---
cited_by:
  - '[[thoughts/papers/elango2026latentmoeoptimalaccuracyflop#15-latent-moe-plus-situ|kimik3 §1.5]]'
  - '[[thoughts/inference/kimi-k3-prs#4-pr-reading-order|kimik3 §4]]'
date: '2026-11-11'
kind: pr
local:
  - file:///Users/aarnphm/workspace/vllm/vllm/envs.py
merged: '2026-08-03'
number: 50656
source:
  - https://github.com/vllm-project/vllm/blob/5df9999fcfaa72d9eb61348a789058fff805f142/vllm/models/kimi_k3/nvidia/model.py
  - https://github.com/vllm-project/vllm/blob/5df9999fcfaa72d9eb61348a789058fff805f142/vllm/envs.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#50656 [Kimi-K3] Add option to shard the shared expert instead of replicating'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/50656
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/50656
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/50656.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
Shard the shared-expert weights to avoid redundant work. In order to do so we need to re-distribute the activations, which adds an extra AllGather and ReduceScatter. · full dump: [[thoughts/inference/vllm-prs/50656|#50656]]

