---
cited_by:
  - '[[thoughts/papers/elango2026latentmoeoptimalaccuracyflop#15-latent-moe-plus-situ|kimik3 §1.5]]'
date: '2026-11-11'
kind: pr
local: []
merged: '2026-10-08'
number: 59591
source:
  - https://github.com/vllm-project/vllm/blob/d80ddaac73217770af1548f43e0cd7e68d4ec33d/vllm/models/kimi_k3/amd/linear.py
  - https://github.com/vllm-project/vllm/blob/d80ddaac73217770af1548f43e0cd7e68d4ec33d/vllm/models/kimi_k3/amd/latent_moe_runner.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#59591 [ROCm][Perf] Kimi-K3 Store only the current rank''s shards in latent MoE up-proj '
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/59591
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/59591
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/59591.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
On the ROCm path, the routed up-projection in Kimi-K3’s latent MoE is built as a `ReplicatedLinear` where each rank holds the full weight. However, the fused tail’s `ROCmLatentMoERunner` doesn’t need the full weight. Instead, each rank… · full dump: [[thoughts/inference/vllm-prs/59591|#59591]]

