refactor(bedrock): remove the dead BedrockLLM invoke code path

[mateo-berri]

TLDR

Problem this solves:

• The pre-(Refactor) - migrate bedrock invoke to BaseLLMHTTPHandler class #8290 BedrockLLM invoke path has been dead since Feb 2025
• Dead path caused the feat: make streaming chunk size configurable in bedrock api #17357 regression: stream_chunk_size popped on the wrong path
• 951 lines of unreachable code invite the same mistake again

How it solves it:

• Deletes the BedrockLLM class from invoke_handler.py
• Live helpers (make_call, make_sync_call, stream decoders, AmazonCohereChatConfig) stay
• Remaining references now use BaseAWSLLM directly
• Ports the openai-route response and error tests onto the live config

Relevant issues

Follow-up to the RCA on stream_chunk_size leaking into invoke request bodies: PR #8290 moved Bedrock Invoke onto AmazonInvokeConfig (base_invoke_transformation.py) but left the old BedrockLLM completion path behind, PR #17357 then popped stream_chunk_size only on that dead path, and PR #30240 fixed the live path. This removes the dead path itself

Linear ticket

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have added meaningful tests (the openai-route response-parsing and error tests are ported onto the live AmazonBedrockOpenAIConfig, and the live-path regression tests from fix(bedrock): stop stream_chunk_size leaking into invoke request bodies #30240 are kept)
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review (Greptile reviews automatically once the PR is opened; only comment @greptileai to re-request a review after pushing changes)

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

Screenshots / Proof of Fix

Live proxy booted from this branch at a71d8d8 with real Bedrock calls (us.anthropic.claude-sonnet-5, us-east-1), config has one invoke-route and one converse-route deployment:

python litellm/proxy/proxy_cli.py --config qa_config.yaml --port 28471

Invoke route, non-streaming:

curl -s http://localhost:28471/v1/chat/completions -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"model": "sonnet5-invoke", "messages": [{"role": "user", "content": "Say INVOKE-OK and nothing else"}], "max_tokens": 20}'

{"id":"chatcmpl-489bb13e-fd05-4482-a6a9-017e2a82763e","created":1785383977,"model":"sonnet5-invoke","object":"chat.completion","choices":[{"finish_reason":"stop","index":0,"message":{"content":"INVOKE-OK","role":"assistant",...

Invoke route, streaming with stream_chunk_size in the request body (the exact trigger from the original regression; a leak into the invoke body would 400):

curl -s http://localhost:28471/v1/chat/completions -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"model": "sonnet5-invoke", "messages": [{"role": "user", "content": "Say INVOKE-STREAM-OK and nothing else"}], "max_tokens": 20, "stream": true, "stream_chunk_size": 4096}'

data: {"id":"chatcmpl-f3c57a91-720b-4208-bea3-03d6a6a212cd","object":"chat.completion.chunk","created":1785383979,"model":"sonnet5-invoke","choices":[{"index":0,"delta":{"role":"assistant","content":"IN"}}]}
data: {"id":"chatcmpl-f3c57a91-720b-4208-bea3-03d6a6a212cd","object":"chat.completion.chunk","created":1785383979,"model":"sonnet5-invoke","choices":[{"index":0,"delta":{"content":"VOKE-STREA"}}]}
data: {"id":"chatcmpl-f3c57a91-720b-4208-bea3-03d6a6a212cd","object":"chat.completion.chunk","created":1785383979,"model":"sonnet5-invoke","choices":[{"index":0,"delta":{"content":"M-OK"}}]}

Converse route, streaming with stream_chunk_size and non-streaming:

curl -s http://localhost:28471/v1/chat/completions -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"model": "sonnet5-converse", "messages": [{"role": "user", "content": "Say CONVERSE-OK and nothing else"}], "max_tokens": 20, "stream": true, "stream_chunk_size": 4096}'

data: {"id":"chatcmpl-095bb2bb-ac7d-483b-8aec-80b1fe929b3c","created":1785383986,"model":"sonnet5-converse","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"content":"CONVERSE-O","role":"assistant"}}],"provider_specific_fields":{}}
data: {"id":"chatcmpl-095bb2bb-ac7d-483b-8aec-80b1fe929b3c","created":1785383986,"model":"sonnet5-converse","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"content":"K"}}],"provider_specific_fields":{}}

curl -s http://localhost:28471/v1/chat/completions -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"model": "sonnet5-converse", "messages": [{"role": "user", "content": "Say CONVERSE-OK and nothing else"}], "max_tokens": 20}'

