fix(bedrock): keep the embeddings endpoint when a batch deployment overrides the model's mode

A deployment with model_info.mode set to batch overwrites the shared
cost-map entry's mode, so the job's model no longer read as an embedding
model. An entry with an output vector size is now also read as one.

litellm/llms/bedrock/batches/transformation.py

@@ -91,10 +91,13 @@ def openai_endpoint_for_bedrock_job_model(
     if not model_id:
         return "/v1/chat/completions"
     try:
-        mode: Final = get_cached_model_info()(model=model_id, custom_llm_provider="bedrock").get("mode")
+        model_info: Final = get_cached_model_info()(model=model_id, custom_llm_provider="bedrock")
     except ModelNotMappedError:
         return "/v1/chat/completions"
-    return "/v1/embeddings" if mode == "embedding" else "/v1/chat/completions"
+    is_embedding_model: Final = (
+        model_info.get("mode") == "embedding" or model_info.get("output_vector_size") is not None
+    )
+    return "/v1/embeddings" if is_embedding_model else "/v1/chat/completions"
 
 
 def titan_embedding_usage_from_batch_output(model_output: Mapping[str, object]) -> Usage | None:

tests/unit/llms/bedrock/batches/test_transformation.py

@@ -24,6 +24,7 @@ from botocore.auth import SigV4Auth
 from botocore.awsrequest import AWSRequest
 from botocore.credentials import Credentials
 
+import litellm
 from litellm.llms.bedrock.batches.transformation import BedrockBatchesConfig
 from litellm.types.utils import LlmProviders
 
@@ -940,6 +941,20 @@ def test_retrieve_response_endpoint_follows_the_job_models_mode(
     assert out.endpoint == expected_endpoint
 
 
+def test_retrieve_response_endpoint_survives_a_batch_deployment_overriding_the_models_mode(
+    config: BedrockBatchesConfig, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    job_model_id = "amazon.titan-embed-text-v2:0"
+    monkeypatch.setitem(litellm.model_cost, job_model_id, {**litellm.model_cost[job_model_id], "mode": "batch"})
+    out = config.transform_retrieve_batch_response(
+        model=None,
+        raw_response=_raw({"jobArn": ARN, "status": "InProgress", "modelId": job_model_id}),
+        logging_obj=MagicMock(),
+        litellm_params={},
+    )
+    assert out.endpoint == "/v1/embeddings"
+
+
 def test_retrieve_response_enriches_metadata(config):
     body = {
         "jobArn": ARN,