Fix xAI usage and Fireworks parallel tool params

litellm/llms/fireworks_ai/chat/transformation.py

@@ -112,11 +112,11 @@ class FireworksAIConfig(OpenAIGPTConfig):
         # Only add tools for models that support function calling
         if supports_function_calling(model=model, custom_llm_provider="fireworks_ai"):
             supported_params.append("tools")
+            supported_params.append("parallel_tool_calls")
 
         # Only add tool_choice for models that explicitly support it
         if supports_tool_choice(model=model, custom_llm_provider="fireworks_ai"):
             supported_params.append("tool_choice")
-            supported_params.append("parallel_tool_calls")
 
         # Only add reasoning_effort for models that support it
         if supports_reasoning(model=model, custom_llm_provider="fireworks_ai"):

litellm/llms/xai/chat/transformation.py

@@ -262,13 +262,13 @@ class XAIChatConfig(OpenAIGPTConfig):
             prompt_tokens = int(usage.get("prompt_tokens") or 0)
             completion_tokens = int(usage.get("completion_tokens") or 0)
             expected_total = prompt_tokens + completion_tokens
-            if int(usage.get("total_tokens") or 0) != expected_total:
+            if int(usage.get("total_tokens") or 0) < expected_total:
                 usage["total_tokens"] = expected_total
             return
         prompt_tokens = int(usage.prompt_tokens or 0)
         completion_tokens = int(usage.completion_tokens or 0)
         expected_total = prompt_tokens + completion_tokens
-        if int(usage.total_tokens or 0) != expected_total:
+        if int(usage.total_tokens or 0) < expected_total:
             usage.total_tokens = expected_total
 
 

tests/test_litellm/llms/fireworks_ai/chat/test_fireworks_ai_chat_transformation.py

@@ -123,7 +123,7 @@ def test_get_supported_openai_params_reasoning_effort():
 
 
 def test_get_supported_openai_params_parallel_tool_calls():
-    """Test that parallel_tool_calls is included for models that support tool_choice."""
+    """Test that parallel_tool_calls is included for models that support function calling."""
     config = FireworksAIConfig()
 
     supported_params = config.get_supported_openai_params(
@@ -137,6 +137,28 @@ def test_get_supported_openai_params_parallel_tool_calls():
     assert "parallel_tool_calls" not in unsupported_params
 
 
+def test_get_supported_openai_params_parallel_tool_calls_without_tool_choice(
+    monkeypatch,
+):
+    """Test that parallel_tool_calls is gated on tools, not tool_choice."""
+    config = FireworksAIConfig()
+    model = "fireworks_ai/test-tools-without-tool-choice"
+    monkeypatch.setitem(
+        litellm.model_cost,
+        model,
+        {
+            "supports_function_calling": True,
+            "supports_tool_choice": False,
+        },
+    )
+
+    supported_params = config.get_supported_openai_params(model)
+
+    assert "tools" in supported_params
+    assert "parallel_tool_calls" in supported_params
+    assert "tool_choice" not in supported_params
+
+
 def test_get_model_info_respects_explicit_fireworks_capabilities():
     """Test that get_model_info preserves explicit capability flags from the model map."""
     model_info = get_model_info("fireworks_ai/accounts/fireworks/models/glm-5p1")

tests/test_litellm/llms/xai/test_xai_chat_transformation.py

@@ -6,6 +6,7 @@ sys.path.insert(
 )  # Adds the parent directory to the system path
 
 from litellm.llms.xai.chat.transformation import XAIChatConfig
+from litellm.types.utils import Usage
 
 
 class TestXAIParallelToolCalls:
@@ -14,9 +15,7 @@ class TestXAIParallelToolCalls:
     def test_get_supported_openai_params_includes_parallel_tool_calls(self):
         """Test that parallel_tool_calls is in supported parameters."""
         config = XAIChatConfig()
-        supported_params = config.get_supported_openai_params(
-            "xai/grok-4.20"
-        )
+        supported_params = config.get_supported_openai_params("xai/grok-4.20")
         assert "parallel_tool_calls" in supported_params
 
     def test_transform_request_preserves_parallel_tool_calls(self):
@@ -37,3 +36,19 @@ class TestXAIParallelToolCalls:
         assert result.get("parallel_tool_calls") is True
         assert len(result["messages"]) == 1
         assert result["messages"][0]["role"] == "user"
+
+
+class TestXAIUsageNormalization:
+    def test_preserves_reasoning_tokens_in_total_usage(self):
+        usage = Usage(prompt_tokens=100, completion_tokens=50, total_tokens=200)
+
+        XAIChatConfig._normalize_openai_compatible_usage_totals(usage)
+
+        assert usage.total_tokens == 200
+
+    def test_preserves_reasoning_tokens_in_streaming_usage(self):
+        usage = {"prompt_tokens": 100, "completion_tokens": 50, "total_tokens": 200}
+
+        XAIChatConfig._normalize_openai_compatible_usage_totals(usage)
+
+        assert usage["total_tokens"] == 200