Although we have set temperature parameter in here, due to this
if do_sample:
sampling_params.use_beam_search = True
- Does it result in using beam_search of beam size 1 and ignoring the temperature parameter in vllm generation?
- In fact, alpaca_eval 0.6.2 with recent vllm does not work with
do_sample = True, related issue here.
- However, we tested successfully with the old VLM version of 0.4.1 and confirmed that the temperature was neglected and that generation is always the same.