Merge PR #32786 (databricks gemini reasoning_effort) into internal_staging, resolving conflicts

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/llms/base_llm/chat/transformation.py

@@ -14,7 +14,6 @@ from typing import (
     Tuple,
     Type,
     Union,
-    cast,
 )
 
 import httpx
@@ -124,7 +123,15 @@ class BaseConfig(ABC):
         if is_thinking_enabled and (
             "max_tokens" not in non_default_params and "max_completion_tokens" not in non_default_params
         ):
-            thinking_token_budget = cast(dict, optional_params["thinking"]).get("budget_tokens", None)
+            # `is_thinking_enabled` is True when `reasoning_effort` is set OR
+            # when `thinking` is set. Providers that pass `reasoning_effort`
+            # through natively (e.g. Databricks-Gemini-3+, Databricks-GPT-5)
+            # never populate an Anthropic-style `thinking` block, so guard
+            # against that case here.
+            thinking = optional_params.get("thinking")
+            if not isinstance(thinking, dict):
+                return
+            thinking_token_budget = thinking.get("budget_tokens", None)
             if thinking_token_budget is not None:
                 optional_params["max_tokens"] = thinking_token_budget + DEFAULT_MAX_TOKENS
 

litellm/llms/databricks/chat/transformation.py

@@ -268,6 +268,24 @@ class DatabricksConfig(DatabricksBase, OpenAILikeChatConfig, AnthropicConfig):
             "thinking",
         ]
 
+    @staticmethod
+    def _databricks_model_uses_anthropic_thinking_param(model: str) -> bool:
+        """
+        Per Databricks docs, Claude and Gemini 2.5 endpoints accept the
+        Anthropic-style `thinking={"type":"enabled","budget_tokens":N}` payload
+        and do NOT accept OpenAI's top-level `reasoning_effort`. Gemini 3+ and
+        GPT-5/GPT-OSS accept `reasoning_effort` natively and need no
+        translation.
+        """
+        from litellm.utils import _supports_factory
+
+        normalized = model.lower().replace(".", "-")
+        return _supports_factory(
+            model=normalized,
+            custom_llm_provider="databricks",
+            key="supports_anthropic_thinking_payload",
+        )
+
     def convert_anthropic_tool_to_databricks_tool(
         self, tool: Optional[AllAnthropicToolsValues]
     ) -> Optional[DatabricksTool]:
@@ -371,7 +389,10 @@ class DatabricksConfig(DatabricksBase, OpenAILikeChatConfig, AnthropicConfig):
                 "response_format", None
             )  # unsupported for claude models - if json_schema -> convert to tool call
 
-        if "reasoning_effort" in non_default_params and "claude" in model:
+        if (
+            "reasoning_effort" in non_default_params
+            and self._databricks_model_uses_anthropic_thinking_param(model)
+        ):
             reasoning_effort_value = non_default_params.get("reasoning_effort")
             mapped_thinking = AnthropicConfig._map_reasoning_effort(
                 reasoning_effort=reasoning_effort_value,
@@ -379,12 +400,14 @@ class DatabricksConfig(DatabricksBase, OpenAILikeChatConfig, AnthropicConfig):
                 custom_llm_provider="databricks",
                 llm_provider="databricks",
             )
+            is_claude = "claude" in model.lower()
             if mapped_thinking is None:
                 optional_params.pop("thinking", None)
                 optional_params.pop("output_config", None)
             else:
                 optional_params["thinking"] = mapped_thinking
-                if AnthropicConfig._is_adaptive_thinking_model(model, "databricks"):
+                # output_config + adaptive thinking is an Anthropic-only feature.
+                if is_claude and AnthropicConfig._is_adaptive_thinking_model(model, "databricks"):
                     mapped_effort: Optional[str] = None
                     if isinstance(reasoning_effort_value, str):
                         mapped_effort = REASONING_EFFORT_TO_OUTPUT_CONFIG_EFFORT.get(reasoning_effort_value)

