Add Prompt caching and reasoning support for MiniMax, GLM, Xiaomi

litellm/llms/minimax/chat/transformation.py

@@ -1,11 +1,12 @@
 """
 MiniMax OpenAI transformation config - extends OpenAI chat config for MiniMax's OpenAI-compatible API
 """
-from typing import Optional
+from typing import List, Optional, Tuple
 
 import litellm
 from litellm.llms.openai.chat.gpt_transformation import OpenAIGPTConfig
 from litellm.secret_managers.main import get_secret_str
+from litellm.types.llms.openai import AllMessageValues, ChatCompletionToolParam
 
 
 class MinimaxChatConfig(OpenAIGPTConfig):
@@ -73,11 +74,33 @@ class MinimaxChatConfig(OpenAIGPTConfig):
         else:
             return f"{base_url}/v1/chat/completions"
 
+    def remove_cache_control_flag_from_messages_and_tools(
+        self,
+        model: str,
+        messages: List[AllMessageValues],
+        tools: Optional[List[ChatCompletionToolParam]] = None,
+    ) -> Tuple[List[AllMessageValues], Optional[List[ChatCompletionToolParam]]]:
+        """
+        Override to preserve cache_control for MiniMax.
+        MiniMax supports cache_control - don't strip it.
+        """
+        # MiniMax supports cache_control, so return messages and tools unchanged
+        return messages, tools
+
     def get_supported_openai_params(self, model: str) -> list:
         """
         Get supported OpenAI parameters for MiniMax.
-        Adds reasoning_split to the list of supported params.
+        Adds reasoning_split and thinking to the list of supported params.
         """
         base_params = super().get_supported_openai_params(model=model)
-        return base_params + ["reasoning_split"]
+        additional_params = ["reasoning_split"]
+        
+        # Add thinking parameter if model supports reasoning
+        try:
+            if litellm.supports_reasoning(model=model, custom_llm_provider="minimax"):
+                additional_params.append("thinking")
+        except Exception:
+            pass
+        
+        return base_params + additional_params
 

litellm/llms/openrouter/chat/transformation.py

@@ -26,6 +26,9 @@ class CacheControlSupportedModels(str, Enum):
     """Models that support cache_control in content blocks."""
     CLAUDE = "claude"
     GEMINI = "gemini"
+    MINIMAX = "minimax"
+    GLM = "glm"
+    ZAI = "z-ai"
 
 
 class OpenrouterConfig(OpenAIGPTConfig):
@@ -39,6 +42,7 @@ class OpenrouterConfig(OpenAIGPTConfig):
                 model=model, custom_llm_provider="openrouter"
             ) or litellm.supports_reasoning(model=model):
                 supported_params.append("reasoning_effort")
+                supported_params.append("thinking")
         except Exception:
             pass
         return list(dict.fromkeys(supported_params))

litellm/llms/zai/chat/transformation.py

@@ -1,6 +1,7 @@
-from typing import Optional, Tuple
+from typing import List, Optional, Tuple
 
 from litellm.secret_managers.main import get_secret_str
+from litellm.types.llms.openai import AllMessageValues, ChatCompletionToolParam
 
 from ...openai.chat.gpt_transformation import OpenAIGPTConfig
 
@@ -19,6 +20,19 @@ class ZAIChatConfig(OpenAIGPTConfig):
         dynamic_api_key = api_key or get_secret_str("ZAI_API_KEY")
         return api_base, dynamic_api_key
 
