# Configuration for PD disaggregation

**URL:** <https://discuss.vllm.ai/t/configuration-for-pd-disaggregation/2760>\
**Category:** General\
**Created:** [July 7, 2026, 2:49am UTC](https://discuss.vllm.ai/t/configuration-for-pd-disaggregation/2760 "2026-07-07T02:49:17Z")\
**Posts on this page:** 3\
**Page:** 1

<div class="post-metadata">

**Author:** ![SJTU-yys](https://yyz2.discourse-cdn.com/flex036/user_avatar/discuss.vllm.ai/sjtu-yys/32/1545_2.png) [@SJTU-yys](https://discuss.vllm.ai/u/SJTU-yys)\
**Post date:** [July 7, 2026, 2:49am UTC](https://discuss.vllm.ai/t/configuration-for-pd-disaggregation/2760/1 "2026-07-07T02:49:17Z")

</div>

现在vllm现在是通过max-num-batched-tokens间接控制chunk prefill的size的吗

---

<div class="post-metadata">

**Author:** ![RunLLM](https://yyz2.discourse-cdn.com/flex036/user_avatar/discuss.vllm.ai/runllm/32/233_2.png) [@RunLLM](https://discuss.vllm.ai/u/RunLLM)\
**Post date:** [July 7, 2026, 2:49am UTC](https://discuss.vllm.ai/t/configuration-for-pd-disaggregation/2760/2 "2026-07-07T02:49:31Z")

</div>

是的，vLLM通过max\_num\_batched\_tokens参数间接控制chunked prefill的大小。当输入序列长度超过max\_num\_batched\_tokens时，vLLM会自动将prefill阶段的输入切分为多个chunk进行处理，从而优化吞吐量和延迟。该参数决定了每个batch中可处理的最大token数，chunked prefill机制会根据这个上限自动分块，无需用户手动干预分块过程。  
详细说明可参考[官方优化文档](https://docs.vllm.ai/en/latest/configuration/optimization/#performance-tuning-with-chunked-prefill)和[源码实现](https://github.com/vllm-project/vllm/blob/main/vllm/config/scheduler.py)。

需要更详细的原理或调优建议吗？

Sources:

- [https://docs.vllm.ai/en/latest/configuration/optimization/#performance-tuning-with-chunked-prefill](https://docs.vllm.ai/en/latest/configuration/optimization/#performance-tuning-with-chunked-prefill)
- [https://github.com/vllm-project/vllm/blob/main/vllm/config/scheduler.py](https://github.com/vllm-project/vllm/blob/main/vllm/config/scheduler.py)

Hint: Mention @Herald in the post for followups.

---

<div class="post-metadata">

**Author:** ![SJTU-yys](https://yyz2.discourse-cdn.com/flex036/user_avatar/discuss.vllm.ai/sjtu-yys/32/1545_2.png) [@SJTU-yys](https://discuss.vllm.ai/u/SJTU-yys)\
**Post date:** [July 7, 2026, 4:31am UTC](https://discuss.vllm.ai/t/configuration-for-pd-disaggregation/2760/4 "2026-07-07T04:31:14Z")

</div>

@RunLLM 启用nixl作为kv connector的配置下，开启pd分离的情况下，p节点是每次推理完max\_num\_batched\_tokens个token再传输这些kvcache吗