fix(azure): match deployment segments case-insensitively and keep relay helpers immutable

litellm/llms/azure/passthrough/transformation.py

@@ -1,5 +1,6 @@
 import re
 from collections.abc import Callable, Collection, Mapping, Sequence
+from types import MappingProxyType
 from typing import TYPE_CHECKING, Final, Optional
 
 import httpx
@@ -55,9 +56,7 @@ def logged_responses_stream(all_chunks: Sequence[str], logging_obj: Logging) ->
     """A streaming logging object assembles the logged response from the terminal event, not from its body."""
     from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig
 
-    terminal_event: Final = OpenAIResponsesAPIConfig.parse_terminal_event_from_stream_chunks(
-        all_chunks=list(all_chunks)
-    )
+    terminal_event: Final = OpenAIResponsesAPIConfig.parse_terminal_event_from_stream_chunks(all_chunks=all_chunks)
     if terminal_event is None:
         return None
     logging_obj.call_type = (
@@ -83,7 +82,11 @@ def foreign_azure_deployment(
     if match is None:
         return None
     deployment: Final = match.group(1)
-    return None if deployment == model_group or deployment in served_models() else deployment
+    folded: Final = deployment.casefold()
+    if folded == model_group.casefold():
+        return None
+    served: Final = frozenset(name.casefold() for name in served_models())
+    return None if folded in served else deployment
 
 
 def without_api_version(api_base: str) -> str:
@@ -119,7 +122,9 @@ class AzurePassthroughConfig(BasePassthroughConfig):
         relay_base: Final = without_api_version(base_target_url) if caller_api_version else base_target_url
         complete_url: Final = BaseAzureLLM._get_base_azure_url(
             api_base=relay_base,
-            litellm_params={**litellm_params, "api_version": caller_api_version or litellm_params.get("api_version")},
+            litellm_params=MappingProxyType(
+                {**litellm_params, "api_version": caller_api_version or litellm_params.get("api_version")}
+            ),
             route=native_endpoint,
         )
         return (

litellm/llms/openai/responses/transformation.py

@@ -620,7 +620,7 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
             return event_pydantic_model.model_construct(**parsed_chunk)
 
     @staticmethod
-    def parse_terminal_event_from_stream_chunks(all_chunks: list[str]) -> ResponsesTerminalEvent | None:
+    def parse_terminal_event_from_stream_chunks(all_chunks: Sequence[str]) -> ResponsesTerminalEvent | None:
         for chunk_str in reversed(all_chunks):
             for event_model in (ResponseCompletedEvent, ResponseIncompleteEvent, ResponseFailedEvent):
                 try:

litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py

@@ -122,6 +122,10 @@ def is_passthrough_request_using_router_model(request_body: dict, llm_router: li
         return False
 
 
+class RelayRejection(TypedDict):
+    error: ReadOnly[str]
+
+
 def _deployment_model_name(litellm_params: LiteLLMParamsTypedDict) -> str:
     model: Final = litellm_params.get("model", "")
     try:
@@ -1543,13 +1547,11 @@ async def _relay_azure_router_model(
         endpoint, model, lambda: _models_served_by_group(llm_router, model)
     )
     if foreign_deployment is not None:
-        raise HTTPException(
-            status_code=400,
-            detail={
-                "error": f"deployment '{foreign_deployment}' in the path is not served by model group '{model}'; "
-                "put the model group name in the deployments segment"
-            },
-        )
+        rejection: Final[RelayRejection] = {
+            "error": f"deployment '{foreign_deployment}' in the path is not served by model group '{model}'; "
+            "put the model group name in the deployments segment"
+        }
+        raise HTTPException(status_code=400, detail=rejection)
     try:
         result: Final = await llm_router.allm_passthrough_route(
             model=model,

tests/test_litellm/llms/azure/passthrough/test_azure_passthrough_transformation.py

@@ -456,16 +456,26 @@ def test_azure_passthrough_is_streaming_request_reads_the_stream_flag(request_da
         ("gpt/openai/deployments/gpt/chat/completions", None),
         ("openai/deployments/gpt/chat/completions", None),
         ("gpt/openai/deployments/gpt-5.4-mini/chat/completions", None),
+        ("gpt/openai/deployments/GPT-5.4-MINI/chat/completions", None),
+        ("gpt/openai/deployments/Gpt/chat/completions", None),
         ("gpt/models/chat/completions", None),
         ("gpt/openai/deployments/gpt-5.4/chat/completions", "gpt-5.4"),
         ("gpt/openai/deployments/other-group/chat/completions", "other-group"),
         ("openai/deployments/victim/gpt/chat/completions", "victim"),
+        ("gpt/openai/deployments/GPT-5.4/chat/completions", "GPT-5.4"),
     ],
 )
 def test_foreign_azure_deployment_names_a_segment_outside_the_group(endpoint, expected):
     assert foreign_azure_deployment(endpoint, "gpt", lambda: frozenset({"gpt-5.4-mini"})) == expected
 
 
+def test_foreign_azure_deployment_skips_the_router_when_the_segment_is_the_group_itself():
+    def served_models():
+        raise AssertionError("the router must not be consulted for the group's own name")
+
+    assert foreign_azure_deployment("gpt/openai/deployments/Gpt/chat/completions", "gpt", served_models) is None
+
+
 @pytest.mark.parametrize(
     "endpoint, expected",
     [