[训练耗时问题] 基于strands-agents魔改的generate训练耗时特别慢

基于example中的strands-agent，魔改并替换上自己的agent框架，训练集用dapo-math-17k.jsonl，模型用qwen3-30B，8卡，发现训练一个step需要2小时，这个耗时是合理的吗？我理解训练集中所有样本的rollout应该是集中起来一个list然后异步执行的吧，而且工具只定义了一个python code，轮次也没那么多，不应该这么久才对。不知道有经验的大佬们能否告知下是否是我魔改的问题，还是说这个训练速度是正常的。

另外我发现rollout_id==0的eval的log，需要运行很久才出来，理论上不应该很早就执行这段代码了吗？

# train loop.
# note that for async training, one can change the position of the sync operation(ray.get).
for rollout_id in range(args.start_rollout_id, args.num_rollout):
    # TODO extract the duplicated eval logic
    if args.eval_interval is not None and rollout_id == 0:
        ray.get(rollout_manager.eval.remote(rollout_id))

Start with the example strands-agent integration and the training loop around rollout_manager.eval.remote(rollout_id) and ray.get. Reproduce the reported timing with dapo-math-17k.jsonl, Qwen3-30B, and 8 cards, then trace rollout scheduling and the rollout_id==0 evaluation path. Done means identifying whether the delay is caused by the customization or expected execution behavior, with evidence from timings or logs.