VERL is super slow compared to TRL, Pls help to understand the timeline trace

[Andcircle]

Currently, full finetune qwenVL2.5 7B, with multi image inputs, the prompt length average 3k tokens, response 500 tokens

I use A100 80GB,
Interestingly, 1 node 8 GPU is faster, once I increase the node number (2x8, 4x8 e.g.), the speed actually are even more slower.

Trying to visualize the timeline, but not sure is it normal or not, can anybody give some hints? really appreciated 感谢感谢

Here I have questions: 1. why there are huge gab between generate sequence, calculate log prob and actor update, it has 2 min gap in between! 2. why update actor takes so long, around 4~5 minutes?

[Andcircle]

Btw, I don't have InfiniBand, but I do have AWS EFA