import json
import time


def completed_outputs(engine, prompts, samplings, stream=False):
    if not stream:
        for index, output in enumerate(engine.generate(prompts,samplings,use_tqdm=False)):
            yield index, output
        return
    identifiers = engine.enqueue(prompts,samplings,use_tqdm=False)
    processor = getattr(engine.llm_engine,'output_processor',None)
    if processor is not None:
        indices = {processor.request_states[request_id].external_req_id:index for index,request_id in enumerate(identifiers)}
    else:
        indices = {request_id:index for index,request_id in enumerate(identifiers)}
    finished = set()
    finished_tokens = 0
    start = last = time.monotonic()
    while engine.llm_engine.has_unfinished_requests():
        for output in engine.llm_engine.step():
            if not output.finished:
                continue
            if output.request_id not in indices or output.request_id in finished:
                raise ValueError('Unknown or duplicate completed request')
            finished.add(output.request_id)
            finished_tokens += sum(len(completion.token_ids) for completion in getattr(output,'outputs',[]))
            yield indices[output.request_id],output
        now = time.monotonic()
        if now-last >= 30:
            progress = {'type':'queue_progress','completed':len(finished),'queued':len(indices),
                        'unfinished':engine.llm_engine.get_num_unfinished_requests(),'seconds':round(now-start,1),
                        'completed_response_tokens':finished_tokens}
            if processor is not None:
                states = [processor.request_states[request_id] for request_id in identifiers
                          if request_id in processor.request_states]
                detokenizers = [getattr(state,'detokenizer',None) for state in states]
                if all(detokenizer is not None for detokenizer in detokenizers):
                    progress['unfinished_response_tokens'] = sum(d.num_output_tokens() for d in detokenizers)
                    progress['generated_response_tokens'] = finished_tokens + progress['unfinished_response_tokens']
            print(json.dumps(progress),flush=True)
            last = now
    if len(finished) != len(indices):
        raise RuntimeError('Engine stopped with incomplete requests')