Function buildSpecDecodeExtras

src/llm/speculative.ts:262–291

(
  draftModel: string | undefined,
  lookahead: number,
  serverHint: SpecServerKind = 'auto',
)

export type SpecServerKind = 'lmstudio' | 'llamacpp' | 'vllm' | 'auto';

export function buildSpecDecodeExtras(
  draftModel: string | undefined,
  lookahead: number,
  serverHint: SpecServerKind = 'auto',
): Record<string, unknown> {
  const extras: Record<string, unknown> = {};
  switch (serverHint) {
    case 'lmstudio':
      if (draftModel) extras.draft_model = draftModel;
      break;
    case 'llamacpp':
      if (draftModel) extras.model_draft = draftModel;
      extras.n_draft = lookahead;
      break;
    case 'vllm':
      extras.num_speculative_tokens = lookahead;
      break;
    case 'auto':
    default:
      // Union of harmless fields — each server reads its own, ignores the rest.
      if (draftModel) {
        extras.draft_model = draftModel;   // LM Studio
        extras.model_draft = draftModel;   // llama.cpp
      }
      extras.n_draft = lookahead;             // llama.cpp
      extras.num_speculative_tokens = lookahead; // vLLM
      break;
  }
  return extras;
}