From 21c8c337f75121ee4f487e242c05e7ee570983cd Mon Sep 17 00:00:00 2001
From: Apricooooot <82204590+Apricooooot@users.noreply.github.com>
Date: Tue, 18 Aug 2026 12:50:06 -0700
Subject: [PATCH] fix: use resolved credentials in health checks

---
 litellm/main.py                               | 15 ++++++---
 .../litellm_utils_tests/test_health_check.py  | 32 +++++++++++++++++++
 2 files changed, 43 insertions(+), 4 deletions(-)

diff --git a/litellm/main.py b/litellm/main.py
index 2a8ed6c87b6..6a0401842f9 100644
--- a/litellm/main.py
+++ b/litellm/main.py
@@ -8385,29 +8385,36 @@ async def ahealth_check(
         api_base_from_params: Final = model_params.get("api_base", None)
         api_key_from_params: Final = model_params.get("api_key", None)
 
-        model, custom_llm_provider, _, _ = get_llm_provider(
+        model, custom_llm_provider, dynamic_api_key, dynamic_api_base = get_llm_provider(
             model=model,
             custom_llm_provider=custom_llm_provider_from_params,
             api_base=api_base_from_params,
             api_key=api_key_from_params,
         )
+        resolved_model_params: Final = dict(  # mutable-ok: health-check handlers require mutable request parameters
+            model_params
+        )
+        if dynamic_api_base is not None:
+            resolved_model_params["api_base"] = dynamic_api_base
+        if dynamic_api_key is not None:
+            resolved_model_params["api_key"] = dynamic_api_key
+        resolved_model_params["cache"] = {"no-cache": True}
         if model in litellm.model_cost and mode is None:
             mode = litellm.model_cost[model].get("mode")
 
-        model_params["cache"] = {"no-cache": True}  # don't used cached responses for making health check calls
         mode = mode or "chat"
         if "*" in model:
             return await HealthCheckHelpers.ahealth_check_wildcard_models(
                 model=model,
                 custom_llm_provider=custom_llm_provider,
-                model_params=model_params,
+                model_params=resolved_model_params,
                 litellm_logging_obj=litellm_logging_obj,
             )
 
         mode_handlers: Final = HealthCheckHelpers.get_mode_handlers(
             model=model,
             custom_llm_provider=custom_llm_provider,
-            model_params=model_params,
+            model_params=resolved_model_params,
             prompt=prompt,
             input=input,
         )
diff --git a/tests/litellm_utils_tests/test_health_check.py b/tests/litellm_utils_tests/test_health_check.py
index de6f7c38fed..321538c0508 100644
--- a/tests/litellm_utils_tests/test_health_check.py
+++ b/tests/litellm_utils_tests/test_health_check.py
@@ -15,6 +15,38 @@ import asyncio
 import litellm
 
 
+@pytest.mark.asyncio
+async def test_hosted_vllm_health_check_uses_resolved_provider_credentials(
+    monkeypatch,
+):
+    monkeypatch.setenv("HOSTED_VLLM_API_BASE", "https://env.example/v1")
+    monkeypatch.setenv("HOSTED_VLLM_API_KEY", "env-key")
+
+    mock_response = litellm.ModelResponse(
+        choices=[{"message": {"role": "assistant", "content": "ok"}}]
+    )
+    with (
+        patch(
+            "litellm.litellm_core_utils.health_check_helpers.HealthCheckHelpers."
+            "_update_model_params_with_health_check_tracking_information",
+            side_effect=lambda model_params: model_params,
+        ),
+        patch(
+            "litellm.acompletion",
+            new_callable=AsyncMock,
+            return_value=mock_response,
+        ) as mock_acompletion,
+    ):
+        response = await litellm.ahealth_check(
+            model_params={"model": "hosted_vllm/org/model"},
+            mode="chat",
+        )
+
+    assert "error" not in response
+    assert mock_acompletion.await_args.kwargs["api_base"] == "https://env.example/v1"
+    assert mock_acompletion.await_args.kwargs["api_key"] == "env-key"
+
+
 @pytest.mark.asyncio
 async def test_azure_health_check():
     response = await litellm.ahealth_check(