[Debugging] Create a stand alone trainer app

[JenniferWang]

Very similar to the stand-alone vllm app, this trainer app is introduced to make investigating trainer OOM faster. This could be very useful for single-node trainer because you can run it locally and the system metrics are much easier to obtain.

Test

Change the activation checkpointing config in apps/grpo/qwen3_32b.yaml and reproduce the OOM.

  activation_checkpoint:
    mode: selective
    selective_ac_option: op

The repro is blazingly fast :)

https://meta.wandb.io/jiyue/grpo-training/runs/8qe73q1b?nw=nwuserjiyue

[allenwang28]

nit, can we re-name this to rl_trainer? The reason is because we have sft, sft_v2 and then if we have a separate trainer it may confuse even further lol

[allenwang28]

just file name change, but looks good to me!

We should consider in the near future another location for this and everything that isn't GRPO, these should probably formulate the basis of many of our integration tests