fix(opd): score teacher logprobs at rollout temperature, not 0

[EazyReal]

Problem

slime/rollout/on_policy_distillation.py scores the teacher with max_new_tokens: 0 (logprobs of the tokens already in the sample, no generation). The payload hardcoded "temperature": 0.

OPD's KL is student (rollout sampling) vs teacher. Teacher logprobs at temperature 0 are a greedy distribution, so the penalty does not match the rollout policy whenever --rollout-temperature is not 0.

Change

"temperature": args.rollout_temperature,

max_new_tokens: 0 is unchanged. Unauthenticated / temp-0 rollouts still send 0.

Test plan

• One OPD run with --rollout-temperature ≠ 0: teacher /generate payload should carry that temperature (not 0)

[EazyReal]

@zhuzilin could you review this one? OPD teacher scoring was using temperature=0 while rollout samples use --rollout-temperature; this makes the teacher logprobs match the actual sampled distribution instead of silently scoring a greedy distribution.

[EazyReal]

@zhuzilin refreshed on latest main and checks are green. This scores OPD teacher logprobs at rollout temperature instead of forcing temperature 0, which keeps teacher scoring aligned with generation. Could you review when you have bandwidth?

[kkyyxhll]

@zhuzilin @EazyReal 这个pr是一个no-op操作，并且这个会引起混淆？默认teacher只需要用到input_token_logprobs,sgl中对于input的logp是temperature=1.0，即保留原始logp。因此不是原来的temperature=0.0 或者 这个pr增加的"temperature": args.rollout_temperature。

我新开了一个pr，#2319 ，辛苦review一下。非常感谢！

另外，sgl 的 SGLANG_RETURN_ORIGINAL_LOGPROB和temperature只影响output的logp，input的logp不受影响（opd在这，max_new_tokens=0）
下面是实际测试结果，不同temperature并不影响input的各种logp。
起serve：

python -m sglang.launch_server \
  --model /root/Qwen3-0.6B \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen25

temperature=1.0的payload

curl -s http://127.0.0.1:30010/generate -H 'Content-Type: application/json' -d '{
  "input_ids": [3838,374,279,6722,315,9625,30,576,6722,315,9625,374,12095],
  "sampling_params": {"temperature": 1, "max_new_tokens": 0},
  "return_logprob": true,
  "logprob_start_len": 0
}'

结果(temperature=1.0)

修改payload的temperature=2.0
结果如下(temperature=2.0)

修改payload的temperature=0.5
结果如下(temperature=0.5)

[EazyReal]

Thanks @kkyyxhll for catching the SGLang side — you are right that sampling_params.temperature does not affect input log-probs, so the change in this PR is a no-op. However, pinning it to 1.0 also leaves the teacher at q_1 while the student is trained with p_T, which optimizes KL(p_T || q_1) rather than the matched objective KL(p_T || q_T). The real-temperature fix therefore needs the larger SGLang input-logprob change; I opened draft #2325 with that implementation and regression coverage.