TTS_BASE_URL
-        base_url = base_url.rstrip("/")
+        base_url: Final = api_base or get_secret_str("ELEVENLABS_API_BASE") or self.TTS_BASE_URL
 
         voice_id: Final = litellm_params.get(self.ELEVENLABS_VOICE_ID_KEY)
         if not isinstance(voice_id, str) or not voice_id.strip():
             raise ValueError("ElevenLabs voice_id is required. Pass `voice` when calling `litellm.speech()`.")
 
         encoded_voice_id: Final = encode_url_path_segment(voice_id, field_name="voice_id")
-        url = f"{base_url}{self.TTS_ENDPOINT_PATH}/{encoded_voice_id}"
+        url = join_api_base(base_url, f"{self.TTS_ENDPOINT_PATH}/{encoded_voice_id}")
 
         query_params: Final = litellm_params.get(self.ELEVENLABS_QUERY_PARAMS_KEY, {})
         if query_params:

litellm/llms/exa_ai/search/transformation.py

@@ -11,6 +11,7 @@ import httpx
 from pydantic import ConfigDict, TypeAdapter
 
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+from litellm.litellm_core_utils.url_utils import ensure_api_route
 from litellm.llms.base_llm.search.transformation import (
     BaseSearchConfig,
     SearchResponse,
@@ -93,11 +94,7 @@ class ExaAISearchConfig(BaseSearchConfig):
         """
         Get complete URL for Search endpoint.
         """
-        resolved_api_base: Final = api_base or get_secret_str("EXA_API_BASE") or self.EXA_AI_API_BASE
-        clean_api_base: Final = resolved_api_base.rstrip("/")
-        if clean_api_base.endswith("/search"):
-            return clean_api_base
-        return f"{clean_api_base}/search"
+        return ensure_api_route(api_base or get_secret_str("EXA_API_BASE") or self.EXA_AI_API_BASE, "search")
 
     def transform_search_request(
         self,

litellm/llms/fal_ai/image_edit/flux_lora_depth_transformation.py

@@ -4,6 +4,7 @@ from typing import Final
 
 from httpx._types import RequestFiles
 
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.secret_managers.main import get_secret_str
 from litellm.types.images.main import ImageEditOptionalRequestParams
 from litellm.types.router import GenericLiteLLMParams
@@ -45,8 +46,8 @@ class FalAIFluxLoraDepthEditConfig(FalAIImageEditConfig):
         api_base: str | None,
         litellm_params: dict,  # mutable-ok: base class contract
     ) -> str:
-        base_url: Final = (api_base or get_secret_str("FAL_AI_API_BASE") or DEFAULT_BASE_URL).rstrip("/")
-        return f"{base_url}/{FLUX_LORA_DEPTH_ENDPOINT}"
+        base_url: Final = api_base or get_secret_str("FAL_AI_API_BASE") or DEFAULT_BASE_URL
+        return join_api_base(base_url, FLUX_LORA_DEPTH_ENDPOINT)
 
     def transform_image_edit_request(
         self,

litellm/llms/fal_ai/image_generation/transformation.py

@@ -5,6 +5,7 @@ import httpx
 from pydantic import ConfigDict, TypeAdapter
 from typing_extensions import ReadOnly, TypedDict
 
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.image_generation.transformation import (
     BaseImageGenerationConfig,
 )
@@ -85,12 +86,10 @@ class FalAIBaseConfig(BaseImageGenerationConfig):
 
         Some providers need `model` in `api_base`
         """
-        complete_url: str = api_base or get_secret_str("FAL_AI_API_BASE") or self.DEFAULT_BASE_URL
-
-        complete_url = complete_url.rstrip("/")
-        if self.IMAGE_GENERATION_ENDPOINT:
-            complete_url = f"{complete_url}/{self.IMAGE_GENERATION_ENDPOINT}"
-        return complete_url
+        base_url: Final = api_base or get_secret_str("FAL_AI_API_BASE") or self.DEFAULT_BASE_URL
+        if not self.IMAGE_GENERATION_ENDPOINT:
+            return base_url.rstrip("/")
+        return join_api_base(base_url, self.IMAGE_GENERATION_ENDPOINT)
 
     def validate_environment(
         self,

litellm/llms/fal_ai/videos/transformation.py

@@ -10,7 +10,7 @@ import httpx
 from httpx._types import FileContent, RequestFiles
 from pydantic import TypeAdapter
 
