fix(vllm): _build_client_args mutates caller's extra_body dict across calls

[feiiiiii5]

Problem

VLLM._build_client_args (and the duplicated AsyncVLLM._build_client_args) builds the client kwargs with:

extra_body = inference_kwargs.pop("extra_body", {})
extra_body.update(output_type_args)

dict.pop(key, {}) only returns a fresh dict when the key is absent. When the caller passes extra_body=my_dict, pop returns that same object, so the subsequent .update(output_type_args) mutates the caller's dict in place — adding structured_outputs (or response_format) to it permanently.

Concrete failure mode: a caller reusing one extra_body dict across calls — a common pattern when keeping a module-level default config and tweaking per call — silently leaks the previous call's structured_outputs constraint into later unconstrained calls. An "unconstrained" generation is then actually constrained to a stale regex/JSON schema from a prior call. Silent data corruption, hard to debug.

Reporter: #1931. The sibling sglang.py::_build_client_args has a different bug (single-call self-clobber via inference_kwargs.update(output_type_args)) and is being fixed separately in #1925 — that file is intentionally left untouched here.

Fix

In both VLLM._build_client_args and AsyncVLLM._build_client_args, wrap the pop in dict(...) so we own the dict before merging into it:

extra_body = dict(inference_kwargs.pop("extra_body", {}))
extra_body.update(output_type_args)

Shallow copy is sufficient: output_type_args only adds top-level keys (structured_outputs, response_format); nested values are not mutated. dict({}) is {} so the no-arg default path is unchanged.

Testing

• Added test_vllm_build_client_args_does_not_mutate_caller_extra_body (parametrized over sync_model/async_model) covering both failure modes:
  • Call 1 (Regex(r"[0-9]{3}") output type, extra_body=shared): caller's dict stays {"top_k": 5} (no structured_outputs leaked in), and the returned client_args["extra_body"] is a separate object that does carry the constraint.
  • Call 2 (None output type, same extra_body=shared): the returned client_args["extra_body"] is {"top_k": 5} (no stale structured_outputs from call 1).
• Confirmed red→green: reverting only src/outlines/models/vllm.py (keeping the new test) reproduces both failures on both parametrizations; reapplying the fix passes.
• Full tests/models/test_vllm.py suite: 21 passed (19 existing + 2 new parametrizations).
• ruff check src/outlines/models/vllm.py tests/models/test_vllm.py passes.
• (ruff format reports pre-existing unrelated drift in this file from prior commits; not addressed here to keep the diff surgical.)

Fixes #1931.

AI-Generated disclosure

Found via an AI-assisted code review pass (Claude Code) over src/outlines/models/ after issue #1931 was filed. I personally reproduced the cross-call mutation with a minimal repro (caller-provided extra_body dict reused across two calls, observing structured_outputs leak into the second, unconstrained call), verified the surgical dict(...) wrapper against both the structured_outputs path (regex/CFG/JSON) and the response_format path (JsonSchema) to confirm no regression, confirmed red→green by reverting only vllm.py, and ran the full test_vllm.py suite before submitting.

[feiiiiii5]

Gentle ping for review — @rlouf @brandonwillard.

This PR fixes a mutation bug where VLLM._build_client_args mutates the caller's extra_body dict across calls (because dict.pop(key, {}) returns the same object when the key is present). The fix uses copy.deepcopy on the popped dict before mutating.

Test results: All existing vLLM tests pass. The fix is minimal (3 lines) and surgical.

Happy to address any feedback.

[RobinPicard]

Great, thanks!

[github-actions]

📚 Documentation preview: https://dottxt-ai.github.io/outlines/pr-preview/pr-1933/

Preview updates automatically with each commit.