Merge branch 'litellm_internal_staging' into litellm_auto-close-low-quality-prs-1f26

Pulls in #28191 (ce87c411bf) which migrates realtime and rerank tests off the
shut-down upstream models that were failing this PR's CI:
- gpt-4o-realtime-preview-* (OpenAI EOL 2026-05-07)
- nvidia/llama-3.2-nv-rerankqa-1b-v2 (NVIDIA EOL 2026-05-18, HTTP 410)

README.md

@@ -292,7 +292,7 @@ curl -X POST 'http://0.0.0.0:4000/v1/chat/completions' \
 | [CompactifAI (`compactifai`)](https://docs.litellm.ai/docs/providers/compactifai) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [Custom (`custom`)](https://docs.litellm.ai/docs/providers/custom_llm_server) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [Custom OpenAI (`custom_openai`)](https://docs.litellm.ai/docs/providers/openai_compatible) | ✅ | ✅ | ✅ |  |  | ✅ | ✅ | ✅ | ✅ |  |
-| [Dashscope (`dashscope`)](https://docs.litellm.ai/docs/providers/dashscope) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
+| [Dashscope (`dashscope`)](https://docs.litellm.ai/docs/providers/dashscope) | ✅ | ✅ | ✅ | ✅ |  |  |  |  |  | ✅ |
 | [Databricks (`databricks`)](https://docs.litellm.ai/docs/providers/databricks) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [DataRobot (`datarobot`)](https://docs.litellm.ai/docs/providers/datarobot) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [Deepgram (`deepgram`)](https://docs.litellm.ai/docs/providers/deepgram) | ✅ | ✅ | ✅ |  |  | ✅ |  |  |  |  |

deploy/charts/litellm-helm/templates/hpa.yaml

@@ -12,6 +12,10 @@ spec:
     name: {{ include "litellm.fullname" . }}
   minReplicas: {{ .Values.autoscaling.minReplicas }}
   maxReplicas: {{ .Values.autoscaling.maxReplicas }}
+  {{- if .Values.autoscaling.behavior }}
+  behavior:
+    {{- toYaml .Values.autoscaling.behavior | nindent 4 }}
+  {{- end }}
   metrics:
     {{- if .Values.autoscaling.targetCPUUtilizationPercentage }}
     - type: Resource

deploy/charts/litellm-helm/tests/hpa_tests.yaml

@@ -0,0 +1,36 @@
+suite: "hpa with behavior"
+templates:
+  - hpa.yaml
+tests:
+  - it: "renders behavior when set"
+    set:
+      autoscaling.enabled: true
+      autoscaling.behavior:
+        scaleUp:
+          stabilizationWindowSeconds: 60
+          policies:
+            - type: Pods
+              value: 2
+              periodSeconds: 60
+        scaleDown:
+          stabilizationWindowSeconds: 90
+          policies:
+            - type: Pods
+              value: 1
+              periodSeconds: 60
+    asserts:
+      - isKind: { of: HorizontalPodAutoscaler }
+      - equal: { path: spec.behavior.scaleUp.stabilizationWindowSeconds, value: 60 }
+      - equal: { path: spec.behavior.scaleDown.stabilizationWindowSeconds, value: 90 }
+
+---
+suite: "hpa without behavior"
+templates:
+  - hpa.yaml
+tests:
+  - it: "does not render behavior when not set"
+    set:
+      autoscaling.enabled: true
+    asserts:
+      - isKind: { of: HorizontalPodAutoscaler }
+      - isNull: { path: spec.behavior }

deploy/charts/litellm-helm/values.yaml

@@ -184,6 +184,7 @@ autoscaling:
   maxReplicas: 100
   targetCPUUtilizationPercentage: 80
   # targetMemoryUtilizationPercentage: 80
+  # behavior: {}
 
 # Autoscaling with keda is mutually exclusive with hpa
 keda:

litellm-proxy-extras/litellm_proxy_extras/migrations/20260514120000_add_blocked_to_proxy_model_table/migration.sql

@@ -0,0 +1,4 @@
+-- AlterTable
+-- Adds the admin-toggleable pause flag used by the router's blocked filter and the
+-- credential lookup helpers; defaults to false so existing rows behave unchanged.
+ALTER TABLE "LiteLLM_ProxyModelTable" ADD COLUMN IF NOT EXISTS "blocked" BOOLEAN NOT NULL DEFAULT false;

litellm-proxy-extras/litellm_proxy_extras/schema.prisma

@@ -48,9 +48,10 @@ model LiteLLM_CredentialsTable {
 // Models on proxy
 model LiteLLM_ProxyModelTable {
   model_id String @id @default(uuid())
-  model_name String 
+  model_name String
   litellm_params Json
-  model_info Json? 
+  model_info Json?
+  blocked Boolean @default(false)
   created_at    DateTime               @default(now()) @map("created_at")
   created_by String
   updated_at    DateTime               @default(now()) @updatedAt @map("updated_at")

litellm/__init__.py

@@ -416,6 +416,7 @@ custom_prometheus_metadata_labels: List[str] = []
 custom_prometheus_tags: List[str] = []
 prometheus_metrics_config: Optional[List] = None
 prometheus_emit_stream_label: bool = False
+prometheus_user_budget_label_include_email_alias: bool = False
 prometheus_end_user_metrics_max_series_per_metric: Optional[int] = 10000
 prometheus_end_user_metrics_ttl_seconds: Optional[float] = 3600.0
 prometheus_end_user_metrics_cleanup_interval_seconds: Optional[float] = 60.0
@@ -1880,6 +1881,12 @@ if TYPE_CHECKING:
     from .llms.dashscope.chat.transformation import (
         DashScopeChatConfig as DashScopeChatConfig,
     )
+    from .llms.dashscope.embed.transformation import (
+        DashScopeEmbeddingConfig as DashScopeEmbeddingConfig,
+    )
+    from .llms.dashscope.rerank.transformation import (
+        DashScopeRerankConfig as DashScopeRerankConfig,
+    )
     from .llms.moonshot.chat.transformation import (
         MoonshotChatConfig as MoonshotChatConfig,
     )

litellm/_redis.py

@@ -100,6 +100,8 @@ def _get_redis_cluster_kwargs(client=None):
         "azure_tenant_id",
         "azure_client_secret",
         "max_connections",
+        "socket_timeout",
+        "socket_connect_timeout",
     }
 
     return available_args

litellm/caching/caching_handler.py

@@ -87,6 +87,16 @@ class CachingHandlerResponse(BaseModel):
 in_memory_cache_obj = InMemoryCache()
 
 
+def _is_chat_completion_cached_dict(cached_result: dict) -> bool:
+    cached_id = cached_result.get("id")
+    if isinstance(cached_id, str) and cached_id.startswith("chatcmpl"):
+        return True
+    obj = cached_result.get("object")
+    if isinstance(obj, str):
+        return obj.startswith("chat.completion")
+    return "choices" in cached_result
+
+
 def _should_defer_streaming_cache_hit_callbacks(*, kwargs: Dict[str, Any]) -> bool:
     """
     When stream=True, do not run success callbacks at cache-hit time.
@@ -861,27 +871,47 @@ class LLMCachingHandler:
         elif (call_type == "aresponses" or call_type == "responses") and isinstance(
             cached_result, dict
         ):
-            from litellm.responses.streaming_iterator import (
-                CachedResponsesAPIStreamingIterator,
-            )
-
-            response_obj = ResponsesAPIResponse(**cached_result)
-            if (
-                hasattr(response_obj, "_hidden_params")
-                and response_obj._hidden_params is not None
-                and isinstance(response_obj._hidden_params, dict)
-            ):
-                response_obj._hidden_params["cache_hit"] = True
-
-            if kwargs.get("stream", False) is True:
-                cached_result = CachedResponsesAPIStreamingIterator(
-                    response=response_obj,
-                    logging_obj=logging_obj,
-                    request_data=kwargs,
-                    call_type=call_type,
-                )
+            use_chat_completion_cache = _is_chat_completion_cached_dict(cached_result)
+            if use_chat_completion_cache:
+                if kwargs.get("stream", False) is True:
+                    bridge_call_type = (
+                        CallTypes.acompletion.value
+                        if call_type == "aresponses"
+                        else CallTypes.completion.value
+                    )
+                    cached_result = self._convert_cached_stream_response(
+                        cached_result=cached_result,
+                        call_type=bridge_call_type,
+                        logging_obj=logging_obj,
+                        model=model,
+                    )
+                else:
+                    cached_result = convert_to_model_response_object(
+                        response_object=cached_result,
+                        model_response_object=ModelResponse(),
+                    )
             else:
-                cached_result = response_obj
+                from litellm.responses.streaming_iterator import (
+                    CachedResponsesAPIStreamingIterator,
+                )
+
+                response_obj = ResponsesAPIResponse(**cached_result)
+                if (
+                    hasattr(response_obj, "_hidden_params")
+                    and response_obj._hidden_params is not None
+                    and isinstance(response_obj._hidden_params, dict)
+                ):
+                    response_obj._hidden_params["cache_hit"] = True
+
+                if kwargs.get("stream", False) is True:
+                    cached_result = CachedResponsesAPIStreamingIterator(
+                        response=response_obj,
+                        logging_obj=logging_obj,
+                        request_data=kwargs,
+                        call_type=call_type,
+                    )
+                else:
+                    cached_result = response_obj
 
         if (
             hasattr(cached_result, "_hidden_params")

litellm/completion_extras/litellm_responses_transformation/handler.py

@@ -37,6 +37,15 @@ class ResponsesToCompletionBridgeHandler:
             stream = litellm_params.get("stream", False)
         return bool(stream)
 
+    @staticmethod
+    def _is_preformatted_cached_chat_stream(result: Any) -> bool:
+        from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper
+
+        return (
+            isinstance(result, CustomStreamWrapper)
+            and result.custom_llm_provider == "cached_response"
+        )
+
     @staticmethod
     def _coerce_response_object(
         response_obj: Any,
@@ -177,6 +186,8 @@ class ResponsesToCompletionBridgeHandler:
             **request_data,
         )
 
+        from litellm.types.utils import ModelResponse
+
         stream = self._resolve_stream_flag(optional_params, litellm_params)
         if isinstance(result, ResponsesAPIResponse):
             return self.transformation_handler.transform_response(
@@ -192,6 +203,8 @@ class ResponsesToCompletionBridgeHandler:
                 api_key=kwargs.get("api_key"),
                 json_mode=kwargs.get("json_mode"),
             )
+        elif isinstance(result, ModelResponse):
+            return result
         elif not stream:
             responses_api_response = self._collect_response_from_stream(result)
             return self.transformation_handler.transform_response(
@@ -208,6 +221,10 @@ class ResponsesToCompletionBridgeHandler:
                 json_mode=kwargs.get("json_mode"),
             )
         else:
+            if self._is_preformatted_cached_chat_stream(result):
+                return self._apply_post_stream_processing(
+                    result, model, custom_llm_provider
+                )
             completion_stream = self.transformation_handler.get_model_response_iterator(
                 streaming_response=result,  # type: ignore
                 sync_stream=True,
@@ -256,6 +273,8 @@ class ResponsesToCompletionBridgeHandler:
             aresponses=True,
         )
 
+        from litellm.types.utils import ModelResponse
+
         stream = self._resolve_stream_flag(optional_params, litellm_params)
         if isinstance(result, ResponsesAPIResponse):
             return self.transformation_handler.transform_response(
@@ -271,6 +290,8 @@ class ResponsesToCompletionBridgeHandler:
                 api_key=kwargs.get("api_key"),
                 json_mode=kwargs.get("json_mode"),
             )
+        elif isinstance(result, ModelResponse):
+            return result
         elif not stream:
             responses_api_response = await self._collect_response_from_stream_async(
                 result
@@ -289,6 +310,10 @@ class ResponsesToCompletionBridgeHandler:
                 json_mode=kwargs.get("json_mode"),
             )
         else:
+            if self._is_preformatted_cached_chat_stream(result):
+                return self._apply_post_stream_processing(
+                    result, model, custom_llm_provider
+                )
             completion_stream = self.transformation_handler.get_model_response_iterator(
                 streaming_response=result,  # type: ignore
                 sync_stream=False,

litellm/completion_extras/litellm_responses_transformation/transformation.py

@@ -1141,6 +1141,14 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
         event_type = parsed_chunk.get("type")
         if isinstance(event_type, ResponsesAPIStreamEvents):
             event_type = event_type.value
+
+        if parsed_chunk.get("object") == "chat.completion.chunk" or (
+            event_type is None
+            and isinstance(parsed_chunk.get("choices"), list)
+            and parsed_chunk.get("choices")
+        ):
+            return ModelResponseStream(**parsed_chunk)
+
         verbose_logger.debug(f"Chat provider: Processing event type: {event_type}")
 
         if event_type == "response.created":

litellm/cost_calculator.py

@@ -173,17 +173,45 @@ def _cost_per_token_custom_pricing_helper(
     prompt_tokens: float = 0,
     completion_tokens: float = 0,
     response_time_ms: Optional[float] = 0.0,
+    cached_tokens: float = 0,
+    cache_creation_tokens: float = 0,
     ### CUSTOM PRICING ###
     custom_cost_per_token: Optional[CostPerToken] = None,
     custom_cost_per_second: Optional[float] = None,
 ) -> Optional[Tuple[float, float]]:
-    """Internal helper function for calculating cost, if custom pricing given"""
+    """Internal helper function for calculating cost, if custom pricing given.
+
+    prompt_tokens is assumed to include both cached_tokens and cache_creation_tokens
+    (OpenAI-compatible convention). Anthropic-style usage where prompt_tokens excludes
+    cache tokens is handled at the caller (cost_per_token) before invoking this helper.
+    """
     if custom_cost_per_token is None and custom_cost_per_second is None:
         return None
 
     if custom_cost_per_token is not None:
-        input_cost = custom_cost_per_token["input_cost_per_token"] * prompt_tokens
-        output_cost = custom_cost_per_token["output_cost_per_token"] * completion_tokens
+        input_cost_per_token = custom_cost_per_token["input_cost_per_token"]
+        output_cost_per_token = custom_cost_per_token["output_cost_per_token"]
+
+        cache_read_input_token_cost = custom_cost_per_token.get(
+            "cache_read_input_token_cost",
+            input_cost_per_token,
+        )
+        cache_creation_input_token_cost = custom_cost_per_token.get(
+            "cache_creation_input_token_cost",
+            input_cost_per_token,
+        )
+
+        regular_prompt_tokens = max(
+            prompt_tokens - cached_tokens - cache_creation_tokens,
+            0,
+        )
+
+        input_cost = (
+            regular_prompt_tokens * input_cost_per_token
+            + cached_tokens * cache_read_input_token_cost
+            + cache_creation_tokens * cache_creation_input_token_cost
+        )
+        output_cost = completion_tokens * output_cost_per_token
         return input_cost, output_cost
     elif custom_cost_per_second is not None:
         output_cost = custom_cost_per_second * response_time_ms / 1000  # type: ignore
@@ -323,10 +351,56 @@ def cost_per_token(  # noqa: PLR0915
         )
 
     ## CUSTOM PRICING ##
+    # Normalize cache token counts across providers:
+    #   - OpenAI-compatible: usage.prompt_tokens_details.cached_tokens
+    #     (prompt_tokens already INCLUDES cached_tokens)
+    #   - Anthropic: usage.cache_read_input_tokens / cache_creation_input_tokens
+    #     (prompt_tokens does NOT include these — adjust before calling helper)
+    _cache_read_tokens: float = 0
+    _cache_creation_tokens: float = 0
+    _is_anthropic_style = False
+
+    if usage_object is not None:
+        _pt_details = getattr(usage_object, "prompt_tokens_details", None)
+        if _pt_details is not None:
+            _cache_read_tokens = float(getattr(_pt_details, "cached_tokens", 0) or 0)
+            # OpenAI-compatible providers report cache-write tokens under
+            # either `cache_write_tokens` (kimi-k2) or `cache_creation_tokens`.
+            # Mirror db_spend_update_writer to stay symmetric.
+            _cache_creation_tokens = float(
+                getattr(_pt_details, "cache_write_tokens", 0)
+                or getattr(_pt_details, "cache_creation_tokens", 0)
+                or 0
+            )
+
+        _anthropic_read = getattr(usage_object, "cache_read_input_tokens", None)
+        _anthropic_create = getattr(usage_object, "cache_creation_input_tokens", None)
+        if _anthropic_read is not None or _anthropic_create is not None:
+            _is_anthropic_style = True
+            if _anthropic_read is not None:
+                _cache_read_tokens = float(_anthropic_read)
+            if _anthropic_create is not None:
+                _cache_creation_tokens = float(_anthropic_create)
+
+    if not _cache_read_tokens and cache_read_input_tokens:
+        _cache_read_tokens = float(cache_read_input_tokens)
+        _is_anthropic_style = True
+    if not _cache_creation_tokens and cache_creation_input_tokens:
+        _cache_creation_tokens = float(cache_creation_input_tokens)
+        _is_anthropic_style = True
+
+    # Anthropic reports prompt_tokens as input_tokens (excluding cache tokens).
+    # Adjust so the helper's "prompt_tokens includes cache tokens" invariant holds.
+    _normalized_prompt_tokens = float(prompt_tokens)
+    if _is_anthropic_style:
+        _normalized_prompt_tokens += _cache_read_tokens + _cache_creation_tokens
+
     response_cost = _cost_per_token_custom_pricing_helper(
-        prompt_tokens=prompt_tokens,
+        prompt_tokens=_normalized_prompt_tokens,
         completion_tokens=completion_tokens,
         response_time_ms=response_time_ms,
+        cached_tokens=_cache_read_tokens,
+        cache_creation_tokens=_cache_creation_tokens,
         custom_cost_per_second=custom_cost_per_second,
         custom_cost_per_token=custom_cost_per_token,
     )

litellm/exceptions.py

@@ -918,9 +918,11 @@ class GuardrailRaisedException(Exception):
         guardrail_name: Optional[str] = None,
         message: str = "",
         should_wrap_with_default_message: bool = True,
+        status_code: int = 400,
     ):
         default_message = f"Guardrail raised an exception, Guardrail: {guardrail_name}, Message: {message}"
         self.guardrail_name = guardrail_name
+        self.status_code = status_code
         self.message = default_message if should_wrap_with_default_message else message
         super().__init__(self.message)
 
@@ -930,12 +932,14 @@ class BlockedPiiEntityError(Exception):
         self,
         entity_type: str,
         guardrail_name: Optional[str] = None,
+        status_code: int = 400,
     ):
         """
         Raised when a blocked entity is detected by a guardrail.
         """
         self.entity_type = entity_type
         self.guardrail_name = guardrail_name
+        self.status_code = status_code
         self.message = f"Blocked entity detected: {entity_type} by Guardrail: {guardrail_name}. This entity is not allowed to be used in this request."
         super().__init__(self.message)
 

litellm/integrations/custom_guardrail.py

@@ -43,7 +43,11 @@ if TYPE_CHECKING:
 dc = DualCache()
 
 
-from litellm.exceptions import ModifyResponseException as ModifyResponseException
+from litellm.exceptions import (
+    BlockedPiiEntityError,
+    GuardrailRaisedException,
+    ModifyResponseException,
+)
 
 
 class CustomGuardrail(CustomLogger):
@@ -737,12 +741,15 @@ class CustomGuardrail(CustomLogger):
         (this was logged previously as an API failure - guardrail_failed_to_respond).
 
         Guardrails signal intentional blocks by raising:
+        - GuardrailRaisedException (generic guardrail API, tool permission)
+        - BlockedPiiEntityError (Presidio PII detection)
         - HTTPException with status 400 (content policy violation)
         - ModifyResponseException (passthrough mode violation)
         """
-
         if isinstance(e, ModifyResponseException):
             return True
+        if isinstance(e, (GuardrailRaisedException, BlockedPiiEntityError)):
+            return True
         if (
             HTTPException is not None
             and isinstance(e, HTTPException)

litellm/integrations/opentelemetry.py

@@ -1107,8 +1107,13 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
             "mcp_tool_call_metadata",
             "vector_store_request_metadata",
         ]:
-            if md.get(key) is not None:
-                common_attrs[f"metadata.{key}"] = str(md[key])
+            value = md.get(key)
+            if value is None:
+                continue
+            if isinstance(value, (dict, list)):
+                common_attrs[f"metadata.{key}"] = safe_dumps(value)
+            else:
+                common_attrs[f"metadata.{key}"] = str(value)
 
         # get hidden params
         hidden_params = getattr(std_log, "hidden_params", None) or (std_log or {}).get(

litellm/integrations/prometheus.py

@@ -3540,6 +3540,10 @@ class PrometheusLogger(CustomLogger):
             user_object.budget_reset_at = user_info.budget_reset_at
             if user_object.max_budget is None and user_info.max_budget is not None:
                 user_object.max_budget = user_info.max_budget
+            if user_info.user_email is not None:
+                user_object.user_email = user_info.user_email
+            if user_info.user_alias is not None:
+                user_object.user_alias = user_info.user_alias
 
         return user_object
 
@@ -3556,6 +3560,8 @@ class PrometheusLogger(CustomLogger):
         """
         enum_values = UserAPIKeyLabelValues(
             user=user.user_id,
+            user_email=user.user_email or "",
+            user_alias=user.user_alias or "",
         )
 
         _labels = prometheus_label_factory(

litellm/interactions/litellm_responses_transformation/streaming_iterator.py

@@ -2,7 +2,7 @@
 Streaming iterator for transforming Responses API stream to Interactions API stream.
 """
 
-from typing import Any, AsyncIterator, Dict, Iterator, Optional, cast
+from typing import Any, AsyncIterator, Dict, Iterator, List, Optional, cast
 
 from litellm.responses.streaming_iterator import (
     BaseResponsesAPIStreamingIterator,
@@ -15,6 +15,7 @@ from litellm.types.interactions import (
     InteractionsAPIStreamingResponse,
 )
 from litellm.types.llms.openai import (
+    ContentPartAddedEvent,
     OutputTextDeltaEvent,
     ResponseCompletedEvent,
     ResponseCreatedEvent,
@@ -51,6 +52,7 @@ class LiteLLMResponsesInteractionsStreamingIterator:
         self.collected_text = ""
         self.sent_interaction_start = False
         self.sent_content_start = False
+        self._pending_events: List[InteractionsAPIStreamingResponse] = []
 
     def _transform_responses_chunk_to_interactions_chunk(
         self,
@@ -80,7 +82,49 @@ class LiteLLMResponsesInteractionsStreamingIterator:
             )
             self.collected_text += delta_text
 
-            # Send interaction.start if not sent
+            # Fallback: emit interaction.start, and queue content.start carrying this
+            # delta so the first token is preserved in the stream.
+            if not self.sent_interaction_start:
+                self.sent_interaction_start = True
+                self.sent_content_start = True
+                self._pending_events.append(
+                    InteractionsAPIStreamingResponse(
+                        event_type="content.start",
+                        id=getattr(responses_chunk, "item_id", None),
+                        object="content",
+                        delta={"type": "text", "text": delta_text},
+                    )
+                )
+                return InteractionsAPIStreamingResponse(
+                    event_type="interaction.start",
+                    id=getattr(responses_chunk, "item_id", None)
+                    or f"interaction_{id(self)}",
+                    object="interaction",
+                    status="in_progress",
+                    model=self.model,
+                )
+
+            # Fallback: emit content.start if ContentPartAddedEvent never arrived
+            if not self.sent_content_start:
+                self.sent_content_start = True
+                return InteractionsAPIStreamingResponse(
+                    event_type="content.start",
+                    id=getattr(responses_chunk, "item_id", None),
+                    object="content",
+                    delta={"type": "text", "text": delta_text},
+                )
+
+            # Normal path: emit content.delta with type field
+            return InteractionsAPIStreamingResponse(
+                event_type="content.delta",
+                id=getattr(responses_chunk, "item_id", None),
+                object="content",
+                delta={"type": "text", "text": delta_text},
+            )
+
+        # Handle ContentPartAddedEvent -> content.start (arrives before text deltas)
+        if isinstance(responses_chunk, ContentPartAddedEvent):
+            # Fallback: emit interaction.start if ResponseCreatedEvent never arrived
             if not self.sent_interaction_start:
                 self.sent_interaction_start = True
                 return InteractionsAPIStreamingResponse(
@@ -91,8 +135,6 @@ class LiteLLMResponsesInteractionsStreamingIterator:
                     status="in_progress",
                     model=self.model,
                 )
-
-            # Send content.start if not sent
             if not self.sent_content_start:
                 self.sent_content_start = True
                 return InteractionsAPIStreamingResponse(
@@ -101,14 +143,7 @@ class LiteLLMResponsesInteractionsStreamingIterator:
                     object="content",
                     delta={"type": "text", "text": ""},
                 )
-
-            # Send content.delta
-            return InteractionsAPIStreamingResponse(
-                event_type="content.delta",
-                id=getattr(responses_chunk, "item_id", None),
-                object="content",
-                delta={"text": delta_text},
-            )
+            return None
 
         # Handle ResponseCreatedEvent or ResponseInProgressEvent -> interaction.start
         if isinstance(responses_chunk, (ResponseCreatedEvent, ResponseInProgressEvent)):
@@ -172,6 +207,10 @@ class LiteLLMResponsesInteractionsStreamingIterator:
             delattr(self, "_pending_interaction_complete")
             return pending
 
+        # Drain events queued from a prior chunk (e.g. content.start emitted alongside
+        # the interaction.start fallback for the first OutputTextDeltaEvent).
+        if self._pending_events:
+            return self._pending_events.pop(0)
         # Use a loop instead of recursion to avoid stack overflow
         sync_iterator = cast(
             SyncResponsesAPIStreamingIterator, self.responses_stream_iterator
@@ -237,6 +276,10 @@ class LiteLLMResponsesInteractionsStreamingIterator:
             delattr(self, "_pending_interaction_complete")
             return pending
 
+        # Drain events queued from a prior chunk (e.g. content.start emitted alongside
+        # the interaction.start fallback for the first OutputTextDeltaEvent).
+        if self._pending_events:
+            return self._pending_events.pop(0)
         # Use a loop instead of recursion to avoid stack overflow
         async_iterator = cast(
             ResponsesAPIStreamingIterator, self.responses_stream_iterator

litellm/litellm_core_utils/prompt_templates/common_utils.py

@@ -20,6 +20,7 @@ from typing import (
     cast,
 )
 
+import litellm
 from litellm import verbose_logger
 from litellm.router_utils.batch_utils import InMemoryFile
 from litellm.types.llms.openai import (
@@ -1170,9 +1171,16 @@ def migrate_file_to_image_url(
         ChatCompletionImageUrlObject,
     )
 
-    file_id = message["file"].get("file_id")
-    file_data = message["file"].get("file_data")
-    format = message["file"].get("format")
+    file_sub = message.get("file")
+    if file_sub is None:
+        raise litellm.BadRequestError(
+            message="Content block has type='file' but is missing the required 'file' field",
+            model=None,
+            llm_provider=None,
+        )
+    file_id = file_sub.get("file_id")
+    file_data = file_sub.get("file_data")
+    format = file_sub.get("format")
     if not file_id and not file_data:
         raise ValueError("file_id and file_data are both None")
     image_url_object = ChatCompletionImageObject(

litellm/litellm_core_utils/prompt_templates/factory.py

@@ -2057,9 +2057,16 @@ def anthropic_process_openai_file_message(
     AnthropicMessagesContainerUploadParam,
 ]:
     file_message = cast(ChatCompletionFileObject, message)
-    file_data = file_message["file"].get("file_data")
-    file_id = file_message["file"].get("file_id")
-    format = file_message["file"].get("format")
+    file_sub = file_message.get("file")
+    if file_sub is None:
+        raise litellm.BadRequestError(
+            message="Content block has type='file' but is missing the required 'file' field",
+            model=None,
+            llm_provider="anthropic",
+        )
+    file_data = file_sub.get("file_data")
+    file_id = file_sub.get("file_id")
+    format = file_sub.get("format")
     if file_data:
         image_chunk = convert_to_anthropic_image_obj(
             openai_image_url=file_data,
@@ -4879,7 +4886,13 @@ class BedrockConverseMessagesProcessor:
 
     @staticmethod
     def _process_file_message(message: ChatCompletionFileObject) -> BedrockContentBlock:
-        file_message = message["file"]
+        file_message = message.get("file")
+        if file_message is None:
+            raise litellm.BadRequestError(
+                message="Content block has type='file' but is missing the required 'file' field",
+                model=None,
+                llm_provider="bedrock",
+            )
         file_data = file_message.get("file_data")
         file_id = file_message.get("file_id")
 
@@ -4900,7 +4913,13 @@ class BedrockConverseMessagesProcessor:
     async def _async_process_file_message(
         message: ChatCompletionFileObject,
     ) -> BedrockContentBlock:
-        file_message = message["file"]
+        file_message = message.get("file")
+        if file_message is None:
+            raise litellm.BadRequestError(
+                message="Content block has type='file' but is missing the required 'file' field",
+                model=None,
+                llm_provider="bedrock",
+            )
         file_data = file_message.get("file_data")
         file_id = file_message.get("file_id")
         format = file_message.get("format")

litellm/llms/bedrock/batches/transformation.py

@@ -5,6 +5,7 @@ from typing import Any, Dict, List, Literal, Optional, Union, cast
 
 from httpx import Headers, Response
 
+from litellm.litellm_core_utils.safe_json_dumps import safe_dumps
 from litellm.llms.base_llm.batches.transformation import BaseBatchesConfig
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.secret_managers.main import get_secret_str
@@ -263,9 +264,32 @@ class BedrockBatchesConfig(BaseAWSLLM, BaseBatchesConfig):
             cancelling_at=None,
             cancelled_at=None,
             request_counts=None,
-            metadata=original_request.get("metadata", {}),
+            metadata=self._get_openai_compatible_batch_metadata(
+                original_request.get("metadata", {})
+            ),
         )
 
+    @staticmethod
+    def _get_openai_compatible_batch_metadata(metadata: Any) -> Dict[str, str]:
+        """
+        OpenAI Batch metadata only accepts string values.
+        """
+        if not isinstance(metadata, dict):
+            return {}
+
+        sanitized_metadata: Dict[str, str] = {}
+        for key, value in metadata.items():
+            if key == "standard_logging_guardrail_information" or value is None:
+                continue
+
+            str_key = str(key)
+            if isinstance(value, str):
+                sanitized_metadata[str_key] = value
+            else:
+                sanitized_metadata[str_key] = safe_dumps(value)
+
+        return sanitized_metadata
+
     def transform_retrieve_batch_request(
         self,
         batch_id: str,

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

@@ -299,9 +299,9 @@ class AmazonInvokeAgentConfig(BaseConfig, BaseAWSLLM):
             )
 
     def _get_response_stream_shape(self):
-        from litellm.llms.bedrock.common_utils import BEDROCK_RESPONSE_STREAM_SHAPE
+        from litellm.llms.bedrock.common_utils import get_bedrock_response_stream_shape
 
-        return BEDROCK_RESPONSE_STREAM_SHAPE
+        return get_bedrock_response_stream_shape()
 
     def _extract_response_content(self, events: InvokeAgentEventList) -> str:
         """Extract the final response content from parsed events."""

litellm/llms/bedrock/chat/invoke_handler.py

@@ -68,9 +68,9 @@ from litellm.utils import CustomStreamWrapper, get_secret
 
 from ..base_aws_llm import BaseAWSLLM
 from ..common_utils import (
-    BEDROCK_RESPONSE_STREAM_SHAPE,
     BedrockError,
     ModelResponseIterator,
+    get_bedrock_response_stream_shape,
     get_bedrock_tool_name,
 )
 
@@ -1828,7 +1828,8 @@ class AWSEventStreamDecoder:
                     yield self._chunk_parser(chunk_data=_data)
 
     def _parse_message_from_event(self, event) -> Optional[str]:
-        if BEDROCK_RESPONSE_STREAM_SHAPE is None:
+        response_stream_shape = get_bedrock_response_stream_shape()
+        if response_stream_shape is None:
             raise BedrockError(
                 status_code=500,
                 message=(
@@ -1837,9 +1838,7 @@ class AWSEventStreamDecoder:
                 ),
             )
         response_dict = event.to_response_dict()
-        parsed_response = self.parser.parse(
-            response_dict, BEDROCK_RESPONSE_STREAM_SHAPE
-        )
+        parsed_response = self.parser.parse(response_dict, response_stream_shape)
 
         if response_dict["status_code"] != 200:
             decoded_body = response_dict["body"].decode()

litellm/llms/bedrock/common_utils.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 Common utilities used across bedrock chat/embedding/image generation
 """
 
+import functools
 import json
 import os
 from typing import TYPE_CHECKING, Any, Dict, List, Literal, Optional, Union
@@ -963,10 +964,8 @@ def _load_bedrock_response_stream_shape():
     """
     Load the ResponseStream shape from botocore's bundled bedrock-runtime schema.
 
-    Called once at module import time; the result is stored in
-    ``BEDROCK_RESPONSE_STREAM_SHAPE`` and reused for the process lifetime.
     Returns ``None`` if botocore is unavailable or the service model cannot be
-    loaded, so the module still imports cleanly.
+    loaded.
     """
     try:
         from botocore.loaders import Loader
@@ -977,15 +976,22 @@ def _load_bedrock_response_stream_shape():
         return ServiceModel(service_dict).shape_for("ResponseStream")
     except Exception as e:
         verbose_logger.warning(
-            "litellm: could not pre-load bedrock-runtime response stream shape "
+            "litellm: could not load bedrock-runtime response stream shape "
             "— Bedrock event-stream decoding will be unavailable. Error: %s",
             e,
         )
         return None
 
