Beam Search may generate more than need.

[TanZhendong]

Hi, thank you for your meticulous work.

In the loop of the beam search iterative process, we find that the condition for early stopping is only:

search-and-learn/src/sal/search/beam_search.py

Lines 151 to 153 in 1cc70e7

	# Early stopping if all beams are completed
	if len(active_beams) == 0:
	break

However, doing so may likely result in the final number of generated completion_beams exceeding config.n, which is not used for aggregation.

Moreover, for the sake of fairness, the number of beams used for aggregation should be the same as in best of n. Otherwise, their generation budget cannot be considered same.

I believe an additional stopping condition should be added as follows:

        # Early stopping if all beams are completed
        if len(active_beams) == 0:
            break
        
        # Early stopping if completions >= config.n
        if len(completed_beams) >= config.n:
            break

On one hand, it speeds up the beam search through early stopping, and on the other hand, it ensures that the number of generated results does not exceed config.n.