unnecessary comment

Author: AmineDiro

trl/experimental/async_grpo/async_grpo_trainer.py

@@ -1266,8 +1266,6 @@ def _log_step_metrics(self) -> None:
         ## `_wall_clock` divides by `perf/step_s`: the whole step, rollout waits included  and says what fraction of the allocation actually became training.
         if self._step_forward_tokens > 0:
             mean_seq_len = self._step_seq_len_weighted / self._step_forward_tokens
-            # `get_text_config()` returns the config itself for text-only models; for a VLM it returns the text
-            # config, which is the one that carries `hidden_size`, `vocab_size`, ... (composite configs do not).
             flops_per_token = compute_flops_per_token(self.model.config.get_text_config(), int(mean_seq_len))
             world_size = self.accelerator.num_processes
             metrics["perf/forwarded_tok_s_fwd_bwd"].append((self._step_forward_tokens, fwd_bwd_s))