+    def remove_cache_control_flag_from_messages_and_tools(
+        self,
+        model: str,
+        messages: List[AllMessageValues],
+        tools: Optional[List[ChatCompletionToolParam]] = None,
+    ) -> Tuple[List[AllMessageValues], Optional[List[ChatCompletionToolParam]]]:
+        """
+        Override to preserve cache_control for GLM/ZAI.
+        GLM supports cache_control - don't strip it.
+        """
+        # GLM/ZAI supports cache_control, so return messages and tools unchanged
+        return messages, tools
+
     def get_supported_openai_params(self, model: str) -> list:
         base_params = [
             "max_tokens",

model_prices_and_context_window.json

@@ -20628,6 +20628,7 @@
         "supports_function_calling": true,
         "supports_tool_choice": true,
         "supports_prompt_caching": true,
+        "supports_reasoning": true,
         "supports_system_messages": true,
         "max_input_tokens": 1000000,
         "max_output_tokens": 8192
@@ -20642,6 +20643,7 @@
         "supports_function_calling": true,
         "supports_tool_choice": true,
         "supports_prompt_caching": true,
+        "supports_reasoning": true,
         "supports_system_messages": true,
         "max_input_tokens": 1000000,
         "max_output_tokens": 8192
@@ -20656,6 +20658,7 @@
         "supports_function_calling": true,
         "supports_tool_choice": true,
         "supports_prompt_caching": true,
+        "supports_reasoning": true,
         "supports_system_messages": true,
         "max_input_tokens": 200000,
         "max_output_tokens": 8192
@@ -23540,7 +23543,7 @@
         "mode": "chat",
         "output_cost_per_token": 1.02e-06,
         "supports_function_calling": true,
-        "supports_prompt_caching": false,
+        "supports_prompt_caching": true,
         "supports_reasoning": true,
         "supports_tool_choice": true
     },
@@ -24232,6 +24235,7 @@
         "output_cost_per_token": 1.75e-06,
         "source": "https://openrouter.ai/z-ai/glm-4.6",
         "supports_function_calling": true,
+        "supports_prompt_caching": true,
         "supports_reasoning": true,
         "supports_tool_choice": true
     },
@@ -24245,6 +24249,7 @@
         "output_cost_per_token": 1.9e-06,
         "source": "https://openrouter.ai/z-ai/glm-4.6:exacto",
         "supports_function_calling": true,
+        "supports_prompt_caching": true,
         "supports_reasoning": true,
         "supports_tool_choice": true
     },
@@ -30872,11 +30877,14 @@
         "max_output_tokens": 128000,
         "mode": "chat",
         "supports_function_calling": true,
+        "supports_prompt_caching": true,
         "supports_reasoning": true,
         "supports_tool_choice": true,
         "source": "https://docs.z.ai/guides/overview/pricing"
     },
     "zai/glm-4.6": {
+        "cache_creation_input_token_cost": 0,
+        "cache_read_input_token_cost": 1.1e-07,
         "input_cost_per_token": 6e-07,
         "output_cost_per_token": 2.2e-06,
         "litellm_provider": "zai",
@@ -30884,6 +30892,8 @@
         "max_output_tokens": 128000,
         "mode": "chat",
         "supports_function_calling": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
         "supports_tool_choice": true,
         "source": "https://docs.z.ai/guides/overview/pricing"
     },