 
-# Eagerly resolved once per process — avoids per-instance or per-request disk I/O.
-BEDROCK_RESPONSE_STREAM_SHAPE = _load_bedrock_response_stream_shape()
+@functools.lru_cache(maxsize=1)
+def get_bedrock_response_stream_shape():
+    """
+    Lazily load and cache the bedrock-runtime ResponseStream shape for the process.
+
+    Avoids importing botocore (and logging warnings) unless Bedrock event-stream
+    decoding is actually needed.
+    """
+    return _load_bedrock_response_stream_shape()
 
 
 class BedrockEventStreamDecoderBase:
@@ -999,7 +1005,8 @@ class BedrockEventStreamDecoderBase:
         self.parser = EventStreamJSONParser()
 
     def _parse_message_from_event(self, event) -> Optional[str]:
-        if BEDROCK_RESPONSE_STREAM_SHAPE is None:
+        response_stream_shape = get_bedrock_response_stream_shape()
+        if response_stream_shape is None:
             raise BedrockError(
                 status_code=500,
                 message=(
@@ -1008,9 +1015,7 @@ class BedrockEventStreamDecoderBase:
                 ),
             )
         response_dict = event.to_response_dict()
-        parsed_response = self.parser.parse(
-            response_dict, BEDROCK_RESPONSE_STREAM_SHAPE
-        )
+        parsed_response = self.parser.parse(response_dict, response_stream_shape)
 
         if response_dict["status_code"] != 200:
             decoded_body = response_dict["body"].decode()

litellm/llms/bedrock/embed/cohere_transformation.py

@@ -22,7 +22,7 @@ class BedrockCohereEmbeddingConfig:
     ) -> dict:
         for k, v in non_default_params.items():
             if k == "encoding_format":
-                optional_params["embedding_types"] = v
+                optional_params["embedding_types"] = v if isinstance(v, list) else [v]
             elif k == "dimensions":
                 optional_params["output_dimension"] = v
         return optional_params

litellm/llms/dashscope/common_utils.py

@@ -0,0 +1,28 @@
+"""
+Common utilities for the DashScope LLM provider.
+"""
+
+from typing import Optional
+
+import httpx
+
+from litellm.llms.base_llm.chat.transformation import BaseLLMException
+
+
+class DashScopeError(BaseLLMException):
+    """Exception class for DashScope provider errors."""
+
+    def __init__(
+        self,
+        status_code: int,
+        message: str,
+        headers: Optional[httpx.Headers] = None,
+    ):
+        self.status_code = status_code
+        self.message = message
+        self.headers = headers or httpx.Headers()
+        super().__init__(
+            status_code=status_code,
+            message=message,
+            headers=dict(self.headers),
+        )

litellm/llms/dashscope/embed/__init__.py

@@ -0,0 +1,7 @@
+"""
+DashScope Embedding Module
+"""
+
+from .transformation import DashScopeEmbeddingConfig
+
+__all__ = ["DashScopeEmbeddingConfig"]

litellm/llms/dashscope/embed/transformation.py

@@ -0,0 +1,191 @@
+"""
+Transformation logic from OpenAI /v1/embeddings format to DashScope's /v1/embeddings format.
+
+Supports
+- text-embedding-v4
+- text-embedding-v3
+
+Endpoint
+- https://dashscope.aliyuncs.com/compatible-mode/v1/embeddings
+
+Docs - https://help.aliyun.com/zh/model-studio/text-embedding-synchronous-api
+"""
+
+from typing import List, Optional, Union
+
+import httpx
+
+from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+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
+from litellm.types.llms.openai import AllEmbeddingInputValues, AllMessageValues
+from litellm.types.utils import EmbeddingResponse, Usage
+
+from ..common_utils import DashScopeError
+
+DEFAULT_API_BASE = "https://dashscope.aliyuncs.com/compatible-mode/v1"
+
+
+class DashScopeEmbeddingConfig(BaseEmbeddingConfig):
+    """
+    Reference: https://help.aliyun.com/zh/model-studio/text-embedding-synchronous-api
+
+    DashScope exposes an OpenAI-compatible /v1/embeddings endpoint, so the
+    request and response shapes are nearly identical to OpenAI's.
+    """
+
+    def __init__(self) -> None:
+        pass
+
+    def get_supported_openai_params(self, model: str) -> List[str]:
+        # DashScope's compatible-mode embeddings API accepts the same params as OpenAI.
+        # `dimensions` / `encoding_format` are only honored by text-embedding-v3 / v4;
+        # earlier versions silently ignore them server-side.
+        return ["dimensions", "encoding_format", "user"]
+
+    def map_openai_params(
+        self,
+        non_default_params: dict,
+        optional_params: dict,
+        model: str,
+        drop_params: bool = False,
+    ) -> dict:
+        supported = self.get_supported_openai_params(model)
+        for k, v in non_default_params.items():
+            if v is None:
+                continue
+            if k in supported:
+                optional_params[k] = v
+            # unsupported params are dropped when drop_params=True;
+            # the upstream _check_valid_arg already raised UnsupportedParamsError
+            # for drop_params=False before this method is called.
+        return optional_params
+
+    def validate_environment(
+        self,
+        headers: dict,
+        model: str,
+        messages: List[AllMessageValues],
+        optional_params: dict,
+        litellm_params: dict,
+        api_key: Optional[str] = None,
+        api_base: Optional[str] = None,
+    ) -> dict:
+        if api_key is None:
+            api_key = get_secret_str("DASHSCOPE_API_KEY")
+        if api_key is None:
+            raise ValueError(
+                "DashScope API key is required. Set 'DASHSCOPE_API_KEY' env var or pass api_key explicitly."
+            )
+        default_headers = {
+            "Content-Type": "application/json",
+            "Authorization": f"Bearer {api_key}",
+        }
+        return {**default_headers, **headers}
+
+    def get_complete_url(
+        self,
+        api_base: Optional[str],
+        api_key: Optional[str],
+        model: str,
+        optional_params: dict,
+        litellm_params: dict,
+        stream: Optional[bool] = None,
+    ) -> str:
+        base = api_base or get_secret_str("DASHSCOPE_API_BASE") or DEFAULT_API_BASE
+        base = base.rstrip("/")
+        if base.endswith("/embeddings"):
+            return base
+        return f"{base}/embeddings"
+
+    def transform_embedding_request(
+        self,
+        model: str,
+        input: AllEmbeddingInputValues,
+        optional_params: dict,
+        headers: dict,
+    ) -> dict:
+        data: dict = {
+            "model": model,
+            "input": input,
+        }
+        for key in ("dimensions", "encoding_format", "user"):
+            value = optional_params.get(key)
+            if value is not None:
+                data[key] = value
+        return data
+
+    def transform_embedding_response(
+        self,
+        model: str,
+        raw_response: httpx.Response,
+        model_response: EmbeddingResponse,
+        logging_obj: LiteLLMLoggingObj,
+        api_key: Optional[str],
+        request_data: dict,
+        optional_params: dict,
+        litellm_params: dict,
+    ) -> EmbeddingResponse:
+        try:
+            response_json = raw_response.json()
+        except Exception as e:
+            raise DashScopeError(
+                status_code=raw_response.status_code,
+                message=f"Failed to parse DashScope response as JSON: {str(e)}",
+            )
+
+        logging_obj.post_call(
+            input=request_data.get("input"),
+            api_key=api_key,
+            additional_args={"complete_input_dict": request_data},
+            original_response=response_json,
+        )
+
+        if "error" in response_json:
+            error = response_json["error"]
+            message = (
+                error.get("message", str(error))
+                if isinstance(error, dict)
+                else str(error)
+            )
+            raise DashScopeError(
+                status_code=raw_response.status_code,
+                message=message,
+            )
+
+        model_response.object = "list"
+        model_response.data = response_json.get("data", [])
+        model_response.model = response_json.get("model", model)
+
+        usage = response_json.get("usage") or {}
+        prompt_tokens = usage.get("prompt_tokens", 0)
+        total_tokens = usage.get("total_tokens", prompt_tokens)
+        setattr(
+            model_response,
+            "usage",
+            Usage(
+                prompt_tokens=prompt_tokens,
+                completion_tokens=0,
+                total_tokens=total_tokens,
+            ),
+        )
+
+        if "id" in response_json:
+            setattr(model_response, "id", response_json["id"])
+
+        return model_response
+
+    def get_error_class(
+        self,
+        error_message: str,
+        status_code: int,
+        headers: Union[dict, httpx.Headers],
+    ) -> BaseLLMException:
+        if isinstance(headers, dict):
+            headers = httpx.Headers(headers)
+        return DashScopeError(
+            status_code=status_code,
+            message=error_message,
+            headers=headers,
+        )

litellm/llms/dashscope/rerank/__init__.py

@@ -0,0 +1,7 @@
+"""
+DashScope Rerank Module
+"""
+
+from .transformation import DashScopeRerankConfig
+
+__all__ = ["DashScopeRerankConfig"]

litellm/llms/dashscope/rerank/transformation.py

@@ -0,0 +1,241 @@
+"""
+Transformation logic for DashScope's OpenAI-compatible /v1/reranks API.
+
+Supports
+- qwen3-rerank
+
+(Other DashScope rerankers — gte-rerank-v2 / qwen3-vl-rerank — share the same
+endpoint but have not been validated against this transformer. Behavior with
+those models is undefined.)
+
+Endpoint
+- https://dashscope.aliyuncs.com/compatible-api/v1/reranks
+
+Note: chat/embed live under `/compatible-mode/v1/`, but DashScope's rerank
+route is exposed under `/compatible-api/v1/reranks` per the docs. Override
+with `DASHSCOPE_API_BASE_RERANK` to point at a different host or path.
+
+Empirically, qwen3-rerank accepts `return_documents=true` and echoes
+`results[].document.text` back, even though the public docs list the flag
+as supported only for gte-rerank-v2 / qwen3-vl-rerank.
+
+Docs - https://help.aliyun.com/zh/model-studio/text-rerank-api
+"""
+
+from typing import Any, Dict, List, Optional, Union
+
+import httpx
+
+from litellm._uuid import uuid
+from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+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
+from litellm.types.rerank import (
+    OptionalRerankParams,
+    RerankBilledUnits,
+    RerankResponse,
+    RerankResponseMeta,
+    RerankTokens,
+)
+
+from ..common_utils import DashScopeError
+
+DEFAULT_RERANK_URL = "https://dashscope.aliyuncs.com/compatible-api/v1/reranks"
+
+
+class DashScopeRerankConfig(BaseRerankConfig):
+    """
+    Reference: https://help.aliyun.com/zh/model-studio/text-rerank-api
+
+    Targets DashScope's qwen3-rerank model. Request fields: model, query,
+    documents, top_n, return_documents. Response: results[].index,
+    results[].relevance_score, optionally results[].document.text (when
+    return_documents=true), plus a top-level usage.total_tokens counter.
+    """
+
+    def __init__(self) -> None:
+        pass
+
+    def get_complete_url(
+        self,
+        api_base: Optional[str],
+        model: str,
+        optional_params: Optional[dict] = None,
+    ) -> str:
+        if api_base is None:
+            api_base = get_secret_str("DASHSCOPE_API_BASE_RERANK") or DEFAULT_RERANK_URL
+
+        if api_base == DEFAULT_RERANK_URL:
+            return DEFAULT_RERANK_URL
+
+        cleaned = api_base.rstrip("/")
+        if cleaned.endswith("/reranks") or cleaned.endswith("/rerank"):
+            return cleaned
+
+        if cleaned.endswith("/v1"):
+            return f"{cleaned}/reranks"
+
+        # Unknown base: append /reranks rather than silently ignoring the caller's api_base.
+        return f"{cleaned}/reranks"
+
+    def validate_environment(
+        self,
+        headers: dict,
+        model: str,
+        api_key: Optional[str] = None,
+        optional_params: Optional[dict] = None,
+    ) -> dict:
+        if api_key is None:
+            api_key = get_secret_str("DASHSCOPE_API_KEY")
+        if api_key is None:
+            raise ValueError(
+                "DashScope API key is required. Set 'DASHSCOPE_API_KEY' env var or pass api_key explicitly."
+            )
+
+        default_headers = {
+            "Authorization": f"Bearer {api_key}",
+            "accept": "application/json",
+            "content-type": "application/json",
+        }
+        return {**default_headers, **headers}
+
+    def get_supported_cohere_rerank_params(self, model: str) -> list:
+        return ["query", "documents", "top_n", "return_documents"]
+
+    def map_cohere_rerank_params(
+        self,
+        non_default_params: Optional[dict],
+        model: str,
+        drop_params: bool,
+        query: str,
+        documents: List[Union[str, Dict[str, Any]]],
+        custom_llm_provider: Optional[str] = None,
+        top_n: Optional[int] = None,
+        rank_fields: Optional[List[str]] = None,
+        return_documents: Optional[bool] = True,
+        max_chunks_per_doc: Optional[int] = None,
+        max_tokens_per_doc: Optional[int] = None,
+    ) -> Dict:
+        # qwen3-rerank accepts query/documents/top_n/return_documents. The
+        # rest (rank_fields, max_*_per_doc) are silently dropped.
+        params: OptionalRerankParams = OptionalRerankParams(
+            query=query,
+            documents=documents,
+        )
+        if top_n is not None:
+            params["top_n"] = top_n
+        if return_documents is not None:
+            params["return_documents"] = return_documents
+        return dict(params)
+
+    def transform_rerank_request(
+        self,
+        model: str,
+        optional_rerank_params: Dict,
+        headers: dict,
+        litellm_params: Optional[dict] = None,
+    ) -> dict:
+        if "query" not in optional_rerank_params:
+            raise ValueError("query is required for DashScope rerank")
+        if "documents" not in optional_rerank_params:
+            raise ValueError("documents is required for DashScope rerank")
+
+        request: Dict[str, Any] = {
+            "model": model,
+            "query": optional_rerank_params["query"],
+            "documents": optional_rerank_params["documents"],
+        }
+        if optional_rerank_params.get("top_n") is not None:
+            request["top_n"] = optional_rerank_params["top_n"]
+        if optional_rerank_params.get("return_documents") is not None:
+            request["return_documents"] = optional_rerank_params["return_documents"]
+        return request
+
+    def transform_rerank_response(
+        self,
+        model: str,
+        raw_response: httpx.Response,
+        model_response: RerankResponse,
+        logging_obj: LiteLLMLoggingObj,
+        api_key: Optional[str] = None,
+        request_data: Optional[dict] = None,
+        optional_params: Optional[dict] = None,
+        litellm_params: Optional[dict] = None,
+    ) -> RerankResponse:
+        request_data = request_data or {}
+        optional_params = optional_params or {}
+        litellm_params = litellm_params or {}
+        try:
+            response_json = raw_response.json()
+        except Exception:
+            raise DashScopeError(
+                status_code=raw_response.status_code,
+                message=raw_response.text,
+            )
+
+        logging_obj.post_call(
+            input=request_data.get("query"),
+            api_key=api_key,
+            additional_args={"complete_input_dict": request_data},
+            original_response=response_json,
+        )
+
+        # DashScope error envelope: {"code": "...", "message": "...", "request_id": "..."}
+        if "code" in response_json and "results" not in response_json:
+            raise DashScopeError(
+                status_code=raw_response.status_code,
+                message=response_json.get("message", str(response_json)),
+            )
+
+        results = response_json.get("results")
+        if results is None:
+            raise DashScopeError(
+                status_code=raw_response.status_code,
+                message=f"No results in DashScope rerank response: {response_json}",
+            )
+
+        # qwen3-rerank returns:
+        #   {"index": int, "relevance_score": float}
+        # plus, when return_documents=true was sent:
+        #   "document": {"text": "..."}
+        # which already matches LiteLLM's RerankResponseDocument shape.
+        transformed_results: List[dict] = []
+        for r in results:
+            item: Dict[str, Any] = {
+                "index": r["index"],
+                "relevance_score": r["relevance_score"],
+            }
+            doc = r.get("document")
+            if isinstance(doc, dict):
+                item["document"] = doc
+            elif isinstance(doc, str):
+                # Defensive: spec says dict, but normalize string-shaped echoes.
+                item["document"] = {"text": doc}
+            transformed_results.append(item)
+
+        usage = response_json.get("usage") or {}
+        total_tokens = usage.get("total_tokens")
+        billed_units = RerankBilledUnits(total_tokens=total_tokens)
+        tokens = RerankTokens(input_tokens=total_tokens)
+        meta = RerankResponseMeta(billed_units=billed_units, tokens=tokens)
+
+        return RerankResponse(
+            id=response_json.get("id") or str(uuid.uuid4()),
+            results=transformed_results,  # type: ignore
+            meta=meta,
+        )
+
+    def get_error_class(
+        self,
+        error_message: str,
+        status_code: int,
+        headers: Union[dict, httpx.Headers],
+    ) -> BaseLLMException:
+        if isinstance(headers, dict):
+            headers = httpx.Headers(headers)
+        return DashScopeError(
+            status_code=status_code,
+            message=error_message,
+            headers=headers,
+        )

litellm/llms/deepseek/chat/transformation.py

@@ -2,13 +2,15 @@
 Translates from OpenAI's `/v1/chat/completions` to DeepSeek's `/v1/chat/completions`
 """
 
-from typing import Any, Coroutine, List, Literal, Optional, Tuple, Union, overload
+from typing import Any, Coroutine, List, Literal, Optional, Tuple, Union, cast, overload
 
+import litellm
 from litellm.litellm_core_utils.prompt_templates.common_utils import (
     handle_messages_with_content_list_to_str_conversion,
 )
 from litellm.secret_managers.main import get_secret_str
 from litellm.types.llms.openai import AllMessageValues
+from litellm.utils import supports_reasoning
 
 from ...openai.chat.gpt_transformation import OpenAIGPTConfig
 
@@ -62,6 +64,48 @@ class DeepSeekChatConfig(OpenAIGPTConfig):
 
         return optional_params
 
+    def _fill_reasoning_content(
+        self, messages: List[AllMessageValues]
+    ) -> List[AllMessageValues]:
+        """
+        DeepSeek thinking mode requires `reasoning_content` to be passed back on
+        every assistant message in multi-turn conversations. If it is missing,
+        the API returns:
+          "The reasoning_content in the thinking mode must be passed back to the API."
+
+        For each assistant message that is missing `reasoning_content`:
+          1. Promote it from `provider_specific_fields["reasoning_content"]` if present
+             (LiteLLM stores provider-specific response fields there).
+          2. Otherwise inject a single space — the minimum value the API accepts.
+        """
+        result: List[AllMessageValues] = []
+        for msg in messages:
+            if msg.get("role") == "assistant" and not msg.get("reasoning_content"):
+                patched = dict(cast(dict, msg))
+                provider_fields = patched.get("provider_specific_fields") or {}
+                stored = provider_fields.get("reasoning_content")
+                if stored:
+                    patched["reasoning_content"] = stored
+                    cleaned = dict(provider_fields)
+                    cleaned.pop("reasoning_content", None)
+                    patched["provider_specific_fields"] = cleaned
+                else:
+                    litellm.verbose_logger.warning(
+                        "DeepSeek thinking mode: assistant message is missing "
+                        "`reasoning_content` and none was saved in "
+                        "`provider_specific_fields`. A single-space placeholder "
+                        "is being injected to satisfy API validation, but the "
+                        "model will receive a blank reasoning chain for this turn, "
+                        "which may silently degrade multi-turn response quality. "
+                        "Preserve `reasoning_content` from the original assistant "
+                        "response when building multi-turn conversation history."
+                    )
+                    patched["reasoning_content"] = " "
+                result.append(cast(AllMessageValues, patched))
+            else:
+                result.append(msg)
+        return result
+
     @overload
     def _transform_messages(
         self, messages: List[AllMessageValues], model: str, is_async: Literal[True]
@@ -91,6 +135,66 @@ class DeepSeekChatConfig(OpenAIGPTConfig):
                 messages=messages, model=model, is_async=False
             )
 
+    def _thinking_mode_active(self, model: str, optional_params: dict) -> bool:
+        """
+        Returns True only when thinking mode is actually active for this request:
+          - model supports reasoning (capability check)
+          - user explicitly passed thinking={"type": "enabled"} (opt-in check)
+        """
+        return (
+            supports_reasoning(model=model, custom_llm_provider="deepseek")
+            and (optional_params.get("thinking") or {}).get("type") == "enabled"
+        )
+
+    def transform_request(
+        self,
+        model: str,
+        messages: List[AllMessageValues],
+        optional_params: dict,
+        litellm_params: dict,
+        headers: dict,
+    ) -> dict:
+        """
+        Ensures `reasoning_content` is forwarded on assistant messages for
+        multi-turn thinking-mode conversations (issue #28045).
+
+        Only runs when thinking mode is actually active - guarded by both
+        supports_reasoning() (model capability) and optional_params["thinking"]
+        (user explicitly enabled it), preventing spurious injection on models
+        like deepseek-v3.2 that support thinking as opt-in but not always-on.
+        """
+        if self._thinking_mode_active(model=model, optional_params=optional_params):
+            messages = self._fill_reasoning_content(messages)
+        return super().transform_request(
+            model=model,
+            messages=messages,
+            optional_params=optional_params,
+            litellm_params=litellm_params,
+            headers=headers,
+        )
+
+    async def async_transform_request(
+        self,
+        model: str,
+        messages: List[AllMessageValues],
+        optional_params: dict,
+        litellm_params: dict,
+        headers: dict,
+    ) -> dict:
+        """
+        Async equivalent of transform_request — applies the same reasoning_content
+        fix for multi-turn thinking-mode conversations.
+        """
+        if self._thinking_mode_active(model=model, optional_params=optional_params):
+            messages = self._fill_reasoning_content(messages)
+        return await super().async_transform_request(
+            model=model,
+            messages=messages,
+            optional_params=optional_params,
+            litellm_params=litellm_params,
+            headers=headers,
+        )
+
     def _get_openai_compatible_provider_info(
         self, api_base: Optional[str], api_key: Optional[str]
     ) -> Tuple[Optional[str], Optional[str]]:

litellm/llms/deepseek/messages/transformation.py

@@ -0,0 +1,133 @@
+"""
+DeepSeek Anthropic-compatible messages transformation config.
+"""
+
+from typing import Any, Dict, List, Optional, Tuple
+
+import litellm
+from litellm.llms.anthropic.experimental_pass_through.messages.transformation import (
+    AnthropicMessagesConfig,
+)
+from litellm.secret_managers.main import get_secret_str
+from litellm.types.router import GenericLiteLLMParams
+
+
+class DeepSeekAnthropicMessagesConfig(AnthropicMessagesConfig):
+    """
+    DeepSeek exposes an Anthropic-compatible Messages API at
+    https://api.deepseek.com/anthropic.
+
+    It accepts the native Anthropic Messages conversation shape, including
+    thinking blocks in assistant history, but rejects Anthropic's explicit
+    custom-tool discriminator (`{"type": "custom"}`).
+    """
+
+    @property
+    def custom_llm_provider(self) -> Optional[str]:
+        return "deepseek"
+
+    @staticmethod
+    def get_api_key(api_key: Optional[str] = None) -> Optional[str]:
+        return api_key or get_secret_str("DEEPSEEK_API_KEY") or litellm.api_key
+
+    @staticmethod
+    def get_api_base(api_base: Optional[str] = None) -> str:
+        return (
+            api_base
+            or get_secret_str("DEEPSEEK_ANTHROPIC_API_BASE")
+            or get_secret_str("DEEPSEEK_API_BASE")
+            or "https://api.deepseek.com/anthropic"
+        )
+
+    def validate_anthropic_messages_environment(
+        self,
+        headers: dict,
+        model: str,
+        messages: List[Any],
+        optional_params: dict,
+        litellm_params: dict,
+        api_key: Optional[str] = None,
+        api_base: Optional[str] = None,
+    ) -> Tuple[dict, Optional[str]]:
+        dynamic_api_key = self.get_api_key(api_key=api_key)
+
+        if (
+            "x-api-key" not in headers
+            and "authorization" not in headers
+            and dynamic_api_key is not None
+        ):
+            headers["x-api-key"] = dynamic_api_key
+
+        if "anthropic-version" not in headers:
+            headers["anthropic-version"] = "2023-06-01"
+        if "content-type" not in headers:
+            headers["content-type"] = "application/json"
+
+        headers = self._update_headers_with_anthropic_beta(
+            headers=headers,
+            optional_params=optional_params,
+            custom_llm_provider=self.custom_llm_provider or "deepseek",
+        )
+
+        return headers, api_base
+
+    def get_complete_url(
+        self,
+        api_base: Optional[str],
+        api_key: Optional[str],
+        model: str,
+        optional_params: dict,
+        litellm_params: dict,
+        stream: Optional[bool] = None,
+    ) -> str:
+        base_url = self.get_api_base(api_base=api_base).rstrip("/")
+
+        if base_url.endswith("/v1/messages") and "/anthropic/" in base_url:
+            return base_url
+        if base_url.endswith("/v1/messages"):
+            base_url = base_url[: -len("/v1/messages")]
+        if base_url.endswith("/v1"):
+            base_url = base_url[: -len("/v1")]
+        if base_url.endswith("/beta"):
+            base_url = base_url[: -len("/beta")]
+
+        if not base_url.endswith("/anthropic") and "/anthropic/" not in base_url:
+            base_url = f"{base_url}/anthropic"
+
+        return f"{base_url}/v1/messages"
+
+    @staticmethod
+    def _sanitize_tools_for_deepseek(tools: Any) -> Any:
+        if not isinstance(tools, list):
+            return tools
+
+        sanitized_tools = []
+        for tool in tools:
+            if isinstance(tool, dict) and tool.get("type") == "custom":
+                sanitized_tool = dict(tool)
+                sanitized_tool.pop("type", None)
+                sanitized_tools.append(sanitized_tool)
+            else:
+                sanitized_tools.append(tool)
+        return sanitized_tools
+
+    def transform_anthropic_messages_request(
+        self,
+        model: str,
+        messages: List[Dict],
+        anthropic_messages_optional_request_params: Dict,
+        litellm_params: GenericLiteLLMParams,
+        headers: dict,
+    ) -> Dict:
+        anthropic_messages_request = super().transform_anthropic_messages_request(
+            model=model,
+            messages=messages,
+            anthropic_messages_optional_request_params=anthropic_messages_optional_request_params,
+            litellm_params=litellm_params,
+            headers=headers,
+        )
+        if "tools" in anthropic_messages_request:
+            anthropic_messages_request["tools"] = self._sanitize_tools_for_deepseek(
+                anthropic_messages_request["tools"]
+            )
+        return anthropic_messages_request

litellm/llms/gemini/chat/transformation.py

@@ -1,5 +1,7 @@
 from typing import List, Optional, cast
 
+import litellm
+
 from litellm.litellm_core_utils.prompt_templates.factory import (
     convert_generic_image_chunk_to_openai_image_obj,
     convert_to_anthropic_image_obj,
@@ -101,7 +103,10 @@ class GoogleAIStudioGeminiConfig(VertexGeminiConfig):
         return supported_params
 
     def _transform_messages(
-        self, messages: List[AllMessageValues], model: Optional[str] = None
+        self,
+        messages: List[AllMessageValues],
+        model: Optional[str] = None,
+        litellm_params: Optional[dict] = None,
     ) -> List[ContentType]:
         """
         Google AI Studio Gemini does not support HTTP/HTTPS URLs for files.
@@ -141,14 +146,23 @@ class GoogleAIStudioGeminiConfig(VertexGeminiConfig):
                                 img_element["image_url"] = converted_image_url  # type: ignore
                     elif element.get("type") == "file":
                         file_element = cast(ChatCompletionFileObject, element)
-                        file_id = file_element["file"].get("file_id")
+                        _file_field = file_element.get("file")
+                        if _file_field is None:
+                            raise litellm.BadRequestError(
+                                message="Content block has type='file' but is missing the required 'file' field",
+                                model=model,
+                                llm_provider="gemini",
+                            )
+                        file_id = _file_field.get("file_id")
                         if file_id and ("http://" in file_id or "https://" in file_id):
                             # Convert HTTP/HTTPS file URL to base64 data
                             try:
                                 base64_data = convert_url_to_base64(file_id)
-                                file_element["file"]["file_data"] = base64_data  # type: ignore
-                                file_element["file"].pop("file_id", None)  # type: ignore
+                                _file_field["file_data"] = base64_data  # type: ignore
+                                _file_field.pop("file_id", None)  # type: ignore
                             except Exception:
                                 # If conversion fails, leave as is and let the API handle it
                                 pass
-        return _gemini_convert_messages_with_history(messages=messages, model=model)
+        return _gemini_convert_messages_with_history(
+            messages=messages, model=model, litellm_params=litellm_params
+        )

litellm/llms/openai/chat/gpt_transformation.py

@@ -287,7 +287,13 @@ class OpenAIGPTConfig(BaseLLMModelInfo, BaseConfig):
                 content_item["image_url"] = new_image_url_obj
         elif content_item.get("type") == "file":
             content_item = cast(ChatCompletionFileObject, content_item)
-            file_obj = content_item["file"]
+            file_obj = content_item.get("file")
+            if file_obj is None:
+                raise litellm.BadRequestError(
+                    message="Content block has type='file' but is missing the required 'file' field",
+                    model=None,
+                    llm_provider="openai",
+                )
             new_file_obj = ChatCompletionFileObjectFile(
                 **{  # type: ignore
                     k: v

litellm/llms/sagemaker/common_utils.py

@@ -1,3 +1,4 @@
+import functools
 import json
 from typing import AsyncIterator, Iterator, List, Optional, Union
 
@@ -22,14 +23,22 @@ def _load_sagemaker_response_stream_shape():
         )
     except Exception as e:
         verbose_logger.warning(
-            "litellm: could not pre-load sagemaker-runtime response stream shape "
+            "litellm: could not load sagemaker-runtime response stream shape "
             "— SageMaker event-stream decoding will be unavailable. Error: %s",
             e,
         )
         return None
 
 
-SAGEMAKER_RESPONSE_STREAM_SHAPE = _load_sagemaker_response_stream_shape()
+@functools.lru_cache(maxsize=1)
+def get_sagemaker_response_stream_shape():
+    """
+    Lazily load and cache the sagemaker-runtime stream shape for the process.
+
+    Avoids importing botocore (and logging warnings) unless SageMaker event-stream
+    decoding is actually needed.
+    """
+    return _load_sagemaker_response_stream_shape()
 
 
 class SagemakerError(BaseLLMException):
@@ -207,7 +216,8 @@ class AWSEventStreamDecoder:
                 verbose_logger.error(f"Final error parsing accumulated JSON: {e}")
 
     def _parse_message_from_event(self, event) -> Optional[str]:
-        if SAGEMAKER_RESPONSE_STREAM_SHAPE is None:
+        response_stream_shape = get_sagemaker_response_stream_shape()
+        if response_stream_shape is None:
             raise SagemakerError(
                 status_code=500,
                 message=(
@@ -216,9 +226,7 @@ class AWSEventStreamDecoder:
                 ),
             )
         response_dict = event.to_response_dict()
-        parsed_response = self.parser.parse(
-            response_dict, SAGEMAKER_RESPONSE_STREAM_SHAPE
-        )
+        parsed_response = self.parser.parse(response_dict, response_stream_shape)
 
         if response_dict["status_code"] != 200:
             raise ValueError(f"Bad response code, expected 200: {response_dict}")

litellm/llms/vertex_ai/gemini/transformation.py

@@ -6,13 +6,16 @@ Why separate file? Make it easy to see how transformation works
 
 import json
 import os
-from typing import TYPE_CHECKING, Dict, List, Literal, Optional, Tuple, Union, cast
+import re
+from typing import TYPE_CHECKING, Any, Dict, List, Literal, Optional, Tuple, Union, cast
+from urllib.parse import quote
 
 import httpx
 from pydantic import BaseModel
 
 import litellm
 from litellm._logging import verbose_logger
+from litellm.litellm_core_utils.asyncify import asyncify
 from litellm.litellm_core_utils.prompt_templates.common_utils import (
     _get_image_mime_type_from_url,
 )
@@ -57,6 +60,45 @@ from ..common_utils import (
     get_supports_system_message,
 )
 
+# Typed as Any to avoid introducing a module-load-time cyclic import to
+# vertex_llm_base. The instance is lazily constructed by _get_vertex_base()
+# the first time GCS metadata needs to be fetched.
+_GCS_METADATA_VERTEX_BASE: Optional[Any] = None
+# Shared sync client for GCS JSON API metadata reads so proxy/SSL settings
+# from litellm's HTTP stack apply (see Greptile review on PR #27278).
+_GCS_METADATA_HTTP_HANDLER: Optional[HTTPHandler] = None
+_GEMINI_MIME_TYPE_ALIASES: Dict[str, str] = {
+    "image/jpg": "image/jpeg",
+}
+
+
+def _apply_gemini_mime_type_aliases(mime_type: str) -> str:
+    """Normalize known MIME aliases only; does not consult the file-type registry.
+
+    Also strips MIME parameters (e.g. ``; charset=utf-8``) so that values
+    sourced from GCS object metadata (``contentType``) validate correctly.
+    """
+    normalized = mime_type.split(";", 1)[0].strip().lower()
+    return _GEMINI_MIME_TYPE_ALIASES.get(normalized, normalized)
+
+
+def _get_vertex_base() -> Any:
+    """Lazily return the shared VertexBase instance to avoid a module-load-time cyclic import."""
+    global _GCS_METADATA_VERTEX_BASE
+    if _GCS_METADATA_VERTEX_BASE is None:
+        from ..vertex_llm_base import VertexBase
+
+        _GCS_METADATA_VERTEX_BASE = VertexBase()
+    return _GCS_METADATA_VERTEX_BASE
+
+
+def _get_gcs_metadata_http_handler() -> HTTPHandler:
+    global _GCS_METADATA_HTTP_HANDLER
+    if _GCS_METADATA_HTTP_HANDLER is None:
+        _GCS_METADATA_HTTP_HANDLER = HTTPHandler(timeout=5.0)
+    return _GCS_METADATA_HTTP_HANDLER
+
+
 if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging as _LiteLLMLoggingObj
 
@@ -171,12 +213,299 @@ def _apply_gemini_metadata(
     return cast(PartType, part_dict)
 
 
+def _parse_gs_uri(gs_uri: str) -> Tuple[str, str]:
+    if not gs_uri.startswith("gs://"):
+        raise ValueError(f"Invalid gs URI: {gs_uri}")
+    uri_without_scheme = gs_uri[5:]  # drop gs://
+    uri_parts = uri_without_scheme.split("/", 1)
+    if len(uri_parts) != 2 or not uri_parts[0] or not uri_parts[1]:
+        raise ValueError(f"Invalid gs URI: {gs_uri}")
+    return uri_parts[0], uri_parts[1]
+
+
+def _is_valid_gcs_bucket_name(bucket: str) -> bool:
+    """
+    Validate bucket name against core GCS naming constraints.
+    """
+    bucket_length = len(bucket)
+    max_bucket_length = 222 if "." in bucket else 63
+    if bucket_length < 3 or bucket_length > max_bucket_length:
+        return False
+    if "." in bucket and any(
+        len(label) == 0 or len(label) > 63 for label in bucket.split(".")
+    ):
+        return False
+    if not re.fullmatch(r"[a-z0-9][a-z0-9._-]*[a-z0-9]", bucket):
+        return False
+    if ".." in bucket:
+        return False
+    if re.fullmatch(r"\d+\.\d+\.\d+\.\d+", bucket):
+        return False
+    return True
+
+
+def _gs_uri_requires_content_type_metadata(url: str) -> bool:
+    """
+    True when _process_gemini_media would call _get_gcs_object_content_type
+    (extension-less gs:// and no explicit format passed into that helper).
+    """
+    if "gs://" not in url:
+        return False
+    extension_with_dot = os.path.splitext(url)[-1]
+    extension = extension_with_dot[1:] if extension_with_dot else ""
+    return len(extension) == 0
+
+
+def _image_url_payload_may_need_sync_gcs_metadata_fetch(
+    raw_image_url: Any,
+) -> bool:
+    """
+    True when this image_url value (content-part image_url or assistant ``images[]``
+    entry) can trigger a blocking GCS metadata read for MIME resolution.
+    """
+    fmt: Optional[str] = None
+    url: Optional[str] = None
+    if isinstance(raw_image_url, dict):
+        url = raw_image_url.get("url")  # type: ignore[assignment]
+        if not isinstance(url, str):
+            return False
+        fmt = (
+            raw_image_url.get("format")
+            or raw_image_url.get("mime_type")
+            or raw_image_url.get("content_type")
+        )
+    elif isinstance(raw_image_url, str):
+        url = raw_image_url
+    else:
+        return False
+    if "gs://" not in url or fmt:
+        return False
+    return _gs_uri_requires_content_type_metadata(url)
+
+
+def _openai_messages_may_need_sync_gcs_metadata_fetch(
+    messages: List[AllMessageValues],
+) -> bool:
+    """
+    Heuristic: True if any message part can trigger a blocking GCS JSON
+    metadata read inside _transform_request_body (extension-less gs:// without
+    explicit MIME hints). Covers user/system ``content`` parts and assistant
+    ``images`` (same paths as ``_gemini_convert_messages_with_history``). Used
+    to decide whether ``async_transform_request_body`` should offload the sync
+    transform via ``asyncify``.
+    """
+    for raw in messages:
+        msg: Any = raw
+        if not isinstance(msg, dict) and hasattr(msg, "model_dump"):
+            msg = msg.model_dump(exclude_none=False)
+        if not isinstance(msg, dict):
+            continue
+        images_field = msg.get("images")
+        if isinstance(images_field, list):
+            for image_item in images_field:
+                if not isinstance(image_item, dict):
+                    continue
+                if _image_url_payload_may_need_sync_gcs_metadata_fetch(
+                    image_item.get("image_url")
+                ):
+                    return True
+
+        content = msg.get("content")
+        if not isinstance(content, list):
+            continue
+        for item in content:
+            if not isinstance(item, dict):
+                continue
+            itype = item.get("type")
+            if itype == "image_url":
+                if _image_url_payload_may_need_sync_gcs_metadata_fetch(
+                    item.get("image_url")
+                ):
+                    return True
+            elif itype == "file":
+                file_obj = item.get("file")
+                if not isinstance(file_obj, dict):
+                    continue
+                fmt = (
+                    file_obj.get("format")
+                    or file_obj.get("mime_type")
+                    or file_obj.get("content_type")
+                )
+                passed = file_obj.get("file_id") or file_obj.get("file_data")
+                if (
+                    isinstance(passed, str)
+                    and "gs://" in passed
+                    and not fmt
+                    and _gs_uri_requires_content_type_metadata(passed)
+                ):
+                    return True
+    return False
+
+
+def _get_gcs_object_content_type(
+    image_url: str,
+    vertex_project: Optional[str] = None,
+    vertex_credentials: Optional[Any] = None,
+) -> Optional[str]:
+    """
+    Resolve content type from GCS object metadata.
+
+    Only attaches a Bearer token when the caller explicitly supplies Vertex
+    credentials, to avoid using the server's default Google credentials on
+    the Gemini API-key (Google AI Studio) path and being used as an oracle
+    for private GCS object metadata. Without explicit credentials we only
+    issue an anonymous request, which only succeeds for publicly-readable
+    objects.
+    """
+    try:
+        bucket, object_name = _parse_gs_uri(image_url)
+    except ValueError:
+        return None
+    if not _is_valid_gcs_bucket_name(bucket):
+        return None
+
+    headers: Dict[str, str] = {}
+    explicit_vertex_auth_provided = (
+        vertex_project is not None or vertex_credentials is not None
+    )
+    if explicit_vertex_auth_provided:
+        try:
+            access_token, _ = _get_vertex_base().get_access_token(
+                credentials=vertex_credentials,
+                project_id=vertex_project,
+            )
+            headers["Authorization"] = f"Bearer {access_token}"
+        except Exception as e:
+            raise litellm.BadRequestError(
+                message=(
+                    "Unable to fetch GCS metadata with provided Vertex credentials/project. "
+                    f"Original error: {str(e)}"
+                ),
+                model=None,
+                llm_provider="vertex_ai",
+            )
+
+    # Build the URL via httpx.URL with a fixed scheme/host and URL-encode both
+    # bucket and object so CodeQL does not flag the interpolation as a
+    # potential SSRF that could resolve to an arbitrary host.
+    encoded_bucket = quote(bucket, safe="")
+    encoded_object = quote(object_name, safe="")
+    metadata_url = httpx.URL(
+        scheme="https",
+        host="storage.googleapis.com",
+        path=f"/storage/v1/b/{encoded_bucket}/o/{encoded_object}",
+        params={"fields": "contentType"},
+    )
+    try:
+        response = _get_gcs_metadata_http_handler().get(
+            url=str(metadata_url),
+            headers=headers or None,
+        )
+    except httpx.RequestError as e:
+        if explicit_vertex_auth_provided:
+            raise litellm.BadRequestError(
+                message=(
+                    "Unable to reach GCS JSON API for object metadata with provided "
+                    f"Vertex credentials. {type(e).__name__}: {e}"
+                ),
+                model=None,
+                llm_provider="vertex_ai",
+            ) from e
+        return None
+
+    if response.is_error:
+        if explicit_vertex_auth_provided:
+            preview = (response.text or "")[:1024]
+            raise litellm.BadRequestError(
+                message=(
+                    "Unable to read GCS object metadata with provided Vertex credentials. "
+                    f"HTTP {response.status_code}. Response body (truncated): {preview!r}"
+                ),
+                model=None,
+                llm_provider="vertex_ai",
+            )
+        return None
+
+    try:
+        payload = response.json()
+    except ValueError as e:
+        if explicit_vertex_auth_provided:
+            raise litellm.BadRequestError(
+                message=(
+                    "GCS metadata response was not valid JSON when using provided "
+                    f"Vertex credentials (HTTP {response.status_code}). Error: {e}"
+                ),
+                model=None,
+                llm_provider="vertex_ai",
+            ) from e
+        return None
+
+    if not isinstance(payload, dict):
+        if explicit_vertex_auth_provided:
+            raise litellm.BadRequestError(
+                message=(
+                    "GCS metadata response was not a JSON object when using provided "
+                    f"Vertex credentials (HTTP {response.status_code})."
+                ),
+                model=None,
+                llm_provider="vertex_ai",
+            )
+        return None
+
+    content_type = payload.get("contentType")
+    if isinstance(content_type, str) and len(content_type) > 0:
+        return content_type
+
+    if explicit_vertex_auth_provided:
+        preview = (response.text or "")[:1024]
+        raise litellm.BadRequestError(
+            message=(
+                "GCS metadata JSON did not include a non-empty contentType field when "
+                f"using provided Vertex credentials (HTTP {response.status_code}). "
+                f"Body (truncated): {preview!r}"
+            ),
+            model=None,
+            llm_provider="vertex_ai",
+        )
+    return None
+
+
+def _normalize_and_validate_gemini_mime_type(
+    mime_type: str, model: Optional[str]
+) -> str:
+    # Import lazily to avoid a module-level cyclic-import alert with
+    # litellm.types.files.
+    from litellm.types.files import get_file_extension_from_mime_type
+
+    normalized_mime_type = _apply_gemini_mime_type_aliases(mime_type)
+    try:
+        file_extension = get_file_extension_from_mime_type(normalized_mime_type)
+        file_type = get_file_type_from_extension(file_extension)
+    except ValueError:
+        raise litellm.BadRequestError(
+            message=f"File type not supported by gemini - {normalized_mime_type}",
+            model=model,
+            llm_provider="vertex_ai",
+        )
+
+    if not is_gemini_1_5_accepted_file_type(file_type):
+        raise litellm.BadRequestError(
+            message=f"File type not supported by gemini - {file_type}",
+            model=model,
+            llm_provider="vertex_ai",
+        )
+
+    return get_file_mime_type_for_file_type(file_type)
+
+
 def _process_gemini_media(
     image_url: str,
     format: Optional[str] = None,
     media_resolution_enum: Optional[Dict[str, str]] = None,
     model: Optional[str] = None,
     video_metadata: Optional[Dict[str, Any]] = None,
+    vertex_project: Optional[str] = None,
+    vertex_credentials: Optional[Any] = None,
 ) -> PartType:
     """
     Given a media URL (image, audio, or video), return the appropriate PartType for Gemini
@@ -193,20 +522,63 @@ def _process_gemini_media(
     try:
         # GCS URIs
         if "gs://" in image_url:
-            # Figure out file type
             extension_with_dot = os.path.splitext(image_url)[-1]  # Ex: ".png"
             extension = extension_with_dot[1:]  # Ex: "png"
 
+            explicit_gcs_format = False
             if not format:
-                file_type = get_file_type_from_extension(extension)
+                mime_type: Optional[str] = None
+                # For extension-less gs:// URIs, we cannot infer from path.
+                # If callers pass `format`/`mime_type`, this branch is skipped.
+                if extension:
+                    file_type = get_file_type_from_extension(extension)
 
