Speculative decoding: candidate generators return a `q` they did not sample from, breaking losslessness with `do_sample=True`

该报错发生在 Transformers 库的投机解码(Speculative decoding)流程中,当候选生成器(candidate generator)返回的 logits 与其实际采样所用的分布不一致时触发。优先检查使用的候选生成器类型(如 DFlashTokenCandidateGener


![[SDK][Python] NaN/Inf floats nested in a pydantic model serialize to invalid JSON (bare NaN/Infinity)](https://www.chat-gpts.plus/wp-content/uploads/2026/08/16048-831ee9c1-768x403.jpg)

![`/config` returns `username` as an unawaited coroutine string (regression from #13749])](https://www.chat-gpts.plus/wp-content/uploads/2026/08/13758-6cb92da7-768x403.jpg)
![[Bug]: DashScope tiered pricing uses graduated slices instead of the request-size tier](https://www.chat-gpts.plus/wp-content/uploads/2026/08/34729-a68deaea-768x403.jpg)
![[Bug]: Nullable DailyTagSpend conflict columns cause duplicate inserts and database CPU exhaustion](https://www.chat-gpts.plus/wp-content/uploads/2026/08/34232-5fa4edae-768x403.jpg)
![[Bug]:amd mi308x gpu, vllm 0.27.0~0.27.1, rocm 7.2.3, Kimi-K2.7-Coder start fails:AssertionError: mla_gluon requires gfx950 (CDNA4), got gfx](https://www.chat-gpts.plus/wp-content/uploads/2026/08/51964-339b289b-768x403.jpg)
