Cohere 论文：细粒度 MoE 上投机解码在 batch size 1 仍有效
New @cohere paper finds that speculative decoding works well on fine-grained MoE models even at batch size 1, where prior work predicted small gains.
Cohere 新论文发现，投机解码在细粒度 MoE 模型上即使 batch size 为 1 也能带来明显收益，推翻了此前认为低 batch 下增益有限的判断。相邻 token 实际共享 20%–45% 的专家，远高于独立路由预期的 4%–12%；在 FP8 下 Qwen3-235B 的验证步骤速度达到普通解码步骤的 89%，而此前预测仅为 52%。研究测试了 30B 至 1T 参数的 10 个 MoE 模型，并在 H100 GPU 上做了性能剖析。
正文
New @cohere paper finds that speculative decoding works well on fine-grained MoE models even at batch size 1, where prior work predicted small gains.
Finds that in mixture-of-experts models, neighboring tokens share 20% to 45% of their experts, against 4% to 12% expected, which makes speculative decoding cheaper than predicted.
Speculative decoding lets a small drafter propose tokens that the big model checks in 1 pass. For MoE models, prior work said checking several tokens loads many extra experts, limiting gains at low batch sizes.
They tested 10 MoE models from 30B to 1T parameters and profiled them on H100 GPUs. Consecutive tokens reuse many experts, and small experts use GPU bandwidth well, so checking extra tokens costs little.
Adjacent tokens shared 20% to 45% of their experts, against 4% to 12% expected from independent routing. For Qwen3-235B in FP8, the verification pass ran at 89% of a normal decoding step's speed, against 52% predicted.
For low-batch MoE serving, such as agent workloads, benchmark speculative decoding on your hardware instead of ruling it out.