-                # Validate the file type is supported by Gemini
-                if not is_gemini_1_5_accepted_file_type(file_type):
-                    raise Exception(f"File type not supported by gemini - {file_type}")
+                    # Validate the file type is supported by Gemini
+                    if not is_gemini_1_5_accepted_file_type(file_type):
+                        raise litellm.BadRequestError(
+                            message=f"File type not supported by gemini - {file_type}",
+                            model=model,
+                            llm_provider="vertex_ai",
+                        )
 
-                mime_type = get_file_mime_type_for_file_type(file_type)
+                    mime_type = get_file_mime_type_for_file_type(file_type)
+                else:
+                    mime_type = _get_gcs_object_content_type(
+                        image_url=image_url,
+                        vertex_project=vertex_project,
+                        vertex_credentials=vertex_credentials,
+                    )
+                    if mime_type is None:
+                        raise litellm.BadRequestError(
+                            message=(
+                                f"Unable to determine mime type for gs URI: {image_url}. "
+                                "This gs:// URI has no file extension and GCS metadata "
+                                "lookup failed. Set it explicitly using image_url.format "
+                                "(or image_url.mime_type/content_type) or "
+                                "message.content[].file.format."
+                            ),
+                            model=model,
+                            llm_provider="vertex_ai",
+                        )
             else:
                 mime_type = format
+                explicit_gcs_format = True
+            if mime_type is None:
+                raise litellm.BadRequestError(
+                    message=f"File type not supported by gemini - {image_url}",
+                    model=model,
+                    llm_provider="vertex_ai",
+                )
+            if explicit_gcs_format:
+                # Callers who pass format/mime_type explicitly for gs:// URIs
+                # rely on pass-through to Gemini (pre-PR behavior). Only apply
+                # known MIME aliases; skip litellm's file-type registry.
+                mime_type = _apply_gemini_mime_type_aliases(mime_type)
+            else:
+                mime_type = _normalize_and_validate_gemini_mime_type(
+                    mime_type=mime_type,
+                    model=model,
+                )
             file_data = FileDataType(mime_type=mime_type, file_uri=image_url)
             part: PartType = {"file_data": file_data}
             return _apply_gemini_metadata(
@@ -258,8 +630,6 @@ def _snake_to_camel(snake_str: str) -> str:
 
 def _camel_to_snake(camel_str: str) -> str:
     """Convert camelCase to snake_case"""
-    import re
-
     return re.sub(r"(?<!^)(?=[A-Z])", "_", camel_str).lower()
 
 
@@ -311,6 +681,7 @@ def check_if_part_exists_in_parts(
 def _gemini_convert_messages_with_history(  # noqa: PLR0915
     messages: List[AllMessageValues],
     model: Optional[str] = None,
+    litellm_params: Optional[dict] = None,
 ) -> List[ContentType]:
     """
     Converts given messages from OpenAI format to Gemini format
@@ -326,6 +697,16 @@ def _gemini_convert_messages_with_history(  # noqa: PLR0915
 
     msg_i = 0
     tool_call_responses = []
+    vertex_project = None
+    vertex_credentials = None
+    if litellm_params:
+        vertex_project = litellm_params.get("vertex_project") or litellm_params.get(
+            "vertex_ai_project"
+        )
+        vertex_credentials = litellm_params.get(
+            "vertex_credentials"
+        ) or litellm_params.get("vertex_ai_credentials")
+
     try:
         while msg_i < len(messages):
             user_content: List[PartType] = []
@@ -351,20 +732,42 @@ def _gemini_convert_messages_with_history(  # noqa: PLR0915
                             img_element = element
                             format: Optional[str] = None
                             media_resolution_enum: Optional[Dict[str, str]] = None
-                            if isinstance(img_element["image_url"], dict):
-                                image_url = img_element["image_url"]["url"]
-                                format = img_element["image_url"].get("format")
-                                detail = img_element["image_url"].get("detail")
+                            raw_image_url = img_element.get("image_url")
+                            if raw_image_url is None:
+                                raise litellm.BadRequestError(
+                                    message="Invalid message content: element type is 'image_url' but 'image_url' field is missing ",
+                                    model=model,
+                                    llm_provider="vertex_ai",
+                                )
+                            if isinstance(raw_image_url, dict):
+                                image_url = raw_image_url.get("url")
+                                if image_url is None:
+                                    raise litellm.BadRequestError(
+                                        message="Invalid message content: element type is 'image_url' but 'url' field is missing inside 'image_url' ",
+                                        model=model,
+                                        llm_provider="vertex_ai",
+                                    )
+                                # TypedDict does not declare mime_type/content_type;
+                                # read via Dict[str, Any] for caller-provided MIME fields.
+                                image_url_dict = cast(Dict[str, Any], raw_image_url)
+                                format = (
+                                    image_url_dict.get("format")
+                                    or image_url_dict.get("mime_type")
+                                    or image_url_dict.get("content_type")
+                                )
+                                detail = image_url_dict.get("detail")
                                 media_resolution_enum = (
                                     _convert_detail_to_media_resolution_enum(detail)
                                 )
                             else:
-                                image_url = img_element["image_url"]
+                                image_url = raw_image_url
                             _part = _process_gemini_media(
                                 image_url=image_url,
                                 format=format,
                                 media_resolution_enum=media_resolution_enum,
                                 model=model,
+                                vertex_project=vertex_project,
+                                vertex_credentials=vertex_credentials,
                             )
                             _parts.append(_part)
                         elif element["type"] == "input_audio":
@@ -390,15 +793,31 @@ def _gemini_convert_messages_with_history(  # noqa: PLR0915
                                     image_url=openai_image_str,
                                     format=audio_format_modified,
                                     model=model,
+                                    vertex_project=vertex_project,
+                                    vertex_credentials=vertex_credentials,
                                 )
                                 _parts.append(_part)
                         elif element["type"] == "file":
                             file_element = cast(ChatCompletionFileObject, element)
-                            file_id = file_element["file"].get("file_id")
-                            format = file_element["file"].get("format")
-                            file_data = file_element["file"].get("file_data")
-                            detail = file_element["file"].get("detail")
-                            video_metadata = file_element["file"].get("video_metadata")
+                            _file_field = file_element.get("file")
+                            if _file_field is None:
+                                raise litellm.BadRequestError(
+                                    message="Content block has type='file' but is missing the required 'file' field",
+                                    model=model,
+                                    llm_provider="vertex_ai",
+                                )
+                            # TypedDict does not declare mime_type/content_type;
+                            # read via Dict[str, Any] for caller-provided MIME fields.
+                            file_dict = cast(Dict[str, Any], _file_field)
+                            file_id = file_dict.get("file_id")
+                            format = (
+                                file_dict.get("format")
+                                or file_dict.get("mime_type")
+                                or file_dict.get("content_type")
+                            )
+                            file_data = file_dict.get("file_data")
+                            detail = file_dict.get("detail")
+                            video_metadata = file_dict.get("video_metadata")
                             passed_file = file_id or file_data
                             if passed_file is None:
                                 raise Exception(
@@ -417,13 +836,23 @@ def _gemini_convert_messages_with_history(  # noqa: PLR0915
                                     model=model,
                                     media_resolution_enum=media_resolution_enum,
                                     video_metadata=video_metadata,
+                                    vertex_project=vertex_project,
+                                    vertex_credentials=vertex_credentials,
                                 )
                                 _parts.append(_part)
-                            except Exception:
-                                raise Exception(
-                                    "Unable to determine mime type for file_id: {}, set this explicitly using message[{}].content[{}].file.format".format(
-                                        file_id, msg_i, element_idx
-                                    )
+                            except litellm.BadRequestError:
+                                raise
+                            except Exception as e:
+                                raise litellm.BadRequestError(
+                                    message=(
+                                        f"Unable to determine mime type for file: "
+                                        f"{file_id or 'provided data'}, set this explicitly "
+                                        f"using message[{msg_i}].content[{element_idx}].file.format "
+                                        f"(or file.mime_type/content_type). "
+                                        f"Original error: {str(e)}"
+                                    ),
+                                    model=model,
+                                    llm_provider="vertex_ai",
                                 )
                     user_content.extend(_parts)
                 elif _message_content is not None and isinstance(_message_content, str):
@@ -528,7 +957,11 @@ def _gemini_convert_messages_with_history(  # noqa: PLR0915
                             image_url_obj = image_item.get("image_url")
                             if isinstance(image_url_obj, dict):
                                 assistant_image_url = image_url_obj.get("url")
-                                format = image_url_obj.get("format")
+                                format = (
+                                    image_url_obj.get("format")
+                                    or image_url_obj.get("mime_type")
+                                    or image_url_obj.get("content_type")
+                                )
                                 detail = image_url_obj.get("detail")
                                 media_resolution_enum = (
                                     _convert_detail_to_media_resolution_enum(detail)
@@ -539,6 +972,8 @@ def _gemini_convert_messages_with_history(  # noqa: PLR0915
                                         format=format,
                                         media_resolution_enum=media_resolution_enum,
                                         model=model,
+                                        vertex_project=vertex_project,
+                                        vertex_credentials=vertex_credentials,
                                     )
                                     assistant_content.append(_part)
 
@@ -713,11 +1148,11 @@ def _transform_request_body(  # noqa: PLR0915
     try:
         if custom_llm_provider == "gemini":
             content = litellm.GoogleAIStudioGeminiConfig()._transform_messages(
-                messages=messages, model=model
+                messages=messages, model=model, litellm_params=litellm_params
             )
         else:
             content = litellm.VertexGeminiConfig()._transform_messages(
-                messages=messages, model=model
+                messages=messages, model=model, litellm_params=litellm_params
             )
         tools: Optional[Tools] = optional_params.pop("tools", None)
         tool_choice: Optional[ToolConfig] = optional_params.pop("tool_choice", None)
@@ -893,6 +1328,20 @@ async def async_transform_request_body(
         vertex_auth_header=vertex_auth_header,
     )
 
+    if _openai_messages_may_need_sync_gcs_metadata_fetch(messages):
+        # _transform_request_body may issue a sync httpx.get (up to 5s timeout)
+        # via _get_gcs_object_content_type to fetch GCS object metadata. Run the
+        # whole sync transformation on a worker thread so it does not block the
+        # async event loop.
+        return await asyncify(_transform_request_body)(
+            messages=messages,
+            model=model,
+            custom_llm_provider=custom_llm_provider,
+            litellm_params=litellm_params,
+            cached_content=cached_content,
+            optional_params=optional_params,
+        )
+
     return _transform_request_body(
         messages=messages,
         model=model,

litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py

@@ -2533,9 +2533,14 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig):
         return model_response
 
     def _transform_messages(
-        self, messages: List[AllMessageValues], model: Optional[str] = None
+        self,
+        messages: List[AllMessageValues],
+        model: Optional[str] = None,
+        litellm_params: Optional[dict] = None,
     ) -> List[ContentType]:
-        return _gemini_convert_messages_with_history(messages=messages, model=model)
+        return _gemini_convert_messages_with_history(
+            messages=messages, model=model, litellm_params=litellm_params
+        )
 
     def get_error_class(
         self, error_message: str, status_code: int, headers: Union[Dict, httpx.Headers]
@@ -3139,6 +3144,31 @@ class ModelResponseIterator:
         self.cumulative_tool_call_index: int = 0
         self.has_seen_tool_calls: bool = False
 
+    @staticmethod
+    def _check_streaming_error(chunk: dict) -> None:
+        """Detect embedded errors (e.g. 429 RESOURCE_EXHAUSTED) in streaming chunks and raise VertexAIError."""
+        if "error" not in chunk:
+            return
+        error_data = chunk["error"]
+        if not isinstance(error_data, dict):
+            raise VertexAIError(
+                status_code=500,
+                message=f"Unexpected error format in mid-stream chunk: {error_data}",
+            )
+        raw_code = error_data.get("code", 500)
+        if raw_code is None:
+            raw_code = 500
+        try:
+            error_code = int(raw_code)
+        except (TypeError, ValueError):
+            error_code = 500
+        error_message = error_data.get("message", "Unknown error")
+        error_status = error_data.get("status", "UNKNOWN")
+        raise VertexAIError(
+            status_code=error_code,
+            message=f"{error_status} - {error_message}",
+        )
+
     def _apply_stream_candidates(
         self,
         _candidates: List[Candidates],
@@ -3256,6 +3286,11 @@ class ModelResponseIterator:
     def chunk_parser(self, chunk: dict) -> Optional["ModelResponseStream"]:
         try:
             verbose_logger.debug(f"RAW GEMINI CHUNK: {chunk}")
+
+            # Detect mid-stream error chunks (e.g. 429 RESOURCE_EXHAUSTED).
+            # Vertex AI can return errors as HTTP 200 but with an "error" field in the SSE body.
+            self._check_streaming_error(chunk)
+
             from litellm.types.utils import ModelResponseStream
 
             processed_chunk = GenerateContentResponseBody(**chunk)  # type: ignore

litellm/llms/vertex_ai/vertex_ai_partner_models/main.py

@@ -292,22 +292,15 @@ class VertexAIPartnerModels(VertexBase):
         Returns:
             Dict containing token count information
         """
-        try:
-            import vertexai
-        except Exception as e:
-            raise VertexAIError(
-                status_code=400,
-                message=f"""vertexai import failed please run `pip install -U "google-cloud-aiplatform>=1.38"`. Got error: {e}""",
-            )
-
-        if not (
-            hasattr(vertexai, "preview") or hasattr(vertexai.preview, "language_models")
-        ):
-            raise VertexAIError(
-                status_code=400,
-                message="""Upgrade vertex ai. Run `pip install "google-cloud-aiplatform>=1.38"`""",
-            )
-
+        # Note: we intentionally do not import `vertexai` (the Gemini SDK shipped
+        # by `google-cloud-aiplatform`) on this path. Partner models such as
+        # Claude on Vertex use the Anthropic Messages API protocol directly via
+        # `:rawPredict`, and `VertexAIPartnerModelsTokenCounter` reaches that
+        # endpoint with an authenticated httpx client — it never touches the
+        # Gemini SDK. Requiring `google-cloud-aiplatform>=1.38` here turned a
+        # SDK-free Anthropic-protocol call into a hard dependency on the Gemini
+        # SDK (see #28084), breaking `/v1/messages/count_tokens` for Claude-on-
+        # Vertex on any LiteLLM install without that extra. Stay SDK-free.
         try:
             from litellm.llms.vertex_ai.vertex_ai_partner_models.count_tokens.handler import (
                 VertexAIPartnerModelsTokenCounter,

litellm/main.py

@@ -5720,6 +5720,33 @@ def embedding(  # noqa: PLR0915
                 aembedding=aembedding,
                 headers=headers,
             )
+        elif custom_llm_provider == "dashscope":
+            dashscope_key = (
+                api_key or litellm.api_key or get_secret_str("DASHSCOPE_API_KEY")
+            )
+            if dashscope_key is None:
+                raise ValueError(
+                    "Missing API key for DashScope. Set DASHSCOPE_API_KEY environment variable or pass api_key parameter."
+                )
+            if extra_headers is not None and isinstance(extra_headers, dict):
+                headers = extra_headers
+            else:
+                headers = {}
+            response = base_llm_http_handler.embedding(
+                model=model,
+                input=input,
+                timeout=timeout,
+                custom_llm_provider=custom_llm_provider,
+                logging_obj=logging,
+                api_base=api_base,
+                optional_params=optional_params,
+                litellm_params={},
+                model_response=EmbeddingResponse(),
+                api_key=dashscope_key,
+                client=client,
+                aembedding=aembedding,
+                headers=headers,
+            )
         elif custom_llm_provider == "ovhcloud":
             api_key = api_key or litellm.api_key or get_secret_str("OVHCLOUD_API_KEY")
             api_base = (

litellm/proxy/_types.py

@@ -4549,6 +4549,7 @@ class PrismaCompatibleUpdateDBModel(TypedDict, total=False):
     model_name: str
     litellm_params: str
     model_info: str
+    blocked: bool
     updated_at: str
     updated_by: str
 

litellm/proxy/db/db_spend_update_writer.py

@@ -69,6 +69,35 @@ else:
     ProxyLogging = Any
 
 
+def _extract_cache_read_tokens(usage_obj: dict) -> int:
+    """
+    Anthropic: top-level cache_read_input_tokens field.
+    OpenAI-compatible (moonshotai, openai, deepseek, etc.): prompt_tokens_details.cached_tokens.
+    """
+    explicit = usage_obj.get("cache_read_input_tokens", 0) or 0
+    if explicit:
+        return int(explicit)
+    details = usage_obj.get("prompt_tokens_details") or {}
+    return int(details.get("cached_tokens", 0) or 0)
+
+
+def _extract_cache_creation_tokens(usage_obj: dict) -> int:
+    """
+    Anthropic: top-level cache_creation_input_tokens field.
+    OpenAI-compatible (kimi-k2 etc.): prompt_tokens_details.cache_write_tokens
+    or prompt_tokens_details.cache_creation_tokens.
+    """
+    explicit = usage_obj.get("cache_creation_input_tokens", 0) or 0
+    if explicit:
+        return int(explicit)
+    details = usage_obj.get("prompt_tokens_details") or {}
+    return int(
+        details.get("cache_write_tokens", 0)
+        or details.get("cache_creation_tokens", 0)
+        or 0
+    )
+
+
 class DBSpendUpdateWriter:
     """
     Module responsible for
@@ -1992,12 +2021,8 @@ class DBSpendUpdateWriter:
                 api_requests=1,
                 successful_requests=1 if request_status == "success" else 0,
                 failed_requests=1 if request_status != "success" else 0,
-                cache_read_input_tokens=usage_obj.get("cache_read_input_tokens", 0)
-                or 0,
-                cache_creation_input_tokens=usage_obj.get(
-                    "cache_creation_input_tokens", 0
-                )
-                or 0,
+                cache_read_input_tokens=_extract_cache_read_tokens(usage_obj),
+                cache_creation_input_tokens=_extract_cache_creation_tokens(usage_obj),
             )
             return daily_transaction
         except Exception as e:

litellm/proxy/management_endpoints/common_utils.py

@@ -1,6 +1,7 @@
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Union
 
 from fastapi import HTTPException, status
+from pydantic import BaseModel
 
 from litellm._logging import verbose_proxy_logger
 from litellm.caching import DualCache
@@ -53,6 +54,37 @@ def require_caller_user_id_for_non_admin(
     return user_api_key_dict.user_id
 
 
+def _check_passthrough_routes_caller_permission(
+    data: BaseModel,
+    user_api_key_dict: UserAPIKeyAuth,
+    *,
+    entity: str = "key",
+) -> None:
+    """
+    Only proxy admins may set `allowed_passthrough_routes` (top-level or under
+    `metadata`) — it short-circuits the role-based route gate, so keys and teams
+    must be gated identically.
+    """
+    # view-only admins excluded by design; blocked upstream from writes anyway
+    if user_api_key_dict.user_role == LitellmUserRoles.PROXY_ADMIN.value:
+        return
+    if getattr(data, "allowed_passthrough_routes", None):
+        raise HTTPException(
+            status_code=403,
+            detail={
+                "error": f"Only proxy admins can set `allowed_passthrough_routes` on a {entity}."
+            },
+        )
+    metadata = getattr(data, "metadata", None)
+    if isinstance(metadata, dict) and metadata.get("allowed_passthrough_routes"):
+        raise HTTPException(
+            status_code=403,
+            detail={
+                "error": f"Only proxy admins can set `metadata.allowed_passthrough_routes` on a {entity}."
+            },
+        )
+
+
 def _is_user_team_admin(
     user_api_key_dict: UserAPIKeyAuth, team_obj: LiteLLM_TeamTable
 ) -> bool:

litellm/proxy/management_endpoints/key_management_endpoints.py

@@ -55,6 +55,7 @@ from litellm.proxy.common_utils.rbac_utils import check_org_admin_can_generate_k
 from litellm.proxy.common_utils.timezone_utils import get_budget_reset_time
 from litellm.proxy.hooks.key_management_event_hooks import KeyManagementEventHooks
 from litellm.proxy.management_endpoints.common_utils import (
+    _check_passthrough_routes_caller_permission,
     _is_user_org_admin_for_team,
     _is_user_team_admin,
     _set_object_metadata_field,
@@ -548,36 +549,6 @@ def _check_allowed_routes_caller_permission(
     )
 
 
-def _check_passthrough_routes_caller_permission(
-    data: BaseModel,
-    user_api_key_dict: UserAPIKeyAuth,
-) -> None:
-    """
-    Only proxy admins may set `allowed_passthrough_routes` on a key, either at
-    the top level of the request or nested under `metadata`.
-
-    The route gate evaluates passthrough access ahead of the standard role
-    gate, so the field is restricted to admins to keep that ordering safe.
-    """
-    if user_api_key_dict.user_role == LitellmUserRoles.PROXY_ADMIN.value:
-        return
-    if getattr(data, "allowed_passthrough_routes", None):
-        raise HTTPException(
-            status_code=403,
-            detail={
-                "error": "Only proxy admins can set `allowed_passthrough_routes` on a key."
-            },
-        )
-    metadata = getattr(data, "metadata", None)
-    if isinstance(metadata, dict) and metadata.get("allowed_passthrough_routes"):
-        raise HTTPException(
-            status_code=403,
-            detail={
-                "error": "Only proxy admins can set `metadata.allowed_passthrough_routes` on a key."
-            },
-        )
-
-
 async def validate_team_id_used_in_service_account_request(
     team_id: Optional[str],
     prisma_client: Optional[PrismaClient],

litellm/proxy/management_endpoints/model_management_endpoints.py

@@ -150,6 +150,9 @@ def update_db_model(
                 model_info[key] = value.isoformat()
         prisma_compatible_model_dict["model_info"] = json.dumps(model_info)
 
+    if updated_patch.blocked is not None:
+        prisma_compatible_model_dict["blocked"] = updated_patch.blocked
+
     return prisma_compatible_model_dict
 
 
@@ -230,6 +233,20 @@ async def patch_model(
             premium_user=premium_user,
         )
 
+        # Pause/resume (`blocked`) is a proxy-admin-only privilege. Team admins
+        # passed the auth check above for team-scoped models, but they must not
+        # be able to unblock (or block) a model their proxy admin has paused.
+        if (
+            patch_data.blocked is not None
+            and user_api_key_dict.user_role != LitellmUserRoles.PROXY_ADMIN
+        ):
+            raise ProxyException(
+                message="Only proxy admins can change a model's blocked flag.",
+                type=ProxyErrorTypes.auth_error.value,
+                code=status.HTTP_403_FORBIDDEN,
+                param="blocked",
+            )
+
         # Handle team model updates with proper alias management
         update_data = await _update_team_model_in_db(
             db_model=db_model,

litellm/proxy/management_endpoints/team_endpoints.py

@@ -73,6 +73,7 @@ from litellm.proxy.auth.auth_checks import (
 )
 from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
 from litellm.proxy.management_endpoints.common_utils import (
+    _check_passthrough_routes_caller_permission,
     _is_user_org_admin_for_team,
     _is_user_team_admin,
     _set_object_metadata_field,
@@ -1049,6 +1050,10 @@ async def new_team(  # noqa: PLR0915
                     Member(role="admin", user_id=user_api_key_dict.user_id)
                 )
 
+        _check_passthrough_routes_caller_permission(
+            data, user_api_key_dict, entity="team"
+        )
+
         ## ADD TO MODEL TABLE
         _model_id = None
         if data.model_aliases is not None and isinstance(data.model_aliases, dict):
@@ -1646,6 +1651,10 @@ async def update_team(  # noqa: PLR0915
             user_api_key_dict=user_api_key_dict,
         )
 
+        _check_passthrough_routes_caller_permission(
+            data, user_api_key_dict, entity="team"
+        )
+
         if data.soft_budget is not None:
             max_budget_to_check = (
                 data.max_budget

litellm/proxy/proxy_server.py

@@ -4727,6 +4727,7 @@ class ProxyConfig:
         if _id is not None:
             model.model_info["id"] = _id
             model.model_info["db_model"] = True
+            model.model_info["blocked"] = bool(getattr(model, "blocked", False))
 
         if premium_user is True:
             # seeing "created_at", "updated_at", "created_by", "updated_by" is a LiteLLM Enterprise Feature
@@ -6916,6 +6917,15 @@ async def async_data_generator(  # noqa: PLR0915
 
             if isinstance(chunk, BaseModel):
                 chunk = _serialize_streaming_chunk(chunk)
+            elif isinstance(chunk, bytes):
+                # Some upstream streaming iterators (e.g. AsyncGoogleGenAIGenerateContentStreamingIterator
+                # for /v1beta/.../streamGenerateContent) yield raw SSE bytes from Gemini.
+                # Decode to str so the f-string below does not emit a Python b'...' literal,
+                # and pass already-formatted SSE through unchanged to avoid double "data:" prefix.
+                chunk = chunk.decode("utf-8", errors="replace")
+                if chunk.startswith(("data:", "event:", ":")):
+                    yield chunk if chunk.endswith("\n\n") else chunk + "\n\n"
+                    continue
             elif isinstance(chunk, str) and chunk.startswith("data: "):
                 error_message = chunk
                 break
@@ -8089,6 +8099,11 @@ async def model_list(
             proxy_logging_obj=proxy_logging_obj,
         )
 
+    # Compute once — used in both branches below to hide paused models from the listing.
+    blocked_names = (
+        llm_router.get_fully_blocked_model_names() if llm_router is not None else set()
+    )
+
     # If scope=expand and user has admin privileges, return all proxy models
     if should_expand_scope:
         # Get all proxy models as if user is a proxy admin
@@ -8121,6 +8136,10 @@ async def model_list(
             only_model_access_groups=only_model_access_groups or False,
         )
 
+        # Hide paused models from the public listing (admins manage them via /model/info)
+        if blocked_names:
+            all_models = [m for m in all_models if m not in blocked_names]
+
         # Build response data with all proxy models
         model_data = []
         for model in all_models:
@@ -8154,6 +8173,10 @@ async def model_list(
         user_api_key_cache=user_api_key_cache,
     )
 
+    # Hide paused models from the public listing (admins manage them via /model/info)
+    if blocked_names:
+        all_models = [m for m in all_models if m not in blocked_names]
+
     # Build response data
     model_data = []
     for model in all_models:

litellm/proxy/route_llm_request.py

@@ -430,7 +430,11 @@ async def route_request(  # noqa: PLR0915 - Complex routing function, refactorin
                         deployment = llm_router.get_deployment_by_model_group_name(
                             model_group_name=model
                         )
-                        if deployment and deployment.litellm_params:
+                        if (
+                            deployment
+                            and deployment.litellm_params
+                            and not llm_router._is_deployment_blocked(deployment)
+                        ):
                             deployment_creds = deployment.litellm_params.model_dump(
                                 exclude_none=True
                             )

litellm/proxy/schema.prisma

@@ -48,9 +48,10 @@ model LiteLLM_CredentialsTable {
 // Models on proxy
 model LiteLLM_ProxyModelTable {
   model_id String @id @default(uuid())
-  model_name String 
+  model_name String
   litellm_params Json
-  model_info Json? 
+  model_info Json?
+  blocked Boolean @default(false)
   created_at    DateTime               @default(now()) @map("created_at")
   created_by String
   updated_at    DateTime               @default(now()) @updatedAt @map("updated_at")

litellm/router.py

@@ -30,6 +30,7 @@ from typing import (
     List,
     Literal,
     Optional,
+    Set,
     Tuple,
     Union,
     cast,
@@ -6957,12 +6958,11 @@ class Router:
         unhealthy_deployments = _get_cooldown_deployments(
             litellm_router_instance=self, parent_otel_span=parent_otel_span
         )
-        healthy_deployments: list = []
-        for deployment in _all_deployments:
-            if deployment["model_info"]["id"] in unhealthy_deployments:
-                continue
-            else:
-                healthy_deployments.append(deployment)
+        unhealthy_set = set(unhealthy_deployments)
+        healthy_deployments: list = [
+            d for d in _all_deployments if d["model_info"]["id"] not in unhealthy_set
+        ]
+        healthy_deployments = self._filter_blocked_deployments(healthy_deployments)
 
         return healthy_deployments, _all_deployments
 
@@ -6990,10 +6990,12 @@ class Router:
         )
         # Convert to set for O(1) lookup instead of O(n)
         unhealthy_deployments_set = set(unhealthy_deployments)
-        healthy_deployments: list = []
-        for deployment in _all_deployments:
-            if deployment["model_info"]["id"] not in unhealthy_deployments_set:
-                healthy_deployments.append(deployment)
+        healthy_deployments: list = [
+            d
+            for d in _all_deployments
+            if d["model_info"]["id"] not in unhealthy_deployments_set
+        ]
+        healthy_deployments = self._filter_blocked_deployments(healthy_deployments)
         return healthy_deployments, _all_deployments
 
     def routing_strategy_pre_call_checks(self, deployment: dict):
@@ -8142,10 +8144,14 @@ class Router:
 
     def get_deployment_credentials(self, model_id: str) -> Optional[dict]:
         """
-        Returns -> dict of credentials for a given model id
+        Returns -> dict of credentials for a given model id.
+
+        Returns None if the deployment is paused via `LiteLLM_ProxyModelTable.blocked`,
+        so file/batch/passthrough callers that resolve credentials directly cannot keep
+        using a paused deployment.
         """
         deployment = self.get_deployment(model_id=model_id)
-        if deployment is None:
+        if deployment is None or self._is_deployment_blocked(deployment):
             return None
         return CredentialLiteLLMParams(
             **deployment.litellm_params.model_dump(exclude_none=True)
@@ -8190,7 +8196,9 @@ class Router:
 
         Returns:
             Dictionary containing api_key, api_base, custom_llm_provider, etc.
-            Returns None if model not found.
+            Returns None if model not found, or if the resolved deployment is
+            paused via `LiteLLM_ProxyModelTable.blocked` (so passthrough callers
+            cannot bypass an admin pause by resolving credentials directly).
 
         Example:
             credentials = router.get_deployment_credentials_with_provider("gpt-4o-litellm")
@@ -8216,7 +8224,7 @@ class Router:
                 elif isinstance(deployment_dict, Deployment):
                     deployment = deployment_dict
 
-        if deployment is None:
+        if deployment is None or self._is_deployment_blocked(deployment):
             return None
 
         # Get basic credentials
@@ -9243,6 +9251,29 @@ class Router:
 
         return model_names
 
+    def get_fully_blocked_model_names(self) -> Set[str]:
+        """
+        Returns the set of model_names where every backing deployment has `blocked=True`.
+
+        Used by `/v1/models` to hide paused models from client listings while still
+        surfacing them on admin endpoints (e.g. `/model/info`). A model with at least
+        one non-blocked deployment is still serviceable and remains visible.
+        """
+        deployments = self.get_model_list() or []
+        blocked_by_name: Dict[str, bool] = {}
+        for deployment in deployments:
+            name = deployment.get("model_name") or ""
+            if not name:
+                continue
+            is_blocked = (deployment.get("model_info") or {}).get("blocked") is True
+            if name in blocked_by_name:
+                blocked_by_name[name] = blocked_by_name[name] and is_blocked
+            else:
+                blocked_by_name[name] = is_blocked
+        return {
+            name for name, fully_blocked in blocked_by_name.items() if fully_blocked
+        }
+
     def _get_team_specific_model(
         self, deployment: DeploymentTypedDict, team_id: Optional[str] = None
     ) -> Optional[str]:
@@ -10131,6 +10162,12 @@ class Router:
             )
 
         if isinstance(healthy_deployments, dict):
+            if (healthy_deployments.get("model_info") or {}).get("blocked") is True:
+                raise litellm.ServiceUnavailableError(
+                    message=f"Model '{model}' is administratively paused. Contact your proxy admin to unblock it.",
+                    model=model,
+                    llm_provider="",
+                )
             return healthy_deployments
 
         # Health-check-based filtering (before cooldown)
@@ -10164,6 +10201,8 @@ class Router:
             )
             healthy_deployments = _pre_cooldown_deployments
 
+        healthy_deployments = self._filter_blocked_deployments(healthy_deployments)
+
         healthy_deployments = await self.async_callback_filter_deployments(
             model=model,
             healthy_deployments=healthy_deployments,
@@ -10387,6 +10426,12 @@ class Router:
 
             # 3. If specific deployment returned, verify if it supports pass-through
             if isinstance(healthy_deployments, dict):
+                if (healthy_deployments.get("model_info") or {}).get("blocked") is True:
+                    raise litellm.ServiceUnavailableError(
+                        message=f"Model '{model}' is administratively paused. Contact your proxy admin to unblock it.",
+                        model=model,
+                        llm_provider="",
+                    )
                 litellm_params = healthy_deployments.get("litellm_params", {})
                 if litellm_params.get("use_in_pass_through"):
                     return healthy_deployments
@@ -10555,6 +10600,12 @@ class Router:
         )
 
         if isinstance(healthy_deployments, dict):
+            if (healthy_deployments.get("model_info") or {}).get("blocked") is True:
+                raise litellm.ServiceUnavailableError(
+                    message=f"Model '{model}' is administratively paused. Contact your proxy admin to unblock it.",
+                    model=model,
+                    llm_provider="",
+                )
             return healthy_deployments
 
         parent_otel_span: Optional[Span] = _get_parent_otel_span_from_kwargs(
@@ -10585,6 +10636,8 @@ class Router:
             )
             healthy_deployments = _pre_cooldown_deployments
 
+        healthy_deployments = self._filter_blocked_deployments(healthy_deployments)
+
         # filter pre-call checks
         if self.enable_pre_call_checks and messages is not None:
             healthy_deployments = self._pre_call_checks(
@@ -10704,6 +10757,12 @@ class Router:
 
         # 2. If the returned is a specific deployment (Dict), verify and return directly
         if isinstance(healthy_deployments, dict):
+            if (healthy_deployments.get("model_info") or {}).get("blocked") is True:
+                raise litellm.ServiceUnavailableError(
+                    message=f"Model '{model}' is administratively paused. Contact your proxy admin to unblock it.",
+                    model=model,
+                    llm_provider="",
+                )
             litellm_params = healthy_deployments.get("litellm_params", {})
             if litellm_params.get("use_in_pass_through"):
                 return healthy_deployments
@@ -10743,6 +10802,9 @@ class Router:
             healthy_deployments=pass_through_deployments,
             cooldown_deployments=cooldown_deployments,
         )
+        pass_through_deployments = self._filter_blocked_deployments(
+            pass_through_deployments
+        )
 
         # 5. Apply pre-call checks (if enabled)
         if self.enable_pre_call_checks and messages is not None:
@@ -10832,6 +10894,36 @@ class Router:
             if deployment["model_info"]["id"] not in cooldown_set
         ]
 
+    def _filter_blocked_deployments(
+        self, healthy_deployments: List[Dict]
+    ) -> List[Dict]:
+        """
+        Filters out deployments that an admin has paused via `LiteLLM_ProxyModelTable.blocked`.
+
+        Applied alongside the cooldown filter on every routing entry point that calls
+        `_common_checks_available_deployment` directly — the primary sync/async path,
+        the sync pass-through path, and the retry / health-check helpers — so paused
+        deployments never serve a request. The async pass-through path inherits this
+        filter through its delegation to `async_get_healthy_deployments`.
+        """
+        return [
+            deployment
+            for deployment in healthy_deployments
+            if (deployment.get("model_info") or {}).get("blocked") is not True
+        ]
+
+    @staticmethod
+    def _is_deployment_blocked(deployment: "Deployment") -> bool:
+        """
+        Returns True when a `Deployment` Pydantic instance carries the admin-paused
+        flag. Used by credential-lookup helpers so passthrough file / batch endpoints
+        cannot bypass the pause by resolving credentials directly.
+        """
+        model_info = getattr(deployment, "model_info", None)
+        if model_info is None:
+            return False
+        return getattr(model_info, "blocked", None) is True
+
     async def _async_filter_health_check_unhealthy_deployments(
         self,
         healthy_deployments: List[Dict],

litellm/types/integrations/prometheus.py

@@ -160,6 +160,7 @@ class UserAPIKeyLabelNames(Enum):
     END_USER = "end_user"
     USER = "user"
     USER_EMAIL = "user_email"
+    USER_ALIAS = "user_alias"
     API_KEY_HASH = "hashed_api_key"
     API_KEY_ALIAS = "api_key_alias"
     TEAM = "team"
@@ -533,17 +534,9 @@ class PrometheusMetricLabels:
         UserAPIKeyLabelNames.USER.value,
     ]
 
-    litellm_user_max_budget_metric = [
-        UserAPIKeyLabelNames.USER.value,
-    ]
+    litellm_user_max_budget_metric = litellm_remaining_user_budget_metric
 
-    litellm_user_budget_remaining_hours_metric = [
-        UserAPIKeyLabelNames.USER.value,
-    ]
-
-    litellm_user_budget_remaining_hours_metric = [
-        UserAPIKeyLabelNames.USER.value,
-    ]
+    litellm_user_budget_remaining_hours_metric = litellm_remaining_user_budget_metric
 
     litellm_remaining_api_key_requests_for_model = [
         UserAPIKeyLabelNames.API_KEY_HASH.value,
@@ -730,6 +723,22 @@ class PrometheusMetricLabels:
         ):
             custom_labels.append(UserAPIKeyLabelNames.STREAM.value)
 
+        _user_budget_metrics = {
+            "litellm_remaining_user_budget_metric",
+            "litellm_user_max_budget_metric",
+            "litellm_user_budget_remaining_hours_metric",
+        }
+        if (
+            label_name in _user_budget_metrics
+            and litellm.prometheus_user_budget_label_include_email_alias is True
+        ):
+            for label in [
+                UserAPIKeyLabelNames.USER_EMAIL.value,
+                UserAPIKeyLabelNames.USER_ALIAS.value,
+            ]:
+                if label not in default_labels and label not in custom_labels:
+                    custom_labels.append(label)
+
         if label_name in PrometheusMetricLabels._org_label_metrics:
             for label in [
                 UserAPIKeyLabelNames.ORG_ID.value,
@@ -759,6 +768,7 @@ class UserAPIKeyLabelValues:
     end_user: Optional[str] = None
     user: Optional[str] = None
     user_email: Optional[str] = None
+    user_alias: Optional[str] = None
     hashed_api_key: Optional[str] = None
     api_key_alias: Optional[str] = None
     team: Optional[str] = None

litellm/types/router.py

@@ -133,6 +133,9 @@ class ModelInfo(BaseModel):
     # the model_name that can be used by the team when making LLM calls
     team_public_model_name: Optional[str] = None
 
+    # admin-toggled pause flag; mirrors LiteLLM_ProxyModelTable.blocked
+    blocked: Optional[bool] = None
+
     def __init__(self, id: Optional[Union[str, int]] = None, **params):
         if id is None:
             id = str(uuid.uuid4())  # Generate a UUID if id is None or not provided
@@ -323,6 +326,7 @@ class updateDeployment(BaseModel):
     model_name: Optional[str] = None
     litellm_params: Optional[updateLiteLLMParams] = None
     model_info: Optional[ModelInfo] = None
+    blocked: Optional[bool] = None
 
     model_config = ConfigDict(protected_namespaces=())
 

litellm/types/utils.py

@@ -107,9 +107,13 @@ class LiteLLMCommonStrings(Enum):
 SupportedCacheControls = ["ttl", "s-maxage", "no-cache", "no-store"]
 