tests/test_litellm/llms/test_cache_control_and_reasoning.py

@@ -0,0 +1,281 @@
+"""
+Test cache_control and reasoning parameter support for MiniMax, GLM/ZAI, and OpenRouter.
+
+This test file verifies the fixes for Issue #19923:
+- cache_control is preserved (not stripped) for MiniMax, GLM, and OpenRouter variants
+- thinking parameter is supported for reasoning-capable models
+- Model metadata correctly reflects capabilities
+"""
+import os
+import sys
+
+import pytest
+
+sys.path.insert(
+    0, os.path.abspath("../../..")
+)  # Adds the parent directory to the system path
+
+from litellm.llms.minimax.chat.transformation import MinimaxChatConfig
+from litellm.llms.openrouter.chat.transformation import OpenrouterConfig
+from litellm.llms.zai.chat.transformation import ZAIChatConfig
+
+
+def test_minimax_preserves_cache_control_in_messages():
+    """MiniMax should NOT strip cache_control from messages."""
+    config = MinimaxChatConfig()
+
+    messages = [
+        {
+            "role": "system",
+            "content": "You are a helpful assistant.",
+            "cache_control": {"type": "ephemeral"},
+        },
+        {
+            "role": "user",
+            "content": "Hello, world!",
+        },
+    ]
+
+    transformed_messages, _ = config.remove_cache_control_flag_from_messages_and_tools(
+        model="minimax/MiniMax-M2.1", messages=messages
+    )
+
+    # cache_control should be preserved
+    assert transformed_messages[0].get("cache_control") == {"type": "ephemeral"}
+
+
+def test_minimax_preserves_cache_control_in_tools():
+    """MiniMax should NOT strip cache_control from tools."""
+    config = MinimaxChatConfig()
+
+    tools = [
+        {
+            "type": "function",
+            "function": {
+                "name": "get_weather",
+                "description": "Get weather information",
+                "parameters": {"type": "object", "properties": {}},
+            },
+            "cache_control": {"type": "ephemeral"},
+        }
+    ]
+
+    _, transformed_tools = config.remove_cache_control_flag_from_messages_and_tools(
+        model="minimax/MiniMax-M2.1", messages=[], tools=tools
+    )
+
+    # cache_control should be preserved
+    assert transformed_tools[0].get("cache_control") == {"type": "ephemeral"}
+
+
+def test_minimax_supports_thinking_param():
+    """MiniMax reasoning models should support thinking parameter."""
+    config = MinimaxChatConfig()
+
+    supported_params = config.get_supported_openai_params(
+        model="minimax/MiniMax-M2.1"
+    )
+
+    # thinking should be in supported params for reasoning models
+    assert "thinking" in supported_params
+    # reasoning_split should also be supported
+    assert "reasoning_split" in supported_params
+
+
+def test_zai_preserves_cache_control_in_messages():
+    """ZAI should NOT strip cache_control from messages."""
+    config = ZAIChatConfig()
+
+    messages = [
+        {
+            "role": "system",
+            "content": "You are a helpful assistant.",
+            "cache_control": {"type": "ephemeral"},
+        },
+        {
+            "role": "user",
+            "content": "Hello, world!",
+        },
+    ]
+
+    transformed_messages, _ = config.remove_cache_control_flag_from_messages_and_tools(
+        model="zai/glm-4.7", messages=messages
+    )
+
+    # cache_control should be preserved
+    assert transformed_messages[0].get("cache_control") == {"type": "ephemeral"}
+
+
+def test_zai_preserves_cache_control_in_tools():
+    """ZAI should NOT strip cache_control from tools."""
+    config = ZAIChatConfig()
+
+    tools = [
+        {
+            "type": "function",
+            "function": {
+                "name": "get_weather",
+                "description": "Get weather information",
+                "parameters": {"type": "object", "properties": {}},
+            },
+            "cache_control": {"type": "ephemeral"},
+        }
+    ]
+
+    _, transformed_tools = config.remove_cache_control_flag_from_messages_and_tools(
+        model="zai/glm-4.7", messages=[], tools=tools
+    )
+
+    # cache_control should be preserved
+    assert transformed_tools[0].get("cache_control") == {"type": "ephemeral"}
+
+
+def test_zai_supports_thinking_param_for_reasoning_models():
+    """ZAI reasoning models (glm-4.7, glm-4.6) should support thinking parameter."""
+    config = ZAIChatConfig()
+
+    # glm-4.7 supports reasoning
+    supported_params_47 = config.get_supported_openai_params(model="zai/glm-4.7")
+    assert "thinking" in supported_params_47
+
+    # glm-4.6 supports reasoning
+    supported_params_46 = config.get_supported_openai_params(model="zai/glm-4.6")
+    assert "thinking" in supported_params_46
+
+
+def test_openrouter_minimax_supports_cache_control():
+    """OpenRouter should preserve cache_control for MiniMax models."""
+    config = OpenrouterConfig()
+
+    messages = [
+        {
+            "role": "user",
+            "content": "Hello, world!",
+            "cache_control": {"type": "ephemeral"},
+        }
+    ]
+
+    # Test that cache_control is not removed
+    transformed_messages, _ = config.remove_cache_control_flag_from_messages_and_tools(
+        model="openrouter/minimax/minimax-m2", messages=messages
+    )
+
+    # The method should preserve cache_control for minimax models
+    assert transformed_messages[0].get("cache_control") == {"type": "ephemeral"}
+
+
+def test_openrouter_glm_supports_cache_control():
+    """OpenRouter should preserve cache_control for GLM models."""
+    config = OpenrouterConfig()
+
+    messages = [
+        {
+            "role": "user",
+            "content": "Hello, world!",
+            "cache_control": {"type": "ephemeral"},
+        }
+    ]
+
+    # Test that cache_control is not removed for GLM models
+    transformed_messages, _ = config.remove_cache_control_flag_from_messages_and_tools(
+        model="openrouter/z-ai/glm-4.6", messages=messages
+    )
+
+    # The method should preserve cache_control for GLM models
+    assert transformed_messages[0].get("cache_control") == {"type": "ephemeral"}
+
+
+def test_openrouter_deepseek_strips_cache_control():
+    """OpenRouter should still strip cache_control for non-supported models."""
+    config = OpenrouterConfig()
+
+    messages = [
+        {
+            "role": "user",
+            "content": "Hello, world!",
+            "cache_control": {"type": "ephemeral"},
+        }
+    ]
+
+    # DeepSeek doesn't support cache_control, so it should be stripped
+    transformed_messages, _ = config.remove_cache_control_flag_from_messages_and_tools(
+        model="openrouter/deepseek/deepseek-chat", messages=messages
+    )
+
+    # cache_control should be removed for non-supported models
+    assert transformed_messages[0].get("cache_control") is None
+
+
+def test_openrouter_minimax_transform_moves_cache_control_to_content():
+    """OpenRouter should move cache_control to content blocks for MiniMax."""
+    config = OpenrouterConfig()
+
+    messages = [
+        {
+            "role": "user",
+            "content": "Analyze this data",
+            "cache_control": {"type": "ephemeral"},
+        }
+    ]
+
+    transformed_request = config.transform_request(
+        model="openrouter/minimax/minimax-m2",
+        messages=messages,
+        optional_params={},
+        litellm_params={},
+        headers={},
+    )
+
+    # cache_control should be moved to content blocks
+    assert "messages" in transformed_request
+    user_message = transformed_request["messages"][0]
+    assert isinstance(user_message["content"], list)
+    assert user_message["content"][0]["cache_control"] == {"type": "ephemeral"}
+    # Message-level cache_control should be removed
+    assert "cache_control" not in user_message
+
+
+def test_openrouter_glm_transform_moves_cache_control_to_content():
+    """OpenRouter should move cache_control to content blocks for GLM."""
+    config = OpenrouterConfig()
+
+    messages = [
+        {
+            "role": "user",
+            "content": "Analyze this data",
+            "cache_control": {"type": "ephemeral"},
+        }
+    ]
+
+    transformed_request = config.transform_request(
+        model="openrouter/z-ai/glm-4.6",
+        messages=messages,
+        optional_params={},
+        litellm_params={},
+        headers={},
+    )
+
+    # cache_control should be moved to content blocks
+    assert "messages" in transformed_request
+    user_message = transformed_request["messages"][0]
+    assert isinstance(user_message["content"], list)
+    assert user_message["content"][0]["cache_control"] == {"type": "ephemeral"}
+
+
+def test_openrouter_supports_thinking_param_for_reasoning_models():
+    """OpenRouter should support thinking parameter for reasoning-capable models."""
+    config = OpenrouterConfig()
+
+    # Test MiniMax (supports reasoning)
+    supported_params_minimax = config.get_supported_openai_params(
+        model="openrouter/minimax/minimax-m2"
+    )
+    assert "thinking" in supported_params_minimax
+    assert "reasoning_effort" in supported_params_minimax
+
+    # Test GLM (supports reasoning)
+    supported_params_glm = config.get_supported_openai_params(
+        model="openrouter/z-ai/glm-4.6"
+    )
+    assert "thinking" in supported_params_glm
+    assert "reasoning_effort" in supported_params_glm