[jperezdealgaba]

Summary

OpenAI request conversion copies caller-controlled vllm_xargs into SamplingParams.extra_args and only overlays typed kv_transfer_params / ec_transfer_params when those top-level fields are present. Planting a reserved key inside vllm_xargs (including a non-dict) therefore reaches EngineCore, where connectors call .get() and the process dies. This change rejects reserved transfer keys at the API merge, type-checks them on SamplingParams, and stores only dicts on Request so EngineCore does not crash if extra_args is mutated after construction.

Changes

• vllm/sampling_params.py: shared merge_request_extra_args that copies xargs, rejects reserved keys, then overlays typed fields; _verify_extra_args requires those keys to be dicts.
• OpenAI completion, chat, and responses protocols, plus STT transcription/translation: use the shared merge helper.
• vllm/v1/request.py: keep kv/ec transfer params only when they are dicts (do not raise in EngineCore).
• Rust HTTP merge helpers: same reserved-key reject for completions, chat, and generate.

Codepath coverage

• /v1/completions, /v1/chat/completions (including batch), /v1/responses, /v1/messages (via ChatCompletionRequest), /invocations, STT transcription/translation, Rust HTTP completions/chat/generate, offline SamplingParams, and EngineCore Request construction.

Duplicate-work check

Searched open and merged PRs for vllm_xargs, kv_transfer_params, and extra_args merge. #50353 copies vllm_xargs to avoid mutating the request dict but still overlays reserved keys and does not reject them. #55182 type-checks nested P2P role dicts only. Neither closes this sink.

Tests

• test_completion_rejects_reserved_xargs_key / chat / responses: reserved keys in vllm_xargs raise VLLMValidationError.
• test_merge_request_extra_args_overlays_typed_fields and test_completion_overlays_typed_kv_transfer_params: typed overlay still works; kv_cache_report_mode is not reserved.
• test_completion_does_not_mutate_request_xargs: overlay does not mutate the request dict.
• test_extra_args_rejects_non_dict_transfer_params: SamplingParams rejects non-dict transfer blobs.
• test_request_ignores_non_dict_transfer_params: EngineCore stores None instead of crashing.
• Rust: merge_kv_rejects_reserved_xargs_key, merge_ec_rejects_reserved_xargs_key, merge_kv_overlays_typed_field.

Commands:

• .venv/bin/python -m pytest tests/test_sampling_params.py tests/v1/test_request.py tests/entrypoints/openai/test_xargs_transfer_params.py tests/entrypoints/openai/test_protocol_responses.py tests/entrypoints/anthropic/test_anthropic.py tests/v1/ec_connector/unit/test_ec_connector_routing.py -v — 69 passed (subset of these files).
• .venv/bin/pre-commit run on the changed files — passed.
• cargo test -p vllm-server merge_kv / merge_ec — 3 passed.

AI assistance

This PR was developed with AI assistance.

Made with Cursor

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @jperezdealgaba.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork