Hybrid HiSparse is a feature I built in vLLM with Lucas Wilkinson and Fares Obeid. It keeps sparse-attention KV cache on the GPU while there is room, then moves the full history to CPU memory as the GPU fills up. Only small hot buffers of recently selected tokens stay on the GPU. With less KV per request on the GPU, a deployment can run more long-context requests at once. For the design and benchmarks, see the vLLM blog post.
I wrote this calculator for that post. It estimates how many concurrent requests a model can serve with and without HiSparse. You choose the model, GPU, parallelism, sequence length and host memory. It also shows the smallest host pool that keeps CPU memory from limiting concurrency, and a breakdown of GPU and CPU memory use. The numbers are planning estimates, not guaranteed serving limits.
Besides GLM-5.3, the calculator covers GLM-5.3-Flash, DeepSeek-V4-Pro, DeepSeek-V4-Flash and DeepSeek-V4.1-Flash. HiSparse support for these is still in open vLLM pull requests, so their numbers come from the KV layouts in those PRs.