Generate this many candidates and return the top n by cumulative logprob. Integer in [1, 128]. Cross-field rule: best_of >= n (otherwise 400 invalid_request). Rejected with 400 unsupported_field when stream: true (mirrors OpenAI).

OpenAI frequency_penalty in [-2.0, 2.0]; out-of-range or non-numeric values yield 400 invalid_request.

SIE extension. true makes the model's reasoning public as reasoning_content on the message and, when streaming, on each delta. A model whose configuration declares no reasoning answers the field with nothing. Absent or false keeps reasoning private; usage is unchanged either way.

Return per-token logprobs. Boolean. When true, the chosen token's logprob (and optionally a top-N list via top_logprobs) rides on each choices[].logprobs entry.

SIE extension: non-empty served-name of a LoRA adapter declared in the model profile's lora_paths. Absent → the base model. Unknown name → 400 with param:"lora_adapter". The gateway forwards it to the worker as a top-level lora_path generation kwarg (SGLang then selects the adapter by served name; the sampling-params placement is an SGLang implementation detail, not part of the SIE wire contract).

Preferred output-token cap. Falls back to max_tokens when absent. When BOTH are omitted the gateway applies a default (1024, override via SIE_GATEWAY_DEFAULT_MAX_TOKENS) rather than rejecting — matching OpenAI, where this field is optional.

Legacy compatibility — max_completion_tokens wins when both are present. Optional; see max_completion_tokens for the behaviour when neither is supplied.

SGLang sampling_params.min_new_tokens: integer >= 0 capping how many tokens the model must emit before any stop condition can fire. Use to work around models that occasionally emit the stop token as the very first decoded token (e.g. Qwen3.6 thinking-off under greedy decode). Absent → sampler default (no minimum).

Number of candidate completions in [1, 128]. n>1 returns a multi-entry choices array; streaming with n>1 is supported (per-choice_index delta chunks + per-choice closure chunks ride before the single global [DONE]).

OpenAI parallel_tool_calls — boolean controlling whether the model may emit multiple tool calls per turn.

OpenAI presence_penalty in [-2.0, 2.0]; same validation as frequency_penalty.

Prompt-cache hint; plumbed onto the work envelope and ignored by the worker on the chat-completions surface.

Non-OpenAI repetition_penalty: float in (0.0, 2.0] (1.0 = no penalty). Absent → sampler default.

Routing affinity hint; same plumbing as prompt_cache_key.

Accepted and silently ignored (never logged, never forwarded).

SIE extension. A json_schema response_format constrains decoding and, as with OpenAI, is also shown to the model: it is appended to the system message as a # Response Formats section with the compact schema, unless a message already contains the schema as JSON. false leaves the prompt as sent (the grammar still applies). Absent: the serving profile decides (default on).

Optional signed 64-bit per-request sampling seed. Reproducibility is best effort, not guaranteed, and depends on the active generation backend and deployment configuration. Non-integer or out-of-range values reject with 400 invalid_request.

SSE streaming. When true, the response is a stream of chat.completion.chunk events terminated by data: [DONE]. For n > 1: per-choice_index delta chunks include a per-choice delta.role:"assistant" once per choice; per-choice closure chunks carry the finish_reason for that choice before the global [DONE]. Non-boolean values reject with 400 invalid_request.

Sampling temperature. Finite number >= 0; non-finite values reject.

Non-OpenAI top_k (Together / Fireworks / vLLM extension): integer >= 1. Absent → top-k disabled (model default).

Number of alternative top tokens to return alongside each chosen token's logprob. Integer in [0, 20]. Requires logprobs: true when > 0 (OpenAI rule; 400 invalid_request otherwise).

Nucleus sampling. Finite number in (0, 1].

OpenAI user — Sensitive PII. Accepted-and-dropped: debug-logged only, never persisted, never forwarded to the worker.

Epoch seconds.

Always "chatcmpl-<request_id>".

Always "chat.completion" on the non-streaming endpoint.