Stricter request validation
The OpenAI-compatible server now refuses request fields it does not understand instead of silently ignoring them, and --prefill-chunk-tokens auto works on the server as well as the CLI.
Notable
- Return
400 unknown_parameternaming any undeclared top-level request field, so a misspelled option such asmax_tokenno longer generates under other settings. - Return
400 unsupported_valuefor OpenAI parameters the server cannot honour, includinglogit_bias,reasoning_effort,verbosity, and the legacyfunctionsandfunction_call; accept and ignoreuser,store,metadata,service_tier,prompt_cache_key, andsafety_identifier. - Refuse
response_formatother than{"type": "text"}explicitly rather than generating free text under a structured-output request. - Bound every echoed field name in bytes, so a key built from combining marks cannot be quoted back whole.
- Accept
--prefill-chunk-tokens autoon the server, where it runs at the 256 cap; help and rejections list the accepted values from one place. - Add nine M5 Pro 24 GB community benchmark rows from issue 170.
- Withdraw the
--experts-per-tokenexperiment from issue 167 after measurement; the decode path is unchanged.
This is a source-only release.