mirror of
https://github.com/BerriAI/litellm.git
synced 2026-09-16 23:41:43 +00:00
898569b48f
Bedrock create_batch through the proxy needs three values that model-based routing was dropping or mis-resolving, so it failed after config resolution inside the Bedrock batch path.
CredentialLiteLLMParams re-validates a deployment's litellm_params and silently drops any undeclared field. aws_batch_role_arn (plus session token, assumed-role, and S3 output/region/endpoint/KMS fields) were undeclared, so the resolved credentials lost the IAM role and Bedrock raised 'AWS IAM role ARN is required'. Declare the AWS/S3 batch fields so they survive resolution.
base_llm_http_handler.create_batch calls the transform with optional_params={} and the deployment credentials in litellm_params, but the AWS signer reads only optional_params, so boto3 saw no keys ('Unable to locate credentials'). Merge litellm_params into the signing params (per-request optional_params still wins) for region resolution and signing.
get_deployment_credentials_with_provider never returned the deployment's underlying model, so the batch endpoints forwarded the proxy model-group name (alias) as the Bedrock modelId, which AWS rejects. Return the deployment model; the endpoints already forward it. This matches what the endpoint tests already assert.
Regression coverage: credential round-trip keeps the role and returns the model, and the transform signs with litellm_params credentials.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
|
||
|---|---|---|
| .. | ||
| access_control | ||
| batches | ||
| claude_code | ||
| coverage_registry | ||
| llm_translation | ||
| load | ||
| logging | ||
| management | ||
| mcp | ||
| quota_management | ||
| router | ||
| CLAUDE.md | ||
| conftest.py | ||
| CONTRIBUTING.md | ||
| e2e_config.py | ||
| e2e_db.py | ||
| e2e_http.py | ||
| junit_properties.py | ||
| lifecycle.py | ||
| models.py | ||
| proxy_client.py | ||
| pytest.ini | ||
| transport.py | ||