[Bugfix] Surface SamplingParams.verify errors as client errors (HTTP 400)

[khluu]

Purpose

Fix the recurring randomized-Schemathesis failures in the H200 Entrypoints Integration (API Server OpenAI - Part 1) CI job: fuzzed POST /inference/v1/generate requests intermittently produced {"error":{"message":"","type":"InternalServerError","param":null,"code":500}} (main builds #84796, #85124, #85154 — the job now fails most runs).

Root cause

SamplingParams.verify() still raises plain ValueError for some request-caused validation failures (e.g. trace_decode_token_ids=[], which passes pydantic and the endpoint's msgspec pre-check). InputProcessor._validate_params runs lazily inside AsyncLLM.generate(), whose exception handling deliberately passes VLLMClientError through — but a plain ValueError falls into the generic branch and is wrapped in a bare EngineGenerateError (async_llm.py), which the entrypoints error mapper classifies as a server error. Result: HTTP 500 with an empty message instead of a 400, and the fuzz check correctly flags every 5xx.

Reproduced deterministically with the exact schemathesis-shrunk payload against the CI image (SmolVLM-256M, the schema-test server config): unpatched → empty 500 with the full EngineGenerateError chain in --log-error-stack; patched → 400 {"message":"trace_decode_token_ids must be a non-empty list.","type":"BadRequestError"}.

Changes

• InputProcessor._validate_params: convert ValueError from both params.verify() call sites (sampling and pooling) into VLLMValidationError at the validation boundary, so current and future plain-ValueError validations surface as 400s with their real message via the existing except VLLMClientError: raise pass-through.
• Migrate the two remaining plain ValueError raises in the sampling branch (return_sampling_mask temperature/top_k gates) to VLLMValidationError with parameter/value, matching the rest of the function.
• Regression test in tests/v1/engine/test_input_processor_trace_replay.py running the real verify() with trace_decode_token_ids=[] and asserting VLLMValidationError; verified it fails without the fix.

Duplicate-work check

Searched open PRs for trace_decode_token_ids, EngineGenerateError 500, and schemathesis inference generate; no open PR addresses this classification bug.

Test plan

• tests/v1/engine/test_input_processor_trace_replay.py: 6/6 passed (new test fails without the fix).
• Live reproduction on the CI image (vllm serve HuggingFaceTB/SmolVLM-256M-Instruct with the schema-test args): exact fuzzed payload returns 400 with message after the fix, empty 500 before.
• All applicable pre-commit hooks pass.
• Exact CI validation: isolated-branch run of entrypoints-integration-api-server-openai-part-1 (the schemathesis job) to be linked in comments.

AI-assistance disclosure

This change was prepared with AI assistance and reviewed before submission.

[khluu]

Exact CI validation is terminal green: isolated build #85158 at head aed5b715 (exact PR metadata, only the schemathesis job) — :nvidia: (H200) Entrypoints Integration (API Server OpenAI - Part 1) passed at 06:58 UTC, including the full test_openapi_stateless fuzz phase that failed in main builds #84796/#85124/#85154.

Caveat stated plainly: the fuzz is randomized, so a single green run is supporting rather than conclusive evidence — the conclusive part is the deterministic live reproduction in the PR description (exact shrunk payload: empty 500 before the fix, 400 BadRequestError with the real validation message after, on the CI image and schema-test server config).

[noooop]

Are you suggesting that exception handling exhibits a degree of nondeterminism?

Our entrypoints appear to be haunted. LOL....

e.g.

• [Bugfix][Frontend] Constrain Anthropic cache_salt to non-empty #50764

It succeeded on the retry.
┓( ´∀` )┏

[AndreasKaratzas]

Just closed #54402 in favor of this PR.