Merge fec5b3d139 into 8d6b8dab3d

litellm/responses/bridge_check.py

@@ -1,3 +1,4 @@
+import re
 from collections.abc import Mapping, Sequence
 from typing import Final, cast  # noqa: TID251  # moved unchanged from litellm/main.py
 from urllib.parse import urlsplit
@@ -61,6 +62,22 @@ def keep_registered_mode(backend_key: str, shared_model_info: Mapping[str, objec
 
 
 _OPENAI_API_HOST: Final = "api.openai.com"
+# Databricks serves OpenAI GPT models as e.g. ``databricks-gpt-5-6-sol``: the minor version is
+# hyphenated, so the OpenAI version parser does not recognise it. The minor version is optional
+# so major-only names (``databricks-gpt-5``) map as well.
+_DATABRICKS_GPT_NAME: Final = re.compile(r"^databricks-gpt-(\d+)(?:-(\d+))?(?=-|$)")
+
+
+def _openai_gpt_name(match: "re.Match[str]") -> str:
+    major: Final = match.group(1)
+    minor: Final = match.group(2)
+    return f"gpt-{major}" if minor is None else f"gpt-{major}.{minor}"
+
+
+def _openai_name_for_databricks_model(model: str) -> str:
+    """``databricks-gpt-5-6-sol`` -> ``gpt-5.6-sol``; other names are returned unchanged."""
+    model_name: Final = model.split("/")[-1]
+    return _DATABRICKS_GPT_NAME.sub(_openai_gpt_name, model_name)
 
 
 def _is_openai_backed_api_base(api_base: str) -> bool:
@@ -243,6 +260,13 @@ def responses_api_bridge_check(
     on_foundry_openai_endpoint: Final = custom_llm_provider == "azure_ai" and azure_ai_supports_native_responses(
         model, api_base
     )
+    # Databricks proxies OpenAI GPT models to the real OpenAI backend, which enforces the
+    # reasoning+tools constraint, and serves them on its own /responses route as well.
+    on_databricks_gpt_endpoint: Final = (
+        custom_llm_provider == "databricks"
+        and OpenAIGPT5Config.is_model_gpt_5_model(model)
+        and not OpenAIGPT5Config.is_model_gpt_5_search_model(model)
+    )
     mantle_series_name: Final = (
         _mantle_responses_series_name(model) if custom_llm_provider == "bedrock_mantle" else None
     )
@@ -254,7 +278,7 @@ def responses_api_bridge_check(
     )
     gpt_series_name: Final = mantle_series_name or model
     on_constraint_enforcing_endpoint: Final = (
-        custom_llm_provider == "azure"
+        custom_llm_provider in ("azure", "databricks")
         or on_mantle_responses_endpoint
         or resolved_api_base == ""
         or _is_openai_backed_api_base(resolved_api_base)
@@ -266,7 +290,9 @@ def responses_api_bridge_check(
             foundry_chat_rejects_function_tools_while_reasoning(model, reasoning_effort)
             if on_foundry_openai_endpoint
             else (
-                OpenAIGPT5Config.is_model_gpt_5_4_plus_model(gpt_series_name)
+                OpenAIGPT5Config.is_model_gpt_5_4_plus_model(
+                    _openai_name_for_databricks_model(model) if on_databricks_gpt_endpoint else gpt_series_name
+                )
                 and (reasoning_effort is not None or on_constraint_enforcing_endpoint)
             )
         )
@@ -282,7 +308,7 @@ def responses_api_bridge_check(
             or (carries_tool_search and on_constraint_enforcing_endpoint)
             or (web_search_options is not None and on_mantle_responses_endpoint)
         )
-    ):
+    ) or (on_databricks_gpt_endpoint and model_info.get("mode") != "responses" and chat_rejects_function_tools):
         model_info["mode"] = "responses"
         model = model.replace("responses/", "")
 

tests/unit/test_main.py

@@ -1810,6 +1810,79 @@ def test_responses_api_bridge_check_azure_ai_without_foundry_responses_route_sta
     assert model_info.get("mode") != "responses"
 