litellm/model_prices_and_context_window_backup.json

@@ -13339,6 +13339,7 @@
         "supports_assistant_prefill": true,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_anthropic_thinking_payload": true,
         "supports_tool_choice": true
     },
     "databricks/databricks-claude-haiku-4-5": {
@@ -13358,6 +13359,7 @@
         "supports_assistant_prefill": true,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_anthropic_thinking_payload": true,
         "supports_tool_choice": true
     },
     "databricks/databricks-claude-opus-4": {
@@ -13377,6 +13379,7 @@
         "supports_assistant_prefill": true,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_anthropic_thinking_payload": true,
         "supports_tool_choice": true
     },
     "databricks/databricks-claude-opus-4-1": {
@@ -13396,6 +13399,7 @@
         "supports_assistant_prefill": true,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_anthropic_thinking_payload": true,
         "supports_tool_choice": true
     },
     "databricks/databricks-claude-opus-4-5": {
@@ -13415,6 +13419,7 @@
         "supports_assistant_prefill": true,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_anthropic_thinking_payload": true,
         "supports_tool_choice": true,
         "supports_output_config": true
     },
@@ -13435,6 +13440,7 @@
         "supports_assistant_prefill": true,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_anthropic_thinking_payload": true,
         "supports_tool_choice": true
     },
     "databricks/databricks-claude-sonnet-4-1": {
@@ -13454,6 +13460,7 @@
         "supports_assistant_prefill": true,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_anthropic_thinking_payload": true,
         "supports_tool_choice": true
     },
     "databricks/databricks-claude-sonnet-4-5": {
@@ -13473,6 +13480,7 @@
         "supports_assistant_prefill": true,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_anthropic_thinking_payload": true,
         "supports_tool_choice": true
     },
     "databricks/databricks-gemini-2-5-flash": {
@@ -13490,6 +13498,7 @@
         "output_dbu_cost_per_token": 3.5714e-05,
         "source": "https://www.databricks.com/product/pricing/proprietary-foundation-model-serving",
         "supports_function_calling": true,
+        "supports_anthropic_thinking_payload": true,
         "supports_tool_choice": true
     },
     "databricks/databricks-gemini-2-5-pro": {
@@ -13507,6 +13516,7 @@
         "output_dbu_cost_per_token": 0.000142857,
         "source": "https://www.databricks.com/product/pricing/proprietary-foundation-model-serving",
         "supports_function_calling": true,
+        "supports_anthropic_thinking_payload": true,
         "supports_tool_choice": true
     },
     "databricks/databricks-gemma-3-12b": {

litellm/types/utils.py

@@ -152,6 +152,7 @@ class ProviderSpecificModelInfo(TypedDict, total=False):
     supports_max_reasoning_effort: Optional[bool]
     supports_output_config: Optional[bool]
     supports_image_size: Optional[bool]
+    supports_anthropic_thinking_payload: Optional[bool]
     bedrock_output_config_effort_ceiling: Optional[Literal["low", "medium", "high", "max", "xhigh"]]
     bedrock_converse_supports_strict_tools: Optional[bool]
 

litellm/utils.py

@@ -5513,6 +5513,7 @@ def _get_model_info_helper(
                 supports_reasoning=_model_info.get("supports_reasoning", None),
                 supports_adaptive_thinking=_model_info.get("supports_adaptive_thinking", None),
                 supports_mid_conversation_system=_model_info.get("supports_mid_conversation_system", None),
+                supports_anthropic_thinking_payload=_model_info.get("supports_anthropic_thinking_payload", None),
                 supports_none_reasoning_effort=_model_info.get("supports_none_reasoning_effort", None),
                 supports_minimal_reasoning_effort=_model_info.get("supports_minimal_reasoning_effort", None),
                 supports_low_reasoning_effort=_model_info.get("supports_low_reasoning_effort", None),

