# vLLM explains PCP and DCP for serving GLM-5.3 on 16 H200s

> Original title: GLM-5.3 Optimizations, Part 2: PCP, DCP, and Hybrid HiSparse on 16 H200s

- Source: vLLM Blog
- Published: 2026-10-23T00:00:00.000Z
- Bad Influence: https://badinfluence.news/items/a1ig5zwypwdv81gjxdxdhsemg
- Original: https://vllm.ai/blog/2026-10-23-glm53-part2-pcp-dcp-hisparse

## Summary

The vLLM team describes running GLM-5.3 on 16 H200s split into a prefill node and a decode node. Prefill Context Parallelism splits each prompt by token, and Decode Context Parallelism shards the KV cache so a 1M-token context fits. On the AgentX benchmark, PCP8 with DCP8 prefill and Hybrid HiSparse decode gives 1.7 to 3.2 times the per-user speed of a DP8 deployment at the same throughput.

## Why it matters

The post explains why PCP and DCP suit sparse attention and KV cache limits, with benchmarks that show where each configuration holds up.