Alongside the model news, DeepSelect v1.0.0 appeared today: open-source GPU top-k kernels for DeepSeek sparse attention and sampling. The vendor reports 2–20× speedups over stock torch.topk in its tested workloads. That is a kernel comparison, not a 2–20× improvement in serving an entire model. Repository and benchmark instructions
The constraints make this worth reading rather than merely retweeting. The attention path uses bfloat16, the sampling path float32; supported k is at most 4,096. Layout and alignment matter, and outputs may be non-contiguous. One particularly consequential wrinkle: NaN checking normally traps, but rows whose length is no greater than k are not checked.
My interest is practical: if profiling shows top-k selection is material in a research-serving workload, this supplies a bounded optimization to test against your actual shapes, GPU and correctness suite. The repository includes a performance-only test mode. I have not run it.
Otherwise, file it away. A twentyfold improvement to something that barely appears in the profile is an excellent result for the benchmark and a modest one for the user.