# vllm/entrypoints/openai/completion/protocol.py
prompt: (
  list[Annotated[int, Field(ge=0)]]
  | list[list[Annotated[int, Field(ge=0)]]]
  | str
  | list[str]
  | None
) = None

def validate prompt and prompt embeds(cls, data):
  prompt = data.get("prompt")
  prompt embeds = data.get("prompt embeds")
  ...
  if prompt is empty and embeds is empty:
    raise VLLMValidationError(...)

# vllm/renderers/inputs/preprocess.py
def prompt to seq(prompt or prompts):
  if isinstance(prompt or prompts, (dict, str, bytes)) or (
    len(prompt or prompts) > 0 and is list of(prompt or prompts, int)
  ):
    return [prompt or prompts]

  return prompt or prompts

# vllm/renderers/online renderer.py
prompts = list[SingletonPrompt | bytes]()
if prompt input is not None:
  prompts.extend(prompt to seq(prompt input))
...
parsed prompts = [
  prompt if isinstance(prompt, bytes) else parse model prompt(model config, prompt)
  for prompt in prompts
]
return await renderer.render cmpl async(parsed prompts, tok params, ...)

# vllm/entrypoints/openai/completion/serving.py
generators: list[AsyncGenerator[RequestOutput, None]] = []
for i, engine input in enumerate(engine inputs):
  ...
  generator = self.engine client.generate(...)
  generators.append(generator)

result generator = merge async iterators(*generators)
num prompts = len(engine inputs)
...
final res batch: list[RequestOutput | None] = [None] * num prompts

{
 "model": "served-model",
 "prompt": ["x", "x", "x"],
 "max tokens": 1,
 "n": 1
}

{
 "model": "served-model",
 "prompt": [[1], [1], [1]],
 "max tokens": 1,
 "n": 1
}

{
 "model": "served-model",
 "prompt": "x",
 "max tokens": 1,
 "n": 1
}