 
-class CostPerToken(TypedDict):
-    input_cost_per_token: float
-    output_cost_per_token: float
+class CostPerToken(TypedDict, total=False):
+    # Required base rates — kept under total=False so we can mark them
+    # Required individually while leaving the cache rates NotRequired.
+    input_cost_per_token: Required[float]
+    output_cost_per_token: Required[float]
+    cache_read_input_token_cost: float
+    cache_creation_input_token_cost: float
 
 
 class ProviderField(TypedDict):

litellm/utils.py

@@ -8402,6 +8402,12 @@ class ProviderConfigManager:
             )
 
             return VolcEngineEmbeddingConfig()
+        elif litellm.LlmProviders.DASHSCOPE == provider:
+            from litellm.llms.dashscope.embed.transformation import (
+                DashScopeEmbeddingConfig,
+            )
+
+            return DashScopeEmbeddingConfig()
         elif litellm.LlmProviders.OVHCLOUD == provider:
             return litellm.OVHCloudEmbeddingConfig()
         elif litellm.LlmProviders.SNOWFLAKE == provider:
@@ -8481,6 +8487,12 @@ class ProviderConfigManager:
             return litellm.VoyageRerankConfig()
         elif litellm.LlmProviders.WATSONX == provider:
             return litellm.IBMWatsonXRerankConfig()
+        elif litellm.LlmProviders.DASHSCOPE == provider:
+            from litellm.llms.dashscope.rerank.transformation import (
+                DashScopeRerankConfig,
+            )
+
+            return DashScopeRerankConfig()
         return litellm.CohereRerankConfig()
 
     @staticmethod
@@ -8527,6 +8539,12 @@ class ProviderConfigManager:
             )
 
             return MinimaxMessagesConfig()
+        elif litellm.LlmProviders.DEEPSEEK == provider:
+            from litellm.llms.deepseek.messages.transformation import (
+                DeepSeekAnthropicMessagesConfig,
+            )
+
+            return DeepSeekAnthropicMessagesConfig()
         return None
 
     @staticmethod

schema.prisma

@@ -48,9 +48,10 @@ model LiteLLM_CredentialsTable {
 // Models on proxy
 model LiteLLM_ProxyModelTable {
   model_id String @id @default(uuid())
-  model_name String 
+  model_name String
   litellm_params Json
-  model_info Json? 
+  model_info Json?
+  blocked Boolean @default(false)
   created_at    DateTime               @default(now()) @map("created_at")
   created_by String
   updated_at    DateTime               @default(now()) @updatedAt @map("updated_at")

tests/_vcr_conftest_common.py

@@ -36,6 +36,75 @@ SAFE_BODY_MATCHER_NAME = "safe_body"
 KEY_FINGERPRINT_MATCHER_NAME = "key_fingerprint"
 KEY_FINGERPRINT_HEADER = "x-litellm-key-fp"
 
+VCR_DIAG_DIR_ENV = "LITELLM_VCR_DIAG_DIR"
+VCR_DIAG_DIR_DEFAULT = "test-results/vcr-diagnostics"
+
+
+def _vcr_diag_dir() -> str:
+    return os.environ.get(VCR_DIAG_DIR_ENV) or VCR_DIAG_DIR_DEFAULT
+
+
+def vcr_diag_write_line(msg: str) -> None:
+    try:
+        directory = _vcr_diag_dir()
+        os.makedirs(directory, exist_ok=True)
+        path = os.path.join(directory, f"{os.getpid()}.log")
+        with open(path, "a", encoding="utf-8") as fh:
+            fh.write(msg.rstrip("\n") + "\n")
+    except OSError:
+        pass
+
+
+def reset_vcr_diag_dir() -> None:
+    if os.environ.get("PYTEST_XDIST_WORKER"):
+        return
+    directory = _vcr_diag_dir()
+    if not os.path.isdir(directory):
+        return
+    try:
+        names = os.listdir(directory)
+    except OSError:
+        return
+    for name in names:
+        if name.endswith(".log"):
+            try:
+                os.remove(os.path.join(directory, name))
+            except OSError:
+                pass
+
+
+def emit_vcr_diagnostic_log(terminalreporter) -> None:
+    directory = _vcr_diag_dir()
+    if not os.path.isdir(directory):
+        return
+    try:
+        files = sorted(f for f in os.listdir(directory) if f.endswith(".log"))
+    except OSError:
+        return
+    if not files:
+        return
+    terminalreporter.write_sep("=", "VCR DIAGNOSTIC LOG", bold=True)
+    terminalreporter.write_line(
+        f"  source dir: {directory}  (also archived as a CI artifact)"
+    )
+    for name in files:
+        path = os.path.join(directory, name)
+        try:
+            with open(path, "r", encoding="utf-8") as fh:
+                content = fh.read()
+        except OSError as exc:
+            terminalreporter.write_line(
+                f"  [failed to read {name}: {type(exc).__name__}: {exc}]"
+            )
+            continue
+        if not content.strip():
+            continue
+        terminalreporter.write_sep("-", name, bold=False)
+        for line in content.splitlines():
+            terminalreporter.write_line(line)
+    terminalreporter.write_sep("=", bold=True)
+
+
 # Intentionally narrower than ``FILTERED_REQUEST_HEADERS``: AWS SigV4 headers
 # carry secrets but their values rotate on every call, so fingerprinting them
 # would defeat caching.
@@ -91,6 +160,32 @@ VCR_IMAGE_B64_PLACEHOLDER = "dGVzdA=="
 VCR_FIXED_MULTIPART_BOUNDARY = "vcr-static-boundary"
 
 
+def pin_httpx_multipart_boundary(monkeypatch) -> None:
+    try:
+        import httpx._multipart as _httpx_multipart
+    except ImportError:
+        return
+
+    _original_init = _httpx_multipart.MultipartStream.__init__
+
+    def _init_with_fixed_boundary(self, data, files, boundary=None, **kwargs):
+        if boundary is None:
+            boundary = VCR_FIXED_MULTIPART_BOUNDARY.encode("ascii")
+        return _original_init(self, data=data, files=files, boundary=boundary, **kwargs)
+
+    monkeypatch.setattr(
+        _httpx_multipart.MultipartStream, "__init__", _init_with_fixed_boundary
+    )
+
+
+@pytest.fixture(scope="session", autouse=True)
+def _pin_multipart_boundary():
+    monkeypatch = pytest.MonkeyPatch()
+    pin_httpx_multipart_boundary(monkeypatch)
+    yield
+    monkeypatch.undo()
+
+
 def _scrub_response(response):
     if not isinstance(response, dict):
         return response
@@ -139,9 +234,17 @@ def _strip_image_b64_payloads(response):
     preserves all those checks while shrinking cassettes by ~99%.
     """
     if not isinstance(response, dict):
+        vcr_diag_write_line(
+            f"[vcr-strip-b64] response is {type(response).__name__!r}, not "
+            "dict; skipping b64 scrub"
+        )
         return response
     body = response.get("body")
     if not isinstance(body, dict):
+        vcr_diag_write_line(
+            f"[vcr-strip-b64] response['body'] is {type(body).__name__!r}, "
+            "not dict; skipping b64 scrub"
+        )
         return response
     raw = body.get("string")
     if raw is None:
@@ -151,12 +254,20 @@ def _strip_image_b64_payloads(response):
         try:
             text = bytes(raw).decode("utf-8")
         except UnicodeDecodeError:
+            vcr_diag_write_line(
+                "[vcr-strip-b64] response body bytes are not valid UTF-8; "
+                "skipping b64 scrub"
+            )
             return response
         was_bytes = True
     elif isinstance(raw, str):
         text = raw
         was_bytes = False
     else:
+        vcr_diag_write_line(
+            f"[vcr-strip-b64] response['body']['string'] is "
+            f"{type(raw).__name__!r}, not bytes/str; skipping b64 scrub"
+        )
         return response
 
     try:
@@ -186,6 +297,35 @@ def _before_record_response(response):
     return filter_non_2xx_response(_scrub_response(_strip_image_b64_payloads(response)))
 
 
+def _canonical_body(request) -> tuple[bytes, str]:
+    pre_type = type(getattr(request, "body", None)).__name__
+    _materialize_iterable_body(request)
+    body = getattr(request, "body", None)
+    if body is None:
+        return b"", pre_type
+    if isinstance(body, bytes):
+        return body, pre_type
+    if isinstance(body, bytearray):
+        return bytes(body), pre_type
+    if isinstance(body, str):
+        return body.encode("utf-8"), pre_type
+    if isinstance(body, (dict, list)):
+        try:
+            return (
+                json.dumps(body, sort_keys=True, separators=(",", ":")).encode("utf-8"),
+                pre_type,
+            )
+        except (TypeError, ValueError):
+            pass
+    method = getattr(request, "method", "?")
+    uri = getattr(request, "uri", getattr(request, "url", "?"))
+    vcr_diag_write_line(
+        f"[vcr-canonical-body] FALLBACK: {method} {uri} body type "
+        f"{type(body).__name__!r} not coerced to bytes; comparing as b''"
+    )
+    return b"", pre_type
+
+
 def _safe_body_matcher(r1, r2) -> None:
     """Compare request bodies as bytes; never invokes ``json.loads``.
 
@@ -195,27 +335,47 @@ def _safe_body_matcher(r1, r2) -> None:
     This matcher is strictly more conservative — the only equivalence
     it gives up vs. the default is "JSON key order doesn't matter".
     """
-    body1 = getattr(r1, "body", None)
-    body2 = getattr(r2, "body", None)
+    body1, pre1 = _canonical_body(r1)
+    body2, pre2 = _canonical_body(r2)
     if body1 == body2:
         return
-
-    def _to_bytes(b):
-        if b is None:
-            return b""
-        if isinstance(b, bytes):
-            return b
-        if isinstance(b, str):
-            return b.encode("utf-8")
-        return None
-
-    n1 = _to_bytes(body1)
-    n2 = _to_bytes(body2)
-    if n1 is not None and n2 is not None and n1 == n2:
-        return
+    _emit_body_mismatch_diagnostic(r1, r2, body1, body2, pre1, pre2)
     raise AssertionError("request bodies differ")
 
 
+def _emit_body_mismatch_diagnostic(r1, r2, body1, body2, pre1, pre2) -> None:
+    def _describe(label, asbytes, pre_type):
+        return (
+            f"  {label}: pre_canonical_type={pre_type!r} length={len(asbytes)} "
+            f"sha256={hashlib.sha256(asbytes).hexdigest()} "
+            f"preview={asbytes[:120]!r}"
+        )
+
+    method_a = getattr(r1, "method", "?")
+    method_b = getattr(r2, "method", "?")
+    url_a = getattr(r1, "uri", getattr(r1, "url", "?"))
+    url_b = getattr(r2, "uri", getattr(r2, "url", "?"))
+    lines = [
+        "[vcr-safe-body-matcher] request body mismatch",
+        f"  request[a]: {method_a} {url_a}",
+        f"  request[b]: {method_b} {url_b}",
+        _describe("body[a]", body1, pre1),
+        _describe("body[b]", body2, pre2),
+    ]
+    if body1 != body2:
+        offset = next(
+            (i for i in range(min(len(body1), len(body2))) if body1[i] != body2[i]),
+            min(len(body1), len(body2)),
+        )
+        start = max(0, offset - 100)
+        end_a = min(len(body1), offset + 100)
+        end_b = min(len(body2), offset + 100)
+        lines.append(f"  first divergent byte offset: {offset}")
+        lines.append(f"  window[a] @ {start}..{end_a}: {body1[start:end_a]!r}")
+        lines.append(f"  window[b] @ {start}..{end_b}: {body2[start:end_b]!r}")
+    vcr_diag_write_line("\n".join(lines))
+
+
 def _iter_header_values(headers, name: str):
     if headers is None:
         return
@@ -271,6 +431,13 @@ def _compute_key_fingerprint(request) -> str:
             stable = _stable_key_value(header_name, text)
             parts.append(f"{header_name}={stable}")
     if not parts:
+        method = getattr(request, "method", "?")
+        uri = getattr(request, "uri", getattr(request, "url", "?"))
+        vcr_diag_write_line(
+            f"[vcr-key-fingerprint] no API key header found on {method} "
+            f"{uri}; falling back to 'no-key'. If this request should have "
+            "carried auth, something earlier in the pipeline stripped it."
+        )
         return "no-key"
     digest = hashlib.sha256("\n".join(parts).encode("utf-8")).hexdigest()
     return digest[:16]
@@ -360,6 +527,13 @@ def _normalize_multipart_boundary(request) -> None:
     elif isinstance(body, str):
         new_body = body.replace(current_boundary, VCR_FIXED_MULTIPART_BOUNDARY)
     else:
+        vcr_diag_write_line(
+            f"[vcr-multipart-normalize] body normalization SKIPPED: "
+            f"body type {type(body).__name__!r} is not bytes/bytearray/str. "
+            f"content-type={content_type_value!r}. "
+            f"Recorded body will retain the random boundary substring "
+            f"and the safe_body matcher will miss on the next run."
+        )
         return
 
     try:
@@ -389,6 +563,7 @@ def _before_record_request(request):
     headers = getattr(request, "headers", None)
     if headers is None:
         return request
+    _materialize_iterable_body(request)
     if not any(_iter_header_values(headers, KEY_FINGERPRINT_HEADER)):
         fingerprint = _compute_key_fingerprint(request)
         try:
@@ -400,6 +575,56 @@ def _before_record_request(request):
     return request
 
 
+def _materialize_iterable_body(request) -> None:
+    body = getattr(request, "body", None)
+    if body is None or isinstance(body, (bytes, bytearray, str)):
+        return
+    if not hasattr(body, "__next__"):
+        return
+    try:
+        chunks = list(body)
+    except TypeError:
+        return
+
+    out = _coalesce_chunks_to_bytes(chunks)
+    if out is None:
+        method = getattr(request, "method", "?")
+        uri = getattr(request, "uri", getattr(request, "url", "?"))
+        first_type = type(chunks[0]).__name__ if chunks else "empty"
+        vcr_diag_write_line(
+            f"[vcr-materialize] FALLBACK: {method} {uri} chunk type "
+            f"{first_type!r} not coerced to bytes; storing b''"
+        )
+        out = b""
+
+    try:
+        request.body = out
+    except (AttributeError, TypeError):
+        pass
+
+    for attr in ("_was_iter", "_was_file"):
+        try:
+            setattr(request, attr, False)
+        except (AttributeError, TypeError):
+            pass
+
+
+def _coalesce_chunks_to_bytes(chunks):
+    if not chunks:
+        return b""
+    first = chunks[0]
+    try:
+        if isinstance(first, int):
+            return bytes(chunks)
+        if isinstance(first, (bytes, bytearray)):
+            return b"".join(c if isinstance(c, bytes) else bytes(c) for c in chunks)
+        if isinstance(first, str):
+            return "".join(chunks).encode("utf-8")
+    except (TypeError, ValueError):
+        return None
+    return None
+
+
 def _key_fingerprint_matcher(r1, r2) -> None:
     def _fp(req):
         for value in _iter_header_values(
@@ -410,7 +635,17 @@ def _key_fingerprint_matcher(r1, r2) -> None:
             return value if isinstance(value, str) else str(value)
         return "no-key"
 
-    if _fp(r1) != _fp(r2):
+    fp1, fp2 = _fp(r1), _fp(r2)
+    if fp1 != fp2:
+        method_a = getattr(r1, "method", "?")
+        method_b = getattr(r2, "method", "?")
+        url_a = getattr(r1, "uri", getattr(r1, "url", "?"))
+        url_b = getattr(r2, "uri", getattr(r2, "url", "?"))
+        vcr_diag_write_line(
+            "[vcr-key-fingerprint-matcher] API key fingerprints differ\n"
+            f"  request[a]: {method_a} {url_a} fingerprint={fp1!r}\n"
+            f"  request[b]: {method_b} {url_b} fingerprint={fp2!r}"
+        )
         raise AssertionError("API key fingerprints differ")
 
 

tests/_vcr_redis_persister.py

@@ -159,9 +159,20 @@ def make_redis_persister(
                 raise CassetteNotFoundError() from exc
             if data is None:
                 raise CassetteNotFoundError()
-            if isinstance(data, bytes):
-                data = data.decode("utf-8")
-            return deserialize(data, serializer)
+            try:
+                if isinstance(data, bytes):
+                    data = data.decode("utf-8")
+                return deserialize(data, serializer)
+            except Exception as exc:
+                _record_cache_failure("load", exc)
+                msg = (
+                    f"VCR redis load failed for {cassette_path}; cached "
+                    f"payload is corrupt, treating as cache miss: "
+                    f"{type(exc).__name__}: {exc}"
+                )
+                _log.warning(msg)
+                warnings.warn(msg, VCRCassetteCacheWarning, stacklevel=2)
+                raise CassetteNotFoundError() from exc
 
         @staticmethod
         def save_cassette(cassette_path, cassette_dict, serializer):

tests/audio_tests/conftest.py

@@ -5,14 +5,17 @@ import pytest
 
 sys.path.insert(0, os.path.abspath("../.."))
 
-from tests._vcr_conftest_common import (  # noqa: E402
+from tests._vcr_conftest_common import (  # noqa: E402,F401
     VerboseReporterState,
+    _pin_multipart_boundary,
     apply_vcr_auto_marker_to_items,
     emit_cassette_cache_session_banner,
     emit_vcr_classification_summary,
+    emit_vcr_diagnostic_log,
     install_live_call_probe,
     record_vcr_outcome,
     register_persister_if_enabled,
+    reset_vcr_diag_dir,
     vcr_config_dict,
 )
 
@@ -44,6 +47,7 @@ def _vcr_outcome_gate(request, vcr):
 
 def pytest_configure(config):
     _verbose_state.remember_pluginmanager(config)
+    reset_vcr_diag_dir()
 
 
 def pytest_runtest_logreport(report):
@@ -57,3 +61,4 @@ def pytest_collection_modifyitems(config, items):
 def pytest_terminal_summary(terminalreporter, exitstatus, config):
     emit_cassette_cache_session_banner(terminalreporter)
     emit_vcr_classification_summary(terminalreporter)
+    emit_vcr_diagnostic_log(terminalreporter)

tests/audio_tests/test_whisper.py

@@ -23,12 +23,21 @@ pwd = os.path.dirname(os.path.realpath(__file__))
 print(pwd)
 
 file_path = os.path.join(pwd, "gettysburg.wav")
-
-audio_file = open(file_path, "rb")
-
-
 file2_path = os.path.join(pwd, "eagle.wav")
-audio_file2 = open(file2_path, "rb")
+
+with open(file_path, "rb") as _f:
+    _GETTYSBURG_BYTES = _f.read()
+with open(file2_path, "rb") as _f:
+    _EAGLE_BYTES = _f.read()
+
+
+def _audio_file():
+    return ("gettysburg.wav", _GETTYSBURG_BYTES, "audio/wav")
+
+
+def _audio_file2():
+    return ("eagle.wav", _EAGLE_BYTES, "audio/wav")
+
 
 load_dotenv()
 
@@ -44,7 +53,7 @@ async def _run_transcription(
 ):
     transcript = await litellm.atranscription(
         model=model,
-        file=audio_file,
+        file=_audio_file(),
         api_key=api_key,
         api_base=api_base,
         response_format=response_format,
@@ -101,7 +110,7 @@ async def test_transcription_caching():
 
     response_1 = await litellm.atranscription(
         model="whisper-1",
-        file=audio_file,
+        file=_audio_file(),
     )
 
     await asyncio.sleep(5)
@@ -110,7 +119,7 @@ async def test_transcription_caching():
 
     response_2 = await litellm.atranscription(
         model="whisper-1",
-        file=audio_file,
+        file=_audio_file(),
     )
 
     print("response_1", response_1)
@@ -122,7 +131,7 @@ async def test_transcription_caching():
 
     response_3 = await litellm.atranscription(
         model="whisper-1",
-        file=audio_file2,
+        file=_audio_file2(),
     )
     print("response_3", response_3)
     print("response3 hidden params", response_3._hidden_params)
@@ -146,7 +155,7 @@ async def test_whisper_log_pre_call():
     with patch.object(custom_logger, "log_pre_api_call") as mock_log_pre_call:
         await litellm.atranscription(
             model="whisper-1",
-            file=audio_file,
+            file=_audio_file(),
         )
         mock_log_pre_call.assert_called_once()
 
@@ -165,7 +174,7 @@ async def test_whisper_log_pre_call():
     with patch.object(custom_logger, "log_pre_api_call") as mock_log_pre_call:
         await litellm.atranscription(
             model="whisper-1",
-            file=audio_file,
+            file=_audio_file(),
         )
         mock_log_pre_call.assert_called_once()
 
@@ -177,7 +186,7 @@ async def test_gpt_4o_transcribe():
     from unittest.mock import patch, MagicMock
 
     await litellm.atranscription(
-        model="openai/gpt-4o-transcribe", file=audio_file, response_format="json"
+        model="openai/gpt-4o-transcribe", file=_audio_file(), response_format="json"
     )
 
 
@@ -187,7 +196,9 @@ async def test_gpt_4o_transcribe_model_mapping():
 
     # Test GPT-4o mini transcribe
     response = await litellm.atranscription(
-        model="openai/gpt-4o-mini-transcribe", file=audio_file, response_format="json"
+        model="openai/gpt-4o-mini-transcribe",
+        file=_audio_file(),
+        response_format="json",
     )
 
     # Check that the response contains the correct model in hidden params
@@ -198,7 +209,7 @@ async def test_gpt_4o_transcribe_model_mapping():
 
     # Test GPT-4o transcribe
     response2 = await litellm.atranscription(
-        model="openai/gpt-4o-transcribe", file=audio_file, response_format="json"
+        model="openai/gpt-4o-transcribe", file=_audio_file(), response_format="json"
     )
 
     # Check that the response contains the correct model in hidden params
@@ -209,7 +220,7 @@ async def test_gpt_4o_transcribe_model_mapping():
 
     # Test traditional whisper-1 still works
     response3 = await litellm.atranscription(
-        model="openai/whisper-1", file=audio_file, response_format="json"
+        model="openai/whisper-1", file=_audio_file(), response_format="json"
     )
 
     # Check that the response contains the correct model in hidden params
@@ -262,7 +273,7 @@ async def test_azure_transcribe_model_mapping():
         # Make the transcription call
         response = await litellm.atranscription(
             model="azure/whisper-1",
-            file=audio_file,
+            file=_audio_file(),
             response_format="json",
             api_key="test-api-key",
             api_base="https://my-endpoint-europe-berri-992.openai.azure.com/",

tests/guardrails_tests/conftest.py

@@ -16,14 +16,17 @@ sys.path.insert(
 )  # Adds the parent directory to the system path
 import litellm
 
-from tests._vcr_conftest_common import (  # noqa: E402
+from tests._vcr_conftest_common import (  # noqa: E402,F401
     VerboseReporterState,
+    _pin_multipart_boundary,
     apply_vcr_auto_marker_to_items,
     emit_cassette_cache_session_banner,
     emit_vcr_classification_summary,
+    emit_vcr_diagnostic_log,
     install_live_call_probe,
     record_vcr_outcome,
     register_persister_if_enabled,
+    reset_vcr_diag_dir,
     vcr_config_dict,
 )
 
@@ -55,6 +58,7 @@ def _vcr_outcome_gate(request, vcr):
 
 def pytest_configure(config):
     _verbose_state.remember_pluginmanager(config)
+    reset_vcr_diag_dir()
 
 
 def pytest_runtest_logreport(report):
@@ -160,3 +164,4 @@ def pytest_collection_modifyitems(config, items):
 def pytest_terminal_summary(terminalreporter, exitstatus, config):
     emit_cassette_cache_session_banner(terminalreporter)
     emit_vcr_classification_summary(terminalreporter)
+    emit_vcr_diagnostic_log(terminalreporter)

tests/image_gen_tests/conftest.py

@@ -9,14 +9,17 @@ sys.path.insert(
 )  # Adds the parent directory to the system path
 import litellm  # noqa: E402,F401
 
-from tests._vcr_conftest_common import (  # noqa: E402
+from tests._vcr_conftest_common import (  # noqa: E402,F401
     VerboseReporterState,
+    _pin_multipart_boundary,
     apply_vcr_auto_marker_to_items,
     emit_cassette_cache_session_banner,
     emit_vcr_classification_summary,
+    emit_vcr_diagnostic_log,
     install_live_call_probe,
     record_vcr_outcome,
     register_persister_if_enabled,
+    reset_vcr_diag_dir,
     vcr_config_dict,
 )
 
@@ -58,6 +61,7 @@ def _vcr_outcome_gate(request, vcr):
 
 def pytest_configure(config):
     _verbose_state.remember_pluginmanager(config)
+    reset_vcr_diag_dir()
 
 
 def pytest_runtest_logreport(report):
@@ -71,3 +75,4 @@ def pytest_collection_modifyitems(config, items):
 def pytest_terminal_summary(terminalreporter, exitstatus, config):
     emit_cassette_cache_session_banner(terminalreporter)
     emit_vcr_classification_summary(terminalreporter)
+    emit_vcr_diagnostic_log(terminalreporter)

tests/image_gen_tests/test_image_edits.py

@@ -103,12 +103,6 @@ class BaseLLMImageEditTest(ABC):
 pwd = os.path.dirname(os.path.realpath(__file__))
 
 
-# Image fixtures must be regenerated per access — module-level
-# ``open(...)`` handles get consumed after a single multipart upload, leaving
-# subsequent tests in the same process to send empty bodies. That non-determinism
-# (a) blows the recorded cassette past ``MAX_EPISODES_PER_CASSETTE`` so the
-# persister refuses to save (see ``tests/_vcr_redis_persister.py``), and
-# (b) re-bills the live image edit endpoint on every CI run.
 def _read_image_bytes(filename: str) -> bytes:
     with open(os.path.join(pwd, filename), "rb") as f:
         return f.read()
@@ -119,32 +113,20 @@ _LITELLM_SITE_BYTES = _read_image_bytes("litellm_site.png")
 
 
 def _make_test_images() -> list:
-    """Return a fresh pair of image streams seeded with the fixture bytes.
+    return [_ISHAAN_GITHUB_BYTES, _LITELLM_SITE_BYTES]
 
-    Use this everywhere you'd previously have used the module-level
-    ``TEST_IMAGES``. Each call returns brand new ``BytesIO`` objects whose
-    file pointers start at 0, so multipart uploads encode the full image
-    bytes on every test invocation. Parametrized and ``flaky``-retried
-    test methods call ``get_base_image_edit_call_args`` once per
-    invocation, so a fresh stream per call is sufficient — the factory
-    must not auto-rewind on EOF or the SDK's multipart writer will read
-    the same bytes forever (worker OOM).
-    """
+
+def _make_single_test_image() -> bytes:
+    return _ISHAAN_GITHUB_BYTES
+
+
+def get_test_images_as_bytesio():
     return [
         BytesIO(_ISHAAN_GITHUB_BYTES),
         BytesIO(_LITELLM_SITE_BYTES),
     ]
 
 
-def _make_single_test_image() -> BytesIO:
-    return BytesIO(_ISHAAN_GITHUB_BYTES)
-
-
-def get_test_images_as_bytesio():
-    """Helper function to get test images as BytesIO objects"""
-    return _make_test_images()
-
-
 class TestOpenAIImageEditGPTImage1(BaseLLMImageEditTest):
     """
     Concrete implementation of BaseLLMImageEditTest for OpenAI image edits.
@@ -710,10 +692,9 @@ async def test_multiple_image_edit_with_different_formats():
     try:
         prompt = "Create a cohesive artistic style across all images"
 
-        # Test with mixed BytesIO and file objects
         mixed_images = [
-            _make_single_test_image(),  # File object
-            get_test_images_as_bytesio()[1],  # BytesIO object
+            _make_single_test_image(),
+            get_test_images_as_bytesio()[1],
         ]
 
         result = await aimage_edit(

tests/litellm_utils_tests/conftest.py

@@ -12,14 +12,17 @@ sys.path.insert(
 )  # Adds the parent directory to the system path
 import litellm  # noqa: E402,F401
 
-from tests._vcr_conftest_common import (  # noqa: E402
+from tests._vcr_conftest_common import (  # noqa: E402,F401
     VerboseReporterState,
+    _pin_multipart_boundary,
     apply_vcr_auto_marker_to_items,
     emit_cassette_cache_session_banner,
     emit_vcr_classification_summary,
+    emit_vcr_diagnostic_log,
     install_live_call_probe,
     record_vcr_outcome,
     register_persister_if_enabled,
+    reset_vcr_diag_dir,
     vcr_config_dict,
 )
 
@@ -86,6 +89,7 @@ def _vcr_outcome_gate(request, vcr):
 
 def pytest_configure(config):
     _verbose_state.remember_pluginmanager(config)
+    reset_vcr_diag_dir()
 
 
 def pytest_runtest_logreport(report):
@@ -116,3 +120,4 @@ def pytest_collection_modifyitems(config, items):
 def pytest_terminal_summary(terminalreporter, exitstatus, config):
     emit_cassette_cache_session_banner(terminalreporter)
     emit_vcr_classification_summary(terminalreporter)
+    emit_vcr_diagnostic_log(terminalreporter)

tests/llm_responses_api_testing/conftest.py

@@ -13,14 +13,17 @@ sys.path.insert(
 
 import litellm  # noqa: E402
 
-from tests._vcr_conftest_common import (  # noqa: E402
+from tests._vcr_conftest_common import (  # noqa: E402,F401
     VerboseReporterState,
+    _pin_multipart_boundary,
     apply_vcr_auto_marker_to_items,
     emit_cassette_cache_session_banner,
     emit_vcr_classification_summary,
+    emit_vcr_diagnostic_log,
     install_live_call_probe,
     record_vcr_outcome,
     register_persister_if_enabled,
+    reset_vcr_diag_dir,
     vcr_config_dict,
 )
 
@@ -52,6 +55,7 @@ def _vcr_outcome_gate(request, vcr):
 
 def pytest_configure(config):
     _verbose_state.remember_pluginmanager(config)
+    reset_vcr_diag_dir()
 
 
 def pytest_runtest_logreport(report):
@@ -116,3 +120,4 @@ def pytest_collection_modifyitems(config, items):
 def pytest_terminal_summary(terminalreporter, exitstatus, config):
     emit_cassette_cache_session_banner(terminalreporter)
     emit_vcr_classification_summary(terminalreporter)
+    emit_vcr_diagnostic_log(terminalreporter)

tests/llm_translation/conftest.py

@@ -18,14 +18,17 @@ sys.path.insert(
 
 import litellm  # noqa: E402
 
-from tests._vcr_conftest_common import (  # noqa: E402
+from tests._vcr_conftest_common import (  # noqa: E402,F401
     VerboseReporterState,
+    _pin_multipart_boundary,
     apply_vcr_auto_marker_to_items,
     emit_cassette_cache_session_banner,
     emit_vcr_classification_summary,
+    emit_vcr_diagnostic_log,
     install_live_call_probe,
     record_vcr_outcome,
     register_persister_if_enabled,
+    reset_vcr_diag_dir,
     vcr_config_dict,
 )
 
@@ -73,6 +76,7 @@ def _vcr_outcome_gate(request, vcr):
 
 def pytest_configure(config):
     _verbose_state.remember_pluginmanager(config)
+    reset_vcr_diag_dir()
 
 
 def pytest_runtest_logreport(report):
@@ -82,6 +86,7 @@ def pytest_runtest_logreport(report):
 def pytest_terminal_summary(terminalreporter, exitstatus, config):
     emit_cassette_cache_session_banner(terminalreporter)
     emit_vcr_classification_summary(terminalreporter)
+    emit_vcr_diagnostic_log(terminalreporter)
 
 
 # ---------------------------------------------------------------------------

tests/llm_translation/realtime/test_openai_realtime.py

@@ -101,7 +101,9 @@ async def test_openai_realtime_direct_call_no_intent():
 
     try:
         await litellm._arealtime(
-            model="openai/gpt-4o-realtime-preview",
+            # OpenAI shut down the gpt-4o-realtime-preview family (incl. the
+            # undated alias) on 2026-05-07; gpt-realtime is the GA successor.
+            model="openai/gpt-realtime",
             websocket=websocket_client,
             api_key=os.environ.get("OPENAI_API_KEY"),
             timeout=60,
@@ -249,14 +251,16 @@ async def test_openai_realtime_direct_call_with_intent():
     websocket_client = RealTimeWebSocketClient()
     caught_exception = None
 
+    # OpenAI shut down the gpt-4o-realtime-preview family (incl. the undated
+    # alias) on 2026-05-07; gpt-realtime is the GA successor.
     query_params: RealtimeQueryParams = {
-        "model": "openai/gpt-4o-realtime-preview",
+        "model": "openai/gpt-realtime",
         "intent": "chat",
     }
 
     try:
         await litellm._arealtime(
-            model="openai/gpt-4o-realtime-preview",
+            model="openai/gpt-realtime",
             websocket=websocket_client,
             api_key=os.environ.get("OPENAI_API_KEY"),
             query_params=query_params,

tests/llm_translation/realtime/test_openai_realtime_simple.py

@@ -21,7 +21,10 @@ class TestOpenAIRealtime(BaseRealtimeTest):
     """
 
     def get_model(self) -> str:
-        return "gpt-4o-realtime-preview"
+        # OpenAI shut down the entire gpt-4o-realtime-preview family
+        # (including the undated alias) on 2026-05-07. gpt-realtime is the
+        # current GA realtime model.
+        return "gpt-realtime"
 
     def get_api_key_env_var(self) -> str:
         return "OPENAI_API_KEY"

tests/llm_translation/realtime/test_realtime_guardrails_openai.py

@@ -26,9 +26,7 @@ from litellm.litellm_core_utils.realtime_streaming import RealTimeStreaming
 from litellm.types.guardrails import GuardrailEventHooks
 
 OPENAI_API_KEY = os.environ.get("OPENAI_API_KEY")
-OPENAI_REALTIME_URL = (
-    "wss://api.openai.com/v1/realtime?model=gpt-4o-realtime-preview-2024-12-17"
-)
+OPENAI_REALTIME_URL = "wss://api.openai.com/v1/realtime?model=gpt-realtime"
 
 pytestmark = pytest.mark.skipif(
     not OPENAI_API_KEY,
@@ -192,10 +190,35 @@ async def test_text_message_blocked_by_guardrail_no_ai_response():
             len(transcript_deltas) >= 1
         ), f"Expected guardrail message in transcript delta, got: {event_types}"
 
-        # 3. No *real* AI response should have been generated.
-        #    The guardrail may produce its own response (e.g. "Content blocked: ...")
-        #    via response.cancel + conversation.item.create + response.create.
-        #    We allow the guardrail's own block message but NOT original AI content.
+        # 3. No *real* AI response to the blocked content should have been
+        #    generated. The original user message is blocked BEFORE it is
+        #    forwarded to OpenAI, so the only thing the model ever sees is the
+        #    guardrail's "say exactly: <block message>" prompt
+        #    (see realtime_streaming.py). Two safe outcomes are possible:
+        #      - the model voices the block message verbatim (older realtime
+        #        snapshots did this -> text contains "blocked"), or
+        #      - the model declines to repeat it (gpt-realtime tends to refuse
+        #        verbatim-repeat instructions, e.g. "I'm sorry, but I can't
+        #        repeat that message.").
+        #    Both mean the blocked prompt itself was never answered, so we
+        #    accept either. The hard invariant is that the blocked phrase must
+        #    never leak into AI output, and the model must not have produced a
+        #    normal answer to the user (which would have neither a block nor a
+        #    refusal marker).
+        safe_markers = (
+            "block",
+            "guardrail",
+            "content filter",
+            "policy",
+            "can't repeat",
+            "cannot repeat",
+            "won't repeat",
+            "can't assist",
+            "can't help",
+            "unable to",
+            "i'm sorry",
+            "i am sorry",
+        )
         done_events = [e for e in client_events if e.get("type") == "response.done"]
         for done in done_events:
             output = done.get("response", {}).get("output", [])
@@ -205,11 +228,19 @@ async def test_text_message_blocked_by_guardrail_no_ai_response():
                 for c in item.get("content", [])
             ]
             real_ai_text = " ".join(ai_texts).strip()
-            # Allow guardrail-generated block messages (contain "Content blocked" or "blocked")
             if real_ai_text:
                 assert (
-                    "blocked" in real_ai_text.lower()
-                    or "guardrail" in real_ai_text.lower()
+                    BLOCKED_PHRASE not in real_ai_text
+                ), f"Blocked phrase leaked into AI response: {real_ai_text!r}"
+                normalized_ai_text = (
+                    real_ai_text.lower()
+                    .replace("\u2019", "'")
+                    .replace("\u2018", "'")
+                    .replace("\u201c", '"')
+                    .replace("\u201d", '"')
+                )
+                assert any(
+                    marker in normalized_ai_text for marker in safe_markers
                 ), f"AI responded with non-guardrail content even though message was blocked: {real_ai_text!r}"
 
     finally:

tests/llm_translation/test_deepseek_completion.py

@@ -176,3 +176,113 @@ def test_completion_cost_deepseek():
         pass
     except Exception as e:
         pytest.fail(f"Error occurred: {e}")
+
+
+def test_deepseek_fill_reasoning_content_multiturn():
+    """
+    Unit test for _fill_reasoning_content.
+    Reproduces issue #28045: DeepSeek thinking mode fails in multi-turn conversations
+    because reasoning_content is not passed back to the API.
+    """
+    from litellm.llms.deepseek.chat.transformation import DeepSeekChatConfig
+
+    config = DeepSeekChatConfig()
+
+    # Case 1: assistant message already has reasoning_content — should be left as-is
+    messages_with_rc = [
+        {"role": "user", "content": "Hello"},
+        {"role": "assistant", "content": "Hi", "reasoning_content": "I thought about it"},
+        {"role": "user", "content": "Follow up"},
+    ]
+    result = config._fill_reasoning_content(messages_with_rc)
+    assert result[1]["reasoning_content"] == "I thought about it"
+
+    # Case 2: assistant message has reasoning_content in provider_specific_fields — should be promoted
+    messages_with_psf = [
+        {"role": "user", "content": "Hello"},
+        {
+            "role": "assistant",
+            "content": "Hi",
+            "provider_specific_fields": {"reasoning_content": "stored thinking"},
+        },
+        {"role": "user", "content": "Follow up"},
+    ]
+    result = config._fill_reasoning_content(messages_with_psf)
+    assert result[1]["reasoning_content"] == "stored thinking"
+    # Should be removed from provider_specific_fields to avoid duplication
+    assert "reasoning_content" not in result[1].get("provider_specific_fields", {})
+
+    # Case 3: assistant message has no reasoning_content anywhere — should inject placeholder
+    messages_no_rc = [
+        {"role": "user", "content": "Hello"},
+        {"role": "assistant", "content": "Hi"},
+        {"role": "user", "content": "Follow up"},
+    ]
+    result = config._fill_reasoning_content(messages_no_rc)
+    assert result[1]["reasoning_content"] == " "
+
+    # Case 4: non-assistant messages should never be touched
+    messages_user_only = [
+        {"role": "user", "content": "Hello"},
+        {"role": "system", "content": "You are helpful"},
+    ]
+    result = config._fill_reasoning_content(messages_user_only)
+    assert "reasoning_content" not in result[0]
+    assert "reasoning_content" not in result[1]
+
+
+def test_deepseek_fill_reasoning_content_guard_in_transform_request():
+    """
+    _fill_reasoning_content must only run when BOTH conditions are true:
+      1. supports_reasoning() is True for the model
+      2. thinking mode is explicitly enabled in optional_params ({"type": "enabled"})
+
+    This prevents spurious injection on models like deepseek-v3.2 that support
+    thinking as opt-in but not always-on. Addresses oss-pr-review-agent feedback
+    on PR #28057.
+    """
+    from litellm.llms.deepseek.chat.transformation import DeepSeekChatConfig
+
+    config = DeepSeekChatConfig()
+
+    messages = [
+        {"role": "user", "content": "Hello"},
+        {"role": "assistant", "content": "Hi"},
+        {"role": "user", "content": "Follow up"},
+    ]
+
+    # Case 1: reasoning model + thinking enabled -> injection should happen
+    result = config.transform_request(
+        model="deepseek-reasoner",
+        messages=messages,
+        optional_params={"thinking": {"type": "enabled"}},
+        litellm_params={},
+        headers={},
+    )
+    assert result["messages"][1].get("reasoning_content") == " ", (
+        "reasoning_content should be injected when thinking is enabled"
+    )
+
+    # Case 2: reasoning model + thinking NOT in optional_params -> no injection
+    result = config.transform_request(
+        model="deepseek-reasoner",
+        messages=messages,
+        optional_params={},
+        litellm_params={},
+        headers={},
+    )
+    assert "reasoning_content" not in result["messages"][1], (
+        "reasoning_content should not be injected when thinking is not enabled"
+    )
+
+    # Case 3: non-reasoning model + thinking enabled -> no injection
+    result = config.transform_request(
+        model="deepseek-chat",
+        messages=messages,
+        optional_params={"thinking": {"type": "enabled"}},
+        litellm_params={},
+        headers={},
+    )
+    assert "reasoning_content" not in result["messages"][1], (
+        "reasoning_content should not be injected for non-reasoning models"
+    )

tests/llm_translation/test_nvidia_nim.py

@@ -262,3 +262,44 @@ class TestNvidiaNim(BaseLLMRerankTest):
     def get_expected_cost(self) -> float:
         """Nvidia NIM rerank models are free (cost = 0.0)"""
         return 0.0
+
+    @pytest.mark.asyncio()
+    @pytest.mark.parametrize("sync_mode", [True, False])
+    async def test_basic_rerank(self, sync_mode, monkeypatch):
+        """
+        Override the base live rerank test with a mocked HTTP layer.
+
+        NVIDIA reached end-of-life for the hosted
+        nvidia/llama-3.2-nv-rerankqa-1b-v2 rerank API on 2026-05-18 and
+        published no replacement model, so a live call now returns HTTP 410
+        ("Gone"). NVIDIA's hosted catalog rotates on a schedule, so pointing
+        at another live model would only defer the same failure. Mock the
+        transport instead (same pattern as
+        test_nvidia_nim_rerank_ranking_endpoint above) so the request/response
+        transformation and cost calculation stay covered offline.
+        """
+        monkeypatch.setenv("NVIDIA_NIM_API_KEY", "fake-api-key")
+
+        mock_response = MagicMock()
+        mock_response.status_code = 200
+        mock_response.headers = {}
+        mock_response.text = ""
+        mock_response.json.return_value = {
+            "rankings": [
+                {"index": 0, "logit": 0.95},
+                {"index": 1, "logit": 0.75},
+            ],
+            "usage": {"total_tokens": 7},
+        }
+
+        with (
+            patch(
+                "litellm.llms.custom_httpx.http_handler.HTTPHandler.post",
+                return_value=mock_response,
+            ),
+            patch(
+                "litellm.llms.custom_httpx.http_handler.AsyncHTTPHandler.post",
+                return_value=mock_response,
+            ),
+        ):
+            await super().test_basic_rerank(sync_mode=sync_mode)

tests/local_testing/conftest.py

@@ -22,14 +22,17 @@ sys.path.insert(
 )  # Adds the parent directory to the system path
 import litellm
 
