should nproc-per-node equals to sequence_parallel_size?

[Edwardmark]

Great work. I have a question, should nproc-per-node equals to sequence_parallel_size when do open-sora inference using DSP?
Thanks in advance.

[oahzxl]

sp size should be as small as possible

[Edwardmark]

@oahzxl why sp size should be as small as possible?

[oahzxl]

because the essense of sequence parallel is to accomodate long sequence at the cost of extra communication cost. so if memory is enough, keep it low