Merge pull request #39166 from BerriAI/litellm_bedrock_bearer_token_converse_crash

fix(bedrock): stop Converse crashing on bearer-token auth without SigV4 credentials

litellm/llms/bedrock/base_aws_llm.py

@@ -1442,7 +1442,7 @@ class BaseAWSLLM:
     @tracer.wrap()
     def get_request_headers(
         self,
-        credentials: Credentials,
+        credentials: Credentials | None,
         aws_region_name: str,
         extra_headers: dict | None,
         endpoint_url: str,
@@ -1469,9 +1469,13 @@ class BaseAWSLLM:
             try:
                 from botocore.auth import SigV4Auth
                 from botocore.awsrequest import AWSRequest
+                from botocore.exceptions import NoCredentialsError
             except ImportError:
                 raise ImportError("Missing boto3 to call bedrock. Run 'pip install boto3'.")
 
+            if credentials is None:
+                raise NoCredentialsError()
+
             # Filter headers for AWS signature calculation
             # AWS SigV4 only includes specific headers in signature calculation
             aws_signature_headers: Final = self._filter_headers_for_aws_signature(headers)

litellm/llms/bedrock/chat/converse_handler.py

@@ -1,4 +1,6 @@
 import json
+from collections.abc import Mapping
+from types import MappingProxyType
 from typing import Any, Final
 
 import httpx
@@ -24,6 +26,22 @@ from ..common_utils import BedrockError, _get_all_bedrock_regions
 from .invoke_handler import AWSEventStreamDecoder, MockResponseIterator, make_call
 
 
+def _sigv4_principal(credentials: Credentials | None) -> Mapping[str, str]:
+    if credentials is None:
+        return MappingProxyType({})
+    return MappingProxyType(
+        {
+            key: value
+            for key, value in (
+                ("aws_access_key_id", credentials.access_key),
+                ("aws_secret_access_key", credentials.secret_key),
+                ("aws_session_token", credentials.token),
+            )
+            if value is not None
+        }
+    )
+
+
 def make_sync_call(
     client: HTTPHandler | None,
     api_base: str,
@@ -95,7 +113,7 @@ class BedrockConverseLLM(BaseAWSLLM):
         stream,
         optional_params: dict,
         litellm_params: dict,
-        credentials: Credentials,
+        credentials: Credentials | None,
         logger_fn=None,
         headers={},
         client: AsyncHTTPHandler | None = None,
@@ -167,7 +185,7 @@ class BedrockConverseLLM(BaseAWSLLM):
         stream,
         optional_params: dict,
         litellm_params: dict,
-        credentials: Credentials,
+        credentials: Credentials | None,
         logger_fn=None,
         headers: dict = {},
         client: AsyncHTTPHandler | None = None,
@@ -331,7 +349,7 @@ class BedrockConverseLLM(BaseAWSLLM):
 
         litellm_params["aws_region_name"] = aws_region_name  # [DO NOT DELETE] important for async calls
 
-        credentials: Final[Credentials] = self.get_credentials(
+        credentials: Final[Credentials | None] = self.get_credentials(
             aws_access_key_id=aws_access_key_id,
             aws_secret_access_key=aws_secret_access_key,
             aws_session_token=aws_session_token,
@@ -368,19 +386,13 @@ class BedrockConverseLLM(BaseAWSLLM):
         # The Rust core owns the whole call for the subset it accepts. Ask
         # before transforming so whichever path runs emits pre_call once, and
         # hand down the credentials, region and endpoint this handler already
-        # resolved so both paths sign as the same principal.
+        # resolved so both paths sign as the same principal. Bearer-token auth
+        # resolves no SigV4 principal at all, and each path reads that token
+        # itself.
         rust_optional_params: Final = {  # mutable-ok: json.dumps in the bridge rejects a mappingproxy
             **optional_params,
-            **{  # mutable-ok: merged into its mutable parent above
-                key: value
-                for key, value in (
-                    ("aws_access_key_id", credentials.access_key),
-                    ("aws_secret_access_key", credentials.secret_key),
-                    ("aws_session_token", credentials.token),
-                    ("aws_region_name", aws_region_name),
-                )
-                if value is not None
-            },
+            **_sigv4_principal(credentials),
+            "aws_region_name": aws_region_name,
         }
         serves_via_rust: Final = rust_chat_completions_accepts(
             model=model,