 
+_DATABRICKS_FUNCTION_TOOL: Final = ({"type": "function", "function": {"name": "get_capital"}},)
+
+
+@pytest.mark.parametrize(
+    "model_name, reasoning_effort",
+    [
+        pytest.param("databricks-gpt-5-6-sol", None, id="gpt-5.6-unset-effort"),
+        pytest.param("databricks-gpt-5-6-luna", "high", id="gpt-5.6-explicit-effort"),
+        pytest.param("databricks-gpt-5-4", None, id="gpt-5.4-unset-effort"),
+        pytest.param("databricks-gpt-5-4-mini", "low", id="gpt-5.4-mini-explicit-effort"),
+        pytest.param("databricks/databricks-gpt-5-6-terra", None, id="provider-prefixed"),
+    ],
+)
+def test_responses_api_bridge_check_databricks_gpt_5_4_plus_tools_route_to_responses(model_name, reasoning_effort):
+    """Databricks forwards GPT models to OpenAI, whose Chat Completions rejects function tools while reasoning."""
+    from litellm.main import responses_api_bridge_check
+
+    model_info, model = responses_api_bridge_check(
+        model=model_name,
+        custom_llm_provider="databricks",
+        tools=_DATABRICKS_FUNCTION_TOOL,
+        reasoning_effort=reasoning_effort,
+    )
+
+    assert model == model_name
+    assert model_info.get("mode") == "responses"
+
+
+@pytest.mark.parametrize(
+    "model_name, tools, reasoning_effort",
+    [
+        pytest.param("databricks-gpt-5-6-sol", _DATABRICKS_FUNCTION_TOOL, "none", id="explicit-none"),
+        pytest.param("databricks-gpt-5-6-sol", None, None, id="no-tools"),
+        pytest.param("databricks-gpt-5-2", _DATABRICKS_FUNCTION_TOOL, None, id="pre-5.4"),
+        pytest.param("databricks-gpt-5", _DATABRICKS_FUNCTION_TOOL, None, id="major-only-pre-5.4"),
+        pytest.param("databricks-gpt-5-6-search-api", _DATABRICKS_FUNCTION_TOOL, None, id="search-model"),
+        pytest.param("databricks-gpt-oss-120b", _DATABRICKS_FUNCTION_TOOL, None, id="gpt-oss"),
+        pytest.param("databricks-claude-sonnet-4-5", _DATABRICKS_FUNCTION_TOOL, None, id="non-gpt"),
+    ],
+)
+def test_responses_api_bridge_check_databricks_without_rejected_tools_stays_chat(model_name, tools, reasoning_effort):
+    from litellm.main import responses_api_bridge_check
+
+    model_info, model = responses_api_bridge_check(
+        model=model_name,
+        custom_llm_provider="databricks",
+        tools=tools,
+        reasoning_effort=reasoning_effort,
+    )
+
+    assert model == model_name
+    assert model_info.get("mode") != "responses"
+
+
+@pytest.mark.parametrize(
+    "model_name, expected",
+    [
+        ("databricks-gpt-5-6-sol", "gpt-5.6-sol"),
+        ("databricks/databricks-gpt-5-4", "gpt-5.4"),
+        ("databricks-gpt-5-4-mini", "gpt-5.4-mini"),
+        ("databricks-gpt-5", "gpt-5"),
+        ("databricks-gpt-5-mini", "gpt-5-mini"),
+        ("databricks-gpt-6", "gpt-6"),
+        ("databricks-gpt-oss-120b", "databricks-gpt-oss-120b"),
+        ("databricks-claude-sonnet-4-5", "databricks-claude-sonnet-4-5"),
+    ],
+)
+def test_openai_name_for_databricks_model(model_name, expected):
+    from litellm.responses.bridge_check import _openai_name_for_databricks_model
+
+    assert _openai_name_for_databricks_model(model_name) == expected
+
+
 def test_responses_api_bridge_check_older_gpt_5_tools_without_reasoning_stays_chat():
     """Pre-5.4 GPT-5 names keep the old boundary: tools alone never bridge."""
     from litellm.main import responses_api_bridge_check