fix: make Bedrock Mantle Responses routing data-driven per model

litellm/litellm_core_utils/get_llm_provider_logic.py

@@ -641,7 +641,7 @@ def _get_openai_compatible_provider_info(
             api_base,
             dynamic_api_key,
         ) = litellm.BedrockMantleChatConfig()._get_openai_compatible_provider_info(
-            api_base, api_key, litellm_params=litellm_params
+            api_base, api_key, litellm_params=litellm_params, model=model
         )
     elif custom_llm_provider == "nvidia_nim":
         # nvidia_nim is openai compatible, we just need to set this to custom_openai and have the api_base be https://api.endpoints.anyscale.com/v1

litellm/llms/bedrock_mantle/chat/transformation.py

@@ -23,6 +23,7 @@
 from litellm.types.llms.openai import AllMessageValues
 from litellm.types.router import GenericLiteLLMParams
 
+from ..common_utils import mantle_base_segment
 from ...openai_like.chat.transformation import OpenAILikeChatConfig
 
 
@@ -48,6 +49,7 @@ def _get_openai_compatible_provider_info(
         api_base: Optional[str],
         api_key: Optional[str],
         litellm_params: Optional[GenericLiteLLMParams] = None,
+        model: str | None = None,
     ) -> Tuple[Optional[str], Optional[str]]:
         region = (
             (litellm_params.aws_region_name if litellm_params else None)
@@ -57,10 +59,13 @@ def _get_openai_compatible_provider_info(
             or BEDROCK_MANTLE_DEFAULT_REGION
         )
         BaseAWSLLM._validate_aws_region_name(region)
+        # The base path segment is data-driven per model (use_openai_responses_path
+        # flag): gemma-4-* and gpt-5.x are served on /openai/v1, everything else on
+        # /v1. An explicit api_base still wins over the derived default.
         api_base = (
             api_base
             or get_secret_str("BEDROCK_MANTLE_API_BASE")
-            or f"https://bedrock-mantle.{region}.api.aws/v1"
+            or f"https://bedrock-mantle.{region}.api.aws/{mantle_base_segment(model, litellm.model_cost)}"
         )
         dynamic_api_key = self._resolve_bearer_token(api_key)
         return api_base, dynamic_api_key

litellm/llms/bedrock_mantle/common_utils.py

@@ -1,115 +1,38 @@
-"""
-Shared auth and region resolution for the Amazon Bedrock Mantle backends.
+"""Shared helpers for the Amazon Bedrock Mantle OpenAI-compatible provider.
 
-Mantle authenticates with a Bearer token when one is available
-(litellm_params.api_key, BEDROCK_MANTLE_API_KEY, or the standard
-AWS_BEARER_TOKEN_BEDROCK); otherwise it falls back to AWS SigV4 (service
-"bedrock") over the standard credential chain (IAM role / access key / profile /
-web identity). The Chat Completions and Responses backends share this behaviour
-through BedrockMantleAuthMixin so the two paths can never drift apart.
+Both helpers are pure functions of (model, model_cost) so the routing rules can be
+unit-tested without patching global state.
 """
 
-import re
-from typing import Tuple
-
-from botocore.exceptions import (
-    CredentialRetrievalError,
-    NoCredentialsError,
-    PartialCredentialsError,
-    ProfileNotFound,
-)
-
-from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
-from litellm.secret_managers.main import get_secret_str
-
-BEDROCK_MANTLE_DEFAULT_REGION = "us-east-1"
-
-# Standard Mantle host: https://bedrock-mantle.<region>.api.aws (group 1 = region).
-MANTLE_HOST_RE = re.compile(
-    r"^https?://bedrock-mantle\.([^/.]+)\.api\.aws", re.IGNORECASE
-)
-
-
-class BedrockMantleAuthMixin:
-    _aws_signer: BaseAWSLLM
-
-    @staticmethod
-    def _resolve_bearer_token(api_key: str | None) -> str | None:
-        return (
-            api_key
-            or get_secret_str("BEDROCK_MANTLE_API_KEY")
-            or get_secret_str("AWS_BEARER_TOKEN_BEDROCK")
-        )
-
-    @staticmethod
-    def _resolve_region(params: dict) -> str:
-        region = params.get("aws_region_name")
-        if region:
-            BaseAWSLLM._validate_aws_region_name(region)
-            return region
-        base = params.get("api_base") or get_secret_str("BEDROCK_MANTLE_API_BASE")
-        if base:
-            match = MANTLE_HOST_RE.match(base.rstrip("/"))
-            if match:
-                return match.group(1)
-        return (
-            get_secret_str("BEDROCK_MANTLE_REGION")
-            or get_secret_str("AWS_REGION_NAME")
-            or get_secret_str("AWS_REGION")
-            or BEDROCK_MANTLE_DEFAULT_REGION
-        )
 
