From 443bee300294796081a201bc1da57833ed77a64b Mon Sep 17 00:00:00 2001 From: cohml <62400541+cohml@users.noreply.github.com> Date: Tue, 23 Jun 2026 11:04:30 -0400 Subject: [PATCH 1/3] fix(databricks): translate reasoning_effort to thinking for Gemini 2.5 models. --- .../llms/databricks/chat/transformation.py | 25 +++- .../test_databricks_chat_transformation.py | 116 ++++++++++++++++++ 2 files changed, 139 insertions(+), 2 deletions(-) diff --git a/litellm/llms/databricks/chat/transformation.py b/litellm/llms/databricks/chat/transformation.py index ba8c312ea51..a79dfe8a59b 100644 --- a/litellm/llms/databricks/chat/transformation.py +++ b/litellm/llms/databricks/chat/transformation.py @@ -264,6 +264,22 @@ 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. + """ + model_lower = model.lower() + if "claude" in model_lower: + return True + if "gemini-2" in model_lower: + return True + return False + def convert_anthropic_tool_to_databricks_tool( self, tool: Optional[AllAnthropicToolsValues] ) -> Optional[DatabricksTool]: @@ -367,19 +383,24 @@ 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, model=model, 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): + # output_config + adaptive thinking is an Anthropic-only feature. + if is_claude and AnthropicConfig._is_adaptive_thinking_model(model): mapped_effort: Optional[str] = None if isinstance(reasoning_effort_value, str): mapped_effort = REASONING_EFFORT_TO_OUTPUT_CONFIG_EFFORT.get(reasoning_effort_value) diff --git a/tests/test_litellm/llms/databricks/chat/test_databricks_chat_transformation.py b/tests/test_litellm/llms/databricks/chat/test_databricks_chat_transformation.py index cfdb76a97f4..e1a126298d7 100644 --- a/tests/test_litellm/llms/databricks/chat/test_databricks_chat_transformation.py +++ b/tests/test_litellm/llms/databricks/chat/test_databricks_chat_transformation.py @@ -8,6 +8,11 @@ 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 +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, @@ -416,3 +421,114 @@ def test_transform_request_keeps_parallel_tool_calls_for_claude(): )["messages"] assert len([m for m in result if m.get("role") == "assistant"]) == 1 + + +# --------------------------------------------------------------------------- +# 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. + + `max_tokens` is included by default to mirror real client behavior. Without + it the base-class `update_optional_params_with_thinking_tokens` helper + KeyErrors on pure pass-through models (a pre-existing issue orthogonal to + this fix — `is_thinking_enabled` returns True whenever `reasoning_effort` is + set, but the helper then assumes `optional_params["thinking"]` exists). + """ + non_default = {"reasoning_effort": reasoning_effort, "max_tokens": 1024} + 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_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 From d7b793c054a184ee9d7784559d959ba124ffe0d5 Mon Sep 17 00:00:00 2001 From: cohml <62400541+cohml@users.noreply.github.com> Date: Tue, 23 Jun 2026 12:02:20 -0400 Subject: [PATCH 2/3] fix(databricks,base_llm): narrow gemini-2.5 match and guard thinking-tokens helper. --- litellm/llms/base_llm/chat/transformation.py | 11 +++++-- .../llms/databricks/chat/transformation.py | 5 +++- .../test_databricks_chat_transformation.py | 30 +++++++++++++------ 3 files changed, 34 insertions(+), 12 deletions(-) diff --git a/litellm/llms/base_llm/chat/transformation.py b/litellm/llms/base_llm/chat/transformation.py index ab901a467e8..693ef38bd29 100644 --- a/litellm/llms/base_llm/chat/transformation.py +++ b/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 diff --git a/litellm/llms/databricks/chat/transformation.py b/litellm/llms/databricks/chat/transformation.py index a79dfe8a59b..e6e63392e9c 100644 --- a/litellm/llms/databricks/chat/transformation.py +++ b/litellm/llms/databricks/chat/transformation.py @@ -276,7 +276,10 @@ class DatabricksConfig(DatabricksBase, OpenAILikeChatConfig, AnthropicConfig): model_lower = model.lower() if "claude" in model_lower: return True - if "gemini-2" in model_lower: + # Match gemini-2-5 / gemini-2.5 only — not the broader 2.x range, which + # could catch hypothetical future 2.0/2.6/etc variants that may use a + # different reasoning contract. + if "gemini-2-5" in model_lower or "gemini-2.5" in model_lower: return True return False diff --git a/tests/test_litellm/llms/databricks/chat/test_databricks_chat_transformation.py b/tests/test_litellm/llms/databricks/chat/test_databricks_chat_transformation.py index e1a126298d7..02936e740f7 100644 --- a/tests/test_litellm/llms/databricks/chat/test_databricks_chat_transformation.py +++ b/tests/test_litellm/llms/databricks/chat/test_databricks_chat_transformation.py @@ -441,15 +441,8 @@ def test_transform_request_keeps_parallel_tool_calls_for_claude(): def _map_reasoning_effort(model: str, reasoning_effort, **extra_non_default): - """Run map_openai_params with reasoning_effort + optional extras. - - `max_tokens` is included by default to mirror real client behavior. Without - it the base-class `update_optional_params_with_thinking_tokens` helper - KeyErrors on pure pass-through models (a pre-existing issue orthogonal to - this fix — `is_thinking_enabled` returns