fix(bedrock): correct modelInput format for Converse API batch models

[hztBUAA]

Summary

Fixes #21596

When litellm.create_file() / litellm.acreate_file() is called with custom_llm_provider="bedrock" for Converse API models (e.g. Amazon Nova), the uploaded JSONL batch file contained incorrect modelInput format:

1. image_url content blocks were passed through unchanged instead of being converted to Bedrock's {"image": {"format": "...", "source": {"bytes": "..."}}} format
2. Inference parameters (max_tokens, temperature, top_p) were left at the top level instead of being wrapped inside inferenceConfig

This caused Bedrock batch jobs to fail at inference time for Nova and other Converse API models.

Root cause: BedrockFilesConfig._map_openai_to_bedrock_params() only had a transformation path for Anthropic models. All other providers fell through to a raw passthrough else branch.

Fix: Added a new code path for Converse API providers (currently Nova) that routes through the existing AmazonConverseConfig.transform_request(), which correctly:

• Converts image_url blocks to Bedrock Converse image format
• Wraps max_tokens/temperature/top_p inside inferenceConfig
• Transforms messages to Converse content block structure

The fix uses a CONVERSE_INVOKE_PROVIDERS tuple that can be extended as other providers adopt the same schema. Anthropic and OpenAI-compatible model paths remain unchanged.

Test plan

• test_nova_text_only_uses_converse_format - Verifies Nova text-only requests get inferenceConfig wrapping
• test_nova_image_content_uses_converse_image_blocks - Verifies image_url blocks are converted to Bedrock image format
• test_anthropic_still_works_after_nova_fix - Regression test for Anthropic models
• test_openai_passthrough_still_works - Regression test for OpenAI-compatible models
• test_transform_openai_jsonl_content_to_bedrock_jsonl_content - Existing test still passes

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Feb 20, 2026 6:03am

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[CLAassistant]

All committers have signed the CLA.

[greptile-apps]

Greptile Summary

This PR fixes a critical bug in Bedrock batch file transformation for Nova (Converse API) models. The issue was that _map_openai_to_bedrock_params() had only Anthropic-specific and generic passthrough paths, causing Nova batch requests to use the wrong format.

Key Changes:

• Added CONVERSE_INVOKE_PROVIDERS = ("nova",) tuple to identify models using Converse API format
• Added new code path routing Nova requests through AmazonConverseConfig.transform_request()
• This correctly converts image_url content blocks to Bedrock's {"image": {"format": "...", "source": {"bytes": "..."}}} format
• Wraps inference parameters (max_tokens, temperature, top_p) inside inferenceConfig as required by Converse API

Test Coverage:

• 4 new tests validate Nova text-only and image transformations
• Regression tests confirm Anthropic and OpenAI-compatible models unchanged
• Tests verify proper inferenceConfig structure and image block format

Architecture:
The fix properly reuses existing transformation infrastructure. Nova models now follow the same transformation path as real-time Converse API requests, ensuring consistency between batch and real-time inference.

Confidence Score: 5/5

• This PR is safe to merge with minimal risk
• The fix correctly addresses the root cause by routing Nova models through the appropriate transformation config. Implementation properly reuses existing, battle-tested transformation logic from AmazonConverseConfig. Comprehensive tests cover all code paths including regressions. Changes are isolated to the files transformation layer with no impact on runtime request handling. The approach is extensible - additional Converse API providers can be easily added to the tuple.
• No files require special attention

Important Files Changed

Filename	Overview
litellm/llms/bedrock/files/transformation.py	Added Converse API transformation path for Nova models. Routes Nova batch requests through AmazonConverseConfig to properly convert image_url blocks and wrap inference params in inferenceConfig. Anthropic and OpenAI-compatible model paths remain unchanged.
tests/test_litellm/llms/bedrock/files/test_bedrock_files_transformation.py	Added 4 comprehensive tests: Nova text-only format validation, Nova image content transformation, Anthropic regression test, and OpenAI passthrough regression test. Tests verify correct inferenceConfig wrapping and image block transformations.
tests/test_litellm/llms/bedrock/files/expected_bedrock_batch_completions.jsonl	Updated expected output to include anthropic_beta: [] field in Anthropic responses, matching current transformation behavior.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[OpenAI JSONL Request] --> B[_map_openai_to_bedrock_params]
    B --> C{Determine Provider}
    C -->|anthropic| D[AmazonAnthropicClaudeConfig]
    C -->|nova| E[AmazonConverseConfig]
    C -->|Other providers| F[Passthrough]
    
    D --> D1[Transform messages]
    D --> D2[Keep max_tokens at top level]
    D --> D3[Add anthropic_version]
    D1 --> G[Bedrock Batch Record]
    D2 --> G
    D3 --> G
    
    E --> E1[Transform messages via<br/>bedrock_converse_messages_pt]
    E --> E2[Convert image_url to<br/>Bedrock image format]
    E --> E3[Wrap params in inferenceConfig]
    E1 --> G
    E2 --> G
    E3 --> G
    
    F --> F1[Keep messages unchanged]
    F --> F2[Keep params at top level]
    F1 --> G
    F2 --> G
    
    G --> H[Upload to S3]

Loading

_{Last reviewed commit: 27d695b}

[greptile-apps]

_{3 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[hung-phan]

This still have some issues and will be rejected by bedrock @hztBUAA . It adds additionalModelRequestFields

{"modelInput":{"messages":[{"role":"user","content":[{"text":"Explain the importance of cash flow management for businesses."}]}],"additionalModelRequestFields":{},"system":[{"text":"You are a business consultant with expertise in strategy, management, and economics."}],"inferenceConfig":{"maxTokens":1000}},"error":{"errorCode":400,"errorMessage":"Malformed input request: #: extraneous key [additionalModelRequestFields] is not permitted, reformat your input and try again.","expired":false,"retryable":false},"recordId":"request-45"}