-    def sign_request(
-        self,
-        headers: dict,
-        optional_params: dict,
-        request_data: dict,
-        api_base: str,
-        api_key: str | None = None,
-        model: str | None = None,
-        stream: bool | None = None,
-        fake_stream: bool | None = None,
-    ) -> Tuple[dict, bytes | None]:
-        bearer = self._resolve_bearer_token(api_key)
-        if not bearer:
-            # SigV4 path. Pin the credential-scope region to the region of the actual
-            # signing URL so the SigV4 scope and the URL host can never disagree, even
-            # when a stale api_base and aws_region_name point at different regions.
-            # Fall back to _resolve_region only for custom proxy hosts that do not
-            # match the standard Mantle URL pattern. Also drop any caller Authorization
-            # so _sign_request's restore-original-Authorization step cannot override
-            # the SigV4 header.
-            host_match = MANTLE_HOST_RE.match(api_base.rstrip("/"))
-            optional_params = {
-                **optional_params,
-                "aws_region_name": (
-                    host_match.group(1)
-                    if host_match
-                    else self._resolve_region({**optional_params, "api_base": api_base})
-                ),
-            }
-            headers = {k: v for k, v in headers.items() if k.lower() != "authorization"}
-        try:
-            return self._aws_signer._sign_request(
-                service_name="bedrock",
-                headers=headers,
-                optional_params=optional_params,
-                request_data=request_data,
-                api_base=api_base,
-                api_key=bearer,
-                model=model,
-                stream=stream,
-                fake_stream=fake_stream,
-            )
-        except (
-            NoCredentialsError,
-            PartialCredentialsError,
-            ProfileNotFound,
-            CredentialRetrievalError,
-        ) as e:
-            raise ValueError(
-                "Bedrock Mantle auth failed: no Bearer token and no usable AWS "
-                "credentials. Set BEDROCK_MANTLE_API_KEY (or AWS_BEARER_TOKEN_BEDROCK) "
-                "or pass api_key for Bearer auth, or provide AWS credentials "
-                "(IAM role / access key / profile / web identity) for SigV4."
-            ) from e
+def mantle_supports_responses(model: str | None, model_cost: dict) -> bool:
+    """Whether a Bedrock Mantle model can serve the native Responses API.
+
+    Purely data-driven from the model's price-map capability signal -- either
+    /v1/responses in supported_endpoints, or mode=responses -- both overridable
+    via register_model and proxy model_info, so onboarding a model is a JSON
+    change, never a code change. There is deliberately NO model-name match here:
+    capability is per-model, not per-family (openai.gpt-oss-120b supports
+    Responses while openai.gpt-oss-safeguard-120b does not, despite sharing the
+    gpt-oss substring), so a substring gate would be wrong. A model absent from
+    model_cost simply has no signal and returns False (chat-completions emulation).
+    """
+    entry = model_cost.get(f"bedrock_mantle/{model}", {})
+    if "/v1/responses" in (entry.get("supported_endpoints") or []):
+        return True
+    return entry.get("mode") == "responses"
+
+
+def mantle_base_segment(model: str | None, model_cost: dict) -> str:
+    """Return the base path segment for a Bedrock Mantle model's OpenAI surface.
+
+    Data-driven from the model's price-map use_openai_responses_path flag
+    (overridable via register_model / proxy model_info). Per the AWS model cards,
+    gpt-5.x and the google gemma-4-* family carry that flag and are served on the
+    /openai/v1 base (.../openai/v1/responses and .../openai/v1/chat/completions);
+    every other model including gpt-oss uses the standard /v1 base. The segment is
+    the base for the model's whole OpenAI-compatible surface, so both the chat and
+    responses configs derive from it -- there is no separate model-name rule.
+    """
+    entry = model_cost.get(f"bedrock_mantle/{model}", {})
+    return "openai/v1" if entry.get("use_openai_responses_path") is True else "v1"

