#1014·mamba

Mamba3.forward() 传递了一个 3-D 张量给 step(), 而 step() 要求 2-D — decode 路径引发 EinopsError

作者: JuiceB0xC0de创建于 2026年8月8日更新于 2026年8月12日

`Mamba3.forward()` routes to `self.step(...)` when `inference_params.seqlen_offset > 0`, but passes `u` through unchanged. `u` is `(batch, seqlen, hidden_dim)` at that point, while `step()` documents and requires `(batch, d_model)`. The sequence axis is never squeezed, so the cached-decode path raises before any kernel runs.

内容来源: state-spaces/mamba