Nadav Timor

> There's likely some bug in handling longer speculations. > > > > I've tried this branch on two datasets: ShareGPT and a synthetic dataset with realistic suffix overlap. And...

> > Does the absolute number of accepted tokens decrease, or only the acceptance _rate_? > > > > > > Also, please share your code for reproducing this. >...

> [@aflah02](https://github.com/aflah02) This is the global random seed. > > [sglang/python/sglang/srt/server_args.py](https://github.com/sgl-project/sglang/blob/39bb49d156f2319d2aec67c458c2db980bb0f4c3/python/sglang/srt/server_args.py#L362) > > Line 362 in [39bb49d](/sgl-project/sglang/commit/39bb49d156f2319d2aec67c458c2db980bb0f4c3) > > "--random-seed", > However, depending on your workloads (e.g., batch size), determinism...