{"id":"chatcmpl-598dc34c-6729-44c6-b681-d977fddbc70e","created":1785383988,"model":"sonnet5-converse","object":"chat.completion","choices":[{"finish_reason":"stop","index":0,"message":{"content":"CONVERSE-OK","role":"assistant"}}],...

pytest tests/test_litellm/llms/bedrock/ passes at this head (1107 passed, 1 skipped), as do the two ported openai-route tests in tests/llm_translation/test_bedrock_completion.py

Type

🧹 Refactoring

Changes

litellm.completion has routed every Bedrock invoke request through AmazonInvokeConfig + base_llm_http_handler since PR #8290; BedrockLLM.completion and everything under it (process_response, convert_messages_to_prompt, async_completion, async_streaming, _async_anthropic_messages_completion) had no callers left. This deletes the class (951 lines) and its now-unused imports from invoke_handler.py, and drops the stale "TODO: DELETE FILE" module docstring since the file still hosts the live make_call/make_sync_call helpers, the AWS event-stream decoders, MockResponseIterator, and AmazonCohereChatConfig

The only live use of the class was common_utils.get_bedrock_chat_config calling the get_bedrock_invoke_provider staticmethod that BedrockLLM inherited from BaseAWSLLM; that call site now uses BaseAWSLLM directly, which is the identical code path as before. litellm/main.py and litellm/llms/bedrock/chat/__init__.py stop importing the class

Tests that exercised only the dead path are removed with it: the legacy-path re-chunk test in tests/test_litellm/llms/bedrock/chat/test_invoke_handler.py (the live-path make_call/make_sync_call chunk-size regression tests from PR #30240 remain), the shallow TestBedrockLLMSSLVerify duplicate of TestBaseAWSLLMSSLVerify, and the tests/llm_translation prompt-conversion test, since the live openai route passes messages through with no prompt conversion. The other two openai-route tests are ported onto the live config instead: test_bedrock_openai_response_parsing now goes through AmazonBedrockOpenAIConfig.transform_response and test_bedrock_openai_error_handling now covers the previously untested get_error_class BedrockError override. Two credential fixtures that only used BedrockLLM as a vessel for BaseAWSLLM.get_credentials now instantiate BaseAWSLLM

Deleting the untyped legacy code lowered the lint and type-check violation counts, so ruff-strict-budget.json, type-discipline-budget.json, and basedpyright-code-budget.json are ratcheted down accordingly (440 violations total)

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

[greptile-apps]

Greptile Summary

Removes the obsolete BedrockLLM invoke implementation while retaining the live streaming helpers and decoders

• Replaces the remaining inherited provider-detection call with BaseAWSLLM
• Removes stale BedrockLLM imports and tests
• Ports OpenAI response and error coverage to AmazonBedrockOpenAIConfig
• Ratchets static-analysis budgets to reflect the deleted legacy code

Confidence Score: 5/5

The PR appears safe to merge

No blocking failure remains

Important Files Changed

Filename	Overview
litellm/llms/bedrock/chat/invoke_handler.py	Deletes the unused BedrockLLM class and its imports while preserving live Bedrock streaming helpers, decoders, and configuration
litellm/llms/bedrock/common_utils.py	Uses BaseAWSLLM directly for the inherited Bedrock invoke-provider lookup
litellm/main.py	Removes the obsolete BedrockLLM import from the provider dispatch entrypoint
tests/llm_translation/test_bedrock_completion.py	Moves OpenAI-on-Bedrock response and error tests onto the live transformation config and removes dead-path coverage

_{Reviews (2): Last reviewed commit: "test(bedrock): port the openai-route inv..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[mateo-berri]

@greptileai

[codspeed]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_remove_dead_bedrock_invoke_path (a71d8d8) with litellm_internal_staging (47f1fb3)}