`eval_rollout_single_dataset` crashes when `generate_and_rm` returns a list

[GuanxingLu]

Description

In miles/rollout/sglang_rollout.py:571, the example logging assumes sample is a single Sample:

if do_print:
    logger.info(
        "eval_rollout_single_dataset example data: "
        f"{[str(sample.prompt) + sample.response]} "
        f"reward={sample.reward}"
    )

However, the isinstance(sample, list) branch a few lines below shows that generate_and_rm can return list[Sample] (multi-turn / multi-agent case). When it does, eval crashes.

Reproduction

bash examples/multi_agent/run-qwen3-30B-A3B-multi-agent.sh

Stack trace

File "miles/rollout/sglang_rollout.py", line 571, in eval_rollout_single_dataset
    f"{[str(sample.prompt) + sample.response]} "
AttributeError: 'list' object has no attribute 'prompt'

Suggested fix

example = sample[0] if isinstance(sample, list) else sample

[mathewjhan]

hi @GuanxingLu

I recently experienced + fixed this bug in my own slime fork 😆 , I was going to make a pr in slime and I can also upstream here too

update: #985