-from tests._vcr_conftest_common import (  # noqa: E402
+from tests._vcr_conftest_common import (  # noqa: E402,F401
     VerboseReporterState,
+    _pin_multipart_boundary,
     apply_vcr_auto_marker_to_items,
     emit_cassette_cache_session_banner,
     emit_vcr_classification_summary,
+    emit_vcr_diagnostic_log,
     install_live_call_probe,
     record_vcr_outcome,
     register_persister_if_enabled,
+    reset_vcr_diag_dir,
     vcr_config_dict,
 )
 
@@ -84,6 +87,7 @@ def _vcr_outcome_gate(request, vcr):
 
 def pytest_configure(config):
     _verbose_state.remember_pluginmanager(config)
+    reset_vcr_diag_dir()
 
 
 def pytest_runtest_logreport(report):
@@ -93,6 +97,7 @@ def pytest_runtest_logreport(report):
 def pytest_terminal_summary(terminalreporter, exitstatus, config):
     emit_cassette_cache_session_banner(terminalreporter)
     emit_vcr_classification_summary(terminalreporter)
+    emit_vcr_diagnostic_log(terminalreporter)
 
 
 # ---------------------------------------------------------------------------

tests/local_testing/test_caching_handler.py

@@ -25,6 +25,7 @@ from unittest.mock import AsyncMock, patch, MagicMock
 from litellm.caching.caching_handler import (
     LLMCachingHandler,
     CachingHandlerResponse,
+    _is_chat_completion_cached_dict,
     _should_defer_streaming_cache_hit_callbacks,
 )
 from litellm.caching.caching import LiteLLMCacheType
@@ -40,6 +41,7 @@ from litellm.types.utils import (
 from litellm.types.llms.openai import ResponsesAPIResponse
 from datetime import timedelta, datetime
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLogging
+from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper
 from litellm._logging import verbose_logger
 import logging
 
@@ -1072,6 +1074,70 @@ def test_convert_cached_streaming_responses_result_to_iterator():
     )
 
 
+def test_is_chat_completion_cached_dict():
+    assert _is_chat_completion_cached_dict(
+        {"id": "chatcmpl-abc", "object": "chat.completion", "choices": []}
+    )
+    assert _is_chat_completion_cached_dict(
+        {"id": "other", "object": "chat.completion.chunk", "choices": []}
+    )
+    assert not _is_chat_completion_cached_dict(
+        {"id": "resp_abc", "object": "response", "output": []}
+    )
+
+
+def test_convert_cached_aresponses_bridge_chat_completion_stream():
+    """
+    openai/responses chat-completions bridge caches ModelResponse JSON on aresponses
+    cache keys; replay must not call ResponsesAPIResponse(**chatcmpl_dict).
+    """
+    caching_handler = LLMCachingHandler(
+        original_function=aresponses, request_kwargs={}, start_time=datetime.now()
+    )
+    logging_obj = LiteLLMLogging(
+        litellm_call_id=str(datetime.now()),
+        call_type=CallTypes.aresponses.value,
+        model="gpt-5.4",
+        messages=[],
+        function_id=str(uuid.uuid4()),
+        stream=True,
+        start_time=datetime.now(),
+    )
+    cached_result = {
+        "id": "chatcmpl-bridge-cache-test",
+        "object": "chat.completion",
+        "created": int(time.time()),
+        "model": "gpt-5.4",
+        "choices": [
+            {
+                "index": 0,
+                "message": {"role": "assistant", "content": "Hi!"},
+                "finish_reason": "stop",
+            }
+        ],
+        "usage": {
+            "prompt_tokens": 7,
+            "completion_tokens": 11,
+            "total_tokens": 18,
+        },
+    }
+
+    result = caching_handler._convert_cached_result_to_model_response(
+        cached_result=cached_result,
+        call_type=CallTypes.aresponses.value,
+        kwargs={
+            "model": "gpt-5.4",
+            "stream": True,
+            "messages": [{"role": "user", "content": "hi"}],
+        },
+        logging_obj=logging_obj,
+        model="gpt-5.4",
+        args=(),
+    )
+
+    assert isinstance(result, CustomStreamWrapper)
+
+
 def test_convert_cached_streaming_reasoning_result_to_iterator():
     caching_handler = LLMCachingHandler(
         original_function=responses, request_kwargs={}, start_time=datetime.now()

tests/logging_callback_tests/conftest.py

@@ -19,14 +19,17 @@ sys.path.insert(
 )  # Adds the parent directory to the system path
 import litellm
 
-from tests._vcr_conftest_common import (  # noqa: E402
+from tests._vcr_conftest_common import (  # noqa: E402,F401
     VerboseReporterState,
+    _pin_multipart_boundary,
     apply_vcr_auto_marker_to_items,
     emit_cassette_cache_session_banner,
     emit_vcr_classification_summary,
+    emit_vcr_diagnostic_log,
     install_live_call_probe,
     record_vcr_outcome,
     register_persister_if_enabled,
+    reset_vcr_diag_dir,
     vcr_config_dict,
 )
 
@@ -79,6 +82,7 @@ def _vcr_outcome_gate(request, vcr):
 
 def pytest_configure(config):
     _verbose_state.remember_pluginmanager(config)
+    reset_vcr_diag_dir()
 
 
 def pytest_runtest_logreport(report):
@@ -229,3 +233,4 @@ def pytest_collection_modifyitems(config, items):
 def pytest_terminal_summary(terminalreporter, exitstatus, config):
     emit_cassette_cache_session_banner(terminalreporter)
     emit_vcr_classification_summary(terminalreporter)
+    emit_vcr_diagnostic_log(terminalreporter)

tests/ocr_tests/conftest.py

@@ -12,14 +12,17 @@ import pytest
 
 sys.path.insert(0, os.path.abspath("../.."))
 
-from tests._vcr_conftest_common import (  # noqa: E402
+from tests._vcr_conftest_common import (  # noqa: E402,F401
     VerboseReporterState,
+    _pin_multipart_boundary,
     apply_vcr_auto_marker_to_items,
     emit_cassette_cache_session_banner,
     emit_vcr_classification_summary,
+    emit_vcr_diagnostic_log,
     install_live_call_probe,
     record_vcr_outcome,
     register_persister_if_enabled,
+    reset_vcr_diag_dir,
     vcr_config_dict,
 )
 
@@ -51,6 +54,7 @@ def _vcr_outcome_gate(request, vcr):
 
 def pytest_configure(config):
     _verbose_state.remember_pluginmanager(config)
+    reset_vcr_diag_dir()
 
 
 def pytest_runtest_logreport(report):
@@ -64,3 +68,4 @@ def pytest_collection_modifyitems(config, items):
 def pytest_terminal_summary(terminalreporter, exitstatus, config):
     emit_cassette_cache_session_banner(terminalreporter)
     emit_vcr_classification_summary(terminalreporter)
+    emit_vcr_diagnostic_log(terminalreporter)

tests/otel_tests/test_prometheus.py

@@ -610,22 +610,18 @@ def extract_user_budget_metrics(metrics_text: str, user_id: str) -> Dict[str, fl
     # Escape user_id for regex pattern matching
     escaped_user_id = re.escape(user_id)
 
-    # Get remaining budget
-    remaining_pattern = (
-        f'litellm_remaining_user_budget_metric{{user="{escaped_user_id}"}} ([0-9.]+)'
-    )
+    # Get remaining budget (user_email and user_alias may also be present as labels)
+    remaining_pattern = rf'litellm_remaining_user_budget_metric{{[^}}]*user="{escaped_user_id}"[^}}]*}} ([0-9.]+)'
     remaining_match = re.search(remaining_pattern, metrics_text)
     metrics["remaining"] = float(remaining_match.group(1)) if remaining_match else None
 
     # Get total budget
-    total_pattern = (
-        f'litellm_user_max_budget_metric{{user="{escaped_user_id}"}} ([0-9.]+)'
-    )
+    total_pattern = rf'litellm_user_max_budget_metric{{[^}}]*user="{escaped_user_id}"[^}}]*}} ([0-9.]+)'
     total_match = re.search(total_pattern, metrics_text)
     metrics["total"] = float(total_match.group(1)) if total_match else None
 
     # Get remaining hours
-    hours_pattern = f'litellm_user_budget_remaining_hours_metric{{user="{escaped_user_id}"}} ([0-9.]+)'
+    hours_pattern = rf'litellm_user_budget_remaining_hours_metric{{[^}}]*user="{escaped_user_id}"[^}}]*}} ([0-9.]+)'
     hours_match = re.search(hours_pattern, metrics_text)
     metrics["remaining_hours"] = float(hours_match.group(1)) if hours_match else None
 

tests/pass_through_unit_tests/conftest.py

@@ -5,14 +5,17 @@ import pytest
 
 sys.path.insert(0, os.path.abspath("../.."))
 
-from tests._vcr_conftest_common import (  # noqa: E402
+from tests._vcr_conftest_common import (  # noqa: E402,F401
     VerboseReporterState,
+    _pin_multipart_boundary,
     apply_vcr_auto_marker_to_items,
     emit_cassette_cache_session_banner,
     emit_vcr_classification_summary,
+    emit_vcr_diagnostic_log,
     install_live_call_probe,
     record_vcr_outcome,
     register_persister_if_enabled,
+    reset_vcr_diag_dir,
     vcr_config_dict,
 )
 
@@ -56,6 +59,7 @@ def _vcr_outcome_gate(request, vcr):
 
 def pytest_configure(config):
     _verbose_state.remember_pluginmanager(config)
+    reset_vcr_diag_dir()
 
 
 def pytest_runtest_logreport(report):
@@ -71,3 +75,4 @@ def pytest_collection_modifyitems(config, items):
 def pytest_terminal_summary(terminalreporter, exitstatus, config):
     emit_cassette_cache_session_banner(terminalreporter)
     emit_vcr_classification_summary(terminalreporter)
+    emit_vcr_diagnostic_log(terminalreporter)

tests/router_unit_tests/conftest.py

@@ -12,14 +12,17 @@ sys.path.insert(
 )  # Adds the parent directory to the system path
 import litellm  # noqa: E402,F401
 
-from tests._vcr_conftest_common import (  # noqa: E402
+from tests._vcr_conftest_common import (  # noqa: E402,F401
     VerboseReporterState,
+    _pin_multipart_boundary,
     apply_vcr_auto_marker_to_items,
     emit_cassette_cache_session_banner,
     emit_vcr_classification_summary,
+    emit_vcr_diagnostic_log,
     install_live_call_probe,
     record_vcr_outcome,
     register_persister_if_enabled,
+    reset_vcr_diag_dir,
     vcr_config_dict,
 )
 
@@ -97,6 +100,7 @@ def _vcr_outcome_gate(request, vcr):
 
 def pytest_configure(config):
     _verbose_state.remember_pluginmanager(config)
+    reset_vcr_diag_dir()
 
 
 def pytest_runtest_logreport(report):
@@ -123,3 +127,4 @@ def pytest_collection_modifyitems(config, items):
 def pytest_terminal_summary(terminalreporter, exitstatus, config):
     emit_cassette_cache_session_banner(terminalreporter)
     emit_vcr_classification_summary(terminalreporter)
+    emit_vcr_diagnostic_log(terminalreporter)

tests/search_tests/conftest.py

@@ -13,14 +13,17 @@ import pytest
 
 sys.path.insert(0, os.path.abspath("../.."))
 
-from tests._vcr_conftest_common import (  # noqa: E402
+from tests._vcr_conftest_common import (  # noqa: E402,F401
     VerboseReporterState,
+    _pin_multipart_boundary,
     apply_vcr_auto_marker_to_items,
     emit_cassette_cache_session_banner,
     emit_vcr_classification_summary,
+    emit_vcr_diagnostic_log,
     install_live_call_probe,
     record_vcr_outcome,
     register_persister_if_enabled,
+    reset_vcr_diag_dir,
     vcr_config_dict,
 )
 
@@ -52,6 +55,7 @@ def _vcr_outcome_gate(request, vcr):
 
 def pytest_configure(config):
     _verbose_state.remember_pluginmanager(config)
+    reset_vcr_diag_dir()
 
 
 def pytest_runtest_logreport(report):
@@ -65,3 +69,4 @@ def pytest_collection_modifyitems(config, items):
 def pytest_terminal_summary(terminalreporter, exitstatus, config):
     emit_cassette_cache_session_banner(terminalreporter)
     emit_vcr_classification_summary(terminalreporter)
+    emit_vcr_diagnostic_log(terminalreporter)

tests/test_callbacks_on_proxy.py

@@ -9,12 +9,155 @@ import pytest
 import asyncio
 import aiohttp
 import os
+import re
 import dotenv
+from collections import Counter
 from dotenv import load_dotenv
 import pytest
 
 load_dotenv()
 
+# A *leak* is sustained, monotonic growth of one callback TYPE across the whole
+# sampling window. A one-time bump that then plateaus is benign pollution from
+# other tests sharing this proxy (this suite runs `pytest -n 4` against a single
+# proxy container, so other workers legitimately add team/key-scoped callbacks
+# while this test sleeps). We therefore sample N times and only flag a type
+# whose normalized count never decreases, grows in >=2 distinct intervals, and
+# nets >= LEAK_MIN_NET_GROWTH overall.
+NUM_SAMPLES = 4
+SAMPLE_INTERVAL_SECONDS = 20
+LEAK_MIN_NET_GROWTH = 5
+LEAK_MIN_GROWING_INTERVALS = 2
+# A routing-strategy switch / alerting config is a *known, bounded, one-time*
+# registration (CCI diagnostic 2026-05-16: total 85->95 on the first interval
+# after switching to latency-based-routing, then flat at 95 for 2.5 min under
+# load). We absorb that step by settling before the baseline sample, so only
+# growth *after* the deliberate perturbation can count as a leak.
+SETTLE_SECONDS = 30
+
+# Strip instance-identity noise so N leaked instances of one class collapse to
+# one rising counter instead of N opaque, unrelated-looking strings.
+_ADDR_RE = re.compile(r" at 0x[0-9a-fA-F]+")
+_OBJ_RE = re.compile(r"<([\w.]+) object")
+
+
+def _normalize_callback(cb_str: str) -> str:
+    """Reduce a callback's str() to a stable type key (drops 0x… addresses)."""
+    s = _ADDR_RE.sub("", cb_str)
+    m = _OBJ_RE.search(s)
+    if m:
+        return m.group(1).split(".")[-1]
+    # bound methods: "<bound method Cls.m of <... at 0x..>>" -> "Cls.m"
+    bm = re.search(r"bound method ([\w.]+)", s)
+    if bm:
+        return bm.group(1)
+    return s.strip()
+
+
+def _summarize(all_litellm_callbacks) -> Counter:
+    return Counter(_normalize_callback(str(c)) for c in all_litellm_callbacks)
+
+
+def _detect_leaks(samples):
+    """
+    samples: list[Counter] taken in time order.
+
+    Returns {callback_type: [counts across samples]} for types that grew
+    monotonically (never decreased), in >=LEAK_MIN_GROWING_INTERVALS intervals,
+    and netted >=LEAK_MIN_NET_GROWTH overall — i.e. a real leak, not a one-shot
+    step from a parallel test.
+    """
+    leaks = {}
+    all_types = set().union(*[set(s) for s in samples]) if samples else set()
+    for t in all_types:
+        series = [s.get(t, 0) for s in samples]
+        deltas = [b - a for a, b in zip(series, series[1:])]
+        net = series[-1] - series[0]
+        non_decreasing = all(d >= 0 for d in deltas)
+        growing_intervals = sum(1 for d in deltas if d > 0)
+        if (
+            non_decreasing
+            and net >= LEAK_MIN_NET_GROWTH
+            and growing_intervals >= LEAK_MIN_GROWING_INTERVALS
+        ):
+            leaks[t] = series
+    return leaks
+
+
+def _terminal_suspects(samples):
+    """
+    Types whose net growth clears the threshold monotonically but is confined
+    to the *final* interval — `growing_intervals == 1` with that one growing
+    interval being the last. `_detect_leaks`' `>= 2` guard silently passes
+    these, so a real leak that accumulates entirely in the last sampled window
+    is indistinguishable from a one-time terminal step *without one more
+    sample*. Returns the set of such types so the caller can re-confirm.
+    """
+    suspects = set()
+    all_types = set().union(*[set(s) for s in samples]) if samples else set()
+    for t in all_types:
+        series = [s.get(t, 0) for s in samples]
+        deltas = [b - a for a, b in zip(series, series[1:])]
+        if not deltas:
+            continue
+        net = series[-1] - series[0]
+        non_decreasing = all(d >= 0 for d in deltas)
+        growing = [i for i, d in enumerate(deltas) if d > 0]
+        if (
+            non_decreasing
+            and net >= LEAK_MIN_NET_GROWTH
+            and growing == [len(deltas) - 1]
+        ):
+            suspects.add(t)
+    return suspects
+
+
+async def _detect_leaks_confirmed(session, samples):
+    """
+    `_detect_leaks`, plus a single confirmation sample when growth is confined
+    to the final interval (see `_terminal_suspects`). A genuine ongoing leak
+    keeps climbing -> now grows in >= 2 intervals -> flagged; a one-time
+    terminal registration plateaus -> still 1 growing interval -> ignored.
+    Returns `(leaks, samples)` (samples may have one extra entry appended).
+    """
+    leaks = _detect_leaks(samples)
+    if not leaks and _terminal_suspects(samples):
+        await asyncio.sleep(SAMPLE_INTERVAL_SECONDS)
+        _, _, all_cb = await get_active_callbacks(session=session)
+        samples = samples + [_summarize(all_cb)]
+        leaks = _detect_leaks(samples)
+    return leaks, samples
+
+
+def _format_report(samples, leaks) -> str:
+    lines = ["Callback count per type across samples (time order):"]
+    all_types = sorted(set().union(*[set(s) for s in samples]))
+    for t in all_types:
+        series = [s.get(t, 0) for s in samples]
+        marker = "  <-- LEAK" if t in leaks else ""
+        lines.append(f"  {t}: {series}{marker}")
+    totals = [sum(s.values()) for s in samples]
+    lines.append(f"TOTAL callbacks per sample: {totals}")
+    if leaks:
+        lines.append(
+            "Leaking callback types (sustained monotonic growth): "
+            + ", ".join(sorted(leaks))
+        )
+    return "\n".join(lines)
+
+
+async def _sample_callbacks(session, num_samples, interval):
+    """Take `num_samples` callback snapshots `interval`s apart."""
+    samples = []
+    alerts = []
+    for i in range(num_samples):
+        if i > 0:
+            await asyncio.sleep(interval)
+        num_cb, num_alert, all_cb = await get_active_callbacks(session=session)
+        samples.append(_summarize(all_cb))
+        alerts.append(num_alert)
+    return samples, alerts
+
 
 async def config_update(session, routing_strategy=None):
     url = "http://0.0.0.0:4000/config/update"
@@ -97,105 +240,65 @@ async def get_current_routing_strategy(session):
 
 @pytest.mark.asyncio
 @pytest.mark.order1
+@pytest.mark.flaky(reruns=2, reruns_delay=5)
 async def test_check_num_callbacks():
     """
-    Test 1:  num callbacks should NOT increase over time
-    -> check current callbacks
-    -> sleep for 30 seconds
-    -> check current callbacks
-    -> sleep for 30 seconds
-    -> check current callbacks
+    PROD invariant: no callback TYPE should grow without bound over time.
+
+    This suite runs `pytest -n 4` against one shared proxy, so the raw count is
+    noisy — other workers legitimately add team/key-scoped callbacks that then
+    plateau. We settle first, then sample several times, and only fail on
+    *sustained, monotonic* per-type growth (a genuine leak), naming the type.
     """
-    from litellm._uuid import uuid
-
     async with aiohttp.ClientSession() as session:
-        await asyncio.sleep(30)
-        num_callbacks_1, _, all_litellm_callbacks_1 = await get_active_callbacks(
-            session=session
-        )
-        assert num_callbacks_1 > 0
-        await asyncio.sleep(30)
+        # Absorb proxy warmup / in-flight parallel registration before baseline.
+        await asyncio.sleep(SETTLE_SECONDS)
 
-        num_callbacks_2, _, all_litellm_callbacks_2 = await get_active_callbacks(
-            session=session
+        samples, _ = await _sample_callbacks(
+            session, NUM_SAMPLES, SAMPLE_INTERVAL_SECONDS
         )
 
-        print("all_litellm_callbacks_1", all_litellm_callbacks_1)
+        assert sum(samples[0].values()) > 0, "expected some callbacks registered"
 
-        print(
-            "diff in callbacks=",
-            set(all_litellm_callbacks_1) - set(all_litellm_callbacks_2),
-        )
-
-        assert abs(num_callbacks_1 - num_callbacks_2) <= 4
-
-        await asyncio.sleep(30)
-
-        num_callbacks_3, _, all_litellm_callbacks_3 = await get_active_callbacks(
-            session=session
-        )
-
-        print(
-            "diff in callbacks = all_litellm_callbacks3 - all_litellm_callbacks2 ",
-            set(all_litellm_callbacks_3) - set(all_litellm_callbacks_2),
-        )
-
-        assert abs(num_callbacks_3 - num_callbacks_2) <= 4
+        leaks, samples = await _detect_leaks_confirmed(session, samples)
+        report = _format_report(samples, leaks)
+        print(report)
+        assert not leaks, f"Callback leak detected.\n{report}"
 
 
 @pytest.mark.asyncio
 @pytest.mark.order2
+@pytest.mark.flaky(reruns=2, reruns_delay=5)
 async def test_check_num_callbacks_on_lowest_latency():
     """
-    Test 1:  num callbacks should NOT increase over time
-    -> Update to lowest latency
-    -> check current callbacks
-    -> sleep for 30s
-    -> check current callbacks
-    -> sleep for 30s
-    -> check current callbacks
-    -> update back to original routing-strategy
+    Same PROD invariant as test_check_num_callbacks, but after switching the
+    router to latency-based-routing. That switch is a *known, bounded* one-time
+    registration (it adds the latency strategy handler + Slack alerting); we
+    settle past it before baselining so only post-switch growth counts as a
+    leak. Also asserts the alerting count is stable.
     """
-    from litellm._uuid import uuid
-
     async with aiohttp.ClientSession() as session:
         await asyncio.sleep(30)
 
         original_routing_strategy = await get_current_routing_strategy(session=session)
         await config_update(session=session, routing_strategy="latency-based-routing")
 
-        await asyncio.sleep(30)
+        try:
+            # Absorb the deliberate one-time config/update registration step.
+            await asyncio.sleep(SETTLE_SECONDS)
 
-        num_callbacks_1, num_alerts_1, all_litellm_callbacks_1 = (
-            await get_active_callbacks(session=session)
-        )
+            samples, alerts = await _sample_callbacks(
+                session, NUM_SAMPLES, SAMPLE_INTERVAL_SECONDS
+            )
 
-        await asyncio.sleep(30)
-
-        num_callbacks_2, num_alerts_2, all_litellm_callbacks_2 = (
-            await get_active_callbacks(session=session)
-        )
-
-        print(
-            "diff in callbacks all_litellm_callbacks_2 - all_litellm_callbacks_1 =",
-            set(all_litellm_callbacks_2) - set(all_litellm_callbacks_1),
-        )
-
-        assert abs(num_callbacks_1 - num_callbacks_2) <= 4
-
-        await asyncio.sleep(30)
-
-        num_callbacks_3, num_alerts_3, all_litellm_callbacks_3 = (
-            await get_active_callbacks(session=session)
-        )
-
-        print(
-            "diff in callbacks all_litellm_callbacks_3 - all_litellm_callbacks_2 =",
-            set(all_litellm_callbacks_3) - set(all_litellm_callbacks_2),
-        )
-
-        assert abs(num_callbacks_2 - num_callbacks_3) <= 4
-
-        assert num_alerts_1 == num_alerts_2 == num_alerts_3
-
-        await config_update(session=session, routing_strategy=original_routing_strategy)
+            leaks, samples = await _detect_leaks_confirmed(session, samples)
+            report = _format_report(samples, leaks)
+            print(report)
+            assert not leaks, f"Callback leak detected.\n{report}"
+            assert (
+                len(set(alerts)) == 1
+            ), f"alerting count changed across samples: {alerts}"
+        finally:
+            await config_update(
+                session=session, routing_strategy=original_routing_strategy
+            )

tests/test_litellm/caching/test_caching_handler.py

@@ -232,3 +232,207 @@ def test_combine_usage_handles_none_details():
     combined = llm_caching_handler.combine_usage(usage_a, usage_c)
     assert combined.prompt_tokens_details is not None
     assert combined.prompt_tokens_details.image_count == 1
+
+
+def test_is_chat_completion_cached_dict():
+    from litellm.caching.caching_handler import _is_chat_completion_cached_dict
+
+    assert _is_chat_completion_cached_dict(
+        {"id": "chatcmpl-abc", "object": "chat.completion", "choices": []}
+    )
+    assert _is_chat_completion_cached_dict(
+        {"id": "other", "object": "chat.completion.chunk", "choices": []}
+    )
+    assert _is_chat_completion_cached_dict(
+        {"id": "no-object", "choices": [{"index": 0}]}
+    )
+    assert not _is_chat_completion_cached_dict(
+        {"id": "resp_abc", "object": "response", "output": []}
+    )
+
+
+def _build_logging_obj(call_type: str, stream: bool):
+    import uuid as _uuid
+
+    from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLogging
+
+    return LiteLLMLogging(
+        litellm_call_id=str(datetime.now()),
+        call_type=call_type,
+        model="gpt-5.4",
+        messages=[],
+        function_id=str(_uuid.uuid4()),
+        stream=stream,
+        start_time=datetime.now(),
+    )
+
+
+def test_convert_cached_aresponses_bridge_chat_completion_stream():
+    """openai/responses chat-completions bridge: streaming cache hit replays as chat stream."""
+    from litellm import aresponses
+    from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper
+    from litellm.types.utils import CallTypes
+
+    caching_handler = LLMCachingHandler(
+        original_function=aresponses, request_kwargs={}, start_time=datetime.now()
+    )
+    cached_result = {
+        "id": "chatcmpl-bridge-cache-test",
+        "object": "chat.completion",
+        "created": int(time.time()),
+        "model": "gpt-5.4",
+        "choices": [
+            {
+                "index": 0,
+                "message": {"role": "assistant", "content": "Hi!"},
+                "finish_reason": "stop",
+            }
+        ],
+        "usage": {"prompt_tokens": 7, "completion_tokens": 11, "total_tokens": 18},
+    }
+
+    result = caching_handler._convert_cached_result_to_model_response(
+        cached_result=cached_result,
+        call_type=CallTypes.aresponses.value,
+        kwargs={
+            "model": "gpt-5.4",
+            "stream": True,
+            "messages": [{"role": "user", "content": "hi"}],
+        },
+        logging_obj=_build_logging_obj(CallTypes.aresponses.value, stream=True),
+        model="gpt-5.4",
+        args=(),
+    )
+
+    assert isinstance(result, CustomStreamWrapper)
+
+
+def test_convert_cached_responses_bridge_chat_completion_nonstream():
+    """openai/responses chat-completions bridge: non-streaming cache hit replays as ModelResponse."""
+    from litellm import responses
+    from litellm.types.utils import CallTypes, ModelResponse
+
+    caching_handler = LLMCachingHandler(
+        original_function=responses, request_kwargs={}, start_time=datetime.now()
+    )
+    cached_result = {
+        "id": "chatcmpl-bridge-nonstream",
+        "object": "chat.completion",
+        "created": int(time.time()),
+        "model": "gpt-5.4",
+        "choices": [
+            {
+                "index": 0,
+                "message": {"role": "assistant", "content": "Hi!"},
+                "finish_reason": "stop",
+            }
+        ],
+        "usage": {"prompt_tokens": 7, "completion_tokens": 11, "total_tokens": 18},
+    }
+
+    result = caching_handler._convert_cached_result_to_model_response(
+        cached_result=cached_result,
+        call_type=CallTypes.responses.value,
+        kwargs={
+            "model": "gpt-5.4",
+            "stream": False,
+            "messages": [{"role": "user", "content": "hi"}],
+        },
+        logging_obj=_build_logging_obj(CallTypes.responses.value, stream=False),
+        model="gpt-5.4",
+        args=(),
+    )
+
+    assert isinstance(result, ModelResponse)
+    assert result.choices[0].message.content == "Hi!"
+
+
+def test_convert_cached_responses_legacy_nonstream_path():
+    """Genuine ResponsesAPIResponse dict (no chatcmpl/choices) falls through legacy path."""
+    from litellm import responses
+    from litellm.types.llms.openai import ResponsesAPIResponse
+    from litellm.types.utils import CallTypes
+
+    caching_handler = LLMCachingHandler(
+        original_function=responses, request_kwargs={}, start_time=datetime.now()
+    )
+    cached_result = {
+        "id": "resp_legacy_nonstream",
+        "created_at": int(time.time()),
+        "status": "completed",
+        "model": "gpt-4o",
+        "object": "response",
+        "output": [
+            {
+                "type": "message",
+                "id": "msg_legacy",
+                "status": "completed",
+                "role": "assistant",
+                "content": [
+                    {
+                        "type": "output_text",
+                        "text": "legacy response",
+                        "annotations": [],
+                    }
+                ],
+            }
+        ],
+    }
+
+    result = caching_handler._convert_cached_result_to_model_response(
+        cached_result=cached_result,
+        call_type=CallTypes.responses.value,
+        kwargs={"model": "gpt-4o", "input": "hi", "stream": False},
+        logging_obj=_build_logging_obj(CallTypes.responses.value, stream=False),
+        model="gpt-4o",
+        args=(),
+    )
+
+    assert isinstance(result, ResponsesAPIResponse)
+    assert result.id == "resp_legacy_nonstream"
+
+
+def test_convert_cached_responses_legacy_stream_path():
+    """Genuine ResponsesAPIResponse dict (no chatcmpl/choices) on stream falls through legacy path."""
+    from litellm import responses
+    from litellm.responses.streaming_iterator import (
+        CachedResponsesAPIStreamingIterator,
+    )
+    from litellm.types.utils import CallTypes
+
+    caching_handler = LLMCachingHandler(
+        original_function=responses, request_kwargs={}, start_time=datetime.now()
+    )
+    cached_result = {
+        "id": "resp_legacy_stream",
+        "created_at": int(time.time()),
+        "status": "completed",
+        "model": "gpt-4o",
+        "object": "response",
+        "output": [
+            {
+                "type": "message",
+                "id": "msg_legacy_stream",
+                "status": "completed",
+                "role": "assistant",
+                "content": [
+                    {
+                        "type": "output_text",
+                        "text": "legacy stream",
+                        "annotations": [],
+                    }
+                ],
+            }
+        ],
+    }
+
+    result = caching_handler._convert_cached_result_to_model_response(
+        cached_result=cached_result,
+        call_type=CallTypes.responses.value,
+        kwargs={"model": "gpt-4o", "input": "hi", "stream": True},
+        logging_obj=_build_logging_obj(CallTypes.responses.value, stream=True),
+        model="gpt-4o",
+        args=(),
+    )
+
+    assert isinstance(result, CachedResponsesAPIStreamingIterator)

tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_handler.py

@@ -0,0 +1,150 @@
+import os
+import sys
+from datetime import datetime
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+sys.path.insert(0, os.path.abspath("../../.."))
+
+from litellm.completion_extras.litellm_responses_transformation.handler import (
+    ResponsesToCompletionBridgeHandler,
+)
+from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLogging
+from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper
+from litellm.types.utils import ModelResponse
+
+
+def test_is_preformatted_cached_chat_stream_true():
+    stream = MagicMock(spec=CustomStreamWrapper)
+    stream.custom_llm_provider = "cached_response"
+    assert (
+        ResponsesToCompletionBridgeHandler._is_preformatted_cached_chat_stream(stream)
+        is True
+    )
+
+
+def test_is_preformatted_cached_chat_stream_false_wrong_provider():
+    stream = MagicMock(spec=CustomStreamWrapper)
+    stream.custom_llm_provider = "openai"
+    assert (
+        ResponsesToCompletionBridgeHandler._is_preformatted_cached_chat_stream(stream)
+        is False
+    )
+
+
+def test_is_preformatted_cached_chat_stream_false_wrong_type():
+    assert (
+        ResponsesToCompletionBridgeHandler._is_preformatted_cached_chat_stream(
+            {"object": "chat.completion.chunk"}
+        )
+        is False
+    )
+
+
+def _bridge_kwargs(stream: bool):
+    logging_obj = LiteLLMLogging(
+        litellm_call_id="test-call",
+        call_type="completion",
+        model="gpt-5.4",
+        messages=[{"role": "user", "content": "hi"}],
+        function_id="fn-id",
+        stream=stream,
+        start_time=datetime.now(),
+    )
+    return {
+        "model": "gpt-5.4",
+        "custom_llm_provider": "openai",
+        "messages": [{"role": "user", "content": "hi"}],
+        "optional_params": {"stream": stream},
+        "litellm_params": {},
+        "headers": {},
+        "model_response": ModelResponse(),
+        "logging_obj": logging_obj,
+    }
+
+
+def test_completion_returns_cached_model_response_directly():
+    """Non-streaming bridge cache hit: responses() returns a ModelResponse -> bridge returns it as-is."""
+    cached = ModelResponse(id="chatcmpl-cached-nonstream", model="gpt-5.4")
+    bridge = ResponsesToCompletionBridgeHandler()
+
+    with (
+        patch.object(
+            bridge.transformation_handler,
+            "transform_request",
+            return_value={"model": "gpt-5.4", "input": "hi"},
+        ),
+        patch("litellm.responses", return_value=cached),
+    ):
+        result = bridge.completion(**_bridge_kwargs(stream=False))
+
+    assert result is cached
+
+
+@pytest.mark.asyncio
+async def test_acompletion_returns_cached_model_response_directly():
+    cached = ModelResponse(id="chatcmpl-cached-nonstream-async", model="gpt-5.4")
+    bridge = ResponsesToCompletionBridgeHandler()
+
+    with (
+        patch.object(
+            bridge.transformation_handler,
+            "transform_request",
+            return_value={"model": "gpt-5.4", "input": "hi"},
+        ),
+        patch("litellm.aresponses", new=AsyncMock(return_value=cached)),
+    ):
+        result = await bridge.acompletion(**_bridge_kwargs(stream=False))
+
+    assert result is cached
+
+
+def test_completion_skips_rewrapping_preformatted_cached_chat_stream():
+    """Streaming bridge cache hit returning CustomStreamWrapper(cached_response) -> bridge skips re-wrapping."""
+    stream = MagicMock(spec=CustomStreamWrapper)
+    stream.custom_llm_provider = "cached_response"
+    bridge = ResponsesToCompletionBridgeHandler()
+
+    with (
+        patch.object(
+            bridge.transformation_handler,
+            "transform_request",
+            return_value={"model": "gpt-5.4", "input": "hi"},
+        ),
+        patch("litellm.responses", return_value=stream),
+        patch.object(
+            bridge,
+            "_apply_post_stream_processing",
+            side_effect=lambda s, *a, **kw: s,
+        ) as post,
+    ):
+        result = bridge.completion(**_bridge_kwargs(stream=True))
+
+    post.assert_called_once()
+    assert result is stream
+
+
+@pytest.mark.asyncio
+async def test_acompletion_skips_rewrapping_preformatted_cached_chat_stream():
+    stream = MagicMock(spec=CustomStreamWrapper)
+    stream.custom_llm_provider = "cached_response"
+    bridge = ResponsesToCompletionBridgeHandler()
+
+    with (
+        patch.object(
+            bridge.transformation_handler,
+            "transform_request",
+            return_value={"model": "gpt-5.4", "input": "hi"},
+        ),
+        patch("litellm.aresponses", new=AsyncMock(return_value=stream)),
+        patch.object(
+            bridge,
+            "_apply_post_stream_processing",
+            side_effect=lambda s, *a, **kw: s,
+        ) as post,
+    ):
+        result = await bridge.acompletion(**_bridge_kwargs(stream=True))
+
+    post.assert_called_once()
+    assert result is stream

tests/test_litellm/completion_extras/test_litellm_responses_transformation_transformation.py

@@ -230,3 +230,30 @@ def test_transform_request_drops_user_metadata_with_additional_drop_params():
 
     assert "metadata" not in result
     assert result["litellm_metadata"]["internal_key"] == "secret"
+
+
+def test_translate_responses_chunk_passthrough_chat_completion_chunk():
+    from litellm.completion_extras.litellm_responses_transformation.transformation import (
+        OpenAiResponsesToChatCompletionStreamIterator,
+    )
+
+    chat_chunk = {
+        "id": "chatcmpl-cache-passthrough",
+        "object": "chat.completion.chunk",
+        "created": 1779104834,
+        "model": "gpt-5.4",
+        "choices": [
+            {
+                "index": 0,
+                "delta": {"role": "assistant", "content": "Hi! How can I help?"},
+                "finish_reason": None,
+            }
+        ],
+    }
+
+    result = OpenAiResponsesToChatCompletionStreamIterator.translate_responses_chunk_to_openai_stream(
+        chat_chunk
+    )
+
+    assert result.choices[0].delta.content == "Hi! How can I help?"
+    assert result.choices[0].finish_reason is None

tests/test_litellm/integrations/test_prometheus_user_team_metrics.py

@@ -460,6 +460,104 @@ async def test_assemble_user_object_does_not_override_metadata_max_budget(
     ), "max_budget from metadata must not be replaced by the DB value"
 
 
+async def test_assemble_user_object_populates_user_email_and_alias_from_db(
+    prometheus_logger,
+):
+    db_user = MagicMock()
+    db_user.max_budget = None
+    db_user.budget_reset_at = None
+    db_user.user_email = "alice@example.com"
+    db_user.user_alias = "Alice"
+
+    with patch("litellm.proxy.auth.auth_checks.get_user_object") as mock_get_user:
+        mock_get_user.return_value = db_user
+        user_object = await prometheus_logger._assemble_user_object(
+            user_id="user-abc-123",
+            spend=10.0,
+            max_budget=None,
+            response_cost=0.5,
+        )
+
+    assert user_object.user_email == "alice@example.com"
+    assert user_object.user_alias == "Alice"
+
+
+def test_set_user_budget_metrics_default_no_email_alias_labels(
+    prometheus_logger,
+):
+    """By default (flag off), only user label is emitted."""
+    import litellm
+    from litellm.proxy._types import LiteLLM_UserTable
+
+    litellm.prometheus_user_budget_label_include_email_alias = False
+
+    user = LiteLLM_UserTable(
+        user_id="user-abc-123",
+        user_email="alice@example.com",
+        user_alias="Alice",
+        spend=25.0,
+        max_budget=100.0,
+        budget_reset_at=datetime(2026, 3, 1, tzinfo=timezone.utc),
+    )
+
+    prometheus_logger.litellm_remaining_user_budget_metric = MagicMock()
+    prometheus_logger.litellm_user_max_budget_metric = MagicMock()
+    prometheus_logger.litellm_user_budget_remaining_hours_metric = MagicMock()
+
+    prometheus_logger._set_user_budget_metrics(user)
+
+    prometheus_logger.litellm_remaining_user_budget_metric.labels.assert_called_once_with(
+        user="user-abc-123",
+    )
+
+
+def test_set_user_budget_metrics_includes_user_email_and_alias_labels_when_opted_in(
+    prometheus_logger,
+):
+    """When prometheus_user_budget_label_include_email_alias=True, email+alias labels appear."""
+    import litellm
+    from litellm.proxy._types import LiteLLM_UserTable
+
+    litellm.prometheus_user_budget_label_include_email_alias = True
+
+    user = LiteLLM_UserTable(
+        user_id="user-abc-123",
+        user_email="alice@example.com",
+        user_alias="Alice",
+        spend=25.0,
+        max_budget=100.0,
+        budget_reset_at=datetime(2026, 3, 1, tzinfo=timezone.utc),
+    )
+
+    prometheus_logger.litellm_remaining_user_budget_metric = MagicMock()
+    prometheus_logger.litellm_user_max_budget_metric = MagicMock()
+    prometheus_logger.litellm_user_budget_remaining_hours_metric = MagicMock()
+
+    try:
+        prometheus_logger._set_user_budget_metrics(user)
+
+        prometheus_logger.litellm_remaining_user_budget_metric.labels.assert_called_once_with(
+            user="user-abc-123",
+            user_email="alice@example.com",
+            user_alias="Alice",
+        )
+        prometheus_logger.litellm_remaining_user_budget_metric.labels().set.assert_called_once_with(
+            75.0
+        )
+        prometheus_logger.litellm_user_max_budget_metric.labels.assert_called_once_with(
+            user="user-abc-123",
+            user_email="alice@example.com",
+            user_alias="Alice",
+        )
+        prometheus_logger.litellm_user_budget_remaining_hours_metric.labels.assert_called_once_with(
+            user="user-abc-123",
+            user_email="alice@example.com",
+            user_alias="Alice",
+        )
+    finally:
+        litellm.prometheus_user_budget_label_include_email_alias = False
+
+
 async def test_set_user_budget_metrics_after_api_request_no_inf_when_metadata_budget_none(
     prometheus_logger,
 ):

tests/test_litellm/interactions/test_gemini_interactions_transformation.py

@@ -9,15 +9,24 @@ Covers credential leak prevention changes:
 
 import os
 import sys
-from unittest.mock import patch
+from unittest.mock import MagicMock, patch
 
 import pytest
 
 sys.path.insert(0, os.path.abspath("../../.."))
 
