Feature request: support logprobs output

[dddgogogo]

Summary

Currently DFlash speculative decoding does not support returning logprobs in the API response. When requesting logprobs=True via the OpenAI-compatible API (/v1/chat/completions), the response either errors out or returns null logprobs.

Use Case

Log probabilities are essential for:

• Confidence estimation — filtering or flagging low-confidence generations
• Calibration & evaluation — measuring model uncertainty on benchmarks
• Routing & cascading — deciding whether to escalate to a larger model based on token-level confidence
• RLHF / reward modeling — computing per-token rewards requires logprobs from the policy model

Expected Behavior

{
  "logprobs": true,
  "top_logprobs": 5
}

should return per-token log probabilities in the response, consistent with the standard SGLang / vLLM behavior for non-speculative decoding.

Notes

• SGLang's native autoregressive decoding and MTP speculative decoding both support logprobs.
• For speculative decoding, logprobs from the target model's verification step would be the correct values to return (since those are the "true" probabilities that determine acceptance).