-from litellm.litellm_core_utils.url_utils import encode_url_path_segment
+from litellm.litellm_core_utils.url_utils import encode_url_path_segment, join_api_base
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.llms.base_llm.videos.transformation import BaseVideoConfig
 from litellm.llms.custom_httpx.http_handler import (
@@ -445,7 +445,7 @@ class FalAIVideoConfig(BaseVideoConfig):
         request_id, model_id = self._decode_video_id(video_id)
         encoded_request_id: Final[str] = encode_url_path_segment(request_id, field_name="video_id")
         return (
-            f"{api_base.rstrip('/')}/{_queue_request_base_path(model_id)}/requests/{encoded_request_id}/status",
+            join_api_base(api_base, f"{_queue_request_base_path(model_id)}/requests/{encoded_request_id}/status"),
             {},
         )
 
@@ -541,7 +541,7 @@ class FalAIVideoConfig(BaseVideoConfig):
         request_id, model_id = self._decode_video_id(video_id)
         encoded_request_id: Final[str] = encode_url_path_segment(request_id, field_name="video_id")
         return (
-            f"{api_base.rstrip('/')}/{_queue_request_base_path(model_id)}/requests/{encoded_request_id}",
+            join_api_base(api_base, f"{_queue_request_base_path(model_id)}/requests/{encoded_request_id}"),
             {},
         )
 

litellm/llms/fastcrw/search/transformation.py

@@ -15,6 +15,7 @@ import httpx
 from pydantic import ConfigDict, TypeAdapter
 
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+from litellm.litellm_core_utils.url_utils import ensure_api_route
 from litellm.llms.base_llm.search.transformation import (
     BaseSearchConfig,
     SearchResponse,
@@ -83,11 +84,7 @@ class FastCRWSearchConfig(BaseSearchConfig):
         """
         Get complete URL for Search endpoint.
         """
-        resolved_api_base: Final = api_base or get_secret_str("CRW_API_BASE") or self.FASTCRW_API_BASE
-        clean_api_base: Final = resolved_api_base.rstrip("/")
-        if clean_api_base.endswith("/search"):
-            return clean_api_base
-        return f"{clean_api_base}/search"
+        return ensure_api_route(api_base or get_secret_str("CRW_API_BASE") or self.FASTCRW_API_BASE, "search")
 
     def transform_search_request(
         self,

litellm/llms/firecrawl/search/transformation.py

@@ -11,6 +11,7 @@ import httpx
 from pydantic import ConfigDict, TypeAdapter
 
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+from litellm.litellm_core_utils.url_utils import ensure_api_route
 from litellm.llms.base_llm.search.transformation import (
     BaseSearchConfig,
     SearchResponse,
@@ -85,11 +86,7 @@ class FirecrawlSearchConfig(BaseSearchConfig):
         """
         Get complete URL for Search endpoint.
         """
-        resolved_api_base: Final = api_base or get_secret_str("FIRECRAWL_API_BASE") or self.FIRECRAWL_API_BASE
-        clean_api_base: Final = resolved_api_base.rstrip("/")
-        if clean_api_base.endswith("/search"):
-            return clean_api_base
-        return f"{clean_api_base}/search"
+        return ensure_api_route(api_base or get_secret_str("FIRECRAWL_API_BASE") or self.FIRECRAWL_API_BASE, "search")
 
     def transform_search_request(
         self,

litellm/llms/fireworks_ai/chat/transformation.py

@@ -17,6 +17,7 @@ from litellm.litellm_core_utils.prompt_templates.common_utils import (
     strip_litellm_response_only_fields,
     unpack_legacy_defs,
 )
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.secret_managers.main import get_secret_str
 from litellm.types.llms.openai import (
     AllMessageValues,
@@ -826,7 +827,7 @@ class FireworksAIConfig(FireworksAIMixin, OpenAIGPTConfig):
         base = api_base.rstrip("/")
         base = base.removesuffix("/v1")
         response: Final = litellm.module_level_client.get(
-            url=f"{base}/v1/accounts/{account_id}/models",
+            url=join_api_base(base, f"v1/accounts/{account_id}/models"),
             headers={"Authorization": f"Bearer {api_key}"},
         )
 

litellm/llms/fireworks_ai/rerank/transformation.py

@@ -12,6 +12,7 @@ from pydantic import ConfigDict, TypeAdapter
 
 from litellm._uuid import uuid
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.rerank.transformation import BaseRerankConfig
 from litellm.llms.fireworks_ai.common_utils import FireworksAIMixin
 from litellm.types.llms.base import LiteLLMBaseModel
@@ -61,9 +62,9 @@ class FireworksAIRerankConfig(FireworksAIMixin, BaseRerankConfig):
             api_base = api_base.rstrip("/")
             if not api_base.endswith("/rerank"):
                 if api_base.endswith("/v1") or api_base.endswith("/inference/v1"):
-                    api_base = f"{api_base}/rerank"
+                    api_base = join_api_base(api_base, "rerank")
                 else:
-                    api_base = f"{api_base}/inference/v1/rerank"
+                    api_base = join_api_base(api_base, "inference/v1/rerank")
             return api_base
         return "https://api.fireworks.ai/inference/v1/rerank"