Benchmark Comparation with Verl

Thank you so much for your awesome code implementation!😘😘

Have we compare the performance with Verl? Since Verl also supports VLM.🥹🥹

I do some DYI experiments: In 4 GPUs, 3B VL model; batch size = 8; response length = 1024

I find that the average time of one step in ROLL is 17s; the average time of one step in VerL is 30s

I wonder why ROLL is much faster than verl.

By the way, I find ROLL is also easier OOM than verl.🥹🥹

Below is the detailed script of ROLL and verl:

[roll.txt](https://github.com/user-attachments/files/21254169/roll.txt)
[verl.txt](https://github.com/user-attachments/files/21254168/verl.txt)

Start by reviewing the attached roll.txt and verl.txt scripts and compare their 4-GPU, 3B VLM, batch-size 8, and response-length 1024 configurations. Reproduce the reported step times and out-of-memory behavior, then document the cause of the performance and memory difference with supporting measurements.

Written by the indexing model from the issue text.