fix(health-check): skip max_tokens for non-chat modes (image, embedding, rerank, etc.)

Health check was unconditionally injecting max_tokens into every
deployment request. OpenAI image endpoints (dall-e-2, dall-e-3,
gpt-image-1) and other non-chat handlers reject max_tokens with a
400 error, causing them to appear permanently unhealthy in the UI.

Guard the max_tokens injection behind a mode check — skip it for
image_generation, video_generation, embedding, rerank, transcription,
and audio_speech modes.

Fixes #26406

litellm/proxy/health_check.py

@@ -366,10 +366,22 @@ def _update_litellm_params_for_health_check(
     - updates the `voice` param with the `health_check_voice` for `audio_speech` mode if it exists Doc: https://docs.litellm.ai/docs/proxy/health#text-to-speech-models
     - for Bedrock models with region routing (bedrock/region/model), strips the litellm routing prefix but preserves the model ID
     """
+    _NON_CHAT_MODES = {
+        "image_generation",
+        "video_generation",
+        "embedding",
+        "rerank",
+        "transcription",
+        "audio_speech",
+    }
+    _mode = model_info.get("mode", None)
     litellm_params["messages"] = _get_random_llm_message()
-    _resolved_max_tokens = _resolve_health_check_max_tokens(model_info, litellm_params)
-    if _resolved_max_tokens is not None:
-        litellm_params["max_tokens"] = _resolved_max_tokens
+    if _mode not in _NON_CHAT_MODES:
+        _resolved_max_tokens = _resolve_health_check_max_tokens(
+            model_info, litellm_params
+        )
+        if _resolved_max_tokens is not None:
+            litellm_params["max_tokens"] = _resolved_max_tokens
 
     _health_check_model = model_info.get("health_check_model", None)
     if _health_check_model is not None:

tests/test_litellm/proxy/test_health_check_max_tokens.py

@@ -86,6 +86,37 @@ async def test_ahealth_check_wildcard_models_respects_max_tokens():
         assert model_params["max_tokens"] == 3
 
 
+@pytest.mark.parametrize(
+    "mode",
+    [
+        "image_generation",
+        "video_generation",
+        "embedding",
+        "rerank",
+        "transcription",
+        "audio_speech",
+    ],
+)
+@pytest.mark.asyncio
+async def test_update_litellm_params_no_max_tokens_for_non_chat_modes(
+    monkeypatch, mode
+):
+    """
+    Non-chat modes (image_generation, embedding, etc.) must not receive max_tokens
+    because those endpoints reject it with a 400 error.
+    """
+    monkeypatch.setattr(hc_module, "BACKGROUND_HEALTH_CHECK_MAX_TOKENS", None)
+    monkeypatch.setattr(hc_module, "BACKGROUND_HEALTH_CHECK_MAX_TOKENS_REASONING", None)
+    model_info = {"mode": mode}
+    litellm_params = {"model": "openai/dall-e-3"}
+
+    updated_params = _update_litellm_params_for_health_check(model_info, litellm_params)
+
+    assert (
+        "max_tokens" not in updated_params
+    ), f"max_tokens should not be set for mode={mode!r}, got {updated_params.get('max_tokens')}"
+
+
 @pytest.mark.asyncio
 async def test_background_health_check_max_tokens_env_var(monkeypatch):
     """