fix(health): splice wildcard prefix onto provider-stripped health check candidates

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/litellm_core_utils/health_check_helpers.py

@@ -9,7 +9,7 @@ from typing import TYPE_CHECKING, Final, Literal
 from litellm.llms.base_llm.ocr.transformation import DocumentType
 from litellm.rust_bridge import runtime
 from litellm.rust_bridge.ocr.entrypoints import NATIVE_OCR_HEALTH_CHECK_DOCUMENT
-from litellm.types.utils import LIST_BATCHES_SUPPORTED_PROVIDERS
+from litellm.types.utils import LIST_BATCHES_SUPPORTED_PROVIDERS, LlmProviders
 
 if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging
@@ -39,16 +39,23 @@ def _ocr_health_check_document(model: str, custom_llm_provider: str) -> Document
     return native(model, custom_llm_provider)
 
 
+def _strip_known_provider_prefix(model: str, known_providers: frozenset[str]) -> str:
+    leading, sep, model_suffix = model.partition("/")
+    return model_suffix if sep and leading in known_providers else model
+
+
 def _wildcard_health_check_models(wildcard_model: str, cheapest_models: Sequence[str]) -> tuple[str, ...]:
-    """Substitute each candidate id into the wildcard's dynamic segment so the health
-    check probes a name the wildcard route actually expands, instead of the bare
-    cost-map id (e.g. `databricks/system.ai.*` + `databricks/databricks-x` ->
-    `databricks/system.ai.databricks-x`)."""
     provider_prefix, sep, wildcard_suffix = wildcard_model.partition("/")
     if not sep or wildcard_suffix == "*":
         return tuple(cheapest_models)
+    known_providers: Final = frozenset(provider.value for provider in LlmProviders)
+    literal_prefix: Final = wildcard_suffix.replace("*", "")
+    stripped_ids: Final = tuple(_strip_known_provider_prefix(model, known_providers) for model in cheapest_models)
     return tuple(
-        f"{provider_prefix}/{wildcard_suffix.replace('*', model.partition('/')[-1], 1)}" for model in cheapest_models
+        f"{provider_prefix}/{stripped}"
+        if stripped.startswith(literal_prefix)
+        else f"{provider_prefix}/{wildcard_suffix.replace('*', stripped, 1)}"
+        for stripped in stripped_ids
     )
 
 

tests/test_litellm/proxy/auth/test_model_checks.py

@@ -12,7 +12,9 @@ def test_get_team_models_for_all_models_and_team_only_models():
     model_access_groups = {}
     include_model_access_groups = False
 
-    result = get_team_models(team_models, proxy_model_list, model_access_groups, include_model_access_groups)
+    result = get_team_models(
+        team_models, proxy_model_list, model_access_groups, include_model_access_groups
+    )
     combined_models = team_models + proxy_model_list
     assert set(result) == set(combined_models)
 
@@ -245,7 +247,9 @@ def test_get_key_models_does_not_mutate_input():
         ),
     ],
 )
-def test_get_complete_model_list_order(key_models, team_models, proxy_model_list, model_list, expected):
+def test_get_complete_model_list_order(
+    key_models, team_models, proxy_model_list, model_list, expected
+):
     """
     Test that get_complete_model_list preserves order
     """
@@ -398,7 +402,9 @@ def test_wildcard_credential_hydration_preserves_deployment_params(
         captured_params["api_key"] = litellm_params.api_key
         captured_params["api_version"] = litellm_params.api_version
         captured_params["credential_name"] = litellm_params.litellm_credential_name
-        captured_params["has_unexpected_field"] = hasattr(litellm_params, "unexpected_field")
+        captured_params["has_unexpected_field"] = hasattr(
+            litellm_params, "unexpected_field"
+        )
         return ["gpt-4o"]
 
     monkeypatch.setattr(model_checks, "get_provider_models", fake_get_provider_models)
@@ -443,7 +449,9 @@ def test_wildcard_custom_prefix_does_not_stack_provider_prefix(monkeypatch):
 
     result = get_known_models_from_wildcard(
         wildcard_model="ollama_server1/*",
-        litellm_params=LiteLLM_Params(model="ollama_chat/*", custom_llm_provider="ollama_chat"),
+        litellm_params=LiteLLM_Params(
+            model="ollama_chat/*", custom_llm_provider="ollama_chat"
+        ),
     )
 
     assert result == ["ollama_server1/gemma3:1b", "ollama_server1/llama3:8b"]
@@ -470,18 +478,15 @@ def test_wildcard_custom_prefix_keeps_org_segment_for_non_provider_first_segment
 
     result = get_known_models_from_wildcard(
         wildcard_model="my_hf/*",
-        litellm_params=LiteLLM_Params(model="huggingface/*", custom_llm_provider="huggingface"),
+        litellm_params=LiteLLM_Params(
+            model="huggingface/*", custom_llm_provider="huggingface"
+        ),
     )
 
     assert result == ["my_hf/meta-llama/Llama-3-8B"]
 
 
 def test_partial_prefix_wildcard_expands_against_stripped_provider_ids(monkeypatch):
-    """A partial wildcard like ``databricks/system.ai.*`` must splice the prefix before
-    ``*`` onto the provider-stripped model id, not onto the whole cost-map key. Prepending
-    to the provider-prefixed key yielded ``databricks/system.ai.databricks/databricks-x``,
-    an id no upstream accepts.
-    """
     import litellm
     from litellm.proxy.auth import model_checks
     from litellm.proxy.auth.model_checks import get_known_models_from_wildcard
@@ -953,7 +958,9 @@ def test_add_known_models_refreshes_models_by_provider_for_wildcard_expansion():
     assert fake_model not in litellm.models_by_provider["vertex_ai"]
     try:
         litellm.add_known_models(
-            model_cost_map={fake_model: {"litellm_provider": "vertex_ai-language-models", "mode": "chat"}}
+            model_cost_map={
+                fake_model: {"litellm_provider": "vertex_ai-language-models", "mode": "chat"}
+            }
         )
         assert fake_model in litellm.models_by_provider["vertex_ai"]
         assert litellm.models_by_provider is captured_reference
@@ -1001,4 +1008,6 @@ def test_transcribe_is_a_known_provider_for_wildcard_expansion():
     assert "transcribe" in litellm.models_by_provider
     assert "transcribe/StartTranscriptionJob" in litellm.models_by_provider["transcribe"]
     assert get_provider_models("transcribe") == ["transcribe/StartTranscriptionJob"]
-    assert get_known_models_from_wildcard("transcribe/*") == ["transcribe/StartTranscriptionJob"]
+    assert get_known_models_from_wildcard("transcribe/*") == [
+        "transcribe/StartTranscriptionJob"
+    ]

tests/unit/litellm_core_utils/test_health_check_helpers.py

@@ -566,3 +566,11 @@ def test_wildcard_health_check_models_bare_provider_and_star_wildcard_unchanged(
     candidates: Final = ("databricks/databricks-gpt-5",)
     assert _wildcard_health_check_models("databricks/*", candidates) == candidates
     assert _wildcard_health_check_models("*", candidates) == candidates
+
+
+def test_wildcard_health_check_models_partial_prefix_matching_literal_keeps_stripped_id():
+    assert _wildcard_health_check_models("openai/gpt-4*", ["gpt-4o-mini"]) == ("openai/gpt-4o-mini",)
+
+
+def test_wildcard_health_check_models_partial_prefix_splices_suffix_around_star():
+    assert _wildcard_health_check_models("openai/ft:*", ["gpt-4o-mini"]) == ("openai/ft:gpt-4o-mini",)