+from litellm.interactions.litellm_responses_transformation.streaming_iterator import (
+    LiteLLMResponsesInteractionsStreamingIterator,
+)
 from litellm.llms.gemini.interactions.transformation import (
     GoogleAIStudioInteractionsConfig,
 )
+from litellm.types.llms.openai import (
+    ContentPartAddedEvent,
+    OutputTextDeltaEvent,
+    ResponseCompletedEvent,
+    ResponseCreatedEvent,
+)
 from litellm.types.router import GenericLiteLLMParams
 
 _PATCH_GET_API_KEY = "litellm.llms.gemini.common_utils.GeminiModelInfo.get_api_key"
@@ -113,6 +122,186 @@ class TestGetCompleteUrl:
                 )
 
 
+class TestStreamingIterator:
+    def _make_iterator(self) -> LiteLLMResponsesInteractionsStreamingIterator:
+        return LiteLLMResponsesInteractionsStreamingIterator(
+            model="gpt-5.4",
+            litellm_custom_stream_wrapper=MagicMock(),
+            request_input="hi",
+            optional_params={},
+        )
+
+    def _make_text_delta(
+        self, text: str, item_id: str = "item_1"
+    ) -> OutputTextDeltaEvent:
+        event = MagicMock(spec=OutputTextDeltaEvent)
+        event.delta = text
+        event.item_id = item_id
+        return event
+
+    def _make_part_added(self, item_id: str = "item_1") -> ContentPartAddedEvent:
+        event = MagicMock(spec=ContentPartAddedEvent)
+        event.item_id = item_id
+        return event
+
+    def _make_response_created(self) -> ResponseCreatedEvent:
+        event = MagicMock(spec=ResponseCreatedEvent)
+        event.response = MagicMock(id="resp_123")
+        return event
+
+    def test_content_delta_includes_type_field(self):
+        """content.delta events must carry delta.type='text' so the UI can display them."""
+        it = self._make_iterator()
+        it.sent_interaction_start = True
+        it.sent_content_start = True
+
+        chunk = it._transform_responses_chunk_to_interactions_chunk(
+            self._make_text_delta("Hello")
+        )
+
+        assert chunk is not None
+        assert chunk.event_type == "content.delta"
+        assert chunk.delta == {"type": "text", "text": "Hello"}
+
+    def test_response_part_added_emits_content_start(self):
+        """ContentPartAddedEvent (arrives before text deltas) should emit content.start
+        so the first OutputTextDeltaEvent immediately emits content.delta without dropping text.
+        """
+        it = self._make_iterator()
+        it.sent_interaction_start = True
+
+        chunk = it._transform_responses_chunk_to_interactions_chunk(
+            self._make_part_added()
+        )
+
+        assert chunk is not None
+        assert chunk.event_type == "content.start"
+        assert it.sent_content_start is True
+
+    def test_first_text_delta_not_dropped_when_part_added_seen(self):
+        """After ContentPartAddedEvent, the first text delta must yield content.delta
+        (not content.start), preserving the token text."""
+        it = self._make_iterator()
+        it.sent_interaction_start = True
+        it._transform_responses_chunk_to_interactions_chunk(self._make_part_added())
+
+        chunk = it._transform_responses_chunk_to_interactions_chunk(
+            self._make_text_delta("Hello")
+        )
+
+        assert chunk is not None
+        assert chunk.event_type == "content.delta"
+        assert chunk.delta is not None
+        assert chunk.delta.get("text") == "Hello"
+
+    def test_part_added_emits_interaction_start_fallback_when_not_sent(self):
+        """If ContentPartAddedEvent arrives before any ResponseCreatedEvent,
+        the iterator must emit interaction.start before content.start to honor
+        the documented event ordering contract."""
+        it = self._make_iterator()
+
+        chunk = it._transform_responses_chunk_to_interactions_chunk(
+            self._make_part_added(item_id="item_42")
+        )
+
+        assert chunk is not None
+        assert chunk.event_type == "interaction.start"
+        assert chunk.id == "item_42"
+        assert chunk.status == "in_progress"
+        assert chunk.model == "gpt-5.4"
+        assert it.sent_interaction_start is True
+        assert it.sent_content_start is False
+
+    def test_part_added_returns_none_when_already_started(self):
+        """A second ContentPartAddedEvent (after content.start was already emitted)
+        should be a no-op so we don't re-emit content.start."""
+        it = self._make_iterator()
+        it.sent_interaction_start = True
+        it.sent_content_start = True
+
+        chunk = it._transform_responses_chunk_to_interactions_chunk(
+            self._make_part_added()
+        )
+
+        assert chunk is None
+
+    def test_part_added_without_item_id_falls_back_to_self_id(self):
+        """When ContentPartAddedEvent has no item_id and we emit the interaction.start
+        fallback, the id must default to an interaction_<id(self)> string."""
+        it = self._make_iterator()
+        event = MagicMock(spec=ContentPartAddedEvent)
+        event.item_id = None
+
+        chunk = it._transform_responses_chunk_to_interactions_chunk(event)
+
+        assert chunk is not None
+        assert chunk.event_type == "interaction.start"
+        assert chunk.id == f"interaction_{id(it)}"
+
+    def test_first_text_delta_not_dropped_when_no_prior_start_events(self):
+        """When OutputTextDeltaEvent arrives before any ResponseCreatedEvent or
+        ContentPartAddedEvent, the iterator must emit interaction.start *and*
+        immediately follow with a content.start that carries this delta's text,
+        so the first token is never silently dropped from the stream."""
+        events = [
+            self._make_text_delta("Hello"),
+            self._make_text_delta(" World"),
+        ]
+        wrapper = MagicMock()
+        wrapper.__iter__ = lambda self: iter(events)
+        wrapper.__next__ = lambda self, _it=iter(events): next(_it)
+        it = LiteLLMResponsesInteractionsStreamingIterator(
+            model="gpt-5.4",
+            litellm_custom_stream_wrapper=wrapper,
+            request_input="hi",
+            optional_params={},
+        )
+
+        first = it._transform_responses_chunk_to_interactions_chunk(events[0])
+        assert first is not None
+        assert first.event_type == "interaction.start"
+        assert it.sent_interaction_start is True
+        assert it.sent_content_start is True
+        assert len(it._pending_events) == 1
+        pending = it._pending_events[0]
+        assert pending.event_type == "content.start"
+        assert pending.delta == {"type": "text", "text": "Hello"}
+
+        second = it._transform_responses_chunk_to_interactions_chunk(events[1])
+        assert second is not None
+        assert second.event_type == "content.delta"
+        assert second.delta == {"type": "text", "text": " World"}
+
+
+class TestTransformRequest:
+    def test_stream_param_included_in_request_body(self, config):
+        """When stream=True is in optional_params, the request body must include it
+        so the proxy forwards the SSE streaming flag to Google's backend."""
+        body = config.transform_request(
+            model="gemini-2.5-flash",
+            agent=None,
+            input="Hello",
+            optional_params={"stream": True},
+            litellm_params=GenericLiteLLMParams(api_key="test-key"),
+            headers={},
+        )
+
+        assert body.get("stream") is True
+        assert body.get("input") == "Hello"
+
+    def test_stream_false_not_included_when_absent(self, config):
+        body = config.transform_request(
+            model="gemini-2.5-flash",
+            agent=None,
+            input="Hello",
+            optional_params={},
+            litellm_params=GenericLiteLLMParams(api_key="test-key"),
+            headers={},
+        )
+
+        assert "stream" not in body
+
+
 class TestInteractionOperationUrls:
     """Test that get/delete/cancel interaction URLs exclude API key."""
 

tests/test_litellm/llms/bedrock/batches/test_batch_metadata_sanitization.py

@@ -0,0 +1,119 @@
+"""
+Test that BedrockBatchesConfig._get_openai_compatible_batch_metadata
+sanitizes non-string metadata values injected by proxy guardrail hooks.
+
+The OpenAI Batch Pydantic model requires metadata: Dict[str, str].
+Proxy hooks (Model Armor, OpenAI Moderations, queue time tracking) inject
+dicts, floats, and other non-string values that cause a ValidationError
+when constructing LiteLLMBatch. This test suite verifies the sanitization
+layer prevents that.
+"""
+
+import os
+import sys
+
+import pytest
+
+sys.path.insert(0, os.path.abspath("../../../../.."))
+
+from litellm.llms.bedrock.batches.transformation import BedrockBatchesConfig
+
+
+class TestGetOpenaiCompatibleBatchMetadata:
+    """Tests for _get_openai_compatible_batch_metadata."""
+
+    def test_string_values_pass_through_unchanged(self):
+        metadata = {"user_key": "user_value", "run_id": "abc123"}
+        result = BedrockBatchesConfig._get_openai_compatible_batch_metadata(metadata)
+        assert result == {"user_key": "user_value", "run_id": "abc123"}
+
+    def test_dict_values_serialized_to_json_string(self):
+        metadata = {
+            "_model_armor_response": {
+                "sanitizationResult": {"filterMatchState": "MATCH_FOUND"}
+            }
+        }
+        result = BedrockBatchesConfig._get_openai_compatible_batch_metadata(metadata)
+        assert "_model_armor_response" in result
+        assert isinstance(result["_model_armor_response"], str)
+        assert "MATCH_FOUND" in result["_model_armor_response"]
+
+    def test_float_values_serialized_to_string(self):
+        metadata = {"queue_time_seconds": 0.5}
+        result = BedrockBatchesConfig._get_openai_compatible_batch_metadata(metadata)
+        assert result == {"queue_time_seconds": "0.5"}
+
+    def test_none_values_excluded(self):
+        metadata = {"key": "value", "empty": None}
+        result = BedrockBatchesConfig._get_openai_compatible_batch_metadata(metadata)
+        assert "empty" not in result
+        assert result == {"key": "value"}
+
+    def test_standard_logging_guardrail_information_excluded(self):
+        metadata = {
+            "standard_logging_guardrail_information": {"some": "logging_data"},
+            "user_key": "keep_me",
+        }
+        result = BedrockBatchesConfig._get_openai_compatible_batch_metadata(metadata)
+        assert "standard_logging_guardrail_information" not in result
+        assert result == {"user_key": "keep_me"}
+
+    def test_non_dict_input_returns_empty_dict(self):
+        assert BedrockBatchesConfig._get_openai_compatible_batch_metadata(None) == {}
+        assert BedrockBatchesConfig._get_openai_compatible_batch_metadata("string") == {}
+        assert BedrockBatchesConfig._get_openai_compatible_batch_metadata(123) == {}
+
+    def test_empty_dict_returns_empty_dict(self):
+        assert BedrockBatchesConfig._get_openai_compatible_batch_metadata({}) == {}
+
+    def test_mixed_metadata_from_guardrails(self):
+        """Simulate real metadata contaminated by proxy guardrails."""
+        metadata = {
+            "_model_armor_response": {"sanitizationResult": {"key": "val"}},
+            "_model_armor_status": "success",
+            "_openai_moderation_response": {"id": "mod-123", "flagged": False},
+            "queue_time_seconds": 1.23,
+            "headers": {"Authorization": "Bearer sk-xxx"},
+            "standard_logging_guardrail_information": {"internal": True},
+            "user_metadata_key": "user_value",
+            "none_field": None,
+        }
+        result = BedrockBatchesConfig._get_openai_compatible_batch_metadata(metadata)
+
+        # All values must be strings
+        for key, value in result.items():
+            assert isinstance(value, str), f"metadata[{key!r}] is {type(value)}, not str"
+
+        # Excluded keys
+        assert "standard_logging_guardrail_information" not in result
+        assert "none_field" not in result
+
+        # Preserved keys
+        assert result["_model_armor_status"] == "success"
+        assert result["user_metadata_key"] == "user_value"
+
+    def test_result_compatible_with_litellm_batch(self):
+        """Verify sanitized metadata can construct a LiteLLMBatch without error."""
+        import time
+
+        from litellm.types.utils import LiteLLMBatch
+
+        metadata = {
+            "_model_armor_response": {"blocked": True},
+            "queue_time_seconds": 0.05,
+            "user_key": "value",
+        }
+        sanitized = BedrockBatchesConfig._get_openai_compatible_batch_metadata(metadata)
+
+        # This would raise ValidationError before the fix
+        batch = LiteLLMBatch(
+            id="arn:aws:bedrock:us-east-1:123:model-invocation-job/test",
+            object="batch",
+            endpoint="/v1/chat/completions",
+            input_file_id="file-123",
+            completion_window="24h",
+            status="validating",
+            created_at=int(time.time()),
+            metadata=sanitized,
+        )
+        assert batch.metadata == sanitized

tests/test_litellm/llms/bedrock/embed/test_bedrock_embedding.py

@@ -957,3 +957,50 @@ def test_titan_image_embedding_cost_uses_per_image_rate():
         assert response.usage is not None
         assert response.usage.prompt_tokens_details is not None
         assert response.usage.prompt_tokens_details.image_count == 1
+
+
+@pytest.mark.parametrize(
+    "encoding_format,expected_embedding_types",
+    [
+        ("float", ["float"]),
+        ("base64", ["base64"]),
+        (["float", "int8"], ["float", "int8"]),
+    ],
+)
+def test_bedrock_cohere_embedding_types_wrapped_as_list(
+    encoding_format, expected_embedding_types
+):
+    """
+    Bedrock Cohere expects `embedding_types` as a JSON array, not a raw string.
+
+    Regression test for: Bedrock returns
+        Malformed input request: #/embedding_types: expected type: JSONArray, found: String
+    when `encoding_format` is passed as a string.
+    """
+    litellm.set_verbose = True
+    client = HTTPHandler()
+    model = "bedrock/cohere.embed-multilingual-v3"
+
+    with patch.object(client, "post") as mock_post:
+        mock_response = Mock()
+        mock_response.status_code = 200
+        mock_response.text = json.dumps(cohere_embedding_response)
+        mock_response.json = lambda: json.loads(mock_response.text)
+        mock_post.return_value = mock_response
+
+        response = litellm.embedding(
+            model=model,
+            input=test_input,
+            encoding_format=encoding_format,
+            client=client,
+            aws_region_name="us-east-1",
+            aws_bedrock_runtime_endpoint="https://bedrock-runtime.us-east-1.amazonaws.com",
+            api_key="test-bearer-token-12345",
+        )
+
+        assert isinstance(response, litellm.EmbeddingResponse)
+
+        request_body = json.loads(mock_post.call_args.kwargs.get("data", "{}"))
+        assert "embedding_types" in request_body
+        assert request_body["embedding_types"] == expected_embedding_types
+        assert isinstance(request_body["embedding_types"], list)

tests/test_litellm/llms/bedrock/test_bedrock_common_utils.py

@@ -14,18 +14,46 @@ from litellm.llms.bedrock.common_utils import BedrockModelInfo
 
 
 # --------------------------------------------------------------------------- #
-# BEDROCK_RESPONSE_STREAM_SHAPE eager-load tests                              #
+# get_bedrock_response_stream_shape lazy-load tests                           #
 # --------------------------------------------------------------------------- #
 
 
-def test_bedrock_response_stream_shape_loaded_at_import():
+@pytest.fixture(autouse=True)
+def _reset_bedrock_response_stream_shape_cache():
+    """Prevent lru_cache leakage between tests in this module."""
+    import litellm.llms.bedrock.common_utils as mod
+
+    mod.get_bedrock_response_stream_shape.cache_clear()
+    yield
+    mod.get_bedrock_response_stream_shape.cache_clear()
+
+
+def test_bedrock_response_stream_shape_lazy_loads_once():
     """
-    BEDROCK_RESPONSE_STREAM_SHAPE is resolved at module import time.
+    get_bedrock_response_stream_shape() loads from botocore at most once per process.
+    """
+    from unittest.mock import MagicMock, patch
+
+    import litellm.llms.bedrock.common_utils as mod
+
+    sentinel = MagicMock()
+    with patch.object(
+        mod, "_load_bedrock_response_stream_shape", return_value=sentinel
+    ) as mock_load:
+        assert mod.get_bedrock_response_stream_shape() is sentinel
+        assert mod.get_bedrock_response_stream_shape() is sentinel
+        mock_load.assert_called_once()
+
+
+def test_bedrock_response_stream_shape_loaded_on_first_access():
+    """
+    get_bedrock_response_stream_shape() loads once on first use.
     In a standard environment with botocore installed it must be non-None.
     """
-    from litellm.llms.bedrock.common_utils import BEDROCK_RESPONSE_STREAM_SHAPE
+    pytest.importorskip("botocore")
+    from litellm.llms.bedrock.common_utils import get_bedrock_response_stream_shape
 
-    assert BEDROCK_RESPONSE_STREAM_SHAPE is not None
+    assert get_bedrock_response_stream_shape() is not None
 
 
 def test_bedrock_response_stream_shape_load_failure_returns_none():
@@ -38,6 +66,7 @@ def test_bedrock_response_stream_shape_load_failure_returns_none():
 
     import litellm.llms.bedrock.common_utils as mod
 
+    pytest.importorskip("botocore")
     with patch(
         "botocore.loaders.Loader.load_service_model",
         side_effect=Exception("no data"),
@@ -51,31 +80,29 @@ def test_bedrock_response_stream_shape_is_structure_shape():
     The loaded shape should be the botocore StructureShape for ResponseStream,
     not a plain dict or any other type.
     """
+    pytest.importorskip("botocore")
     from botocore.model import StructureShape
 
-    from litellm.llms.bedrock.common_utils import BEDROCK_RESPONSE_STREAM_SHAPE
+    from litellm.llms.bedrock.common_utils import get_bedrock_response_stream_shape
 
-    assert BEDROCK_RESPONSE_STREAM_SHAPE is not None, (
-        "BEDROCK_RESPONSE_STREAM_SHAPE is None — botocore may not be installed"
-    )
-    shape: StructureShape = BEDROCK_RESPONSE_STREAM_SHAPE  # remove Optional
+    loaded_shape = get_bedrock_response_stream_shape()
+    assert (
+        loaded_shape is not None
+    ), "get_bedrock_response_stream_shape() is None — botocore may not be installed"
+    shape: StructureShape = loaded_shape
     assert isinstance(shape, StructureShape)
     assert shape.name == "ResponseStream"
 
 
-def test_bedrock_response_stream_shape_same_object_across_imports():
+def test_bedrock_response_stream_shape_same_object_across_calls():
     """
-    Both bedrock modules that use the shape must reference the identical object —
-    confirming the constant is not re-loaded per import.
+    Repeated calls must return the identical cached object.
     """
-    from litellm.llms.bedrock.chat.invoke_handler import (
-        BEDROCK_RESPONSE_STREAM_SHAPE as invoke_shape,
-    )
-    from litellm.llms.bedrock.common_utils import (
-        BEDROCK_RESPONSE_STREAM_SHAPE as common_shape,
-    )
+    from litellm.llms.bedrock.common_utils import get_bedrock_response_stream_shape
 
-    assert common_shape is invoke_shape
+    first = get_bedrock_response_stream_shape()
+    second = get_bedrock_response_stream_shape()
+    assert first is second
 
 
 def test_bedrock_event_stream_decoder_base_uses_module_shape():
@@ -95,19 +122,23 @@ def test_bedrock_event_stream_decoder_base_uses_module_shape():
 
 def test_bedrock_parse_message_from_event_raises_on_none_shape():
     """
-    When BEDROCK_RESPONSE_STREAM_SHAPE is None (botocore unavailable),
+    When get_bedrock_response_stream_shape() returns None (botocore unavailable),
     _parse_message_from_event must raise BedrockError before touching the
     botocore parser — not an opaque AttributeError from inside botocore.
     """
     from unittest.mock import MagicMock, patch
 
     import litellm.llms.bedrock.common_utils as mod
-    from litellm.llms.bedrock.common_utils import BedrockError, BedrockEventStreamDecoderBase
+    from litellm.llms.bedrock.common_utils import (
+        BedrockError,
+        BedrockEventStreamDecoderBase,
+    )
 
-    decoder = BedrockEventStreamDecoderBase()
+    decoder = BedrockEventStreamDecoderBase.__new__(BedrockEventStreamDecoderBase)
+    decoder.parser = MagicMock()
     mock_event = MagicMock()
 
-    with patch.object(mod, "BEDROCK_RESPONSE_STREAM_SHAPE", None):
+    with patch.object(mod, "get_bedrock_response_stream_shape", return_value=None):
         with pytest.raises(BedrockError) as exc_info:
             decoder._parse_message_from_event(mock_event)
 

tests/test_litellm/llms/dashscope/test_dashscope_embedding_transformation.py

@@ -0,0 +1,141 @@
+"""
+Unit tests for DashScope embedding transformation.
+"""
+
+import json
+import os
+import sys
+from unittest.mock import MagicMock
+
+import httpx
+import pytest
+
+sys.path.insert(0, os.path.abspath("../../../../.."))
+
+from litellm.llms.dashscope.common_utils import DashScopeError
+from litellm.llms.dashscope.embed.transformation import (
+    DEFAULT_API_BASE,
+    DashScopeEmbeddingConfig,
+)
+from litellm.types.utils import EmbeddingResponse
+
+
+def test_validate_environment_and_url():
+    config = DashScopeEmbeddingConfig()
+    headers = config.validate_environment(
+        headers={},
+        model="text-embedding-v4",
+        messages=[],
+        optional_params={},
+        litellm_params={},
+        api_key="sk-test",
+    )
+    assert headers["Authorization"] == "Bearer sk-test"
+
+    url = config.get_complete_url(
+        api_base=None,
+        api_key="sk-test",
+        model="text-embedding-v4",
+        optional_params={},
+        litellm_params={},
+    )
+    assert url == f"{DEFAULT_API_BASE}/embeddings"
+
+
+def test_transform_embedding_request():
+    config = DashScopeEmbeddingConfig()
+    data = config.transform_embedding_request(
+        model="text-embedding-v4",
+        input=["风急天高猿啸哀"],
+        optional_params={"dimensions": 1024, "encoding_format": "float"},
+        headers={},
+    )
+    assert data == {
+        "model": "text-embedding-v4",
+        "input": ["风急天高猿啸哀"],
+        "dimensions": 1024,
+        "encoding_format": "float",
+    }
+
+
+def test_transform_embedding_response_success():
+    config = DashScopeEmbeddingConfig()
+    payload = {
+        "data": [
+            {"embedding": [0.1, 0.2], "index": 0, "object": "embedding"},
+        ],
+        "model": "text-embedding-v4",
+        "object": "list",
+        "usage": {"prompt_tokens": 5, "total_tokens": 5},
+        "id": "73591b79-xxxx",
+    }
+    raw = httpx.Response(
+        status_code=200,
+        content=json.dumps(payload).encode("utf-8"),
+        request=httpx.Request("POST", "https://example.com"),
+    )
+    result = config.transform_embedding_response(
+        model="text-embedding-v4",
+        raw_response=raw,
+        model_response=EmbeddingResponse(),
+        logging_obj=MagicMock(),
+        api_key="sk-x",
+        request_data={"input": ["a"]},
+        optional_params={},
+        litellm_params={},
+    )
+    assert result.model == "text-embedding-v4"
+    assert len(result.data) == 1
+    assert result.usage.prompt_tokens == 5
+
+
+def test_transform_embedding_request_user_param():
+    config = DashScopeEmbeddingConfig()
+    data = config.transform_embedding_request(
+        model="text-embedding-v4",
+        input=["hello"],
+        optional_params={"user": "user-123"},
+        headers={},
+    )
+    assert data["user"] == "user-123"
+
+
+def test_map_openai_params_drops_unsupported_with_drop_params():
+    config = DashScopeEmbeddingConfig()
+    result = config.map_openai_params(
+        non_default_params={"dimensions": 512, "unknown_param": "value"},
+        optional_params={},
+        model="text-embedding-v4",
+        drop_params=True,
+    )
+    assert result == {"dimensions": 512}
+    assert "unknown_param" not in result
+
+
+def test_transform_embedding_response_error():
+    config = DashScopeEmbeddingConfig()
+    payload = {
+        "error": {
+            "message": "Incorrect API key provided.",
+            "type": "invalid_request_error",
+            "code": "invalid_api_key",
+        }
+    }
+    raw = httpx.Response(
+        status_code=401,
+        content=json.dumps(payload).encode("utf-8"),
+        request=httpx.Request("POST", "https://example.com"),
+    )
+    with pytest.raises(DashScopeError) as exc:
+        config.transform_embedding_response(
+            model="text-embedding-v4",
+            raw_response=raw,
+            model_response=EmbeddingResponse(),
+            logging_obj=MagicMock(),
+            api_key="sk-bad",
+            request_data={"input": ["a"]},
+            optional_params={},
+            litellm_params={},
+        )
+    assert exc.value.status_code == 401
+    assert "Incorrect API key" in exc.value.message

tests/test_litellm/llms/dashscope/test_dashscope_rerank_transformation.py

@@ -0,0 +1,328 @@
+"""
+Unit tests for DashScope rerank transformation.
+"""
+
+import json
+import os
+import sys
+from unittest.mock import MagicMock
+
+import httpx
+import pytest
+
+sys.path.insert(0, os.path.abspath("../../../../.."))
+
+from litellm.llms.dashscope.common_utils import DashScopeError
+from litellm.llms.dashscope.rerank.transformation import (
+    DEFAULT_RERANK_URL,
+    DashScopeRerankConfig,
+)
+from litellm.types.rerank import RerankResponse
+
+
+class TestDashScopeRerankURL:
+    def setup_method(self):
+        self.config = DashScopeRerankConfig()
+
+    def test_default_url(self):
+        url = self.config.get_complete_url(api_base=None, model="qwen3-rerank")
+        assert url == DEFAULT_RERANK_URL
+
+    def test_explicit_v1_base_appends_reranks(self):
+        url = self.config.get_complete_url(
+            api_base="https://dashscope.aliyuncs.com/compatible-mode/v1",
+            model="qwen3-rerank",
+        )
+        assert url == "https://dashscope.aliyuncs.com/compatible-mode/v1/reranks"
+
+    def test_intl_v1_base_appends_reranks(self):
+        url = self.config.get_complete_url(
+            api_base="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
+            model="qwen3-rerank",
+        )
+        assert url == "https://dashscope-intl.aliyuncs.com/compatible-mode/v1/reranks"
+
+    def test_already_complete_url_passthrough(self):
+        full = "https://dashscope.aliyuncs.com/compatible-api/v1/reranks"
+        assert self.config.get_complete_url(api_base=full, model="qwen3-rerank") == full
+
+    def test_trailing_slash_stripped(self):
+        full = "https://dashscope.aliyuncs.com/compatible-api/v1/reranks/"
+        assert self.config.get_complete_url(
+            api_base=full, model="qwen3-rerank"
+        ) == full.rstrip("/")
+
+    def test_custom_v1_base_appends_reranks(self):
+        url = self.config.get_complete_url(
+            api_base="https://my-proxy.example.com/v1", model="qwen3-rerank"
+        )
+        assert url == "https://my-proxy.example.com/v1/reranks"
+
+
+class TestDashScopeRerankRequest:
+    def setup_method(self):
+        self.config = DashScopeRerankConfig()
+
+    def test_validate_environment_with_explicit_key(self):
+        headers = self.config.validate_environment(
+            headers={}, model="qwen3-rerank", api_key="sk-test"
+        )
+        assert headers["Authorization"] == "Bearer sk-test"
+        assert headers["content-type"] == "application/json"
+
+    def test_validate_environment_missing_key(self, monkeypatch):
+        monkeypatch.delenv("DASHSCOPE_API_KEY", raising=False)
+        with pytest.raises(ValueError, match="DASHSCOPE_API_KEY"):
+            self.config.validate_environment(
+                headers={}, model="qwen3-rerank", api_key=None
+            )
+
+    def test_validate_environment_falls_back_to_env(self, monkeypatch):
+        monkeypatch.setenv("DASHSCOPE_API_KEY", "env-key")
+        headers = self.config.validate_environment(
+            headers={}, model="qwen3-rerank", api_key=None
+        )
+        assert headers["Authorization"] == "Bearer env-key"
+
+    def test_supported_params(self):
+        assert self.config.get_supported_cohere_rerank_params("qwen3-rerank") == [
+            "query",
+            "documents",
+            "top_n",
+            "return_documents",
+        ]
+
+    def test_map_params_drops_unsupported(self):
+        # qwen3-rerank accepts query/documents/top_n/return_documents.
+        # rank_fields and max_*_per_doc are silently dropped.
+        params = self.config.map_cohere_rerank_params(
+            non_default_params={},
+            model="qwen3-rerank",
+            drop_params=False,
+            query="什么是文本排序模型",
+            documents=["d1", "d2"],
+            top_n=2,
+            rank_fields=["title"],
+            return_documents=True,
+            max_chunks_per_doc=5,
+            max_tokens_per_doc=100,
+        )
+        assert params == {
+            "query": "什么是文本排序模型",
+            "documents": ["d1", "d2"],
+            "top_n": 2,
+            "return_documents": True,
+        }
+
+    def test_transform_request_full(self):
+        body = self.config.transform_rerank_request(
+            model="qwen3-rerank",
+            optional_rerank_params={
+                "query": "如何制作美味的苹果派?",
+                "documents": ["a", "b"],
+                "top_n": 5,
+                "return_documents": True,
+            },
+            headers={},
+        )
+        assert body == {
+            "model": "qwen3-rerank",
+            "query": "如何制作美味的苹果派?",
+            "documents": ["a", "b"],
+            "top_n": 5,
+            "return_documents": True,
+        }
+
+    def test_transform_request_omits_unset_optional(self):
+        body = self.config.transform_rerank_request(
+            model="qwen3-rerank",
+            optional_rerank_params={"query": "q", "documents": ["a"]},
+            headers={},
+        )
+        assert "top_n" not in body
+        assert "return_documents" not in body
+
+    def test_transform_request_requires_query(self):
+        with pytest.raises(ValueError, match="query"):
+            self.config.transform_rerank_request(
+                model="qwen3-rerank",
+                optional_rerank_params={"documents": ["a"]},
+                headers={},
+            )
+
+    def test_transform_request_requires_documents(self):
+        with pytest.raises(ValueError, match="documents"):
+            self.config.transform_rerank_request(
+                model="qwen3-rerank",
+                optional_rerank_params={"query": "q"},
+                headers={},
+            )
+
+
+class TestDashScopeRerankResponse:
+    def setup_method(self):
+        self.config = DashScopeRerankConfig()
+        self.logging = MagicMock()
+
+    def _resp(self, body, status_code=200):
+        return httpx.Response(
+            status_code=status_code,
+            content=json.dumps(body).encode(),
+            request=httpx.Request("POST", "https://example.com"),
+        )
+
+    def test_success_response(self):
+        body = {
+            "object": "list",
+            "results": [
+                {"index": 0, "relevance_score": 0.93},
+                {"index": 2, "relevance_score": 0.34},
+            ],
+            "model": "qwen3-rerank",
+            "id": "85ba5752",
+            "usage": {"total_tokens": 79},
+        }
+        out = self.config.transform_rerank_response(
+            model="qwen3-rerank",
+            raw_response=self._resp(body),
+            model_response=RerankResponse(),
+            logging_obj=self.logging,
+            api_key="sk",
+            request_data={"query": "q"},
+        )
+        assert out.id == "85ba5752"
+        assert out.results == [
+            {"index": 0, "relevance_score": 0.93},
+            {"index": 2, "relevance_score": 0.34},
+        ]
+        assert out.meta == {
+            "billed_units": {"total_tokens": 79},
+            "tokens": {"input_tokens": 79},
+        }
+
+    def test_response_with_return_documents_real_payload(self):
+        # Verbatim sample from a real qwen3-rerank call with return_documents=true.
+        body = {
+            "object": "list",
+            "results": [
+                {
+                    "document": {
+                        "text": "苹果派的制作步骤包括准备面团、切苹果、调制馅料、组装和烘烤。"
+                    },
+                    "index": 1,
+                    "relevance_score": 0.8304247466067356,
+                },
+                {
+                    "document": {
+                        "text": "制作苹果派时，预先煮软苹果可以缩短烘烤时间。"
+                    },
+                    "index": 3,
+                    "relevance_score": 0.7142660211908354,
+                },
+            ],
+            "model": "qwen3-rerank",
+            "id": "e191b077-97c4-9929-b121-c2fbd2c7b0af",
+            "usage": {"total_tokens": 192},
+        }
+        out = self.config.transform_rerank_response(
+            model="qwen3-rerank",
+            raw_response=self._resp(body),
+            model_response=RerankResponse(),
+            logging_obj=self.logging,
+            request_data={"query": "如何制作美味的苹果派?"},
+        )
+        assert out.id == "e191b077-97c4-9929-b121-c2fbd2c7b0af"
+        assert out.results == [
+            {
+                "index": 1,
+                "relevance_score": 0.8304247466067356,
+                "document": {
+                    "text": "苹果派的制作步骤包括准备面团、切苹果、调制馅料、组装和烘烤。"
+                },
+            },
+            {
+                "index": 3,
+                "relevance_score": 0.7142660211908354,
+                "document": {"text": "制作苹果派时，预先煮软苹果可以缩短烘烤时间。"},
+            },
+        ]
+        assert out.meta == {
+            "billed_units": {"total_tokens": 192},
+            "tokens": {"input_tokens": 192},
+        }
+
+    def test_response_string_document_normalized(self):
+        # Defensive path: if a future API revision returns a bare string,
+        # normalize to {"text": ...} so downstream code stays consistent.
+        body = {
+            "results": [{"index": 0, "relevance_score": 0.9, "document": "hello"}],
+            "model": "qwen3-rerank",
+            "usage": {"total_tokens": 5},
+        }
+        out = self.config.transform_rerank_response(
+            model="qwen3-rerank",
+            raw_response=self._resp(body),
+            model_response=RerankResponse(),
+            logging_obj=self.logging,
+        )
+        assert out.results[0]["document"] == {"text": "hello"}
+
+    def test_missing_id_generates_uuid(self):
+        body = {"results": [{"index": 0, "relevance_score": 0.5}], "usage": {}}
+        out = self.config.transform_rerank_response(
+            model="qwen3-rerank",
+            raw_response=self._resp(body),
+            model_response=RerankResponse(),
+            logging_obj=self.logging,
+        )
+        assert out.id is not None and len(out.id) > 0
+
+    def test_error_envelope_raises(self):
+        body = {
+            "code": "InvalidApiKey",
+            "message": "Invalid API-key provided.",
+            "request_id": "fb53",
+        }
+        with pytest.raises(DashScopeError) as exc_info:
+            self.config.transform_rerank_response(
+                model="qwen3-rerank",
+                raw_response=self._resp(body, status_code=401),
+                model_response=RerankResponse(),
+                logging_obj=self.logging,
+            )
+        assert "Invalid API-key provided." in str(exc_info.value)
+
+    def test_non_json_response_raises(self):
+        bad = httpx.Response(
+            status_code=500,
+            content=b"<html>bad gateway</html>",
+            request=httpx.Request("POST", "https://example.com"),
+        )
+        with pytest.raises(DashScopeError):
+            self.config.transform_rerank_response(
+                model="qwen3-rerank",
+                raw_response=bad,
+                model_response=RerankResponse(),
+                logging_obj=self.logging,
+            )
+
+    def test_get_error_class(self):
+        err = self.config.get_error_class(
+            error_message="boom", status_code=500, headers={}
+        )
+        assert isinstance(err, DashScopeError)
+        assert err.status_code == 500
+
+
+class TestProviderConfigManagerDispatch:
+    def test_dashscope_returns_rerank_config(self):
+        import litellm
+        from litellm.utils import ProviderConfigManager
+
+        cfg = ProviderConfigManager.get_provider_rerank_config(
+            model="qwen3-rerank",
+            provider=litellm.LlmProviders.DASHSCOPE,
+            api_base=None,
+            present_version_params=[],
+        )
+        assert isinstance(cfg, DashScopeRerankConfig)

tests/test_litellm/llms/deepseek/messages/test_deepseek_anthropic_messages_transformation.py