tests/test_litellm/llms/bedrock/chat/test_bedrock_converse_handler.py

@@ -96,10 +96,8 @@ def _completion_kwargs(**overrides):
     return kwargs
 
 
-def _run(**overrides):
-    with patch.object(
-        BedrockConverseLLM, "get_credentials", return_value=RESOLVED_CREDENTIALS
-    ):
+def _run(*, credentials: Credentials | None = RESOLVED_CREDENTIALS, **overrides):
+    with patch.object(BedrockConverseLLM, "get_credentials", return_value=credentials):
         return BedrockConverseLLM().completion(**_completion_kwargs(**overrides))
 
 
@@ -360,7 +358,7 @@ async def test_async_completion_logs_pre_call_by_default():
 
 def _sync_client_returning_converse_response():
     client = MagicMock()
-    client.post = lambda **_kwargs: httpx.Response(
+    client.post.side_effect = lambda **_kwargs: httpx.Response(
         200,
         json=CONVERSE_RESPONSE,
         request=httpx.Request("POST", "https://bedrock-runtime.us-west-2.amazonaws.com"),
@@ -487,3 +485,31 @@ def test_post_call_is_not_logged_twice_when_the_sync_rust_call_declines():
     assert response.choices[0].message.content == "hi"
     assert len(calls["post_call"]) == 1
     assert "hi" in calls["post_call"][0]["original_response"]
+
+
+def test_bearer_token_auth_serves_when_boto3_resolves_no_sigv4_credentials(monkeypatch):
+    """With only `AWS_BEARER_TOKEN_BEDROCK` configured boto3 resolves no
+    credentials at all. Preparing the Rust handoff must not dereference that
+    None: the bearer token signs the request on its own."""
+    monkeypatch.setenv("AWS_BEARER_TOKEN_BEDROCK", "bedrock-bearer-token")
+    client = _sync_client_returning_converse_response()
+
+    response = _run(credentials=None, litellm_params={}, client=client)
+
+    assert response.choices[0].message.content == "hi"
+    sent_headers = client.post.call_args.kwargs["headers"]
+    assert sent_headers["Authorization"] == "Bearer bedrock-bearer-token"
+
+
+def test_the_rust_opt_in_needs_no_sigv4_principal():
+    """The core resolves the bearer token itself, so a bearer-only deployment
+    keeps its opt-in and the gate sees no aws_* credential keys to sign with."""
+    seen = _inject()
+
+    response = _run(credentials=None, api_key="bedrock-bearer-token")
+
+    assert response.choices[0].message.content == "hello from rust"
+    params = seen["call"][0]["optional_params"]
+    assert not {"aws_access_key_id", "aws_secret_access_key", "aws_session_token"} & params.keys()
+    assert params["aws_region_name"] == "us-east-1"
+    assert seen["call"][0]["api_key"] == "bedrock-bearer-token"

tests/test_litellm/llms/bedrock/test_base_aws_llm.py

@@ -15,6 +15,7 @@ from unittest.mock import MagicMock, patch
 from botocore.awsrequest import AWSPreparedRequest, AWSRequest
 from botocore.auth import SigV4Auth
 from botocore.credentials import Credentials
+from botocore.exceptions import NoCredentialsError
 
 import litellm
 from litellm.llms.bedrock.base_aws_llm import (
@@ -801,6 +802,23 @@ def test_get_request_headers_with_sigv4():
         assert result == mock_request.prepare.return_value
 
 
+def test_get_request_headers_without_credentials_or_bearer_token_raises_no_credentials():
+    """Bearer-token auth needs no SigV4 principal, so `credentials` may be None.
+    Reaching the SigV4 branch with neither must fail the way botocore always
+    has instead of signing with a missing principal."""
+    llm = BaseAWSLLM()
+
+    with patch.dict(os.environ, {}, clear=True), pytest.raises(NoCredentialsError):
+        llm.get_request_headers(
+            credentials=None,
+            aws_region_name="us-west-2",
+            extra_headers=None,
+            endpoint_url="https://api.example.com",
+            data='{"prompt": "test"}',
+            headers={"Content-Type": "application/json"},
+        )
+
+
 def test_sigv4_matches_rust_golden_vector():
     request = AWSRequest(
         method="POST",