Fix: skip trainer config validation in bench mode

[shiweijiezero]

Bug

In bench mode, cluster.trainer_gpu_num is left at its default 0 because the cluster validator deliberately skips trainer GPU allocation for bench / explore / serve (these modes don't train) — see config_validator.py:244.

But the trainer config validator (config_validator.py:1168) still keeps bench in its whitelist. So for any local-model bench run (i.e. external_model.enable=false), the call chain reaches:

trinity/trainer/verl/verl_config.py:430
    if train_batch_size % (world_size // sp_size) != 0:
                              ↑ trainer_gpu_num = 0 → ZeroDivisionError

Repro: any yaml with mode: bench + a non-external model + engine_num × tensor_parallel_size = total GPU.

Fix

Drop bench from the trainer-config-check whitelist. Bench mode runs explorer-only and never touches the trainer, so its trainer parallelism config doesn't need validation. Same fast-path semantics as the existing external_model.enable=true check immediately below at line 1170.

After the fix, all 6 modes behave correctly:

mode	trainer GPU allocated	trainer config check
train	yes	yes
both	yes	yes
colocate	yes (1 GPU placeholder)	yes
bench	no	no (was: yes, broken)
explore	no	no (already)
serve	no	no (already)

Test

A local-model bench run (Qwen3.6-27B + frozen_lake_obscure eval, 1 node × 8 GPU, engine_num=4 × TP=2) reproduces the ZeroDivisionError on main and runs cleanly with this patch.

config.trainer.trainer_config is not accessed by any module outside config_validator.py (verified via repo-wide grep), so skipping synchronize_config() in bench mode has no downstream effect.