model_prices_and_context_window.json

@@ -13339,6 +13339,7 @@
         "supports_assistant_prefill": true,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_anthropic_thinking_payload": true,
         "supports_tool_choice": true
     },
     "databricks/databricks-claude-haiku-4-5": {
@@ -13358,6 +13359,7 @@
         "supports_assistant_prefill": true,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_anthropic_thinking_payload": true,
         "supports_tool_choice": true
     },
     "databricks/databricks-claude-opus-4": {
@@ -13377,6 +13379,7 @@
         "supports_assistant_prefill": true,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_anthropic_thinking_payload": true,
         "supports_tool_choice": true
     },
     "databricks/databricks-claude-opus-4-1": {
@@ -13396,6 +13399,7 @@
         "supports_assistant_prefill": true,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_anthropic_thinking_payload": true,
         "supports_tool_choice": true
     },
     "databricks/databricks-claude-opus-4-5": {
@@ -13415,6 +13419,7 @@
         "supports_assistant_prefill": true,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_anthropic_thinking_payload": true,
         "supports_tool_choice": true,
         "supports_output_config": true
     },
@@ -13435,6 +13440,7 @@
         "supports_assistant_prefill": true,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_anthropic_thinking_payload": true,
         "supports_tool_choice": true
     },
     "databricks/databricks-claude-sonnet-4-1": {
@@ -13454,6 +13460,7 @@
         "supports_assistant_prefill": true,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_anthropic_thinking_payload": true,
         "supports_tool_choice": true
     },
     "databricks/databricks-claude-sonnet-4-5": {
@@ -13473,6 +13480,7 @@
         "supports_assistant_prefill": true,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_anthropic_thinking_payload": true,
         "supports_tool_choice": true
     },
     "databricks/databricks-gemini-2-5-flash": {
@@ -13490,6 +13498,7 @@
         "output_dbu_cost_per_token": 3.5714e-05,
         "source": "https://www.databricks.com/product/pricing/proprietary-foundation-model-serving",
         "supports_function_calling": true,
+        "supports_anthropic_thinking_payload": true,
         "supports_tool_choice": true
     },
     "databricks/databricks-gemini-2-5-pro": {
@@ -13507,6 +13516,7 @@
         "output_dbu_cost_per_token": 0.000142857,
         "source": "https://www.databricks.com/product/pricing/proprietary-foundation-model-serving",
         "supports_function_calling": true,
+        "supports_anthropic_thinking_payload": true,
         "supports_tool_choice": true
     },
     "databricks/databricks-gemma-3-12b": {

tests/test_litellm/llms/databricks/chat/test_databricks_chat_transformation.py

@@ -8,6 +8,12 @@ from fastapi.testclient import TestClient
 sys.path.insert(0, os.path.abspath("../../../../.."))  # Adds the parent directory to the system path
 from unittest.mock import MagicMock, patch
 
+import litellm
+from litellm.constants import (
+    DEFAULT_REASONING_EFFORT_HIGH_THINKING_BUDGET,
+    DEFAULT_REASONING_EFFORT_LOW_THINKING_BUDGET,
+    DEFAULT_REASONING_EFFORT_MEDIUM_THINKING_BUDGET,
+)
 from litellm.llms.databricks.chat.transformation import (
     DatabricksChatResponseIterator,
     DatabricksConfig,
@@ -15,6 +21,12 @@ from litellm.llms.databricks.chat.transformation import (
 )
 