litellm/model_prices_and_context_window_backup.json

@@ -42383,6 +42383,7 @@
         "max_output_tokens": 32768,
         "max_tokens": 32768,
         "mode": "chat",
+        "supported_endpoints": ["/v1/chat/completions", "/v1/responses"],
         "supports_function_calling": true,
         "supports_parallel_function_calling": true,
         "supports_reasoning": true,
@@ -42397,6 +42398,7 @@
         "max_output_tokens": 32768,
         "max_tokens": 32768,
         "mode": "chat",
+        "supported_endpoints": ["/v1/chat/completions", "/v1/responses"],
         "supports_function_calling": true,
         "supports_parallel_function_calling": true,
         "supports_reasoning": true,
@@ -42411,6 +42413,7 @@
         "max_output_tokens": 65536,
         "max_tokens": 65536,
         "mode": "chat",
+        "supported_endpoints": ["/v1/chat/completions"],
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_response_schema": true,
@@ -42424,6 +42427,7 @@
         "max_output_tokens": 65536,
         "max_tokens": 65536,
         "mode": "chat",
+        "supported_endpoints": ["/v1/chat/completions"],
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_response_schema": true,
@@ -42477,6 +42481,8 @@
         "max_output_tokens": 256000,
         "max_tokens": 256000,
         "mode": "chat",
+        "use_openai_responses_path": true,
+        "supported_endpoints": ["/v1/chat/completions", "/v1/responses"],
         "supports_function_calling": true,
         "supports_parallel_function_calling": false,
         "supports_reasoning": true,
@@ -42491,6 +42497,8 @@
         "max_output_tokens": 256000,
         "max_tokens": 256000,
         "mode": "chat",
+        "use_openai_responses_path": true,
+        "supported_endpoints": ["/v1/chat/completions", "/v1/responses"],
         "supports_function_calling": true,
         "supports_parallel_function_calling": false,
         "supports_reasoning": true,
@@ -42505,6 +42513,8 @@
         "max_output_tokens": 128000,
         "max_tokens": 128000,
         "mode": "chat",
+        "use_openai_responses_path": true,
+        "supported_endpoints": ["/v1/chat/completions", "/v1/responses"],
         "supports_function_calling": true,
         "supports_parallel_function_calling": false,
         "supports_reasoning": true,