True whenever `reasoning_effort` is - set, but the helper then assumes `optional_params["thinking"]` exists). - """ - non_default = {"reasoning_effort": reasoning_effort, "max_tokens": 1024} + """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, @@ -506,6 +499,25 @@ def test_gemini_2_5_pro_translates_to_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") From 2bff743f438c2abec4809ab4cfdcb742f2a9173f Mon Sep 17 00:00:00 2001 From: cohml <62400541+cohml@users.noreply.github.com> Date: Fri, 10 Jul 2026 12:37:59 -0400 Subject: [PATCH 3/3] refactor(databricks): drive anthropic-thinking routing via JSON flag. --- litellm/llms/databricks/chat/transformation.py | 17 ++++++++--------- .../model_prices_and_context_window_backup.json | 10 ++++++++++ litellm/types/utils.py | 1 + litellm/utils.py | 1 + model_prices_and_context_window.json | 10 ++++++++++ .../chat/test_databricks_chat_transformation.py | 7 +++++++ 6 files changed, 37 insertions(+), 9 deletions(-) diff --git a/litellm/llms/databricks/chat/transformation.py b/litellm/llms/databricks/chat/transformation.py index e6e63392e9c..8902f4cb061 100644 --- a/litellm/llms/databricks/chat/transformation.py +++ b/litellm/llms/databricks/chat/transformation.py @@ -273,15 +273,14 @@ class DatabricksConfig(DatabricksBase, OpenAILikeChatConfig, AnthropicConfig): GPT-5/GPT-OSS accept `reasoning_effort` natively and need no translation. """ - model_lower = model.lower() - if "claude" in model_lower: - return True - # Match gemini-2-5 / gemini-2.5 only — not the broader 2.x range, which - # could catch hypothetical future 2.0/2.6/etc variants that may use a - # different reasoning contract. - if "gemini-2-5" in model_lower or "gemini-2.5" in model_lower: - return True - return False + 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] diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 70b6b05e6ec..1cd41a40ed4 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -12690,6 +12690,7 @@ "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": { @@ -12709,6 +12710,7 @@ "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": { @@ -12728,6 +12730,7 @@ "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": { @@ -12747,6 +12750,7 @@ "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": { @@ -12766,6 +12770,7 @@ "supports_assistant_prefill": true, "supports_function_calling": true, "supports_reasoning": true, + "supports_anthropic_thinking_payload": true, "supports_tool_choice": true, "supports_output_config": true }, @@ -12786,6 +12791,7 @@ "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": { @@ -12805,6 +12811,7 @@ "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": { @@ -12824,6 +12831,7 @@ "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": { @@ -12841,6 +12849,7 @@ "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": { @@ -12858,6 +12867,7 @@ "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": { diff --git a/litellm/types/utils.py b/litellm/types/utils.py index 908f5b76424..10af3fc8866 100644 --- a/litellm/types/utils.py +++ b/litellm/types/utils.py @@ -151,6 +151,7 @@ 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] diff --git a/litellm/utils.py b/litellm/utils.py index 19c2fe16085..5f7985b84fd 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -5472,6 +5472,7 @@ def _get_model_info_helper( supports_url_context=_model_info.get("supports_url_context", None), supports_reasoning=_model_info.get("supports_reasoning", None), supports_adaptive_thinking=_model_info.get("supports_adaptive_thinking", 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), diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index b961c326625..fb9b192bac5 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -12690,6 +12690,7 @@ "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": { @@ -12709,6 +12710,7 @@ "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": { @@ -12728,6 +12730,7 @@ "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": { @@ -12747,6 +12750,7 @@ "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": { @@ -12766,6 +12770,7 @@ "supports_assistant_prefill": true, "supports_function_calling": true, "supports_reasoning": true, + "supports_anthropic_thinking_payload": true, "supports_tool_choice": true, "supports_output_config": true }, @@ -12786,6 +12791,7 @@ "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": { @@ -12805,6 +12811,7 @@ "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": { @@ -12824,6 +12831,7 @@ "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": { @@ -12841,6 +12849,7 @@ "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": { @@ -12858,6 +12867,7 @@ "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": { diff --git a/tests/test_litellm/llms/databricks/chat/test_databricks_chat_transformation.py b/tests/test_litellm/llms/databricks/chat/test_databricks_chat_transformation.py index 02936e740f7..59872e9962f 100644 --- a/tests/test_litellm/llms/databricks/chat/test_databricks_chat_transformation.py +++ b/tests/test_litellm/llms/databricks/chat/test_databricks_chat_transformation.py @@ -8,6 +8,7 @@ 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, @@ -20,6 +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 = [