 
+@pytest.fixture(autouse=True)
+def _use_local_model_cost_map(monkeypatch):
+    monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True")
+    monkeypatch.setattr(litellm, "model_cost", litellm.get_model_cost_map(url=""))
+
+
 def test_transform_choices():
     config = DatabricksConfig()
     databricks_choices = [
@@ -423,3 +435,126 @@ def test_databricks_config_probes_capabilities_under_databricks_namespace():
     without this override they probed the ``anthropic`` cost-map namespace and
     ignored the exact ``databricks/databricks-claude-*`` entries."""
     assert DatabricksConfig().custom_llm_provider == "databricks"
+
+
+# ---------------------------------------------------------------------------
+# reasoning_effort translation
+#
+# Databricks foundation-model endpoints take reasoning controls via different
+# payload shapes depending on the underlying model family:
+#
+#   Claude:        Anthropic-style `thinking={"type":"enabled","budget_tokens":N}`
+#   Gemini 2.5:    Same Anthropic-style `thinking` payload as Claude
+#                  (per docs.databricks.com/.../query-reason-models)
+#   Gemini 3+:     Native OpenAI-style top-level `reasoning_effort`
+#   GPT-5/GPT-OSS: Native OpenAI-style top-level `reasoning_effort`
+#
+# LiteLLM should translate `reasoning_effort` into the right shape for the
+# first two families and pass it through unchanged for the latter two.
+# ---------------------------------------------------------------------------
+
+
+def _map_reasoning_effort(model: str, reasoning_effort, **extra_non_default):
+    """Run map_openai_params with reasoning_effort + optional extras."""
+    non_default = {"reasoning_effort": reasoning_effort}
+    non_default.update(extra_non_default)
+    return DatabricksConfig().map_openai_params(
+        non_default_params=non_default,
+        optional_params={},
+        model=model,
+        drop_params=False,
+    )
+
+
+def test_claude_translates_reasoning_effort_to_thinking():
+    """Regression: Claude path must still translate to Anthropic-style thinking."""
+    params = _map_reasoning_effort("databricks-claude-3-7-sonnet", "low")
+    assert params.get("thinking") == {
+        "type": "enabled",
+        "budget_tokens": DEFAULT_REASONING_EFFORT_LOW_THINKING_BUDGET,
+    }
+    assert "reasoning_effort" not in params
+
+
+def test_gemini_2_5_low_translates_to_thinking_budget():
+    params = _map_reasoning_effort("databricks-gemini-2-5-flash", "low")
+    assert params.get("thinking") == {
+        "type": "enabled",
+        "budget_tokens": DEFAULT_REASONING_EFFORT_LOW_THINKING_BUDGET,
+    }
+    assert "reasoning_effort" not in params
+
+
+def test_gemini_2_5_medium_translates_to_thinking_budget():
+    params = _map_reasoning_effort("databricks-gemini-2-5-flash", "medium")
+    assert params.get("thinking") == {
+        "type": "enabled",
+        "budget_tokens": DEFAULT_REASONING_EFFORT_MEDIUM_THINKING_BUDGET,
+    }
+    assert "reasoning_effort" not in params
+
+
+def test_gemini_2_5_high_translates_to_thinking_budget():
+    params = _map_reasoning_effort("databricks-gemini-2-5-flash", "high")
+    assert params.get("thinking") == {
+        "type": "enabled",
+        "budget_tokens": DEFAULT_REASONING_EFFORT_HIGH_THINKING_BUDGET,
+    }
+    assert "reasoning_effort" not in params
+
+
+def test_gemini_2_5_pro_translates_to_thinking_budget():
+    """Cover the gemini-2-5-pro endpoint too, not just flash."""
+    params = _map_reasoning_effort("databricks-gemini-2-5-pro", "high")
+    assert params.get("thinking") == {
+        "type": "enabled",
+        "budget_tokens": DEFAULT_REASONING_EFFORT_HIGH_THINKING_BUDGET,
+    }
+    assert "reasoning_effort" not in params
+
+
+def test_gemini_2_5_with_dot_notation_translates():
+    """A user passing the upstream Google-style `gemini-2.5-...` form should
+    still trigger the Anthropic-thinking translation, not pass through."""
+    params = _map_reasoning_effort("databricks-gemini-2.5-flash", "low")
+    assert params.get("thinking") == {
+        "type": "enabled",
+        "budget_tokens": DEFAULT_REASONING_EFFORT_LOW_THINKING_BUDGET,
+    }
+    assert "reasoning_effort" not in params
+
+
+def test_gemini_2_0_does_not_match():
+    """Guard against over-matching: `gemini-2-0` (hypothetical or future) is
+    NOT a Gemini 2.5 endpoint and must not get the thinking translation."""
+    params = _map_reasoning_effort("databricks-gemini-2-0-flash", "low")
+    assert "thinking" not in params
+    assert params.get("reasoning_effort") == "low"
+
+
+def test_gemini_2_5_none_drops_thinking_and_reasoning_effort():
+    """`reasoning_effort='none'` mirrors the Claude behavior: no thinking emitted."""
+    params = _map_reasoning_effort("databricks-gemini-2-5-flash", "none")
+    assert "thinking" not in params
+    assert "reasoning_effort" not in params
+
+
+def test_gemini_3_passes_reasoning_effort_through():
+    """Databricks-Gemini-3+ accepts reasoning_effort natively — do not translate."""
+    params = _map_reasoning_effort("databricks-gemini-3-1-pro", "low")
+    assert params.get("reasoning_effort") == "low"
+    assert "thinking" not in params
+
+
+def test_gpt_5_passes_reasoning_effort_through():
+    """Databricks-GPT-5 family accepts reasoning_effort natively."""
+    params = _map_reasoning_effort("databricks-gpt-5-1", "low")
+    assert params.get("reasoning_effort") == "low"
+    assert "thinking" not in params
+
+
+def test_gpt_oss_passes_reasoning_effort_through():
+    """Databricks-GPT-OSS accepts reasoning_effort natively."""
+    params = _map_reasoning_effort("databricks-gpt-oss-120b", "high")
+    assert params.get("reasoning_effort") == "high"
+    assert "thinking" not in params