---
cited_by:
  - '[[thoughts/inference/kv-transfer#5-reading-order|reading-map §5]]'
date: '2026-11-11'
kind: pr
local:
  - file:///Users/aarnphm/workspace/vllm/vllm/distributed/kv_transfer/kv_connector/v1/mooncake/mooncake_connector.py
merged: '2026-09-09'
number: 52516
source:
  - https://github.com/vllm-project/vllm/blob/d2906091bfc579cebefe3d8e8fb9077397ce9882/vllm/distributed/kv_transfer/kv_connector/v1/mooncake/mooncake_connector.py
state: merged
tags:
  - vllm
  - vllm-pr
  - hetero-tp-kv-layout
title: '#52516 [Bugfix] Fix Mooncake heterogeneous TP with replicated GQA heads'
topic: hetero-tp-kv-layout
url: https://github.com/vllm-project/vllm/pull/52516
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/52516
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/52516.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
Prefill TP > `num_kv_heads` >= Decode TP, where producer ranks contain replicated head slices that must be gathered into the Decode KV block; · full dump: [[thoughts/inference/vllm-prs/52516|#52516]]