@@ -0,0 +1,189 @@
+import litellm
+from litellm.llms.anthropic.experimental_pass_through.messages.transformation import (
+    AnthropicMessagesConfig,
+)
+from litellm.llms.deepseek.messages.transformation import (
+    DeepSeekAnthropicMessagesConfig,
+)
+from litellm.types.router import GenericLiteLLMParams
+from litellm.utils import ProviderConfigManager
+
+
+def test_deepseek_provider_uses_anthropic_messages_config():
+    config = ProviderConfigManager.get_provider_anthropic_messages_config(
+        model="deepseek-v4-pro",
+        provider=litellm.LlmProviders.DEEPSEEK,
+    )
+
+    assert isinstance(config, DeepSeekAnthropicMessagesConfig)
+    assert config.custom_llm_provider == "deepseek"
+
+
+def test_deepseek_anthropic_messages_config_defaults():
+    config = DeepSeekAnthropicMessagesConfig()
+
+    assert config.custom_llm_provider == "deepseek"
+    assert config.get_api_base() == "https://api.deepseek.com/anthropic"
+
+
+def test_anthropic_provider_keeps_default_config_for_deepseek_named_model():
+    config = ProviderConfigManager.get_provider_anthropic_messages_config(
+        model="deepseek-v4-pro",
+        provider=litellm.LlmProviders.ANTHROPIC,
+    )
+
+    assert isinstance(config, AnthropicMessagesConfig)
+    assert not isinstance(config, DeepSeekAnthropicMessagesConfig)
+
+
+def test_deepseek_anthropic_messages_url_defaults_to_anthropic_endpoint():
+    config = DeepSeekAnthropicMessagesConfig()
+
+    assert (
+        config.get_complete_url(
+            api_base=None,
+            api_key=None,
+            model="deepseek-v4-pro",
+            optional_params={},
+            litellm_params={},
+        )
+        == "https://api.deepseek.com/anthropic/v1/messages"
+    )
+    assert (
+        config.get_complete_url(
+            api_base="https://api.deepseek.com/anthropic/v1",
+            api_key=None,
+            model="deepseek-v4-pro",
+            optional_params={},
+            litellm_params={},
+        )
+        == "https://api.deepseek.com/anthropic/v1/messages"
+    )
+    assert (
+        config.get_complete_url(
+            api_base="https://api.deepseek.com/anthropic",
+            api_key=None,
+            model="deepseek-v4-pro",
+            optional_params={},
+            litellm_params={},
+        )
+        == "https://api.deepseek.com/anthropic/v1/messages"
+    )
+    assert (
+        config.get_complete_url(
+            api_base="https://api.deepseek.com",
+            api_key=None,
+            model="deepseek-v4-pro",
+            optional_params={},
+            litellm_params={},
+        )
+        == "https://api.deepseek.com/anthropic/v1/messages"
+    )
+    assert (
+        config.get_complete_url(
+            api_base="https://api.deepseek.com/v1",
+            api_key=None,
+            model="deepseek-v4-pro",
+            optional_params={},
+            litellm_params={},
+        )
+        == "https://api.deepseek.com/anthropic/v1/messages"
+    )
+    assert (
+        config.get_complete_url(
+            api_base="https://api.deepseek.com/v1/messages",
+            api_key=None,
+            model="deepseek-v4-pro",
+            optional_params={},
+            litellm_params={},
+        )
+        == "https://api.deepseek.com/anthropic/v1/messages"
+    )
+
+
+def test_deepseek_anthropic_messages_headers_use_deepseek_key():
+    config = DeepSeekAnthropicMessagesConfig()
+
+    headers, api_base = config.validate_anthropic_messages_environment(
+        headers={},
+        model="deepseek-v4-pro",
+        messages=[],
+        optional_params={},
+        litellm_params={},
+        api_key="sk-deepseek",
+        api_base="https://example.test/anthropic",
+    )
+
+    assert api_base == "https://example.test/anthropic"
+    assert headers["x-api-key"] == "sk-deepseek"
+    assert headers["anthropic-version"] == "2023-06-01"
+    assert headers["content-type"] == "application/json"
+
+
+def test_deepseek_anthropic_messages_preserves_thinking_and_sanitizes_custom_tools():
+    config = DeepSeekAnthropicMessagesConfig()
+    messages = [
+        {
+            "role": "user",
+            "content": "Use the tool.",
+        },
+        {
+            "role": "assistant",
+            "content": [
+                {
+                    "type": "thinking",
+                    "thinking": "I should call the tool.",
+                    "signature": "sig",
+                },
+                {
+                    "type": "tool_use",
+                    "id": "toolu_123",
+                    "name": "get_weather",
+                    "input": {"city": "Sao Paulo"},
+                },
+            ],
+        },
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "tool_result",
+                    "tool_use_id": "toolu_123",
+                    "content": "Sunny",
+                }
+            ],
+        },
+    ]
+
+    request = config.transform_anthropic_messages_request(
+        model="deepseek-v4-pro",
+        messages=messages,
+        anthropic_messages_optional_request_params={
+            "max_tokens": 100,
+            "thinking": {"type": "enabled", "budget_tokens": 1024},
+            "tools": [
+                {
+                    "type": "custom",
+                    "name": "get_weather",
+                    "description": "Get weather",
+                    "input_schema": {"type": "object"},
+                },
+                {
+                    "type": "web_search_20260209",
+                    "name": "web_search",
+                    "max_uses": 1,
+                },
+            ],
+        },
+        litellm_params=GenericLiteLLMParams(),
+        headers={},
+    )
+
+    assert request["messages"] == messages
+    assert request["thinking"] == {"type": "enabled", "budget_tokens": 1024}
+    assert request["tools"][0] == {
+        "name": "get_weather",
+        "description": "Get weather",
+        "input_schema": {"type": "object"},
+    }
+    assert request["tools"][1]["type"] == "web_search_20260209"

tests/test_litellm/llms/sagemaker/test_sagemaker_common_utils.py

@@ -12,18 +12,46 @@ from litellm.llms.sagemaker.completion.transformation import SagemakerConfig
 
 
 # --------------------------------------------------------------------------- #
-# SAGEMAKER_RESPONSE_STREAM_SHAPE eager-load tests                            #
+# get_sagemaker_response_stream_shape lazy-load tests                         #
 # --------------------------------------------------------------------------- #
 
 
-def test_sagemaker_response_stream_shape_loaded_at_import():
+@pytest.fixture(autouse=True)
+def _reset_sagemaker_response_stream_shape_cache():
+    """Prevent lru_cache leakage between tests in this module."""
+    import litellm.llms.sagemaker.common_utils as mod
+
+    mod.get_sagemaker_response_stream_shape.cache_clear()
+    yield
+    mod.get_sagemaker_response_stream_shape.cache_clear()
+
+
+def test_sagemaker_response_stream_shape_lazy_loads_once():
     """
-    SAGEMAKER_RESPONSE_STREAM_SHAPE is resolved at module import time.
+    get_sagemaker_response_stream_shape() loads from botocore at most once per process.
+    """
+    from unittest.mock import MagicMock, patch
+
+    import litellm.llms.sagemaker.common_utils as mod
+
+    sentinel = MagicMock()
+    with patch.object(
+        mod, "_load_sagemaker_response_stream_shape", return_value=sentinel
+    ) as mock_load:
+        assert mod.get_sagemaker_response_stream_shape() is sentinel
+        assert mod.get_sagemaker_response_stream_shape() is sentinel
+        mock_load.assert_called_once()
+
+
+def test_sagemaker_response_stream_shape_loaded_on_first_access():
+    """
+    get_sagemaker_response_stream_shape() loads once on first use.
     In a standard environment with botocore installed it must be non-None.
     """
-    from litellm.llms.sagemaker.common_utils import SAGEMAKER_RESPONSE_STREAM_SHAPE
+    pytest.importorskip("botocore")
+    from litellm.llms.sagemaker.common_utils import get_sagemaker_response_stream_shape
 
-    assert SAGEMAKER_RESPONSE_STREAM_SHAPE is not None
+    assert get_sagemaker_response_stream_shape() is not None
 
 
 def test_sagemaker_response_stream_shape_load_failure_returns_none():
@@ -36,6 +64,7 @@ def test_sagemaker_response_stream_shape_load_failure_returns_none():
 
     import litellm.llms.sagemaker.common_utils as mod
 
+    pytest.importorskip("botocore")
     with patch(
         "botocore.loaders.Loader.load_service_model",
         side_effect=Exception("no data"),
@@ -49,14 +78,16 @@ def test_sagemaker_response_stream_shape_is_structure_shape():
     The loaded shape should be the botocore StructureShape for
     InvokeEndpointWithResponseStreamOutput, not a plain dict or any other type.
     """
+    pytest.importorskip("botocore")
     from botocore.model import StructureShape
 
-    from litellm.llms.sagemaker.common_utils import SAGEMAKER_RESPONSE_STREAM_SHAPE
+    from litellm.llms.sagemaker.common_utils import get_sagemaker_response_stream_shape
 
-    assert SAGEMAKER_RESPONSE_STREAM_SHAPE is not None, (
-        "SAGEMAKER_RESPONSE_STREAM_SHAPE is None — botocore may not be installed"
-    )
-    shape: StructureShape = SAGEMAKER_RESPONSE_STREAM_SHAPE  # remove Optional
+    shape = get_sagemaker_response_stream_shape()
+    assert (
+        shape is not None
+    ), "get_sagemaker_response_stream_shape() is None — botocore may not be installed"
+    shape: StructureShape = shape  # remove Optional
     assert isinstance(shape, StructureShape)
     assert shape.name == "InvokeEndpointWithResponseStreamOutput"
 
@@ -64,29 +95,25 @@ def test_sagemaker_response_stream_shape_is_structure_shape():
 def test_sagemaker_response_stream_shape_not_reloaded_on_new_decoder():
     """
     Creating multiple AWSEventStreamDecoder instances must not trigger
-    additional botocore Loader calls — the shape is resolved once at import
-    time and reused.
+    additional botocore Loader calls — the shape is cached after first access.
     """
-    from litellm.llms.sagemaker.common_utils import SAGEMAKER_RESPONSE_STREAM_SHAPE
+    from litellm.llms.sagemaker.common_utils import get_sagemaker_response_stream_shape
 
-    decoder_a = AWSEventStreamDecoder(model="test-model-a")
-    decoder_b = AWSEventStreamDecoder(model="test-model-b")
+    decoder_a = AWSEventStreamDecoder.__new__(AWSEventStreamDecoder)
+    decoder_b = AWSEventStreamDecoder.__new__(AWSEventStreamDecoder)
 
-    # Both decoders should use the same pre-loaded shape object (identity check)
     assert "_response_stream_shape_cache" not in decoder_a.__dict__
     assert "_response_stream_shape_cache" not in decoder_b.__dict__
-    # The module constant is still the same object
-    from litellm.llms.sagemaker.common_utils import (
-        SAGEMAKER_RESPONSE_STREAM_SHAPE as shape_after,
-    )
 
-    assert SAGEMAKER_RESPONSE_STREAM_SHAPE is shape_after
+    first = get_sagemaker_response_stream_shape()
+    second = get_sagemaker_response_stream_shape()
+    assert first is second
 
 
 def test_sagemaker_parse_message_from_event_raises_on_none_shape():
     """
-    When SAGEMAKER_RESPONSE_STREAM_SHAPE is None (botocore unavailable),
-    _parse_message_from_event must raise ValueError before touching the
+    When get_sagemaker_response_stream_shape() returns None (botocore unavailable),
+    _parse_message_from_event must raise SagemakerError before touching the
     botocore parser — not an opaque AttributeError from inside botocore.
     """
     from unittest.mock import MagicMock, patch
@@ -94,10 +121,14 @@ def test_sagemaker_parse_message_from_event_raises_on_none_shape():
     import litellm.llms.sagemaker.common_utils as mod
     from litellm.llms.sagemaker.common_utils import SagemakerError
 
-    decoder = AWSEventStreamDecoder(model="test-model")
+    decoder = AWSEventStreamDecoder.__new__(AWSEventStreamDecoder)
+    decoder.model = "test-model"
+    decoder.parser = MagicMock()
+    decoder.content_blocks = []
+    decoder.is_messages_api = None
     mock_event = MagicMock()
 
-    with patch.object(mod, "SAGEMAKER_RESPONSE_STREAM_SHAPE", None):
+    with patch.object(mod, "get_sagemaker_response_stream_shape", return_value=None):
         with pytest.raises(SagemakerError) as exc_info:
             decoder._parse_message_from_event(mock_event)
 

tests/test_litellm/llms/test_file_content_block.py

@@ -0,0 +1,433 @@
+"""
+Tests for handling malformed or invalid 'file' content blocks (missing or null
+`file` sub-field, HTTP file_id URLs for Google AI Studio).
+
+Regression tests for:
+- litellm/llms/vertex_ai/gemini/transformation.py
+- litellm/llms/gemini/chat/transformation.py
+- litellm/litellm_core_utils/prompt_templates/common_utils.py
+  (migrate_file_to_image_url raises on missing `file`; file-id helpers skip non-OpenAI shapes)
+- litellm/litellm_core_utils/prompt_templates/factory.py (Bedrock + Anthropic)
+- litellm/llms/openai/chat/gpt_transformation.py
+"""
+
+import asyncio
+import copy
+from typing import List, cast
+
+import pytest
+
+import litellm
+from litellm.litellm_core_utils.prompt_templates.common_utils import (
+    get_file_ids_from_messages,
+    migrate_file_to_image_url,
+    update_messages_with_model_file_ids,
+)
+from litellm.litellm_core_utils.prompt_templates.factory import (
+    BedrockConverseMessagesProcessor,
+    anthropic_process_openai_file_message,
+)
+from litellm.llms.gemini.chat.transformation import GoogleAIStudioGeminiConfig
+from litellm.llms.openai.chat.gpt_transformation import OpenAIGPTConfig
+from litellm.llms.vertex_ai.gemini.transformation import (
+    _gemini_convert_messages_with_history,
+)
+from litellm.types.llms.openai import (
+    AllMessageValues,
+    ChatCompletionFileObject,
+    OpenAIMessageContentListBlock,
+)
+
+_MALFORMED_MESSAGES_RAW = [
+    {
+        "role": "user",
+        "content": [
+            {"type": "text", "text": "hello"},
+            {"type": "file"},  # Missing required "file" sub-field
+        ],
+    }
+]
+
+_WELL_FORMED_MESSAGES_RAW = [
+    {
+        "role": "user",
+        "content": [
+            {"type": "text", "text": "hello"},
+            {
+                "type": "file",
+                "file": {"file_id": "file-abc123", "format": "pdf"},
+            },
+        ],
+    }
+]
+
+MALFORMED_FILE_OBJECT: ChatCompletionFileObject = cast(
+    ChatCompletionFileObject, {"type": "file"}
+)
+
+EXPLICIT_NULL_FILE_OBJECT: ChatCompletionFileObject = cast(
+    ChatCompletionFileObject,
+    {"type": "file", "file": None},
+)
+
+
+def _malformed() -> List[AllMessageValues]:
+    return copy.deepcopy(cast(List[AllMessageValues], _MALFORMED_MESSAGES_RAW))
+
+
+def _well_formed() -> List[AllMessageValues]:
+    return copy.deepcopy(cast(List[AllMessageValues], _WELL_FORMED_MESSAGES_RAW))
+
+
+def _explicit_null_file_in_content() -> List[AllMessageValues]:
+    return copy.deepcopy(
+        cast(
+            List[AllMessageValues],
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "text", "text": "hello"},
+                        {"type": "file", "file": None},
+                    ],
+                }
+            ],
+        )
+    )
+
+
+# ---------------------------------------------------------------------------
+# vertex_ai/gemini/transformation.py
+# ---------------------------------------------------------------------------
+
+
+def test_gemini_convert_messages_malformed_file_raises_bad_request():
+    """_gemini_convert_messages_with_history should raise BadRequestError (not KeyError)
+    when a content block has type='file' but no 'file' sub-field."""
+    with pytest.raises(litellm.BadRequestError, match="missing the required 'file' field"):
+        _gemini_convert_messages_with_history(
+            messages=_malformed(),
+            model="gemini-2.0-flash",
+        )
+
+
+def test_gemini_convert_messages_explicit_null_file_field_raises_bad_request():
+    """Explicit JSON null for `file` must be rejected like a missing `file` key."""
+    with pytest.raises(litellm.BadRequestError, match="missing the required 'file' field"):
+        _gemini_convert_messages_with_history(
+            messages=_explicit_null_file_in_content(),
+            model="gemini-2.0-flash",
+        )
+
+
+# ---------------------------------------------------------------------------
+# gemini/chat/transformation.py - GoogleAIStudioGeminiConfig
+# ---------------------------------------------------------------------------
+
+
+def test_google_ai_studio_transform_messages_malformed_file_raises_bad_request():
+    """GoogleAIStudioGeminiConfig._transform_messages should raise BadRequestError
+    when a content block has type='file' but no 'file' sub-field."""
+    config = GoogleAIStudioGeminiConfig()
+    with pytest.raises(litellm.BadRequestError, match="missing the required 'file' field"):
+        config._transform_messages(messages=_malformed(), model="gemini-2.0-flash")
+
+
+def test_google_ai_studio_transform_messages_explicit_null_file_field_raises_bad_request():
+    """Explicit JSON null for `file` must be rejected like a missing `file` key."""
+    config = GoogleAIStudioGeminiConfig()
+    with pytest.raises(litellm.BadRequestError, match="missing the required 'file' field"):
+        config._transform_messages(
+            messages=_explicit_null_file_in_content(), model="gemini-2.0-flash"
+        )
+
+
+def test_google_ai_studio_transform_messages_http_file_id_converts_to_base64(monkeypatch):
+    """Google AI Studio rejects raw HTTP(S) file URLs; _transform_messages should
+    fetch and replace them with base64 `file_data` before conversion."""
+    # Data URL shape so downstream Gemini media parsing accepts the inlined bytes
+    # (mirrors real `convert_url_to_base64` output from `_process_image_response`).
+    fake_file_data = "data:application/pdf;base64,aGVsbG8="
+
+    def _fake_convert_url_to_base64(url: str) -> str:
+        assert url == "https://example.com/doc.pdf"
+        return fake_file_data
+
+    monkeypatch.setattr(
+        "litellm.llms.gemini.chat.transformation.convert_url_to_base64",
+        _fake_convert_url_to_base64,
+    )
+    messages = cast(
+        List[AllMessageValues],
+        [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "hello"},
+                    {
+                        "type": "file",
+                        "file": {
+                            "file_id": "https://example.com/doc.pdf",
+                            "format": "pdf",
+                        },
+                    },
+                ],
+            }
+        ],
+    )
+    config = GoogleAIStudioGeminiConfig()
+    config._transform_messages(messages=messages, model="gemini-2.0-flash")
+    content = messages[0].get("content")
+    assert isinstance(content, list)
+    file_block = next(c for c in content if isinstance(c, dict) and c.get("type") == "file")
+    file_field = file_block.get("file")
+    assert isinstance(file_field, dict)
+    assert file_field.get("file_data") == fake_file_data
+    assert "file_id" not in file_field
+
+
+def test_google_ai_studio_transform_messages_http_file_id_convert_failure_leaves_file_unchanged(
+    monkeypatch,
+):
+    """If convert_url_to_base64 fails, the Studio prep step must not mutate the block
+    (see try/except in GoogleAIStudioGeminiConfig._transform_messages)."""
+    https_id = "https://example.com/missing.pdf"
+
+    def _raise(_url: str) -> str:
+        raise litellm.ImageFetchError("simulated fetch failure")
+
+    monkeypatch.setattr(
+        "litellm.llms.gemini.chat.transformation.convert_url_to_base64",
+        _raise,
+    )
+    messages = cast(
+        List[AllMessageValues],
+        [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "hello"},
+                    {
+                        "type": "file",
+                        "file": {
+                            "file_id": https_id,
+                            "format": "application/pdf",
+                        },
+                    },
+                ],
+            }
+        ],
+    )
+    config = GoogleAIStudioGeminiConfig()
+    config._transform_messages(messages=messages, model="gemini-2.0-flash")
+    content = messages[0].get("content")
+    assert isinstance(content, list)
+    file_block = next(c for c in content if isinstance(c, dict) and c.get("type") == "file")
+    file_field = file_block.get("file")
+    assert isinstance(file_field, dict)
+    assert file_field.get("file_id") == https_id
+    assert file_field.get("format") == "application/pdf"
+    assert "file_data" not in file_field
+
+
+# ---------------------------------------------------------------------------
+# common_utils.py - update_messages_with_model_file_ids
+# ---------------------------------------------------------------------------
+
+
+def test_update_messages_with_model_file_ids_malformed_skips_non_openai_file_block():
+    """Non-OpenAI file blocks (e.g. missing nested `file` dict) are skipped so callers
+    relying on LangChain v1 / provider-native shapes are not rejected here."""
+    messages = _malformed()
+    result = update_messages_with_model_file_ids(
+        messages=messages,
+        model_id="some-model",
+        model_file_id_mapping={},
+    )
+    assert result == messages
+    content = result[0].get("content")
+    assert isinstance(content, list)
+    file_block = next(c for c in content if isinstance(c, dict) and c.get("type") == "file")
+    assert "file" not in file_block
+
+
+def test_update_messages_with_model_file_ids_well_formed_updates():
+    """update_messages_with_model_file_ids should update file_id for well-formed blocks."""
+    mapping = {"file-abc123": {"some-model": "provider-file-xyz"}}
+    result = update_messages_with_model_file_ids(
+        messages=_well_formed(),
+        model_id="some-model",
+        model_file_id_mapping=mapping,
+    )
+    content = result[0].get("content")
+    assert isinstance(content, list)
+    file_block = next(c for c in content if c.get("type") == "file")
+    assert file_block.get("file", {}).get("file_id") == "provider-file-xyz"
+
+
+# ---------------------------------------------------------------------------
+# common_utils.py - get_file_ids_from_messages
+# ---------------------------------------------------------------------------
+
+
+def test_get_file_ids_from_messages_malformed_skips_non_openai_file_block():
+    """Blocks with type='file' but no OpenAI `file` sub-dict yield no extracted ids."""
+    assert get_file_ids_from_messages(messages=_malformed()) == []
+
+
+def test_get_file_ids_from_messages_well_formed_returns_ids():
+    """get_file_ids_from_messages should extract file_id from well-formed blocks."""
+    messages: List[AllMessageValues] = cast(
+        List[AllMessageValues],
+        [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "hello"},
+                    {"type": "file", "file": {"file_id": "file-abc123", "format": "pdf"}},
+                ],
+            }
+        ],
+    )
+    result = get_file_ids_from_messages(messages=messages)
+    assert result == ["file-abc123"]
+
+
+# ---------------------------------------------------------------------------
+# factory.py - BedrockConverseMessagesProcessor (sync + async)
+# ---------------------------------------------------------------------------
+
+
+def test_bedrock_process_file_message_malformed_raises_bad_request():
+    """_process_file_message should raise BadRequestError (not KeyError)
+    when the file object is missing the 'file' sub-field."""
+    with pytest.raises(litellm.BadRequestError, match="missing the required 'file' field"):
+        BedrockConverseMessagesProcessor._process_file_message(MALFORMED_FILE_OBJECT)
+
+
+def test_bedrock_process_file_message_explicit_null_file_field_raises_bad_request():
+    with pytest.raises(litellm.BadRequestError, match="missing the required 'file' field"):
+        BedrockConverseMessagesProcessor._process_file_message(EXPLICIT_NULL_FILE_OBJECT)
+
+
+def test_bedrock_async_process_file_message_malformed_raises_bad_request():
+    """_async_process_file_message should raise BadRequestError (not KeyError)
+    when the file object is missing the 'file' sub-field."""
+
+    async def _run() -> None:
+        with pytest.raises(
+            litellm.BadRequestError, match="missing the required 'file' field"
+        ):
+            await BedrockConverseMessagesProcessor._async_process_file_message(
+                MALFORMED_FILE_OBJECT
+            )
+
+    asyncio.run(_run())
+
+
+def test_bedrock_async_process_file_message_explicit_null_file_field_raises_bad_request():
+    async def _run() -> None:
+        with pytest.raises(
+            litellm.BadRequestError, match="missing the required 'file' field"
+        ):
+            await BedrockConverseMessagesProcessor._async_process_file_message(
+                EXPLICIT_NULL_FILE_OBJECT
+            )
+
+    asyncio.run(_run())
+
+
+# ---------------------------------------------------------------------------
+# openai/chat/gpt_transformation.py
+# ---------------------------------------------------------------------------
+
+
+def test_openai_apply_common_transform_malformed_file_raises_bad_request():
+    """_apply_common_transform_content_item should raise BadRequestError (not KeyError)
+    when a content block has type='file' but no 'file' sub-field."""
+    config = OpenAIGPTConfig()
+    malformed_block: OpenAIMessageContentListBlock = cast(
+        OpenAIMessageContentListBlock, {"type": "file"}
+    )
+    with pytest.raises(litellm.BadRequestError, match="missing the required 'file' field"):
+        config._apply_common_transform_content_item(malformed_block)
+
+
+def test_openai_apply_common_transform_explicit_null_file_field_raises_bad_request():
+    config = OpenAIGPTConfig()
+    explicit_null_block: OpenAIMessageContentListBlock = cast(
+        OpenAIMessageContentListBlock,
+        {"type": "file", "file": None},
+    )
+    with pytest.raises(litellm.BadRequestError, match="missing the required 'file' field"):
+        config._apply_common_transform_content_item(explicit_null_block)
+
+
+def test_openai_apply_common_transform_well_formed_file_does_not_raise():
+    """_apply_common_transform_content_item should not raise for well-formed file blocks."""
+    config = OpenAIGPTConfig()
+    well_formed_block: OpenAIMessageContentListBlock = cast(
+        OpenAIMessageContentListBlock,
+        {"type": "file", "file": {"file_id": "file-abc123"}},
+    )
+    result = config._apply_common_transform_content_item(well_formed_block)
+    assert result.get("type") == "file"
+    file_field = cast(ChatCompletionFileObject, result).get("file", {})
+    assert file_field.get("file_id") == "file-abc123"
+
+
+# ---------------------------------------------------------------------------
+# factory.py - anthropic_process_openai_file_message
+# ---------------------------------------------------------------------------
+
+
+def test_anthropic_process_openai_file_message_malformed_raises_bad_request():
+    """anthropic_process_openai_file_message should raise BadRequestError (not KeyError)
+    when the file object is missing the 'file' sub-field."""
+    with pytest.raises(litellm.BadRequestError, match="missing the required 'file' field"):
+        anthropic_process_openai_file_message(MALFORMED_FILE_OBJECT)
+
+
+def test_anthropic_process_openai_file_message_explicit_null_file_field_raises_bad_request():
+    with pytest.raises(litellm.BadRequestError, match="missing the required 'file' field"):
+        anthropic_process_openai_file_message(EXPLICIT_NULL_FILE_OBJECT)
+
+
+def test_anthropic_process_openai_file_message_well_formed_file_id_does_not_raise():
+    """anthropic_process_openai_file_message should not raise for a well-formed file_id block."""
+    well_formed: ChatCompletionFileObject = cast(
+        ChatCompletionFileObject,
+        {"type": "file", "file": {"file_id": "file-abc123"}},
+    )
+    result = anthropic_process_openai_file_message(well_formed)
+    assert result.get("type") in ("document", "image", "container_upload")
+
+
+# ---------------------------------------------------------------------------
+# common_utils.py - migrate_file_to_image_url
+# ---------------------------------------------------------------------------
+
+
+def test_migrate_file_to_image_url_malformed_raises_bad_request():
+    """migrate_file_to_image_url should raise BadRequestError (not KeyError)
+    when the file object is missing the 'file' sub-field."""
+    with pytest.raises(litellm.BadRequestError, match="missing the required 'file' field"):
+        migrate_file_to_image_url(MALFORMED_FILE_OBJECT)
+
+
+def test_migrate_file_to_image_url_explicit_null_file_field_raises_bad_request():
+    with pytest.raises(litellm.BadRequestError, match="missing the required 'file' field"):
+        migrate_file_to_image_url(EXPLICIT_NULL_FILE_OBJECT)
+
+
+def test_migrate_file_to_image_url_well_formed_returns_image_url():
+    """migrate_file_to_image_url should return an image_url block for a well-formed file."""
+    well_formed: ChatCompletionFileObject = cast(
+        ChatCompletionFileObject,
+        {"type": "file", "file": {"file_id": "file-abc123", "format": "png"}},
+    )
+    result = migrate_file_to_image_url(well_formed)
+    assert result.get("type") == "image_url"
+    image_url = result.get("image_url", {})
+    assert isinstance(image_url, dict)
+    assert image_url.get("url") == "file-abc123"

tests/test_litellm/llms/vertex_ai/gemini/test_gemini_image_url_missing_field.py

@@ -0,0 +1,52 @@
+import pytest
+from typing import List, cast
+
+import litellm
+from litellm.llms.vertex_ai.gemini.transformation import (
+    _gemini_convert_messages_with_history,
+)
+from litellm.types.llms.openai import AllMessageValues
+
+
+def test_missing_image_url_field_raises_bad_request_error():
+    """When element type is 'image_url' but 'image_url' field is missing, a BadRequestError is raised."""
+    messages = cast(
+        List[AllMessageValues],
+        [{"role": "user", "content": [{"type": "image_url"}]}],
+    )
+    with pytest.raises(litellm.BadRequestError) as exc_info:
+        _gemini_convert_messages_with_history(messages, model="gemini-1.5-pro")
+    assert "'image_url' field is missing" in str(exc_info.value)
+
+
+def test_missing_url_inside_image_url_dict_raises_bad_request_error():
+    """When image_url is a dict but 'url' key is absent, a BadRequestError is raised."""
+    messages = cast(
+        List[AllMessageValues],
+        [{"role": "user", "content": [{"type": "image_url", "image_url": {"detail": "high"}}]}],
+    )
+    with pytest.raises(litellm.BadRequestError) as exc_info:
+        _gemini_convert_messages_with_history(messages, model="gemini-1.5-pro")
+    assert "'url' field is missing inside" in str(exc_info.value)
+
+
+def test_explicit_null_image_url_raises_bad_request_error():
+    """When image_url key is present but explicitly null, a BadRequestError is raised."""
+    messages = cast(
+        List[AllMessageValues],
+        [{"role": "user", "content": [{"type": "image_url", "image_url": None}]}],
+    )
+    with pytest.raises(litellm.BadRequestError) as exc_info:
+        _gemini_convert_messages_with_history(messages, model="gemini-1.5-pro")
+    assert "'image_url' field is missing" in str(exc_info.value)
+
+
+def test_empty_dict_image_url_raises_bad_request_error():
+    """When image_url is an empty dict (no url), a BadRequestError is raised."""
+    messages = cast(
+        List[AllMessageValues],
+        [{"role": "user", "content": [{"type": "image_url", "image_url": {}}]}],
+    )
+    with pytest.raises(litellm.BadRequestError) as exc_info:
+        _gemini_convert_messages_with_history(messages, model="gemini-1.5-pro")
+    assert "'url' field is missing inside" in str(exc_info.value)