litellm/utils.py

@@ -9070,34 +9070,26 @@ def _get_python_responses_api_config(
         elif litellm.LlmProviders.HOSTED_VLLM == provider:
             return litellm.HostedVLLMResponsesAPIConfig()
         elif litellm.LlmProviders.BEDROCK_MANTLE == provider:
-            # Mantle serves Responses on two upstream paths. A model takes the
-            # /openai/v1/responses path when its price-map entry declares
-            # use_openai_responses_path (data-driven, so a non-gpt-named frontier
-            # model can be onboarded by JSON alone), or, as a fallback needing no
-            # price-map entry, when its name matches the openai.gpt- frontier
-            # convention (minus gpt-oss) -- this keeps a future gpt-6 routing
-            # correctly before its entry loads. Any other model declared
-            # mode=responses takes the standard /v1/responses path. Everything
-            # else returns None and keeps the chat-completions emulation (see
-            # responses/main.py "config is None").
-            if not model:
+            # Both decisions are data-driven from the model's price-map entry, with
+            # no model-name logic. Capability (can it serve Responses?) comes from
+            # mantle_supports_responses (supported_endpoints / mode);
+            # chat-only models (gpt-oss safeguard, nvidia, ...) return None and keep
+            # the chat-completions emulation (responses/main.py "config is None").
+            # The wire path comes from mantle_base_segment, which reads the
+            # use_openai_responses_path flag: gpt-5.x and gemma-4-* on
+            # /openai/v1/responses, everything else (incl. gpt-oss) on
+            # /v1/responses.
+            from litellm.llms.bedrock_mantle.common_utils import (
+                mantle_base_segment,
+                mantle_supports_responses,
+            )
+
+            if not model or not mantle_supports_responses(model, litellm.model_cost):
                 return None
-            model_lower = model.lower()
-            entry = litellm.model_cost.get(f"bedrock_mantle/{model}", {})
-            on_openai_path = entry.get("use_openai_responses_path") is True
-            name_is_frontier = (
-                "openai.gpt-" in model_lower and "gpt-oss" not in model_lower
-            )
-            if on_openai_path or name_is_frontier:
-                return litellm.BedrockMantleResponsesAPIConfig(use_openai_path=True)
-            try:
-                if get_model_info(model, "bedrock_mantle").get("mode") == "responses":
-                    return litellm.BedrockMantleResponsesAPIConfig(
-                        use_openai_path=False
-                    )
-            except Exception:
-                pass
-            return None
+            return litellm.BedrockMantleResponsesAPIConfig(
+                use_openai_path=mantle_base_segment(model, litellm.model_cost)
+                == "openai/v1"
+            )
         return None
 
     @staticmethod

model_prices_and_context_window.json

@@ -42585,6 +42585,7 @@
         "max_output_tokens": 32768,
         "max_tokens": 32768,
         "mode": "chat",
+        "supported_endpoints": ["/v1/chat/completions", "/v1/responses"],
         "supports_function_calling": true,
         "supports_parallel_function_calling": true,
         "supports_reasoning": true,
@@ -42599,6 +42600,7 @@
         "max_output_tokens": 32768,
         "max_tokens": 32768,
         "mode": "chat",
+        "supported_endpoints": ["/v1/chat/completions", "/v1/responses"],
         "supports_function_calling": true,
         "supports_parallel_function_calling": true,
         "supports_reasoning": true,
@@ -42613,6 +42615,7 @@
         "max_output_tokens": 65536,
         "max_tokens": 65536,
         "mode": "chat",
+        "supported_endpoints": ["/v1/chat/completions"],
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_response_schema": true,
@@ -42626,6 +42629,7 @@
         "max_output_tokens": 65536,
         "max_tokens": 65536,
         "mode": "chat",
+        "supported_endpoints": ["/v1/chat/completions"],
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_response_schema": true,
@@ -42679,6 +42683,8 @@
         "max_output_tokens": 256000,
         "max_tokens": 256000,
         "mode": "chat",
+        "use_openai_responses_path": true,
+        "supported_endpoints": ["/v1/chat/completions", "/v1/responses"],
         "supports_function_calling": true,
         "supports_parallel_function_calling": false,
         "supports_reasoning": true,
@@ -42693,6 +42699,8 @@
         "max_output_tokens": 256000,
         "max_tokens": 256000,
         "mode": "chat",
+        "use_openai_responses_path": true,
+        "supported_endpoints": ["/v1/chat/completions", "/v1/responses"],
         "supports_function_calling": true,
         "supports_parallel_function_calling": false,
         "supports_reasoning": true,
@@ -42707,6 +42715,8 @@
         "max_output_tokens": 128000,
         "max_tokens": 128000,
         "mode": "chat",
+        "use_openai_responses_path": true,
+        "supported_endpoints": ["/v1/chat/completions", "/v1/responses"],
         "supports_function_calling": true,
         "supports_parallel_function_calling": false,
         "supports_reasoning": true,