fix(health_check): skip max_tokens for image_generation mode

`_update_litellm_params_for_health_check` injected `max_tokens` for
every deployment. OpenAI `/v1/images/generations` strictly rejects
unknown fields, so health checks for dall-e-* and gpt-image-1 always
failed with `400 "Unknown parameter: 'max_tokens'"` even though the
actual image endpoint calls succeed. Skip the `max_tokens` injection
when `model_info.mode == "image_generation"`. `messages` still gets
injected (downstream `_filter_model_params` already strips it for
non-chat handlers).

litellm/proxy/health_check.py

@@ -362,14 +362,21 @@ def _update_litellm_params_for_health_check(
     Update the litellm params for health check.
 
     - gets a short `messages` param for health check
+    - adds a bounded `max_tokens`, except when
+      `model_info.mode == "image_generation"` — OpenAI
+      `/v1/images/generations` rejects unknown fields with
+      400 "Unknown parameter: 'max_tokens'".
     - updates the `model` param with the `health_check_model` if it exists Doc: https://docs.litellm.ai/docs/proxy/health#wildcard-routes
     - 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
     """
     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 model_info.get("mode", None) != "image_generation":
+        _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

@@ -225,3 +225,59 @@ def test_wildcard_ignores_reasoning_split_model_info(monkeypatch):
     litellm_params = {"model": "openai/*"}
 
     assert _resolve_health_check_max_tokens(model_info, litellm_params) is None
+
+
+# ---------------------------------------------------------------------------
+# image_generation must not receive max_tokens.
+#
+# _update_litellm_params_for_health_check injected `max_tokens` for every
+# deployment. For `mode: image_generation` that leaked into OpenAI
+# `/v1/images/generations`, which strictly rejects unknown fields with
+# `400 "Unknown parameter: 'max_tokens'"`, marking dall-e-* and
+# gpt-image-1 as permanently unhealthy even though their actual image
+# calls succeed. `messages` still gets injected (downstream
+# `_filter_model_params` already strips it for non-chat handlers).
+# ---------------------------------------------------------------------------
+
+
+def test_image_generation_mode_skips_max_tokens():
+    """image_generation must not receive max_tokens."""
+    model_info = {"mode": "image_generation"}
+    litellm_params = {"model": "openai/dall-e-3", "api_key": "sk-test"}
+
+    updated = _update_litellm_params_for_health_check(model_info, litellm_params)
+
+    assert "max_tokens" not in updated
+    # connection-level params must still pass through unchanged
+    assert updated["api_key"] == "sk-test"
+
+
+def test_image_generation_ignores_explicit_health_check_max_tokens():
+    """Even an explicit `health_check_max_tokens` override must not be
+    forwarded to image endpoints — the field is invalid there."""
+    model_info = {"mode": "image_generation", "health_check_max_tokens": 50}
+    litellm_params = {"model": "openai/dall-e-3"}
+
+    updated = _update_litellm_params_for_health_check(model_info, litellm_params)
+
+    assert "max_tokens" not in updated
+
+
+def test_chat_mode_still_injects_max_tokens():
+    """Regression guard: the chat-style probe payload is unchanged."""
+    model_info = {"mode": "chat"}
+    litellm_params = {"model": "gpt-4"}
+
+    updated = _update_litellm_params_for_health_check(model_info, litellm_params)
+
+    assert updated["max_tokens"] == 5
+
+
+def test_no_mode_still_injects_max_tokens():
+    """Regression guard: model_info without `mode` keeps the legacy path."""
+    model_info: dict = {}
+    litellm_params = {"model": "gpt-4"}
+
+    updated = _update_litellm_params_for_health_check(model_info, litellm_params)
+
+    assert updated["max_tokens"] == 5