tests/test_litellm/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py

@@ -4290,3 +4290,444 @@ def test_transform_response_does_not_leak_body_on_parse_failure():
     msg = str(exc_info.value)
     assert "secret content" not in msg
     assert "Error converting to valid response block" in msg
+
+
+def test_chunk_parser_raises_on_429_error_chunk():
+    """Test chunk_parser raises VertexAIError on 429 RESOURCE_EXHAUSTED error chunk"""
+    from unittest.mock import Mock
+
+    from litellm.llms.vertex_ai.common_utils import VertexAIError
+    from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
+        ModelResponseIterator,
+    )
+
+    error_chunk = {
+        "error": {
+            "code": 429,
+            "message": "Resource exhausted. Please try again later. Please refer to https://cloud.google.com/vertex-ai/generative-ai/docs/error-code-429 for more details.",
+            "status": "RESOURCE_EXHAUSTED",
+        }
+    }
+
+    logging_obj = Mock()
+    logging_obj.optional_params = {}
+
+    streaming_obj = ModelResponseIterator(
+        streaming_response=iter([]),
+        sync_stream=True,
+        logging_obj=logging_obj,
+    )
+
+    with pytest.raises(VertexAIError) as exc_info:
+        streaming_obj.chunk_parser(error_chunk)
+
+    assert exc_info.value.status_code == 429
+    assert "RESOURCE_EXHAUSTED" in exc_info.value.message
+    assert "Resource exhausted" in exc_info.value.message
+
+
+def test_chunk_parser_raises_on_500_error_chunk():
+    """Test chunk_parser raises VertexAIError on 500 INTERNAL error chunk"""
+    from unittest.mock import Mock
+
+    from litellm.llms.vertex_ai.common_utils import VertexAIError
+    from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
+        ModelResponseIterator,
+    )
+
+    error_chunk = {
+        "error": {
+            "code": 500,
+            "message": "Internal error encountered.",
+            "status": "INTERNAL",
+        }
+    }
+
+    logging_obj = Mock()
+    logging_obj.optional_params = {}
+
+    streaming_obj = ModelResponseIterator(
+        streaming_response=iter([]),
+        sync_stream=True,
+        logging_obj=logging_obj,
+    )
+
+    with pytest.raises(VertexAIError) as exc_info:
+        streaming_obj.chunk_parser(error_chunk)
+
+    assert exc_info.value.status_code == 500
+    assert "INTERNAL" in exc_info.value.message
+
+
+def test_chunk_parser_raises_on_error_chunk_with_minimal_fields():
+    """Test chunk_parser handles error chunks with missing optional fields"""
+    from unittest.mock import Mock
+
+    from litellm.llms.vertex_ai.common_utils import VertexAIError
+    from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
+        ModelResponseIterator,
+    )
+
+    error_chunk = {
+        "error": {
+            "code": 429,
+            "message": "Resource exhausted.",
+        }
+    }
+
+    logging_obj = Mock()
+    logging_obj.optional_params = {}
+
+    streaming_obj = ModelResponseIterator(
+        streaming_response=iter([]),
+        sync_stream=True,
+        logging_obj=logging_obj,
+    )
+
+    with pytest.raises(VertexAIError) as exc_info:
+        streaming_obj.chunk_parser(error_chunk)
+
+    assert exc_info.value.status_code == 429
+
+
+def test_chunk_parser_normal_chunk_unaffected_by_error_check():
+    """Test that normal streaming chunks still work correctly after error check addition"""
+    from unittest.mock import Mock
+
+    from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
+        ModelResponseIterator,
+    )
+
+    normal_chunk = {
+        "candidates": [
+            {
+                "content": {
+                    "role": "model",
+                    "parts": [{"text": "Hello"}],
+                },
+                "index": 0,
+            }
+        ],
+        "usageMetadata": {
+            "promptTokenCount": 5,
+            "candidatesTokenCount": 1,
+            "totalTokenCount": 6,
+        },
+    }
+
+    logging_obj = Mock()
+    logging_obj.optional_params = {}
+
+    streaming_obj = ModelResponseIterator(
+        streaming_response=iter([]),
+        sync_stream=True,
+        logging_obj=logging_obj,
+    )
+
+    result = streaming_obj.chunk_parser(normal_chunk)
+    assert result is not None
+    assert len(result.choices) > 0
+    assert result.choices[0].delta.content == "Hello"
+
+
+def test_chunk_parser_raises_on_non_dict_error():
+    """Test chunk_parser raises VertexAIError when chunk['error'] is not a dict"""
+    from unittest.mock import Mock
+
+    from litellm.llms.vertex_ai.common_utils import VertexAIError
+    from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
+        ModelResponseIterator,
+    )
+
+    error_chunk = {"error": "something went wrong"}
+
+    logging_obj = Mock()
+    logging_obj.optional_params = {}
+
+    streaming_obj = ModelResponseIterator(
+        streaming_response=iter([]),
+        sync_stream=True,
+        logging_obj=logging_obj,
+    )
+
+    with pytest.raises(VertexAIError) as exc_info:
+        streaming_obj.chunk_parser(error_chunk)
+
+    assert exc_info.value.status_code == 500
+    assert "Unexpected error format" in exc_info.value.message
+
+
+def test_chunk_parser_raises_on_string_error_code():
+    """Test chunk_parser correctly converts string error code to int"""
+    from unittest.mock import Mock
+
+    from litellm.llms.vertex_ai.common_utils import VertexAIError
+    from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
+        ModelResponseIterator,
+    )
+
+    # code field is a string "429" rather than an int
+    error_chunk = {
+        "error": {
+            "code": "429",
+            "message": "Resource exhausted.",
+            "status": "RESOURCE_EXHAUSTED",
+        }
+    }
+
+    logging_obj = Mock()
+    logging_obj.optional_params = {}
+
+    streaming_obj = ModelResponseIterator(
+        streaming_response=iter([]),
+        sync_stream=True,
+        logging_obj=logging_obj,
+    )
+
+    with pytest.raises(VertexAIError) as exc_info:
+        streaming_obj.chunk_parser(error_chunk)
+
+    assert exc_info.value.status_code == 429
+    assert isinstance(exc_info.value.status_code, int)
+
+
+def test_chunk_parser_error_chunk_explicit_null_code_uses_500():
+    """JSON null for code must not call int(None); status defaults to 500."""
+    from unittest.mock import Mock
+
+    from litellm.llms.vertex_ai.common_utils import VertexAIError
+    from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
+        ModelResponseIterator,
+    )
+
+    error_chunk = {
+        "error": {
+            "code": None,
+            "message": "Something went wrong.",
+            "status": "UNKNOWN",
+        }
+    }
+
+    logging_obj = Mock()
+    logging_obj.optional_params = {}
+
+    streaming_obj = ModelResponseIterator(
+        streaming_response=iter([]),
+        sync_stream=True,
+        logging_obj=logging_obj,
+    )
+
+    with pytest.raises(VertexAIError) as exc_info:
+        streaming_obj.chunk_parser(error_chunk)
+
+    assert exc_info.value.status_code == 500
+    assert "Something went wrong" in exc_info.value.message
+
+
+def test_chunk_parser_error_chunk_non_numeric_code_defaults_to_500():
+    """Non-numeric code must not become ValueError -> RuntimeError in __next__."""
+    from unittest.mock import Mock
+
+    from litellm.llms.vertex_ai.common_utils import VertexAIError
+    from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
+        ModelResponseIterator,
+    )
+
+    error_chunk = {
+        "error": {
+            "code": "NOT_A_NUMBER",
+            "message": "Malformed.",
+            "status": "INVALID",
+        }
+    }
+
+    logging_obj = Mock()
+    logging_obj.optional_params = {}
+
+    streaming_obj = ModelResponseIterator(
+        streaming_response=iter([]),
+        sync_stream=True,
+        logging_obj=logging_obj,
+    )
+
+    with pytest.raises(VertexAIError) as exc_info:
+        streaming_obj.chunk_parser(error_chunk)
+
+    assert exc_info.value.status_code == 500
+    assert "Malformed" in exc_info.value.message
+
+
+def test_chunk_parser_error_chunk_empty_dict_defaults_to_500():
+    """Empty error object {} uses default code 500 and default message/status strings."""
+    from unittest.mock import Mock
+
+    from litellm.llms.vertex_ai.common_utils import VertexAIError
+    from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
+        ModelResponseIterator,
+    )
+
+    error_chunk = {"error": {}}
+
+    logging_obj = Mock()
+    logging_obj.optional_params = {}
+
+    streaming_obj = ModelResponseIterator(
+        streaming_response=iter([]),
+        sync_stream=True,
+        logging_obj=logging_obj,
+    )
+
+    with pytest.raises(VertexAIError) as exc_info:
+        streaming_obj.chunk_parser(error_chunk)
+
+    assert exc_info.value.status_code == 500
+    assert "UNKNOWN" in exc_info.value.message
+    assert "Unknown error" in exc_info.value.message
+
+
+def test_chunk_parser_error_chunk_non_dict_int_value():
+    """Non-dict error payloads (e.g. bare JSON number) must raise with status 500, not TypeError."""
+    from unittest.mock import Mock
+
+    from litellm.llms.vertex_ai.common_utils import VertexAIError
+    from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
+        ModelResponseIterator,
+    )
+
+    error_chunk = {"error": 503}
+
+    logging_obj = Mock()
+    logging_obj.optional_params = {}
+
+    streaming_obj = ModelResponseIterator(
+        streaming_response=iter([]),
+        sync_stream=True,
+        logging_obj=logging_obj,
+    )
+
+    with pytest.raises(VertexAIError) as exc_info:
+        streaming_obj.chunk_parser(error_chunk)
+
+    assert exc_info.value.status_code == 500
+    assert "Unexpected error format" in exc_info.value.message
+    assert "503" in exc_info.value.message
+
+
+def test_chunk_parser_error_chunk_non_dict_null_value():
+    """JSON null for error must hit the non-dict branch (same as int/string)."""
+    from unittest.mock import Mock
+
+    from litellm.llms.vertex_ai.common_utils import VertexAIError
+    from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
+        ModelResponseIterator,
+    )
+
+    error_chunk = {"error": None}
+
+    logging_obj = Mock()
+    logging_obj.optional_params = {}
+
+    streaming_obj = ModelResponseIterator(
+        streaming_response=iter([]),
+        sync_stream=True,
+        logging_obj=logging_obj,
+    )
+
+    with pytest.raises(VertexAIError) as exc_info:
+        streaming_obj.chunk_parser(error_chunk)
+
+    assert exc_info.value.status_code == 500
+    assert "Unexpected error format" in exc_info.value.message
+
+
+def test_mid_stream_429_error_raises_during_iteration():
+    """
+    Simulate a full streaming scenario: normal thinking chunks arrive first,
+    then a 429 RESOURCE_EXHAUSTED error chunk arrives mid-stream.
+    Verify that ModelResponseIterator raises VertexAIError during iteration.
+    """
+    import json
+    from unittest.mock import Mock
+
+    from litellm.llms.vertex_ai.common_utils import VertexAIError
+    from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
+        ModelResponseIterator,
+    )
+
+    # Simulate Vertex AI SSE stream: normal chunks followed by a 429 error chunk
+    normal_chunk_1 = json.dumps(
+        {
+            "candidates": [
+                {
+                    "content": {
+                        "role": "model",
+                        "parts": [{"text": "Let me think about this...", "thought": True}],
+                    },
+                    "index": 0,
+                }
+            ],
+            "usageMetadata": {
+                "promptTokenCount": 10,
+                "candidatesTokenCount": 5,
+                "totalTokenCount": 15,
+            },
+            "modelVersion": "gemini-3.1-flash-image-preview",
+        }
+    )
+
+    normal_chunk_2 = json.dumps(
+        {
+            "candidates": [
+                {
+                    "content": {
+                        "role": "model",
+                        "parts": [{"text": "I'll generate the image now.", "thought": True}],
+                    },
+                    "index": 0,
+                }
+            ],
+            "usageMetadata": {
+                "promptTokenCount": 10,
+                "candidatesTokenCount": 12,
+                "totalTokenCount": 22,
+            },
+        }
+    )
+
+    error_chunk = json.dumps(
+        {
+            "error": {
+                "code": 429,
+                "message": "Resource exhausted. Please try again later. Please refer to https://cloud.google.com/vertex-ai/generative-ai/docs/error-code-429 for more details.",
+                "status": "RESOURCE_EXHAUSTED",
+            }
+        }
+    )
+
+    # Build a mock SSE stream (lines returned by iter_lines)
+    sse_lines = iter([normal_chunk_1, normal_chunk_2, error_chunk])
+
+    logging_obj = Mock()
+    logging_obj.optional_params = {}
+
+    streaming_obj = ModelResponseIterator(
+        streaming_response=sse_lines,
+        sync_stream=True,
+        logging_obj=logging_obj,
+    )
+
+    # Iterate the stream: first chunks should succeed, then 429 error should be raised
+    results = []
+    with pytest.raises(VertexAIError) as exc_info:
+        for chunk in streaming_obj:
+            if chunk is not None:
+                results.append(chunk)
+
+    # Verify: received normal chunks before the error
+    assert (
+        len(results) >= 1
+    ), "Should have received at least 1 normal chunk before the error"
+
+    # Verify: 429 error is properly raised
+    assert exc_info.value.status_code == 429
+    assert "RESOURCE_EXHAUSTED" in str(exc_info.value.message)

tests/test_litellm/llms/vertex_ai/test_vertex.py

@@ -1219,6 +1219,32 @@ def test_process_gemini_media():
         mime_type="image/jpeg", file_uri="gs://bucket/image"
     )
 
+    # Test gs url without extension using mime_type from image_url object
+    image_message = [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "image_url",
+                    "image_url": {
+                        "url": "gs://bucket/image-without-extension",
+                        "mime_type": "image/png",
+                    },
+                }
+            ],
+        }
+    ]
+    from litellm.llms.vertex_ai.gemini.transformation import (
+        _gemini_convert_messages_with_history,
+    )
+
+    converted = _gemini_convert_messages_with_history(
+        messages=image_message, model="gemini-2.5-flash"
+    )
+    assert converted[0]["parts"][0]["file_data"] == FileDataType(
+        mime_type="image/png", file_uri="gs://bucket/image-without-extension"
+    )
+
     # Test HTTPS JPG URL
     https_result = _process_gemini_media("https://example.com/image.jpg")
     print("https_result JPG", https_result)
@@ -1256,6 +1282,7 @@ def test_process_gemini_media():
     assert base64_result["inline_data"]["data"] == "/9j/4AAQSkZJRg..."
 
 
+
 def test_get_image_mime_type_from_url():
     """Test the _get_image_mime_type_from_url function for different image URLs"""
     from litellm.llms.vertex_ai.gemini.transformation import (

tests/test_litellm/llms/vertex_ai/test_vertex_gemini_gcs_uri_mime.py

@@ -0,0 +1,466 @@
+"""Vertex Gemini: extensionless gs:// MIME + GCS metadata tests.
+
+Split from test_vertex.py to satisfy CI per-file size limits.
+"""
+import asyncio
+import os
+import sys
+import time
+
+from dotenv import load_dotenv
+
+load_dotenv()
+
+import pytest
+
+import litellm
+from unittest.mock import MagicMock, patch
+
+sys.path.insert(0, os.path.abspath("../.."))
+
+from litellm.llms.vertex_ai.gemini.transformation import _process_gemini_media
+
+
+def test_process_gemini_media_gcs_explicit_format_octet_stream_and_alias():
+    """Explicit format bypasses registry; image/jpg alias still applies."""
+    from litellm.types.llms.vertex_ai import FileDataType
+
+    r1 = _process_gemini_media(
+        "gs://bucket/object-no-ext",
+        format="application/octet-stream",
+    )
+    assert r1["file_data"] == FileDataType(
+        mime_type="application/octet-stream",
+        file_uri="gs://bucket/object-no-ext",
+    )
+    r2 = _process_gemini_media("gs://bucket/object-no-ext", format="image/jpg")
+    assert r2["file_data"] == FileDataType(
+        mime_type="image/jpeg",
+        file_uri="gs://bucket/object-no-ext",
+    )
+
+
+def test_process_gemini_media_gcs_without_extension_errors_and_metadata_mock():
+    with patch(
+        "litellm.llms.vertex_ai.gemini.transformation._get_gcs_object_content_type",
+        return_value=None,
+    ):
+        with pytest.raises(litellm.BadRequestError) as exc:
+            _process_gemini_media("gs://bucket/image-without-extension")
+    assert "Unable to determine mime type for gs URI" in str(exc.value)
+
+    from litellm.types.llms.vertex_ai import FileDataType
+
+    with patch(
+        "litellm.llms.vertex_ai.gemini.transformation._get_gcs_object_content_type",
+        return_value="image/jpeg",
+    ) as m:
+        r = _process_gemini_media("gs://bucket/image-without-extension")
+    assert r["file_data"] == FileDataType(
+        mime_type="image/jpeg", file_uri="gs://bucket/image-without-extension"
+    )
+    m.assert_called()
+
+    with patch(
+        "litellm.llms.vertex_ai.gemini.transformation._get_gcs_object_content_type",
+        return_value="image/jpg",
+    ):
+        r_alias = _process_gemini_media("gs://bucket/image-without-extension")
+    assert r_alias["file_data"]["mime_type"] == "image/jpeg"
+
+
+def test_process_gemini_media_rejects_gcs_metadata_mime_not_supported_by_gemini():
+    """Non-empty GCS contentType that fails _normalize_and_validate_gemini_mime_type."""
+    with patch(
+        "litellm.llms.vertex_ai.gemini.transformation._get_gcs_object_content_type",
+        return_value="application/x-litellm-unit-test-unknown-mime",
+    ):
+        with pytest.raises(
+            litellm.BadRequestError,
+            match="File type not supported by gemini",
+        ):
+            _process_gemini_media("gs://bucket/object-without-extension")
+
+
+def test_file_block_uses_mime_type_alias_for_extensionless_gcs():
+    from litellm.llms.vertex_ai.gemini.transformation import (
+        _gemini_convert_messages_with_history,
+    )
+    from litellm.types.llms.vertex_ai import FileDataType
+
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "file",
+                    "file": {
+                        "file_id": "gs://bucket/no-extension-object",
+                        "mime_type": "application/pdf",
+                    },
+                }
+            ],
+        }
+    ]
+    converted = _gemini_convert_messages_with_history(
+        messages=messages, model="gemini-2.5-flash"
+    )
+    assert converted[0]["parts"][0]["file_data"] == FileDataType(
+        mime_type="application/pdf", file_uri="gs://bucket/no-extension-object"
+    )
+
+
+@pytest.mark.parametrize(
+    "bucket,expected",
+    [
+        (("a." * 110) + "aa", True),
+        ("ab", False),
+        ("a" * 64, False),
+        ("ab..cd", False),
+        ("1.2.3.4", False),
+        ("192.168.0.1", False),
+        ("Bucket-Upper", False),
+        ("bucket@name", False),
+        ("bucket name", False),
+        ("-mybucket", False),
+        ("mybucket-", False),
+        (".mybucket", False),
+        ("mybucket.", False),
+    ],
+)
+def test_is_valid_gcs_bucket_name_matrix(bucket, expected):
+    from litellm.llms.vertex_ai.gemini.transformation import _is_valid_gcs_bucket_name
+
+    assert _is_valid_gcs_bucket_name(bucket) is expected
+
+
+def test_get_gcs_object_content_type_explicit_vertex_success_and_token_failure():
+    from litellm.llms.vertex_ai.gemini import transformation as gt
+
+    mock_v = MagicMock()
+    mock_v.get_access_token.return_value = ("test-token", "test-project")
+    resp = MagicMock()
+    resp.is_error = False
+    resp.status_code = 200
+    resp.json.return_value = {"contentType": "image/png"}
+    http = MagicMock()
+    http.get.return_value = resp
+
+    with (
+        patch.object(gt, "_GCS_METADATA_VERTEX_BASE", mock_v),
+        patch(
+            "litellm.llms.vertex_ai.gemini.transformation._get_gcs_metadata_http_handler",
+            return_value=http,
+        ),
+    ):
+        assert (
+            gt._get_gcs_object_content_type(
+                image_url="gs://my-bucket/path/to/image-without-extension",
+                vertex_project="project-123",
+                vertex_credentials="credential-json",
+            )
+            == "image/png"
+        )
+    mock_v.get_access_token.assert_called_once_with(
+        credentials="credential-json",
+        project_id="project-123",
+    )
+
+    mock_v2 = MagicMock()
+    mock_v2.get_access_token.side_effect = Exception("token failure")
+    with patch.object(gt, "_GCS_METADATA_VERTEX_BASE", mock_v2):
+        with pytest.raises(
+            litellm.BadRequestError,
+            match="Unable to fetch GCS metadata with provided Vertex credentials/project",
+        ):
+            gt._get_gcs_object_content_type(
+                image_url="gs://my-bucket/path/to/image-without-extension",
+                vertex_project="project-123",
+                vertex_credentials="credential-json",
+            )
+
+
+def test_get_gcs_object_content_type_http_error_explicit_vs_anonymous():
+    from litellm.llms.vertex_ai.gemini import transformation as gt
+
+    mock_v = MagicMock()
+    mock_v.get_access_token.return_value = ("t", "p")
+    err_resp = MagicMock()
+    err_resp.is_error = True
+    err_resp.status_code = 403
+    err_resp.text = '{"error":{"message":"Permission denied"}}'
+    http = MagicMock()
+    http.get.return_value = err_resp
+
+    with (
+        patch.object(gt, "_GCS_METADATA_VERTEX_BASE", mock_v),
+        patch(
+            "litellm.llms.vertex_ai.gemini.transformation._get_gcs_metadata_http_handler",
+            return_value=http,
+        ),
+    ):
+        with pytest.raises(litellm.BadRequestError, match="HTTP 403") as ei:
+            gt._get_gcs_object_content_type(
+                image_url="gs://my-bucket/path/to/obj",
+                vertex_project="project-123",
+                vertex_credentials="credential-json",
+            )
+    assert "Permission denied" in str(ei.value)
+
+    mock_v2 = MagicMock()
+    anon_err = MagicMock()
+    anon_err.is_error = True
+    anon_err.status_code = 403
+    anon_err.text = "Forbidden"
+    http2 = MagicMock()
+    http2.get.return_value = anon_err
+    with (
+        patch.object(gt, "_GCS_METADATA_VERTEX_BASE", mock_v2),
+        patch(
+            "litellm.llms.vertex_ai.gemini.transformation._get_gcs_metadata_http_handler",
+            return_value=http2,
+        ),
+    ):
+        assert (
+            gt._get_gcs_object_content_type(image_url="gs://public-bucket/public-object")
+            is None
+        )
+    mock_v2.get_access_token.assert_not_called()
+
+
+def test_get_gcs_object_content_type_anonymous_success_no_auth_header():
+    from litellm.llms.vertex_ai.gemini import transformation as gt
+
+    mock_v = MagicMock()
+    ok = MagicMock()
+    ok.is_error = False
+    ok.status_code = 200
+    ok.json.return_value = {"contentType": "image/jpeg"}
+    http = MagicMock()
+    http.get.return_value = ok
+
+    with (
+        patch.object(gt, "_GCS_METADATA_VERTEX_BASE", mock_v),
+        patch(
+            "litellm.llms.vertex_ai.gemini.transformation._get_gcs_metadata_http_handler",
+            return_value=http,
+        ),
+    ):
+        assert (
+            gt._get_gcs_object_content_type(image_url="gs://public-bucket/public-object")
+            == "image/jpeg"
+        )
+    mock_v.get_access_token.assert_not_called()
+    hdrs = http.get.call_args.kwargs.get("headers")
+    assert hdrs is None or "Authorization" not in hdrs
+
+
+def test_async_transform_request_body_offloads_extensionless_gs_not_plain_text():
+    from litellm.llms.vertex_ai.gemini import transformation as gemini_transformation
+
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "image_url",
+                    "image_url": {"url": "gs://bucket/image-without-extension"},
+                }
+            ],
+        }
+    ]
+
+    def slow_http_get(*args, **kwargs):
+        time.sleep(0.5)
+        response = MagicMock()
+        response.is_error = False
+        response.status_code = 200
+        response.raise_for_status.return_value = None
+        response.json.return_value = {"contentType": "image/png"}
+        return response
+
+    async def fake_check_and_create_cache(self, **kwargs):
+        return kwargs["messages"], kwargs["optional_params"], None
+
+    mock_v = MagicMock()
+    mock_v.get_access_token.return_value = ("token", "project")
+    mock_http = MagicMock()
+    mock_http.get.side_effect = slow_http_get
+
+    async def run_scenario() -> float:
+        async def concurrent_sleep() -> float:
+            start = time.monotonic()
+            await asyncio.sleep(0.05)
+            return time.monotonic() - start
+
+        task = asyncio.create_task(
+            gemini_transformation.async_transform_request_body(
+                gemini_api_key=None,
+                messages=messages,
+                api_base=None,
+                model="gemini-2.5-flash",
+                client=None,
+                timeout=None,
+                extra_headers=None,
+                optional_params={},
+                logging_obj=MagicMock(),
+                custom_llm_provider="vertex_ai",
+                litellm_params={},
+                vertex_project=None,
+                vertex_location=None,
+                vertex_auth_header=None,
+            )
+        )
+        elapsed = await concurrent_sleep()
+        await task
+        return elapsed
+
+    with (
+        patch.object(gemini_transformation, "_GCS_METADATA_VERTEX_BASE", mock_v),
+        patch(
+            "litellm.llms.vertex_ai.gemini.transformation._get_gcs_metadata_http_handler",
+            return_value=mock_http,
+        ),
+        patch(
+            "litellm.llms.vertex_ai.context_caching.vertex_ai_context_caching."
+            "ContextCachingEndpoints.async_check_and_create_cache",
+            new=fake_check_and_create_cache,
+        ),
+    ):
+        sleep_elapsed = asyncio.run(run_scenario())
+
+    assert sleep_elapsed < 0.4, (
+        f"Event loop blocked for {sleep_elapsed:.3f}s; "
+        "async_transform_request_body did not offload sync GCS metadata"
+    )
+
+    async def fake_cache2(self, **kwargs):
+        return kwargs["messages"], kwargs["optional_params"], None
+
+    async def run_plain():
+        with patch(
+            "litellm.llms.vertex_ai.gemini.transformation.asyncify",
+            side_effect=AssertionError("asyncify must not run without extensionless gs://"),
+        ):
+            return await gemini_transformation.async_transform_request_body(
+                gemini_api_key=None,
+                messages=[{"role": "user", "content": "hello"}],
+                api_base=None,
+                model="gemini-2.5-flash",
+                client=None,
+                timeout=None,
+                extra_headers=None,
+                optional_params={},
+                logging_obj=MagicMock(),
+                custom_llm_provider="vertex_ai",
+                litellm_params={},
+                vertex_project=None,
+                vertex_location=None,
+                vertex_auth_header=None,
+            )
+
+    with patch(
+        "litellm.llms.vertex_ai.context_caching.vertex_ai_context_caching."
+        "ContextCachingEndpoints.async_check_and_create_cache",
+        new=fake_cache2,
+    ):
+        body = asyncio.run(run_plain())
+    assert body is not None and "contents" in body
+
+
+@pytest.mark.parametrize(
+    "messages,expected",
+    [
+        ([{"role": "user", "content": "hello"}], False),
+        (
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "image_url",
+                            "image_url": {"url": "gs://bucket/image-without-extension"},
+                        }
+                    ],
+                }
+            ],
+            True,
+        ),
+        (
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "image_url",
+                            "image_url": {"url": "gs://bucket/image.png"},
+                        }
+                    ],
+                }
+            ],
+            False,
+        ),
+        (
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "image_url",
+                            "image_url": {
+                                "url": "gs://bucket/image-without-extension",
+                                "mime_type": "image/png",
+                            },
+                        }
+                    ],
+                }
+            ],
+            False,
+        ),
+        (
+            [
+                {
+                    "role": "assistant",
+                    "content": [],
+                    "images": [
+                        {"image_url": {"url": "gs://bucket/gen-without-extension"}},
+                    ],
+                }
+            ],
+            True,
+        ),
+        (
+            [
+                {
+                    "role": "assistant",
+                    "content": [],
+                    "images": [{"image_url": {"url": "gs://bucket/gen.png"}}],
+                }
+            ],
+            False,
+        ),
+        (
+            [
+                {
+                    "role": "assistant",
+                    "content": [],
+                    "images": [
+                        {
+                            "image_url": {
+                                "url": "gs://bucket/gen-no-ext",
+                                "mime_type": "image/png",
+                            },
+                        }
+                    ],
+                }
+            ],
+            False,
+        ),
+    ],
+)
+def test_openai_messages_may_need_sync_gcs_metadata_fetch_matrix(messages, expected):
+    from litellm.llms.vertex_ai.gemini.transformation import (
+        _openai_messages_may_need_sync_gcs_metadata_fetch,
+    )
+
+    assert _openai_messages_may_need_sync_gcs_metadata_fetch(messages) is expected

tests/test_litellm/llms/vertex_ai/vertex_ai_partner_models/count_tokens/test_count_tokens_no_vertexai_sdk.py

@@ -0,0 +1,121 @@
+"""
+Regression tests for #28084:
+
+`VertexAIPartnerModels.count_tokens` (for Claude / Mistral / Llama on Vertex)
+used to gate on `import vertexai` even though the actual count-tokens path goes
+through `VertexAIPartnerModelsTokenCounter.handle_count_tokens_request`, which
+talks to the publisher's `:rawPredict` endpoint over plain httpx and never
+touches the Gemini SDK. The unused gate broke `/v1/messages/count_tokens` for
+any LiteLLM install that did not pull in `google-cloud-aiplatform` (which is
+not in the default `proxy` / `proxy-dev` extras).
+
+These tests pin the absence of that gate by:
+
+1. simulating `vertexai` being unimportable and verifying the partner-model
+   path does not raise the historical "vertexai import failed" error before
+   reaching the network/auth layer, and
+2. asserting that import of the partner-model count-tokens handler module by
+   itself does not pull `vertexai` into `sys.modules`.
+"""
+
+import sys
+
+import pytest
+
+from litellm.llms.vertex_ai.vertex_ai_partner_models.count_tokens.handler import (
+    VertexAIPartnerModelsTokenCounter,
+)
+from litellm.llms.vertex_ai.vertex_ai_partner_models.main import VertexAIPartnerModels
+
+
+@pytest.mark.asyncio
+async def test_count_tokens_does_not_require_vertexai_sdk(monkeypatch):
+    """Even when `import vertexai` would fail, count_tokens must not raise the
+    historical "vertexai import failed" gate. The downstream handler talks to
+    `:rawPredict` over httpx with an access token — no Gemini SDK needed."""
+
+    # Simulate `vertexai` being unimportable, regardless of what is actually on
+    # the test environment's sys.path.
+    monkeypatch.setitem(sys.modules, "vertexai", None)
+    monkeypatch.setitem(sys.modules, "vertexai.preview", None)
+
+    captured = {}
+
+    async def fake_ensure_access_token(
+        self, credentials, project_id, custom_llm_provider
+    ):
+        return "fake-token", "fake-project"
+
+    def fake_build_endpoint(self, model, project_id, vertex_location, api_base=None):
+        captured["model_to_endpoint"] = model
+        return "https://fake-endpoint"
+
+    monkeypatch.setattr(
+        VertexAIPartnerModelsTokenCounter,
+        "_ensure_access_token_async",
+        fake_ensure_access_token,
+    )
+    monkeypatch.setattr(
+        VertexAIPartnerModelsTokenCounter,
+        "_build_count_tokens_endpoint",
+        fake_build_endpoint,
+    )
+
+    class FakeResponse:
+        status_code = 200
+
+        def json(self):
+            return {"input_tokens": 9}
+
+    class FakeClient:
+        async def post(self, url, headers=None, json=None, **kwargs):
+            captured["url"] = url
+            captured["headers"] = headers
+            captured["json"] = json
+            return FakeResponse()
+
+    import litellm.llms.vertex_ai.vertex_ai_partner_models.count_tokens.handler as handler_mod
+
+    monkeypatch.setattr(
+        handler_mod, "get_async_httpx_client", lambda **kwargs: FakeClient()
+    )
+
+    result = await VertexAIPartnerModels().count_tokens(
+        model="claude-sonnet-4-6",
+        messages=[{"role": "user", "content": "hello"}],
+        litellm_params={"vertex_location": "us-east5"},
+        vertex_project="test-project",
+        vertex_location="us-east5",
+        vertex_credentials=None,
+    )
+
+    # We should reach the publisher endpoint and parse its response, not raise
+    # the vertexai-import gate.
+    assert result == {
+        "input_tokens": 9,
+        "tokenizer_used": "vertex_ai_partner_models",
+    }
+    assert captured["headers"] == {"Authorization": "Bearer fake-token"}
+    assert captured["model_to_endpoint"] == "claude-sonnet-4-6"
+
+
+def test_handler_module_does_not_import_vertexai_sdk():
+    """Importing the partner-model count-tokens handler must not load the
+    Gemini SDK into sys.modules. Operators who only need Claude-on-Vertex
+    token counting should not pay for `google-cloud-aiplatform`."""
+
+    # Force-evict any prior load so this assertion measures what THIS module
+    # pulls in, not what an unrelated earlier test did.
+    for mod in list(sys.modules):
+        if mod == "vertexai" or mod.startswith("vertexai."):
+            sys.modules.pop(mod, None)
+
+    # Re-import the handler module to verify it stays SDK-free.
+    import importlib
+
+    import litellm.llms.vertex_ai.vertex_ai_partner_models.count_tokens.handler as handler_mod
+
+    importlib.reload(handler_mod)
+
+    leaked = [m for m in sys.modules if m == "vertexai" or m.startswith("vertexai.")]
+    assert leaked == [], f"unexpected vertexai SDK imports: {leaked}"

tests/test_litellm/proxy/guardrails/guardrail_hooks/test_generic_guardrail_api.py

@@ -553,6 +553,7 @@ class TestGuardrailActions:
             # Verify the exception has the clean error message (no wrapper)
             assert str(exc_info.value) == "Content contains harmful instructions"
             assert exc_info.value.guardrail_name == "generic_guardrail_api"
+            assert exc_info.value.status_code == 400
 
     @pytest.mark.asyncio
     async def test_action_intervened_modifies_content(

tests/test_litellm/proxy/management_endpoints/test_model_management_endpoints.py

@@ -1446,3 +1446,175 @@ class TestGetTeamDeployments:
         result = await _get_team_deployments(team_id, prisma_client)
         assert len(result) == 1
         assert result[0] is dep1
+
+
+def _build_db_model_for_blocked_test():
+    from litellm.types.router import Deployment, LiteLLM_Params, ModelInfo
+
+    return Deployment(
+        model_name="gpt-4o",
+        litellm_params=LiteLLM_Params(model="openai/gpt-4o"),
+        model_info=ModelInfo(id="dep-0"),
+    )
+
+
+class TestUpdateDBModelBlocked:
+    """`update_db_model` must thread `blocked` through to the Prisma payload only
+    when the caller explicitly set it — PATCH semantics: an absent field means
+    "leave the stored value untouched"."""
+
+    def test_update_db_model_passes_blocked_true_to_db(self):
+        from litellm.proxy.management_endpoints.model_management_endpoints import (
+            update_db_model,
+        )
+
+        result = update_db_model(
+            db_model=_build_db_model_for_blocked_test(),
+            updated_patch=updateDeployment(blocked=True),
+        )
+        assert result["blocked"] is True
+
+    def test_update_db_model_passes_blocked_false_to_db(self):
+        from litellm.proxy.management_endpoints.model_management_endpoints import (
+            update_db_model,
+        )
+
+        result = update_db_model(
+            db_model=_build_db_model_for_blocked_test(),
+            updated_patch=updateDeployment(blocked=False),
+        )
+        assert result["blocked"] is False
+
+    def test_update_db_model_omits_blocked_when_patch_is_none(self):
+        from litellm.proxy.management_endpoints.model_management_endpoints import (
+            update_db_model,
+        )
+
+        result = update_db_model(
+            db_model=_build_db_model_for_blocked_test(),
+            updated_patch=updateDeployment(),
+        )
+        assert "blocked" not in result
+
+
+class TestGetModelInfoWithIdBlocked:
+    """`ProxyConfig.get_model_info_with_id` must propagate the DB-level `blocked`
+    column into the in-memory `model_info` dict so the router filter can read it."""
+
+    def test_get_model_info_with_id_propagates_blocked_true(self):
+        from litellm.proxy.proxy_server import ProxyConfig
+
+        model = MagicMock()
+        model.model_id = "dep-1"
+        model.model_info = {}
+        model.blocked = True
+        info = ProxyConfig().get_model_info_with_id(model=model, db_model=True)
+        assert info.id == "dep-1"
+        assert getattr(info, "blocked") is True
+
+    def test_get_model_info_with_id_defaults_blocked_to_false_when_missing(self):
+        from litellm.proxy.proxy_server import ProxyConfig
+
+        model = MagicMock(spec=["model_id", "model_info"])
+        model.model_id = "dep-2"
+        model.model_info = {}
+        info = ProxyConfig().get_model_info_with_id(model=model, db_model=True)
+        assert getattr(info, "blocked") is False
+
+
+class TestPatchModelBlockedAuthGate:
+    """Only proxy admins may flip `blocked` — team admins authorized for
+    team-scoped models via `can_user_make_model_call` must still be rejected
+    when they attempt to toggle the pause flag."""
+
+    @pytest.mark.asyncio
+    async def test_team_admin_cannot_toggle_blocked(self):
+        from litellm.proxy.management_endpoints.model_management_endpoints import (
+            patch_model,
+        )
+
+        non_admin = UserAPIKeyAuth(
+            user_id="team_admin",
+            user_role=LitellmUserRoles.INTERNAL_USER,
+        )
+        existing_row = MagicMock()
+        existing_row.litellm_params = {"model": "openai/gpt-4o-mini"}
+        existing_row.model_dump.return_value = {
+            "model_name": "gpt-4o-mini",
+            "litellm_params": existing_row.litellm_params,
+            "model_info": {"id": "m1"},
+        }
+        existing_row.model_dump_json.return_value = "{}"
+
+        mock_prisma = MagicMock()
+        mock_prisma.db.litellm_proxymodeltable.find_unique = AsyncMock(
+            return_value=existing_row
+        )
+
+        with (
+            patch("litellm.proxy.proxy_server.prisma_client", mock_prisma),
+            patch("litellm.proxy.proxy_server.llm_router", MagicMock()),
+            patch("litellm.proxy.proxy_server.store_model_in_db", True),
+            patch("litellm.proxy.proxy_server.premium_user", True),
+            patch(
+                "litellm.proxy.management_endpoints.model_management_endpoints.ModelManagementAuthChecks.can_user_make_model_call",
+                new=AsyncMock(return_value=None),
+            ),
+        ):
+            with pytest.raises(Exception) as exc_info:
+                await patch_model(
+                    model_id="m1",
+                    patch_data=updateDeployment(blocked=True),
+                    user_api_key_dict=non_admin,
+                )
+            err = exc_info.value
+            assert getattr(err, "param", "") == "blocked"
+            assert "proxy admin" in getattr(err, "message", "").lower()
+
+    @pytest.mark.asyncio
+    async def test_proxy_admin_can_toggle_blocked(self):
+        from litellm.proxy.management_endpoints.model_management_endpoints import (
+            patch_model,
+        )
+
+        admin = UserAPIKeyAuth(user_id="admin", user_role=LitellmUserRoles.PROXY_ADMIN)
+        existing_row = MagicMock()
+        existing_row.litellm_params = {"model": "openai/gpt-4o-mini"}
+        existing_row.model_dump.return_value = {
+            "model_name": "gpt-4o-mini",
+            "litellm_params": existing_row.litellm_params,
+            "model_info": {"id": "m1"},
+        }
+        existing_row.model_dump_json.return_value = "{}"
+        updated_row = MagicMock()
+        updated_row.model_dump_json.return_value = "{}"
+
+        mock_prisma = MagicMock()
+        mock_prisma.db.litellm_proxymodeltable.find_unique = AsyncMock(
+            return_value=existing_row
+        )
+        mock_prisma.db.litellm_proxymodeltable.update = AsyncMock(
+            return_value=updated_row
+        )
+
+        with (
+            patch("litellm.proxy.proxy_server.prisma_client", mock_prisma),
+            patch("litellm.proxy.proxy_server.llm_router", MagicMock()),
+            patch("litellm.proxy.proxy_server.store_model_in_db", True),
+            patch("litellm.proxy.proxy_server.premium_user", True),
+            patch(
+                "litellm.proxy.management_endpoints.model_management_endpoints.ModelManagementAuthChecks.can_user_make_model_call",
+                new=AsyncMock(return_value=None),
+            ),
+            patch(
+                "litellm.proxy.management_endpoints.model_management_endpoints.clear_cache",
+                new=AsyncMock(return_value=None),
+            ),
+        ):
+            result = await patch_model(
+                model_id="m1",
+                patch_data=updateDeployment(blocked=True),
+                user_api_key_dict=admin,
+            )
+            assert result is updated_row
+            mock_prisma.db.litellm_proxymodeltable.update.assert_awaited_once()

tests/test_litellm/proxy/management_endpoints/test_team_endpoints.py

@@ -7943,3 +7943,103 @@ async def test_team_member_me_returns_404_for_unknown_team(mock_db_client):
                 user_api_key_dict=caller_auth,
             )
     assert exc_info.value.status_code == 404
+
+
+def _non_admin_auth():
+    return UserAPIKeyAuth(
+        user_id="u-team-admin", user_role=LitellmUserRoles.INTERNAL_USER
+    )
+
+
+def test_check_passthrough_routes_caller_permission_team():
+    from litellm.proxy._types import NewTeamRequest
+    from litellm.proxy.management_endpoints.common_utils import (
+        _check_passthrough_routes_caller_permission,
+    )
+
+    admin = UserAPIKeyAuth(user_role=LitellmUserRoles.PROXY_ADMIN)
+    non_admin = _non_admin_auth()
+
+    _check_passthrough_routes_caller_permission(
+        NewTeamRequest(allowed_passthrough_routes=["/foo/*"]), admin, entity="team"
+    )
+
+    _check_passthrough_routes_caller_permission(
+        NewTeamRequest(), non_admin, entity="team"
+    )
+    _check_passthrough_routes_caller_permission(
+        NewTeamRequest(allowed_passthrough_routes=[]), non_admin, entity="team"
+    )
+
+    with pytest.raises(HTTPException) as exc:
+        _check_passthrough_routes_caller_permission(
+            NewTeamRequest(allowed_passthrough_routes=["/admin/*"]),
+            non_admin,
+            entity="team",
+        )
+    assert exc.value.status_code == 403
+    assert "allowed_passthrough_routes" in str(exc.value.detail)
+    assert "team" in str(exc.value.detail)
+
+    with pytest.raises(HTTPException) as exc:
+        _check_passthrough_routes_caller_permission(
+            NewTeamRequest(metadata={"allowed_passthrough_routes": ["/admin/*"]}),
+            non_admin,
+            entity="team",
+        )
+    assert exc.value.status_code == 403
+    assert "metadata.allowed_passthrough_routes" in str(exc.value.detail)
+
+
+@pytest.mark.asyncio
+async def test_new_team_blocks_non_admin_passthrough_routes(mock_db_client):
+    """A non-proxy-admin cannot self-grant pass-through routes via /team/new."""
+    mock_db_client.db.litellm_teamtable.count = AsyncMock(return_value=0)
+    from fastapi import Request
+
+    from litellm.proxy._types import NewTeamRequest, ProxyException
+    from litellm.proxy.management_endpoints.team_endpoints import new_team
+
+    with patch(
+        "litellm.proxy.management_endpoints.team_endpoints._check_user_team_limits",
+        AsyncMock(return_value=None),
+    ):
+        with pytest.raises(ProxyException) as exc:
+            await new_team(
+                data=NewTeamRequest(
+                    team_alias="t", allowed_passthrough_routes=["/admin/*"]
+                ),
+                http_request=MagicMock(spec=Request),
+                user_api_key_dict=_non_admin_auth(),
+            )
+    assert str(exc.value.code) == "403"
+    assert "allowed_passthrough_routes" in str(exc.value.message)
+
+
+@pytest.mark.asyncio
+async def test_update_team_blocks_non_admin_passthrough_routes(mock_db_client):
+    """Even a team manager (non-proxy-admin) cannot set pass-through routes via
+    /team/update — the gate runs after _verify_team_access."""
+    from fastapi import Request
+
+    from litellm.proxy._types import ProxyException, UpdateTeamRequest
+    from litellm.proxy.management_endpoints.team_endpoints import update_team
+
+    existing = MagicMock()
+    existing.model_dump.return_value = {"team_id": "t1"}
+    mock_db_client.db.litellm_teamtable.find_unique = AsyncMock(return_value=existing)
+
+    with patch(
+        "litellm.proxy.management_endpoints.team_endpoints._verify_team_access",
+        AsyncMock(return_value=None),
+    ):
+        with pytest.raises(ProxyException) as exc:
+            await update_team(
+                data=UpdateTeamRequest(
+                    team_id="t1", allowed_passthrough_routes=["/admin/*"]
+                ),
+                http_request=MagicMock(spec=Request),
+                user_api_key_dict=_non_admin_auth(),
+            )
+    assert str(exc.value.code) == "403"
+    assert "allowed_passthrough_routes" in str(exc.value.message)