From aec84020e5a2882e6ad6515354e55e25ccd8b6fa Mon Sep 17 00:00:00 2001
From: Graham Neubig <398875+neubig@users.noreply.github.com>
Date: Thu, 21 May 2026 08:54:53 -0400
Subject: [PATCH] Normalize Lemonade runtime model metadata

---
 litellm/llms/lemonade/chat/transformation.py  | 24 +++++++++++++--
 .../llms/lemonade/test_lemonade.py            | 30 +++++++++++++++++++
 2 files changed, 51 insertions(+), 3 deletions(-)

diff --git a/litellm/llms/lemonade/chat/transformation.py b/litellm/llms/lemonade/chat/transformation.py
index ff8069395b0..1576dbd0a16 100644
--- a/litellm/llms/lemonade/chat/transformation.py
+++ b/litellm/llms/lemonade/chat/transformation.py
@@ -121,16 +121,31 @@ class LemonadeChatConfig(OpenAILikeChatConfig):
                 return parsed
         return None
 
+    @staticmethod
+    def _get_provider_specific_entry(model_info: dict) -> dict:
+        provider_specific_entry = model_info.get("provider_specific_entry")
+        if not isinstance(provider_specific_entry, dict):
+            provider_specific_entry = {}
+        else:
+            provider_specific_entry = provider_specific_entry.copy()
+
+        for key in ("recipe_options", "context_window", "max_context_window"):
+            if key in model_info:
+                provider_specific_entry[key] = model_info[key]
+
+        return provider_specific_entry
+
     def _get_context_window(self, model_info: dict) -> Optional[int]:
-        recipe_options = model_info.get("recipe_options")
+        provider_specific_entry = self._get_provider_specific_entry(model_info)
+        recipe_options = provider_specific_entry.get("recipe_options")
         if not isinstance(recipe_options, dict):
             recipe_options = {}
 
         for value in (
             recipe_options.get("ctx_size"),
             model_info.get("max_input_tokens"),
-            model_info.get("context_window"),
-            model_info.get("max_context_window"),
+            provider_specific_entry.get("context_window"),
+            provider_specific_entry.get("max_context_window"),
         ):
             parsed = self._get_positive_int(value)
             if parsed is not None:
@@ -172,6 +187,7 @@ class LemonadeChatConfig(OpenAILikeChatConfig):
         max_input_tokens = self._get_context_window(model_info)
         max_output_tokens = self._get_positive_int(model_info.get("max_output_tokens"))
         max_tokens = self._get_positive_int(model_info.get("max_tokens"))
+        provider_specific_entry = self._get_provider_specific_entry(model_info)
 
         model_info_response = self._get_default_model_info(model)
         model_info_response.update(
@@ -181,6 +197,8 @@ class LemonadeChatConfig(OpenAILikeChatConfig):
                 "max_output_tokens": max_output_tokens,
             }
         )
+        if provider_specific_entry:
+            model_info_response["provider_specific_entry"] = provider_specific_entry
         return model_info_response
 
     def _get_openai_compatible_provider_info(
diff --git a/tests/test_litellm/llms/lemonade/test_lemonade.py b/tests/test_litellm/llms/lemonade/test_lemonade.py
index 491e4aac18d..d6aae6e53f9 100644
--- a/tests/test_litellm/llms/lemonade/test_lemonade.py
+++ b/tests/test_litellm/llms/lemonade/test_lemonade.py
@@ -109,6 +109,10 @@ def test_get_model_info_uses_loaded_context_size():
     assert model_info["key"] == "lemonade/Qwen3.6-35B-A3B-GGUF"
     assert model_info["litellm_provider"] == "lemonade"
     assert model_info["max_input_tokens"] == 65536
+    assert model_info["provider_specific_entry"] == {
+        "recipe_options": {"ctx_size": 65536},
+        "max_context_window": 262144,
+    }
     assert "supports_function_calling" not in model_info
     assert "supports_response_schema" not in model_info
     assert "supports_tool_choice" not in model_info
@@ -140,6 +144,32 @@ def test_get_model_info_falls_back_when_server_unavailable():
     assert "supports_tool_choice" not in model_info
 
 
+def test_get_model_info_reads_context_from_provider_specific_entry():
+    """Test that Lemonade model info uses provider-specific runtime metadata."""
+    config = LemonadeChatConfig()
+    response = MagicMock()
+    response.status_code = 200
+    response.json.return_value = {
+        "id": "Qwen3.6-35B-A3B-GGUF",
+        "provider_specific_entry": {
+            "recipe_options": {"ctx_size": "32768"},
+            "max_context_window": 262144,
+        },
+    }
+
+    with patch.object(litellm.module_level_client, "get", return_value=response):
+        model_info = config.get_model_info(
+            model="lemonade/Qwen3.6-35B-A3B-GGUF",
+            api_base="http://lemonade.test/v1",
+        )
+
+    assert model_info["max_input_tokens"] == 32768
+    assert model_info["provider_specific_entry"] == {
+        "recipe_options": {"ctx_size": "32768"},
+        "max_context_window": 262144,
+    }
+
+
 def test_get_model_info_sends_lemonade_api_key(monkeypatch):
     """Test that Lemonade model info uses auth for authenticated servers."""
     monkeypatch.setenv("LEMONADE_API_KEY", "test-key")